Datasets:

memray

/

krapivin

like 1

502528

Data mining with sparse grids using simplicial basis functions.

Recently we presented a new approach [18] to the classification problem arising in data mining. It is based on the regularization network approach but, in contrast to other methods which employ ansatz functions associated to data points, we use a grid in the usually high-dimensional feature space for the minimization process. To cope with the curse of dimensionality, we employ sparse grids [49]. Thus, only O(hn-1nd-1) instead of O(hn-d) grid points and unknowns are involved. Here d denotes the dimension of the feature space and gives the mesh size. We use the sparse grid combination technique [28] where the classification problem is discretized and solved on a sequence of conventional grids with uniform mesh sizes in each dimension. The sparse grid solution is then obtained by linear combination. In contrast to our former work, where d-linear functions were used, we now apply linear basis functions based on a simplicial discretization. This allows to handle more dimensions and the algorithm needs less operations per data point.We describe the sparse grid combination technique for the classification problem, give implementational details and discuss the complexity of the algorithm. It turns out that the method scales linearly with the number of given data points. Finally we report on the quality of the classifier built by our new method on data sets with up to 10 dimensions. It turns out that our new method achieves correctness rates which are competitive to that of the best existing methods.

INTRODUCTION Data mining is the process of nding patterns, relations and trends in large data sets. Examples range from scien- tic applications like the post-processing of data in medicine or the evaluation of satellite pictures to nancial and commercial applications, e.g. the assessment of credit risks or the selection of customers for advertising campaign letters. For an overview on data mining and its various tasks and approaches see [5, 12]. In this paper we consider the classication problem arising in data mining. Given is a set of data points in a d-dimensional feature space together with a class label. From this data, a classier must be constructed which allows to predict the class of any newly given data point for future decision making. Widely used approaches are, besides others, decision tree induction, rule learning, adaptive multivariate regression splines, neural networks, and support vector machines. Interestingly, some of these techniques can be interpreted in the framework of regularization networks [21]. This approach allows a direct description of the most important neural networks and it also allows for an equivalent description of support vector machines and n-term approximation schemes [20]. Here, the classication of data is interpreted as a scattered data approximation problem with certain additional regularization terms in high-dimensional spaces. In [18] we presented a new approach to the classication problem. It is also based on the regularization network approach but, in contrast to the other methods which employ mostly global ansatz functions associated to data points, we use an independent grid with associated local ansatz functions in the minimization process. This is similar to the numerical treatment of partial dierential equations. Here, a uniform grid would result in O(h d denotes the dimension of the feature space and gives the mesh size. Therefore the complexity of the problem would grow exponentially with d and we encounter the curse of dimensionality. This is probably the reason why conventional grid-based techniques are not used in data mining up to now. However, there is the so-called sparse grids approach which allows to cope with the complexity of the problem to some extent. This method has been originally developed for the solution of partial dierential equations [2, 8, 28, 49] and is now used successfully also for integral equations [14, 27], interpolation and approximation [3, 26, 39, 42], eigenvalue problems [16] and integration problems [19]. In the information based complexity community it is also known as 'hyper- bolic cross points' and the idea can even be traced back to [41]. For a d-dimensional problem, the sparse grid approach employs only O(h 1 points in the dis- cretization. The accuracy of the approximation however is nearly as good as for the conventional full grid methods, provided that certain additional smoothness requirements are fullled. Thus a sparse grid discretization method can be employed also for higher-dimensional problems. The curse of the dimensionality of conventional 'full' grid methods affects sparse grids much less. In this paper, we apply the sparse grid combination technique [28] to the classication problem. For that the regularization network problem is discretized and solved on a certain sequence of conventional grids with uniform mesh sizes in each coordinate direction. In contrast to [18], where d-linear functions stemming from a tensor-product approach were used, we now apply linear basis functions based on a simplicial discretization. In comparison, this approach allows the processing of more dimensions and needs less operations per data point. The sparse grid solution is then obtained from the solutions on the dierent grids by linear combination. Thus the classier is build on sparse grid points and not on data points. A discussion of the complexity of the method gives that the method scales linearly with the number of instances, i.e. the amount of data to be classied. Therefore, our method is well suited for realistic data mining applications where the dimension of the feature space is moderately high (e.g. after some preprocessing steps) but the amount of data is very large. Furthermore the quality of the classier build by our new method seems to be very good. Here we consider standard test problems from the UCI repository and problems with huge synthetical data sets in up to 10 dimensions. It turns out that our new method achieves correctness rates which are competitive to those of the best existing methods. Note that the combination method is simple to use and can be parallelized in a natural and straightforward way. The remainder of this paper is organized as follows: In Section 2 we describe the classication problem in the frame-work of regularization networks as minimization of a (qua- dratic) functional. We then discretize the feature space and derive the associated linear problem. Here we focus on grid-based discretization techniques. Then, we introduce the sparse grid combination technique for the classication problem and discuss its properties. Furthermore, we present a new variant based on a discretization by simplices and discuss complexity aspects. Section 3 presents the results of numerical experiments conducted with the sparse grid combination method, demonstrates the quality of the classier build by our new method and compares the results with the ones from [18] and with the ones obtained with dierent forms of SVMs [33]. Some nal remarks conclude the paper. 2. THE PROBLEM Classication of data can be interpreted as traditional scattered data approximation problem with certain additional regularization terms. In contrast to conventional scattered data approximation applications, we now encounter quite high-dimensional spaces. To this end, the approach of regularization networks [21] gives a good framework. This approach allows a direct description of the most important neural networks and it also allows for an equivalent description of support vector machines and n-term approximation schemes [20]. Consider the given set of already classied data (the training d Rg Assume now that these data have been obtained by sampling of an unknown function f which belongs to some function space V dened over R d . The sampling process was disturbed by noise. The aim is now to recover the function f from the given data as good as possible. This is clearly an ill-posed problem since there are innitely many solutions possible. To get a well-posed, uniquely solvable problem we have to assume further knowledge on f . To this end, regularization theory [43, 47] imposes an additional smoothness constraint on the solution of the approximation problem and the regularization network approach considers the variational problem min f2V with Here, C(:; :) denotes an error cost function which measures the interpolation error and (f) is a smoothness functional which must be well dened for . The rst term enforces closeness of f to the data, the second term enforces smoothness of f and the regularization parameter balances these two terms. Typical examples are and 2 with with r denoting the gradient and the Laplace operator. The value of can be chosen according to cross-validation techniques [13, 22, 37, 44] or to some other principle, such as structural risk minimization [45]. Note that we nd exactly this type of formulation in the case scattered data approximation methods, see [1, 31], where the regularization term is usually physically motivated. 2.1 Discretization We now restrict the problem to a nite dimensional sub-space . The function f is then replaced by Here the ansatz functions f' j g N should span VN and preferably should form a basis for VN . The coecients f j g N j=1 denote the degrees of freedom. Note that the restriction to a suitably chosen nite-dimensional subspace involves some additional regularization (regularization by discretization) which depends on the choice of VN . In the remainder of this paper, we restrict ourselves to the choice and for some given linear operator P . This way we obtain from the minimization problem a feasible linear system. We thus have to minimize with fN in the nite dimensional space VN . We plug (2) into (4) and obtain after dierentiation with respect to k , @k 'k This is equivalent to In matrix notation we end up with the linear system Here C is a square N N matrix with entries C is a rectangular N M matrix with entries B . The vector y contains the data labels y i and has length M . The unknown vector contains the degrees of freedom j and has length N . Depending on the regularization operator we obtain different minimization problems in VN . For example if we use the gradient (fN in the regularization expression in (1) we obtain a Poisson problem with an additional term which resembles the interpolation problem. The natural boundary conditions for such a partial dierential equation are Neumann conditions. The discretization (2) gives us then the linear system (7) where C corresponds to a discrete Laplacian. To obtain the classier fN we now have to solve this system. 2.2 Grid based discrete approximation Up to now we have not yet been specic what nite- dimensional subspace VN and what type of basis functions we want to use. In contrast to conventional data mining approaches which work with ansatz functions associated to data points we now use a certain grid in the attribute space to determine the classier with the help of these grid points. This is similar to the numerical treatment of partial dierential equations. For reasons of simplicity, here and in the the remainder of this paper, we restrict ourself to the case x i= [0; 1] d . This situation can always be reached by a proper rescaling of the data space. A conventional nite element discretization would now employ an equidistant grid n with mesh size for each coordinate direction, where n is the renement level. In the following we always use the gradient in the regularization expression (3). Let j denote the d . A nite element method with piecewise d-linear, i.e. linear in each dimension, test- and trial-functions n;j (x) on grid now would give n;j n;j (x) and the variational procedure (4) - (6) would result in the discrete linear system of size and matrix entries corresponding to (7). Note that fn lives in the space Vn := spanf n;j The discrete problem (8) might in principle be treated by an appropriate solver like the conjugate gradient method, a multigrid method or some other suitable ecient iterative method. However, this direct application of a nite element discretization and the solution of the resulting linear system by an appropriate solver is clearly not possible for a d-dimensional problem if d is larger than four. The number of grid points is of the order O(h d the best case, the number of operations is of the same order. Here we encounter the so-called curse of dimensionality: The complexity of the problem grows exponentially with d. At least for d > 4 and a reasonable value of n, the arising system can not be stored and solved on even the largest parallel computers today. 2.3 The sparse grid combination technique Therefore we proceed as follows: We discretize and solve the problem on a certain sequence of grids l l 1 ;:::;l d with uniform mesh sizes h in the t-th coordinate direction. These grids may possess dierent mesh sizes for dierent coordinate directions. To this end, we consider all grids l with For the two-dimensional case, the grids needed in the combination formula of level 4 are shown in Figure 1. The - nite element approach with piecewise d-linear test- and trial- functions l;j (x) := d Y on grid l now would give f l l d l;j l;j (x) and the variational procedure (4) - (6) would result in the discrete system l with the matrices M; and the unknown vector d. We then solve these c Figure 1: Combination technique with level two dimensions problems by a feasible method. To this end we use here a diagonally preconditioned conjugate gradient algorithm. But also an appropriate multigrid method with partial semi- coarsening can be applied. The discrete solutions f l are contained in the spaces of piecewise d-linear functions on grid l . Note that all these problems are substantially reduced in size in comparison to (8). Instead of one problem with size nd ), we now have to deal with problems of size dim(V l Moreover, all these problems can be solved independently, which allows for a straightforward parallelization on a coarse grain level, see [23]. There is also a simple but eective static load balancing strategy available [25]. Finally we linearly combine the results f l l;j l;j (x); from the dierent grids l as follows: f (c) The resulting function f (c) n lives in the sparse grid space This space has dim(V It is spanned by a piecewise d-linear hierarchical tensor product basis, see [8]. Note that the summation of the discrete functions from dierent spaces V l in (13) involves d-linear interpolation which resembles just the transformation to a representation in the hierarchical basis. For details see [24, 28, 29]. However we never explicitly assemble the function f (c) but keep instead the solutions f l on the dierent grids l which arise in the combination formula. Now, any linear operation F on f (c) can easily be expressed by means of the combination Figure 2: Two-dimensional sparse grid (left) and three-dimensional sparse grid acting directly on the functions f l , i.e. F(f (c) l 1 +:::+l d =n+(d 1) q F(f l Therefore, if we now want to evaluate a newly given set of data points f ~ (the test or evaluation set) by ~ we just form the combination of the associated values for f l according to (13). The evaluation of the dierent f l in the test points can be done completely in parallel, their summation needs basically an all-reduce/gather operation. For second order elliptic PDE model problems, it was proven that the combination solution f (c) n is almost as accurate as the full grid solution fn , i.e. the discretization error jje (c) provided that a slightly stronger smoothness requirement on f than for the full grid approach holds. We need the seminorm 1 to be bounded. Furthermore, a series expansion of the error is necessary for the combination technique. Its existence was shown for PDE model problems in [10]. The combination technique is only one of the various methods to solve problems on sparse grids. Note that there exist also nite dierence [24, 38] and Galerkin nite element approaches [2, 8, 9] which work directly in the hierarchical product basis on the sparse grid. But the combination technique is conceptually much simpler and easier to implement. Moreover it allows to reuse standard solvers for its dierent subproblems and is straightforwardly parallelizable. 2.4 Simplicial basis functions So far we only mentioned d-linear basis functions based on a tensor-product approach, this case was presented in detail in [18]. But on the grids of the combination technique linear basis functions based on a simplicial discretization are also possible. For that we use the so-called Kuhn's triangulation [15, 32] for each rectangular block, see Figure 3. Now, the summation of the discrete functions for the dierent spaces l in (13) only involves linear interpolation. Table 1: Complexities of the storage, the assembly and the matrix-vector multiplication for the dierent matrices arising in the combination method on one grid l for both discretization approaches. C l and G l can be stored together in one matrix structure. d-linear basis functions linear basis functions l B l storage O(3 d N) O(3 d N) O(2 d M) O((2 d assembly O(3 d N) O(d 2 2d M) O(d 2 d M) O((2 d mv-multiplication O(3 d N) O(3 d N) O(2 d M) O((2 d Figure 3: Kuhn's triangulation of a three-dimensional unit cube The theroetical properties of this variant of the sparse grid technique still has to be investigated in more detail. However the results which are presented in section 3 warrant its use. We see, if at all, just slightly worse results with linear basis functions than with d-linear basis functions and we believe that our new approach results in the same approximation order. Since in our new variant of the combination technique the overlap of supports, i.e. the regions where two basis functions are both non-zero, is greatly reduced due to the use of a simplicial discretization, the complexities scale signicantly better. This concerns both the costs of the assembly and the storage of the non-zero entries of the sparsely populated matrices from (8), see Table 1. Note that for general operators P the complexities for C l scale with O(2 d N ). But for our choice of zero-entries arise, which need not to be considered, and which further reduce the complex- ities, see Table 1 (right), column C l . The actual iterative solution process (by a diagonally preconditioned conjugate gradient method) scales independent of the number of data points for both approaches. Note however that both the storage and the run time complexities still depend exponentially on the dimension d. Presently, due to the limitations of the memory of modern workstations (512 MByte - 2 GByte), we therefore can only deal with the case d 8 for d-linear basis functions and d 11 for linear basis functions. A decomposition of the matrix entries over several computers in a parallel environment would permit more dimensions. 3. NUMERICAL RESULTS We now apply our approach to dierent test data sets. Here we use both synthetical data and real data from practical data mining applications. All the data sets are rescaled to [0; 1] d . To evaluate our method we give the correctness rates on testing data sets, if available, or the ten-fold cross-validation results otherwise. For further details and a criti- Figure 4: Spiral data set, sparse grid with level 5 (top left) to 8 (bottom right) cal discussion on the evaluation of the quality of classica- tion algorithms see [13, 37]. 3.1 Two-dimensional problems We rst consider synthetic two-dimensional problems with small sets of data which correspond to certain structures. 3.1.1 Spiral The rst example is the spiral data set, proposed by Alexis Wieland of MITRE Corp [48]. Here, 194 data points describe two intertwined spirals, see Figure 4. This is surely an articial problem which does not appear in practical ap- plications. However it serves as a hard test case for new data mining algorithms. It is known that neural networks can have severe problems with this data set and some neural networks can not separate the two spirals at all [40]. In Table 2 we give the correctness rates achieved with the leave-one-out cross-validation method, i.e. a 194-fold cross- validation. The best testing correctness was achieved on level 8 with 89.18% in comparison to 77.20% in [40]. In Figure 4 we show the corresponding results obtained with our sparse grid combination method for the levels 5 to 8. With level 7 the two spirals are clearly detected and resolved. Note that here 1281 grid points are contained in the sparse grid. For level 8 (2817 sparse grid points) the shape of the two reconstructed spirals gets smoother and Table 3: Results for the Ripley data set linear basis d-linear basis best possible % level ten-fold test % test data % test data % linear d-linear 9 87.7 0.0015 90.1 90.9 91.1 91.0 level training correctness testing correctness 9 0.0006 100.00 % 88.14 % Table 2: Leave-one-out cross-validation results for the spiral data set the reconstruction gets more precise. 3.1.2 Ripley This data set, taken from [36], consists of 250 training data and 1000 test points. The data set was generated synthetically and is known to exhibit 8 % error. Thus no better testing correctness than 92 % can be expected. Since we now have training and testing data, we proceed as follows: First we use the training set to determine the best regularization parameter per ten-fold cross-validation. The best test correctness rate and the corresponding are given for dierent levels n in the rst two columns of Table 3. With this we then compute the sparse grid classier from the 250 training data. The column three of Table 3 gives the result of this classier on the (previously unknown) test data set. We see that our method works well. Already level 4 is sucient to obtain results of 91.4 %. The reason is surely the relative simplicity of the data, see Figure 5. Just a few hyperplanes should be enough to separate the classes quite properly. We also see that there is not much need to use any higher levels, on the contrary there is even an overtting eect visible in Figure 5. In column 4 we show the results from [18], there we achieve almost the same results with d-linear functions. To see what kind of results could be possible with a more sophisticated strategy for determing we give in the last two columns of Table 3 the testing correctness which is achieved for the best possible . To this end we compute for all (discrete) values of the sparse grid classiers from the 250 data points and evaluate them on the test set. We then pick the best result. We clearly see that there is not much of a dierence. This indicates that the approach to determine the value of from the training set by cross-validation works well. Again we have almost the same results with linear and d-linear basis functions. Note that a testing correctness of Figure 5: Ripley data set, combination technique with linear basis functions. Left: level 4, Right: level 8, 90.6 % and 91.1 % was achieved in [36] and [35], respectively, for this data set. 3.2 6-dimensional problems 3.2.1 BUPA Liver The BUPA Liver Disorders data set from Irvine Machine Learning Database Repository [6] consists of 345 data points with 6 features and a selector eld used to split the data into 2 sets with 145 instances and 200 instances respectively. Here we have no test data and therefore can only report our ten-fold cross-validation results. We compare with our d-linear results from [18] and with the two best results from [33], the therein introduced smoothed support vector machine (SSVM) and the classical support vector machine (SVM jj:jj 2) [11, 46]. The results are given in Table 4. As expected, our sparse grid combination approach with linear basis functions performs slightly worse than the d- linear approach. The best test result was 69.60% on level 4. The new variant of the sparse grid combination technique performs only slightly worse than the SSVM, whereas the d-linear variant performs slighly better than the support vector machines. Note that the results for other SVM approaches like the support vector machine using the 1-norm approach (SVM jj:jj 1 were reported to be somewhat worse in [33]. Table 4: Results for the BUPA liver disorders data set linear d-linear For comparison with % % other methods level 1 10-fold train. correctness 0.012 76.00 0.020 76.00 SVM [33] 10-fold test. correctness 69.00 67.87 SSVM SVM jj:jj 2level 2 10-fold train. correctness 0.040 76.13 0.10 77.49 70.37 70.57 10-fold test. correctness 66.01 67.84 70.33 69.86 level 3 10-fold train. correctness 0.165 78.71 0.007 84.28 10-fold test. correctness 66.41 70.34 level 4 10-fold train. correctness 0.075 92.01 0.0004 90.27 10-fold test. correctness 69.60 70.92 3.2.2 Synthetic massive data set in 6D To measure the performance on a massive data set we produced with DatGen [34] a 6-dimensional test case with 5 million training points and 20 000 points for testing. We used the call datgen -r1 -X0/100,R,O:0/100,R,O:0/100,R,O: -O5020000 -p -e0.15. The results are given in Table 5. Note that already on level 1 a testing correctness of over 90 % was achieved with just 0:01. The main observation on this test case concerns the execution time, measured on a Pentium III 700 MHz machine. Besides the total run time, we also give the CPU time which is needed for the computation of the matrices l . We see that with linear basis functions really huge data sets of 5 million points can be processed in reasonable time. Note that more than 50 % of the computation time is spent for the data matrix assembly only and, more importantly, that the execution time scales linearly with the number of data points. The latter is also the case for the d-linear func- tions, but, as mentioned, this approach needs more operations per data point and results in a much longer execution time, compare also Table 5. Especially the assembly of the data matrix needs more than 96 % of the total run time for this variant. For our present example the linear basis approach is about 40 times faster than the d-linear approach on the same renement level, e.g. for level 2 we need 17 minutes in the linear case and 11 hours in the d-linear case. For higher dimensions the factor will be even larger. 3.3 10-dimensional problems 3.3.1 Forest cover type The forest cover type dataset comes from the UCI KDD Archive [4], it was also used in [30], where an approach similar to ours was followed. It consists of cartographic variables for meter cells and a forest cover type is to be pre- dicted. The 12 originally measured attributes resulted in 54 attributes in the data set, besides 10 quantitative variables there are 4 binary wilderness areas and 40 binary soil type variables. We only use the quantitative variables. The class label has 7 values, Spruce/Fir, Lodgepole Pine, Ponderosa Pine, Cottonwood/Willow, Aspen, Douglas-r and Krummholz. Like [30] we only report results for the classi- cation of Ponderosa Pine, which has 35754 instances out of the total 581012. Since far less than 10 % of the instances belong to Ponderosa Pine we weigh this class with a factor of 5, i.e. Ponderosa Pine has a class value of 5, all others of -1 and the treshold value for separating the classes is 0. The data set was randomly separated into a training set, a test set, and a evaluation set, all similar in size. In [30] only results up to 6 dimensions could be reported. In Table 6 we present our results for the 6 dimensions chosen there, i.e. the dimensions 1,4,5,6,7, and 10, and for all 10 dimensions as well. To give an overview of the behavior over several 's we present for each level n the overall correctness results, the correctness results for Ponderosa Pine and the correctness result for the other class for three values of . We then give results on the evaluation set for a chosen . We see in Table 6 that already with level 1 we have a testing correctness of 93.95 % for the Ponderosa Pine in the 6 dimensional version. Higher renement levels do not give better results. The result of 93.52% on the evaluation set is almost the same as the corresponding testing correctness. Note that in [30] a correctness rate of 86.97 % was achieved on the evaluation set. The usage of all 10 dimensions improves the results slightly, we get 93.81 % as our evaluation result on level 1. As before higher renement levels do not improve the results for this data set. Note that the forest cover example is sound enough as an example of classication, but it might strike forest scientists as being amusingly supercial. It has been known for years that the dynamics of forest growth can have a dominant eect on which species is present at a given location [7], yet there are no dynamic variables in the classier. This one can see as a warning that it should never be assumed that the available data contains all the relevant information. 3.3.2 Synthetic massive data set in 10D To measure the performance on a still higher dimensional massive data set we produced with DatGen [34] a 10-dimen- sional test case with 5 million training points and 50 000 points for testing. We used the call datgen -r1 -X0/200,R,O: Like in the synthetical 6-dimensional example the main observations concern the run time, measured on a Pentium III 700 MHz machine. Besides the total run time, we also give the CPU time which is needed for the computation of the matrices G l . Note that the highest amount of memory needed (for level 2 in the case of 5 million data points) was 500 MBytes, about 250 MBytes for the matrix and about 250 MBytes for keeping the data points in memory More than 50 % of the run time is spent for the assembly Table 5: Results for a 6D synthetic massive data set, training testing total data matrix # of # of points correctness correctness time (sec) time (sec) iterations linear basis functions level 1 500 000 90.5 90.5 25 8 25 5 million 90.5 90.6 242 77 28 level 2 500 000 91.2 91.1 110 55 204 5 million 91.1 91.2 1086 546 223 50 000 92.2 91.4 48 23 869 level 3 500 000 91.7 91.7 417 226 966 5 million 91.6 91.7 4087 2239 1057 d-linear basis functions level 1 500 000 90.7 90.8 597 572 91 5 million 90.7 90.7 5897 5658 102 level 2 500 000 91.5 91.6 4285 4168 656 5 million 91.4 91.5 42690 41596 742 of the data matrix and the time needed for the data matrix scales linearly with the number of data points, see Table 7. The total run time seems to scale even better than linear. 4. CONCLUSIONS We presented the sparse grid combination technique with linear basis functions based on simplices for the classication of data in moderate-dimensional spaces. Our new method gave good results for a wide range of problems. It is capable to handle huge data sets with 5 million points and more. The run time scales only linearly with the number of data. This is an important property for many practical applications where often the dimension of the problem can substantially be reduced by certain preprocessing steps but the number of data can be extremely huge. We believe that our sparse grid combination method possesses a great potential in such practical application problems. We demonstrated for the Ripley data set how the best value of the regularization parameter can be determined. This is also of practical relevance. A parallel version of the sparse grid combination technique reduces the run time signicantly, see [17]. Note that our method is easily parallelizable already on a coarse grain level. A second level of parallelization is possible on each grid of the combination technique with the standard techniques known from the numerical treatment of partial differential equations. Since not necessarily all dimensions need the maximum renement level, a modication of the combination technique with regard to dierent renement levels in each dimension along the lines of [19] seems to be promising. Note furthermore that our approach delivers a continuous classier function which approximates the data. It therefore can be used without modication for regression problems as well. This is in contrast to many other methods like e.g. decision trees. Also more than two classes can be handled by using isolines with just dierent values. Finally, for reasons of simplicity, we used the operator r. But other dierential (e.g. operators can be employed here with their associated regular nite element ansatz functions. 5. ACKNOWLEDGEMENTS Part of the work was supported by the German Bundesministerium fur Bildung und Forschung (BMB+F) within the project 03GRM6BN. This work was carried out in cooperation with Prudential Systems Software GmbH, Chemnitz. The authors thank one of the referees for his remarks on the forest cover data set. 6. --R Adaptive Verfahren f The UCI KDD archive. UCI repository of machine learning databases Some ecological consequences of a computer model of forest growth. Tensor product approximation spaces for the e Learning from Data - Concepts Data Mining Methods for Knowledge Discovery. Approximate statistical tests for comparing supervised classi Information Complexity of Multivariate Fredholm Integral Equations in Sobolev Classes. Simplizialzerlegungen von beschr On the computation of the eigenproblems of hydrogen and helium in strong magnetic and electric On the parallelization of the sparse grid approach for data mining. Data mining with sparse grids. Numerical Integration using Sparse Grids. An equivalence between sparse approximation and support vector machines. Regularization theory and neural networks architectures. Generalized cross validation as a method for choosing a good ridge parameter. The combination technique for the sparse grid solution of PDEs on multiprocessor machines. Adaptive sparse grid multilevel methods for elliptic PDEs based on Optimized tensor-product approximation spaces Sparse grids for boundary integral equations. A combination technique for the solution of sparse grid problems. High dimensional smoothing based on multilevel analysis. Grundlagen der goemetrischen Datenverarbeitung Some combinatorial lemmas in topology. SSVM: A smooth support vector machine for classi A program that creates structured data. Bayesian neural networks for classi Neural networks and related methods for classi On comparing classi Die Methode der Finiten Di Interpolation on sparse grids and Nikol'skij-Besov spaces of dominating mixed smoothness 2d spiral pattern recognition with possibilistic measures. Quadrature and interpolation formulas for tensor products of certain classes of functions. Approximation of functions with bounded mixed derivative. Solutios of ill-posed problems Estimation of dependences based on empirical data. The Nature of Statistical Learning Theory. Spline models for observational data Spiral data set. Sparse grids. --TR Regularization theory and neural networks architectures Approximation of scattered data using smooth grid functions The nature of statistical learning theory Information complexity of multivariate Fredholm integral equations in Sobolev classes 2D spiral pattern recognition with possibilistic measures An equivalence between sparse approximation and support vector machines Data mining methods for knowledge discovery Adaptive sparse grid multilevel methods for elliptic PDEs based on finite differences Approximate statistical tests for comparing supervised classification learning algorithms Bayesian neural networks for classification On the computation of the eigenproblems of hydrogen helium in strong magnetic and electric fields with the sparse grid combination technique Learning from Data On Comparing Classifiers On the Parallel Solution of 3D PDEs on a Network of Workstations and on Vector Computers --CTR Jochen Garcke, Regression with the optimised combination technique, Proceedings of the 23rd international conference on Machine learning, p.321-328, June 25-29, 2006, Pittsburgh, Pennsylvania Deepak K. Agarwal, Shrinkage estimator generalizations of Proximal Support Vector Machines, Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining, July 23-26, 2002, Edmonton, Alberta, Canada J. Garcke , M. Griebel , M. Thess, Data mining with sparse grids, Computing, v.67 n.3, p.225-253, November 2001

data mining;classification;approximation;simplicial discretization;sparse grids;combination technique

502529

Mining time-changing data streams.

Most statistical and machine-learning algorithms assume that the data is a random sample drawn from a stationary distribution. Unfortunately, most of the large databases available for mining today violate this assumption. They were gathered over months or years, and the underlying processes generating them changed during this time, sometimes radically. Although a number of algorithms have been proposed for learning time-changing concepts, they generally do not scale well to very large databases. In this paper we propose an efficient algorithm for mining decision trees from continuously-changing data streams, based on the ultra-fast VFDT decision tree learner. This algorithm, called CVFDT, stays current while making the most of old data by growing an alternative subtree whenever an old one becomes questionable, and replacing the old with the new when the new becomes more accurate. CVFDT learns a model which is similar in accuracy to the one that would be learned by reapplying VFDT to a moving window of examples every time a new example arrives, but with O(1) complexity per example, as opposed to O(w), where w is the size of the window. Experiments on a set of large time-changing data streams demonstrate the utility of this approach.

INTRODUCTION Modern organizations produce data at unprecedented rates; among large retailers, e-commerce sites, telecommunications providers, and scientic projects, rates of gigabytes per day are common. While this data can contain valuable knowledge, its volume increasingly outpaces practitioners' ability to mine it. As a result, it is now common practice either to mine a subsample of the available data or to mine for models drastically simpler than the data could support. In some cases, the volume and time span of accumulated data is such that just storing it consistently and reliably for future use is a challenge. Further, even when storage is not problematic, it is often di-cult to gather the data in one place, at one time, in a format appropriate for mining. For all these reasons, in many areas the notion of mining a xed-sized database is giving way to the notion of mining an open-ended data stream as it arrives. The goal of our re-search is to help make this possible with a minimum of eort for the data mining practitioner. In a previous paper [9] we presented VFDT, a decision tree induction system capable of learning from high-speed data streams in an incremental, anytime fashion, while producing models that are asymptotically arbitrarily close to those that would be learned by traditional decision tree induction systems. Most statistical and machine-learning algorithms, including VFDT, make the assumption that training data is a random sample drawn from a stationary distribution. Un- fortunately, most of the large databases and data streams available for mining today violate this assumption. They exist over months or years, and the underlying processes generating them changes during this time, sometimes radically. For example, a new product or promotion, a hacker's attack, a holiday, changing weather conditions, changing economic conditions, or a poorly calibrated sensor could all lead to violations of this assumption. For classication systems, which attempt to learn a discrete function given examples of its inputs and outputs, this problem takes the form of changes in the target function over time, and is known as concept drift. Traditional systems assume that all data was generated by a single concept. In many cases, however, it is more accurate to assume that data was generated by a series of concepts, or by a concept function with time-varying parameters. Traditional systems learn incorrect models when they erroneously assume that the underlying concept is stationary if in fact it is drifting. One common approach to learning from time-changing data is to repeatedly apply a traditional learner to a sliding window of w examples; as new examples arrive they are inserted into the beginning of the window, a corresponding number of examples is removed from the end of the win- dow, and the learner is reapplied [27]. As long as w is small relative to the rate of concept drift, this procedure assures availability of a model re ecting the current concept generating the data. If the window is too small, however, this may result in insu-cient examples to satisfactorily learn the concept. Further, the computational cost of reapplying a learner may be prohibitively high, especially if examples arrive at a rapid rate and the concept changes quickly. To meet these challenges we propose the CVFDT system, which is capable of learning decision trees from high-speed, time changing data streams. CVFDT works by e-ciently keeping a decision tree up-to-date with a window of exam- ples. In particular, it is able to keep its model consistent with a window using only a constant amount of time for each new example (more precisely, time proportional to the number of attributes in the data and the depth of the induced tree). CVFDT grows an alternate subtree whenever an old one seems to be out-of-date, and replaces the old one when the new one becomes more accurate. This allows it to make smooth, ne-grained adjustments when concept drift occurs. In eect, CVFDT is able to learn a nearly equivalent model to the one VFDT would learn if repeatedly reapplied to a window of examples, but in O(1) time instead of O(w) time per new example. In the next section we discuss the basics of the VFDT sys- tem, and in the following section we introduce the CVFDT system. We then present a series of experiments on synthetic data which demonstrate how CVFDT can outperform traditional systems on high-speed, time-changing data streams. Next, we apply CVFDT to mining the stream of web page requests for the entire University of Washington campus. We conclude with a discussion of related and future work. 2. THE VFDT SYSTEM The classication problem is generally dened as follows. A set of N training examples of the form (x; y) is given, where y is a discrete class label and x is a vector of d at- tributes, each of which may be symbolic or numeric. The goal is to produce from these examples a model which will predict the classes y of future examples x with high accuracy. For example, x could be a description of a client's recent purchases, and y the decision to send that customer a catalog or not; or x could be a record of a cellular- telephone call, and y the decision whether it is fraudulent or not. One of the most eective and widely-used classi- cation methods is decision tree learning [4, 20]. Learners of this type induce models in the form of decision trees, where each node contains a test on an attribute, each branch from a node corresponds to a possible outcome of the test, and each leaf contains a class prediction. The label for an example x is obtained by passing the example down from the root to a leaf, testing the appropriate attribute at each node and following the branch corresponding to the attribute's value in the example. A decision tree is learned by recursively replacing leaves by test nodes, starting at the root. The attribute to test at a node is chosen by comparing all the available attributes and choosing the best one according to some heuristic measure. Classic decision tree learners like C4.5 [20], CART, SLIQ [17], and SPRINT [24] Table 1: The VFDT Algorithm. Inputs: S is a stream of examples, X is a set of symbolic attributes, G(:) is a split evaluation function, - is one minus the desired probability of choosing the correct attribute at any given node, is a user-supplied tie threshold, nmin is the # examples between checks for growth. Output: HT is a decision tree. Procedure VFDT (S; X;G; -; ) Let HT be a tree with a single leaf l 1 (the root). g. be the G obtained by predicting the most frequent class in S. For each class yk For each value x ij of each attribute X For each example (x; y) in S Sort (x; y) into a leaf l using HT . For each x ij in x such that X Increment n ijy (l). Label l with the majority class among the examples seen so far at l. Let n l be the number of examples seen at l. If the examples seen so far at l are not all of the same class and n l mod nmin is 0, then Compute using the counts n ijk (l). Let Xa be the attribute with highest G l . Let X b be the attribute with second-highest G l . Compute using Equation 1. If ((G l > ) or (G l <= < Replace l by an internal node that splits on Xa . For each branch of the split Add a new leaf l m , and let be the G obtained by predicting the most frequent class at l m . For each class yk and each value x ij of each attribute Return HT . use every available training example to select the best attribute for each split. This policy is necessary when data is scarce, but it has two problems when training examples are abundant: it requires all examples be available for consideration throughout their entire runs, which is problematic when data does not t in RAM or on disk, and it assumes that the process generating examples remains the same during the entire period over which the examples are collected and mined. In previous work [9] we presented the VFDT (Very Fast Decision Tree learner) system, which is able to learn from abundant data within practical time and memory constrai- nts. It accomplishes this by noting, with Catlett [5] and others [12, 19], that it may be su-cient to use a small sample of the available examples when choosing the split attribute at any given node. Thus, only the rst examples to arrive on the data stream need to be used to choose the split attribute at the root; subsequent ones are passed through the induced portion of the tree until they reach a leaf, are used to choose a split attribute there, and so on recursively. To determine the number of examples needed for each decision, VFDT uses a statistical result known as Hoeding bounds or additive Cherno bounds [13]. After n independent observations of a real-valued random variable r with range R, the Hoeding bound ensures that, with condence 1 -, the true mean of r is at least r , where r is the observed mean of the samples and r 2n (1) This is true irrespective of the probability distribution that generated the observations. Let G(X i ) be the heuristic measure used to choose test attributes (we use information gain). After seeing n samples at a leaf, let Xa be the attribute with the best heuristic measure and X b be the attribute with the second best. Let be a new random variable, the dierence between the observed heuristic val- ues. Applying the Hoeding bound to G, we see that if G > (as calculated by Equation 1 with a user-supplied -), we can condently say that the dierence between G(Xa) and G(X b ) is larger than zero, and select Xa as the split at- tribute. 1;2 Table 1 contains pseudo-code for VFDT's core al- gorithm. The counts n ijk are the su-cient statistics needed to compute most heuristic measures; if other quantities are required, they can be similarly maintained. When the sucient statistics ll the available memory, VFDT reduces its memory requirements by temporarily deactivating learning in the least promising nodes; these nodes can be reactivated later if they begin to look more promising than currently active nodes. VFDT employs a tie mechanism which precludes it from spending inordinate time deciding between 1 This is valid as long as G (and therefore G) can be viewed as an average over all examples seen at the leaf, which is the case for most commonly-used heuristics. For example, if information gain is used, the quantity being averaged is the reduction in the uncertainty regarding the class membership of the example. 2 In this paper we assume that the third-best and lower attributes have su-ciently smaller gains that their probability of being the true best choice is negligible. We plan to lift this assumption in future work. If the attributes at a given node are (pessimistically) assumed independent, it simply involves a Bonferroni correction to - [18]. attributes whose practical dierence is negligible. That is, VFDT declares a tie and selects Xa as the split attribute any time G < < (where is a user-supplied tie thresh- old). Pre-pruning is carried out by considering at each node a \null" attribute X ; that consists of not splitting the node. Thus a split will only be made if, with condence 1 -, the best split found is better according to G than not splitting. Notice that the tests for splits and ties are only executed once for every nmin (a user supplied value) examples that arrive at a leaf. This is justied by the observation that VFDT is unlikely to make a decision after any given exam- ple, so it is wasteful to carry out these calculations for each one of them. The pseudo-code shown is only for symbolic attributes; we are currently developing its extension to numeric ones. The sequence of examples S may be innite, in which case the procedure never terminates, and at any point in time a parallel procedure can use the current tree HT to make class predictions. Using o-the-shelf hardware, VFDT is able to learn as fast as data can be read from disk. The time to incorporate an example is O(ldvc) where l is the maximum depth of HT , d is the number of attributes, v is the maximum number of values per attribute, and c is the number of classes. This time is independent of the total number of examples already seen (assuming the size of the tree depends only on the \true" concept, and not on the dataset). Because of the use of Hoeding bounds, these speed gains do not necessarily lead to a loss of accuracy. It can be shown that, with high condence, the core VFDT system (without ties or de- activations due to memory constraints) will asymptotically induce a tree arbitrarily close to the tree induced by a traditional batch learner. Let DT1 be the tree induced by a version of VFDT using innite data to choose each node's split attribute, HT - be the tree learned by the core VFDT system given an innite data stream, and p be the probability that an example passed through DT1 to level i will fall into a leaf at that point. Then the probability that an arbitrary example will take a dierent path through DT1 and HT - is bounded by -=p [9]. A corollary of this result states that the tree learned by the core VFDT system on a nite sequence of examples will correspond to a subtree of DT1 with the same bound of -=p. See Domingos and Hulten [9] for more details on VFDT and this -=p bound. 3. THE CVFDT SYSTEM CVFDT (Concept-adapting Very Fast Decision Tree learner) is an extension to VFDT which maintains VFDT's speed and accuracy advantages but adds the ability to detect and respond to changes in the example-generating process. Like other systems with this capability, CVFDT works by keeping its model consistent with a sliding window of ex- amples. However, it does not need to learn a new model from scratch every time a new example arrives; instead, it updates the su-cient statistics at its nodes by incrementing the counts corresponding to the new example, and decrementing the counts corresponding to the oldest example in the window (which now needs to be forgotten). This will statistically have no eect if the underlying concept is sta- tionary. If the concept is changing, however, some splits that previously passed the Hoeding test will no longer do so, because an alternative attribute now has higher gain (or the two are too close to tell). In this case CVFDT begins to grow an alternative subtree with the new best attribute at Table 2: The CVFDT algorithm. Inputs: S is a sequence of examples, X is a set of symbolic attributes, G(:) is a split evaluation function, - is one minus the desired probability of choosing the correct attribute at any given node, is a user-supplied tie threshold, w is the size of the window, nmin is the # examples between checks for growth, f is the # examples between checks for drift. Output: HT is a decision tree. Procedure CVFDT(S;X; G; -; ; w; nmin) /* Initialize */ Let HT be a tree with a single leaf l 1 (the root). Let ALT (l 1) be an initially empty set of alternate trees for l 1 . be the G obtained by predicting the most frequent class in S. g. Let W be the window of examples, initially empty. For each class yk For each value x ij of each attribute X /* Process the examples */ For each example (x; y) in S Sort (x; y) into a set of leaves L using HT and all trees in ALT of any node (x; y) passes through. Let ID be the maximum id of the leaves in L. Add ((x; y); ID) to the beginning of W . If be the last element of W removed G; (x; y); -; nmin ; ) If there have been f examples since the last checking of alternate trees CheckSplitValidity(HT; n; -) Return HT . Table 3: The CVFDTGrow procedure. Procedure CVFDTGrow(HT; n; G; (x; y); -; nmin ; ) Sort (x; y) into a leaf l using HT . Let P be the set of nodes traversed in the sort. For each node l pi in P For each x ij in x such that X Increment n ijy (l p ). For each tree Ta in ALT (l p) G; (x; y); -; nmin ; ) Label l with the majority class among the examples seen so far at l. Let n l be the number of examples seen at l. If the examples seen so far at l are not all of the same class and n l mod nmin is 0, then Compute using the counts n ijk (l). Let Xa be the attribute with highest G l . Let X b be the attribute with second-highest G l . Compute using Equation 1 and -. If ((G l > ) or (G l <= < Replace l by an internal node that splits on Xa . For each branch of the split Add a new leaf l m , and let Let ALT (l m) = fg. be the G obtained by predicting the most frequent class at l m . For each class yk and each value x ij of each attribute its root. When this alternate subtree becomes more accurate on new data than the old one, the old subtree is replaced by the new one. Table contains a pseudo-code outline of the CVFDT algorithm. CVFDT does some initializations, and then processes examples from the stream S indenitely. As each example arrives, it is added to the window 3 , an old example is forgotten if needed, and (x; y) is incorporated into the current model. CVFDT periodically scans HT and all alternate trees looking for internal nodes whose su-cient statistics indicate that some new attribute would make a better test than the chosen split attribute. An alternate subtree is started at each such node. Table 3 contains pseudo-code for the tree-growing portion of the CVFDT system. It is similar to the Hoeding Tree algorithm, but CVFDT monitors the validity of its old decisions by maintaining su-cient statistics at every node in HT (instead of only at the leaves like VFDT). Forgetting an old example is slightly complicated by the fact that HT may have grown or changed since the example was initially incorporated. Therefore, nodes are assigned a unique, monotonically increasing ID as they are created. When an example is added to W , the maximum ID of the leaves it reaches in HT and all alternate trees is recorded with it. An example's eects are forgotten by decrementing the counts in the su-cient statistics of every node the example reaches 3 The window is stored in RAM if resources are available, otherwise it will be kept on disk. Table 4: The ForgetExample procedure. Procedure while it traverses leaves with id IDw , Let P be the set of nodes traversed in the sort. For each node l in P For each x ij in x such that X Decrement n ijk (l). For each tree T alt in ALT (l) in HT whose ID is the stored ID. See the pseudo-code in Table 4 for more detail about how CVFDT forgets examples. CVFDT periodically scans the internal nodes of HT looking for ones where the chosen split attribute would no longer be selected; that is, where G(Xa) G(X b ) and > . When it nds such a node, CVFDT knows that it either initially made a mistake splitting on Xa (which should happen less than -% of the time), or that something about the process generating examples has changed. In either case, CVFDT will need to take action to correct HT . CVFDT grows alternate subtrees to changed subtrees of HT , and only modies HT when the alternate is more accurate than the original. To see why this is needed, let l be a node where change was detected. A simple solution is to replace l with a leaf predicting the most common class in l 's sufcient statistics. This policy assures that HT is always as current as possible with respect to the process generating examples. However, it may be too drastic, because it initially forces a single leaf to do the job previously done by a whole subtree. Even if the subtree is outdated, it may still be better than the best single leaf. This is particularly true when l is at or near the root of HT , as it will result in drastic short-term reductions in HT 's predictive accuracy { clearly not acceptable when a parallel process is using HT to make critical decisions. Each internal node in HT has a list of alternate subtrees being considered as replacements for the subtree rooted at the node. Table 5 contains pseudo-code for the CheckSplit- procedure. CheckSplitValidity starts an alternate subtree whenever it nds a new winning attribute at a node; that is, when there is a new best attribute and G > or if < and G =2. This is very similar to the procedure used to choose initial splits, except the tie criteria is tighter to avoid excessive alternate tree creation. CVFDT supports a parameter which limits the total number of alternate trees being grown at any one time. Alternate trees are grown the same way HT is, via recursive calls to the CVFDT pro- cedures. Periodically, each node with a non-empty set of alternate subtrees, l test , enters a testing mode to determine if it should be replaced by one of its alternate subtrees. Once in this mode, l test collects the next m training examples that arrive at it and, instead of using them to grow its children or alternate trees, uses them to compare the accuracy of the subtree it roots with the accuracies of all of its alternate subtrees. If the most accurate alternate subtree is more accurate than the l test , l test is replaced by the alternate. During the test phase, CVFDT also prunes alternate subtrees that are not making progress (i.e., whose accuracy is not in- Table 5: The CheckSplitValidity procedure. Procedure CheckSplitValidity(HT; n; -) For each node l in HT that is not a leaf For each tree T alt in ALT (l) Let Xa be the split attribute at l. Let Xn be the attribute with the highest G l other than Xa . Let X b be the attribute with the highest G l other than Xn . If G l 0 and no tree in ALT (l) already splits on at its root Compute using Equation 1 and -. If (G l > ) or ( < and G l =2), then Let l new be an internal node that splits on Xn . Let ALT For each branch of the split Add a new leaf l m to l new Let ALT (l m) = fg. be the G obtained by predicting the most frequent class at l m . For each class yk and each value x ij of each attribute creasing over time). For each alternate subtree of l test , l i alt , CVFDT remembers the smallest accuracy dierence ever achieved between the two, min (l test ; l i alt ). CVFDT prunes any alternate whose current test phase accuracy dierence is at least min (l test ; l i One window size w will not be appropriate for every concept and every type of drift; it may be benecial to dynamically change w during a run. For example, it may make sense to shrink w when many of the nodes in HT become questionable at once, or in response to a rapid change in data rate, as these events could indicate a sudden concept change. Similarly, some applications may benet from an increase in w when there are few questionable nodes because this may indicate that the concept is stable { a good time to learn a more detailed model. CVFDT is able to dynamically adjust the size of its window in response to user-supplied events. Events are specied in the form of hook functions which monitor S and HT and can call the SetWindowSize function when appropriate. CVFDT changes the window size by updating w and immediately forgetting any examples that no longer t in W . We now discuss a few of the properties of the CVFDT system and brie y compare it with VFDT-Window, a learner that reapplies VFDT to W for every new example. CVFDT requires memory proportional to O(ndvc) where n is the number of nodes in CVFDT's main tree and all alternate trees, d is the number of attributes, v is the maximum number of values per attribute, and c is the number of classes. The window of examples can be in RAM or can be stored on 4 When RAM is short, CVFDT is more aggressive about pruning unpromising alternate subtrees. disk at the cost of a few disk accesses per example. There- fore, CVFDT's memory requirements are dominated by the su-cient statistics and are independent of the total number of examples seen. At any point during a run, CVFDT will have available a model which re ects the current concept generating W . It is able to keep this model up-to-date in time proportional to O(lcdvc) per example, where l c is the length of the longest path an example will have to take through HT times the number of alternate trees. VFDT- Window requires O(lvdvcw) time to keep its model up-to- date for every new example, where l v is the maximum depth of HT . VFDT is a factor of wlv =lc worse than CVFDT; em- pirically, we observed l c to be smaller than l v in all of our experiments. Despite this large time dierence, CVFDT's drift mechanisms allow it to produce a model of similar ac- curacy. The structure of the models induced by the two may, however, be signicantly dierent, for the following reason. VFDT-Window uses the information from each training example at one place in the tree it induces: the leaf where the example falls when it arrives. This means that VFDT- Window uses the rst examples from W to make a decision at its root, the next to make a decision at the rst level of the tree, and so on. After an initial building phase, CVFDT will have a fully induced tree available. Every new example is passed through this induced tree, and the information it contains is used to update statistics at every node it passes through. This dierence can be an advantage for CVFDT, as it allows the induction of larger trees with better probability estimates at the leaves. It can also be a disadvantage and VFDT-Window may be more accurate when there is a large concept shift part-way through W . This is because VFDT-Window's leaf probabilities will be set by examples near the end of W while CVFDT's will re ect all of W . Also notice that, even when the structure of the induced tree does not change, CVFDT and VFDT-Window can out-perform VFDT simply because their leaf probabilities (and therefore class predictions) are updated faster, without the \dead weight" of all the examples that fell into leaves before the current window. 4. EMPIRICAL STUDY We conducted a series of experiments comparing CVFDT to VFDT and VFDT-Window. Our goals were to evaluate CVFDT's ability to scale up, to evaluate CVFDT's ability to deal with varying levels of drift, and to identify and characterize the situations where CVFDT outperforms the other systems. 4.1 Synthetic Data The experiments with synthetic data used a changing concept based on a rotating hyperplane. A hyperplane in d-dimensional space is the set of points x that satisfy d where x i is the ith coordinate of x. Examples for which are labeled positive, and examples for which are labeled negative. Hyperplanes are useful for simulating time-changing concepts because we can change the orientation and position of the hyperplane in a smooth manner by changing the relative size of the weights. In particular, sorting the weights by their magnitudes provides a good indication of which dimensions contain the most information; in the limit, when all but one of the weights are zero, the dimension associated with the non-zero weight is the only one that contains any information about the concept. This allows us to control the relative information content of the attributes, and thus change the optimal order of tests in a decision tree representing the hyperplane, by simply changing the relative sizes of the weights. We sought a concept that maintained the advantages of a hy- perplane, but where the weights could be randomly modied without potentially causing the decision frontier to move outside the range of the data. To meet these goals we used a series of alternating class bands separated by parallel hyper- planes. We start with a reference hyperplane whose weights are initialized to :2 except for w0 which is :25d. To label an example, we substitute its coordinates into the left hand side of Equation 2 to obtain a sum s. If jsj :1 w0 the example is labeled positive, otherwise if jsj :2 w0 the example is labeled negative, and so on. Examples were generated uniformly in a d-dimensional unit hypercube (with the value of each x i ranging from [0, 1]). They were then labeled using the concept, and their continuous attributes were uniformly discretized into ve bins. Noise was added by randomly switching the class labels of p% of the exam- ples. Unless otherwise stated, each experiment used the following settings: ve million training examples; window on disk; no memory limits; no pre-pruning; a test set of 50,000 examples; and alternate tree test mode after 9,000 examples and used test samples of 1,000 examples. All runs were done on a 1GHz Pentium III machine with 512 MB of RAM, running Linux. The rst series of experiments compares the ability of CVFDT and VFDT to deal with large concept-drifting data- sets. Concept drift was added to the datasets in the following manner. Every 50,000 examples w1 was modied by adding 0:01d to it, and the test set was relabeled with the updated concept (with p% noise as before). was initially 1 and was multiplied by 1 at 5% of the drift points and also just before w1 fell below 0 or rose above :25d. Figure 1 compares the accuracy of the algorithms as a function of d, the dimensionality of the space. The reported values are obtained by testing the accuracy of the learned models every 10,000 examples throughout the run and averaging these results. Drift level, reported on the minor axis, is the average percentage of the test set that changes label at each point the concept changes. CVFDT is substantially more accurate than VFDT, by approximately 10% on average, and CVFDT's performance improves slightly with increasing d. Figure 2 compares the average size of the models induced during the run shown in Figure 1 (the reported values are generated by averaging after every 10,000 examples, as before). CVFDT's trees are substantially smaller than VFDT's, and the advantage is consistent across all the values of d we tried. This simultaneous accuracy and size advantage derives from the fact that CVFDT's tree is built on the 100,000 most relevant examples, while VFDT's is built on millions of outdated examples. We next carried out a more detailed evaluation of CVFDT's concept drift mechanism. Figure 3 shows a detailed view of one of the runs from Figures 1 and 2, the one for 50. The minor axis shows the portion of the test Figure 1: Error rates as a function of the number of attributes. Figure 2: Tree sizes as a function of the number of attributes. Figure 3: Error rates of learners as a function of the number of examples seen. set that is labeled negative at each test point (computed before noise is added to the test set) and is included to illustrate the concept drift present in the dataset. CVFDT is able to quickly respond to drift, while VFDT's error rate often rises drastically before reacting to the change. Further, VFDT's error rate seems to peak at worse values as the run goes on, while CVFDT's error peaks seem to have constant height. We believe this happens because VFDT has more trouble responding to drift when it has induced a larger tree and must replicate corrections across more outdated struc- ture. CVFDT does not face this problem because it replaces subtrees when they become outdated. We gathered some detailed statistics about this run. CVFDT took 4.3 times longer than VFDT (5.7 times longer if including time to do the disk I/O needed to keep the window on disk). VFDT's average memory allocation over the course of the run was 23 MB while CVFDT's was 16.5 MB. The average number of nodes in VFDT's tree was 2696 and the average number in CVFDT's tree was 677, of which 132 were in alternate trees and the remainder were in the main tree. Next we examined how CVFDT responds to changing levels of concept drift on ve datasets with added using a parameter D. Every 75,000 examples, D of the concept hyperplane's weights were selected at random and updated as before, w has a 25% chance of ipping signs, chosen to prevent too many weights from drifting in the same pattern). Figure 4 shows the comparison on these datasets. CVFDT substantially outperformed VFDT at every level of drift. Notice that VFDT's error rate approaches 50% for D > 2, and that the variance in VFDT's data points is large. CVFDT's error rate seems to grow smoothly with increasing levels of concept change, suggesting that its drift adaptations are robust and eective. We wanted to gain some insight into the way CVFDT starts new alternate subtrees, prunes existing ones, and replaces portions of HT with alternates. For this purpose, we instrumented a run of CVFDT on the from Figure 4 to output a token in response to each of these events. We aggregated the events in chunks of 100,000 training examples, and generated data points for all non-zero values Figure 5 shows the results of this experiment. There are a large number of events during the run. For example, alternate subtrees were swapped into HT . Most of the swaps seem to occur when the examples in the test set are changing labels quickly. We also wanted to see how well CVFDT would compare to a system using traditional drift-tracking methods. We thus compared CVFDT, VFDT, and VFDT-Window. We simulated VFDT-Window by running VFDT on W for every 100,000 examples instead of for every example. The dataset for the experiment had used the same drift settings used to generate Figure 4 with 6 shows the results. CVFDT's error rate was the same as VFDT-Window's, except for a brief period during the middle of the run when class labels were changing most rapidly. CVFDT's average error rate for the run was 16.3%, VFDT's was 19.4%, and VFDT-Window's was 15.3%. The dierence in runtimes was very large. VFDT took about 10 minutes, CVFDT took about 46 minutes, and we estimate that VFDT-Window would have taken 548 days to do its complete run if applied to every new example. Put another way, VFDT-Window provides a 4% accuracy gain compared Figure 4: Error rates as a function of the amount of concept drift. Figure 5: CVFDT's drift characteristics. Figure rates over time of CVFDT, VFDT, and VFDT-Window. to VFDT, at a cost of increasing the running time by a factor of 17,000. CVFDT provides 75% of VFDT-Window's accuracy gain, and introduces a time penalty of less than 0.1% of VFDT-Window's. CVFDT's alternate trees and additional su-cient statistics do not use too much RAM. For example, none of runs ever grew to more than 70MB. We never observed CVFDT to use more RAM than VFDT; in fact it often used as little as half the RAM of VFDT. The systems' RAM requirements are dominated by the su-cient statistics which are kept at the leaves in VFDT, and at every node in CVFDT. We observed that VFDT often had twice as many leaves as there were nodes in CVFDT's tree and all alternate trees combined. This is what we expected: VFDT considers many more examples and is forced to grow larger trees to make up for the fact that its early decisions become incorrect due to concept drift. CVFDT's alternate tree pruning mechanism seems to be eective at trading memory for smooth transitions between concepts. Further, there is room for more aggressive pruning if CVFDT exhausts available RAM. Exploring this tradeo is an area for future work. 4.2 Web Data We are currently applying CVFDT to mining the stream of Web page requests emanating from the whole University of Washington main campus. The nature of the data is described in detail in Wolman et al. [29]. In our experiments so far we have used a one-week anonymized trace of all the external web accesses made from the university campus. There were 23,000 active clients during this one-week trace period, and the entire university population is estimated at 50,000 people (students, faculty and sta). The trace contains million requests, which arrive at a peak rate of 17,400 per minute. The size of the compressed trace le is about 20 GB. 5 Each request is tagged with an anonymized organization ID that associates the request with one of the organizations (colleges, departments, etc.) within the university. One purpose this data can be used for is to improve Web caching. The key to this is predicting as accurately as possible which hosts and pages will be requested in the near future, given recent requests. We applied decision-tree learning to this problem in the following manner. We split the campus-wide request log into a series of equal time slices in the experiments we report, each time slice is an hour. For each organization O1 ; and each of the 244k hosts appearing in the logs maintained a count of how many times the organization accessed the host in the time slice, C ijt . We discretized these counts into four buckets, representing \no requests," \1 { 12 requests," \13 { 25 requests" and \26 or more requests." Then for each time slice and host accessed in that time slice (T t generated an example with attributes if H j is requested in time slice T t+1 and 0 if it is not. This can be carried out in real time using modest resources by keeping statistics on the last and current time slices C t 1 and C t in memory, only keeping counts for hosts that actually appear in a time slice (we never needed more than 30k counts), and outputting the examples for C t 1 as soon as C t is complete. Using this procedure we obtained a dataset containing 1.89 million examples, 60.9% of which were la- 5 This log is from May 1999. Tra-c in May 2000 was more than double this size. beled with the most common class (that the host did not appear again in the next time slice). Our exploration was designed to determine if CVFDT's concept drift features would provide any benet to this ap- plication. As each example arrived, we tested the accuracy of the learners' models on it, and then allowed the learners to update their models with the example. We kept statistics about how the aggregated accuracies changed over time. VFDT and CVFDT were both run with and additional parameters were achieved 72.7% accuracy over the whole dataset and CVFDT achieved 72.3%. However, CVFDT's aggregated accuracy was higher for the rst 70% of the run, at times by as much as 1.0%. CVFDT's accuracy fell behind only near the end of the run, for (we believe) the following reason. Its drift tracking kept it ahead throughout the rst part of the run, but its window was too small for it to learn as detailed a model of the data as VFDT did by the end. This experiment shows that the data does indeed contain concept drift, and that CVFDT's ability to respond to the drift gives it an advantage over VFDT. The next step is to run CVFDT with dierent, perhaps dynamic, window sizes to further evaluate the nature of the drift. We also plan to evaluate CVFDT over traces longer than a week. 5. RELATED WORK Schlimmer and Granger's [23] STAGGER system was one of the rst to explicitly address the problem of concept drift. Salganico [21] studied drift in the context of nearest-neighbor learning. Widmer and Kubat's [27] FLORA system used a window of examples, but also stored old concept descriptions and reactivated them if they seemed to be appropriate again. All of these systems were only applied to small databases (by today's standards). Kelly, Hand, and Adams [14] addressed the issue of drifting parameters in probability distributions. Theoretical work on concept drift includes [16] and [3]. Ganti, Gehrke, and Ramakrishnan's [11] DEMON frame-work is designed to help adapt incremental learning algorithms to work eectively with time-changing data streams. DEMON diers from CVFDT by assuming data arrives pe- riodically, perhaps daily, in large blocks, while CVFDT deals with each example as it arrives. The framework uses o-line processing time to mine interesting subsets of the available data blocks. In earlier work [12] Gehrke, Ganti, and Ramakrishnan presented an incremental decision tree induction algorithm, BOAT, which works in the DEMON framework. BOAT is able to incrementally maintain a decision tree equivalent to the one that would be learned by a batch decision tree induction system. When the underlying concept is stable, BOAT can perform this maintenance extremely quickly. When drift is present, BOAT must discard and regrow portions of its induced tree. This can be very expensive when the drift is large or aects nodes near the root of the tree. CVFDT avoids the problem by using alternate trees and removing the restriction that it learn exactly the tree that a batch system would. A comparison between BOAT and CVFDT is an area for future work. There has been a great deal of work on incrementally maintaining association rules. Cheung, Han, Ng, and Wong and Fazil, Tansel, and Arkun [2] propose algorithms for maintaining sets of association rules when new transactions are added to the database. Sarda and Srinivas [22] have also done some work in the area. DEMON's contribution [11] is particularly relevant, as it addresses association rule maintenance specically in the high-speed data stream domain where blocks of transactions are added and deleted from the database on a regular basis. Aspects of the concept drift problem are also addressed in the areas of activity monitoring [10], active data mining [1] and deviation detection [6]. The main goal here is to explicitly detect changes, rather than simply maintain an up-to-date concept, but techniques for the latter can obviously help in the former. Several pieces of research on concept drift and context-sensitive learning are collected in a special issue of the journal Machine Learning [28]. Other relevant research appeared in the ICML-96 Workshop on Learning in Context-Sensitive Domains [15], the AAAI-98 Workshop on AI Approaches to Time-Series Problems [8], and the NIPS-2000 Workshop on Real-Time Modeling for Complex Learning Tasks [26]. Turney [25] maintains an online bibliography on context-sensitive learning. 6. FUTURE WORK We plan to apply CVFDT to more real-world problems; its ability to adjust to concept changes should allow it to perform very well on a broad range of tasks. CVFDT may be a useful tool for identifying anomalous situations. Currently CVFDT discards subtrees that are out-of-date, but some concepts change periodically and these subtrees may become useful again { identifying these situations and taking advantage of them is another area for further study. Other areas for study include: comparisons with related systems; continuous attributes; weighting examples; partially forgetting examples by allowing their weights to decay; simulating weights by subsampling; and controlling the weight decay function according to external information about drift. 7. CONCLUSION This paper introduced CVFDT, a decision-tree induction system capable of learning accurate models from the most demanding high-speed, concept-drifting data streams. CVFDT is able to maintain a decision-tree up-to-date with a window of examples by using a small, constant amount of time for each new example that arrives. The resulting accuracy is similar to what would be obtained by reapplying a conventional learner to the entire window every time a new example arrives. Empirical studies show that CVFDT is eectively able to keep its model up-to-date with a massive data stream even in the face of large and frequent concept shifts. A preliminary application of CVFDT to a real world domain shows promising results. 8. ACKNOWLEDGMENTS This research was partly supported by a gift from the Ford Motor Company, and by NSF CAREER and IBM Faculty awards to the third author. 9. --R Active data mining. Learning changing concepts by exploiting the structure of change. Megainduction: Machine Learning on Very Large Databases. Mining surprising patterns using temporal description length. Maintenance of discovered association rules in large databases: An incremental updating technique. Mining high-speed data streams Activity monitoring: Noticing interesting changes in behavior. DEMON: Mining and monitoring evolving data. BOAT: optimistic decision tree construction. The impact of changing populations on classi The complexity of learning according to two models of a drifting environment. SLIQ: A fast scalable classi Decision theoretic subsampling for induction on large databases. An adaptive algorithm for incremental mining of association rules. SPRINT: A scalable parallel classi Learning in the presence of concept drift and hidden contexts. Special issue on context sensitivity and concept drift. --TR C4.5: programs for machine learning Learning in the presence of concept drift and hidden contexts BOATMYAMPERSANDmdash;optimistic decision tree construction Activity monitoring An efficient algorithm to update large itemsets with early pruning The impact of changing populations on classifier performance The Complexity of Learning According to Two Models of a Drifting Environment Mining high-speed data streams Learning Changing Concepts by Exploiting the Structure of Change Maintenance of Discovered Association Rules in Large Databases Mining Surprising Patterns Using Temporal Description Length An Adaptive Algorithm for Incremental Mining of Association Rules DEMON --CTR Ying Yang , Xindong Wu , Xingquan Zhu, Combining proactive and reactive predictions for data streams, Proceeding of the eleventh ACM SIGKDD international conference on Knowledge discovery in data mining, August 21-24, 2005, Chicago, Illinois, USA Charu C. Aggarwal , Jiawei Han , Jianyong Wang , Philip S. Yu, On demand classification of data streams, Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining, August 22-25, 2004, Seattle, WA, USA Francisco Ferrer-Troyano , Jesus S. Aguilar-Ruiz , Jose C. Riquelme, Incremental rule learning based on example nearness from numerical data streams, Proceedings of the 2005 ACM symposium on Applied computing, March 13-17, 2005, Santa Fe, New Mexico Francisco Ferrer-Troyano , Jesus S. Aguilar-Ruiz , Jose C. Riquelme, Data streams classification by incremental rule learning with parameterized generalization, Proceedings of the 2006 ACM symposium on Applied computing, April 23-27, 2006, Dijon, France Joong Hyuk Chang , Won Suk Lee, Finding recently frequent itemsets adaptively over online transactional data streams, Information Systems, v.31 n.8, p.849-869, December 2006 Joo Gama , Pedro Medas , Pedro Rodrigues, Learning decision trees from dynamic data streams, Proceedings of the 2005 ACM symposium on Applied computing, March 13-17, 2005, Santa Fe, New Mexico Joo Gama , Ricardo Rocha , Pedro Medas, Accurate decision trees for mining high-speed data streams, Proceedings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining, August 24-27, 2003, Washington, D.C. Yi Zhang , Xiaoming Jin, An automatic construction and organization strategy for ensemble learning on data streams, ACM SIGMOD Record, v.35 n.3, p.28-33, September 2006 Francisco Ferrer-Troyano , Jess S. Aguilar-Ruiz , Jos C. Riquelme, Prototype-based mining of numeric data streams, Proceedings of the ACM symposium on Applied computing, March 09-12, 2003, Melbourne, Florida Brain Babcock , Mayur Datar , Rajeev Motwani , Liadan O'Callaghan, Maintaining variance and k-medians over data stream windows, Proceedings of the twenty-second ACM SIGMOD-SIGACT-SIGART symposium on Principles of database systems, p.234-243, June 09-11, 2003, San Diego, California Francisco Ferrer-Troyano , Jess S. Aguilar-Ruiz , Jos C. Riquelme, Discovering decision rules from numerical data streams, Proceedings of the 2004 ACM symposium on Applied computing, March 14-17, 2004, Nicosia, Cyprus Nilesh Dalvi , Pedro Domingos , Mausam , Sumit Sanghai , Deepak Verma, Adversarial classification, Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining, August 22-25, 2004, Seattle, WA, USA M. Otey , S. Parthasarathy , A. Ghoting , G. Li , S. Narravula , D. Panda, Towards NIC-based intrusion detection, Proceedings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining, August 24-27, 2003, Washington, D.C. George Forman, Tackling concept drift by temporal inductive transfer, Proceedings of the 29th annual international ACM SIGIR conference on Research and development in information retrieval, August 06-11, 2006, Seattle, Washington, USA Shi Zhong, Efficient streaming text clustering, Neural Networks, v.18 n.5-6, p.790-798, June 2005 Wei Fan, StreamMiner: a classifier ensemble-based engine to mine concept-drifting data streams, Proceedings of the Thirtieth international conference on Very large data bases, p.1257-1260, August 31-September 03, 2004, Toronto, Canada Geoff Hulten , Pedro Domingos, Mining complex models from arbitrarily large databases in constant time, Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining, July 23-26, 2002, Edmonton, Alberta, Canada Anand Narasimhamurthy , Ludmila I. Kuncheva, A framework for generating data to simulate changing environments, Proceedings of the 25th conference on Proceedings of the 25th IASTED International Multi-Conference: artificial intelligence and applications, p.384-389, February 12-14, 2007, Innsbruck, Austria Orna Raz , Philip Koopman , Mary Shaw, Semantic anomaly detection in online data sources, Proceedings of the 24th International Conference on Software Engineering, May 19-25, 2002, Orlando, Florida Yunyue Zhu , Dennis Shasha, Efficient elastic burst detection in data streams, Proceedings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining, August 24-27, 2003, Washington, D.C. Rouming Jin , Gagan Agrawal, Efficient decision tree construction on streaming data, Proceedings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining, August 24-27, 2003, Washington, D.C. Graham Cormode , Mayur Datar , Piotr Indyk , S. Muthukrishnan, Comparing data streams using Hamming norms (how to zero in), Proceedings of the 28th international conference on Very Large Data Bases, p.335-345, August 20-23, 2002, Hong Kong, China Jimeng Sun , Dacheng Tao , Christos Faloutsos, Beyond streams and graphs: dynamic tensor analysis, Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining, August 20-23, 2006, Philadelphia, PA, USA Wei-Guang Teng , Ming-Syan Chen , Philip S. Yu, A regression-based temporal pattern mining scheme for data streams, Proceedings of the 29th international conference on Very large data bases, p.93-104, September 09-12, 2003, Berlin, Germany Haixun Wang , Jian Yin , Jian Pei , Philip S. Yu , Jeffrey Xu Yu, Suppressing model overfitting in mining concept-drifting data streams, Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining, August 20-23, 2006, Philadelphia, PA, USA Hillol Kargupta , Byung-Hoon Park , Sweta Pittie , Lei Liu , Deepali Kushraj , Kakali Sarkar, MobiMine: monitoring the stock market from a PDA, ACM SIGKDD Explorations Newsletter, v.3 n.2, January 2002 Wei Fan, Systematic data selection to mine concept-drifting data streams, Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining, August 22-25, 2004, Seattle, WA, USA Lilian Harada, Detection of complex temporal patterns over data streams, Information Systems, v.29 n.6, p.439-459, September 2004 Themistoklis Palpanas , Dimitris Papadopoulos , Vana Kalogeraki , Dimitrios Gunopulos, Distributed deviation detection in sensor networks, ACM SIGMOD Record, v.32 n.4, December Joong Hyuk Chang , Won Suk Lee, Efficient mining method for retrieving sequential patterns over online data streams, Journal of Information Science, v.31 n.5, p.420-432, October 2005 Sreenivas Gollapudi , D. Sivakumar, Framework and algorithms for trend analysis in massive temporal data sets, Proceedings of the thirteenth ACM international conference on Information and knowledge management, November 08-13, 2004, Washington, D.C., USA Kevin B. Pratt , Gleb Tschapek, Visualizing concept drift, Proceedings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining, August 24-27, 2003, Washington, D.C. Joo Gama , Ricardo Fernandes , Ricardo Rocha, Decision trees for mining data streams, Intelligent Data Analysis, v.10 n.1, p.23-45, January 2006 Charu C. Aggarwal, On Change Diagnosis in Evolving Data Streams, IEEE Transactions on Knowledge and Data Engineering, v.17 n.5, p.587-600, May 2005 Daniel Kifer , Shai Ben-David , Johannes Gehrke, Detecting change in data streams, Proceedings of the Thirtieth international conference on Very large data bases, p.180-191, August 31-September 03, 2004, Toronto, Canada Mohamed Medhat Gaber , Shonali Krishnaswamy , Arkady Zaslavsky, Cost-efficient mining techniques for data streams, Proceedings of the second workshop on Australasian information security, Data Mining and Web Intelligence, and Software Internationalisation, p.109-114, January 01, 2004, Dunedin, New Zealand Graham Cormode , Mayur Datar , Piotr Indyk , S. Muthukrishnan, Comparing Data Streams Using Hamming Norms (How to Zero In), IEEE Transactions on Knowledge and Data Engineering, v.15 n.3, p.529-540, March Malu Castellanos , Fabio Casati , Umeshwar Dayal , Ming-Chien Shan, A Comprehensive and Automated Approach to Intelligent Business Processes Execution Analysis, Distributed and Parallel Databases, v.16 n.3, p.239-273, November 2004 Joong Hyuk Chang , Won Suk Lee, estWin: Online data stream mining of recent frequent itemsets by sliding window method, Journal of Information Science, v.31 n.2, p.76-90, April 2005 Tho Manh Nguyen , Josef Schiefer , A. Min Tjoa, Sense & response service architecture (SARESA): an approach towards a real-time business intelligence solution and its use for a fraud detection application, Proceedings of the 8th ACM international workshop on Data warehousing and OLAP, November 04-05, 2005, Bremen, Germany Haixun Wang , Wei Fan , Philip S. Yu , Jiawei Han, Mining concept-drifting data streams using ensemble classifiers, Proceedings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining, August 24-27, 2003, Washington, D.C. Yixin Chen , Guozhu Dong , Jiawei Han , Benjamin W. Wah , Jianyong Wang, Multi-dimensional regression analysis of time-series data streams, Proceedings of the 28th international conference on Very Large Data Bases, p.323-334, August 20-23, 2002, Hong Kong, China Jiawei Han , Yixin Chen , Guozhu Dong , Jian Pei , Benjamin W. Wah , Jianyong Wang , Y. Dora Cai, Stream Cube: An Architecture for Multi-Dimensional Analysis of Data Streams, Distributed and Parallel Databases, v.18 n.2, p.173-197, September 2005 Weng-Keen Wong , Andrew Moore , Gregory Cooper , Michael Wagner, What's Strange About Recent Events (WSARE): An Algorithm for the Early Detection of Disease Outbreaks, The Journal of Machine Learning Research, 6, p.1961-1998, 12/1/2005 Yasushi Sakurai , Spiros Papadimitriou , Christos Faloutsos, BRAID: stream mining through group lag correlations, Proceedings of the 2005 ACM SIGMOD international conference on Management of data, June 14-16, 2005, Baltimore, Maryland Chang-Tien Lu , Yufeng Kou , Jiang Zhao , Li Chen, Detecting and tracking regional outliers in meteorological data, Information Sciences: an International Journal, v.177 n.7, p.1609-1632, April, 2007 Yang , Li Lee , Wynne Hsu, Finding hot query patterns over an XQuery stream, The VLDB Journal The International Journal on Very Large Data Bases, v.13 n.4, p.318-332, December 2004 Marcus A. Maloof , Ryszard S. Michalski, Incremental learning with partial instance memory, Artificial Intelligence, v.154 n.1-2, p.95-126, April 2004 Jrgen Beringer , Eyke Hllermeier, Online clustering of parallel data streams, Data & Knowledge Engineering, v.58 n.2, p.180-204, August 2006 Streaming pattern discovery in multiple time-series, Proceedings of the 31st international conference on Very large data bases, August 30-September 02, 2005, Trondheim, Norway Brian Babcock , Shivnath Babu , Mayur Datar , Rajeev Motwani , Jennifer Widom, Models and issues in data stream systems, Proceedings of the twenty-first ACM SIGMOD-SIGACT-SIGART symposium on Principles of database systems, June 03-05, 2002, Madison, Wisconsin Zhiyuan Chen , Chen Li , Jian Pei , Yufei Tao , Haixun Wang , Wei Wang , Jiong Yang , Jun Yang , Donghui Zhang, Recent progress on selected topics in database research: a report by nine young Chinese researchers working in the United States, Journal of Computer Science and Technology, v.18 n.5, p.538-552, September single-pass mining of path traversal patterns over streaming web click-sequences, Computer Networks: The International Journal of Computer and Telecommunications Networking, v.50 n.10, p.1474-1487, 14 July 2006 Mohamed Medhat Gaber , Arkady Zaslavsky , Shonali Krishnaswamy, Mining data streams: a review, ACM SIGMOD Record, v.34 n.2, June 2005 Shivnath Babu , Jennifer Widom, Continuous queries over data streams, ACM SIGMOD Record, v.30 n.3, September 2001 Lukasz Golab , M. Tamer zsu, Issues in data stream management, ACM SIGMOD Record, v.32 n.2, p.5-14, June Venkatesh Ganti , Johannes Gehrke , Raghu Ramakrishnan, Mining data streams under block evolution, ACM SIGKDD Explorations Newsletter, v.3 n.2, January 2002

data streams;incremental learning;concept drift;decision trees;subsampling;hoeffding bounds

502532

Robust space transformations for distance-based operations.

For many KDD operations, such as nearest neighbor search, distance-based clustering, and outlier detection, there is an underlying &kgr;-D data space in which each tuple/object is represented as a point in the space. In the presence of differing scales, variability, correlation, and/or outliers, we may get unintuitive results if an inappropriate space is used.The fundamental question that this paper addresses is: "What then is an appropriate space?" We propose using a robust space transformation called the Donoho-Stahel estimator. In the first half of the paper, we show the key properties of the estimator. Of particular importance to KDD applications involving databases is the stability property, which says that in spite of frequent updates, the estimator does not: (a) change much, (b) lose its usefulness, or (c) require re-computation. In the second half, we focus on the computation of the estimator for high-dimensional databases. We develop randomized algorithms and evaluate how well they perform empirically. The novel algorithm we develop called the Hybrid-random algorithm is, in most cases, at least an order of magnitude faster than the Fixed-angle and Subsampling algorithms.

INTRODUCTION For many KDD operations, such as nearest neighbor search, distance-based clustering, and outlier detection, there is an underlying k-D data space in which each tuple/object is represented as a point in the space. Often times, the tuple represented simply as the point in the k-D space. More formally, the transformation from the tuple t to the point p t is the identity matrix. We begin by arguing that the identity transformation is not appropriate for many distance-based operations, Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. SIGKDD 2001 San Francisco, California USA particularly in the presence of variability, correlation, out- liers, and/or diering scales. Consider a dataset with the following attributes: systolic blood pressure (typical range: 100-160 mm of mercury, with mean =120) body temperature degrees Celsius, with a very small standard deviation (e.g., 1-2 degrees for sick pa- age (range: 20-50 years of age in this example) Note that dierent attributes have dierent scales and units (e.g., mm of Hg vs. degree Celsius), and dierent variability (e.g., high variability for blood pressure vs. low variability for body temperature). Also, attributes may be correlated (e.g., age and blood pressure), and there may be outliers. Example Operation 1 (nearest neighbor search). Consider a nearest neighbor search using the Euclidean distance function in the original data space, i.e., the identity transformation. The results are likely to be dominated by blood pressure readings, because their variability is much higher than that of the other attributes. Consider the query point (blood pressure Using Euclidean distance, the point (120, 40, 35) is nearer to the query point than (130, 37, 35) is. But, in terms of similarity/dissimilarity, this nding is not very meaningful because, intuitively, a body temperature of 40 degrees is far away from a body temperature of 37 degrees; in fact, a person with a body temperature of 40 degrees needs medical attention immediately! A simple x to the above problem is to somehow weight the various attributes. One common approach is to apply a \normalization" transformation, such as normalizing each attribute into the range [0,1]. This is usually not a satisfactory solution because a single outlier (e.g., blood pressure = could cause virtually all other values to be contained in a small subrange, again making the nearest neighbor search produce less meaningful results. Another common x is to apply a \standardization" trans- formation, such as subtracting the mean from each attribute and then dividing by its standard deviation. While this transformation is superior to the normalization transforma- tion, outliers may still be too in uential in skewing the mean and the standard deviation. Equally importantly, this transformation does not take into account possible correlation between attributes. For example, older people tend to have higher blood pressure than younger people. This means that we could be \double counting" when determining distances. Example Operation 2 (Data Mining Operations). Data clustering is one of the most studied operations in data mining. As an input to a clustering algorithm, a distance function is specied. Although some algorithms can deal with non-metric distance functions (e.g., CLARANS [20]), most algorithms require metric ones. Among those, a sub-class of algorithms that has received a lot of attention recently is the class of density-based algorithms (e.g., DBSCAN [11] and DENCLUE [14]). The density of a region is computed based on the number of points contained in a xed size neighborhood. Thus, density calculation can be viewed as a xed radius search. Hence, all the concerns raised above for nearest neighbor search apply just the same for density-based clustering. Outlier detection is another important operation in data mining, particularly for surveillance applications. Many outlier detection algorithms are distance- or density-based [16, 21, 7]. Again, the issues of diering scale, variability, corre- lation, and outliers could seriously aect the eectiveness of those algorithms. At rst glance, the statement that outliers could impact the eectiveness of outlier detection algorithms may seem odd. But if attention is not paid to outliers, it is possible that the outliers may aect the quantities used to scale the data, eectively masking (hiding) themselves [3]. Contributions of this Paper The fundamental question addressed in this paper is: \What is an appropriate space in the presence of diering scale, variability, correla- tion, and outliers?" So far, we have seen that the spaces associated with the identity, normalization, and standardization transformations are inadequate. In this paper, we focus on robust space transformations, or robust estimators, so that for distance computation, all points in the space are treated \fairly". Specically: Among many robust space estimators that have been studied in statistics, we propose using the Donoho- Stahel estimator (DSE). In Section 3, we show two important properties of the DSE. The rst is the Euclidean property. It says that while inappropriate in the original space, the Euclidean distance function becomes reasonable in the DSE transformed space. The second, and arguably the more important, property is the stability property. It says that the transformed space is robust against updates. That is, in spite of frequent updates, the transformed space does not lose its usefulness and requires no re-computation. Stability is a particularly meaningful property for KDD applications. If an amount of eort x was spent to set up an index in the transformed space, we certainly would not like to spend another amount x after every single update to the database. In Section 3, we give experimental results showing that the DSE transformed space is so stable that it can easily withstand adding many more tuples to the database (e.g., 50% of the database size). Having shown its key properties, in the second half of this paper, we focus on the computation of the DSE for high-dimensional (e.g., 10 attributes) databases. The original DSE algorithm was dened independently by both Donoho and Stahel [25]; we refer to it as the Fixed-angle algorithm. In Section 4, we show that the original algorithm does not scale well with dimension- ality. Stahel also proposed a version of the algorithm which uses subsampling (i.e., taking samples of sam- ples) [25]. However, the number of subsamples to be used in order to obtain good results is not well known. We follow the work of Rousseeuw on least median of squares [22], and come up with a heuristic that seems to work well, as shown in Section 6. For comparison purposes, we have implemented this algorithm, applied some heuristics (e.g., number of subsamples), and evaluated eectiveness and e-ciency. Last but not least, in Section 5, we develop a new algorithm, which we refer to as the Hybrid-random algorithm, for computing the DSE. Our experimental results show that the Hybrid-random algorithm is at least an order of magnitude more e-cient than the Fixed-angle and Subsampling algorithms. Fur- thermore, to support the broader claim that the DSE transformation should be used for KDD operations, the Hybrid-random algorithm can run very e-ciently (e.g., compute the estimator for 100,000 5-D tuples in tens of seconds of total time). Related Work Space transformations have been studied in the database and KDD literature. However, they are from the class of distance-preserving transformations (e.g., [12]), where the objective is to reduce the dimensionality of the space. As far as space transformations go, our focus is not so much on preserving distances, but on providing robustness and stability. Principal component analysis (PCA) is useful for data re- duction, and is well-studied in the statistics literature [17, 15, 10]. The idea is to nd linear combinations of the at- tributes, while either maximizing or minimizing the variabil- ity. Unfortunately, PCA is not robust since a few outliers can radically aect the results. Outliers can also be masked (hidden by other points). Moreover, PCA lacks the stability requirements that we desire (cf: Section 3). SVD is not robust either; it, too, may fail to detect outliers due to masking. Many clustering algorithms have been proposed in recent years, and most are distance-based or density-based [11, 26, 1, 14]. The results presented in this paper will improve the eectiveness of all these algorithms in producing more meaningful clusters. Outlier detection has received considerable attention in recent years. Designed for large high-dimensional datasets, the notion of DB-outliers introduced in [16] is distance- based. A variation of this notion is considered in [21]. The notion of outliers studied in [7] is density-based. Again, all of these notions and detection algorithms will benet from the results presented in this paper. Developing eective multi-dimensional indexing structures is the subject of numerous studies [13, 4, 6]. However, this paper is not about indexing structures. Instead, we focus on determining an appropriate space within which an index is to be created. In [24], nearest neighbor search based on quadratic form distance functions is considered, that is, distances computed using some matrix A. That study assumes prior knowledge of A. For some applications, A may be data-independent and well-known. For example, for a distance between two color histograms, each entry in A represents the degree of perceptual similarity between two colors [12]. However, for most applications, it is far from clear what a suitable A could be. The focus of this paper is to propose a meaningful way of picking A in a data-dependent fashion. 2. BACKGROUND: DONOHO-STAHEL If two similar attributes are being compared, and those attributes are independent and have the same scale and vari- ability, then all points within distance D of a point P lie within the circle of radius D centered at P . In the presence of diering scales, variability, and correlation, all points within distance D of a point P lie within an ellipse. If there is no correlation, then the major and minor axes of the ellipse lie on the standard coordinate axes; but, if there is correlation, then the ellipse is rotated through some angle . (See Fig. 4 later.) This generalizes to higher dimen- sions. In 3-D, the ellipsoid resembles a football, with the covariance determining the football's size, and the correlation determining its orientation. An estimator A, also called a scatter matrix, is a k k square matrix, where k is the dimensionality of the original data space. An estimator is related to an ellipsoid as follows. Suppose x and y are k-dimensional column vectors. The Euclidean distance between x and y can be expressed as where T denotes the transpose operator. A quadratic form distance function can be expressed as dA(x; y). For x 6= 0, and x T Ax > 0, A is called a positive denite matrix, and x T Ax yields an ellipsoid. Donoho-Stahel Estimator and Fixed-angle Algorithm The DSE is a robust multivariate estimator of location and scatter. Essentially, it is an \outlyingness-weighted" mean and covariance, which downweights any point that is many robust standard deviations away from the sample in some univariate projection [19, 22]. This estimator also possesses desirable statistical properties such as a-ne equivariance, which not all transformations (e.g., principal component analysis) possess. The DSE is our estimator of choice, although there are many estimators to choose from [22]. For example, we chose the DSE over the Minimum Volume Ellipsoid estimator because the DSE is easier to compute, scales better, and has much less bias (especially for dimensions > 2) [18]. In our extended work, we consider other robust estimators, but the one that seems to perform the best (based on numerous simulations and analyses) is the DSE. In the interest of space, we only deal with the DSE in this paper. Although the application we focus on here is outlier detection, we add that the DSE is a general transformation that is useful for many applications, such as those described in Section 1. Fig. 1 gives a skeleton of the initial algorithm proposed by Stahel [25] for computing the estimator in 2-D. Let us step through the algorithm to understand how the estimator is dened. The input is a dataset containing N 2-D points of the form y In step 1, we iterate through the unit circle, to consider a large number of possible an- gles/directions , on which to project. We iterate through degrees rather than 360 degrees since the 180-360 degree range is redundant. Hereafter, we call this algorithm The Fixed-angle Donoho Stahel Algorithm 1. For to (i.e., 0 < ) using some small increment (e.g., 1 degree), do: (a) For is the unit vector) (b) Compute (c) Compute (The MAD is dened to be 1:4826 (medianjx i () m()j)). (d) For 2. For 3. Compute the robust multivariate centre ^ and the weighting function w(t) is dened as follows: 2:5 4. Compute the robust covariance matrix ^ 5. Return the Donoho-Stahel estimator of location and scatter: Figure 1: The DSE Fixed-angle Algorithm for 2-D the Fixed-angle algorithm. For each , each point is projected onto the line corresponding to rotating the x-axis by , giving the value x i (). Mathematically, this is given by the dot product between y i and u, which is the unit vector We call u the projection vector. In step 1(b), we compute m(), which is the median of all the x i () values. MAD is an acronym for median absolute deviation from the median. It is a better estimator of scatter than the standard deviation in the presence of out- liers. Finally, step 1(d) yields d i (), which measures how outlying the projection of y i is with respect to . Note that d i () is analogous to classical standardization, where each value x i () is standardized to x i () , with and being the mean and the standard deviation of x i (), respectively. By replacing the mean with the median, and the standard deviation with the MAD, d i () is more robust against the inuence of outliers than the value obtained by classical standardization Robustness is achieved by rst identifying outlying points, and then downweighting their in uence. Step 1 computes for each point and each angle , the degree of outlyingness of the point with respect to . As a measure of how outlying each point is over all possible angles, step 2 computes, for each point, the maximum degree of outlyingness over all possible 's. In step 3, if this maximum degree for a point is too high (our threshold is 2.5), the in uence of this point is weakened by a decreasing weight function. Finally, with all points weighted accordingly, the location center ^ R and the covariance matrix ^ R are computed. 3. KEY PROPERTIES OF THE DSE In this section, we examine whether the estimator is useful for distance-based operations in KDD applications. In Section 6, we provide experimental results showing the difference the estimator can make. But rst, in this section, we conduct a more detailed examination of the properties of the estimator. We show that the estimator possesses the Euclidean property and the stability property, both of which are essential for database applications. Euclidean Property In this section, we show that once the DSE transformation has been applied, the Euclidean distance function becomes readily applicable. This is what we call the Euclidean property. Lemma 1. The Donoho-Stahel estimator of scatter, ^ R , is a positive denite matrix. The proof is omitted for brevity. According to standard matrix algebra [2], the key implication of the above lemma is that the matrix ^ R can be decomposed into: ^ where is a diagonal matrix whose entries are the eigen- values, and Q is the matrix containing the eigenvectors of R . This decomposition is critical to the following lemma. It says that the quadratic form distance wrt ^ R between two vectors x and y is the same as the Euclidean distance between the transformed vectors in the transformed space. Lemma 2. Let x; y be two vectors in the original space. Suppose they are transformed into the space described by ^ R , Then, the quadratic form distance wrt ^ R is equal to the Euclidean distance between xR and yR . Proof: R The proof is rather standard, but we include it to provide a context for these comments: For each vector x in the original space (or tuple in the relation), each vector is transformed only once, i.e., Future operations do not require any extra transformations. For example, for indexing, all tuples are transformed once and can be stored in an indexing structure. When a query point z is given, z is similarly transformed to zR . From that point on, the Euclidean distance function can be used for the transformed vectors (e.g., xR and zR ). Furthermore, many existing distance-based structures are the most e-cient or eective when dealing with Euclidean-based calculations. Examples include R-trees and variants for indexing [13, 4], and the outlier detection algorithm studied in [16]. The key message here is that space transformation wrt ^ R is by itself not expensive to compute, and can bring further e-ciency/eectiveness to subsequent processing. Stability Property The second property we analyze here for the DSE concerns stability. A transformation is stable if the transformed space does not lose its usefulness| even in the presence of frequent updates. This is an important issue for database applications. If an amount of eort x was spent in setting up an index in the transformed space, we certainly would not like to spend another amount x after every single update to the index. In statistics, there is the notion of a breakdown point of an estimator, which quanties the proportion of the dataset that can be contaminated without causing the estimator to become \arbitrarily absurd" [27]. But we do not pursue this formal approach regarding breakdown points; instead, we resort to experimental evaluation. In our experiments, we used a real dataset D and computed the R (D). We then inserted or deleted tuples from D, thereby changing D to Dnew . To measure stabil- ity, we compared matrix ^ R (D) with ^ R(Dnew ). In the numerical computation domain, there are a few heuristics for measuring the dierence between matrices; but there is no universally agreed-upon metric [9]. To make our comparison more intuitive, we instead picked a distance-based operation|outlier detection|and compared the results. Section 6 gives the details of our experiments, but in brief, we proceeded as follows: (a) We used the old estimator ^ to transform the space for Dnew and then found all the outliers in Dnew ; and (b) We used the updated estimator (Dnew ) to transform the space for Dnew , and then found all the outliers in Dnew . To measure the dierence between the two sets of detected outliers, we use standard precision and recall [23], and we dene: (i) the answer set as the set of outliers found by a given algorithm, and (ii) the target set as the \o-cial" set of outliers that are found using a su-ciently exhaustive search (i.e., using the Fixed-angle algorithm with a relatively small angular increment). Precision is the percentage of the answer set that is actually found in the target set. Recall is the percentage of the target set that is in the answer set. Ideally, we want 100% precision and 100% recall. Fig. 2 shows the results when there were 25%, 50%, 75% and 100% new tuples added to D, and when 25%, 50% and 75% of the tuples in D were deleted from D. The new tuples were randomly chosen and followed the distribution of the tuples originally in D. The deleted tuples were randomly chosen from D. The second and third columns show the number of outliers found, with the third column giving the \real" answer, and the second column giving the \approximated" answer using the old transformation. The fourth and fth columns show the precision and recall. They clearly show that the DSE transformation is stable. Even a 50% change in the database does not invalidate the old transformation, and re-computation appears unnecessary. For the results shown in Fig. 2, the newly added tuples followed the same distribution as the tuples originally in D. For the results shown in Fig. 3, we tried a more drastic sce- nario: the newly added tuples, called junk tuples, followed a totally dierent distribution. This is re ected by the relatively higher numbers in the second and third columns of Fig. 3. Nevertheless, despite the presence of tuples from two distributions, the precision and recall gures are still close to 100%. This again shows the stability of the DSE. % Change in # of Outliers in Dnew , using: Precision Recall (Dnew 25% Inserts 17 15 88.2% 100% 50% Inserts 17 75% Inserts 100% Inserts 37 29 78.4% 100% 25% Deletes 13 13 100% 100% 50% Deletes 75% Deletes 15 19 100% 78.9% Figure 2: Precision and Recall: Same Distribution of Outliers in Inserted when: Dnew , using: Precision Recall (Dnew 25% 53 52 94.3% 96.2% 37.5% 74 70 91.9% 97.1% 50% 95 90 92.6% 97.8% 62.5% 108 100 90.7% 98.0% Figure 3: Precision and Recall: Drastically Dierent Distribution 4. K-D SUBSAMPLING ALGORITHM In the previous section, we showed that the DSE possesses the desirable Euclidean and stability properties for KDD applications. The remaining question is whether the associated cost is considerable. Let us consider how to compute R e-ciently, for k > 2 dimensions. Complexity of the Fixed-angle Algorithm Recall that Fig. 1 gives the 2-D Fixed-angle algorithm proposed by Donoho and Stahel. The extension of this algorithm to 3-D and beyond is straightforward. Instead of using a unit circle, we use a unit sphere in 3-D. Thus, there are two angles|1 and 2|through which to iterate. Similarly, in k-D, we deal with a unit hypersphere, and there are k 1 angles through which to iterate: To understand the performance of the Fixed-angle algo- rithm, we conduct a complexity analysis. In step 1 of Fig. 1, each angle requires nding the median of N values, where N is the size of the dataset. Finding the median takes O(N) time, which is the time that a selection algorithm can partition an array to nd the median entry. (Note that sorting is not needed.) Thus, in 2-D, if there are a increments to iterate through, the complexity of the rst step is O(aN ). For k-D, there are k 1 angles to iterate through. If there are a increments for each of these angles, the total complexity of the rst step is O(a k 1 kN ). In step 2, in the k-D case, there are a k 1 projection vectors to evaluate. Thus, the complexity of this step is nds a robust center, which can be done in O(kN) time. Step 4 sets up the k k robust covariance matrix, which takes O(k 2 N) time. Hence, the total complexity of the Fixed-angle algorithm is O(a k 1 kN ). Su-ce it to say that running this basic k-D algorithm is impractical for larger values of a and k. Intuition behind the Subsampling Algorithm in 2-D The rst two steps of the Fixed-angle algorithm compute, for each point y i , the degree of \outlyingness" d i . The value of d i is obtained by taking the maximum value of d i () over all 's, where d i () measures how outlying the projection of y i is wrt . In the Fixed-angle algorithm, there is an exhaustive enumeration of 's. For high dimensions, this approach is infeasible. Let us see if there is a better way to determine \good" projection vectors. Consider points A, B, and C in Fig. 4(a), which shows a 2-D scenario involving correlated attributes. Fig. 4(b) shows the projection of points onto a line orthogonal to the major axis of the ellipse. (Not all points are projected in the gure.) Note that B's projection appears to belong to the bulk of the points projected down from the ellipse; it does not appear to be outlying at all in this projection. Also, although A is outlying on the projection Algorithm Subsampling For is the number of iterations chosen, do: (a) Select k 1 random points from the dataset. Together with the origin, these points form a hyper-plane through the origin, and a subspace V . i. Compute a basis for the orthogonal complement of V . ii. Choose a unit vector u i from the orthogonal complement to use as vector u in the Fixed- angle algorithm. (b) For (c) (continue with step 1(b) and beyond of the Fixed- angle algorithm shown in Fig. 1, where i takes the role of ) Figure 5: DSE Subsampling Algorithm for k-D line, C is not. In Fig. 4(c), A and C are not outlying, but B clearly is. Fig. 4(d) shows yet another projection. As can be seen, the projection vectors which are chosen greatly in uence the values for d i in the Donoho-Stahel al- gorithm. In applying subsampling, our goal is to use lines orthogonal to the axes of an ellipse (or ellipsoid, in k-D), to improve our odds of obtaining a good projection line. While there may be better projection vectors in which to identify outliers, these are good choices. There is an increased chance of detecting outliers using these orthogonal lines because many outliers are likely to stand out after the orthogonal projection (see Fig. 4). Non-outlying points, especially those within the ellipsoids, are unlikely to stand out because they project to a common or relatively short interval on the line. If we knew what the axes of the ellipse were, then there would be no need to do subsampling. However, since: (a) we do not know the parameters of the ellipsoid, and (b) in general, there will be too many points and too many dimensions involved in calculating the parameters of the ellipsoid, we use the following approach called subsampling. In 2-D, the idea is to rst pick a random point P from the set of N input points. Then compute a line orthogonal to the line joining P and the origin. Note that, with reasonable proba- bility, we are likely to pick a point P in the ellipse, and the resulting orthogonal line may approximate one of the axes of the ellipse. This is the essence of subsampling. More Details of the Subsampling Algorithm in k-D In k-D, we rst nd a random sample of k 1 points. Together with the origin, they form a subspace V . Next, we need to nd a subspace that is orthogonal to V , which is called the orthogonal complement of V [2]. From this point on, everything else proceeds as in the Fixed-angle algorithm. Fig. 5 outlines the Subsampling algorithm for k-D DSE computation One key detail of Fig. 5 deserves elaboration: how to compute m. To determine m, we begin by analyzing the probability of getting a \bad" subsample. For each subsample, are randomly chosen. A subsample is likely to be good if all k 1 points are within the ellipsoid. Let (a user-chosen parameter) be the fraction of points outside the ellipsoid. Typically, varies between 0.01 to 0.5; the bigger the value, the more conservative or demanding the user is on the quality of the subsamples. A (a) A (b) A (c) A (d) Figure 4: Bivariate Plots Showing the Eect of Dierent Projection Lines. (a) Data points only. (b) Projection onto a line orthogonal to the major axis of the ellipse. (c) Projection onto a line orthogonal to the minor axis. (d) Projection onto another line. Let us say that m can be the smallest number of subsamples such that there is at least a 95% probability that we get at least one good subsample out of the m subsamples. Given , the probability of getting a \good" subsample is the probability of picking all k 1 random points within the ellipsoid, which is (1 Conversely, the probability of getting a bad subsample is 1 (1 . Thus, the probability of all m subsamples being bad is (1 (1 Hence, we can determine a base value of m by solving the following inequality for m: 1 (1 (1 0:95. For example, In Section 6, we show how the quality of the estimator varies with m. Complexity of the Subsampling Algorithm In k-D, we can determine a basis for the orthogonal complement of a hyperplane through the origin and through k 1 non-zero points in O(k 3 ) time, using Gauss-Jordan elimination [2, 9]. Using this basis, we simply pick any unit vector u as our projection vector, and then continue with the basic Fixed-angle algorithm. Recall from Section 4 that the basic algorithm runs in O(a k 1 kN) time for step 1 and O(k 2 N) time for the remaining steps. For the Subsampling algorithm, however, we perform a total of m iterations, where each iteration consists of k 1 randomly selected points, and thus step 1 of the Subsampling algorithm runs in O(mk 3 ) time. Thus, following the analysis in Section 4, the entire algorithm runs in O(mk 3 5. K-D RANDOMIZED ALGORITHMS The Subsampling algorithm is more scalable with respect to k than the Fixed-angle algorithm is, but the mk 3 complexity factor is still costly when the number of subsamples m is large (i.e., for a high quality estimator). Thus, in this section, we explore how the k-D DSE estimator can be computed more e-ciently. First, we implement a simple alternative to the Fixed-angle algorithm, called Pure-random. After evaluating its strengths and weaknesses, we develop a new algorithm called Hybrid-random, which combines part of the Pure-random algorithm with part of the Subsampling algorithm. In Section 6, we provide experimental results, showing eectiveness and e-ciency. Pure-random Algorithm Recall from Fig. 1 that in the Fixed-angle algorithm, the high complexity is due to the a k 1 factor, where a k 1 denotes the number of projection unit vectors examined. However, for any given projection unit vector, the complexity of step 1 reduces drastically to O(kN ). It is certainly possible for an algorithm to do well if it randomly selects r projections to examine, and if some of those projections happen to be \good" or in uential pro- jections. A skeleton of this algorithm called Pure-random Algorithm Pure-random For where r is the number of projection vectors chosen, do: (a) Select a k-D projection unit vector u i randomly (i.e., pick k 1 random angles) (b) For (c) (continue with step 1(b) and beyond of the Fixed- angle algorithm, where i takes the role of ) Figure is presented in Fig. 6. Following the analysis shown in Section 4, it is easy to see that the complexity of the Pure- random algorithm is O(rkN randomization is also used in the Subsampling algorithm. But there, each random \draw" is a subspace V formed by 1 points from the dataset, from which the orthogonal complement of V is computed. In the Pure-random case, however, each random draw is a projection vector. In order for the Pure-random algorithm to produce results comparable to that of the Subsampling algorithm, it is very likely that r m. Hybrid-random Algorithm Conceptually, the Pure- random algorithm probes the k-D space blindly. This is the reason why the value of r may need to be high for acceptable quality. The question is whether random draws of projection vectors can be done more intelligently. More specically, are there areas of the k-D space over which the randomization can skip, or equivalently, are there areas on which the randomization should focus? In a new algorithm that we develop called Hybrid-random, we rst apply the Subsampling algorithm for a very small number of subsamples. Consider the orthogonal complement of V that passes through the origin. Imagine rotating this line through a small angle anchored at the origin, thus creating a cone. This rotation yields a \patch" on the surface of a k-D unit hypersphere. From the Fixed-angle al- gorithm, we know that projection vectors too close to each other do not give markedly dierent results. So, in the second phase of the Hybrid-random algorithm, we will restrict the random draws of projection vectors to stay clear of previously examined cones/patches. Using the Euclidean inner product and the Law of Cosines, a collision between two vectors a and b occurs if dist 2 (a; is the radius of a patch on the surface of the k-D unit hypersphere. To determine -, we used the following heuristic. We say that vectors a and b are too close to each other if cos 0:95, where is the angle between the vectors. Thus, (2-) hence, as an upper bound, we use 0:1 Two observations are in order. First, patches that are too large are counterproductive because many promising projection vectors may be excluded. Second, although increasing the number of patches improves accuracy, favourable results can be obtained with relatively few patches (e.g., 100), as will be shown in Section 6. Fig. 7 gives a skeleton of the Hybrid-random algorithm. Steps 1 to 3 use the Subsampling algorithm to nd some initial projection vectors (including the eigenvectors of the scatter matrix) and keep them in S. In each iteration of step 4, a new random projection vector is generated in such a way that it stays clear of existing projection vectors. Algorithm Hybrid-random 1. Run the Subsampling algorithm for a small number m of iterations (e.g., 2. Compute the k eigenvectors of the resulting scatter matrix. This gives us an approximation for the axes of the ellipsoid. 3. Initialize the set S of previously examined projection vectors to consist of the m projection vectors from step 1 and the k eigenvectors from step 2. 4. For where r is the number of extra random patches desired, do: (a) From S, randomly select 2 unique vectors a and b that are at least 2- radians apart. (b) Compute a new vector u i that is a linear combination of a and b. In particular, u )b, where is randomly chosen between [-, 1-]. (c) If u i is within - radians from an existing vector in S, then redo the previous step with a new If these two vectors are still too close during the second attempt, then go back to step 4(a). (d) Normalize u i so that it is a unit vector, and add it to S. (f) (continue with step 1(b) and beyond of the Fixed- angle algorithm, where i takes the role of ) Figure 7: DSE Hybrid-random Algorithm for k-D Recall from our earlier discussion that the complexity of the Subsampling algorithm is O(m1k 3 m1 is the number of subsamples taken. As for the Pure- random algorithm, the complexity is O(r1kN ), where r1 is the number of random projections probed. It is easy to see that the Hybrid-random algorithm requires a complexity of We expect that m2 m1 , and r2 r1 . Experimental results follow. 6. EXPERIMENTAL EVALUATION Experimental Setup To evaluate the Donoho-Stahel transformation, we picked the distance-based outlier detection operation described in [16]. As explained in Section 3, we use precision and recall [23], to compare the results. Our base dataset is an 855-record dataset consisting of 1995-96 National Hockey League (NHL) player performance statistics. These publicly available statistics can be down-loaded from sites such as the Professional Hockey Server at http://maxwell.uhh.hawaii.edu/hockey/. Since this real-life dataset is quite small, we created a number of synthetic datasets mirroring the distribution of statistics within the NHL dataset. Specically, we determined the distribution of each attribute in the original dataset by using a 10-partition histogram. Then, we generated datasets containing up to 100,000 tuples|whose distribution mirrored that of the base dataset. As an optional preprocessing step, we applied the Box and Cox transformation to normality [8] to nd appropriate parameters p and D for the distance-based outliers implementation. Unless otherwise stated, we used a 5-D case of 100,000 tuples as our default, where the attributes are goals, assists, penalty minutes, shots on goal, and games played. Our tests were run on Sun Microsystems Ultra-1 proces- sor, running SunOS 5.7, and having 256 MB of main mem- ory. Of the four DSE algorithms presented, only the Fixed- angle algorithm is deterministic. The other three involve randomization, so we used the median results of several runs. Precision was almost always 100%, but recall often varied. Usefulness of Donoho-Stahel Transformation In the introduction, we motivated the usefulness of the Donoho- Stahel transformation by arguing that the identity transformation (i.e., raw data), as well as the normalization and standardization transformations, may not give good results. In the experiment reported below, we show a more concrete situation based on outlier detection. Based on the 1995-96 NHL statistics, we conducted an experiment using the two attributes: penalty-minutes and goals-scored. We note that the range for penalty-minutes was [0,335], and the range for goals-scored was [0,69]. Fig. 8 compares the top outliers found using the identity, standardization, and Donoho-Stahel transformations. Also shown are the actual penalty-minutes and goals-scored by the identied players. With the identity transformation (i.e., no transformation), players with the highest penalty- minutes dominate. With classical standardization, the dominance shifts to the players with the highest goals-scored (with Matthew Barnaby appearing on both lists). How- ever, in both cases, the identied outliers are \trivial", in the sense that they are merely extreme points for some at- tribute. Barnaby, May, and Simon were all in the top-5 for penalty-minutes; Lemieux and Jagr were the top-2 for goals-scored. With the Donoho-Stahel transformation, the identied outliers are a lot more interesting and surprising. Donald Transform- Top Outliers Penalty-mins. Goals-scored ation Found (raw data) (raw data) Matthew Barnaby 335 15 Chris Simon 250 Matthew Barnaby 335 15 Standard- Jaromir Jagr 96 62 ization Mario Lemieux 54 69 Matthew Barnaby 335 15 Donoho- Donald Brashear 223 0 Stahel Jan Caloun 0 8 Joe Mullen 0 8 Figure 8: Identied Outliers: Usefulness of Donoho- Stahel Transformation Brashear was not even in the top-15 as far as penalty- minutes goes, and his goals-scored performance was unim- pressive, that is, penalty-minutes = 223 and goals-scored Yet, he has a unique combination. This is because to amass a high number of penalty minutes, a player needs to play a lot, and if he plays a lot, he is likely to score at least some goals. (Incidentally, 0 goals is an extreme univariate point; however, well over 100 players share this value.) Similar comments apply to Jan Caloun and Joe Mullen; both had 0 penalty-minutes but 8 goals-scored. While their raw gures look unimpressive, the players were exceptional in their own ways. 1 The point is, without an appropriate space transformation, these outliers would likely be missed. Internal Parameters of the Algorithms Every algorithm presented here has key internal parameters. In the Fixed-angle case, it is the parameter a, the number of angles tested per dimension. For the randomization algorithms, there are m, the number of subsamples, and r, the number of random projection vectors. Let us now examine how the choices of these parameters aect the quality of the estimator computed. Precision and recall will be used to evaluate quality. However, for the results presented below, precision was always at 100%. Thus, we only report the recall values. The four graphs in Fig. 9 each contrast: (i) CPU times, (ii) recall values, and (iii) number of iterations (or patches used) for one of the four algorithms. The left hand y-axis denes CPU times (in minutes for the top two graphs, and in seconds for the bottom two graphs). The right hand y-axis, in conjunction with the recall curve (see each gure's legend) denes recall values. Note, however, that the recall range varies from one graph to another. Fig. 9(a) measures CPU time in minutes, and shows that the Fixed-angle algorithm can take a long time to nish, especially as the number of random angles a tested increases. The horizontal axis is in tens of thousands of iterations. Recall that a small decrease in the angle increment for each dimension can cause a very large number of additional iterations to occur. For many of our datasets, it was necessary to use increments as small as degrees (e.g., 75 hours of CPU time, for 100,000 tuples in 5-D), before determining the number of outliers present. We omit these very long runs from our graphs, to allow us to more clearly contrast CPU times and recall values. Compared to the Fixed-angle algorithm, the Pure-random algorithm achieves a given level of recall more quickly, al- though, as Fig. 9(b) shows, it can still take a long time to achieve high levels of recall. Recall that, for the Subsampling algorithm, a key issue was how many subsamples to use. Based on the heuristic presented in Section 4, the base value of m was determined to be 47, and multiples of 47 subsamples were used. From the recall curve in Fig. 9(c), it is clear that below 47 sub- samples, the recall value is poor. But even with 3 * 141 subsamples, the recall value becomes rather acceptable. This is the strength of the Subsampling algorithm, which 1 Actually, we did not even hear of Jan Caloun before our experiment. During 1995-96, Caloun played a total of 11 games, and scored 8 goals|almost a goal per game, which is a rarity in the NHL. A search of the World Wide Web reveals that Caloun played a grand total of 13 games in the NHL|11 games in 1995-96, and 2 games in 1996- 97|before disappearing from the NHL scene. We also learned that he scored on his rst four NHL shots to tie an NHL record. can give acceptable results in a short time. But, the recall curve has a diminishing rate of return, and it may take a very long time for Subsampling to reach a high level of recall, as conrmed in Fig. 10. Since the Hybrid-random algorithm uses the Subsampling algorithm in its rst phase (with d 47 expected that the Hybrid-random algorithm behaves about as well as the Subsampling algorithm, at the beginning, for mediocre levels of recall, such as 70-75% (cf: Fig. 10). But, as shown in Fig. 9(d), if the Hybrid-random algorithm is allowed to execute longer, it steadily and quickly improves the quality of its computation. Thus, in terms of CPU time, we start with the Subsampling curve, but quickly switch to the Pure-random curve to reap the benets of a fast algorithm and pruned randomization. Achieving a Given Rate of Recall The above experiment shows how each algorithm trades o e-ciency with quality. Having picked a reasonable set of parameter values for each algorithm, let us now compare the algorithms head- to-head. Specically, for xed recall rates, we compare the time taken for each algorithm to deliver that recall rate. Because the run time of the Fixed-angle algorithm is typically several orders of magnitude above the others (for comparable quality), we omit the Fixed-angle algorithm results from now on. Fig. 10 compares the Hybrid-random algorithm with both the Pure-random and Subsampling algorithms, for higher rates of recall. In general, the Subsampling algorithm is very eective for quick, consistent results. However, to improve further on the quality, it can take a very long time. In contrast, when the Hybrid-random algorithm is allowed to run just a bit longer, it can deliver steady improvement on quality. As a case in point, to achieve about 90% recall in the current example, it takes the Subsampling algorithm almost 14 hours to achieve the same level of recall produced by the Hybrid-random algorithm in about two minutes. Never- theless, we must give the Subsampling algorithm credit for giving the Hybrid-random algorithm an excellent base from which to start its computation. In Fig. 10, the Pure-random algorithm signicantly out-performs the Subsampling algorithm, but this is not always the case. We expect the recall rate for Pure-random to be volatile, and there are cases where the Pure-random algorithm returns substantially dierent outliers for large numbers of iterations. The Hybrid-random algorithm tends to be more focused and consistent. Scalability in Dimensionality and Dataset Size Fig. 11(a) shows scalability of dimensionality for the Subsampling and Hybrid-random algorithms. We used moderate levels of recall (e.g., 75%) and 60,000 tuples for this anal- ysis. High levels of recall would favor the Hybrid-random algorithm. The results shown here are for 282 iterations for the Subsampling algorithm, and 90 Patches for the Hybrid- random algorithm. Our experience has shown that these numbers of iterations and patches are satisfactory, assuming we are satised with conservative levels of recall. Fig. 11(a) shows that both algorithms scale well, and this conrms our complexity analysis of Section 4. Fig. 11(b) shows how the Subsampling and Hybrid-random algorithms scale with dataset size, in 5-D, for conservative levels of recall. Again, both algorithms seem to scale well, and again the Hybrid-random algorithm outperforms the Run Time and Recall for the Fixed-angle Algorithm: 5-D, 100,000 Tuples Number of Angles Tested (Iterations) CPU Time in Minutes CPU Time Recall Recall Run Time and Recall for the Pure-random Algorithm: 5-D, 100,000 Tuples Number of Random Angles (Iterations) CPU Time in Minutes CPU Time Recall Recall 5002060100Run Time and Recall for the Subsampling Algorithm: 5-D, 100,000 Tuples Number of Subsamples (Iterations) CPU Time in Seconds CPU Time Recall Recall 700100Run Time and Recall for the Hybrid-random Algorithm: 5-D, 100,000 Tuples; delta=0.0800 Number of Patches CPU Time in Seconds CPU Time Recall Recall Figure 9: Plots of Run Time and Recall: (a) Top left: Fixed-angle. (b) Top right: Pure-random. (c) Bottom left: Subsampling. (d) Bottom right: Hybrid-random. 9050150250350Run Times to Achieve a Given Level of Recall, for 3 Algorithms, 5-D, & 100,000 Tuples Percent Recall CPU Time in Seconds Pure-random Subsampling Hybrid-random Figure 10: Run Time vs. Recall for Subsampling, Pure-random, and Hybrid-random Algorithms Subsampling algorithm. High levels of recall would favor the Hybrid-random algorithm, even more so than shown. 7. AND CONCLUSION The results returned by many types of distance-based KDD operations/queries tend to be less meaningful when when no attention is paid to scale, variability, correlation, and outliers in the underlying data. In this paper, we presented the case for robust space transformations to support operations such as nearest neighbor search, distance-based clustering, and outlier detection. An appropriate space is one that: (a) preserves the Euclidean property, so that ecient Euclidean distance operations can be performed without sacricing quality and meaningfulness of results, and (b) is stable in the presence of a non-trivial number of updates. We saw that distance operations which ordinarily would be inappropriate when operating on the raw data (and even on normalized or standardized data), are actually appropriate in the transformed space. Thus, the end user sees results which tend to be more intuitive or meaningful for a given application. We presented a data mining case study on the detection of outliers to support these claims. After considering issues such as eectiveness (as measured by precision and recall, especially the latter) and efciency (as measured by scalability both in dimensionality and dataset size), we believe that the Hybrid-random algorithm that we have developed in this paper is an excellent choice among the Donoho-Stahel algorithms. In tens of seconds of CPU time, a robust estimator can be computed which not only accounts for scale, variability, correlation, and outliers, but is also able to withstand a signicant number of database updates (e.g., 50% of the tuples) without losing eectiveness or requiring re-computation. For many cases involving high levels of recall, the randomized algo- rithms, and in particular, the Hybrid-random algorithm can be at least an order of magnitude faster (and sometimes several orders of magnitude faster) than the alternatives. In conclusion, we believe that our results have shown that robust estimation has a place in the KDD community, and can nd value in many KDD applications. 8. --R Automatic Subspace Clustering of High Dimensional Data for Data Mining Applications. Elementary Linear Algebra: Applications Version. Outliers in Statistical Data. LOF: Identifying Density-Based Local Outliers Box and D. Numerical Analysis. A fast algorithm for robust principal components based on projection pursuit. A Density-based Algorithm for Discovering Clusters in Large Spatial Databases with Noise R-trees: a dynamic index structure for spatial searching. Algorithms for Mining Distance-Based Outliers in Large Datasets Bias robust estimation of scale. The behaviour of the Stahel-Donoho robust multivariate estimator Robust Regression and Outlier Detection. Introduction to Modern Information Retrieval. Breakdown of Covariance Estimators. STING: A statistical information grid approach to spatial data mining. High breakdown point estimates of regression by means of the minimization of an e-cient scale --TR Robust regression and outlier detection The R*-tree: an efficient and robust access method for points and rectangles Efficient and effective querying by image content Distance-based indexing for high-dimensional metric spaces Automatic subspace clustering of high dimensional data for data mining applications Density-based indexing for approximate nearest-neighbor queries Efficient algorithms for mining outliers from large data sets Introduction to Modern Information Retrieval R-trees Algorithms for Mining Distance-Based Outliers in Large Datasets Efficient and Effective Clustering Methods for Spatial Data Mining Efficient User-Adaptable Similarity Search in Large Multimedia Databases --CTR Peng Sun , Robert M. Freund, Computation of Minimum-Volume Covering Ellipsoids, Operations Research, v.52 n.5, p.690-706, Sep. - Oct. 2004 S. Cateni , V. Colla , M. Vannucci, A fuzzy logic-based method for outliers detection, Proceedings of the 25th conference on Proceedings of the 25th IASTED International Multi-Conference: artificial intelligence and applications, p.561-566, February 12-14, 2007, Innsbruck, Austria Leejay Wu , Christos Faloutsos, Making every bit count: fast nonlinear axis scaling, Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining, July 23-26, 2002, Edmonton, Alberta, Canada Jaideep Vaidya , Chris Clifton, Privacy-preserving k-means clustering over vertically partitioned data, Proceedings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining, August 24-27, 2003, Washington, D.C.

data mining;robust statistics;space transformations;outliers;robust estimators;distance-based operations

502544

Evaluating the novelty of text-mined rules using lexical knowledge.

In this paper, we present a new method of estimating the novelty of rules discovered by data-mining methods using WordNet, a lexical knowledge-base of English words. We assess the novelty of a rule by the average semantic distance in a knowledge hierarchy between the words in the antecedent and the consequent of the rule - the more the average distance, more is the novelty of the rule. The novelty of rules extracted by the DiscoTEX text-mining system on Amazon.com book descriptions were evaluated by both human subjects and by our algorithm. By computing correlation coefficients between pairs of human ratings and between human and automatic ratings, we found that the automatic scoring of rules based on our novelty measure correlates with human judgments about as well as human judgments correlate with one another. @Text mining

Introduction A data-mining system may discover a large body of rules; however, relatively few of these may convey useful new knowledge to the user. Several metrics for evaluating the \interestingness" of mined rules have been proposed [BA99, HK01]. These metrics can be used to lter out a large percentage of the less interesting rules, thus yielding a more manageable number of higher quality rules to be presented to the user. However, most of these measure simplicity (e.g. rule size), certainty (e.g. condence), or utility (e.g. support). Another important aspect of interestingness is novelty: does the rule represent an association that is currently unknown. For example, a text-mining system we developed that discovers rules from computer-science job announcements posted to a local newsgroup [NM00] induced the rule: \SQL ! database". A knowledgeable computer scientist may nd this rule uninteresting because it conveys a known association. Evaluating the novelty of a rule requires comparing it to an existing body of knowledge the user is assumed to already possess. For text mining [Hea99, Fel99, Mla00], in which rules consist of words in natural language, a relevant body of common knowledge is basic lexical semantics, i.e. the meanings of words and the semantic relationships between them. A number of lexical knowledge bases are now available. WordNet [Fel98] is a semantic network of about 130,000 English words linked to about 100,000 lexical senses (synsets) that are interconnected by relations such as antonym, generalization (hypernym), and part-of (holonym). We present and evaluate a method for measuring the novelty of text-mined rules using such lexical knowledge. We dene a measure of the semantic distance, d(w words based on the length of the shortest path connecting w i and w j in WordNet. The novelty of a rule is then dened as the average value of d(w all pairs of words (w is in the antecedent and w j is in the consequent of the rule. Intuitively, the semantic dissimilarity of the terms in a rule's antecedent and in its consequent is an indication of the rule's novelty. For example, \beer ! diapers" would be considered more novel than \beer ! pretzels" since beer and pretzels are both food products and therefore closer in WordNet. We present an experimental evaluation of this novelty metric by applying it to rules mined from book descriptions extracted from Amazon.com. Since novelty is fundamentally subjective, we compared the metric to human judgments. We have developed a web-based tool that allows human subjects to enter estimates of the novelty of rules. We asked multiple human subjects to score random selections of mined rules and compared the results to those obtained by applying our metric to the same rules. We found that the average correlation between the scoring of our algorithm and that of the human users, using both raw score correlation (Pearson's metric) and rank correlation (Spearman's metric), was comparable to the average score correlation between the human users. This suggests that the algorithm has a rule scoring judgment similar to that of human users. Background 2.1 Text Mining Traditional data mining algorithms are generally applied to structured databases, but text mining algorithms try to discover knowledge from unstructured or semi-structured textual data, e.g. web-pages. Text mining is a relatively new research area at the intersection of natural language processing, machine learning and information retrieval. Various new useful techniques are being developed by researchers for discovering knowledge from large text corpora, by appropriately integrating methods from these dierent disciplines. DiscoTEX [NM00] is one such system, that discovers prediction rules from natural language corpora using a combination of information extraction and data mining. It learns an information extraction system to transform text into more structured data, and this structured data is then mined for interesting relationships. For our experiments, we have used rules mined by DiscoTEX from book descriptions extracted from Amazon.com, in the \science", \romance" and \literature" categories. DiscoTEX rst extracts a structured template from the Amazon.com book description web-pages. It constructs a template for each book description, with pre-dened slots (e.g. title, author, subject, etc.) that are lled with words extracted from the text. DiscoTEX then uses a rule mining technique to extract prediction rules from this template database. An example extracted rule is shown in Figure 1, where the slot is predicted from the other slots. For our purpose, we only use the ller words in the slot, ignoring the slotnames | in our algorithm, the rule in Figure 1 would be used in the form \daring love woman romance historical ction story read wonderful". daring, love woman romance, historical, fiction -> story, read, wonderful Figure 1: DiscoTEX rule mined from Amazon.com \romance" book descriptions 2.2 WordNet WordNet [Fel98] is an online lexical knowledge-base of 130,000 English words, developed at Princeton University. In WordNet, English nouns, adjectives, verbs and adverbs are organized into synonym sets or synsets, each representing an underlying lexical concept. A synset contains words of similar meaning pertaining to a common semantic concept. But since a word can have dierent meanings in dierent contexts, a word can be present in multiple synsets. A synset contains associated pointers representing its relation to other synsets. WordNet supports many pointer types e.g. antonyms, synonyms, etc. The pointer types we used in our algorithm are explained below: 1. Synonym: This pointer is implicit. Since words in the same synset are synonymous, e.g. life and existence, the synonym of a synset is itself. 2. Antonym: This pointer type refers to another synset that is quite opposite in meaning to the given synset, e.g. front is the antonym of back. 3. Attribute: This pointer type refers to another synset that is implicated by this synset, e.g. benevolence is an attribute of good. 4. Pertainym: This pointer refers to a relation from a noun to an adjective, an adjective to a noun, or an adverb to an adjective, indicating morphological relation, e.g. alphabetical is a pertainym of alphabet. 5. Similar: This pointers refers to another adjective that is very close in terms of meaning to the current adjective, although not enough to be part of the same synset, e.g. unquestioning is similar to absolute. 6. Cause: This pointer type refers to a cause and eect relation, e.g. kill is cause to die. 7. Entailment: This pointer refers to the implication of another action e.g. breathe is an entailment of inhale. 8. Holonym: This pointer refers to a part in a part-whole relation, e.g. chapter is a holonym of text. There are three kinds of holonyms | by member, by substance and by part. 9. Meronym: This pointer refers to a whole in a part-whole relation, e.g. computer is a meronym of cpu. There are three kinds of meronyms | by member, by substance and by part. 10. Hyponym: This pointer refers to a specication of the concept, e.g. fungus is a hyponym of plant. 11. Hypernym: This pointer refers to a generalization of the concept, e.g. fruit is a hypernym of apple. 2.3 Semantic Similarity of Words Several measures of semantic similarity based on distance between words in WordNet have been used by dierent researchers. Leacock and Chodorow [LC98] have used the negative logarithm of the normalized shortest path length as a measure of similarity between two words, where the path length is measured as the number of nodes in the path between the two words and the normalizing factor is the maximum depth in the taxonomy. In this metric, the greater the semantic distance between two words in the WordNet hierarchy, the less is their semantic similarity. Lee et al. [LKY93] and Rada et al. [RMBB89] have used conceptual distance, based on an edge counting metric, to measure similarity of a query to documents. Resnick [Res92] observed that two words deep in the WordNet are more closely related than two words higher up in the tree, both pairs having the same path length (number of nodes) between them. Sussna [Sus93] took this into account in his semantic distance measure that uses depth-relative scaling. Hirst et al. [HSO98] classied the relations of WordNet into the three broad directional categories and used a distance measure where they took into account not only the path length but also the number of direction changes in the semantic relations along the path. Resnick [Res95] has used an information-based measure instead of path length to measure the similarity, where the similarity of two words is estimated from the information content of the least probable class to which both words belong. 3 Scoring the Novelty of Rules 3.1 Semantic Distance Measure We have dened the semantic distance between two words w i and w j as: where is the distance along path p according to our weighting scheme, Dir(p) is the number of direction changes of relations along path p, and K is a suitably chosen constant. The second component of the formula is derived from the denition of Hirst et al. [HSO98], where the relations of WordNet are divided into three direction classes | \up", \down" and \horizontal", depending on how the two words in the relation are lexically related. Table 1 summarizes the direction information for the relation types we use. The more direction changes in the path from one word to another, the greater the semantic distance between the words, since changes of direction along the path re ect large changes in semantic context. The path distance component of the above formula is based on the semantic distance de- nition of Sussna [Sus93]. It is dened as the shortest weighted path between w i and w j , where every edge in the path is weighted according to the WordNet relation corresponding to that edge, and is normalized by the depth in the WordNet tree where the edge occurs. We have used 15 dierent WordNet relations in our framework, and we have assigned dierent weights to dierent link types, e.g. hypernym represents a larger semantic change than synonym, so hypernym has a higher weight than synonym. The weight chosen for the dierent relations are given in Table 1. One point to note here is that Sussna's denition of semantic distance calculated the weight of an edge between two nouns w i and w j as the average of the two relations w corresponding to the edge, relation r 0 being the inverse of relation r. This made the semantic distance between two words a symmetric measure. He had considered the noun hierarchy, where every relation between nouns has an inverse relation. But in our framework, where we have considered all the four types of words in WordNet (nouns, adverbs, adjectives and verbs) and 15 dierent relation types between these words, all of these relations do not have inverses, e.g. the entailment relation has no direct inverse. So, we have used only the weight of the relation w as a measure of the weight of the edge between w i and w j . This gives a directionality to our semantic measure, which is also conceptually compatible with the fact that w i is a word in the antecedent of the rule and w j is a word in the consequent of the rule. 3.2 Rule Scoring Algorithm The scoring algorithm of rules according to novelty is outlined in Figure 2. The algorithm calculates the semantic distance d(w is in the antecedent and w j is in the consequent of the rule, based on the length of the shortest path Relation Direction Weight Synonym, Attribute, Pertainym, Similar Horizontal 0.5 Antonym Horizontal 2.5 Hypernym, (MemberjPartjSubstance) Meronym Up 1.5 Hyponym, (MemberjPartjSubstance) Holonym, Down 1.5 Cause, Entailment Table 1: Direction and weight information for the 15 WordNet relations used connecting w i and w j in WordNet. The novelty of a rule is then calculated as the average value of all pairs of words (w The noun hierarchy of the WordNet is disconnected | there are 11 trees with distinct root nodes. The verb hierarchy is also disconnected, with 15 distinct root nodes. For our purpose, following the method of Leacock and Chodorow [LC98], we have connected the 11 root nodes of the noun hierarchy to a single root node R noun so that a path can always be found between two nouns. Similarly, we have connected the verb root nodes by a single root node R verb . R noun and R verb are further connected to a top-level root node, R top . This connects all the verbs and nouns in the WordNet database. Adjectives and adverbs are not hierarchically arranged in WordNet, but they are related to their corresponding nouns. In this composite connected hierarchy derived from the WordNet hierarchy, we nd the shortest weighted path between two words by performing a branch and bound search. In this composite word hierarchy, any two words are connected by a path. However, we have used 15 dierent WordNet relations while searching for the path between two words | this creates a combinatorial explosion while performing the branch and bound search on the composite hierarchy. So, for e-cient implementation, we have a user-specied time-limit (set to 3 seconds in our experiments) within which we try to nd the shortest path between the words w i and w j . If the shortest path cannot be found within the time-limit, the algorithm nds a default path between w i and w j by going up the hierarchy from both w i and w j , using hypernym links, till a common root node is reached. The function PathViaRoot in Figure 2 computes the distance of the default path. For nouns and verbs, the PathViaRoot function calculates the distance of the path between the two words as the sum of the path distances of each word to its root. If the R noun or the R verb node are For each rule in a rule le set of antecedent words, set of consequent words For each word w and w j are not a valid words in WordNet Score (w Elseif w j is not a valid word in WordNet Score (w is not a valid word in WordNet Score (w Elseif path not found between w i and w j (in user-specied time-limit) Score (w Else Score (w Score of rule = Average of all (w Sort scored rules in descending order Figure 2: Rule Scoring Algorithm a part of this path, it adds a penalty term POSRootPenalty = 3.0 to the path distance. If the R top node is a part of this path, it adds a larger penalty TopRootPenalty = 4.0 to the path distance. These penalty terms re ect the large semantic jumps in paths which go through the root nodes R noun , R verb and R top . If one of the words is an adjective or an adverb, and the shortest path method does not terminate within the specied time-limit, then the algorithm nds the path from the adjective or adverb to the nearest noun, through relations like \pertainym", \attribute", etc. It then nds the default path up the noun hierarchy, and the PathViaRoot function incorporates the distance of the path from the adjective or adverb to the noun form into the path distance measurement. Some of the words extracted from the rules are not valid words in WordNet e.g. abbrevi- ations, names like Philip, domain specic terms like booknews, etc. We assigned such words the average depth of a word (d avg in Figure 2) in the WordNet hierarchy, which was estimated by sampling techniques to be about 6, and then estimated its path distance to the root of the combined hierarchy by using the PathViaRoot function. 4 Experimental Results We performed experiments to compare the novelty judgment of human users to the automatic ratings of our algorithm. The objective here is that if the automatic ratings correlate with human High score (9.5): romance love heart -> midnight Medium score (5.8): author romance -> characters love Low astronomy science -> space Figure 3: Examples of rules scored by our novelty measure judgments about as well as human judgments correlate with each other, then the novelty metric can be considered successful. 4.1 Methodology For the purpose of our experiments, we took rules generated by DiscoTEX from 9000 Ama- zon.com book descriptions: 2000 in the \literature" category, 3000 in the \science" category and 4000 in the \romance" category. From the total set of rules, we selected a subset of rules that had less than a total of 10 words in the antecedent and consequent of the rule | this was done so that the rules were not too large for human users to rank. Further pruning was performed to remove duplicate words from the rules. For the Amazon.com book description do- main, we also created a stoplist of commonly occurring words, e.g. book, table, index, content, etc., and removed them from the rules. There were 1258 rules in the nal pruned rule-set. We sampled this pruned rule-set to create 4 sets of random rules, each containing 25 rules. We created a web-interface, which the subjects used to rank these rules with scores in the range from (least interesting) to 10.0 (most interesting), according to their judgment. The 48 subjects were randomly divided into 4 groups and each group scored one of the rule-sets. For each of the rule-sets, two types of average correlation were calculated. The rst average correlation was measured between the human subjects, to nd the correlation in the judgment of novelty between human users. The second average correlation measure was measured between the algorithm and the users in each group, to nd the correlation between the novelty scoring of the algorithm and that of the human subjects. We used both Pearson's raw score correlation metric and Spearman's rank correlation metric to compute the correlation measures. One of the rule-sets was used as a training set, to tune the parameters of the algorithm. The results on the 3 other rule-sets, used as test sets for our experiment, are summarized in Table 2. Human - Human Algorithm - Human Correlation Correlation Raw Rank Raw Rank Group2 Average Table 2: Summary of experimental results 4.2 Results and Discussion Some of the rules scorings generated by our algorithm are shown in Figure 3. The high-scoring rule and the low-scoring rule were rated by the human subjects, on the average, as high scoring and low-scoring too. From the results, considering both the raw and the rank correlation measures, we see that the correlation between the human subjects and the algorithm is comparable to that between the human subjects, averaging over the three random rule-sets considered. The average raw correlation values among the human subjects and between the human subjects and the algorithm are both not very high. This is because for some rules, the human subjects diered a lot in their novelty assessment. This is also due to the fact that these are initial experiments, and we are working on improving the methodology. In later experiments, we intend to apply our method to domains where we can expect human users to agree more in their novelty judgment of rules. However, it is important to note that it is very unlikely that these correlations are due to random chance, since both the average raw correlation values are above the minimum signicant r at the p < 0:1 level of signicance determined by a t-test. The correlation between the human subjects and the algorithm was low for the rst rule- set. For the second and the third rule-sets, the algorithm-human correlation is better than the human-human correlation. On closer analysis of the results of Group1, we noticed that this rule-set contained many rules involving proper names. Our algorithm currently uses only semantic information from WordNet, so it's scoring on these rules diered from that of human subjects. For example, one rule many users scored as uninteresting was \ieee society ! science mathematics", but since WordNet does not have an entry for \ieee", our algorithm gave the overall rule a high score. Another rule to which some users gave a low score was \physics science nature ! john wiley publisher sons", presumably based on their background knowledge about publishing houses. In this case, our algorithm found the name John in the WordNet hierarchy (synset lemma: disciple of Jesus), but there was no short path between John and the words in the antecedent of the rule. As a result, the algorithm gave this rule a high score. A point to note here is that some names like Jesus, John, James, etc. have entries in WordNet, but others like Sandra, Robert, etc. do not | this makes it di-cult to use any kind of consistent handling of names using lters like name lists. In the training rule-set, we had also noticed that the rule \sea ! oceanography" had been given a large score by our algorithm, while most subjects in that group had rated that rule as uninteresting. This happened because there is no short path between sea and oceanography in WordNet | these two words are related thematically, and WordNet does not have thematic connections, an issue which is discussed in detail in Section 6. 5 Related Work Soon after the Apriori algorithm for extracting association rules was proposed, researchers in the data mining area realized that even modest settings for support and condence typically resulted in a large number of rules. So much eort has gone into reducing such rule-sets by applying both objective and subjective criteria. Klemettinen et al. [KMR proposed the use of rule templates to describe the structure of relevant rules and constrain the search space. Another notable attempt in using objective measures was by Bayardo and Agrawal [BA99], who dened a partial order, in terms of both support and condence, to identify a smaller set of rules that were more interesting than the rest. Sahar [Sah99] proposed an iterative elimination of uninteresting rules, limiting user interaction to a few simple classication questions. Hussain et al. [HLSL00] developed a method for identifying exception rules, with the interestingness of a rule being estimated relative to common sense rules and reference rules. In a series of papers, Tuzhilin and his co-researchers [ST96, PT98, AT99] argued the need for subjective measures for the interestingness of rules. Rules that were not only actionable but also unexpected in that they con icted with the existing system of beliefs of the user, were preferred. Liu et al. [LHMH99] have further built on this theme, implementing it as an interactive, post-processing routine. They have also analyzed classication rules, such as those extracted from C4.5, dening a measure of rule interestingness in terms of the syntactic distance between a rule and a belief. A rule and a belief are \dierent" if either the consequents of the rule and the belief are \similar" but the antecedents are far apart, or vice versa. In contrast, in this paper we have analyzed information extracted from unstructured or semi-structured data such as web-pages, and extracted rules depicting important relations and regularities in such data. The nature of rules as well as of prior domain knowledge is quite dierent from those extracted, say, from market baskets. We have proposed an innovative use of WordNet to estimate the semantic distance between the antecedents and consequents of a rule, which is used as an indication of the novelty of the rule. Domain-specic concept hierarchies have previously been used to lter redundant mined rules [HF95, FD95]; however, to our knowledge they have not been used to evaluate novelty quantitatively, or applied to rules extracted from text data. 6 Future Work An important issue that we want to address in future is the selection of the parameters of the algorithm, e.g. the weights of the relations, and values of K, POSRootPenalty and Top- RootPenalty. These constants are now chosen experimentally. We would like to learn these parameters automatically from training data, using a machine learning technique. The novelty score could then be adaptively learnt for a particular user and tailored to suit the user's expectation. We are using the average of the pairwise word similarity measures as the novelty score of a rule. The average measure smoothes out the skewing eect due to large distances between any two pairs of word in a rule. This is ne for most rules, except for some special cases, e.g. if we have a rule \science ! scientic home", then the distance between \science" and \scientic" is small, but that between \science" and \home" is large. Using average here gives the whole rule a medium novelty score, which does not re ect the fact that a part of the rule involving the words \science" and \home" is highly interesting, while the other part involving the words \science" and \scientic" is uninteresting. In this case, a combination method like maximum might be more useful. A suitable combination of the average and the maximum metrics would hopefully give a better novelty scoring. Unfortunately, WordNet fails to capture all semantic relationships between words, such as general thematic connections like that between \pencil" and \paper". However, other approaches to lexical semantic similarity, such as statistical methods based on word co-occurrence [MS99], can capture such relationships. In these methods, a word is typically represented by a vector in which each component is the number of times the word co-occurs with another specied word within a particular corpus. Co-occurrence can be based on appearing within a xed-size window of words, or in the same sentence, paragraph, or document. The similarity of two words is then determined by a vector-space metric such as the cosine of the angle between their corresponding vectors [MS99]. In techniques such as Latent Semantic Analysis the dimensionality of word vectors is rst reduced using singular value decomposition (SVD) in order to produce lexical representations with a small number of highly-relevant dimensions. Such methods have been shown to accurately model human lexical-similarity judgments [LD97]. By utilizing a co-occurrence-based metric for d(w rules could be ranked by novelty using statistical lexical knowledge. In the end, some mathematical combination of WordNet and co-occurrence based metrics may be the best approach to measuring lexical semantic distance. To the extent that the names of relations, attributes, and values in a traditional database are natural-language words (or can be segmented into words), our approach could also be applied to traditional data mining as well as text mining. The algorithm can be easily generalized for scoring the novelty of other types of rules, e.g. association rules derived from market-basket data. In that case, we would require a knowledge-base for the corresponding domain, e.g. a concept hierarchy of the company products. The domain-specic concept hierarchies and knowledge-bases could be used to nd semantic connections between rule antecedents and consequents and thereby contribute to evaluating novelty. Finally, the overall interestingness of a rule might be best computed as a suitable mathematical combination of novelty and more traditional metrics such as condence and support. 7 Conclusion This paper proposes a methodology for extracting, analyzing and ltering rules extracted from unstructured or semi-structured data such as web pages. These rules can underscore novel and useful relations and regularities in textual sources of information such as web pages, email and usenet postings. Note that the nature of rules as well as of prior domain knowledge is quite dierent from those extracted, say, from market baskets. A salient contribution of this paper is a new approach for measuring the novelty of rules mined from text data, based on the lexical knowledge in WordNet. This algorithm can also be extended to rules in other domains, where a domain-specic knowledge hierarchy is available. We have also introduced a systematic method of empirically evaluating interestingness measures for rules, based on average correlation statistics, and have successfully shown that the automatic scoring of rules based on our novelty measure correlates with human judgments about as well as human judgments correlate with each other. Acknowledgments We would like to thank Un Yong Nahm for giving us the DiscoTEX rules sets on which we ran our experiments. We are grateful to John Didion for providing the JWNL Java interface to WordNet, which we used to develop the software, and for giving us useful feedback about the package. We are also grateful to all the people who volunteered to take part in our experiments. The rst author was supported by the Microelectronics and Computer Development (MCD) Fellowship, awarded by the University of Texas at Austin, while doing this research. --R User pro Bayardo Jr. Indexing by latent semantic analysis. Knowledge discovery in textual databases (KDT). An Electronic Lexical Database. Untangling text data mining. Discovery of multiple-level association rules from large databases Data Mining: Concepts and Techniques. Exception rule mining with a relative interestingness measure. Lexical chains as representations of context for the detection and correction of malapropims. Finding interesting rules from large sets of discovered association rules. Combining local context and WordNet similarity for word sense identi Finding interesting patterns using user expectations. Information retrieval based on a conceptual distance in IS-A heirarchy Dunja Mladeni A mutually bene A belief-driven method for discovering unexpected patterns WordNet and distribution analysis: A class-based approach to lexical discovery Using information content to evaluate semantic similarity in a taxon- omy Development and application of a metric on semantic nets. Interestingness via what is not interesting. What makes patterns interesting in knowledge discovery systems. Word sense disambiguation for free-text indexing using a massive semantic network --TR Word sense disambiguation for free-text indexing using a massive semantic network Finding interesting rules from large sets of discovered association rules Foundations of statistical natural language processing Mining the most interesting rules Interestingness via what is not interesting Data mining What Makes Patterns Interesting in Knowledge Discovery Systems Finding Interesting Patterns Using User Expectations Discovery of Multiple-Level Association Rules from Large Databases Exception Rule Mining with a Relative Interestingness Measure A Mutually Beneficial Integration of Data Mining and Information Extraction --CTR Xin Chen , Yi-fang Brook Wu, Web mining from competitors' websites, Proceeding of the eleventh ACM SIGKDD international conference on Knowledge discovery in data mining, August 21-24, 2005, Chicago, Illinois, USA Raz Tamir , Yehuda Singer, On a confidence gain measure for association rule discovery and scoring, The VLDB Journal The International Journal on Very Large Data Bases, v.15 n.1, p.40-52, January 2006 B. Shekar , Rajesh Natarajan, A Framework for Evaluating Knowledge-Based Interestingness of Association Rules, Fuzzy Optimization and Decision Making, v.3 n.2, p.157-185, June 2004 Combining Information Extraction with Genetic Algorithms for Text Mining, IEEE Intelligent Systems, v.19 n.3, p.22-30, May 2004

wordnet;interesting rules;knowledge hierarchy;novelty;semantic distance

502555

Generalized clustering, supervised learning, and data assignment.

Clustering algorithms have become increasingly important in handling and analyzing data. Considerable work has been done in devising effective but increasingly specific clustering algorithms. In contrast, we have developed a generalized framework that accommodates diverse clustering algorithms in a systematic way. This framework views clustering as a general process of iterative optimization that includes modules for supervised learning and instance assignment. The framework has also suggested several novel clustering methods. In this paper, we investigate experimentally the efficacy of these algorithms and test some hypotheses about the relation between such unsupervised techniques and the supervised methods embedded in them.

INTRODUCTION AND MOTIVATION Although most research on machine learning focuses on induction from supervised training data, there are many situations in which class labels are not available and which thus require unsupervised methods. One widespread approach to unsupervised induction involves clustering the training cases into groups that reflect distinct regions of the decision space. There exists a large literature on clustering methods (e.g., Everitt [3]), a long history of their development, and increasing interest in their application, yet there is still Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. KDD-2001 San Francisco, CA USA Copyright 2001 ACM . $5.00. little understanding of the relation between supervised and unsupervised approaches to induction. In this paper, we begin to remedy that oversight by examining situations in which a supervised induction method occurs as a subroutine in a clustering algorithm. This suggests two important ideas. First, one should be able to generate new clustering methods from existing techniques by replacing the initial supervised technique with a different supervised technique. Second, one would expect the resulting clustering methods to behave well (e.g., form desirable clusters) in the same domains for which their supervised components behave well, provided the latter have labeled training data available. In the pages that follow, we explore both ideas in the context of iterative optimization, a common scheme for clustering that includes K-means and expectation maximization as special cases. After reviewing this framework in Section 2, we describe an approach to embedding any supervised algorithm and its learned classifier in an iterative optimizer, and in Section 3 we examine four supervised methods for which we have taken this step. In Section 4, we report on experimental studies designed to test our hypotheses about the relations between behavior of the resulting clustering methods and that of their supervised components. In closing, we review related work on generative frameworks for machine learning and consider some directions for future research. Update class models Assign instances to classes Clustered data Data instances labeled data model parameters Figure 1: The iterative optimization procedure. 2. GENERALIZED CLUSTERING Many clustering systems rely on the notion of iterative optimization. As Figure 1 depicts, such a system iterates between two steps - class model creation and data reassignment - until reaching a predetermined iteration limit or until no further changes occur in reassignments. There are many variations within this general framework, but the basic idea is best illustrated with some well-known example methods. 2.1 K-means and EM as Iterative Optimizers Two clustering algorithms that are popular for their simplicity and flexibility are K-means [2] and expectation maximization (EM) [1]. Both methods have been studied experimentally on many problems and have been used widely in applied settings. Here we review the algorithms briefly, note their key similarities, and show how their differences suggest a more general clustering framework. The K-means algorithm represents each class by a cen- troid, which it computes by taking the mean for each attribute over all the instances belonging to that class. In geometric terms, this corresponds to finding the center of mass for the cases associated with that class. Data reassignment involves assigning each instance to the class of the closest centroid. In contrast, EM models each class by a probability distribution that it extracts from the training data in the class model creation step. If the data are continuous, each class is generally modeled by an n-dimensional Gaussian distribution that consists of a mean and variance for each attribute. In the discrete case, P (a extracted for each possible combination of class ck , attribute a j , and attribute value v jl . In both cases, when finding these parameters, the contribution of each instance x i is weighted by P (ck jx i ). Data reassignment is done by recalculating P (ck jx i ) for each instance x i and class ck using the new class models. 2.2 A General Framework Although both of the above clustering algorithms incorporate iterative optimization, they employ different methods for developing class models. Thus, we can view them as invoking a different supervised learning technique to distinguish among the classes. The two algorithms also differ in how they assign instances to classes: K-means assigns each instance to a single class, whereas EM uses partial as- signment, in that each instance is distributed among the classes. We will refer to the absolute method as the "strict" paradigm and to the partial method as "weighted". These observations lead to a general framework for clustering that involves selecting a supervised learning algorithm and selecting one of these assignment paradigms. In the context of K-means and EM, this framework immediately suggests some variants. By using the weighted paradigm with the K-means classifier, we obtain a weighted K-means algorithm. Similarly, combining EM's probabilistic classifier with the strict paradigm produces a variant in which each instance is assigned entirely to its most probable class. This variant has been explored under the name of "strict- assignment EM", although the partial assignment method is more commonly used. Although the classifiers utilized in K-means and EM can be easily modified to operate with either assignment method, other supervised algorithms can require more sophisticated adaptations, as we will see shortly. 3. SUPERVISED LEARNING METHODS As we have argued, it should be possible to embed any supervised learning method within our generalized clustering framework. However, our evaluation has focused on four simple induction algorithms that have limited representational power [6], because the clustering process itself aims to generate the disjoint decision regions that more powerful supervised methods are designed to produce. Below we describe these algorithms in some detail, including the adaptations we made for use in the weighted paradigm. These adaptations involve altering model production to take into account the weights of instances and revising instance reassignment to generate class weights for every instance, which are then used to produce the next generation of class models. 3.1 Prototype Modeler Our first supervised algorithm, which plays a role in K- means, creates a prototype [13] or centroid for each class by extracting the mean of each attribute from training cases for that class. Such a prototype modeler classifies an instance by selecting the class with the centroid closest to it in n-dimensional space. Because the distance metric is sensitive to variations in scale, our version normalizes all data to values between zero and one before creating the prototypes. In the weighted paradigm, the mean for each attribute becomes a weighted average of the training cases. The relative proximity of each instance to a given centroid determines the associated weight for that centroid's class. Formally, we can express this by where jCj is the number of classes. The new centroid is then composed of the weighted mean for each attribute, with the mean of attribute a j for cluster ck being calculated by is the value of the jth attribute of instance x i and where jXj is the total number of instances. 3.2 Naive Bayesian Modeler We selected naive Bayes [2] as our second induction algo- rithm. As described in the context of EM, this technique models each class as a probability distribution described by for each class ck , attribute a j , and attribute value v jl . For nominal attributes, naive Bayes represents as a discrete conditional probability distribution, which it estimates from counts in the training data, and it estimates the class probability P (ck ) in a similar manner. For continuous attributes, it typically uses a conditional Gaussian distribution that it estimates by computing the mean and variance for each attribute from training data for each class. To calculate the relative probability that a new instance belongs to a given class ck , naive Bayes employs the expression Y which assumes that the distribution of values for each attribute are independent given the class. When operating normally as a strict classifier, naive Bayes returns the class with the highest probability for each in- stance. In the weighted case, the conditional distributions are calculated using a weighted sum rather than a strict sum, while the expression determines the weight used in the data reassignment process. 3.3 Perceptron List Modeler Another simple induction method, the perceptron algorithm [12], also combines evidence from attributes during classification, but uses the expression ae threshold to assign a test case to the positive (1) or negative (0) class. Each weight wm specifies the relative importance of an attribute m; taken together, these weights determine a hyper-plane that attempts to separate the two classes. The learning algorithm invokes an error-driven scheme to adjust the weights associated with each attribute. 1 Because a perceptron can only differentiate between two classes, we employed an ordered list of perceptrons that operates much like a decision list. The algorithm first learns to discriminate between the majority class and others, generating the first percep- tron. Instances in the majority class are removed, and the system trains to distinguish the new majority class from the rest, producing another perceptron. This process continues until one class remains, which is treated as a default. Although the perceptron traditionally assumes all-or-none assignment, it seems natural to interpret the scaled difference between the sum and the threshold as a likelihood. The weighted variant multiplies the update for each attribute weight by the weight for each instance, so that an instance with a smaller weight has a smaller effect on learning. To prevent small weights from causing endless oscillations, it triggers an updating cycle through the data only if an incorrectly classified instance has a weight of greater than 0.5, although all instances are used for the actual update. In reassignment, the weighted method calculates the difference between the instance value and the threshold, scaled by the sigmoid which produces bounds on the weight size. If an instance were evaluated as being perfectly at the threshold, the function would return 0.5. The factor 5 in the exponent of e distributes the resulting weights over a larger range, so the algorithm will not give a weight close to 0.5 for all instances. Otherwise the sigmoid is not tight enough to be useful for a generally small range of values. 3.4 Decision Stump Modeler For our final supervised learning algorithm, we selected decision-stump induction (e.g., Holte [4]), which differs from the others in selecting a single attribute to classify instances. To this end, it uses the information-theoretic measure is the frequency of class ck in a training set S with jCj classes. If the attribute is continuous, the algorithm orders its observed values and considers splitting between each successive pair, selecting the split with the highest score. The method applies this process recursively to 1 For the purposes of this study, we used a learning rate of iterations through the training data, which did well on all our classification tasks. the values in each subset, continuing until further divisions gain no more information, as measured by where T is the training set, Tm is a given subset of T , and jP j is the number of branches. If the attribute is nominal, the algorithm creates a separate branch for each attribute value. Each branch of the stump is then associated with the majority class of those training cases that are sorted to that branch. To accommodate weighted assignment, we adjust the equations above to sum over the weights of instances, rather than over strict frequencies, and keep simple statistical information for each branch. The reassignment weight given to each instance for class ck is calculated by where jBj is the number of instances associated with the branch to which that instance is sorted. 4. EXPERIMENTAL STUDIES We had two intuitions about our clustering framework that suggested corresponding formal hypotheses. 2 First, we expected that each algorithm would exhibit a "preference" for one of the data assignment paradigms by demonstrating better performance in that paradigm across different data sets. Second, we anticipated that, across data sets, high (low) predictive accuracy by a supervised method would be associated with relatively high (low) accuracy for the corresponding clustering algorithm. In this section, we describe our designs for the experiments to test these hypotheses and the results we obtained. 4.1 Experiments with Natural Data To test these hypotheses, we ran the generalized clustering system with each algorithm-paradigm combination on a battery of natural data sets. We also evaluated each supervised algorithm independently by training it and measuring its predictive accuracy on a separate test set. The independent variables were the assignment paradigm (for the clustering tests), the supervised learning algorithm, the data set, and the number of instances used in training. The dependent variables were the classification accuracies on unseen data. We used a standard accuracy metric to evaluate both the supervised classifiers and the clustering algorithms: where T is the test set, and where ffi(x classified correctly and 0 otherwise. When evaluating accuracy, we trained each classifier on the labeled data set with the test set removed. Because the clustering algorithms create their own classes, we added a step in which each completed cluster is assigned the actual Naturally, we also expected that no single algorithm combination would outperform all others on all data sets, but this is consistent with general findings in machine learning, and so hardly deserves the status of an hypothesis. Table 1: Supervised accuracies on four data sets. Prototype Bayes Perceptron Stump Promoters 86.0 87.0 76.0 70.0 Iris 49.3 94.7 46.0 93.3 Hayes-Roth 32.3 61.5 79.2 43.1 Glass 84.8 79.0 39.0 97.6 class of its majority population. For example, if a given cluster consists of instances that are actually class A and 10 that are actually class B, all instances in the cluster will be declared members of class A, with an accuracy of 75% for that cluster. This approach loses detail, but it let us evaluate each clustering algorithm against the "correct" clusters. We selected four data sets from the UCI repository - Pro- moters, Iris, Hayes-Roth, and Glass - that involved different numbers of classes (two to seven), different numbers of attributes (five to 57), and different attribute types (nominal, continuous, or mixed). Another factor in their selection was that each led to high classification accuracy for one of the supervised methods but (typically) to lower accuracy for the others, as shown with bold font in Table 1. This differentiation on supervised training data seemed a prerequisite for testing the predicted correlation between accuracies for supervised learning and clustering. Moreover, remember that our four supervised methods each has restricted representational power that is generally limited to one decision region per class. As a result, the fact that one such method obtains high accuracy in each of these domains suggests that each of their classes maps onto to a single cluster. This lets us assume that the number of classes in each data set corresponds to the number of clus- ters, further increasing the chances of meaningful results. For each data set, we collected a learning curve using ten-fold cross-validation, recording results for each increment of 25 data points. Typically, clustering accuracy ceased to improve early in the curve, although the supervised accuracy often continued to increase. The results we report here all involve accuracy as measured at the last point on each curve. Table 2: Unsupervised accuracies for two alternative data assignment paradigms (strict/weighted). Prototype Bayes Perceptron Stump Promoters 62.0/77.0 52.0/41.0 49.0/57.0 19.0/26.0 Iris 27.3/51.3 83.3/88.0 26.7/32.0 55.3/53.3 Hayes-Roth 37.7/39.2 30.0/40.0 38.5/38.5 34.6/36.2 Glass 84.8/51.0 44.8/61.9 26.2/34.3 77.1/74.3 Recall that our first hypothesis predicted each supervised method would construct more accurate clusters when combined with its preferred data assignment paradigm. The results in Table 2, which shows the classification accuracies for each method-paradigm combination on the four do- mains, disconfirms this hypothesis. In general, each supervised algorithm sometimes did better with one assignment scheme and sometimes with the other, depending on the do- main. Both naive Bayes and the prototype learner showed Supervised accuracy2060100 Unsupervised accuracy Decision stump Perceptron list Naive Bayes Prototype Figure 2: Supervised and unsupervised accuracies, using strict data assignment, for four algorithms on four natural data sets large shifts of this sort, though swings for the decision-stump learner were less drastic. Only the perceptron list method showed any support for our prediction, favoring weighted assignment on three data sets and a tied result on the fourth. After addressing our first hypothesis, we proceeded to test our second claim, that relatively higher (lower) accuracy in supervised mode is associated with relatively higher (lower) accuracy on unsupervised data, i.e., that they are correlated positively. Our original plan was to measure the unsupervised accuracy of each learning algorithm when combined with its preferred data assignment paradigm. Having rejected the notion of such preference, we resorted instead to measuring the relation between supervised accuracy and that achieved by clustering with strict assignment, followed by a separate measure between the accuracy of supervised learning and weighted assignment. To this end, we computed the correlation between the supervised accuracies using the 16 algorithm-domain combinations in Table 1 and the analogous strict accuracies from Table 2. The resulting correlation coefficient, was significant at the 0.01 level and explained 55 percent of the variance. Figure 2 shows that supervised accuracy is a reasonable predictor of unsupervised accuracy, thus generally supporting our hypothesis. We also calculated the correlation between supervised accuracies and the weighted accuracies from Table 2. Here the correlation was which was also significant at the 0.01 level and explained 43 percent of the variance. 4.2 Experiments with Synthetic Data Our encouraging results with natural data sets show that our framework has relevance to real-world clustering prob- lems, but they can give only limited understanding for the reasons underlying the phenomena. For this reason, we decided to carry out another study that employed synthetic data designed to reveal the detailed causes of these effects. One standard explanation for some induction methods outperforming others relies on the notion of inductive bias, which reflects the fact that some formalisms can represent certain decision regions more easily than others. Since our four supervised learning methods have quite different inductive biases, we designed four separate learning tasks, each intended to be easily learned by one of these methods but not by others. Each learning task incorporated two continuous variables and three classes, with a single contiguous decision region for each class. Thus, the domain designed with decision stumps in mind involved splits along one relevant attribute, the prototype-friendly domain involved three distinct proto- types, and so forth. The naive Bayesian classifier is difficult to foil, but for every other supervised method, we had at least one domain on which it should do relatively poorly. For each domain, we devised a generator that produced 125 random instances from either a uniform or, for the Bayes- friendly domain, a Gaussian distribution for every class, creating the same number of instances for each one. The geometric metaphor clarifies one reason that a given method should outperform others in both supervised and unsupervised mode, but it also suggests a reason why the correlation between behavior on these two tasks is imper- fect. Conventional wisdom states that clustering is easy when clusters are well separated but difficult when they are not. Thus, our data generator also included a parameter S that let us vary systematically the separation between the boundaries of each class. The predictive variables for each domain ranged from 0 to 1, so we varied the separation distance from Although we expected our synthetic domains to reproduce the positive correlation we observed with natural data, we also predicted that cluster separation should influence this effect. In particular, we thought the correlation would be lower when the gap was small, since iterative optimization would have difficulty assigning instances to the "right" unlabeled classes, whereas supervised learning would have no such difficulty. However, the correlation should increase monotonically with cluster distance, since the process of finding well-separated clusters should then be dominated by the inductive bias of the supervised learning modules. Our experimental runs with synthetic data did not support these predictions. 3 Despite our attempts to design data sets that would distinguish among the supervised learning methods, correlations between supervised and unsupervised accuracies when cluster separation considerably lower strict and weighted) than for our studies with natural domains, though still marginally significant at the 0.1 level. Moreover, our experiments showed no evidence that correlation increases with cluster separation, giving strict and weighted when for strict and weighted when Figure 3, which plots the accuracies for strict unsupervised learning against supervised accuracy when cluster separation suggests one reason for this negative result. Apparently, the correlations are being reduced by a "ceiling effect" in which the supervised accuracies (generally much higher than for our results on natural domains) show little variation, whereas the unsupervised accuracies still range widely. The supervised methods typically learn very accurate classifiers across all four synthetic domains, even though 3 This study also revealed no evidence for a preferred data assignment scheme, with the best combinations shifting across both domain and separation level. Supervised accuracy2060100 Unsupervised accuracy Decision stump Perceptron list Naive Bayes Prototype Figure 3: Supervised and unsupervised accuracies, using strict data assignment, for four algorithms with four synthetic data sets we did our best to design them otherwise. Analogous plots for higher values of the separation parameter S show even stronger versions of this effect, indicating that supervised induction benefits more from cluster separation than does unsupervised clustering, which explains why the correlation does not increase as predicted. Our expectations rested on the intuition that inductive bias and cluster separation are the dominant factors in determining the behavior of an iterative optimizer. From these negative results, and from the high correlations on natural domains, we can infer that other factors we did not vary in this experiment play an equal or more important role. Likely candidates include the number of relevant attributes, the number of irrelevant attributes, the amount of attribute noise, and the number of classes, all of which are known to affect the predictive accuracy of learned classifiers. These domain characteristics should be varied systematically in future studies that draw on synthetic data to explore the relation between clustering and supervised learning. 5. RELATED WORK As we noted earlier, there exists a large literature on clustering that others (e.g., Everitt [3]) have reviewed at length. Much of this work relies on iterative optimization to group training cases, and there exist many variants beyond the K-means and expectation-maximization algorithms familiar to most readers. For instance, Michalski and Stepp's [11] used logical rule induction to characterize its clusters and assign cases to them. More recently, Zhang, Hsu, and Dayal [15] have described the K-harmonic means method, which operates like K-means but invokes a different distance metric that usually speeds convergence. How- ever, despite this diversity, researchers have not proposed either theoretical frameworks for characterizing the space of iterative optimization methods or software frameworks to support their rapid construction and evaluation. In the broader arena, there have been some efforts to link methods for supervised and unsupervised learning. For ex- ample, Langley and Sage [8] adapted a method for inducing univariate decision trees to operate on unsupervised data and thus generate taxonomy, and, more recently, Langley [6] and Liu et al. [9] have described similar but more sophisticated approaches. The relationship between supervised and unsupervised algorithms for rule learning is more transpar- Martin [10] has reported one approach that adapts supervised techniques to construct association rules from unlabeled data. But again, such research has focused on specific algorithms rather than on general or generative frameworks. However, other areas of machine learning have seen a few frameworks of this sort. Langley and Neches [7] developed Prism, a flexible language for production-system architectures that supported many combinations of performance and learning algorithms, and later versions of Prodigy [14] included a variety of mechanisms for learning search-control knowledge. For classification problems, Kohavi et al.'s [5] MLC++ supported a broad set of supervised induction algorithms that one could invoke with considerable flexibility. The generative abilities of MLC++ are apparent from its use for feature selection and its support for novel combinations of existing algorithms. This effort comes closest to our own in spirit, both in its goals and its attempt to provide a flexible software infrastructure for machine learning. 6. CONCLUDING REMARKS In this paper, we presented a framework for iterative optimization approaches to clustering that lets one embed any supervised learning algorithm as a model-construction com- ponent. This approach produces some familiar clustering techniques, like K-means and EM, but it also generates some novel methods that have not appeared in the literature. The framework also let us evaluate some hypotheses about the relation between the resulting clustering methods and their supervised modules, which we tested using both natural and synthetic data. Our first hypothesis, that each supervised method had a preferred data assignment scheme with which it produced more accurate clusters, was not borne out the experiments. Clustering practitioners can continue to combine prototype learning with strict assignment (giving K-means) and naive Bayes with weighted assignment (giving EM), but we found no evidence that these combinations are superior to the al- ternatives. However, our experiments did support our second hypothesis by revealing strong correlations between the accuracy of supervised algorithms on natural data sets and the accuracy of iterative optimizers in which they were em- bedded. We augmented these results with experiments on synthetic data, which gave us control over decision regions and separation of clusters. These studies also produced positive correlations between supervised and unsupervised ac- curacy, but failed to reveal an effect of cluster separation. Clearly, there remains considerable room for additional research. The framework supports a variety of new clustering algorithms, each interesting in its own right but also important for testing further our hypotheses about relations between supervised and unsupervised learning. We should also carry out experiments with synthetic data that vary systematically other factors that can affect predictive accu- racy, such as irrelevant features and attribute noise. Finally, we should explore further the role of cluster separation and the reason it had no apparent influence in our studies. Although our specific results are intriguing, we attach more importance to the framework itself, which supports a new direction for studies of clustering mechanisms. We encourage other researchers to view existing techniques as examples of some generative framework and to utilize that framework both to explore the space of clustering methods and to reveal underlying relations between supervised and unsupervised approaches to induction. Ultimately, this strategy should produce a deeper understanding of the clustering process and its role in the broader science of machine learning. 7. --R Maximum likelihood from incomplete data via the EM algo- rithm Pattern Classification and Scene Analysis. Analysis. Very simple classification rules perform well on most commonly used data sets. Elements of Machine Learning. Prism user's manual. Conceptual clustering as discrimination learning. Clustering through decision tree construction. Focusing attention for observational learn- ing: The importance of context Learning from ob- servation: Conceptual clustering Principles of Neurodynamics: Perceptrons and the Theory of Brain Mechanisms. Categories and Concepts. Derivational analogy in Prodigy: Automating case acquisition --TR Very Simple Classification Rules Perform Well on Most Commonly Used Datasets Derivational Analogy in PRODIGY Elements of machine learning Clustering through decision tree construction K-Harmonic Means - A Spatial Clustering Algorithm with Boosting --CTR Tadashi Nomoto , Yuji Matsumoto, Supervised ranking in open-domain text summarization, Proceedings of the 40th Annual Meeting on Association for Computational Linguistics, July 07-12, 2002, Philadelphia, Pennsylvania Greg Hamerly , Charles Elkan, Alternatives to the k-means algorithm that find better clusterings, Proceedings of the eleventh international conference on Information and knowledge management, November 04-09, 2002, McLean, Virginia, USA Shi Zhong , Joydeep Ghosh, A unified framework for model-based clustering, The Journal of Machine Learning Research, 4, p.1001-1037, 12/1/2003

iterative optimization;clustering;supervised learning

502567

Detecting graph-based spatial outliers.

of outliers can lead to the discovery of unexpected, interesting, and useful knowledge. Existing methods are designed for detecting spatial outliers in multidimensional geometric data sets, where a distance metric is available. In this paper, we focus on detecting spatial outliers in graph structured data sets. We define statistical tests, analyze the statistical foundation underlying our approach, design several fast algorithms to detect spatial outliers, and provide a cost model for outlier detection procedures. In addition, we provide experimental results from the application of our algorithms on a Minneapolis-St.Paul(Twin Cities) traffic dataset to show their effectiveness and usefulness.

Introduction Data mining is a process to extract nontrivial, previously unknown and potentially useful infor- mation(such as knowledge rules, constraints, regularities) from data in databases [11, 4]. The explosive growth in data and databases used in business management, government administra- tion, and scientific data analysis has created a need for tools that can automatically transform the processed data into useful information and knowledge. Spatial data mining is a process of discovering interesting and useful but implicit spatial patterns. With the enormous amounts of spatial data obtained from satellite images, medical images, GIS, etc., it is a nontrivial task for humans to explore spatial data in detail. Spatial data sets and patterns are abundant in many application domains related to NASA, the National Imagery and Mapping Agency(NIMA), the National Cancer Institute(NCI), and the Unite States Department of Transportation(USDOT). Data Mining tasks can be classified into four general categories: (a) dependency detection (e.g., association rules) (b) class identification (e.g., classification, clustering) (c) class description (e.g., concept generalization), and (d) exception/outlier detection [9]. The objective of the first three categories is to identify patterns or rules from a significant portion of a data set. On the other hand, the outlier detection problem focuses on the identification of a very small subset of data objects often viewed as noises, errors, exceptions, or deviations. Outliers have been informally defined as observations which appear to be inconsistent with the remainder of that set of data [2], or which deviate so much from other observations so as to arouse suspicions that they were generated by a different mechanism [6]. The identification of outliers can lead to the discovery of unexpected knowledge and has a number of practical applications in areas such as credit card fraud, the performance analysis of athletes, voting irregularities, bankruptcy, and weather prediction. Outliers in a spatial data set can be classified into three categories: set-based outliers, multi-dimensional space-based outliers, and graph-based outliers. A set-based outlier is a data object whose attributes are inconsistent with attribute values of other objects in a given data set regardless of spatial relationships. Both multi-dimensional space-based outliers and graph-based outliers are spatial outliers, that is, data objects that are significantly different in attribute values from the collection of data objects among spatial neighborhoods. However, multi-dimension space-based outliers and graph-based outliers are based on different spatial neighborhood defini- tions. In multi-dimensional space-based outlier detection, the definition of spatial neighborhood is based on Euclidean distance, while in graph-based spatial outlier detections, the definition is based on graph connectivity. Many spatial outlier detection algorithms have been recently proposed; however, spatial outlier detection remains challenging for various reasons. First, the choice of a neighborhood is nontrivial. Second,the design of statistical tests for spatial outliers needs to account for the distribution of the attribute values at various locations as well as the aggregate distribution of attribute values over the neighborhoods. In addition, the computation cost of determining parameters for a neighborhood-based test can be high due to the possibility of join computations In this paper, we formulate a general framework for detecting outliers in spatial graph data sets, and propose an efficient graph-based outlier detection algorithm. We provide cost models for outlier detection queries, and compare underlying data storage and clustering methods that facilitate outlier query processing. We also use our basic algorithm to detect spatial and temporal outliers in a Minneapolis-St.Paul(Twin Cities) traffic data set, and show the correctness and effectiveness of our approach. 1.1 An Illustrative Application Domain: Traffic Data Set In 1995, the University of Minnesota and the Traffic Management Center(TMC) Freeway Operations group started the development of a database to archive sensor network measurements from the freeway system in the Twin Cities. The sensor network includes about nine hundred stations, each of which contains one to four loop detectors, depending on the number of lanes. Sensors embedded in the freeways monitor the occupancy and volume of traffic on the road. At regular intervals, this information is sent to the Traffic Management Center for operational purposes, e.g., ramp meter control, and research on traffic modeling and experiments. Figure 1 shows a map of the stations on highways within the Twin-Cities metropolitan area, where each polygon represents one station. Interstate freeways include I-35W, I35E, I-94, I-394, I-494, and I-694. State trunk highways include TH-100, TH-169, TH-212, TH-252, TH-5, TH-55, TH-62, TH-65, and TH-77. I-494 and I-694 together forming a ring around the Twin Cities. I-94 passes from East to North-West, while I-35W and I-35E run in a North-South direction. Downtown Minneapolis is located at the intersection of I-94, I-394, and I-35W, and downtown Saint Paul is located at the intersection of I-35E and I-94. I- I- I- I- I- I- I- I- I- I-3 Figure 1: Detector map in station level Figure 2(a) demonstrates the relationship between a station and its encompassing detectors. For each station, there is one detector installed in each lane. The traffic flow information measured by each detector can then be aggregated to the station level. Figure 2(b) shows the three basic data tables for the traffic data. The station table stores the geographical location and some related attributes for each station. The relationship between each detector and its corresponding station is captured in the detector table. The value table records all the volume and occupancy information within each 5-minute time slot at each particular station. Detector 50 Detector 51 Detector 52 Station 20 (a) Relationship between detectors and stations2 Detector Station1 Detector Table Time Detector Volume Occupancy260 12Value Table Polygon Boundary (3,5),(4,10),. (5,7),(6,4),. Station Location 26th St. 28th St. Freeway Direction Q4 Q4 Station Table Zone (b) Three basic tables Figure 2: Detector-station Relationship and Basic Tables In this application, each station exhibits both graph and attribute properties. The topological space is the map, where each station represents a node and the connection between each station and its surrounding stations can be represented as an edge. The attribute space for each station is the traffic flow information (e.g., volume, occupancy) stored in the value table. In this application, we are interested in discovering the location of stations whose measurements are inconsistent with those of their graph-based spatial neighbors and the time periods when those abnormalities arise. This outlier detection task is to: ffl Build a statistical model for a spatial data set ffl Check whether a specific station is an outlier ffl Check whether stations on a route are outliers We use three neighborhood definitions in this application as shown in Figure 3. First, we define a neighborhood based on spatial graph connectivity as a spatial graph neighborhood. In Figure are the spatial neighbors of (s are connected to s 2 in a spatial graph. Second, we define a neighborhood based on time series as a temporal neighborhood. In Figure 3, (s are the temporal neighbors of (s t 3 are consecutive time slots. In addition, we define a neighborhood based on both space and time series as a spatial-temporal neighborhood. In Figure 3, (s are the spatial-temporal neighbors of (s are connected to s 2 in a spatial graph, and t 1 , t 2 , and t 3 are consecutive time slots. 1.2 Problem Formulation In this section, we formally define the spatial outlier detection problem. Given a spatial frame-work S for the underlying spatial graph G, an attribute f over S, and neighborhood relationship R, we can build a model and construct statistical tests for spatial outliers based on a spatial graph according to the given confidence level threshold. The problem is formally defined as follows. Spatial Outlier Detection Problem Given: ffl A spatial graph is a spatial framework consisting of locations Temporal Neighborhood Spatial-Temporal Neighborhood Time Space Legend Neighbors Spatial Temporal Neighbors Additional Neighbors in Spatial-Temporal Neighborhood Spatial-Temporal Window of Neighborhood Neighborhood Figure 3: Spatial and Temporal outlier in traffic data \Theta S) is a collection of edges between locations in S. ffl A neighborhood relationship R ' S \Theta S consistent with E ffl An attribute function f a set of real numbers ffl An aggregate function f aggr : R N ! a set of real numbers to summarize values of attribute f over a neighborhood relationship R N ' R ffl Confidence level threshold ' Find: A set O of spatial outliers O Objective: ffl Correctness: outliers identified by a method have significantly different attribute values with those of their neighborhood ffl Efficiency: to minimize the computation time Constraints: ffl Attribute values for different locations in S have a normal distribution ffl The size of the data set is much greater than the main memory size ffl The range of attribute function f is the set of real numbers The formulation shows two subtasks in this spatial outlier detection problem: (a) the design of a statistical model M and a test for spatial outliers (b) the design of an efficient computation method to estimate parameters of the test, test whether a specific spatial location is an outlier, and test whether spatial locations on a given path are outliers. 1.3 Paper Scope and Outline This paper focuses on graph-based spatial outlier detection using a single attribute. Outlier detection in a multi-dimensional space using multiple attributes is outside the scope of this paper. The rest of the paper is organized as follows. Section 2 reviews related work and discusses our contributions. In Section 3, we propose our graph-based spatial outlier detection algorithm and discuss its computational complexity. The cost models for different outlier query processing are analyzed in Section 4. Section 5 presents our experimental design. The experimental observation and results are shown in Section 6. We summarize our work in Section 7. Related Work and Our Contribution Many outlier detection algorithms [1, 2, 3, 8, 9, 12, 14, 16] have been recently proposed. As shown in Figure 4, these methods can be broadly classified into two categories, namely set-based outlier detection methods and spatial-set-based outlier detection methods. The set-based outlier detection algorithms [2, 7] consider the statistical distribution of attribute values, ignoring the spatial relationships among items. Numerous outlier detection tests, known as discordancy tests [2, 7], have been developed for different circumstances, depending on the data distribution, the number of expected outliers, and the types of expected outliers. The main idea is to fit the data set to a known standard distribution, and develop a test based on distribution properties. Outlier Detection Methods Distance-based Wavelet Depth threshold Distance to k-th Neighbor Density in Neighborhood Multi-dimensional metric spatial data set Graph-based spatial data set attribute value Statistical distribution of based Spatial set based based Spatial Graph Detection Based Outlier Figure 4: Classification of outlier detection methods Spatial-set-based outlier detection methods consider both attribute values and spatial rela- tionships. They can be further grouped into two categories, namely multi-dimensional metric space-based methods and graph-based methods. The multi-dimensional metric space-based methods model data sets as a collection of points in a multidimensional space, and provide tests based on concepts such as distance, density, convex-hull depth. We discuss different example tests now. Knorr and Ng presented the notion of distance-based outliers [8, 9]. For a dimensional data set T with N objects; an object O in T is a DB(p;D)-outlier if at least a fraction p of the objects in T lies greater than distance D from O. Ramaswamy et al. [13] proposed a formulation for distance-based outliers based on the distance of a point from its th nearest neighbor. After ranking points by the distance to its k th nearest neighbor, the top n points are declared as outliers. Breunig et al. [3] introduced the notion of a "local" outlier where the outlier-degree of an object is determined by taking into account the clustering structure in a bounded neighborhood of the object, e.g., k nearest neighbors. They formally defined the outlier factor to capture this relative degree of isolation or outlierness. Their notions of outliers are based on the same theoretical foundation as density-based cluster analysis [1]. In computational geometry, some depth-based approaches [14, 12] organize data objects in convex hull layers in data space according to their peeling depth [12], and outliers are expected to be found from data objects with a shallow depth value. Conceptually, depth-based outlier detection methods are capable of processing multidimensional datasets. However, with the best case computational complexity of \Omega\Gamma N dk=2e ) for computing a convex hull, where N is the number of objects and k is the dimensionality of the dataset, depth-based outlier detection methods may not be applicable for high dimensional data sets. Yu et al. [16] introduced an outlier detection approach, called FindOut, which identifies outliers by removing clusters from the original data. Its key idea is to apply signal processing techniques to transform the space and find the dense regions in the transformed space. The remaining objects in the non-dense regions are labeled as outliers. Methods for detecting outliers in multi-dimensional Euclidean space have some limitation. First, multi-dimensional approaches assume that the data items are embedded in a isometric metric space and do not capture the spatial graph structure. Consider the application domain of traffic data analysis. A multi-dimensional method may put a detector station in the neighborhood of another detector even if they are on opposite sides of the highway (e.g., I-35W north bound at exit 230, and I-35W south bound at exit 230), leading to the potentially incorrect identification of bad detector. Secondly, they do not exploit apriori information about the statistical distribution of attribute data. Last, they seldom provide a confidence measure of the discovered outliers. In this paper, we formulate a general framework for detecting spatial outliers in a spatial data set with an underlying graph structure. We define neighborhood-based statistics and validate the statistical distribution. We then design a statistically correct test for discovering spatial outliers, and develop a fast algorithm to estimate model parameters, as well as to determine the results of a spatial outlier test on a given item. In addition, we evaluate our method in Twin Cities traffic data set and show the effectiveness and usefulness of our approach. 3 Our Approach: Spatial Outlier Detection Algorithm In this section, we list the key design decisions and propose an I/O efficient algorithm for spatial graph based outliers. 3.1 Choice of Spatial Statistic For spatial statistics, several parameters should be pre-determined before running the spatial outlier test. First, the neighborhood can be selected based on a fixed cardinality or a fixed graph distance or a fixed Euclidean distance. Second, the choice of neighborhood aggregate function can be mean, variance, and auto-correlation. Third, the choice for comparing a location with its neighbors can use either just a number or a vector of attribute values. Finally, the statistic for base distribution can be selected from various choices. The statistics we used are is the attribute value for a data record x, N(x) is the fixed cardinality set of neighbors of x, and E y2N(x) (f(y)) is the average attribute value for neighbors of x. Statistic S(x) denotes the difference of the attribute value of each data object x and the average attribute value of x 0 s neighbors. 3.2 Characterizing the Distribution of the Statistic normally distributed if attribute value f(x) is normally distributed. Proof: Given the definition of neighborhood, for each data record x, the average attribute values neighbors can be calculated. Since attribute values f(x) are normally distributed and an average of normal variables is also normally distributed, the average attribute values neighbors is also a normal distribution for a fixed cardinality neighborhood. Since the attribute value and the average attribute value over neighbors are two normal variable, the distribution of the difference of S(x) of each data object x and the average attribute value of x 0 s neighbors is also normally distributed. \Xi 3.3 Test for Outlier Detection The test for detecting an outlier can be described as j S(x)\Gamma- s oe s For each data object x with an attribute value f(x), the S(x) is the difference of the attribute value of data object x and the average attribute value of its neighbors; - s is the mean value of all S(x), and oe s is the standard deviation of all S(x). The choice of ' depends on the specified confidence interval. For example, a confidence interval of 95 percent will lead to ' - 2. 3.4 Computation of Test Parameters We now propose an I/O efficient algorithm to calculate the test parameters, e.g., mean and standard deviation for the statistics, as shown in Algorithm 1. The computed mean and standard deviation can then be used to detect the outlier in the incoming data set. Given an attribute data set V and the connectivity graph G, the TPC algorithm first retrieves the neighbor nodes from G for each data object x. It then computes the difference of the attribute value of x and the average of the attribute values of x 0 s neighbor nodes. These different values are then stored as a set in the AvgDist Set. Finally, the AvgDist Set is used to get the distribution value - s and oe s . Note that the data objects are processed on a page basis to reduce redundant I/O. 3.5 Computation of Test Results The neighborhood aggregate statistics value, e.g., mean and standard deviation, computed in the TPC algorithm can be used to verify the outliers in an incoming data set. The two verification procedures are Route Outlier Detection(ROD) and Random Node Verification(RNV). The Test Parameters Computation(TPC) Algorithm Input: S is the multidimensional attribute space; D is the attribute data set in S; F is the distance function in ND is the depth of neighbor; E) is the spatial graph; O i =Get One Object(i,D); /* Select each object from D */ NNS=Find Neighbor Nodes Set(O i ,ND,G); Find neighbor nodes of O i from G */ Accum Dist=0; O k =Get One Object(j,NNS); /* Select each object from NNS */ Accum Dist += F (O Dist / jNNSj; Add Element(AvgDist Set,i); /* Add the element to AvgDist Set */ Mean(AvgDist Dev(AvgDist Set); /* Compute Standard Deviation */ return (- s ,oe s ). Algorithm 1: Pseudo-code for test parameters computation ROD procedure detects the spatial outliers from a user specified route, as shown in Algorithm 2. The RNV procedure check the outlierness from a set of randomly generated nodes. Given route RN in the data set D with graph structure G, the ROD algorithm first retrieves the neighboring nodes from G for each data object x in the route RN , then it computes the difference S(x) between the attribute value of x and the average of attribute values of x 0 s neighboring nodes. Each S(x) can then be tested using the spatial outlier detection test j S(x)\Gamma- s oe s ' is predetermined by the given confidence interval. The steps to detect outliers in both ROD and RNV are similar, except that the RNV has no shared data access needs across tests for different nodes. The I/O operations for Find Neighbor Nodes Set() in different iterations are independent of each other in RNV. We note that the operation Find Neighbor Nodes Set() is executed once in each iteration and dominates the I/O cost of the entire algorithm. The storage of the data set should support the I/O efficient computation of this operation. We discuss the choice for storage structure and provide an experimental comparison in Section 5 and 6. The I/O cost of ROD and RNV are also dominated by the I/O cost of Find Neighbor Nodes Set() operation. Analytical Evaluation and Cost Models In this section, we provide simple algebraic cost models for the I/O cost of outlier detection operations, using the Connectivity Residue Ratio(CRR) measure of physical page clustering methods. The CRR value is defined as follows. Route Outlier Detection(ROD) Algorithm Input: S is the multidimensional attribute space; D is the attribute data set in S; F is the distance function in ND is the depth of neighbor; E) is the spatial graph; CI is the confidence interval; are mean and standard deviation calculated in TPC; RN is the set of node in a route; Output: Outlier Set. O i =Get One Object(i,D); /* Select each object from D */ NNS=Find Neighbor Nodes Set(O i ,ND,G); Find neighbor nodes of O i from G */ Accum Dist=0; O k =Get One Object(j,NNS); /* Select each object from NNS */ Accum Dist += F (O oe s /*Check the normal distribution table */ Check Normal Table(T value ,CI)== True)f Add Element(Outlier Set,i); /* Add the element to Outlier Set */ return Outlier Set. Algorithm 2: Pseudo-code for route outlier detection number of unsplit edges otal numbe of edges The CRR value is determined by the page clustering method, the data record size, and the page size. Figure 5 gives an example of CRR value calculation. The blocking factor, i.e., the number of data records within a page is three, and there are nine data records. The data records are clustered into three pages. There are a total of nine edges and six unsplit edges. The CRR value of this graph can be calculated as Table 1 lists the symbols used to develop our cost formulas. ff is the CRR value. fi denotes the blocking factor, which is the number of data records that can be stored in one memory page. is the average number of nodes in the neighbor list of a node. N is the total number of node in the data set, L is the number of node along a route, and R is the number of nodes randomly generated by users for spatial outlier verification. Page B Page C Page A Figure 5: Example of CRR Symbol Meaning ff The CRR value Average blocking factor Total number of nodes L Number of nodes in a route R Number of nodes in a random set Average number of neighbors for each node Table 1: Symbols used in Cost Analysis 4.1 Cost Modeling for Test Parameters Computation(TPC) Algorithm The TPC algorithm is a nest loop index join. Suppose that we use two memory buffers. If one memory buffer stores the data object x used in the outer loop and the other memory buffer is reserved for processing the neighbors of x, we get the following cost function to estimate the number number of page accesses. The outer loop retrieves all the data records on a page basis, and has an aggregated cost of N fi . For each node x, on average, ff neighbors are in the same page as x, and can be processed without redundant I/O. Additional data page accesses are needed to retrieve the other neighbors, and it takes at most (1 \Gamma ff) data page accesses. Thus the expected total cost for the inner loop is N 4.2 Cost Modeling for Route Outlier Detection(ROD) algorithm We get the following cost function to estimate the number of page accesses with two memory buffers for ROD algorithm. One memory buffer is reserved for processing the node x to be verified, and the other is used to process the neighbors of x. For each node x, on the average, its successor node y is on the same page as x with probability ff, and can be processed with no redundant page accesses. The cost to access all the nodes along a route is L To process the neighbors of each node, ff neighbors are on the same page as x. Additional data page accesses are needed to retrieve the other neighbors, and it takes at most (1 \Gamma ff) data page accesses. 4.3 Cost Modeling for Random Node Verification(RNV) algorithm We get the following cost function to estimate the number of page accesses with two memory buffers for the RNV algorithm. One memory buffer is reserved for processing the node x to be verified, the other is used to process the neighbors of x. Since the memory buffer is assumed to be cleared for each consecutive random node, we need R page accesses to process all these random nodes. For each node x, ff neighbors are on the same page as x, and can be processed without extra I/O. Additional data page accesses are needed to retrieve the other neighbors, and it takes at most (1 \Gamma ff) data page accesses. Thus, the expected total cost to process the neighbor of R nodes is R 5 Experiment Design In this section, we describe the layout of our experiments and then illustrate the candidate clustering methods. 5.1 Experimental Layout The design of our experiments is shown in Figure 6. Using the Twin Cities Highway Connectivity Graph(TCHCG), we took data from the TCHCG and physically stored the data set into data pages using different clustering strategies and page sizes. These data pages were then processed to generate the global distribution or sampling distribution, depending on the size of the data sets. We compared different data page clustering schemes: CCAM [15], Z-ordering [10], and Cell-tree [5]. Other parameters of interest were the size of the memory buffer, the buffering strategies, the memory block size(page size), and the number of neighbors. The measures of our experiments were the CRR values and I/O cost for each outlier detection procedure. Clustering method Page Size Sets of pages of data Sets of pages of data Highway Connectivity Graph Twin-Cities Buffering Size Buffer No of Neighbors Buffering strategy No of neighbors Buffer size Z-order Cell-tree Test Parameters Test Parameters Computation (TPC) (Nest loop index join) CRR I/O Cost Route Outlier Detection(ROD). Verification (RNV). Random Node Figure Experimental Layout The experiments were conducted on many graphs. We present the results on a representative graph, which is a spatial network with 990 nodes that represents the traffic detector stations for a 20-square-mile section of the Twin Cities area. This data set is provided by the Minnesota Dept. of Transportation(MnDot). We used a common record type for all the clustering methods. Each record contains a node and its neighbor-list, i.e., a successor-list and a predecessor-list. We also conducted performance comparisons of the I/O cost for outlier-detection query processing. 5.2 Candidate Clustering Methods In this section we describe the candidate clustering methods used in the experiments. Connectivity-Clustered Access Method(CCAM): CCAM [15] clusters the nodes of the graph via graph partitioning, e.g., Metis. Other graph-partitioning methods can also be used as the basis of our scheme. In addition, an auxiliary secondary index is used to support query operations. The choice of a secondary index can be tailored to the application. We used the tree with Z-order in our experiments, since the benchmark graph was embedded in graphical space. Other access methods such as the R-tree and Grid File can alternatively be created on top of the data file, as secondary indices in CCAM to suit the application. Linear Clustering by Z-order: Z-order [10] utilizes spatial information while imposing a total order on the points. The Z-order of a coordinate (x,y) is computed by interweaving the bits in the binary representation of the two values. Alternatively, Hilbert ordering may be used. A conventional one-dimensional primary index (e.g. B + -tree) can be used to facilitate the search. Cell Tree: A cell tree [5] is a height-balanced tree. Each cell tree node corresponds, not necessarily to a rectangular box, but to a convex polyhedron. A cell tree restricts polyhedra to partitions of a BSP(Binary Space Partitioning), to avoid overlaps among sibling polyhedra. Each cell tree node corresponds to one disk space, and the leaf nodes contain all the information required to answer a given search query. The cell tree can be viewed as a combination of a BSP- and R + -tree, or as a BSP-tree mapped on paged secondary memory. 6 Experimental Results In this section, we illustrate the outlier examples detected in the traffic data set, present the results of our experiments, and test the effectiveness of the different page clustering methods. To simplify the comparison, the I/O cost represents the number of data pages accessed. This represents the relative performance of the various methods for very large databases. For smaller databases, the I/O cost associated with the indices should be measured. Here we present the evaluation of I/O cost for the TPC algorithm. The evaluations of I/O cost for the RNV and ROD algorithms are available in the full version paper. 6.1 Outliers Detected We tested the effectiveness of our algorithm on the Twin Cities traffic data set and detected numerous outliers, as described in the following examples. Time Volume Traffic Volume v.s. Time for Station 138 on 1/12 1997 (a) Station 138 Volume Traffic Volume v.s. Time for Station 139 on 1/12 1997 (b) Outlier Station 139 Time Volume Traffic Volume v.s. Time for Station 140 on 1/12 1997 (c) Station 140 Figure 7: Outlier station 139 and its neighbor stations on 1/12 1997 In Figure 7, the abnormal station(Station 139) was detected with volume values significantly inconsistent with the volume values of its neighboring stations 138 and 140. Note that our basic algorithm detects outlier stations in each time slot; the detected outlier stations in each time slot are then aggregated on a daily basis.4080120160Average Traffic Volume(Time v.s. Station) Time Station ID(North Bound) (a) I-35W North Bound2060100140Average Traffic Volume(Time v.s. Station) Time Station ID(South Bound) (b) I-35W South Bound Figure 8: An example of outliers Figure 8 shows another example of traffic flow outliers. Figures 8(a) and (b) are the traffic volume maps for I-35W North Bound and South Bound, respectively, on 1/21/1997. The X-axis is a 5-minute time slot for the whole day and the Y-axis is the label of the stations installed on the highway, starting from 1 on the north end to 61 on the south end. The abnormal white line at 2:45pm and the white rectangle from 8:20am to 10:00am on the X-axis and between stations 29 to 34 on the Y-axis can be easily observed from both (a) and (b). The white line at 2:45pm is an instance of temporal outliers, where the white rectangle is a spatial-temporal outlier. Moreover, station 9 in Figure 8(a) exhibits inconsistent traffic flow compared with its neighboring stations, and was detected as a spatial outlier. 6.2 Testing Statistical Assumption In this traffic data set, the volume values of all stations at one moment are approximately a normal distribution. The histogram of stations on different volumes are shown in Figure 9(a) with a normal probability distribution superimposed. As can be seen in Figure 9(a), the normal distribution approximates the volume distribution very well. We calculated the interval of v and oe are the mean and standard deviation of the volume distribution, and the percentages of measurements falling in the three intervals are equal to 68.27%, 95.45%, and 99.73%, respectively. This pattern fits well with a normal distribution since the expected values in a normal distribution are 68%, 95%, and 100%. Moreover, we plot the normal probability plot in Figure9(b), and it appears linear. Hence the volume values of all stations at the same time are approximately a normal distribution. Given the definition of neighborhood, we then calculate the average volume value ( - around its k neighbors according to topological relationship for each station. Since the volume values are normally distributed, the average of the normal variables is also a normal distribution. Volume Distribution at 10:00am on 1/15/1997 Volume Value Number of Stations (a) Histogram of traffic volume distribution Probability Volume Normal Distribution Probability Plot at 10:00am on 1/15/1997 (b) Normal probability plot for traffic volume distribu- tion 4100Normalized Volume Difference over Spatial Neighborhood at 10:00am on 1/15/1997 Normalized Volume Difference over Spatial Neighborhood Number of Stations (c) Histogram of volume difference over neighborhood Figure 9: Verification of normal distribution for traffic volumes and volume difference over neighbors Since the volume values and the average volume values over neighborhoods are normally dis- tributed, the difference(v \Gamma - v) between these volumes and their corresponding average volume values over neighborhoods is also a normal distribution since the difference of the two random normal random variables is always normal, as shown in Figure 9(c). Given the confidence level 100(1-ff)%, we can calculate the confidence interval for the difference distributions, i.e., 100(1- ff) percentage of difference value distribution lies between \Gammaz ff=2 and z ff=2 standard deviation of the mean in the sample space. So we can classify the spatial outliers at the given confidence level threshold. 6.3 Evaluation of Proposed Cost Model We evaluated the I/O cost for different clustering methods for outlier detection procedures, namely, Test Parameters Computation(TPC), Route Outlier Detection(ROD) and Random Node Verification(RNV). The experiments used Twin-Cities traffic data with page size 1K bytes, and two memory buffers. Table 2 shows the number of data page accesses for each procedure under various clustering methods. The CRR value for each method is also listed in the table. The cost function for TPC is C ff). The cost function for RNV is ff). The cost function for ROD is as described in Section 4.2. Clustering Parameters Computation Random Node Verification Route Outlier Detect Method Actual Predicted Actual Predicted Actual Predicted CRR Zord 1263 1269 349 357 78 79 0.31 Table 2: The actual I/O cost and predicted cost model for different clustering methods As shown in Table 1, CCAM produced the lowest number of data page accesses for the outlier detection procedures. This is to be expected, since CCAM generated the highest CRR value. 6.4 Evaluation of I/O cost for TPC algorithm In this section, we present the results of our evaluation of the I/O cost and CRR value for alternative clustering methods while computing the test parameters. The parameters of interest are buffer size, page size, number of neighbors, and neighborhood depth. 6.4.1 The effect of page size and CRR value Figures (a) and (b) show the number of data pages accessed and the CRR values respectively, for different page clustering methods as the page sizes change. The buffer size is fixed at Kbytes. As can be seen, a higher CRR value implies a lower number of data page accesses, as predicted in the cost model. CCAM outperforms the other competitors for all four page sizes, and CELL has better performance than Z-order clustering. 6.4.2 The effect of neighborhood cardinality We evaluated the effect of varying the number of neighbors and the depth of neighbors for different page clustering methods. We fixed the page size at 1K, and the buffer size at 4K, and used the LRU buffering strategy. Figure 11 shows the number of page accesses as the number of neighbors for each node increases from 2 to 10. CCAM has better performance than Z-order and CELL. The performance ranking for each page clustering method remains the same for different numbers of neighbors. Figure 11 shows the number of page accesses as the Number of page Page Size (K bytes) Zord (a) Page Accesses0.40.60.811 2 3 4 5 6 7 8 CRR Value Page Size (K Bytes) Zord (b) CRR Figure 10: Effect of page size on data page accesses and CRR (buffer size = 32K) neighborhood depth increases from 1 to 5. CCAM has better performance than Z-order and CELL for all the neighborhood depths.5001500250035004500 Number of page No of Neighbors Zord (a) Number of Neighbors1000300050007000 Number of page Neighborhood Depth Zord (b) Neighborhood Depth Figure 11: Effect of neighborhood cardinality on data page accesses (Page size = 1K, Buffer Conclusions In this paper, we focused on detecting outliers in spatial graph data sets. We proposed the notion of a neighbor outlier in graph structured data sets, designed a fast algorithm to detect outliers, analyzed the statistical foundation underlying our approach, provided the cost models for different outlier detection procedures, and compared the performance of our approach using different data clustering approaches. In addition, we provided experimental results from the application of our algorithm on Twin Cities traffic archival to show its effectiveness and usefulness. We have evaluated alternative clustering methods for neighbor outlier query processing, including model construction, random node verification, and route outlier detection. Our experimental results show that the CCAM, which achieves the highest CRR, provides the best overall performance. Acknowledgment We are particularly grateful to Professor Vipin Kumar, and our Spatial Database Group members, Weili Wu, Yan Huang, Xiaobin Ma,and Hui Xiong for their helpful comments and valuable discussions. We would also like to express our thanks to Kim Koffolt for improving the readability and technical accuracy of this paper. This work is supported in part by USDOT grant on high performance spatial visualization of traffic data, and is sponsored in part by the Army High Performance Computing Research Center under the auspices of the Department of the Army, Army Research Laboratory cooperative agreement number DAAH04-95-2-0003/contract number DAAH04-95-C-0008, the content of which does not necessarily reflect the position or the policy of the government, and no official endorsement should be inferred. This work was also supported in part by NSF grant #9631539. --R Optics: Ordering points to identify the clustering structure. Outliers in Statistical Data. Optics-of: Identifying local outliers. Advances in Knowledge Discovery and Data Mining. The Design of the Cell Tree: An Object-Oriented Index Structure for Geometric Databases of Outliers. Applied Multivariate Statistical Analysis. A unified notion of outliers: Properties and computation. Algorithms for mining distance-based outliers in large datasets A Class of Data Structures for Associative Searching. Knowledge Discovery in Databases. Computatinal Geometry: An Introduction. Efficient algorithms for mining outliers from large data sets. Computing depth contours of bivariate point clouds. A connectivity-clustered access method for aggregate queries on transportation networks-a summary of results Finding outliers in very large datasets. --TR Computational geometry: an introduction Applied multivariate statistical analysis Computing depth contours of bivariate point clouds OPTICS Efficient algorithms for mining outliers from large data sets A class of data structures for associative searching The Design of the Cell Tree OPTICS-OF Algorithms for Mining Distance-Based Outliers in Large Datasets --CTR Yan Huang , Hui Xiong , Shashi Shekhar , Jian Pei, Mining confident co-location rules without a support threshold, Proceedings of the ACM symposium on Applied computing, March 09-12, 2003, Melbourne, Florida a <u>L</u>inear <u>S</u>emantic <u>S</u>can <u>S</u>tatistic technique for detecting anomalous windows, Proceedings of the 2005 ACM symposium on Applied computing, March 13-17, 2005, Santa Fe, New Mexico Sanjay Chawla , Pei Sun, SLOM: a new measure for local spatial outliers, Knowledge and Information Systems, v.9 n.4, p.412-429, April 2006 Shashi Shekhar , Yan Huang , Judy Djugash , Changqing Zhou, Vector map compression: a clustering approach, Proceedings of the 10th ACM international symposium on Advances in geographic information systems, November 08-09, 2002, McLean, Virginia, USA Ian Davidson , Goutam Paul, Locating secret messages in images, Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining, August 22-25, 2004, Seattle, WA, USA Jeffrey Xu Yu , Weining Qian , Hongjun Lu , Aoying Zhou, Finding centric local outliers in categorical/numerical spaces, Knowledge and Information Systems, v.9 n.3, p.309-338, March 2006 Chang-Tien Lu , Yufeng Kou , Jiang Zhao , Li Chen, Detecting and tracking regional outliers in meteorological data, Information Sciences: an International Journal, v.177 n.7, p.1609-1632, April, 2007 Victoria Hodge , Jim Austin, A Survey of Outlier Detection Methodologies, Artificial Intelligence Review, v.22 n.2, p.85-126, October 2004

outlier detection;spatial graphs;Spatial Data Mining

502591

Bipartite graph partitioning and data clustering.

Many data types arising from data mining applications can be modeled as bipartite graphs, examples include terms and documents in a text corpus, customers and purchasing items in market basket analysis and reviewers and movies in a movie recommender system. In this paper, we propose a new data clustering method based on partitioning the underlying bipartite graph. The partition is constructed by minimizing a normalized sum of edge weights between unmatched pairs of vertices of the bipartite graph. We show that an approximate solution to the minimization problem can be obtained by computing a partial singular value decomposition (SVD) of the associated edge weight matrix of the bipartite graph. We point out the connection of our clustering algorithm to correspondence analysis used in multivariate analysis. We also briefly discuss the issue of assigning data objects to multiple clusters. In the experimental results, we apply our clustering algorithm to the problem of document clustering to illustrate its effectiveness and efficiency.

INTRODUCTION analysis is an important tool for exploratory data mining applications arising from many diverse disciplines. Informally, cluster analysis seeks to partition a given data set into compact clusters so that data objects within a cluster are more similar than those in distinct clusters. The literature on cluster analysis is enormous including contributions from many research communities. (see [6, 9] for recent surveys of some classical approaches.) Many traditional clustering algorithms are based on the assumption that the given dataset consists of covariate information (or attributes) for each individual data object, and cluster analysis can be cast as a problem of grouping a set of n-dimensional vectors each representing a data object in the dataset. A familiar example is document clustering using the vector space model [1]. Here each document is represented by an n-dimensional vector, and each coordinate of the vector corresponds to a term in a vocabulary of size n. This formulation leads to the so-called term-document matrix for the representation of the collection of documents, where a ij is the so-called term frequency, i.e., the number of times term i occurs in document j. In this vector space model terms and documents are treated asymmetrically with terms considered as the covariates or attributes of documents. It is also possible to treat both terms and documents as first-class citizens in a symmetric fashion, and consider a ij as the frequency of co-occurrence of term i and document j as is done, for example, in probabilistic latent semantic indexing [12]. 1 In this paper, we follow this basic principle and propose a new approach to model terms and documents as vertices in a bipartite graph with edges of the graph indicating the co-occurrence of terms and documents. In addition we can optionally use edge weights to indicate the frequency of this co-occurrence. Cluster analysis for document collections in this context is based on a very intuitive notion: documents are grouped by topics, on one hand documents in a topic tend to more heavily use the same subset of terms which form a term cluster, and on the other hand a topic usually is characterized by a subset of terms and those documents heavily using those terms tend to be about that particular topic. It is this interplay of terms and documents which gives rise to what we call bi-clustering by which terms and documents are simultaneously grouped into semantically co- Our clustering algorithm computes an approximate global optimal solution while probabilistic latent semantic indexing relies on the EM algorithm and therefore might be prone to local minima even with the help of some annealing process. herent clusters. Within our bipartite graph model, the clustering problem can be solved by constructing vertex graph partitions. Many criteria have been proposed for measuring the quality of graph partitions of undirected graphs [4, 14]. In this pa- per, we show how to adapt those criteria for bipartite graph partitioning and therefore solve the bi-clustering problem. A great variety of objective functions have been proposed for cluster analysis without efficient algorithms for finding the (approximate) optimal solutions. We will show that our bipartite graph formulation naturally leads to partial SVD problems for the underlying edge weight matrix which admit efficient global optimal solutions. The rest of the paper is organized as follows: in section 2, we propose a new criterion for bipartite graph partitioning which tends to produce balanced clusters. In section 3, we show that our criterion leads to an optimization problem that can be approximately solved by computing a partial SVD of the weight matrix of the bipartite graph. In section 4, we make connection of our approximate solution to correspondence analysis used in multivariate data analysis. In section 5, we briefly discuss how to deal with clusters with overlaps. In section 6, we describe experimental results on bi-clustering a dataset of newsgroup articles. We conclude the paper in section 7 and give pointers to future research. 2. We denote a graph by G(V; E), where V is the vertex set and E is the edge set of the graph. A graph G(V; E) is bipartite with two vertex classes X and Y if and each edge in E has one endpoint in X and one endpoint in Y . We consider weighted bipartite graph denotes the weight of the edge between vertex i and j. We let if there is no edge between vertices i and j. In the context of document clustering, X represents the set of terms and Y represents the set of documents, and w ij can be used to denote the number of times term i occurs in document j. A vertex partition of G(X;Y; W ) denoted by \Pi(A; B) is defined by a partition of the vertex sets X and Y , respectively: , where for a set S, S c denotes its compliment. By convention, we pair A with B, and A c with . We say that a pair of vertices x 2 X and y 2 Y is matched with respect to a partition \Pi(A; B) if there is an edge between x and y, and either x 2 A and y 2 B or . For any two subsets of vertices and T ae Y , define i.e., W (S; T ) is the sum of the weights of edges with one endpoint in S and one endpoint in T . The quantity W (S; T ) can be considered as measuring the association between the vertex sets S and T . In the context of cluster analysis edge weight measures the similarity between data objects. To partition data objects into clusters, we seek a partition of such that the association unmatched vertices is as small as possible. One possibility is to consider for a partition \Pi(A; B) the following quantity (1) Intuitively, choosing \Pi(A; B) to minimize cut(A; B) will give rise to a partition that minimizes the sum of all the edge weights between unmatched vertices. In the context of document clustering, we try to find two document clusters B and B c which have few terms in common, and the documents in B mostly use terms in A and those in B c use terms in A c . Unfortunately, choosing a partition based entirely on cut(A; B) tends to produce unbalanced clusters, i.e., the sizes of A and/or B or their compliments tend to be small. Inspired by the work in [4, 5, 14], we propose the following normalized variant of the edge cut in (1) The intuition behind this criterion is that not only we want a partition with small edge cut, but we also want the two subgraphs formed between the matched vertices to be as dense as possible. This latter requirement is partially satisfied by introducing the normalizing denominators in the above equation. 2 Our bi-clustering problem is now equivalent to the following optimization problem min i.e., finding partitions of the vertex sets X and Y to minimize the normalized cut of the bipartite graph G(X;Y; W ). 3. APPROXIMATESOLUTIONSUSING SINGULAR VECTORS Given a bipartite graph G(X;Y; W ) and the associated partition \Pi(A; B). Let us reorder the vertices of X and Y so that vertices in A and B are ordered before vertices in A c respectively. The weight matrix W can be written in a block format i.e., the rows of W11 correspond to the vertices in the vertex set A and the columns of W11 correspond to those in B. Therefore G(A;B;W11 ) denotes the weighted bipartite graph corresponding to the vertex sets A and B. For any i.e., s(H) is the sum of all the elements of H. It is easy to see from the definition of Ncut, 2 A more natural criterion seems to be However, it can be shown that it will leads to an SVD problem with the same set of left and right singular vectors. In order to make connections to SVD problems, we first consider the case when W is symmetric. 3 It is easy to see that with W symmetric (denoting Ncut(A; A) by Ncut(A)), we have Let e be the vector with all its elements equal to 1. Let D be the diagonal matrix such that W be the vector with ae It is easy to verify that Then and Notice that (D \Gamma W scalar s, we have (se To cast (4) in the form of a Rayleigh quotient, we need to find s such that (se it follows from the above equation that is easy to see that y T Thus min A ae y T (D \Gamma W )y oe where If we drop the constraints y let the elements of y take arbitrary continuous values, then the optimal y can be approximated by the following relaxed continuous minimization problem, min ae y T (D \Gamma W )y oe Notice that it follows from W D \Gamma1=2 WD \Gamma1=2 (D 1=2 3 A different proof for the symmetric case was first derived in [14]. However, our derivation is simpler and more transparent and leads naturally to the SVD problems for the rectangular case. and therefore D 1=2 e is an eigenvector of D \Gamma1=2 WD \Gamma1=2 corresponding to the eigenvalue 1. It is easy to show that all the eigenvalues of D \Gamma1=2 WD \Gamma1=2 have absolute value at most 1 (See the Appendix). Thus the optimal y in (5) can be computed as y, where - y is the second largest eigenvector of D \Gamma1=2 WD \Gamma1=2 . Now we return to the rectangular case for the weight matrix W , and let DX and DY be diagonal matrices such that e: (6) Consider a partition \Pi(A; B), and define ae ae Let W have the block form as in (2), and consider the augmented 22 If we interchange the second and third block rows and columns of the above matrix, we obtain6 6 4 22 07 7 5 j W22 and the normalized cut can be written as a form that resembles the symmetric case (3). Define Then we have v. It is also easy to see that Therefore, min x6=0;y 6=0 ae 2x T Wy oe In [11], the Laplacian of - W is used for partitioning a rectangular matrix in the context of designing load-balanced matrix-vector multiplication algorithms for parallel compu- tation. However, the eigenvalue problem of the Laplacian of W does not lead to a simpler singular value problem. Ignoring the discrete constraints on the elements of x and y, we have the following continuous maximization problem, x6=0;y 6=0 ae 2x T Wy oe Without the constraints x T the above problem is equivalent to computing the largest singular triplet of D \Gamma1=2 Y (see the Appendix). From (6), we have Y (D 1=2 (D \Gamma1=2 and similarly to the symmetric case, it is easy to show that all the singular values of D \Gamma1=2 Y are at most 1. Therefore, an optimal pair fx; yg for (8) can be computed as x and y, where - x and - y are the second largest left and right singular vectors of D \Gamma1=2 Y , respectively (see the Appendix). With the above discus- sion, we can now summerize our basic approach for bipartite graph clustering incorporating a recursive procedure. Algorithm. Spectral Recursive Embedding (SRE) Given a weighted bipartite graph E) with its edge weight matrix W : 1. Compute DX and DY and form the scaled weight Y . 2. Compute the second largest left and right singular vectors of - x and - y. 3. Find cut points cx and cy for x and y, respectively. 4. Form partitions vertex set Y . 5. Recursively partition the sub-graphs G(A;B) and necessary. Two basic strategies can be used for selecting the cut points cx and cy . The simplest strategy is to set and cy = 0. Another more computing-intensive approach is to base the selection on Ncut: Check N equally spaced splitting points of x and y, respectively, find the cut points cx and cy with the smallest Ncut [14]. Computational complexity. The major computational cost of SRE is Step 2 for computing the left and right singular vectors which can be obtained either by power method or more robustly by Lanczos bidiagonalization process [8, Chapter 9]. Lanczos method is an iterative process for computing partial SVDs in which each iterative step involves the computation of two matrix-vector multiplications - and - vectors u and v. The computational cost of these is roughly proportional to nnz( - W ), the number of nonzero elements of - W . The total computational cost of SRE is O(c sre k svd nnz( - sre the the level of recursion and k svd is the number of Lanczos iteration steps. In general, k svd depends on the singular value gaps of - W . Also notice that nnz( - is the average number of terms per document and n is the total number of document. Therefore, the total cost of SRE is in general linear in the number of documents to be clustered. 4. CONNECTIONSTOCORRESPONDENCE ANALYSIS In its basic form correspondence analysis is applied to an m-by-n two-way table of counts W [2, 10, 16]. Let the sum of all the elements of W , DX and DY be diagonal matrices defined in section 3. Correspondence analysis seeks to compute the largest singular triplets of the (DX (i; i)=w)(DY (j; j)=w) The matrix Z can be considered as the correlation matrix of two group indicator matrices for the original W [16]. We now show that the SVD of Z is closely related to the SVD of - Y . In fact, in section 3, we showed that D 1=2 Y e are the left and right singular vectors of - W corresponding to the singular value one, and it is also easy to show that all the singular values of - W are at most 1. Therefore, the rest of the singular values and singular vectors of - W can be found by computing the SVD of the following rank-one modification of - Y Y k2 which has (i; and is a constant multiple of the (i; j) element of Z. There- fore, normalized-cut based cluster analysis and correspondence analysis arrive at the same SVD problems even though they start with completely different principles. It is worth-while to explore more deeply the interplay between these two different points of views and approaches, for example, using the statistical analysis of correspondence analysis to provide better strategy for selecting cut points and estimating the number of clusters. 5. PARTITIONS WITH OVERLAPS So far in our discussion, we have only looked at hard clus- tering, i.e., a data object belongs to one and only one cluster. In many situations, especially when there are much overlap among the clusters, it is more advantageous to allow data objects to belong to different clusters. For example, in document clustering, certain groups of words can be shared by two clusters. Is it possible to model this overlap using our bipartite graph model and also find efficient approximate solutions? The answer seems to be yes, but our results at this point are rather preliminary and we will only illustrate the possibilities. Our basic idea is that when computing B), we should disregard the contributions of the set of vertices that is in the overlap. More specifically, let A and B, where OX denotes the Figure 1: Sparsity patterns of a test matrix before clustering (left) and after clustering (right) overlap between the vertex subsets A[OX and - A[OX , and OY the overlap between B [ OY and - However, we can make Ncut(A; B; - smaller simply by putting more vertices in the overlap. Therefore, we need to balance these two competing quantities: the size of the overlap and the modified normalized cut by minimizing where ff is a regularization parameter. How to find an efficient method for computing the (approximate) optimal solution to the above minimization problem still needs to be investigated. We close this section by presenting an illustrative example showing that in some situations, the singular vectors already automatically separating the overlap sets while giving the coordinates for carrying out clustering. Example 1. We construct a sparse m-by-n rectangular matrix so that W11 and W22 are relatively denser than W12 and W21 . We also add some dense rows and columns to the matrix W to represent row and column overlaps. The left panel of Figure 1 shows the sparsity pattern of - W , a matrix obtained by randomly permuting the rows and columns of W . We then compute the second largest left and right singular vectors of D \Gamma1=2 WD \Gamma1=2 Y , say x and y, then sort the rows and columns of - W according to the values of the entries in Y y, respectively. The sparsity pattern of this permuted - W is shown on the right panel of Figure 1. As can be seen that the singular vectors not only do the job of clustering but at the same time also concentrate the dense rows and columns at the boundary of the two clusters. 6. EXPERIMENTS In this section we present our experimental results on clustering a dataset of newsgroup articles submitted to 20 news- groups. 5 This dataset contains about 20,000 articles (email messages) evenly divided among the 20 newsgroups. We list the names of the newsgroups together with the associated group labels (the labels will be used in the sequel to identify the newsgroups). NG2: comp.graphics NG3: comp.os.ms-windows.misc NG4: comp.sys.ibm.pc.hardware NG5:comp.sys.mac.hardware NG7:misc.forsale NG8: rec.autos NG9:rec.motorcycles NG10: rec.sport.baseball NG11:rec.sport.hockey NG12: sci.crypt NG13:sci.electronics NG14: sci.med NG15:sci.space soc.religion.christian NG17:talk.politics.guns NG18: talk.politics.mideast NG20: talk.religion.misc We used the bow toolkit to construct the term-document matrix for this dataset, specifically we use the tokenization option so that the UseNet headers are stripped, and we also applied stemming [13]. Some of the newsgroups have large overlaps, for example, the five newsgroups comp.* about computers. In fact several articles are posted to multiple newsgroups. Before we apply clustering algorithms to the dataset, several preprocessing steps need to be consid- ered. Two standard steps are weighting and feature selec- tion. For weighting, we considered a variant of tf.idf weighting scheme, tf log 2 (n=df); where tf is the term frequency and df is the document frequency and several other variations listed in [1]. For feature selection, we looked at three approaches 1) deleting terms that occur less than certain number of times in the dataset; 2) deleting terms that occur in less than certain number of documents in the dataset; selecting terms according to mutual information of terms and documents defined as x where y represents a term and x a document [15]. In general we found out that the traditional tf.idf based weighting schemes do not improve performance for SRE. One possible explanation comes from the connection with correspondence analysis, the raw frequencies are samples of co-occurrence probabilities, and the pre- and post-multiplication by D \Gamma1=2 and D \Gamma1=2 Y in D \Gamma1=2 Y automatically taking into account of weighting. We did, however, found out that trimming the raw frequencies can sometimes improve performance for SRE, especially for the anomalous cases where some words can occur in certain documents an unusual number of times, skewing the clustering process. 5 The newsgroup dataset together with the bow toolkit for processing it can be downloaded from http://www.cs.cmu.edu/afs/cs/project/theo-11/www/ naive-bayes.html. Table 1: Comparison of spectral embedding (SRE), PDDP, and K-means (NG1/NG2) Mixture SRE PDDP K-means 50/100 90:57 \Sigma 3:11% 86:11 \Sigma 3:94% (86; 5; Table 2: Comparison of spectral embedding (SRE), PDDP, and K-means (NG10/NG11) Mixture SRE PDDP K-means For the purpose of comparison, we consider two other clustering methods: 1) K-means method [9]; 2) Principal direction divisive partion (PDDP) method [3]. K-means method is a widely used cluster analysis tool. The variant we used employs the Euclidean distance when comparing the dissimilarity between two documents. When applying K-means, we normalize the length of each document so that it has Euclidean length one. In essence, we use the cosine of the angle between two document vectors when measuring their similarity. We have also tried K-means without document length normalization, the results are far worse and therefore we will not report the corresponding results. Since K-means method is an iterative method, we need to specify a stopping criterion. For the variant we used, we compare the centroids between two consecutive iterations, and stop when the difference is smaller than a pre-defined tolerance. PDDP is another clustering method that utilizes singular vectors. It is based on the idea of principal component analysis and has been shown to outperform several standard clustering methods such as hierarchical agglomerative algorithm [3]. First each document is considered as a multivariate data point. The set of document is normalized to have unit Euclidean length and then centered, i,e., let W be the term-document matrix, and w be the average of the columns of W . Compute the largest singular value triplet fu; oe; vg of split the set of documents based on their values of the simple scheme is to let those with positive v i go into one cluster and those with nonnegative v i inot another cluster. Then the whole process is repeated on the term-document matrices of the two clusters, respectively. Although both our clustering method SRE and PDDP make use of the singular vectors of some versions of the term-document matrices, they are derived from fundamentally different principles. PDDP is a feature-based clustering method, projecting all the data points to the one-dimensional subspace spanned by the first principal axis; SRE is a similarity-based clustering method, two co-occurring variables (terms and documents in the context of document clustering) are simultaneously clustered. Unlike SRE, PDDP does not have a well-defined objective function for minimization. It only partitions the columns of the term-document matrices while SRE partitions both of its rows and columns. This will have significant impact on the computational costs. PDDP, however, has an advantage that it can be applied to dataset with both positive and negative values while SRE can only be applied to datasets with nonnegative data values. Example 2. In this example, we examine binary clustering with uneven clusters. We consider three pairs of news- groups: newsgroups 1 and 2 are well-separated, 10 and 11 are less well-separated and have a lot of overlap. We used document frequency as the feature selection criterion and delete words that occur in less than 5 documents in each datasets we used. For both K-means and PDDP we apply tf.idf weighting together with document length normalization so that each document vector will have Euclidean norm one. For SRE we trim the raw frequency so that the maximum is 10. For each newsgroup pair, we select four types of mixture of articles from each newsgroup: x=y indicates that x articles are from the first group and y articles are from the second group. The results are listed in Table 1 for groups 1 and 2, Table 2 for groups 10 and 11 and Table 3 for groups We list the means and standard deviations for 100 random samples. For PDDP and K-means we also include a triplet of numbers which indicates how many of the 100 samples SRE performs better (the first number), the same (the second number) and worse (the third num- ber) than the corresponding methods (PDDP or K-means). We should emphasize that K-means method can only find local minimum, and the results depend on initial values and stopping criteria. This is also reflected by the large standard deviations associated with K-means method. From the three tests we can conclude that both SRE and PDDP outperform K-means method. The performance of SRE and PDDP are similar in balanced mixtures, but SRE is superior to PDDP in skewed mixtures. Example 3. In this example, we consider an easy multi-cluster case, we examine five newsgroups 2; 9; 10; 15; was also considered in [15]. We sample 100 articles from each newsgroups, we use mutual information for feature selection. We use minimum normalized cut as cut point for each level of the recursion. For one sample, Table 4 gives the confusion matrix. The accuracy for this sample is 88:2%. We also tested two other samples with accuracy 85:4% and 81:2% which compare favorably with those obtained for three samples with accuracy 59%, 58% and 53% reported in [15]. In the following we also listed the top few words for each clusters computed by mutual information. Table 3: Comparison of spectral embedding (SRE), PDDP, and K-means (NG18/NG19) Mixture SRE PDDP K-means Table 4: Confusion matrix for newsgroups f2; 9; 10; 15; 18g mideast graphics space baseball motorcycles cluster cluster cluster cluster cluster armenian israel arab palestinian peopl jew isra iran muslim kill turkis war greek iraqi adl call 2: imag file bit green gif mail graphic colour group version comput jpeg blue xv ftp ac uk list 3: univers space nasa theori system mission henri moon cost sky launch orbit shuttl physic work clutch year game gant player team hirschbeck basebal won hi lost ball defens base run win 5: bike dog lock ride don wave drive black articl write apr motorcycl ca turn dod insur 7. CONCLUSIONS AND FUTURE WORK In this paper, we formulate a class of clustering problems as bipartite graph partitioning problems, and we show that efficient optimal solutions can be found by computing the partial singular value decomposition of some scaled edge weight matrices. However, we have also shown that there still remain many challenging problems. One area that needs further investigation is the selection of cut points and number of clusters using multiple left and right singular vec- tors, and the possibility of adding local refinements to improve clustering quality. 6 Another area is to find efficient algorithms for handling overlapping clusters. Finally, the treatment of missing data under our bipartite graph model especially when we apply our spectral clustering methods to the problem of data analysis of recommender systems also deserves further investigation. 8. ACKNOWLEDGMENTS 6 It will be difficult to use local refinement for PDDP because it does not have a global objective function for minimization. The work of Hongyuan Zha and Xiaofeng He was supported in part by NSF grant CCR-9901986. The work of Xiaofeng He, Chris Ding and Horst Simon was supported in part by Department of Energy through an LBL LDRD fund. 9. --R Finding Out About: A Cognitive Perspective on Search Engine Technology and the WWW. Correspondence analysis handbook. Principal Direction Divisive Partitioning. Spectral Graph Theory. An improved spectral bisection algorithm and its application to dynamic load balancing. Analysis. Algebraic connectivity of graphs. Matrix computations Second Edition Correspondence analysis in practice. Partitioning sparse rectangular and structurally nonsymmetric matrices for parallel computation. Probabilistic Latent Semantic Indexing. A toolkit for statistical language modeling Normalized cuts and image segmentation. Modern Applied Statistics with S-plus --TR An improved spectral bisection algorithm and its application to dynamic load balancing Matrix computations (3rd ed.) Probabilistic latent semantic indexing Document clustering using word clusters via the information bottleneck method Partitioning Rectangular and Structurally Unsymmetric Sparse Matrices for Parallel Processing Finding out about Principal Direction Divisive Partitioning Normalized Cuts and Image Segmentation --CTR Bhushan Mandhani , Sachindra Joshi , Krishna Kummamuru, A matrix density based algorithm to hierarchically co-cluster documents and words, Proceedings of the 12th international conference on World Wide Web, May 20-24, 2003, Budapest, Hungary Hongyuan Zha , Xiang Ji, Correlating multilingual documents via bipartite graph modeling, Proceedings of the 25th annual international ACM SIGIR conference on Research and development in information retrieval, August 11-15, 2002, Tampere, Finland Wei Xu , Yihong Gong, Document clustering by concept factorization, Proceedings of the 27th annual international ACM SIGIR conference on Research and development in information retrieval, July 25-29, 2004, Sheffield, United Kingdom S. H. Srinivasan, Features for unsupervised document classification, proceeding of the 6th conference on Natural language learning, p.1-7, August 31, 2002 Long , Zhongfei (Mark) Zhang , Xiaoyun W , Philip S. Yu, Spectral clustering for multi-type relational data, Proceedings of the 23rd international conference on Machine learning, p.585-592, June 25-29, 2006, Pittsburgh, Pennsylvania Long , Xiaoyun Wu , Zhongfei (Mark) Zhang , Philip S. Yu, Unsupervised learning on k-partite graphs, Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining, August 20-23, 2006, Philadelphia, PA, USA Chris Ding , Tao Li, Adaptive dimension reduction using discriminant analysis and K-means clustering, Proceedings of the 24th international conference on Machine learning, p.521-528, June 20-24, 2007, Corvalis, Oregon Hua Yan , Keke Chen , Ling Liu, Efficiently clustering transactional data with weighted coverage density, Proceedings of the 15th ACM international conference on Information and knowledge management, November 06-11, 2006, Arlington, Virginia, USA Zheng-Yu Niu , Dong-Hong Ji , Chew-Lim Tan, Document clustering based on cluster validation, Proceedings of the thirteenth ACM international conference on Information and knowledge management, November 08-13, 2004, Washington, D.C., USA Bin Gao , Tie-Yan Liu , Guang Feng , Tao Qin , Qian-Sheng Cheng , Wei-Ying Ma, Hierarchical Taxonomy Preparation for Text Categorization Using Consistent Bipartite Spectral Graph Copartitioning, IEEE Transactions on Knowledge and Data Engineering, v.17 n.9, p.1263-1273, September 2005 Bin Gao , Tie-Yan Liu , Xin Zheng , Qian-Sheng Cheng , Wei-Ying Ma, Consistent bipartite graph co-partitioning for star-structured high-order heterogeneous data co-clustering, Proceeding of the eleventh ACM SIGKDD international conference on Knowledge discovery in data mining, August 21-24, 2005, Chicago, Illinois, USA Gne Erkan, Language model-based document clustering using random walks, Proceedings of the main conference on Human Language Technology Conference of the North American Chapter of the Association of Computational Linguistics, p.479-486, June 04-09, 2006, New York, New York Zheng-Yu Niu , Dong-Hong Ji , Chew Lim Tan, Using cluster validation criterion to identify optimal feature subset and cluster number for document clustering, Information Processing and Management: an International Journal, v.43 n.3, p.730-739, May, 2007 Long , Zhongfei (Mark) Zhang , Xiaoyun Wu , Philip S. Yu, Relational clustering by symmetric convex coding, Proceedings of the 24th international conference on Machine learning, p.569-576, June 20-24, 2007, Corvalis, Oregon Navaratnasothie Selvakkumaran , George Karypis, Multi.Objective Hypergraph Partitioning Algorithms for Cut and Maximum Subdomain Degree Minimization, Proceedings of the IEEE/ACM international conference on Computer-aided design, p.726, November 09-13, Bin Gao , Tie-Yan Liu , Tao Qin , Xin Zheng , Qian-Sheng Cheng , Wei-Ying Ma, Web image clustering by consistent utilization of visual features and surrounding texts, Proceedings of the 13th annual ACM international conference on Multimedia, November 06-11, 2005, Hilton, Singapore Chris Ding , Tao Li , Wei Peng , Haesun Park, Orthogonal nonnegative matrix t-factorizations for clustering, Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining, August 20-23, 2006, Philadelphia, PA, USA Han , Lee Giles , Hongyuan Zha , Cheng Li , Kostas Tsioutsiouliklis, Two supervised learning approaches for name disambiguation in author citations, Proceedings of the 4th ACM/IEEE-CS joint conference on Digital libraries, June 07-11, 2004, Tuscon, AZ, USA Ying Zhao , George Karypis , Usama Fayyad, Hierarchical Clustering Algorithms for Document Datasets, Data Mining and Knowledge Discovery, v.10 n.2, p.141-168, March 2005 Ying Zhao , George Karypis, Empirical and Theoretical Comparisons of Selected Criterion Functions for Document Clustering, Machine Learning, v.55 n.3, p.311-331, June 2004 Han , Hongyuan Zha , C. Lee Giles, Name disambiguation in author citations using a K-way spectral clustering method, Proceedings of the 5th ACM/IEEE-CS joint conference on Digital libraries, June 07-11, 2005, Denver, CO, USA

singular value decomposition;graph partitioning;correspondence analysis;document clustering;spectral relaxation;bipartite graph

502609

Using navigation data to improve IR functions in the context of web search.

As part of the process of delivering content, devices like proxies and gateways log valuable information about the activities and navigation patterns of users on the Web. In this study, we consider how this navigation data can be used to improve Web search. A query posted to a search engine together with the set of pages accessed during a search task is known as a search session. We develop a mixture model for the observed set of search sessions, and propose variants of the classical EM algorithm for training. The model itself yields a type of navigation-based query clustering. By implicitly borrowing strength between related queries, the mixture formulation allows us to identify the "highly relevant" URLs for each query cluster. Next, we explore methods for incorporating existing labeled data (the Yahoo! directory, for example) to speed convergence and help resolve low-traffic clusters. Finally, the mixture formulation also provides for a simple, hierarchical display of search results based on the query clusters. The effectiveness of our approach is evaluated using proxy access logs for the outgoing Lucent proxy.

INTRODUCTION Searching for information on the Web can be tedious. Traditional search engines like Lycos and Google now routinely return tens of thousands of resources per query. Navigating these lists can be time consuming and frustrating. In this paper, we propose narrowing search results by observing the browsing patterns of users during search tasks. The data to support our work comes from devices in the network that log requests as they serve content. We will focus primarily on proxy access logs, but it is easy to see how these ideas will carry over to the kinds of data collected by an Internet Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. service provider (ISP). From these logs, we first extract the search path that a user follows. Then, we apply a statistical model to combine related searches to help provide guidance to people beginning searches on related topics. We capture the interesting part of the search path in a search session, which is a user's query together with the URLs of the Web pages they visit in response to their query. Implicit in our approach is a form of query clustering that combines similar search terms on the basis of the Web pages visited during a search session. These clusters are then used to improve the display of search engine results. In this paper, we illustrate the technique by creating a directory consisting of two levels; each query is related to one or more groups of URLs. When a user submits a query to a search engine, we display the most relevant URLs from the most relevant directory groups. Here "relevance" is based on the data gathered from previous searches. While clustering has been proposed previously in the IR literature, our use of passively collected data to build search sessions is new. With data of this kind we have access to orders of magnitude more searching activity than is possible with specialty search engines or opt-in systems. In addition to exploring this new source of search session data, we also propose techniques for leveraging existing (manually de- rived) content hierarchies or labeled URLs to improve the relevance of identified resources. Finally, to make this work practical in a real proxy, we must consider a number of new implementation issues, including online versions of our clustering algorithms. The balance of the paper is organized as follows. In Section 2 we discuss our search session extractor. In Section 3 we describe a particular mixture model for query cluster- ing, and illustrate some of the groups it finds. In Section 4 we examine the kind of improvement our recommendations represent over a standard search engine. Related work is presented in Section 5, and we conclude with a discussion of future work in Section 6. 2. SEARCH SESSION EXTRACTION To formalize the extraction process, we introduce our basic data element, the "search session." A search session is the collection of words a user submits to a search engine (also known as a "query string") together with the URLs of the Web pages they visit in response to their request. As users browse the Web, a proxy server will record every URL they access, including HTML and PDF documents, embedded GIF and JPEG images, and Java class files. Because search engines deal primarily in Web pages, we filter xxx.xxx.xxx.xxx 02/Dec/2000:01:48:55 "GET http://www.google.com/search?q=infocom+2001" xxx.xxx.xxx.xxx 02/Dec/2000:01:49:03 "GET http://www.ieee-infocom.org/2001/" xxx.xxx.xxx.xxx 02/Dec/2000:01:49:27 "GET http://www.ieee-infocom.org/2001/program.html" Figure 1: A subset of the fields available in the proxy server log corresponding to a search session. The fields are: the IP address of the client, a timestamp recording when the request was handled, the HTTP request method, and the URL. user requests by file type and keep only HTML documents. Therefore, from this point on we will use the terms "Web page," "HTML document" and "URL" synonymously. In Figure 1 we present three lines of a proxy log that represent a single search session. (The IP address of the user's computer has been masked.) The query "infocom 2001" is extracted from the URL in the first line of this figure using a set of manually-derived rules for www.google.com. We maintain a list of such rules for each of the most popular search engines, allowing us to automatically tag and parse log lines corresponding to search queries. Every time a user posts a query to one of the search engines in our list, we initiate a new search session. Once a new session has been initiated, we then examine the link structure of the URLs subsequently requested by the user. Those pages that can be traced back to the list of search results (either directly or indirectly through one or more links) are included in the search session. We determine that the session has ended if either the user makes another request to a search engine, or if an hour has passed since the last addition to the session. 1 In Figure 2 we list the search session extracted from the log fragment in Figure 1. We refer to a search session as completed if it contains at least one URL. In the remaining cases, we assume the user examined the search engine results and decided that none of the Web pages were relevant. Search session session id 1001 user id 873 query infocom+2001 1914.59305 urls www.ieee./2001/ 1914.59340 1 Figure 2: Sample search session. The fields include a session ID, user ID (based on the IP address of the client), the query and the time it was submitted (a truncated Julian date), and the requested URLs, again with timestamps. The last column records whether or not the URL was linked directly to the search results page (1 or 0, respectively). We have implemented a search session extractor that takes as its input the access log file from a proxy server. By using a historical log, however, we are forced to "replay" part of a user's actions to reconstruct the path they followed. 1 Unfortunately, this process is complicated by the fact that some users maintain several browser windows and launch multiple concurrent (but related) search sessions. The simplified algorithm presented here is sufficient for the purposes of the present paper, but the precise details of our extraction algorithm are given in [14]. This means we must re-retrieve pages previously visited and hence the extractor process is usually scheduled when the server is not busy handling primary requests. This overhead can be reduced by examining pages as they are being served by a proxy. The link structure as well as various other aspects of the requested pages can be extracted by a background process having access to the proxy's cache. We are currently building a system of this kind and examining how data collection impacts proxy performance. We consider data collected by a proxy server handling all of the outgoing HTTP requests made by a portion of the Lucent employees based in Murray Hill, New Jersey. In the log maintained by this proxy, we find an essentially complete record of the sites visited by a population of 2,000 researchers, developers and support staff. Between January and May of 2001, this server logged over 42 million requests. In Table 1 we list the ten most frequently accessed search engines by the Lucent employees during this period. 2 We also recorded 13,657 search sessions, of which 44% were com- pleted. Roughly 60% of the completed sessions consisted of only 1 or 2 URLs, with 20% having length 3 or 4. In gen- eral, the longer the search sessions the larger the percentage of pages that were not linked directly to the results page. For sessions of length 5 or more (comprising 20% of the total number of search sessions), 55% of the pages were not among the search results seen by the user. This number is for sessions of length 3 or 4, and only 12% for those with only 1 or 2 URLs. Table 1: A list of frequently accessed search engines compiled from the Lucent proxy logs for January 1, 2001 through May 23, 2001. % of queries Search engine 53.0 www.google.com search.yahoo.com 11.2 www.altavista.com 4.4 hotbot.lycos.com 4.2 www.lycos.com 2.7 search.excite.com 0.3 search.msn.com 0.3 www.northernlight.com In terms of time, of the queries that were exactly repeated in our session data set, about 50% were issued within a day of each other. These tend to be made by the same user, the only exceptions being queries related to paper deadlines for 2 It is worth noting that www.google.com is much more popular among this research community than one would expect from standard ratings offered by Media Metrix or NetRatings [15] which rank yahoo.com, msn.com and search.aol.com as the clear market leaders. Query: bridal+dresses bridesmaid+dress flower+girl+dresses URLs: www.priscillaofboston.com www.martasbridal.com www.bestbuybridal.com www.bestbuybridal.com weddingworld.net www.martasbridal.com weddingworld.net weddingworld.net/htm/. www.ldsweddings.com/bridal dress. www.usedweddingdresses.com Figure 3: Three search sessions initiated by three different users. The query strings are all related to wedding dresses of some kind, and the URLs visited by the users are similar. major computer science conferences and searches for online greeting card services near major holidays. On larger time scales, 80% of the repeated queries were posted by different users. So far, we have focused on proxy logs as the primary source of data for constructing search sessions. Many search engines collect abbreviated navigation data to help improve their service [6]. So-called "click-through" statistics record the pages that users select from the list of search results. By design, the redirection mechanism used to collect these data cannot capture any information about a user's activities beyond their selections from the search results list. Given the fact that long search sessions consist mainly of pages not returned by the search engine, click-through data does not have the same potential to uncover relevant pages. To see this, we applied the heuristic relevance measures introduced at the end of this paper, and found that the "desired pages" in over half of the long sessions were not linked directly to the list of search results examined by a user. Even if a page is included among the search results, click-through statistics cannot identify requests for the page that are referred by other URLs linked either directly or indirectly to the search results list, missing valuable information about the relevance of sources in the search engine database. For these reasons, we see our approach as potentially much more powerful than traditional schemes that rely on tracking click-throughs. 3. CLUSTERING QUERIES Preliminary data analysis on our collection of search sessions led to the simple observation that semantically related query terms often draw users to the same sets of URLs. In Figure 3 we present three search sessions (each initiated by different users) that all relate to wedding dresses of different kinds, and all produce requests for many of the same Web pages. In this section, we consider improving search results by first forming groups of queries based on the similarity of their associated search sessions. In turn, by combining search sessions with queries in a given group, we can better identify relevant URLs. In the IR literature, there are several examples in which the pattern of retrieved items is used to form so-called "query clusters," see [16, 17]. In our context, these schemes would involve the queries submitted by users together with the top L relevant pages returned by a given search engine. Our technique is different in that we consider only those pages that were actually selected by a user during a search task. This has the effect of reducing spurious associations between queries. In addition, we include pages that are not listed by the search engine at all, effectively making up for deficiencies in spidering. By including user choices, we have developed an improved search scheme that is similar in spirit to collaborative filtering. When this kind of approach has been discussed previously in the IR literature, it has typically required the user to report details of their search and manually tag pages according to their relevance, for example, [9]. This level of involvement is unrealistic in the context of Web searching, and hence the impact of these schemes is somewhat limited. We avoid such problems by basing our work on passively collected proxy logs. Our initial intuition for navigation-based query clustering came not from the IR literature, but rather from an often-cited motivation for the popular PageRank algorithm [5]. Broadly speaking, PageRank is based on the amount of time a "random surfer" would spend on each page. Random walks on the link structure of the Web are also discussed in [11]. We felt that actual user data would be a better measure of importance than this random walk idea. 3.1 The mixture model In this section, we consider models for both forming query groups as well as determining the most relevant URLs for each group. We construct a statistical mixture model to describe the search session data. This model has as its parameters the probability that a given query belongs to a particular group as well as a set of group-specific relevance weights assigned to collections of URLs. The algorithms we present all attempt to fit the same model to the data. Our first approach makes use of the standard EM (Expectation- Maximization) algorithm to find maximum likelihood estimates of the model parameters. At the end of the paper, we discuss various alternatives that work in real time and can be incorporated in a proxy implementation. To help guide the cluster process, we also introduce labeled data from an existing topic hierarchy that contains over 1.2 million Web sites. We present an ad hoc algorithm for dealing with labeled pages, and at the end of the paper discuss a more formal statistical approach that uses the content hierarchy to specify a prior distribution for the model's parameters. Let q denote queries and u URLs. Each query q i is associated with one or more groups, where the subscript i runs from 1 to the number of queries seen in the training data, I. For the moment, we assume that the number of groups, K, is fixed. The group relation is captured by the triple denotes a group ID and w ik is the probability that q i belongs to group k. Then, for each group, we identify a number of relevant URLs. This is described by the triple (k; is a URL and -kj is a weight that determines how likely it is that u j is associated with the queries belonging to group k. We let the index j range from 1 to the number of URLs seen in the training data, J (which might include URLs from a content hierarchy). An example of a query-group triple (q while the associated group-relevance triples (k; be As mentioned above, sets of such triples constitute the parameters in a statistical model for the search sessions. These triples can be used by a search engine to improve page rank- ings. When a user initiates a new search, we present them with a display of query groups most related to their search terms. For each such group, we select the most relevant URLs arranged in a display like that in Figure 4. Here, we arrange the query groups and URLs by weight, with the most relevant appearing at the top. In this example, we have used the data from a content hierarchy to name the separate query groups (e.g., Medications in Figure 4). At the end of the paper, we discuss model extensions that will purely automated group naming. Finally, we see our clustering as an addition to a standard page of search results. By presenting the user with a small, organized set of URLs from our system together with a spider-based search list, we allow new resources to be added to our system in a natural way. In fact, estimates of confidence can be used to suppress our display entirely, forcing the user to help train the system when we have insufficient navigation data. Our clustering can also be used to modify the rankings of results from a traditional search engine; to guarantee that new resources will be visible to the user, the modified rankings can be displayed only a fraction of the time. Figure 4: Example of cluster results for a query of 'hypertension'. A mixture model is employed to form both the query groups as well as the relevance weights. Assume that in our dataset we have I queries which we would like to assign to K groups, and, in turn, determine group-specific relevance weights for each of J URLs. For the moment, we simplify our data structure and let n ij denote the number of times the URL u j was selected by some user during a search session under the query q i . Let n the vector of counts associated with query q i . We model this vector as coming from a mixture of the form where the terms ff k sum to one and denote the proportion of the population coming from the kth component. Also associated with the kth component in the mixture is a vector of parameters From a sampling perspec- tive, one can consider the entire dataset as being generated in two steps: first, we pick one of the K groups k according to the probabilities ff k and then use the associated distribution p(\Deltaj- k ) to generate the vector n i . We now consider the specification of each component in the mixture (1). We assume that in the kth component the data come from a Poisson distribution with mean -kj , where the counts for each different URL u j are independent. Then, setting the likelihood of the kth component associated with a vector of counts n i is given by Y e \Gamma- kj To fit a model of this type, we introduce a set of unobserved (or missing) indicator variables fl ik , where group k, and zero otherwise. Then, the so-called complete data likelihood for both the set of counts n i and the indicator variables can be expressed as Y Y Y e \Gamma- kj We refer to the parameters -kj as relevance weights, and use the probability that fl as the kth group weight for query q i (the w ik mentioned at the beginning of this section). 3.2 Basic EM algorithm The Expectation-Maximization (EM) algorithm is a convenient statistical tool for finding maximum likelihood estimates of the parameters in a mixture model [8]. The EM algorithm alternates between two steps; an E-step in which we compute the expectation of the complete data log-likelihood conditional on the observed data and the current parameter estimates, and an M-step in which the parameters maximizing the expected log-likelihood from the E-step are found. In our context, the E-step consists of calculating the conditional expectation of the indicator variables fl ik , which we l - ff l where p(\Deltaj -k ) is given in (2). In this expression, the quantities -k denote our current parameter estimates. Note that - fl ik is an estimate of the probability that query q i belongs to group k, and will be taken to be our query group weights. Then, for the M-step, we substitute - (3), and maximize with respect to the parameters ff k and -k . In this case, a closed form expression is available, giving science:math:statistics:conferences www.beeri.org.il/srtl/ computers:computer science:conferences www.computer.org/conferen/conf.htm computers:computer science:conferences www.acm.org/sigmod/ computers:computer science:conferences www.acm.org/sigmm/Events/./sigmm conferences.html computers:computer science:conferences www.acm.org/sigkdd/events.html Figure 5: A subset of the fields available in the DMOZ data. The fields are directory label and URL. us the updates l Clearly, these simple updates make the EM algorithm a convenient tool for determining query group weights and relevance weights. Unfortunately, the convergence of this algorithm can be slow and it is guaranteed only to converge to a local maximum. To obtain a good solution, we start the EM process from several random initial conditions and take the best of the converged fits. In Figure 6 we present an example of three of the groups found by this standard EM approach. For each group we display those query terms with the highest weights. It is arguable that the last group is fairly loose, combining different countries in Africa. From the standpoint of searches typically conducted by the Lucent employees served by our proxy, however, this degree of resolution is not surprising. With proxy logs from an ISP, we would have sufficient sample size to further subdivide this topic area. 3.3 Approximate algorithm with labeled data The query groups formed by the mixture model introduced in Section 3.2 allow us to borrow strength from search sessions initiated with different, but semantically, related query strings. The mixture approach, however, is highly unstructured in the sense that we only incorporate user data to learn the groups and relevance weights. In this section, we consider a simple method for incorporating existing information about related URLs from, say, a directory like DMOZ (www.dmoz.com) or Yahoo!. Essentially, we use the directory labels obtained from these sources to seed our query groups. In Figure 5 we present a subset of the data available from the DMOZ hierarchy. We assume that the data in such a structure can be represented as a set of pairs (l; u lj ) where l indexes groups of URLs and u lj is a URL in the lth group (here j runs from 1 to J l , the number of URLs in group l). The weights associated with these data, - jl , are not specified in the directory so we assume they have a value of ff. In Figure 7, we present a simple algorithm to establish mappings between queries and nj+transit bridal+dresses kenya njtransit bridesmaid+dress tanzania nj+train+warren flower+girl+dresses africa new+jersey+train Figure Three query groups found by fitting the mixture model via the standard EM algorithm. In each case, only the top ranking queries per group are displayed. URLs when either the query or the URL has been seen in either the labeled data or the sessions that have already been processed. The remaining sessions are processed in a batch using the basic EM algorithm in Section 3.2. The algorithm can be tuned with a threshold value T (0 - T - 1) to force more of the URLs in the session to exist in a previously created group. The approximate algorithm has the advantage of incorporating labeled data, but has the disadvantage of slowly adding to the set of clusters when a new topic is found in the data. At the end of the paper, we describe a more formal statistical approach to using content hierarchies that avoids this problem. 4. EXPERIMENTAL RESULTS To evaluate our methods, we used Lucent proxy data and computed search session lengths with and without query clustering. We began by selecting a "desired URL" from each search session. Since the users did not provide relevance feedback on the pages they viewed, we developed a simple heuristic for making this choice. In our first set of experiments, we took the desired URL to be the last URL that the user visited before moving on to a new task. Since users may continue to browse after they have found a relevant page, this simple choice is not correct for all of the search sessions. In subsequent experiments, we defined the second-to-last URL to be the desired URL, and so on. We found that the results of these experiments were all very sim- ilar, each suggesting that clustering can considerably reduce session length. Therefore, in this paper, we present only the experiments with the the last URL being the desired URL. We consider two metrics to judge the effectiveness of our algorithms: the percent of queries where the desired URL is in a cluster; and the position of the desired URL in a group (i.e., the session length). Our claim is that the location of the desired URL in our system's output can be compared against the number of URLs that a user visited during their assign each URL in labeled data to a URL group based on its directory label with weight 0 for each session s largest overlap with URLs in s if (query in s was NOT seen before) and (# URLs in s existing in g/# URLs in s ! T) add s to B else for each URL in s if URL is not in g, add it increment weight of URL by ff if query is not in query groups associated with g, add it output mappings cluster sessions in B using the BasicEM algorithm Figure 7: Pseudo-code for approximate algorithm. number of URLs visited Figure 8: Experiment results: the number of URLs visited without and with clustering on 46 days of testing data. searching task to measure the improvement of our system. Averaging these values across search sessions provides us with a measure of expected search length. To study our algorithms, we used a portion of the search sessions extracted from the access logs of the Lucent Murray Hill proxies to train with, and then tested with another portion, distinct in time. Given our performance metrics, we conducted a number of experiments involving different time-based divisions into training and test sets and varying the parameters required for the basic and approximate EM algorithms. We then used our experience with these different trials to choose a reasonable set of parameters for the experiments (on new data) reported here. We used groups when training with the basic EM algorithm. We decided on 0:1, and the labeled data from the DMOZ image from May 19, 2001 when training with our approximate algorithm. The results presented below are representative of our many experiments. They correspond with training on 95 days worth of data, and testing on 46 days of data. The percent of queries where the desired URL is in a cluster is 93%. Thus, 93% of the time, when both the query and desired URL were seen in the training data, our algorithms display the desired URL. Figure 8 presents our position results; we see that both algorithms reduce the number of URLs visited by 43% for the basic EM algorithm (38% for the approximate algorithm). The means are slightly misleading due to the presence of a small number of outliers (queries with very large search sessions). Thus, we also computed median positions; without our clustering algorithms, the median number of URLs visited is 2, and with either clustering algorithm, the median position of the desired URL is 1. Perhaps most importantly, for 43% of the cases (29% for the approximate algorithm) we provide a strictly shorter search session length. Figure 9 presents our results with testing with six months of data. (The difference in the height of the "without" bars is due to different ranges of training data being used.) We found the value of the percentage of queries where the desired URL is in a cluster is dependent on the length of the number of URLs visited Figure 9: Experiment results: the number of URLs visited without and with clustering on 6 months of testing data. testing data. For example, when 6 months of testing data was used, the percent matched is 60-69%, and when only were used, the percent matched is 93%. In some sense, this effect is an obvious byproduct of our experimental refitting the model periodically, the set of URLs become out-of-date. In a real system, we would incorporate frequent model updates. At the end of this paper, we consider an online version of the EM algorithm that would provide incremental updates with each search session. While the results with the basic EM algorithm are en- couraging, it does not scale well at all. Even the modest data sizes entertained here can take many hours to con- verge. For the approximate algorithm, roughly 60% of the search sessions were passed into the basic EM. This allowed the approximate algorithm to run in much less time (1 hour compared to half a day). Since a smaller set of data is clustered in the approximate setup, a smaller value of K can be used. We note that the percentage of queries on the list for the approximate algorithm is only a few points less than the basic EM algorithm. Varying the value of K affects both the time needed to form the clusters and the quality of results. For example, when clustering 142 days of data, twice as long to form the clusters as times as long. 5. RELATED WORK There are many systems which involve users ranking or judging Web sites as they visit them; [9, 2] are examples. We feel that systems which require the users to explicitly comment on the Web sites place too burden on the user. Therefore, while the data can be considerably cleaner than our search sessions, it is necessarily of limited coverage. The database in [9], for example, contains very few of the queries seen in the Lucent logs. Many techniques exist for automatically determining the category of a document based on its content (e.g., [18] and its references, [10] and its references, [1]) and the in- and out-links of the document (e.g., [7, 12]). We are currently investigating techniques to include content in our clustering algorithms, with the advantage that by working with the proxy cache we do not require extra spidering. Another approach to document categorization is "content ignorant." [3], for example, uses click-through data to discover disjoint sets of similar queries and disjoint sets of similar URLs. Their algorithm represents each query and URL as a node in a graph and creates edges representing the user action of selecting a specified URL in response to a given query. Nodes are then merged in an iterative fashion until some termination condition is reached. While similar in spirit to our algorithms, this algorithm forces a hard clustering of queries, limiting its ability to easily incorporate prior information. In addition, our system relies on much richer data, namely proxy logs. Finally, the literature contains a large number of distance-based methods for clustering; [4, 19] present two well-known algorithms for handling large amounts of data. The approach taken by these types of algorithms might not work well on our problem where there should be tens of thousands of clusters. 6. CONCLUSIONS AND FUTURE WORK We have developed a method to extract search-related navigation information from proxy logs, and a mixture model which uses these data to perform clustering of queries. We have shown that this kind of clustering can improve the display of search engine results by placing the target URL high in the displayed list. The basic mixture approach using the basic EM algorithm is highly unstructured in the sense that we only incorporate user data to learn the groups and relevance weights. Having taken a probabilistic approach to grouping, we can easily incorporate prior information about related URLs from DMOZ in a more formal way. A natural prior for our coefficients - lj (the relevance weights) is a Gamma distribution. Suppose that for each URL contained in category l, we assign - lj a prior Gamma distribution with parameters ff and while those URLs not listed under the directory in category l receive a Gamma distribution with parameters ff 0 and fi, where ff ff. The ingredients necessary for the EM algorithm in Section 3 can be carried out under this simple model as well. In this way, we can force a stronger tendency toward maintaining the existing hierarchy topics, while still allowing new URLs to be added. The approximate algorithm presented in the paper can be viewed as a very rough approximation to this approach. Next, while the standard EM algorithm is sufficient for the relatively small data sets presented in this paper, it is known not to scale very well as either the number of queries or the number of query clusters increases. To combat this, we have started working with on-line versions of the EM algorithm [13] that can process individual search sessions as they arrive. This requires a reformulation of the model as presented here, but we believe it will give reasonable perfor- mance. We are also extending our model to treat the query string as a collection of query terms and not simply a sorted list as we have done here. The same kind of Poisson structure used for the collection of URLs in a search session is applied to the query terms. This allows us to be much more flexible in how we form query clusters and how we treat new queries. Finally, by extending our probability models to include content of the pages beyond the link structures, we hope to generate even greater improvements in cluster- ing. In a proxy-based implementation, processing the pages to extract content represents very little overhead as we are already examining pages for links. Another direction that we are exploring avoids the aggregation across users that we described in Section 3. In this case, a model for user's searching habits will replace the simplifying assumption that URLs associated with a query group are sampled independently. Directly describing how a user moves within and between sites will improve our calculation of relevance weights. In addition, user-level parameters can be introduced that capture the time scale over which a user is likely to be refining a query and not changing topics. Acknowledgments . Tom Limoncelli and Tommy Reingold were invaluable in helping us access the proxy server logs that we used for our experiments. 7. --R Theseus: categorization by context. Agglomerative clustering of a search engine query log. Scaling clustering algorithms to large databases. The anatomy of a large-scale hypertextual web search engine User popularity ranked search engines. Finding related web pages in the World Wide Web. Maximum likelihood for incomplete data via the EM algorithm (with discussion). Capturing human intelligence in the Net. Information retrieval on the Web. The stochastic approach for link-structure analysis (SALSA) and the TKC effect Clustering hypertext with applications to web searching. Mining Web proxy logs: a user model of searching. Nielsen//netratings search engine ratings Learning collection fusion strategies. Multiple search engines in database merging. BIRCH: an efficient data clustering method for very large databases. --TR Learning collection fusion strategies Fab Multiple search engines in database merging The anatomy of a large-scale hypertextual Web search engine A re-examination of text categorization methods Finding related pages in the World Wide Web Clustering hypertext with applications to web searching Capturing human intelligence in the net The stochastic approach for link-structure analysis (SALSA) and the TKC effect Agglomerative clustering of a search engine query log Information retrieval on the web --CTR Bernard J. Jansen , Amanda Spink , Chris Blakely , Sherry Koshman, Defining a session on Web search engines: Research Articles, Journal of the American Society for Information Science and Technology, v.58 n.6, p.862-871, April 2007 Mathias Gry , Hatem Haddad, Evaluation of web usage mining approaches for user's next request prediction, Proceedings of the 5th ACM international workshop on Web information and data management, November 07-08, 2003, New Orleans, Louisiana, USA Mark Truran , James Goulding , Helen Ashman, Co-active intelligence for image retrieval, Proceedings of the 13th annual ACM international conference on Multimedia, November 06-11, 2005, Hilton, Singapore Bernard J. Jansen , Amanda Spink, How are we searching the world wide web?: a comparison of nine search engine transaction logs, Information Processing and Management: an International Journal, v.42 n.1, p.248-263, January 2006

web searching;proxy access logs;model-based clustering;query clustering;expectation-maximization algorithm

502661

Scaling replica maintenance in intermittently synchronized mobile databases.

To avoid the high cost of continuous connectivity, a class of mobile applications employs replicas of shared data that are periodically updated. Updates to these replicas are typically performed on a client-by-client basis--that is, the server individually computes and transmits updates to each client--limiting scalability. By basing updates on replica groups (instead of clients), however, update generation complexity is no longer bound by client population size. Clients then download updates of pertinent groups. Proper group design reduces redundancies in server processing, disk usage and bandwidth usage, and dimininishes the tie between the complexity of updating replicas and the size of the client population. In this paper, we expand on previous work done on group design, include a detailed I/O cost model for update generation, and propose a heuristic-based greedy algorithm for group computation. Experimental results with an adapted commercial replication system demonstrate a significant increase in overall scalability over the client-centric approach.

INTRODUCTION Intermittently Synchronized Database (ISDB) systems allow mobile data sharing applications to reduce cost by forgoing continuous connectivity. To allow sharing, a dedicated update server maintains the primary copy (global database) of all data held by mobile clients while each client maintains its own replica of some subset of the global database schema. The update server maintains the primary copy of the global database by receiving updates from the clients and distributing them on demand to the clients based on knowledge of their subscriptions. (See Figure 1.) Typically, the server generates update les for clients on an individual basis: the server scans through a set of updates, and for each client and update, decides if the client should receive the update. In this way, each client receives a customized update le containing only the relevant data. This technique is simple and straightforward but requires the server to do work in direct proportion to the number of clients, limiting the scalability of the system and resulting in greater time to maintain local replicas. In this paper, we build on previous work that proposed organizing updates into groups shared by clients [5]. Using this approach, the server manages update processing for a limited and controllable number of groups, irrespective of the number of clients. Instead of receiving updates customized for its specic needs, each client accesses the updates for the groups containing data it needs. The group-based approach results in better server scalability because the work of the server is decoupled from the actual number of clients maintained [5, 11]. For example, a salesperson may enter her o-ce in the morning before visiting customers and load sales data from the server onto her laptop. Throughout the day, the salesperson updates her local copy of the sales data regardless of the connectivity with the server. When conditions permit it, a reliable connection can be established with the server to synchronize local data using updates generated for the proper groups. Connectivity options include conventional wireless modems, high speed wireless LANs (e.g., or docking the mobile device wired base station. Furthermore, her coworkers may concurrently update their replicated data. In short, we develop a detailed cost model for group design and oer a remedy to a scalability problem natural to ISDBs that share data. We then present extensive experimental results demonstrating the improved e-ciency using our design techniques. We begin in Section 2 by placing our work in the context of related work. We dene the architecture and an I/O-based cost model for evaluating a spe- cic data grouping in Section 3. In Section 4, we propose a heuristic set of operators for modifying data groupings and a greedy algorithm to apply these operators. We also outline how to handle changes in the system conguration. In Section 5, we compare our approach to intuitive grouping alternatives, and demonstrate that our grouping algorithm provides signicantly greater scalability in client population size. Finally, we summarize our observations and describe future work in Section 6. Note that for the sake of brevity, we do not include the proofs of theorems in this paper, and present only a subset of our experimental results. For more detail, the reader may refer to [12]. 2. RELATED WORK An ISDB is an instance of a distributed computing sys- tem. Multiple databases independently operate on shared data. The assumptions that clients are mobile and commonly suer long periods of disconnection from the server, however, make traditional concurrency control protocols in- applicable. Traditional distributed database systems use the two-phase-commit protocol to help ensure the ACID properties of transactions [7]. This protocol is communication intensive and is therefore impractical when clients can be unreachable for long periods of time{e.g., if a mobile client powers down in order to save energy. In response, researchers have proposed replicating data among multiple clients and allowing them to operate on the replicas independently [1]. This allows quicker response time, reduces the possibility of deadlock, reduces the need for energy-consuming communication with the server, and allows mobility resistant to network outages, with the down-side of relaxing the ACID properties [3]. To aid such functionality, the architecture must include a dedicated centralized server that collects updates, and resolves con icts as described in [2]. This server increases the availability and reliability of shared data, but may suffer performance problems because the amount of work the server must do increases with the number of clients served. The architecture and goal of the CODA intermittently connected le system is similar to those of ISDBs. Researchers of CODA predicted (but did not experience) the possibility of a reintegration storm, which occurs when multiple clients simultaneously try to synchronize their local caches with the server [9]. This results in unmanageably long reintegration times for each client. A similar project called DBMate supports intermittently connected database systems. Experiments with DBMate show that server design can improve synchronization performance [8]. The work in this paper generalizes these results. Group design is similar to materialized view design for large database systems. Both try to reorganize monolithic sets of data in order to speed up response time for clients. However, view design has slightly dierent goals and as- sumptions. The utility of view design is measured by how closely the resultant views cover the expected queries; the main cost is the disk space consumed in storing the views [4]. In group design for ISDBs, the utility of groups is measured by how quickly they are generated, and the cost is roughly how much extra data must be transmitted to the clients. Furthermore, since views are assumed to contain subsets of relational data and groups contain updates, they can be manipulated in dierent ways. For these reasons, algorithms for view design are inapplicable to group design. 3. MODEL Standard ISDB architecture includes the database server, update server, le server, network, and clients. The database server stores the global database. The update server generates sets of updates for the replicas stored on the clients. These updates are made available to clients through a set of le servers. Clients intermittently connect to the le server via a network. The client is composed of a client update agent, and a local DBMS. When the client downloads up- dates, the client update agent processes them by applying them to the replicas contained on the local DBMS. Data ow from client to server is not discussed in this paper, but the interested reader may refer to [10] for more details. The global database is divided into publications or frag- ments, which are horizontal partitions of the database. Each client subscribes to a subset of fragments based on its data needs. The server maintains multiple, possibly overlapping datagroups of fragments, designed based on client subscrip- tions. The DBMS records updates from clients in a log, which is modied so that it also stores the scope [3]{the which the update aects{of each update. An update session is dened as the process during which the update server scans the log for outstanding updates, and for each datagroup, generates an update le (called the update log) containing the updates that are in that data- group's scope. The resultant update logs are then placed on a le server for use by the clients. The frequency of up-date sessions is application dependent and controlled by the system administrator. A client downloads the update log(s) that correspond to its local subscription. The client's up-date server then processes and applies the contents of the update log(s) to its local fragment replicas. In the basic approach to update processing practiced by industry, which we call client-centric, one datagroup is dened for each client to exactly match its local subscription. In the proposed data-centric approach [5], datagroups are created according to how data are shared and the number of datagroups is generally independent of the number of clients. We consider the following two steps critical in synchronizing a client: generating update les at the server and transmitting them to the clients. We do not emphasize client-side processing because the availability of powerful mobile computing platforms (e.g., Pentium-class laptops) means that client-side processing is not a performance bottleneck, especially when the client is able to \install" updates while being disconnected. These two critical steps are tightly coupled. If an up-date server can generate update logs more quickly, then they are available for download sooner, and if update logs take more time to generate, a client must wait longer to down-load them. It is therefore reasonable to increase the cost of one of these steps if the decrease in the other is greater. Problem Statement: We are given a global database, divided into a set of fragments, g. The proportion of updates applied to F i during an update session is estimated by a weight, W i . Fragment weights can be determined by either using database statistics or as a function of the volume of data a fragment denition spans (i.e., the fragment's size) and the number of clients that subscribe to it (i.e., the fragment's subscription level): subscription level) of fragment i (size subscription level) of all fragments The server stores a set of datagroups generates an update log for each, containing the updates to the group's fragments. The size of each update log is a function of the sum of the weights of its corresponding datagroup. Each client, i, has a set of interests, C i where C i F . Each client subscribes to a set of datagroups such that the groups to which the client subscribe contain a superset of the fragments of interest to the client. Stated formally, if C is the set of all C i , given a mapping from clients to datagroups, (the power set of G), for each client i, [(i) . This is the covering constraint. Our goal is to generate a set of datagroups G and a mapping that minimizes the total cost function (see Equation below). With client-centric grouping, jCj. The data-centric grouping approach determines its grouping based on the aggregated interests of the entire client population and the capabilities of the system architecture. Figure 2 gives a simple example of data-centric redesign in the case where the interests of one client are a subset of those of another. Intuitively, if the number of fragments in the database is xed, as the number of clients increases, the absolute amount of overlap in interests increases. This suggests that data-centric redesign increases in usefulness with a growing client population. 3.1 Cost Model Our cost model assumes that I/O time is the dominant cost factor and is therefore a good approximation of the over-all update processing cost of a particular grouping scheme. Three server and network activities make up the total cost: update mapping (mapping of updates to their respective datagroups 1 ), update log storage (storage of all update logs An update operation is one of three data manipulating onto disk), and update log propagation (loading and transmission of update logs). Total cost is therefore: Total Cost = Update Mapping Cost Update Storage Cost Update Propagation Cost (2) The components of Equation 2 are explicitly modeled below. The variables we use are shown in the following table Variable Description (units, if appropriate) CS server disk seek time (secs) CL server disk latency time (secs) CD CS +CL CT server disk transmission rate (secs/byte) rate between client k and server VB buer size for each update log le (bytes) VD average update operation record size (bytes) VP average fragment denition record size (bytes) VT average temporary table record size (bytes) VG average datagroup denition record size (bytes) VS number of operations in the update le total size of update le M number of datagroups N number of clients Update (to datagroup) mapping cost is the time required to map updates to datagroups. We assume that the log of updates to be distributed and an update-to-fragment mapping table (the publication information) are sequentially read into memory (2CD +CT (VF )). The results of the join are saved into a temporary le (CD The temporary le and the fragment-to-datagroup mapping table (the subscription information) are then read (2CD joined in memory to produce the nal result. Hence, Update Mapping Cost Update storage (to disk I/O) cost measures the time required to store all the update logs onto disk at the server. We assume that a main-memory buer is maintained for each update log, and whenever a buer is lled, its contents are written to disk. This happens until all update logs are The left term in the parentheses below indicates the time spent on disk latencies that are experienced for each update log when the buer is lled, whereas the right term indicates how much time is required to store the actual data on disk. Recall that the weight of fragment estimates the proportion of operations that commands, namely INSERT, UPDATE and DELETE. Combinations of these operations constitute the transactions contained in update logs. 2 The term \record" in the table refers to the data structure (e.g., row) containing the respective information Salesman Northeast Manager Salesman Northeast Manager Data Data Data Data Data Data Data Data North East South West North East South West Database Server F Definition G Mapping F Client Population CLIENT-CENTRIC GROUPING DATA-CENTRIC GROUPING REDESIGN Figure 2: Example of client-centric to data-centric redesign using aggregated interests. are applied to that fragment. Update Storage Cost (Server to client) update propagation cost measures the time required to load into memory and then transmit the appropriate update logs to all clients, assuming unicast communication between the server and each client. Each log the client downloads is sequentially read, then joined in memory, then transmitted over the network at the client's bandwidth. Update Propagation Cost In the equation above, disk latencies are experienced while reading each client k's update logs (CD j(k)j). For each client k, the volume of data read and transmitted are the same (VF although the server's disk rate is xed (CT ), the clients' transmission rates are independent 3.2 Some Potential Cost Reducing Alternative For illustrative purposes, we now discuss certain extreme datagroup design strategies. A single large datagroup the amount of disk resources used. However, all clients must then receive all updates regardless of their particular interests. If transmission costs are high, this is a very costly solution. At the other ex- treme, one datagroup can be generated for each fragment This is similar to the server-side organization proposed in [8]. Note that if the number of fragments is high, then managing all the datagroups becomes costly. Another solution is to generate one datagroup per client C). This is the client-centric solution standard used by industry that, as shown in [5], results in high levels of redundant work as the client population grows. The alternative we propose lies somewhere in between these solutions and is based on how the clients share data. Another option is to ameliorate the system architecture itself with multiple update servers, in which each can work in parallel to generate the update logs for a subset of clients, as described in [10]. Instead of precluding it, such architectural solutions complement data-centric design. For exam- ple, clients can be allocated to each update server based on the potential eectiveness of the resultant data-centric grouping. One way to do this is to cluster clients based on interest a-nities as in [6]. However, because we assume a single update server, such work is outside the scope of this paper. 4. A HEURISTIC APPROACH TO GROUPING Because of the complexity of the grouping problem (it can be modeled as an NP complete mathematical programming problem), we oer a heuristic algorithm. We start by introducing three operators that perform grouping operations on fragments. Each has dierent cost proles, and thus are applicable in dierent situations. At the end of this section, we introduce a means of greedily applying these operators. For brevity, we do not include a formal cost analysis but one can be found in [12]. Based on the cost equations introduced in Section 3.1, certain conclusions can be drawn about ways of manipulating the grouping in order to reduce update processing costs. Namely, we can manipulate the number of datagroups, the composition of datagroups, and the subscription of clients to datagroups in order to change update processing costs. Common ways of redesigning datagroups, although with different side eects, include merging, splitting, and subtracting overlapping datagroups [7]. Although similar operators are used in materialized view design, these operators are based on algorithms that, as we explained in Section 2, are generally inapplicable here. We therefore dene our own operators and give their denitions, side-eects, and applicability below. See Figure 3. 4.1 Operators for Redesigning Datagroups Merging two datagroups involves replacing them with their union. Clients subscribing to at least one of the merged datagroups instead subscribe to their union, preserving the covering constraint (See Section 3). If there is overlap between the merged datagroups, then storage cost is reduced in proportion to the size of the overlap. If a client originally subscribed to only one of the merged datagroups, the client must receive super uous updates for fragments contained in the \other" merged datagroup, resulting in increased update log transmission costs. Applicability of Merge - Because this operator typically increases the amount of data that must be transmitted to a client, we do not expect this operator to be used much, unless the network bandwidth is very high, or the degree of overlap between two datagroups is very high. Splitting involves nding two non-totally overlapping, but partially intersecting datagroups, and splitting to form a third datagroup. Subscribers to either datagroup must also subscribe to the third datagroup. Splitting reduces overlap in datagroups, but does not increase the amount of data transmitted to the respective subscribers. There is, however, overhead in terms of disk seeks and latencies for each additional datagroup generated. Applicability of Split - Splitting should increase in relevance as the volume of updates increases or the degree of overlap between two datagroups increases, because the time saved in not rewriting large sets of updates osets the increase in disk seek and scan times. Subtracting two datagroups applies only if one is a sub-set (either proper or not) of the other. The smaller of the two is subtracted from the larger one, eliminating their over- lap. If the datagroup subtracted from becomes empty (in the case where the subset relationship is not proper), then it is discarded. Subscribers to the larger datagroup must also subscribe to the smaller one (if a smaller one exists); overhead in terms of disk latencies is incurred by clients having to subscribe to additional datagroups. The number of datagroups is not increased by this operation, and savings are proportional to the degree of overlap between the subtracted datagroups. Subscribers to these datagroups do not need to receive extra data. Applicability of Subtract - This operator diers from the other two, because of the subset restriction on the operands. Nonetheless, this operator typically has fewer side eects: It neither increases the amount of data sent to clients the way merge does, nor the number of datagroups the way split does. Example: Consider two datagroups, A and B, that each generate update logs of 10KB. If, due to the denitions of A and B, these update logs always have a single byte dif- ference, then it may make sense to merge A and B. This saves the server some work by maintaining one fewer data- group and not storing its corresponding update log. The cost however, is that clients subscribing to one of A or B, must download an additional byte. But at 19Kbps, the additional byte adds less than a millisecond in transfer time. Alternatively, splitting A and B forces the server to maintain the denition of an additional datagroup and generate an additional byte-sized update log, which, depending on the server load, may not be cost eective. On the other hand, if A and B both generate update logs of 10KB, which only have 5KB in common, it may make sense to split the datagroups. Although splitting in the above case would force the server to maintain the denition of an additional datagroup and gererate an additional byte- sized update log, here it would save the server some storage time (because of the magnitude of the overlap) without increasing the volume of data sent to the client. The drawback is that splitting forces the server to maintain an additional datagroup. Savings in storage, however, increase in importance as datagroups grow. If these two datagroups were merged instead, then clients subscribing to one of A or B would have to spend 2 seconds downloading super uous data in addition to the 4 seconds spent downloading pertinent data at 19Kbps. 4.2 Greedy Heuristic The application strategy proposed here pursues local cost minima until no cost reduction can be made by applying the above operators. The greedy algorithm starts with the client-centric solution and applies all possible subtraction operations on the set of datagroups until no cost-reduction is possible with this operator. We then search for the most cost-eective merge or split. If one exists, we perform it, and repeat the cycle: GREEDY HEURISTIC while TRUE do Perform all possible cost-reducing subtraction operations Let m be the most benecial merge and let s be the most benecial split. If either m or s results in a cost reduction, then perform the one that reduces cost the most. neither m nor s is performed, then quit. od Subtraction is given precedence because it typically has the fewest side-eects in terms of cost penalties and reduces the search space for the other two operators. Furthermore, merge and split can increase the applicability of subtract. 4.3 Redesign As time passes, an ISDB changes its conguration. New clients may be periodically added to the ISDB, with a given subscription. The problem is deciding the proper data- groups to assign to them. This problem is similar to the NP - hard set-covering problem. Moreover, the existing clients may change their subscriptions or their type of connection with the server may change. A client may move from one location to another and change its subscription to match its locale, or another client may acquire a faster, higher band-width connection. The problem of redesign therefore has two levels: redesign only subscriptions (groups remain xed); redesign both the groups as well as the subscriptions. We address these problems heuristically. Such solutions are necessary because the greedy heuristic described in this paper has complexity We therefore oer techniques that reduce the need of running it. These techniques and relevant proofs are fully described in [12]. 4.3.1 Addition of Clients In this section, we roughly describe how to map data- groups to the subscriptions of new clients. This problem is similar to the NP -complete weighted set-covering problem, A A A' A A' merge split subtract Figure 3: The Merge, Split, and Subtract Operators. The boxes represent datagroups. The dashed lines indicate overlap. and we solve the client-addition problem in a similar way. To each subscription, we greedily map the datagroup that has the best (lowest) cost-eectiveness. Cost-eectiveness is the ratio of the size of the datagroup (size is dened in Section and the total size of the fragments covered for the rst time by the datagroup. This process is repeated until the entire subscription is covered. We have a ratio bound for the results of this algorithm: Theorem 4.1. The solution achieved by the greedy algorithm is within a factor of HN approximation of the minimal cost cover, where is the N th harmonic number and N is the size of the subscription. 4.3.2 Rerunning the Redesign Algorithm As time passes, the conguration of the ISDB changes making a data-centric grouping less cost-eective. Examples of changes include changing of client connectivity, client subscriptions, or server speed. Since we can keep track of these changes, recomputing current client-centric cost (Ccc ) is straight-forward using the cost model from Section 3. By comparing this cost with the actual current cost of the current data-centric grouping (Cdc ), an administrator can decide to redesign the datagroups if the cost improvement o- sets the estimated redesign time (Cr This is a conservative rule of thumb for deciding the benet of redesign and varies depending on the application. 5. EXPERIMENTS 5.1 Goals The goal of our experiments is to show that data-centric grouping of updates using our greedy heuristic (denoted by dc), results in faster refresh times for the average client than other, more intuitive methods. Refresh time includes the time required for the server to generate an update log(s) (including computation and storage) and transmit it to the proper client. These costs correspond to the ones described in Section 3. The other grouping methods we consider are described in Table 1. In our experimental design, the database is composed of 100 fragments. Each fragment i is assigned a value, p i (0 Each client subscribes to fragment i with a probability p i . To assign values to each p i , we consider two probability distributions: one highly skewed (Zipan, with one uniform. This results in the average client subscribing to 1% of the database. The total volume of updates is a linear function of the number of clients. They are allocated to each fragment in proportion to that fragment's p i value. Each client is assumed to have the same network bandwidth (C i The client population and client bandwidth. Varying these parameters over these values re ects a wide range of appli- cations, such as the OLTP-class mobile o-ce application described in the beginning of this paper. We omit experiments varying update volume for the sake of brevity. See Table 2. We show that as any of the experimental parameters in- creases, the advantage of using dc increases over the other techniques. For example, the dierence between dc and cc increases with the number of clients. Performance gains generally come from detecting and removing redundant update processing (i.e., overlapping subscriptions), but as we show, under some circumstances, can also come by ooading work onto other components in the architecture. 5.2 Experimental Set-Up We ran our experiment on an Ethernet LAN consisting of Pentium II computers running Windows NT. Update services are provided by Synchrologic iMobile on a 266MHz PC with 64MB RAM. (Performance trends using alternative ISDB middleware, such as those by Oracle, IBM or Sybase, should be similar to those reported here.) The database server is Sybase SQL Anywhere v.6, running on a 200MHz PC with 128MB RAM. The database stores a universal relation to which we apply updates that are distributed to the clients. For the data-centric experiments, extensions based on those described in [5] have been incorporated into iMobile. One of the consequences of these extensions is that an extra set of meta-data must be sent to each client. The size of the meta-data le has been empirically estimated as ( 24 where jGj is the number of datagroups generated. The meta-data grows with the number of groups because increasing the number of groups results in the need to store more mapping information for them. This extra metadata increases the refresh times of dc, og and op generally by adding transmission time. This extra cost becomes insignicant however, as workloads increase and does not aect the trends of the results. 5.3 Experiment 1: Varying Client Population The results for both data distributions are nearly identi- cal. Method op is bad with low populations because it generates update logs regardless of whether they are subscribed to or not. But, as the population increases, the probability that a datagroup is not subscribed to becomes low. Method og makes sense when there are few clients, because it saves Grouping Methods (notation) comments data-centric (dc) groups generated with the greedy heuristic client-centric (cc) employed in industry, creates a unique group for each client one-giant-group (og) minimizes costs at the server by storing updates for all clients in a single group one-per-fragment (op) minimizes network costs and eliminates some storage redundancy by generating a single group for each fragment Table 1: Grouping Methods Parameter Description Values (control values in fg) Number of clients 1, 5, 50, f100g, 200, 500, 1000 Workload (updates/client) f50g Client bandwidth (bps) 19:2 Distribution of updates over fragments Zipf Table 2: Parameter values for experiments.20601001401 5 50 100 200 500 1000 clients refresh time og op cc clients refresh time og op cc dc A. Uniform Distribution B. Zipf Distribution Figure 4: Per-Client Refresh Time (seconds) With Varying Client Population.51525354519200 57600 512000 1024000 10240000 bandwidth (bps) refresh time og op cc bandwidth (bps) refresh time og op cc dc A. Uniform Distribution B. Zipf Distribution Figure 5: Per-Client Refresh Time (seconds) With Varying Client Bandwidth. the server some work. The amount of super uous data sent to each client grows, however, with the population, making this grouping infeasible. Method cc breaks down with a high population because of the increasing amount of redundant work it must do with a growing population. In these tests, dc, which uses the proposed greedy heuris- tic, has consistently good performance, resulting in the lowest or nearly the lowest refresh times over all populations. Method dc outperforms op, because dc groups together fragments that are often subscribed to together. Method dc avoids the problems associated with og and cc by generating many datagroups with little overlap, but not so many as to compromise server performance. See Figure 4. Please note that the total volume of data distributed is not scaled up per client as the client population increases. We keep the total volume of data constant in order to isolate client-population eects. However, experimental results given a greater total volume of data make the dc results even more favorable with respect to the others. 5.4 Experiment 2: Varying Client Bandwidth In practice, clients may choose among many connectivity options, including wireless modem, conventional modem, high-speed wireless LAN, or simply docking the portable device at the o-ce LAN. We therefore study how these changing bandwidths aect the eectiveness of the various grouping methods. Although og performs poorly when there is little band- width, it gains the most as bandwidth increases. With og, work at the server is already minimized, so more bandwidth reduces the most harmful eects of shipping super uous data. Method op has good performance as well, but fails to take advantage of the increased bandwidth by generating fewer groups in order to save the server some work. It therefore ultimately performs worse than og. With high bandwidth, cc has the worst performance of all, because of the redundant work that must be done at the server, regardless of network performance. This is an important result, and indicates that, regardless of the capability of the network, per-client ISDB refresh processing using cc has a performance oor. Method dc does the best, by generating multiple disjoint groups which conserve network resources when they are poor, and generating fewer groups which conserve server resources when the network is fast. See Figure 5. 6. CONCLUSION In this paper, we dene a detailed model of ISDBs, describe how this model inherently leads to performance problems with client refresh, and oer a grouping solution. We propose redesigning updates into groups based on the aggregated interests of the entire client population (data-centric). This allows clients to share groups intended for many clients. To formulate a redesign technique we dene a detailed cost dene operators that manipulate costs based on the system conguration, and devise a way of applying them, based on greedy heuristics. We tested our greedy heuristic against the client-centric approach, and two other intuitive solutions: a monolithic group containing all updates and individuals groups for each fragment. Overall, our heuristic outperforms all the others in terms of refresh time because it greedily allocates resources where they are needed{either to the server or the network depending on the conguration of the ISDB. Fur- thermore, the relative benet of data-centric grouping increases with the client population, improving scalability. Our work on ISDBs is ongoing. For example, we limited client connectivity to unicast because this is currently the dominant form of communication in practice. We are currently exploring the use of multicast as a means of improving network scalability. 7. ACKNOWLEDGMENTS The authors would like to acknowledge Ms. Mireille Jacobson for her valuable editorial assistance. 8. --R Replication and consistency: Being lazy helps sometimes. Replicating and allocation data in a distributed database system for workstations. The dangers of replication and a solution. Implementing data cubes e-ciently Grouping techniques for update propagation in intermittently connected databases. Vertical partitioning algorithms for database design. Principles of Distributed Database Systems. Data partitioning for disconnected client server databases. Experience with disconnected operation in a mobile computing environment. A framework for server data fragment grouping to improve scalability in intermittently synchronized databases. Minimizing redundant work in lazily updated replicated databases. --TR Vertical partitioning algorithms for database design The dangers of replication and a solution Implementing data cubes efficiently Replication and consistency Principles of distributed database systems (2nd ed.) Data partitioning for disconnected client server databases Replicating and allocating data in a distributed database system for workstations A framework for designing update objects to improve server scalability in intermittently synchronized databases Grouping Techniques for Update Propagation in Intermittently Connected Databases --CTR Efficient synchronization for mobile XML data, Proceedings of the eleventh international conference on Information and knowledge management, November 04-09, 2002, McLean, Virginia, USA

distributed databases;mobile databases;intermittent synchronization

502784

Controllable morphing of compatible planar triangulations.

Two planar triangulations with a correspondence between the pair of vertex sets are compatible (isomorphic) if they are topologically equivalent. This work describes methods for morphing compatible planar triangulations with identical convex boundaries in a manner that guarantees compatibility throughout the morph. These methods are based on a fundamental representation of a planar triangulation as a matrix that unambiguously describes the triangulation. Morphing the triangulations corresponds to interpolations between these matrices.We show that this basic approach can be extended to obtain better control over the morph, resulting in valid morphs with various natural properties. Two schemes, which generate the linear trajectory morph if it is valid, or a morph with trajectories close to linear otherwise, are presented. An efficient method for verification of validity of the linear trajectory morph between two triangulations is proposed. We also demonstrate how to obtain a morph with a natural evolution of triangle areas and how to find a smooth morph through a given intermediate triangulation.

Figure 7 shows this on a concrete example. Conjecture 3.5 allows to choose the maximum m, namely that guarantees a valid morph. A simple algorithm to flnd M sequentially checks morphs for every m > 1, incrementing m by 2. The number of morphs checked may be signiflcantly reduced to ACM Transactions on Graphics, Vol. X, No. X, xxx 2001. Controllable Morphing of Compatible Planar Triangulations 11 (a) (b) (c) (d) (e) Fig. 7. Morphs generated by raising neighborhood matrices to various powers. (a), (b) The source and the target triangula- tions. Correspondence is color coded. (c) The convex combination morph at 0:5. (d) An invalid morph is generated by raising neighborhood matrices to power valid morph is generated by raising neighborhood matrices to power Zoom in on the triangulations in (d) and (e). (h) Trajectories of the convex combination morph. (i), (j) Trajectories of morphs generated by raising neighborhood matrices to power Note the positions of the trajectories relative to the straight lines (the linear morph). O(log M). First, we flnd an upper bound mmax by doubling m until the morph is invalid. Then, the resulting M is found by binary search in the interval [mm2ax ; mmax]. Furthermore, a morph that is even closer to the linear morph than a morph deflned by may be obtained. Consider the following deflnition of A(t): This equation averages the neighborhood matrices of the valid morph deflned by with the neighborhood matrices of the invalid morph, deflned by 2. Using power m for neighborhood matrices and a parameter d, equation (14) may be viewed as a morph with a non-integer power for neighborhood matrices m 1). In order to obtain a morph, which is the closest possible to the linear morph, the maximal parameter d may be chosen by binary search in the interval [0; 1], verifying the morph validity at every step. See Figure 11(b) for a morph generated by this scheme. 3.4 Morphing with an Intermediate Triangulation This section demonstrates how to flnd a morph between two triangulations T0 and T1 such that at a the morph interpolates a given triangulation Tm. The triangulations T0, T1 and Tm are compatible and with identical boundaries. A naive solution is to flnd two convex combination morphs independently: the flrst|between T0 and Tm, and the second morph|between Tm and T1. The problem with this is that while the two independent morphs are continuous and smooth, the combined morph will usually have a C1 discontinuity at the intermediate vertices. In order to flnd a smooth morph, it is necessary to smoothly interpolate A0, Am and A1 in A(T0). Consequently, the corresponding elements of the three matrices should be smoothly interpolated. Given three points (0; i;j(0)), (tm; i;j(tm)) and (1; i;j(1)) in R2, it is necessary to flnd an interpolation i;j(t) for all t 2 [0; 1], see Figure 8. Since the entries of the matrices are barycentric coordinates, the interpolation must satisfy 0 i;j(t) 1. An interpolation within the bounded region [0; 1] [0; 1] may be found as a piecewise Bezier curve, since any Bezier curve is located in the convex hull of its control points. An important point is that interpolations for the matrix entries are performed independently. But every row i, 1 i n, of A(t), being barycentric coordinates of the interior vertex i, should sum to unity. Due to the independent interpolations, this might not be the case. Normalizing the elements of each row can solve this problem. The normalized entry i;j(t) is deflned as follows: 12 V. Surazhsky and C. Gotsman i;j(t)PSfrag replacements i;j(tm) Fig. 8. An interpolation of three points (0; i;j(0)), (tm; i;j(tm)) and (1; i;j(1)) when 0 tm 1 is in the bounded region [0; 1] [0; 1]. Fig. 9. (top) A smooth morph interpolates an intermediate triangulation given at The morph trajectories. The dashed lines are the edges of the source, target and intermediate triangulations. Since i;j(t) is smooth, the sum of i;j(t)'s is also smooth. Therefore the normalized i;j(t) is a smooth interpolation. See an example demonstrating a smooth morph in Figure 9. 4. MORPHING WITH LOCAL CONTROL A well-behaved morphing scheme should have properties like those described in Section 3.2. Trajectories traveled by the interior vertices should be smooth and even (not jerky and not bumpy). It would be useful if the scheme would be linear-reducible. When the linear morph is invalid, the natural requirement is to generate a morph as close as possible to the linear one. It would also be useful to be able to control triangle areas in such a way that they transform naturally (uniformly) during the morph. This may help to prevent shrinking/swelling of triangles, that result in an unnatural-looking morph. The schemes presented in this section allow the control of trajectories of the interior vertices, triangle areas etc. in a local manner. To flnd a morph between two triangulations T0 and T1 means to flnd a curve A(t) for 0 t 1 in A(T0) with endpoints in A(T0) and A(T1). We will do this by constructing each row of A(t) (corresponding to each interior vertex) separately. We deflne T 0(G0; }0) to be a subtriangulation of a triangulation T (G; }) if T 0 is a valid triangulation, G0 is a subgraph of G and the coordinates of the corresponding vertices of T 0 and T are identical. The triangulations T0 and T1 may be decomposed into n subtriangulations in the following manner: each interior vertex corresponds to a subtriangulation that consists of the interior vertex, its neighbors and edges connecting these vertices, see Figure 10. A subtriangulation deflned above is said to Sn be a star denoted by Zi. Namely, every star Zi corresponds to the interior vertex i, and be stars of the triangulation T0; stars Zi(1) are deflned analogously for T1. Clearly, Zi(0) and Zi(1) are two isomorphic triangulations, since they are the same subgraph of two isomorphic triangulations T0 and T1. Barycentric coordinates of the interior vertex in star Zi with respect to the boundary vertices of that star are also barycentric coordinates of the interior vertex i Controllable Morphing of Compatible Planar Triangulations 13 Fig. 10. A triangulation is decomposed into stars; each interior vertex deflnes a separate star. in a triangulation T with respect to its neighbors. Thus all Zi(0)'s when 1 i n together deflne a neighborhood matrix A0 in A(T0); Zi(1) for 1 i n deflne A1 respectively. In the same manner, we would like to deflne A(t) for a speciflc t using stars Zi(t), 1 i n. The question is how to flnd Zi(t) for 0 < t < 1 such that it will deflne barycentric coordinates with some intermediate values between barycentric coordinates of Zi(0) and Zi(1). Obviously, a smoth morph of two stars Zi(0) and should su-ce to obtain this. But only a smooth A(t) will deflne a smooth morph between the triangulations. For that reason it is important to use a method to generate barycentric coordinates of the interior vertices that is at least C1-continuous, such as that described in Section 2.1. A morphing scheme that generates A(t) by morphing separately the stars of two triangulations is said to be a local scheme. There is a simple way to morph two stars Z(0) and Z(1). First, translate the two stars in such a way that the interior vertices of the both stars are at the origin. Then a morph may be deflned by linear interpolation of the polar coordinates of the corresponding boundary vertices. One can flnd a proof for the correctness of this morph in [Shapira and Rappoport 1995; Floater and Gotsman 1999; Surazhsky 1999], where there are also recommendations on how to choose the polar coordinates in order to obtain a valid morph. Note that the validity of star morphs, based on the translation of interior vertices to the origin, depends only on how angle components of the polar coordinates of the boundary vertices vary during the morphs. Arbitrary variations in the radial direction of the boundary vertices do not afiect the validity of the morph. 4.1 The Local Linear-Reducible Scheme The local scheme morphs separately the corresponding stars of T0 and T1 and is based on the translation of the source and target stars to the origin. Two translated stars Zi(0) and Zi(1) are morphed in the following manner. If the linear morph of two stars is valid, we adopt it. Otherwise, an arbitrary valid morph is taken. It can be the morph that averages the polar coordinates of the boundary vertices, as described in Section 4; or translated trajectories of the boundary vertices during the convex combination morph. In the latter case the row corresponding to the star Zi in A(t) is equal to the i'th row of A(t) generated by the convex combination morph. The following theorem shows that this scheme is linear-reducible. Theorem 4.1 The linear morph of two triangulations T0 and T1 is valid ifi the linear morphs of all component stars are valid. Proof. Validity: If the linear morph between T0 and T1 is valid, then any triangulation T (t) for compatible with T0 (and T1) and thus may be decomposed into valid stars. Each is a valid morph between Zi(0) and Zi(1). If all morphs of the stars Zi(t) are valid, then we have a legal neighborhood matrix function A(t) for 0 t 1, and thus T (t) is valid. Linearity: Let pi;j be the coordinates of the vertex i in the star Zj. First, we prove that if the morph between T0 and T1 is linear then the morphs of all stars are linear. We have for 0 t 1. Every star Zj(t) is translated in such a way that the interior vertex is at the origin. Thus, 14 V. Surazhsky and C. Gotsman Combining both equations: For the opposite direction, we prove that if the morphs of all stars are linear then the morph between T0 and T1 is linear. The linear morphs of the stars imply that: Let T (t) be the linear morph between T0 and T1, namely, It remains to show that A(t), deflned by the stars Zj(t) for 1 j n, satisfles will show that every star Zj(t) deflnes the same barycentric coordinates of the interior vertex i as the corresponding star j of T (t), and thus A(t) 2 A(T (t)). Clearly, Zj(t) is isomorphic with the corresponding star j of T (t). The following states that the vertex coordinates of Zj(t) are translated coordinates of the corresponding star j of T (t). Due to the initial translations of the interior vertices to the origin: We can now express (16) as: Hence, after the translation of pj(t) the coordinates of every vertex in the star Zj(t) are equal to the co-ordinates of the corresponding vertex in T (t). Since barycentric coordinates are invariant to a translation (as a special case of a-ne transformations), the stars Zj(t) for 1 j n deflne A(t) 2 A(T (t)). This work presents two linear-reducible schemes: the scheme described in Section 3.3 and the scheme introduced in this section. It is important to emphasize the principal difierence between these two schemes. The flrst scheme approaches the linear morph using neighborhood matrices raised to a power. This approaching signiflcantly afiects all trajectories of the interior vertices and is a global approach to the linear morph. It allows to choose a degree of approximation to the linear morph by specifying the power of the neighborhood matrices. However, even the morph closest to the linear morph does not allow each individual trajectory to be as 'linear' as possible. The global convergence may be blocked by a single problematic trajectory (which invalidates the morph), preventing the others from being straightened further, see Figures 11(a) and 11(b). On the other hand, the local linear-reducible scheme, morphing the component stars separately, may afiect a group of trajectories of adjacent vertices almost independently of other trajectories of the morph. However, the local scheme does not attempt to approximate the linear morph for stars for which the linear morph is invalid. Thus, vertices of triangulation regions that cannot be morphed linearly have trajectories similar to those generated by the convex combination morph, and vertices of regions that may be morphed linearly have trajectories very close to straight lines, see Figure 11(c). Knowing the properties of both the linear-reducible schemes, it is possible to choose the most suitable for speciflc triangulations and speciflc applications. These two schemes may also be combined to obtain a morph that is closer to the linear morph than a morph generated separately by each of the schemes. First, the scheme of Section 3.3 is used to generate a valid morph TP (t) with a maximal power for neighborhood matrices. Then the scheme of this section is applied, morphing each of the stars separately. For stars for which the linear morph is invalid, corresponding trajectories from TP (t) are used. For the rest of the stars, the linear morph is used. The resulting morph is valid, since the morphs of all component stars are valid. The morph in Figure 11(d) was generated using this combined scheme. Controllable Morphing of Compatible Planar Triangulations 15 (a) (b) (c) (d) Fig. 11. Trajectories of various morphs approaching the linear morph. The dashed lines are the edges of the source and target triangulations. (a) Trajectories of the convex combination morph. (b) Trajectories of the valid morph generated by raising neighborhood matrices to power trajectories are closer to straight lines than the trajectories of the convex combination morph. However, the two lower trajectories could still potentially be straight lines without afiecting the validity of the morph. (c) Trajectories of the valid morph generated using the local linear-reducible scheme. The two lower trajectories are straight lines, but the rest are identical to the corresponding trajectories of the convex combination morph. (d) Trajectories of the valid morph generated by combination of two linear-reducible schemes. The two lower trajectories are linear. The rest approach straight lines similar to the corresponding trajectories of the morph with power 1:3. 4.2 Testing Validity of the Linear Morph The linear-reducible scheme, described in the previous section, morphs an individual star linearly only if the linear morph is valid. A natural question is how to determine whether the linear morph between Z(0) and Z(1) will be valid or not. Clearly, the naive test which verifles whether Z(1 ) is a valid triangulationis not enough. To verify whether Z(t) is a valid triangulation for all 0 t 1 is impossible in practice, since [0; 1] is a continuum. Appendix A presents a robust and fast (linear time complexity) method to perform the test. This method can also be applied to check the validity of linear morphs for general triangulations. According to Theorem 4.1 it is su-cient to check the validity of the linear morphs for all corresponding stars of the two triangulations. The complexity of this test is O(V (T )), namely, linear in the size of the triangulations. 4.3 Improving Triangle Area Behavior This section describes a method for improving the behavior of the triangle areas during the morph. The triangle areas do not always evolve uniformly during the morph when using the methods described in the previous sections. In fact, the triangle areas may evolve linearly only when the triangulations have a single interior vertex. For a speciflc triangle i, we would like its area, denoted by Si, to behave for This cannot be satisfled for all triangles of the triangulation for all 0 < t < 1. Consider the following equation for the area of a triangle with vertices i, j and k: Areas of triangles with two boundary vertices are transformed uniformly only when the third vertex travels linearly with a constant velocity. Areas of triangles with a single boundary vertex are quadratically (not uniformly) transformed, since the two non-boundary vertices travel linearly with constant velocities, by (21). The problem of the triangle area improvement may be formulated as follows. Denote by Si(t) the ACM Transactions on Graphics, Vol. X, No. X, xxx 2001. Gotsman desired area of a triangle i that evolves linearly: Thus, a morph between two triangulations should minimize a cost function such as: We now show how to improve the triangle area evolution using the local scheme, such that the resulting morph is at least closer to (23) than the convex combination morph. Since it is di-cult to improve the triangle areas for the entire triangulation, the improvement may be done separately for the stars of the triangulation. This is performed after a morph of a speciflc star is deflned. To preserve the validity of the morph, '-components of the boundary vertices are preserved. The improvement is done by a variation of the boundary vertices in the radial direction relative to the origin. First, we consider an improvement such that all triangles within the star have exactly the desired areas, namely, the area of each triangle is Si(t). This approach, however, has a serious drawback. While every triangle has its desired area within the star, its shape signiflcantly difiers from the shape it assumes in the entire triangulation. Furthermore, since a triangle belongs to a number of stars, its shapes in the difierent stars might contradict each other considerably. Therefore the resulting morph is very unstable. The trajectories that the interior vertices travel are tortuous. The triangle areas are far from uniform and hardly better than those generated by the convex combination morph. All this means that the evolution of the triangle areas within the stars must also take into account the triangle shapes. For a speciflc triangle i, one of its vertices is the interior vertex of the star, that is placed at the origin for 0 t 1. The angle adjacent to the interior vertex cannot be changed, because it may afiect the validity of the morph. Therefore an improvement of the triangle area is achieved by a variation of the lengths of its two edges adjacent to the interior vertex. Every edge adjacent to the interior vertex belongs to exactly two triangles. Consequently, the length of the edge after an improvement for one triangle does not always coincide with the length of the edge within the second triangle. We improve the triangle areas separately for each triangle, and the resulting length of the edge is the average of the two lengths. We propose a simple method that improves the area of a single triangle and also preserves the triangle shape. This method changes the positions of the triangle vertices in the radial direction such that the triangle area evolves linearly and the lengths of the radial edges maintain the proportions they would have had, had the edge lengths evolved linearly. Let a and b be the lengths of the edges, and be the angle between them. The area of the triangle is:S = a b sin() (24)Denote by a(t) and b(t) the lengths of the edges as they evolve linearly: The resulting a(t) and b(t) are the lengths of the edges such that the triangle area is S(t) deflned by (22). In order to flnd a(t) and b(t) it is necessary to solve the following system of equations with the unique solution.2 S(t) to satisfy preserving the relation between the edges: See an example of a morph generated using this method in Figure 12. 5. EXPERIMENTAL RESULTS: MORPHING POLYGONS In practice, morphing is performed more frequently on planar flgures than on planar triangulations. Luckily, many types of planar flgures may be embedded in triangulations as a subset of the triangulation edges. Thus the problem of morphing planar flgures may be reduced to that of morphing triangulations, and the edges not part of the flgure are ignored in the resulting morph. Two popular cases of planar flgures are planar polygons and planar stick flgures. The former is a cycle of edges, and the latter a ACM Transactions on Graphics, Vol. X, No. X, xxx 2001. Controllable Morphing of Compatible Planar Triangulations 17 Fig. 12. A morph with good area behavior is generated using the local scheme with the method for area improvement. Compare with Figure 7 showing the convex combination morph between these triangulations. connected straight line graph. Embedding these types of flgures in a triangulation in an e-cient manner is a di-cult problem in itself, and has been treated separately by us in [Gotsman and Surazhsky 2001] and [Surazhsky and Gotsman 2001]. Here we will assume that these embeddings have been done, and investigate the efiect of the various triangulation morphing techniques described in the previous sections on the results. Figure 13 shows morphs between two polygons|the shapes of the two letters U and S. Figure 14 shows morphs between two stick flgures|the shapes of a scorpion and a dragony. These examples have been embedded in planar convex tilings [Floater and Gotsman 1999] (the faces are not necessarily triangles), for which all the theory developed in this paper holds too. Being convex, these tilings may be easily triangulated if needed. In both examples the linear morph self-intersects (Figure 13(a) and Figure 14(a)). The convex combination morph is valid, but has an unpleasant behavior (Figure 13(b) and Figure 14(b)). The local scheme, which averages the polar coordinates of star boundary vertices, provides good results when parts of the flgures should be rotated during the morph. Figure 13(c) and Figure 14(c) demonstrate that this results in a rather natural morph. Unfortunately, the morph of Figure 13(c) undergoes some exaggerated shrinking. This may be avoided by using the local scheme with area improvement, as in Figure 13(d). Figure 14(d) shows how to approach the linear morph while still preserving the morph validity, by using the scheme that raises the neighborhood matrices to power 17. Note that the tail travels a path similar to that of the linear morph, but by shrinking avoids self-intersection. Also note that some parts of the rest of the triangulations self-intersect, since the trajectories of all interior vertices approach the linear ones. But since we are interested only in the validity of the stick flgure itself, we can ignore the behavior of other edges. 6. CONCLUSION We have described a robust approach for morphing planar triangulations. This approach always yields a valid morph, free of self-intersections, based on the only known analytical method for generating morphs guaranteed to be valid [Floater and Gotsman 1999]. The approach, having many degrees of freedom, may be used to produce a variety of morphs, and, thus, can be tuned to obtain morphs with many desirable characteristics. 6.1 Discussion Morphing thru an intermediate triangulation poses the following interesting problem. Find a morph through an intermediate triangulation at a given time tm, in which only a subset of the interior vertices have prescribed positions. This contrasts with the scenario treated in Section 3.4, where all vertices of the intermediate triangulation have prescribed positions. While constraining only a subset of the vertices might seem easier than constraining all the vertices, it is actually more di-cult, since if all vertices are constrained, the user supplies a complete geometry compatible with the triangulation. Supplying only part of the vertex geometry leaves the algorithm the task of flnding compatible geometries for the other vertices, which is di-cult, especially since they might not exist. This (static) problem is interesting in its own right, and has applications in the generation of texture coordinates. It is only recently that Eckstein et al. [2001] have shown how to solve this problem by the introduction of (extraneous) Steiner vertices. The solution with a minimal number of Steiner vertices, Gotsman (a) (b) (c) (d) Fig. 13. Morphing simple polygons|the shapes of two letters S and U: (a) The linear morph is invalid | the polygon self-intersects. (b) The convex combination morph is valid, but unnatural. (c) Morph generated by the local scheme that averages polar coordinates. It behaves naturally, accounting for the rotation of the lower part of the S, but shrinks in an exaggerated manner. (d) Morph generated by the local scheme with area improvement. It is similar to the morph in (c), but with much less shrinking of the shape. and in particular, with none when it is possible, is still open. In general, the main di-culty stems from the fact that our morphing techniques use neighborhood matrices, which always result in a global solution to the morphing problem, making it virtually impossible to precisely control an individual vertex location (or trajectory). In Section 3.3, two conjectures are used to generate a morph that approaches the linear morph. Numerous examples support these conjectures, but a proof still eludes us. Since the matrices used to generate morphs by that method are not legal neighborhood matrices, the proof requires more a profound comprehension of the method. Section 4.3 presents a heuristic for improving the evolution of triangle areas. Further analysis of the correlation between vertex trajectories as well as triangle area behavior within the stars and behavior of these elements in the triangulation, may provide insight to more successful heuristics, perhaps even some optimal approximation to the desired triangle areas. It is important to make the techniques presented in this work applicable to real-world scenarios. As mentioned in Section 5, the techniques have already been applied to morph simple planar polygons and stick flgures by embedding them in triangulations. In practice, the triangulations are built around them. For example, the triangulations in which simple polygons are embedded are constructed by compatibly triangulating the interior of the polygons and an annular region in the exterior of the polygon between the polygon boundary and a flxed convex enclosure. See [Gotsman and Surazhsky 2001; Surazhsky and Gotsman 2001] for more details. These works have yet to be extended to morph planar flgures with arbitrary (e.g. disconnected) topologies. Controllable Morphing of Compatible Planar Triangulations 19 (a) (b) (c) (d) Fig. 14. Morphing between flgures of a scorpion and a dragony: (a) The linear morph is invalid | the flgure self- intersects. (b) The convex combination morph is valid, but unnatural. (c) Morph generated by the local scheme that averages polar coordinates. It behaves naturally, accounting for the rotation of the tail. (d) Morph generated by the raising the neighborhood matrices to power 17. Note that the tail travels a path similar to that of the linear morph (a), but it shrinks in order to avoid self-intersection. Morphing triangulations is usually useful as a means to morph planar flgures, and in that case the flxed convex boundary is not restrictive. However, if the objective is to actually morph two triangulations (e.g. for image warping), then a flxed common convex boundary might be restrictive. Fortunately, using the methods of [Gotsman and Surazhsky 2001; Surazhsky and Gotsman 2001], it is possible to overcome this by embedding the source and target triangulations with difierent boundaries, in two larger triangulations with a common flxed boundary. In practice this is done by compatibly triangulating the annulus between the original and new boundary, possibly introducing Steiner vertices. See Figure 15. 6.2 Future Work A challenging research subject would be to extend the techniques of this work to three dimensions, certainly, starting from an extension of [Tutte 1963]. Furthermore, it would be interesting to address the problems in [Aronov et al. 1993; Babikov et al. 1997; Souvaine and Wenger 1994; Etzion and Rappoport 1997] for 3D. A. TESTING VALIDITY OF THE LINEAR MORPH OF A STAR Let v0 be the interior vertex of the stars with degree d. The corresponding boundary vertices are indexed without loss of generality as in a counterclockwise order with respect to the interior vertex. Let be the boundary vertex coordinates during the linear morph. We denote by i(t); 'i(t) polar coordinates of the vertex i. Let 'i(t) be the angle of a triangle i adjacent to the interior vertex. Note that all calculations with indices are performed modulo d. It is ACM Transactions on Graphics, Vol. X, No. X, xxx 2001. 20 V. Surazhsky and C. Gotsman Fig. 15. Embedding two compatible triangulations (shaded regions) with difierent boundaries into larger triangulations with a common flxed boundary. assumed that the polar coordinates of the vertices for are chosen in such a way that is necessary to check that the linear morph preserves the triangle orientations, namely, it should be verifled that 0 < i(t) < for 0 t 1. To verify this, it is su-cient to check the extrema of i(t) on [0; 1]. The extremum points may be found by solving the equation For notational simplicity, we denote b. Due to the linear traversals of the vertices, we have: The '-component of the polar coordinates is expressed as: The next step is to derive '0b(t)'0a(t). But the sign(z) function is not convenient for the derivation. To overcome this problem we perform some substitutions for '(t). Since j'a(1) 'a(0)j < we can rotate both vertices va(0) and va(1) by the same angle !a round the origin such that the vertices are placed in the upper half plane, namely, the y-components of the vertices are positive. We denote the rotated coordinates by (x~; y~) with the polar '-component '~. Thus, we have Since the rotation is an a-ne transformation, it is easy to see that for the line segment deflned by (27): Clearly, due to y~a(0) > 0, y~a(1) > 0 and (27). Consequently, 'a(t) deflned as in (29) may now be expressed as: may easily be derived and after the simpliflcation we get: Due to the rotational invariance of the nominator and the denominator, we can return to the original (not rotated) coordinates: Controllable Morphing of Compatible Planar Triangulations 21 The similar procedure of a rotation may be performed for the vertex b, since for b it also holds that Therefore we can write The expression x~2(t) being -components of the polar coordinates, is strictly positive. Hence, 35 is equivalent to: Since x(t) and y(t) are linear in t, (36) is a quadratic equation in t and may be solved analytically. --R On compatible triangulations of simple polygons. Constructing piecewise linear homeomorphisms of polygons with holes. Texture mapping with hard constraints. On compatible star decompositions of simple polygons. Parameterization and smooth approximation of surface triangulation. How to morph tilings injectively. Polygon morphing using a multiresolution representation. Guaranteed intersection-free polygon morphing Introduction to linear and nonlinear programming. Joint triangulations and triangulation maps. 2D shape blending: an intrinsic solution to the vertex path problem. A physically based approach to 2D shape blending. Shape blending using the star-skeleton representation Constructing piecewise linear homeomorphisms. Surface interpolation based on new local coordinates. Morphing planar triangulations. Morphing stick flgures using optimized compatible triangulations. Image morphing with feature preserving texture. How to draw a graph. accepted May --TR Joint triangulations and triangulation maps A physically based approach to 2MYAMPERSANDndash;D shape blending Feature-based image metamorphosis On compatible triangulations of simple polygons 2-D shape blending Parametrization and smooth approximation of surface triangulations Three-dimensional distance field metamorphosis Foldover-free image warping How to morph tilings injectively As-rigid-as-possible shape interpolation On Compatible Star Decompositions of Simple Polygons Morphing Stick Figures Using Optimized Compatible Triangulations --CTR David Vronay , Shuo Wang, Designing a compelling user interface for morphing, Proceedings of the SIGCHI conference on Human factors in computing systems, p.143-149, April 24-29, 2004, Vienna, Austria Jeff Danciger , Satyan L. Devadoss , Don Sheehy, Compatible triangulations and point partitions by series-triangular graphs, Computational Geometry: Theory and Applications, v.34 n.3, p.195-202, July 2006 Anna Lubiw , Mark Petrick , Michael Spriggs, Morphing orthogonal planar graph drawings, Proceedings of the seventeenth annual ACM-SIAM symposium on Discrete algorithm, p.222-230, January 22-26, 2006, Miami, Florida Hayley N. Iben , James F. O'Brien , Erik D. Demaine, Refolding planar polygons, Proceedings of the twenty-second annual symposium on Computational geometry, June 05-07, 2006, Sedona, Arizona, USA Vitaly Surazhsky , Joseph (Yossi) Gil, Type-safe covariance in C++, Proceedings of the 2004 ACM symposium on Applied computing, March 14-17, 2004, Nicosia, Cyprus

linear Morph;morphing;compatible triangulations;self-intersection elemination;controllable Morphing;isomorphic triangulations;local Control

502796

Application of aboutness to functional benchmarking in information retrieval.

Experimental approaches are widely employed to benchmark the performance of an information retrieval (IR) system. Measurements in terms of recall and precision are computed as performance indicators. Although they are good at assessing the retrieval effectiveness of an IR system, they fail to explore deeper aspects such as its underlying functionality and explain why the system shows such performance. Recently, inductive (i.e., theoretical) evaluation of IR systems has been proposed to circumvent the controversies of the experimental methods. Several studies have adopted the inductive approach, but they mostly focus on theoretical modeling of IR properties by using some metalogic. In this article, we propose to use inductive evaluation for functional benchmarking of IR models as a complement of the traditional experiment-based performance benchmarking. We define a functional benchmark suite in two stages: the evaluation criteria based on the notion of "aboutness," and the formal evaluation methodology using the criteria. The proposed benchmark has been successfully applied to evaluate various well-known classical and logic-based IR models. The functional benchmarking results allow us to compare and analyze the functionality of the different IR models.

Summary In summary, the probabilistic model has the highest degree of potential precision, followed by the threshold vector space model, then the Boolean model and the nave vector space model. This conclusion is consistent with the experimental results. The motivation for this judgment lies in the varying degrees to which they respectively support (or don't support) conservative monotonicity. 4. Inductive Evaluation of Logical IR Models In the past decade, a number of logic based IR models have been proposed (see [Bruza and Lalmas 1996; Lalmas 1998; Lalmas and Bruza 1998] for detailed surveys). These models can be generally classified into three types: Situation Theory based, Possible World based, and other types. In what follows, we investigate three well-known logic IR models. In the following analyses, the fact of a document D consisting of information ~ carrier i is represented by Dfi i. For example, Guarded Left Compositional Monotonicity (i.e., postulate 7) means that if a document consisting of i is about k (i.e. under the guarded condition that i doesn't preclude j (i ^/ j), we can conclude that a document consisting of i j is about k (i j |= k). In the following benchmarking exercise, we adopt this interpretation for logical IR models for reasons of simplicity. For the classical models, we treated the document and the query as information carriers directly, for there are no term semantic relationships involved in classical models. 4.1 Situation theory based model 4.1.1 Background Rijsbergen and Lalmas developed a situation theory based model [Lalmas 1996; Rijsbergen and Lalmas 1996]. In their model, a document and the information it contains are modeled as a situation and types. A situation s supports the type j, denoted by s|=j, means that j is a part of the information content of the situation. The flow of information is modeled by constraints (fi ). Here, we assume jfi j. A query is one type (single type query) or a set of types (complex query), e.g., a query f For a situation s and a set of types f, there are two methods to determine whether d supports f. The first is that d supports f if and only if s supports j for all types j f [Barwise 1989]. Later Lalmas relaxed the condition to represent partial relevance: any situation supports f if it supports at least one type in f [Lalmas 1996]. IR system is to determine to which extent a document d supports the query f, denoted by d|=f. If d|=f, then the document is relevant to the query with certainty. Otherwise, constraints from the knowledge set will be used to find the flow that lead to the information f. The uncertainty attached to this flow is used to compute the degree of relevance. A channel is to link situations. The flow of information circulates in the channel, where the combination of constraints in sequence (c ;c ) and in parallel (c ||c ) can be represented. Given two situations s1, s2, s1| cfi s2 means that s1 contains the information about s2 due to the existence of the channel c. A channel c supports constraint jfi y, denoted c|=jfi y, if and only if for all situations s1 and s2, if s1|=j, s1|fi s2, and jfi y, then s2|=y. The notation s1|=j | cfi s2|=y stands for c|=jfi y and s1|fi s2, which means that s1|=j carries the information that s2|=y, due to channel c. If s1|=j | cfi s2|=y and s1=s2, then c is replaced by a special channel 1, and j logically entails y. 4.1.2 Situation Theory Based Aboutness (|= ) Let U be the set of documents, S be the set of situations, T be the set of types, C be the set of channels. Furthermore, let DU is a document, and Q is a query. Then, D is modeled as a situation. Q is modeled as a set of types Given two set of types f1 and f2: Dfi~ f1 iff ("jf1)(D|=j). f1 |= f2 iff ($ c C) ("D|Dfi~ f1) ($jf1) ($yf2) (D |=j | cfi D' |=y). Note that D' could be D itself, i.e. c=1. A more special case is D |=y | 1fi D |=y. (Aboutness) f1 |=/ ST f2 iff ($/ c C) ("D|Dfi~ f1) ($jf1) ($yf2) (D |=j | cfi D' |=y). f1 sfi f2 iff f1 f2 (Surface f1 d fi f2 iff ($y1f1) ($y2 f2) (jfi y). (Deep f1 f2 f1 f2 (Composition) precludes its negation, e.g., (s| s|=<<hit, john, john, x; 0>>). Suppose the negation of a set of types Q is the set of the negations of every component type, then Q^Q. 4.1.3 Inductive Evaluation Situation theory based IR model supports R, C, LM, RM, M, C-FA, GLM, GRM, QLM and QRM. The proofs are provided as follows: 1. R: Reflexivity is supported. Given Dfi~ f, Q=f ($ c C and c=1) ("D|Dfi~ f) ($jf) (D |=j | 1fi D |=y). 2. C: Containment is supported. Surface containment is supported. Given f1 sfi f2 f1 f2 ($ c C and c=1) ("D|Dfi~ f1) ($jf1) ($jf2) (D |=j | 1fi D |=j). Deep containment is supported. Given f1 d fi f2 ($ c C) ("D|Dfi~ f1) ($y1f1) ($y2 f2) (D |=y1 | cfi D' |=y2). 3. RCM: Right Containment Monotonicity is not supported. Surface containment: Given f1 |= f2 and f2 sfi f3 ($ c1 C)("D|Dfi~ f1)($y1f1)($y2 f2) (D |=y1 | c1fi D' |=y2) and But it not necessary that y2 f3 It is not necessary that f1 |= f3. Deep containment: Given f1 |= f2 and f2 d fi f3 ($ c1 C)("D|Dfi~ f1)($y1f1)($y2 f2) (D |=y1 | c1fi D' |=y2) and But it is not necessary that j1=y2 It is not necessary that ($c2C) ("D|Dfi~ f1) ($y1f1) ($j2f3) (D|=y1 | c1fi D' |=j2) It is not necessary that f1 |= f3. 4. LM: Left Compositional Monotonicity is supported. Given f1 |= f2 ($ c1 C) ("D|Dfi~ f1) ($y1f1) ($y2 f2) (D |=y1 | c1fi D' |=y2), f1 f3 f1 f3, and {"D| D ~ f1 f3} {"D| D fi~ f1} ("D|Dfi~ f1 f3) ($y1f1 f3) ($y2 f2) (D |=y1 | c1fi D' |=y2), 5. RM: Right Compositional Monotonicity is supported. Given f1 |= f2 ($ c1 C) ("D|Dfi~ f1) ($y1f1) ($y2 f2) (D |=y1 | c1fi D' |=y2), f2 f3 f2 f3, and {"D| D ~ f2 f3} {"D| D fi~ f2} ("D|Dfi~ f1) ($y1f1) ($y2 f2 f3) (D |=y1 | c1fi D' |=y2), 6. Mix (M) is supported. Given f1 |= f2 and f3 |= f2 ($ c1 C) ("D|Dfi~ f1) ($y1f1) ($y2 f2) (D |=y1 | c1fi D' |=y2) and Recall f1 f3 f1 f3 (D |=y1 | c1fi D' |=y2) and (D |=j1 | c2fi D' |=j2) 6 If Q1 and Q2 are single types, RCM would be supported. Here, however, we consider Q as a set of types, which is a more general case. 7. C-FA: Context Free And is supported. Given f1 |= f2 and f1 |= f3 ($ c1 C) ("D|Dfi~ f1) ($y1f1) ($y2 f2) (D |=y1 | c1fi D' |=y2) and ($ c2 C) ("D|Dfi~ f1) ($j1f1) ($j2 f3) (D |=j1 | c2fi D' |=j2) Recalling f1 f2 f1 f1 ($ c1 C) ("D|Dfi~ f1) ($y1f1) ($y2 f2 f3) (D |=y1 | c1fi D' |=y2) and ($ c2 C) ("D|Dfi~ f1) ($j1f1) ($j2 f2 f3) (D |=j1 | c2fi D' |=j2) 8. GLM: Guarded Left Compositional Monotonicity is trivially supported, as LM is supported. 9. GRM: Guarded Right Compositional Monotonicity is trivially supported, as RM is supported. 10. QLM: Qualified Left Monotonicity is trivially supported, as LM is supported. 11. QRM: Qualified Right Monotonicity is trivially supported, as RM is supported. 12. NR: Negation Rational is not supported. Given f1|=/ ST f2 ($/ c C) ("D|Dfi~ f1) ($jf1) ($yf2) (D |=j | cfi D' |=y) This does not imply that ($/ c C) ("D|Dfi~ f1) ($jf1) ($yf2 f3) (D |=j | cfi D' |=y) \ f1|=/ ST f2 f3 can not be guaranteed. 13. CWA: Close World Assumption is not supported. Given f1|=/ ST f2, f2^f2 ($/ c C) ("D|Dfi~ f1) ($jf1) ($yf2) (D |=j | cfi D' |=y) But it doesn't mean it is necessary that ($ c C) ("D|Dfi~ f1) ($jf1) (D |=j | cfi D' |= y) \ We cannot conclude f1|= f2. 4.2 Terminological Logic based model 4.2.1 Background Meghini, et al. proposed an IR model based on Terminological Logic (TL) [Meghini et al. 1993]. An object-oriented approach is used to represent documents, queries, and lexical, thesaural knowledge. The representations are not confined to describing the content by a set of keywords. Instead, the contextual attributes, the layout characterizations, the structure organizations and the information contents of the documents are also taken into account. uses terms as the primary syntactic expressions, which are concepts (monadic relations), roles (dyadic relations), and individuals (see [Meghini et al. 1993] for formal definitions). Documents are modeled as individual constants and a query is a concept entailing a class of individuals, while a document can be an instance of a set of concepts describing the properties of the document. The assertion C(i) means that an individual i is an instance of a concept C. Concepts can be partially ordered by subsumption which is specified by terminological postulates comprising the thesaural knowledge base W. A terminological postulate is an expression of the form connotation (<.) or definition . (= ). The semantics of TL is defined by the interpretation I over the nonempty set of individuals U i.e., the domain of discourse. I is a function that maps individual constants into elements of U such that I(i )I(i subsets of D and roles into subsets of U.U (see (Meghini et al. 1993) for details). An algorithm called constraint propagation is also proposed for reasoning the complete constraint set on the knowledge base by using a set of completion rules. This inference is performed at KB construction time rather than at query time. When a query C is formulated, it is added to the constraint set and the completion rules are applied; then for every individual constant i occurring in the set, C(i) is checked by simple table lookup techniques. 4.2.2 Let U be the set of all the documents, C be an alphabet of concepts. Furthermore, let DU be a document, and QC be a query, then the aboutness in Terminological Logic is defined as follows: D is modeled as an individual constant. Q is modeled as a concept. For C1, C2 C, Dfi~ C1 implies that C1(D) is satisfied, i.e. DI(C1). C1fi C2 implies C1 is subsumed by C2, i.e. I(Q1) I(Q2). (Containment) Note that if there are no terminological postulates involved, they are surface containment. Otherwise, they are deep containment. As the surface containment and deep containment have the same semantics under the interpretation I, we need not distinguish them in the following proofs. C1|= C2 Given any Dfi~ C1 and Q=C2, Q(D) is satisfied, i.e. D I(Q). (Aboutness) C1|=/ For every D such that Dfi~ C1 and Q=C2, Q(D) is unsatisfied, i.e. D I(Q). C1^C1. Note that I(a-not 4.2.3 Inductive Evaluation based model supports R, C, RCM, C-FA, LM, M, GLM, QLM, NR and CWA. The proofs are shown as follows: 7 (and C1 C2 Cn) denotes the set of those individuals that are denoted by C1 and C2 and Cn. I(and C1 C2 8 (a-not C) denotes the set of all individuals that are not denoted by C. I(a-not C)=U\I(C). 1. R: Reflexivity is supported. ~ Given Dfi C1 and Q=C1 D I(C1) D I(Q) C1|= C1 2. C: Containment is supported. ~ Given C1 fi C2, Dfi C1 and Q=C2 D I(C1) I(C2) D I(C2) C1|= C2 3. RCM: Right Containment Monotonicity is supported. Given C1|= C2, and C2fi C3 C1|= C2 For every D such that D fi~ C1, there is D I(C2) D I(C3) 4. LM is supported. Given C1|= C2 C1|= C2 For every D such that D fi~ C1, there is D I(C2) \ For every D such that D fi~ C1 C3, there is D I(C2) C1 C3|= C2. 5. RM: Right Compositional Monotonicity is not supported. Given C1|= C2 C1|= C2 For every D such that D fi~ C1, there is D I(C2). But it is not necessary that D I(C2 C3), as I(C2 C3) is a subset of I(C2). \It is not necessary that C1 |= C2 C3. 6. Mix (M) is supported. Given C1|= C2 and C3|= C2 C1|= C2 For every D such that D fi~ C1, there is D I(C2) C3|= C2 For every D such that D fi~ C3, there is D I(C2) \ For every D such that D fi~ C1 C3, there is D I(C2) C1 C3|= C2. 7. C-FA is supported. Given C1|= C2, C1|= C3 C1|= C2 For every D such that D fi~ C1, there is D I(C2) C1|= C3 For every D such that D fi~ C1, there is D I(C3) \ For every D such that D fi~ C1, there is D I(C2). I(C3), i.e. D I(C2 C3) C1 |= C2 C3. 8. GLM: Guarded Left Compositional Monotonicity is trivially supported, as LM is supported. 9. GRM: Guarded right Compositional Monotonicity is not supported for the similar reason of RM. 10. QLM: Qualified Left Monotonicity is trivially supported as LM is supported. 11. QRM: Qualified Right Monotonicity is not supported for the similar reason of RM. 12. NR: Negation Rational is supported. C1|=/ For every D such that D fi~ C1, there is D I(C2) D I(C2 C3) 13. CWA: Close World Assumption is supported. Given C1 | C2, C2^C2 C1|=/ For every D such that D fi~ C1, there is D I(C2) C2^C2 I(a-not Q1)=U\I(Q1) D I(not Q1) 4.3 Possible world based model 4.3.1 Background A number of possible world based logical IR models have been proposed. As stated in [Lalmas and Bruza 1998], these systems are founded on a structure <W, R>, where W is the set of worlds and R W.W is the accessibility relation. They can be classified according to the choice made for the worlds wW and accessibility relation R. For example, w can be a document (or its variation) and R is the similarity between two documents w1 and w2 [Nie 1989; Nie 1992], or w is a term and R is the similarity between two terms w1 and w2 [Crestani and van Rijsbergen 1995(a); Crestani and van Rijsbergen 1995(b); Crestani and Van Rijsbergen 1998], or w is the retrieval situation and R is the similarity between two situations w1 and w2 [Nie et al. 1995], etc. Most of these systems use a technique called imaging. To obtain P(dfi q), where the connective fi represents conditional, we can move the probability from non-d- world to d-world by a shift from the original probability distribution P of the world w to a new probability distribution Pd of its closest world wd where d is true. This process is called deriving P from P by imaging on d. The truth of dfi q at w will d then be measured by the truth of q at wd . To simplify the analysis, let's suppose that the truth of q in a world is binary and the closest world of a world w is unique . P(dfi q) can be computed as follows: P(d 1, if q is true in w 0, otherwise 0, otherwise wd is the closest world of w where d is true Now, we study in detail Crestani and van Rijsbergen's model which models the terms as possible worlds to see some properties of the possible world based approach. In this model, term is considered as vector of documents, while the document and query are vectors of terms. The accessibility relations between terms are estimated by the co-occurrence of terms. P(dfi q) can be computed as: P(d occurs in q 0, otherwise 0, otherwise td is the closest term of t where d is true(td occurs in d) (12) Generally, d is deemed relevant to q when P(dfi q) is greater than a threshold value, e.g., a positive real number . Similar to the vector space model (see section 3.3.2), the simplest case is that at least one term which occurs in both d and q, or it is also the closest term of some other terms occurring in d and q. This case is referred to as nave possible world based model and the general case as threshold possible world based model. 9 Actually, it can be multi-valued in an interval. There is also an approach called General Logical Imaging that does not rely on this assumption. 4.3.2 Nave Possible World Aboutness Based on Crestani and van Rijsbergen's Model (|=NAIVE-PW -CV ) Let U be the set of all the documents, T be the set of all the index terms, Furthermore, let DU be a document, Q be a query, and t be a term. The aboutness in the nave Possible World based models is defined as follows: D and Q are sets of terms (Surface containment) is the closest term of t2 (Deep containment) Q1 Q2 Q1 Q2 Preclusion is foreign to this model. 4.3.3 Inductive evaluation This model supports R, C (surface containment), LM, RM, M and C-FA. Proofs are given as follows: 1. R: Reflexivity is supported. Given P(D fi 2. C: Surface containment is supported. Given DQ Deep containment is not supported. Given t1 D, t2 Q, and t1fi t2 t2 can be imaged to its closest D-world t1 But this cannot imply P(D fi Q may not occur in Q. \D|= Q cannot be guaranteed. 3. RCM: Right Containment Monotonicity is not supported. For surface containment: Given D|= Q1, and Q1fi Q2 P(D fi But it does not imply ($t Q2) ($t' T) (I(t ,t' ) =1) It is not necessary that P(D fi \ It is not necessary that D|= Q2 For deep containment: Given P(D fi Q1) > 0, t1 occurs in Q1, t1fi t2, and t2 occurs in Q2. This does not mean that there must exist a term which is the closest term of some terms where D is true, and occurs in Q2. Thus, it is not necessary that D|= Q2 4. LM: Left Compositional Monotonicity is supported. Given D1|= Q, and D= D1 D2 At least one term t is the closest term of some terms where is true and t Q, and D1 D2=D1 D2 t is also true in D1 D2, and t Q 5. RM: Right Compositional Monotonicity is supported. Given D|= Q1, and P(D fi ($t Q1) ($t' T) (I(t ,t' )=1) and t Q P(D fi Q1 \D|= Q1 Q2 trivially supported, as LM is supported. 7. C-FA: Context Free And is trivially supported, as RM is supported. 8. GLM: Guarded Left Compositional Monotonicity is inapplicable, as preclusion is foreign to this model. 9. GRM: Guarded Right Compositional Monotonicity is inapplicable, as preclusion is foreign to this model. 10. QLM: Qualified Left Monotonicity is inapplicable, as preclusion is foreign to this model. 11. QRM: Qualified Right Monotonicity is inapplicable, as preclusion is foreign to this model. 12. NR: Negation Rational is not supported. Given D|NAIVE-PW -CV Q1, P(D fi ($t Q1) ($t' T) (I(t ,t' )=1), and Q =Q1 Q2 But it is possible that ($t Q) ($t' T) ( I(t ,t' )=1), It is possible that P(D fi Q1 \ It's possible that D|= Q1 Q2. 13. CWA: Close World Assumption is inapplicable, as preclusion is foreign to this model. 4.3.4 Threshold Possible World Aboutness Based on Crestani and van Rijsbergen's Model (|=T -PW -CV ) Let U be the set of all the documents, T be the set of all the index terms, Furthermore, let DU be a document, Q be a query, and t be a term. The aboutness in this models is then defined as follows: D and Q are sets of terms D|=T -PW -CV Q iff P(Dfi Q), where is a positive real number in the interval (0, 1]. (aboutness) The mappings of containment, composition and preclusion are same as those in Section 4.3.2. 4.3.5 Inductive evaluationThis model supports R, LM, RM, M, C-FA, and conditionally supports C, RCM and NR. Proofs are given as follows: 1. R: Reflexivity is supported. The proof is the same as that of R for |=Z -PW -CV . 2. C: Surface containment is conditionally supported. Given DQ, This does not imply P(D fi Q It depends on the sum of probability of the index terms shared by D and Q and the index terms which 11 The comparison between nave and threshold PW based models are similar to that between nave and threshold vector space models. can be imaged to those shared terms. Only under the condition that the threshold is not greater than that sum, P(D fi Q D|=T -PW -CV Q) can be guaranteed. Deep containment is conditionally supported. Given t1 D, t2 Q, and t1fi t2 t2 can be imaged to its closest D-world t1 But this cannot imply P(D fi Q . Only under the condition that the threshold is not greater than the probability of the index terms shared by D and Q and the index terms which can be imaged to those shared can be guaranteed. 3. RCM: Right Containment Monotonicity is conditionally supported. For surface containment: Given D|=T -PW -CV Q1, and Q1fi Q2 P(D fi But this does not imply enough index terms shared by D and Q1 and the index terms which can be imaged to those shared terms to make Only under the condition that the threshold is set to be not greater than the sum of probability of the index terms shared by D and Q and the index terms which can be imaged to those shared terms, P(D fi (i.e. D|=T -PW -CV Q2) can be guaranteed. For deep containment: The proof and the condition are similar to those of surface containment. 4. LM: Left Compositional Monotonicity is supported. Given D1|=T -PW -CV Q, and D= D1 D2 The number of index terms which are the closest terms of certain terms where true must be not less than that of index terms which are the closest terms of certain terms where D1 is true. This implies that P (t) P (t) . 5. RM: Right Compositional Monotonicity is supported. The proof is similar to that of C-FA. 6. Mix (M) is supported. Given D1|=T -PW -CV Q, D2|=T -PW -CV Q 7. C-FA: Context Free And is supported. Given D|=T -PW -CV Q1, D|=T -PW -CV Q2, and P(D fi Q=Q1 Q2=Q1 Q2 (Q1 Q and Q2 Q), P(D fi P(D fi Q1 \D|=T -PW -CV Q1 Q2 8. GLM: Guarded Left Compositional Monotonicity is inapplicable, as preclusion is foreign to this model. 9. GRM: Guarded right Compositional Monotonicity is inapplicable, as preclusion is foreign to this model. 10. QLM: Qualified Left Monotonicity is inapplicable, as preclusion is foreign to this model. 11. QRM: Qualified Right Monotonicity is inapplicable, as preclusion is foreign to this model. 12. NR: Negation Rational is conditionally supported. Given D|T -PW -CV Q1, P(D fi But it's possible that ($ t Q) ($ t' T) ( I(t ,t' )=1) and the number of t is large It's possible that P(D fi Q1 \ It's possible that D|=T -PW -CV Q1 Q2. Only under the condition that the threshold is set to be greater than the sum of probability of the index terms shared by D and Q and the index terms which can be imaged to those shared terms, P(D fi D|T -PW -CV Q) can be guaranteed. 13. CWA: Close World Assumption is inapplicable, as preclusion is foreign to this model. 4.4 Discussion Deep containment is not relevant to classical models, unless they are augmented by thesauri from which deep containment relationships like penguin ELUG can be extracted. Logical models, by their very nature, directly handle deep containment relationships. This means logical models are able to capture information transformation e.g., logical imaging in the possible world models. This is a major advantage of logical models. Moreover, they provide stronger expressive power, e.g. based model provides the structured representation of information, while concepts such as situation, type and channel, etc. in situation theory based model make it more flexible. The properties of an IR model are largely determined by the matching function it supports. Two classes of matching function are widely used: containment andoverlapping (nave and non-zero threshold). The Boolean and based models have similar properties (except that some properties inapplicable to Boolean model are supported by based model), due to their common retrieval mechanism, namely containment, which requires that all the information of the query must be contained in or can be transformed to the information of the document. The nave vector space model and nave possible world based model have similar properties (except that deep containment is applicable to possible world based model only) due to their simple overlapping retrieval mechanism (i.e., a document is judged to be relevant if it shares at least one term with the query). Compared with Boolean and based models, the nave vector space and the nave possible world based model support Left and Right Compositional Monotonicity, which causes imprecision. The Boolean and based models support Right Containment Monotonicity, which promotes recall and supports the Negation Rationale, which can improve precision. In the nave vector space and possible world based models, Right Containment Monotonicity and Negation rational are not supported. In summary, there is evidence to support the assumption that the Boolean and based models are more effective models than the nave vector space and the nave possible world based model. The nave possible world model uses imaging (i.e., imaging from non-D world to D-world) besides simple overlapping. Even though there may exist a containment relation between a term t1 in the document and another one t2 in the query, if t1 is not shared by the document and the query, then this transformation from t2 to t1 is ineffective to establish the relevance. This explains why nave possible world model does not support Containment (deep). The mechanics of imaging is dependent on a notion of similarity between worlds. Experimental evidence shows 12 The discussion of Boolean model is based on the assumption that the information composition is modeled by logical AND as we adopted in 3.1. a relation between retrieval performance and the way in which the relationship between worlds is defined [Crestani and Van Rijsbergen 1998]. As the underlying framework for inductive evaluation presented in this paper does not explicitly support a concept of similarity, it can be argued that the mapping of the possible worlds based model into the inductive framework is incomplete. More will be said about this point in the conclusions. The threshold possible world model is (surprisingly) both left and right monotonic. As a consequence there is some grounds to conclude that this model would be imprecise in practice, and also be insensitive to document length. As mentioned in the previous point, retrieval performance depends on how the similarity between worlds is defined. As both LM and RM are supported, it can be hypothesized that the baseline performance for the threshold possible world model would be similar to the nave overlap model. More sophisticated similarity metrics between worlds would improve performance above this baseline. Crestani and Rijsbergen allude to this point as follows: . it is possible to obtain higher levels of retrieval effectiveness by taking into consideration the similarity between the objects involved in the transfer of probability. However, the similarity information should not be used too drastically since similarity is often based on cooccurrence and such a source of similarity information is itself uncertain [Crestani and Van Rijsbergen 1998]. When the threshold possible world model judges a document D relevant to the query Q, this implies that D shares a number of terms with Q or a number of terms can be transformed to the shared terms so that P(Dfi Q) is not less than the threshold . The expansion of D or Q can only increase P(Dfi Q). This judgment is not true for threshold vector space model, for after the expansion of D (or Q), the increase of the space of D (or Q), i.e. number of terms in D and Q, may be much more than the increase of the shared terms. Thus the degree of overlapping may be decreased. The threshold possible worlds model and situation theory using Lalmas' relaxed condition support LM and RM, and based model supports LM. This implies that these models turn out to be less precise than probabilistic and threshold vector space models. This in turn reflects the fact that logical models have not yet shown the performance hoped for since their inception. 5. Results Summary and Conclusions Table 1: Summary of the results of the evaluation. Models Postulates Boolean Nave Vector Space Threshold Vector Space Probabilistic Model Situation Theory Based Terminological Logic Based Nave Possible World Threshold Possible World R CS C (Deep) NA NA NA NA . CS RCM (Surface) . CS CS . . CS RCM (Deep) NA NA NA NA . . CS RM . CS CS . C-FA CS CS GLM NA NA NA NA NA NA GRM . NA NA NA . NA NA QLM NA NA NA NA NA NA QRM . NA NA NA . NA NA NR . CS CS . . CS CWA NA NA NA . NA NA Note: NA means not applicable, CS means conditionally support, means support; and . means not support. 5.2 Conclusion The functional benchmarking exercise presented in this paper indicates that functional benchmarking is both feasible and useful. It has been used to analyze and compare the functionality of various classical and logical IR models. Through the functional benchmarking, phenomenaoccurring in the experimental IR research can be explained from a theoretical view. The theoretical analysis could in turn help us better understand IR and provide guideline to investigate more effective IR models. A major point could be drawn here is that IR is conservatively monotonic in nature. It is important that the conservatively monotonic model be studied and developed, as it would help get an optimal tradeoff between precision and recall. The postulates GLM, GRM, QLM, QRM, etc. guarantee the conservatively monotonic properties, but they are foreign to some models. Even in those models, which support some of conservatively monotonic properties, preclusion is only based on the assumption that an information carrier precludes its negation. Moreover, GLM, QLM and MIX are the special cases of LM, and GRM, QRM and C-FA are the special case of RM. As such, if a model supports LM, and GLM and CautM are applicable, then it must also support GLM and CautM. In this case, these conservative monotonicity properties have no effect. Therefore, a model supporting conservative monotonicity should embody conservatively monotonic properties and without also supporting LM and RM. The probabilistic model and threshold vector space model show the good performance in practice because they mimic the conservatively monotonicity. However, they are dependant of factors set extraneously, e.g. the threshold value. This is undesirable from theoretical point of view. Current logical IR models have advantage of modeling information transformation and their expressive power, however, they are still insufficient to model conservative monotonicity. A primary reason is that the important concepts, such as (deep and containment, information preclusion, etc., upon which conservative monotonicity is based, are not sufficiently modeled. For example, semantics of information preclusion is not explicitly defined in current logical models. We just simply assume that an information carrier precludes its negation during the benchmarking. It is interesting to show that if we add some kind of semantics of preclusion to the logical IR models, the conservative monotonicity could be partially realized. For example, we could add the following definition to the model: Preclusion: Given two types j1 and j2, j1^j2, s1|=j1 and s2|=j2, there does not exist any channel between s1 and s2. The Left composition monotonicity (LM) is no longer supported: Given f1 |= f2 ($ c1 C) ("D|Dfi~ f1) ($y1f1) ($y2 f2) (D |=y1 | c1fi D' |=y2), Assume LM is supported, i.e. ("D|Dfi~ f1 f3) ($y1f1 f3) ($y2f2) (D |=y1 | c1fi D' |=y2). Consider the case of f2^f3. This implies for D|=f3 and D' |=f2, there does not exist a channel between D and D'. This contradicts the above assumption, because {"D|D~ f1 f3} {"D| D|=f3}. \ It is not necessary that f1 f2 |= f2. On the other hand, RM is not supported for the similar reason of LM. However, by applying the conservative forms of monotonicity, QLM and QRM, with the qualifying non-preclusion conditions, the above-like counter example will no longer exist. The above definition of preclusion is simple and just for the purpose of illustration. It is true that current IR systems are not defined in terms of these concepts mainly because they do not view retrieval as an aboutness reasoning process. However informational concepts are in the background. Preclusion relationships can be derived via relevance feedback [Amati and Georgatos 1996, Bruza et al.1998]. For restricted domains, information containment relationships can be derived from ontologies, and the like. For example, we have been investigating automatic knowledge discovery from text corpus based on Barwise and Seligman's theory of information flow [Barwise and Seligman 1997, Bruza and song 2001; Song and Bruza 2001]. When language processing tools have advanced further, the concepts under the aboutness theory could be applied to IR more easily and more directly. More sensitive IR systems would then result; in particular those which are conservatively monotonic with respect to composition. Therefore, more investigations about how to achieve conservative monotonicity in current logical IR models are necessary. Finally, we reflect on the strengths and weaknesses of the inductive theory of information retrieval evaluation. The strengths are summarized below: Enhanced perspective: Matching functions can be characterized qualitatively in terms of aboutness properties that are, or are not implied, by the matching function in question. It may not be obvious what the implications are of a given numeric formulation of a matching function. The inductive analysis allows some of these implications to be teased out. By way of illustration, models based on overlap may imply monotonicity (left or right), which is precision degrading. In addition, inductive analysis allows one to compute under what conditions a particular aboutness property is supported. It has been argued that a conservatively monotonic aboutness relationship promotes effective retrieval. The analysis in this paper revealed that although both of these models support conservative monotonicity, the fundaments of this support are very different: The thresholded vector space model support for conservative monotonicity depends on overlap between document and query terms modulo the size of the document [check this because our formulations don't include document length normalization]. Support for conservative monotonicity in the probabilistic model depends on whether the terms being added have a high enough probability of occurring in relevant documents. Form an intuitive point of view, the latter condition would seem a more sound basis for support because it is directly tied to relevance. Transparency: One may disagree with a given functional benchmark (as represented by a set of aboutness properties), or with how a given matching function has been mapped into the inductive framework, however, the assumptions made have been explicitly stated. This differs from some experimental studies where the underlying assumptions (e.g., the import of certain constants) are not, or insufficiently, motivated. New insights: The use of an abstract framework allows new insights to be gleaned. Inductive evaluation has highlighted the import of monotonicity in retrieval functions, and its affect on retrieval performance. Designers of new matching functions should provide functions that are conservatively monotonic with respect to the composition of information. More sensitive IR systems would then result. The lack of such systems currently can be attributed in part to the inability to effectively "operationalize" information preclusion. Most common IR models are either monotonic or non-monotonic - another class of IR models, namely those that are conservatively monotonic is missing. Such models are interesting for purposes for producing symbolic inference foundation to query expansion and perhaps even relevance feedback. The weaknesses of an inductive theory for evaluation are: Difficulty in dealing with weights: Much of the subtlety of IR models remains buried in different weighting schemes. Due to its symbolic nature, the inductive approach can abstract too much, thereby losing sensitivity in the final analysis. For example, the nuances of document length normalization, term independence assumptions, probabilistic weighting schemes are difficult, if not impossible, to map faithfully into a symbolic, inductive framework Difficulties with mapping: For an arbitrary model, it may not be obvious how to map the model into an inductive framework. This is particularly true for heavily numeric models such as probabilistic models. It is often the case that such models do not support many symbolic properties - they are like black holes defying analysis [Bruza, Song & Wong 2000]. However, by analysing the conditions under which given properties are supported allow us to peak at the edges of the black hole. Incompleteness of framework: In order to pursue functional benchmarking, a sufficiently expressive framework is necessary in order to represent salient aspects of the model in question. This is an issue of completeness. In the inductive analysis of the possible worlds based models presented in this paper, we have seen that the notion of similarity inherent to these models cannot be directly translated into the underlying inductive framework. This suggests that the framework presented in this paper should be extended. One could also argue that not all salient aspects of aboutness have been captured by the properties used for the benchmark. These are not criticisms of inductive evaluation, but of the tools being used. It is noteworthy that conventional experimental IR evaluation approaches are good performance indicators but fail to reflect the functionality of an IR system, i.e. which types of IR operation the system supports. From an application point of view, the experimental approaches could serve as the performance benchmark (e.g. TREC). Practically, it is complementary to the functional benchmark proposed in this paper. Acknowledgement This project is partially supported by the Chinese University of Hong Kong's strategic grant (project ID: 44M5007), and by the Cooperative Research Centres Program through the Department of the Prime Minister and Cabinet of Australia. --R Information Retrieval and Hypertext. Relevance as deduction: A logical view of information retrieval. Modern Information Retrieval. The Situation in Logic. Investigating aboutness axioms using information fields. Logic based information retrieval: Is it really worth it? Preferential models of query by navigation. Informational Inference Via Information Flow. Commonsense aboutness for information retrieval. Fundamental properties of aboutness. Information Retrieval An Axiomatic Theory for Information Retrieval. Information retrieval and situation theory. Using default logic in information retrieval. Intelligent text handling using default logic On the problem of 'aboutness' in document analysis. Theories of Information and Uncertainty for the Modeling of Information Retrieval: An Application of Situation Theory and Dempster-Shafer's Theory of Evidence Information retrieval and Dempster-Shafer's theory of evidence Logical models in information retrieval: Introduction and overview. The use of logic in information retrieval modeling. Towards a theory of information. Comparing boolean and probabilistic information retrieval systems across queries and disciplines. Text Retrieval and Filtering: Analytic Models of Performance. On indexing A model of information retrieval based on terminological logic. A relevance terminological logic for information retrieval. An information retrieval model based on modal logic. Towards a probabilistic modal logic for semantic-based information retrieval Information retrieval as counterfactual. What is information discovery about? A new theoretical framework for information retrieval. The state of information retrieval: logic and information. An information calculus for information retrieval. Retrieval of complex objects using a four-valued logic A probabilistic terminological logic for modeling information retrieval. On the role of logic in information retrieval. Discovering Information Flow using a High Dimensional Conceptual Space. Fundamental properties of the core matching functions for information retrieval. Towards a commonsense aboutness theory for information retrieval modeling. A comparison of text retrieval models. --TR Automatic text processing Towards an information logic Nonmonotonic reasoning, preferential models and cumulative logics Information retrieval Towards a probabilistic modal logic for semantic-based information retrieval A comparison of text retrieval models Investigating aboutness axioms using information fields Probability kinematics in information retrieval Information calculus for information retrieval Query expansion using local and global document analysis Pivoted document length normalization Retrieval of complex objects using a four-valued logic A study of aboutness in information retrieval Comparing Boolean and probabilistic information retrieval systems across queries and disciplines (invited paper) A new theoretical framework for information retrieval Information flow On the role of logic in information retrieval Logical models in information retrieval A study of probability kinematics in information retrieval Text retrieval and filtering What is information discovery about? Fundamental properties of aboutness (poster abstract) Aboutness from a commonsense perspective Information retrieval and situation theory Discovering information flow suing high dimensional conceptual space Information Retrieval Modern Information Retrieval Informal Inference via Information Flow Information retrieval and Dempster-Shafer''s theory of evidence Using Default Logic in Information Retrieval Towards Functional Benchmarking of Information Retrieval Models Fundamental Properties of the Core Matching Functions for Information Retrieval Intelligent Text Handling Using Default Logic A commonsense aboutness theory for information retrieval modeling --CTR Tobias Blanke , Mounia Lalmas, Theoretical benchmarks of XML retrieval, Proceedings of the 29th annual international ACM SIGIR conference on Research and development in information retrieval, August 06-11, 2006, Seattle, Washington, USA Dawei Song , Jian-Yun Nie, Introduction to special issue on reasoning in natural language information processing, ACM Transactions on Asian Language Information Processing (TALIP), v.5 n.4, p.291-295, December 2006 D. Song , P. D. Bruza, Towards context sensitive information inference, Journal of the American Society for Information Science and Technology, v.54 n.4, p.321-334, February 15, Fang , ChengXiang Zhai, An exploration of axiomatic approaches to information retrieval, Proceedings of the 28th annual international ACM SIGIR conference on Research and development in information retrieval, August 15-19, 2005, Salvador, Brazil Raymond Y.K. Lau , Peter D. Bruza , Dawei Song, Belief revision for adaptive information retrieval, Proceedings of the 27th annual international ACM SIGIR conference on Research and development in information retrieval, July 25-29, 2004, Sheffield, United Kingdom Jian-Yun Nie , Guihong Cao , Jing Bai, Inferential language models for information retrieval, ACM Transactions on Asian Language Information Processing (TALIP), v.5 n.4, p.296-322, December 2006

functional benchmarking;inductive evaluation;aboutness;logic-based information retrieval

503036

From checking to inference via driving and dag grammars.

Abramov and Glck have recently introduced a technique called URA for inverting first order functional programs. Given some desired output value, URA computes a potentially infinite sequence of substitutions/restrictions corresponding to the relevant input values. In some cases this process does not terminate.In the present paper, we propose a new program analysis for inverting programs. The technique works by computing a finite grammar describing the set of all input that relate to a given output. During the production of the grammar, the original program is implicitly transformed using so-called driving steps. Whereas URA is sound and complete, but sometimes fails to terminate, our technique always terminates and is complete, but not sound. As an example, we demonstrate how to derive type inference from type checking.The idea of approximating functional programs by grammars is not new. For instance, the second author has developed a technique using tree grammars to approximate termination behaviour of deforestation. However, for the present purposes it has been necessary to invent a new type of grammar that extends tree grammars by permitting a notion of sharing in the productions. These dag grammars seem to be of independent interest.

INTRODUCTION The program-transformation techniques collectively called supercompilation have been shown to eectively handle problems that partial evaluation and deforestation cannot han- dle. The apparent strength of supercompilation stems from driving the object programs, that is, speculatively unfolding expressions containing variables, based on the possible executions described by the program. As an example of driving, consider the Haskell-like program where, by convention, the main denition serves as the interface to the program. It is not possible to execute this program because of the variables v and vs in main v vs, but we can drive the program, resulting in a process tree describing all possible computations of the program. For the above program, one possible process tree is - append (Cons v vs) Nil Cons v (append vs Nil) - append vs Nil let u= - append Nil Nil vs=Nil append (Cons x xs) Nil - vs=Cons x xs Cons v u - in (1) In general, the process tree is constructed as follows. The root of the process tree is labelled with right-hand side of the main denition. New leaves are added to the tree repeatedly by inspecting the label of some leaf, and either 1. Unfolding an outermost call (-). 2. Instantiating a variable that hinders unfolding (-). 3. Generalising by creating two disjoint labels (-). Whenever the label of a leaf is identical to the label of an ancestor (up to renaming of variables), that leaf is not unfolded further (-9 9 K-). A new (slightly more e-cient) program can easily be extracted from a process tree, namely Our main interest in this paper is to transform a checker into a description of the input that will satisfy the checker. That is, given a program that answers true or false, we will transform the program into a description of the input for which the checker answers true. The above activity is generally known as program inversion when the description of the satisfying input is yet another program. It is, however, a non-trivial task to perform program inversion, as the following example shows. Example 1. Consider a program that checks whether two lists are identical. The auxiliary functions isnil and iscons are needed because we only allow pattern matching on one argument at a time. The reason for having this restriction is that it associates every pattern match with a particular function denition. The result of inverting the above program w.r.t. true should be another program that produces all pairs of identical lists. However, it is unreasonable to assume that we can produce such a program: Even though it is easy to imagine a program that produces an innite stream of pairs of lists with identical spines, where should the elements come from? Based on their type, these elements could be enumerated, but such an enumeration clearly leads to non-termination in the general case. What is worse still, the imagined program will not per se give us a good description of the input set; we can merely get an indication of the input set by running it and observing its ever-increasing output. Instead of inverting a program, one might perform its computation backwards. A general method to perform inverse computation has been proposed by Abramov & Gluck [1], namely the Universal Resolving Algorithm (URA). The URA constructs a process tree for the object program, and produces from the process tree a potentially innite set of constraints on the uninstantiated input (variables xs and ys in the above example). Each constraint describes a set of input values by means of a substitution/restriction pair. The produced constraints are pairwise disjoint, in the sense that the sets they describe are pairwise disjoint. Variables can appear several times in each constraint, indicating identical values. For the above example, the URA would produce something like ([xs 7! Nil; ys 7! Nil]; []) ([xs 7! Cons x1 Nil; ys 7! Cons x1 Nil]; []) ([xs 7! Cons x1 (Cons x2 Nil); ys 7! Cons x1 (Cons x2 Nil)]; []) (2) Here the URA would never terminate. The merit of the URA is that it is sound and complete, so if it terminates, the result precisely captures the input set. In this paper we will sacrice soundness to develop an approach that always produces a nite description of the satisfying input. 1.2 Overview In Section 2, we present a formalisation of a certain kind of context-free grammars, namely dag grammars. For instance, the above checker can be described by the grammar> > < Nil Nil Cons x Cons The grammar consists of two productions, each formed as an acyclic directed graph (also known as a dag). The rst says that an S can be rewritten into a dag consisting of two single nodes labelled Nil. The second says that an S can be rewritten into a more complex dag with two roots. The two productions can be viewed as a nite representation of (2). Such dag grammars can precisely express the data and control ow of a program, something which is not possible with standard tree grammars. In Section 4, we present an automatic method to extract a dag grammar from a program. Conceptually, the extraction works in two steps: First we drive the object program to produce a process tree, second we extract a grammar from this process tree. A precursor for driving the program is a precise formulation of the semantics of our programming language, which we present in Section 3. The extracted dag grammars are approximations of the input set, and in Section 6, we consider various ways to improve the precision of our method. As an application, we will in Section 5 show that, given a type checker and a -term, it is possible to derive a type scheme for the -term. 2. DAG GRAMMARS We denote by fs0 the set containing s0 we g. For a binary relation * S T , we denote by D* the domain S of *. The set of deterministic binary relations (i.e., partial functions) from S to T is denoted by S - T , and such partial functions can be g. By S $ T we denote the set of bijections between S and T . Denition 1. Given a set S, a graph G over S consists of a label function lab 2 N - S and a connected-by relation means that node i's kth successor is node j and j's 'th predecessor is i. The relation should satisfy the properties that there is a label to each node, and that the successors and predecessors are numbered consecutively from 0. Formally, When the order of successors and predecessors is immaterial, we simply use - as a binary relation and write i - j whenever 9k ' 2 N [i k ' - j]. In the following, we will use subscripts like lab G or - G when it is otherwise not obvious which graph we refer to. By - we denote the transitive closure of - . In general, we will superscript any binary relation with + to denote its transitive closure, and we will superscript with to denote its re exive closure. Denition 2. A graph G is a dag when it contains no cycles (i.e., @i 2 N [i i]). For dags, it is natural to talk of roots and leaves: roots G G i] leaves G G j] Two dags D and E are equivalent, denoted D E, when D and E can be transformed into one another by renumbering the nodes, that is, The set of dags over S is denoted DS . Example 2. Each of the two structures (3) depicts an equivalence class of dags over fS Nil Cons xg | in the sense that the structure describes a family of equivalent dags | because the node set ( N) is left unspecied; the order of successors and predecessors, however, is specied by adopting the convention that heads and tails of arcs are ordered from left to right. A concretisation of the right dag is, for example, and we have that leaves = f3 4g and roots = f0g. Denition 3. Given a set , a dag grammar over is a set of nite dags over . The set of dag grammars over is denoted G. Example 3. The two dags (3) comprise a dag grammar over fS x Nil Consg. A dag grammar describes a dag language, in which every dag has been generated by a number of rewrites on an initial dag. Before we give the formal denition of rewrites and languages, an example is in order. Example 4. The dag can be rewritten by the dag grammar (3) as Nil Nil Cons x Cons Cons x Nil Cons Cons x Cons x Cons Cons The symbol is here used to maintain an order between an otherwise unordered set of roots. Below you will nd the formal denition of graph grammar rewrites. The example above hopefully illustrates how such rewrites work. Informally, a dag can be rewritten if it has a leaf node i that matches a root node j in a dag in the graph grammar. By matches, we mean that i and j have the same label, and that the number of predecessors of i matches the number of successors of j. The result of the rewrite is that the leaf i and root j dissolve, as it were, and the predecessors of i become the predecessors of the successors of j, in the right order, as illustrated below.6 6k1 kn 6 6 6k1 kn denitions generalise the above notion of rewriting to several leaves We will need this generality in Section 4. Denition 4. We denote by S P the set of all nite sequences over set S. Both sequences and tuples are denoted by angles h i. Denition 5. Given dags D and E, we dene match ng (lab and maxmatch Denition 6. Given sets S and T , if S and T are disjoint, we write S T , and in that case the disjoint union S ] T is dened (and undened, otherwise). Given a relation * S T and a set S 0 , we dene the removal of S 0 from * as * nS 0 . The disjoint union of two binary relations is dened if their domains are disjoint. In the following, we carefully pay attention to the exact set of nodes ( N) that each particular dag comprises. To avoid node clashes when we rewrite graphs, we use the fact that, given a nite dag G 0 , there are innitely many equivalent dags G. Denition 7. Given a dag grammar 2 G , the one- step-rewrite relation ! D D is dened by 9 hI Ji 2 maxmatch D G6 6 6 I Denition 8. Given a dag grammar and an initial dag I, the dag language LI is the set of normal forms: LI We now have a grammar framework that can readily express sets of input. The next step is to show how a dag grammar can be extracted from a program. We start by dening the semantics of the object language, and, on top of the semantics, we then build the machinery that calculates the desired grammars. 3. OBJECT LANGUAGE Our object language is a rst-order functional language with pattern matching and operators for term equality and boolean and. Denition 9. Given a set of variables X, function names F , pattern-function names G, and constructors C (including the constants true and false), we dene d G. As usual, we require that no variable occurs more than once in the left-hand side of a function denition (left-linear), and that all variables in the right-hand side of a function definition is a subset of those in the left-hand side (closed). Finally we require that the patterns dened for each g-function are pairwise distinct modulo variable names (non- overlapping); in particular, a g-function can have at most one \catch-all" pattern. In concrete programs we use a Haskell-like syntax, including data-type denitions. The above syntax can be viewed as the intermediate language obtained after type checking, which rules out terms like Nil & Nil and Nil true. Denition 10. Given a term t, we denote by V t the variables of t, collected from left to right. Term t is Example 5. (Triple x y Denition 11. A function 2 X - T can be regarded as a substitution T - T in the usual way, and, for t 2 T , we will write t for the result of such a substitution. Given program p, writing, say, g (c means that in p there is a function denition identical to t, up to variable naming. We will now give the normal-order semantics of the object language by dening a small-step relation that takes redexes into their contracta. We can separate redexes from their contexts by the following syntactic classes. if if true Figure 1: Normal-order semantics Denition 12. e O 3 By d([t]) we denote the result of replacing ([ ]) in d with term t. Any ground term t is either a value or can be uniquely decomposed into a context and a redex (i.e., decomposition property allows us to dene the semantics as below. The semantics imposes left-to-right evaluation, except for the equality operator, which evaluates both its arguments until two outermost constructors can be compared. Denition 13. Given a program p, the small-step semantics of p is dened by the smallest relation ! RT on closed terms as dened by Fig. 1. In the following we will use subscripts like ! p when it is otherwise not obvious which program we refer to. To get the full semantics of the language, we simply need to close the !-relation under all contexts. The semantics is deterministic: Lemma 1. Proof sketch. By induction and case analysis of the syntactic structure of terms: Each term either cannot be decomposed into a context and a redex, or it can be uniquely decomposed, in which case the redex has at most one small-step derivation. 4. EXPLICITATION The previous section aloows us to deal with execution of ground terms. To be able to drive a program, however, we need to handle terms containing variables. We use the following syntactic class, in combination with the previously dened ones, to describe all terms that cannot be decomposed into contexts and redexes. Denition 14. As for the variables of a let-term, V (let The use of let-terms will be described below. As in the previous section, we obtain a unique-decomposition property: Any term t is either a value, can be uniquely decomposed into a context and a redex, or can be uniquely decomposed into a context and a speculative (i.e., d([s])). The extra syntactic class s enables us to identify terms that need to be driven (i.e., instantiated). In supercompilation, driving is used to obtain a process tree from the object program. The process tree serves two orthogonal purposes: It keeps track of data and control ow (essentially variable instantiations and recursive calls), and it provides a convenient way to monitor the driving process and perform the generalisations needed to ensure a - nite process tree. When a generalisation is needed for a t, t is replaced by a term \let " such that . The point of making such a generalisation is to be able to treat t1 and t2 independently. For an exam- ple, you might want to revisit (1) in the introduction. For a thorough introduction to these techniques, see Srensen & Gluck [13]. In our approach to program inversion, called explicitation, we will assume that the generalisations necessary to ensure termination have been computed in advance by an o-line generalisation analysis. To be more specic, we assume that some terms have been replaced by terms of the form let t1 in t2 in the program of interest. With respect to the data and control ow of the program, the ow can be expressed by dag grammars, which we will elaborate on later in this section. Since we have thus eliminated the need for a process tree, we will, to keep things simple, drive the object program without constructing the actual process tree. The construction of a process tree, although important in practice, is not the essence in our approach. Remark 1. We should note here that the existence of an o-line generalisation analysis is not essential: The explicitation process described in the following could incorporate well-known non-termination detection and perform the necessary generalisations. 1 But because such an incorporation would induce unnecessary clutter in our presentation, we will concentrate on the description of how to extract a dag- grammar by driving. From a bird's-eye view, the explicitation process of a single branch of speculative execution works as follows. Starting with the main term of the object program, a dag grammar is gradually built up by driving the term: each time a speculative (cf. Def. 14) hinders normal reduction, we perform an instantiation to both the term and the dag gram- mar, such that reduction of the term can proceed and the re ects the structure of the input variables. In fact, each driving step of a term results in a new production in the grammar, such that every term we meet while driving the program has its own production. When we meet a term which we have seen before, a loop is introduced into the grammar, and driving of the term stops. A term t that cannot be driven any further is compared to the output we desired, namely non-false output: If t is false, then the result is an empty grammar; otherwise it is the accumulated grammar . In general, we parameterise the explicitation process over a discrimination function h. For our purpose, In the above fashion, we can produce a grammar for every speculative execution of the program: Each possible instantiation of a term gives rise to a slightly dierent term and grammar. The nal grammar is then the union of the grammars produced for all executions. As an example of an instantiation on a dag grammar, consider the dags same xs ys 6 6 6xs ys Cons iscons ys x xs7 7 same xs ys Cons iscons ys x xs7 D represents a call to same where the arguments are un- known. E represents the body of same: a pattern match on the variable xs and a call to iscons with three arguments (cf. Example 1). The order of the arrows is important, since it have presented a method for preventing non-termination and performing generalisations of dangerous terms, as it were, based on certain quasi-orders. With a few extensions, this method can be applied to our lan- guage. We can prove that such an extended method will indeed guarantee termination of the explicitation, if we apply the general framework for proving termination of program transformers presented by the second author [14]. denes the data ow. If we view fEg as a dag grammar, D can be rewritten into D 0 by means of fEg, as shown above. Formally, Denition 15. Given dags D and E, the dag substitution fEgD is dened as (cf. Def. 5) D; otherwise. Substitutions are extended to grammars in the obvious way: To construct dags from terms, we use the following shorthands Denition 16. Given a term t, and ? t The full explicitation process will be explained in detail be- low. Formally, it is summed up in the following denition. Denition 17. Given a program p and a function h 2 the explicitation of p is a dag grammar Eh [[p]], as described in Fig. 2. The following explanation of the explicitation process carries a number of references that hopefully will guide the understanding of Fig. 2. Explicitation starts with the main term t of the program, an empty dag grammar , and an empty set of previously seen terms In each step, we inspect the structure of the current term t, and either stop, or add a production for t to and make a new step with t added to the seen-before set. If t has been seen before, a production for t is already present in the grammar , so we return the accumulated unchanged. - If a renaming of t has been seen before (captured by a bijection ), we introduce recursion in the grammar by adding a production that connects t to the (pre- viously seen) t, respecting the number of variables. - If a redex can be unfolded using the standard semantics (cf. Defs. 12 & 13), a production linking t to its unfolding is added to , and the process continues with the unfolding. - If a generalised term hinders unfolding, that is t1 in t2 ]), d([t 2 ]) and t1 are processed independently. There- fore, a production is added to the grammar such that t is linked to both d([t 2 ]) and t1 . This production will have some dangling roots 2 (namely x and V t1 \ V t2 ) which re ect that the data ow is approximated. Because the traces of t1 will tell us nothing about the output of t, the function h (cf. (4)) Unmatched roots are allowed in dag rewrites, cf. Def. 7. Eh let [ ftg in - if tthen then [< @> < c AC A @> < r x y AC A and t and x c and t x true #) and t x false #) Figure 2: Explicitation by driving that is supposed to discriminate between various output is replaced by the function xy:y which does not discriminate: It always returns the accumulated grammar. - For a pattern matching function, the process is continued for all dened patterns. For each pattern q, we substitute the arguments into the matching body, and put it back into the context, which in turn receives the instantiation fx 7! qg, and we add to the grammar a production re ecting this instantiation. For comparisons, there are three cases. - The rst simply makes sure that variables are on the left side of the com- parison. That settled, - if the right-hand side is another variable, two possibilities are explored: Either the comparison will fail, and hence we replace the speculative with false; or the comparison will succeed, and we replace the speculative with true and update the grammar and the context to re ect that both sides must be identical. In the grammar, the equal variables are coalesced by means of a special symbol r, which is needed to maintain the invariant that the in/out degree of terms correspond to the number of distinct variables. - If the right-hand side is an n-ary constructor, either the comparison will fail (as above), or it will succeed, in which case we will propagate that the variable must have a particular outermost constructor, of which the children must be tested for equality with the children of the constructor. - If a boolean expression depends on a variable, then the variable will evaluate to either true or false, and this information is propagated to each branch. - Terms that cannot be driven any further (i.e., t 2 V ) are fed to the function h, which in turn decides what to do with the accumulated grammar. Example 6. Let p be the program from Example 1 and h dened as (4). The explicitation Eh is depicted in Fig. 3. The grammar produced is fA B D F I Jg. The derived grammars are sub-optimal: Most of the productions are intermediate, in the sense that they can be directly unfolded, or in-lined as it were. We say that a grammar is normalised if it contains no intermediate productions, and we can easily normalise a grammar. Denition 18. A dag grammar can be normalised, denoted b , as follows. roots G leaves G Example 7. If we normalise the grammar from Example 6, we almost get the grammar we promised in the introduction:4 same xs ys same xs ys Cons Cons r same xs ys In this particular grammar, the bookkeeping symbol r can be eliminated by yet another normalisation process, if so desired. Given that the object program contains the necessary gen- eralisations, the lemma below is not surprising. However, if we incorporated a standard on-line termination strategy into the explication process, as explained in Remark 1, the following lemma would still hold. Lemma 2. Any explicitation produces a nite grammar. Proof. The process tree of the program is nite (since the program contains the necessary generalisations), and thus a nite number of dags will be produced, assuming the h function is to be total and computable. More interestingly, the explicitation returns a dag grammar that contains all solutions. To express such a completeness theorem, we need a precise formulation of the set of terms generated by a dag grammar, as given below. Infor- mally, a term is extracted from a dag simply by following the edges from left to right, collecting the labels | except when the label is a variable or the bookkeeping symbol r: Every variable is given a distinct name, and r is treated as an indirection and left out in the term. Denition 19. Given a dag grammar , a label S with arity n, and a set of variables g, the term language T n S is the set of tuples of terms that can be extracted from the underlying dag language: I =4 0: S5 => < and fhi 0 '0 We can now relate all possible executions of the program to the set of terms generated by its explicitation. Eh [[same xs ys fsame xs ysg Eh f(same xs ys) (isnil ys)g Eh f(same xs ys) (isnil ys)g Eh [[iscons ys x xs fsame xs ysg Eh f(same xs ys) (iscons ys x xs)g f(same xs ys) (iscons ys x xs)g f:::(if (xy) (same xs ys) false)g f:::(if (xy) (same xs ys) false) (if false (same xs ys) f:::(if (xy) (same xs ys) false)g f(same xs ys)::: 6 6same xs ys Nil 6 6isnil ys Cons iscons ys x xs Nil x xs7G 2 (same xs ys)(false) x y if false (same xs ys)(false)7 7 if (same xs ys)(false) r x if true (same xs ys) isnil ys same xs ys Cons iscons ys x xs7 7 75 26 6 6 6 6iscons ys x xs Cons if 6 6if false (same xs ys) xs ys7 6 6if true (same xs ys) same xs ys7 Figure 3: Explicitation of the same-program Theorem 1. Given a program p where main holds that 5. APPLICATION: TYPE INFERENCE We now show that a type checking algorithm can be transformed into a type inference algorithm by explicitation. Specif- ically, we check that a -calculus expression has a given type in a given environment, using the standard relation As an example, consider the expression meaning \what is the type of x:y:z:(xz)(yz) in the empty environment?". We would expect the answer to be something like We will now perform explicitation of the above expression. The type checker below takes an encoding of a proof tree P , an expression M , and an environment , and returns true if P is indeed a valid proof of the type of M in . In this encoding, 3 x:y:z:(xz)(yz) becomes and ? becomes Empty. If we want to explicitate the above expression, we can add the denition main to the implementation of the type checker (which we then refer to as the specialised typechecker): data 3 | Abs Int Term 4 data 5 data Proof = Infer Premise Type 6 data 7 | Intro Proof 9 expchk (Var x) y z match z x y 14 elimchk (Elim x y) z v w 3 The implementation assumes a primitive integer type Int. 19 arrowcheck v x y z w match (Bind x y z) v 22 if (v == x) (w == y) (match z v w) match x y arrowcheck (Ar x y) z v w 26 y == conclusion z 28 conclusion (Infer x 29 if false x Explicitation of the specialised typechecker gives a term language that consists of a single pair <(Ar (Ar x (Ar y z)) (Ar (Ar x y) (Ar x z))) (.)> The rst element is indeed the encoding of type (5), and the second is the proof tree, which we have left out. 6. IMPROVING SOUNDNESS Any automatic method for inversion of Turing-complete programs will have to make a compromise with respect to com- pleteness, soundness and termination. We have made a compromise that will result in possibly unsound results: The explicitation of a program can produce a grammar that generates too many terms. From a practical point of view, we feel that this is the right compromise: It is better to obtain an approximation than not obtaining an answer at all. Moreover, explicitation can produce grammars that precisely identies the input set, as seen from the two examples is previous sections, which indicates that explicitation is tight enough for at least some practical problems. However, it still remains to identify exactly what causes loss of soundness in the general case. Our method is inherently imprecise for three reasons. Generalisations cause terms to be split up in separate parts (by means of let-terms). This prevents instantiations in the left branch of the process tree to aect the right branch, and vice versa. Negative information is not taken in to account while driving. For example, driving the speculative term x y will not propagate the fact that x 6= y to the right branch of the process tree (although the fact that propagated to the left branch, by means of a substitution). Moreover, the dag grammars cannot represent negative information. Occur check is not performed on speculative terms like that is, a situations where for some i n is not discovered. Obviously, such an equality would never imply that Similarly, the symmetric property that x x never implies x 6= x is not used either. The occur check and its counterpart can easily be incorporated into rules - and -, respectively, in Fig. 2. Interest- ingly, explicitation of the type checker (specialised to a - would be sound, had these checks been incorporated. As for negative information, we have described how to handle and propagate such information in another paper [11]. Incorporating negative information as proposed in that paper would be a simple task. Incorporating negative information into the dag grammars, however, would destroy their simplicity and thus severely cripple their usability. Hence, generalisation and the inability of dag grammars to represent negative information are the true culprits. One could therefore imagine a variant of the explicitation algorithm | lets call it EXP | where one has incorporated the extensions we have suggested above. To improve soundness of EXP, one should target the way generalisations are carried out. We are now in a position to conjecture the following roundabout relationship between EXP and the URA [1] described in the introduction: Given a program p without generalisa- tions, EXP terminates on p whenever URA terminates on p. Moreover, if the result of URA contains no restrictions (neg- ative information), the resulting grammar of EXP is sound. 7. RELATED WORK Program inversion In the literature, most program inversion is carried out by hand. One exception is Romanenko [9], who describes a pseudo algorithm for inverting a small class of programs written in Refal. In a later paper, Romanenko [10] extends the Refal language with non-deterministic construct, akin to those seen in logic programming, to facilitate exhaustive search. He also considers inverting a program with respect to a subset of its parameters, so-called partial inversion. We would like to extend our method to handle partial inversion, but as of this writing it is unclear how this should be done. The only fully automated program inversion method we know of, is the one presented by Abramov & Gluck [1]. Their method constructs a process tree from the object pro- gram, and solutions are extracted from the leaves of the tree in form of substitution/restriction pairs. The process trees are perfect (Gluck & Klimov [3]) in the sense that no information is lost in any branch (completeness), and every branching node divides the information in disjoint sets (soundness). Unfortunately, soundness together with completeness implies non-termination for all but a small class of programs. The method we have presented here sacrices soundness for termination. What is common to both ours and the above methods is that they all build upon the ground breaking work of Turchin and co-workers. The rst English paper that contains examples of program inversion by driving seems to be [15]. For more references, see Srensen & Gluck [4]. Grammars The idea of approximating functional programs by grammars is not new. For instance, Jones [5] presents a ow analysis for non-strict languages by means of tree gram- mars. Based on this work, the second author has developed a technique using tree grammars to approximate termination behaviour of deforestation [12]. Tree grammars, how- ever, cannot capture the precise data and control ow: By denition, branches in a tree grammar are developed inde- pendent, which renders impossible the kind of synchronisation we need between variable and function calls. The dag grammars we have presented, precisely capture the data and control ow for any single speculative computation trace; synchronisation can only be lost when several alternative recursive productions exist in the grammar. It seems possible to devise a more precise ow analysis based on dag grammars, which could be used along the lines of [12]. Indeed, the way we use dag grammars to trace data and control ow, turns out to have striking similarities with the use of size-change graphs, presented by Lee, Jones & Ben- Amram [7]. Size-change graphs approximate the data ow from one function to another, by capturing size-changes of parameters. The dag-rewrite mechanism we have presented turns out to have a lot in common with the fan-in/fan-out rewrites presented by Lamping [6], in the quest for optimal reduction in the -calculus. The fan-in/fan-outs represent a complex way of synchronising dierent parts of a -graph, whereas our dag rewrites only perform a simple use-once synchroni- sation. The rewriting mechanism is also akin to graph substitution in hyper-graph grammars (see Bauderon & Courcelle [2] for a non-category-theory formulation), except that we allow any number of leaves to be rewritten and do not allow inner nodes to be rewritten. Strength-wise, hyper-graph grammars are apparently equivalent to attribute grammars. At present, we are not sure of the generative power of dag grammars. 8. CONCLUSION AND FUTURE WORK We have developed a method for inverting a given program with respect to a given output. The result of the inversion is a nite dag grammar that gives a complete description of the input: \Running" the dag grammar produces a (possibly innite) set of terms that will contain all tuples of terms that result in the given output. The method seems to be particularly useful when the object program is a checker, that is, one that returns either true or false. We have exemplied this by deriving a type scheme from a type checker and a given -term, thus eectively synthesising a type inference algorithm. Following this line, one could imagine a program that checks whether a document is valid XML. Inverting this program would result in a dag grammar, which could then be compared to a speci- cation for valid XML, as a means of verifying the program. Inverting the program when specialised to a particular doc- ument, would result in a Document Type Denition. One could even imagine inverting a proof-carrying-code verier [8] with respect to a particular program, thus obtaining a proof skeleton for the correctness of the code. Further experiments with the above kinds of applications should be carried out to establish the strength and usability of our method. 9. --R Graph expressions and graph rewritings. Flow analysis of lazy higher-order functional programs An algorithm for optimal lambda calculus reduction. The size-change principle for program termination Compiling with Proofs. The generation of inverse functions in Refal. Inversion and metacomputation. de --TR An algorithm for optimal lambda calculus reduction Inversion and metacomputation Convergence of program transformers in the metric space of trees The size-change principle for program termination Introduction to Supercompilation On Perfect Supercompilation Occam''s Razor in Metacompuation A Roadmap to Metacomputation by Supercompilation Semantic definitions in REFAL and the automatic production of compilers The Universal Resolving Algorithm Grammar-Based Data-Flow Analysis to Stop Deforestation Compiling with proofs --CTR Siau-Cheng Khoo , Kun Shi, Output-constraint specialization, Proceedings of the ASIAN symposium on Partial evaluation and semantics-based program manipulation, p.106-116, September 12-14, 2002, Aizu, Japan Aaron Tomb , Cormac Flanagan, Automatic type inference via partial evaluation, Proceedings of the 7th ACM SIGPLAN international conference on Principles and practice of declarative programming, p.106-116, July 11-13, 2005, Lisbon, Portugal Siau-Cheng Khoo , Kun Shi, Program Adaptation via Output-Constraint Specialization, Higher-Order and Symbolic Computation, v.17 n.1-2, p.93-128, March-June 2004 Morten Heine Srensen , Jens Peter Secher, From type inference to configuration, The essence of computation: complexity, analysis, transformation, Springer-Verlag New York, Inc., New York, NY, 2002 Principles of inverse computation and the Universal resolving algorithm, The essence of computation: complexity, analysis, transformation, Springer-Verlag New York, Inc., New York, NY, 2002

supercompilation;program inversion;inference

503042

Mixed-initiative mixed computation.

We show that partial evaluation can be usefully viewed as a programming model for realizing mixed-initiative functionality in interactive applications. Mixed-initiative interaction between two participants is one where the parties can take turns at any time to change and steer the flow of interaction. We concentrate on the facet of mixed-initiative referred to as 'unsolicited reporting' and demonstrate how out-of-turn interactions by users can be modeled by 'jumping ahead' to nested dialogs (via partial evaluation). Our approach permits the view of dialog management systems in terms of their support for staging and simplifying inter-actions; we characterize three different voice-based interaction technologies using this viewpoint. In particular, we show that the built-in form interpretation algorithm (FIA) in the VoiceXML dialog management architecture is actually a (well disguised) combination of an interpreter and a partial evaluator.

INTRODUCTION Mixed-initiative interaction [8] has been studied for the past years in the areas of artificial intelligence (AI) planning [17], human-computer interaction [5], and discourse analysis [6]. As Novick and Sutton point out [13], it is 'one of those things that people think that they can recognize when they see it even if they can't define it.' It can Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. PEPM '02, Jan. 14-15, 2002 Portland, OR, USA be broadly viewed as a flexible interaction strategy between participants where the parties can take turns at any time to change and steer the flow of interaction. Human conversations are typically mixed-initiative and, interestingly, so are interactions with some modern computer systems. Consider the two dialogs in Figs. 1 and 2 with a telephone pizza delivery service that has voice-recognition capability (the line numbers are provided for ease of refer- ence). Both these conversations involve the specification of a {size,topping,crust} tuple to complete the pizza ordering procedure but di#er in important ways. In Fig. 1, the caller responds to the questions in the order they are posed by the system. The system has the initiative at all times (other than, perhaps, on line 0) and such an interaction is thus referred to as system-initiated. In Fig. 2, when the system prompts the caller about pizza size, he responds with information about his choice of topping instead (sausage; see line 3 of Fig. 2). Nevertheless, the conversation is not stalled and the system continues with the other aspects of the information gathering activity. In particular, the system registers that the caller has specified a topping, skips its default question on this topic, and repeats its question about the size (see line 5 of Fig. 2). The caller thus gained the initiative for a brief period during the conversation, before returning it to the system. Such a conversation that mixes system-initiated and user-initiated modes of interaction is said to be mixed-initiative. 1.1 Tiers of Mixed-Initiative Interaction It is well accepted that mixed-initiative provides a more natural and personalized mode of interaction. A matter of debate, however, are the qualities that an interaction must possess to merit its classification as mixed-initiative [13]. In determining who has the initiative at a given point in an interaction can itself be a contentious issue! The role of intention in an interaction and the underlying task goals also a#ect the characterization of initiative. We will not attempt to settle this debate here but a few preliminary observations will be useful. One of the basic levels of mixed-initiative is referred to as unsolicited reporting in [3] and is illustrated in Fig. 2. In this facet, a participant provides information out-of-turn (in our case the caller, about his choice of topping). Furthermore, the out-of-turn interaction is not agreed upon in advance by the two participants. Novick and Sutton [13] stress that the unanticipated nature of out-of-turn interactions is important and that mere turn-taking (perhaps in a hardwired order) does not constitute mixed-initiative. Finally, notice that in Fig. 2 there is a resumption of the original questioning task Caller: #calls Joe's Pizza on the phone# calling Joe's pizza ordering system. What size pizza would you like? 3 Caller: I'd like a medium, please. 4 System: What topping would you like on your pizza? 5 Caller: Pepperoni. 6 System: What type of crust do you want? 7 Caller: Uh, deep-dish. 8 System: So that is a medium pepperoni pizza with deep-dish crust. Is this correct? 9 Caller: Yes. (conversation continues to get delivery and payment information) Figure 1: Example of a system-directed conversation. Caller: #calls Joe's Pizza on the phone# calling Joe's pizza ordering system. What size pizza would you like? 3 Caller: I'd like a sausage pizza, please. 4 System: Okay, sausage. 5 System: What size pizza would you like? 6 Caller: Medium. 7 System: What type of crust do you want? 8 Caller: Deep-dish. 9 System: So that is a medium sausage pizza with deep-dish crust. Is this correct? (conversation continues to get delivery and payment information) Figure 2: Example of a mixed-initiative conversation. once processing of the unsolicited response is completed. In other applications, an unsolicited response might shift the control to a new interaction sequence and/or abort the current interaction. Another level of mixed-initiative involves subdialog invo- cation; for instance, the computer system might not have understood the user's response and could ask for clarifications (which amounts to it having the initiative). A final, sophisticated, form of mixed-initiative is one where participants negotiate with each other to determine initiative (as opposed to merely 'taking the initiative') [3]. An example is given in Fig. 3. In addition to models that characterize initiative, there are models for designing dialog-based interaction systems. Allen et al. [2] provide a taxonomy of such software models - finite-state machines, slot-and-filler structures, frame-based methods, and more sophisticated models involving planning, agent-based programming, and exploiting contextual information. While mixed-initiative interaction can be studied in any of these models, it is beyond the scope of this paper to address all or even a majority of them. Instead, we concentrate on the view of (i) a dialog as a task-oriented information assessment activity requiring the filling of a set of slots, (ii) where one of the participants in the dialog is a computer system and the other is a human, and (iii) where mixed-initiative arises from unsolicited reporting (by the human), involving out-of-turn interactions. This characterization includes many voice-based interfaces to information (our pizza ordering dialog is an example) and web sites modeling interaction by hyperlinks [15]. In Section 2, we show that partial evaluation can be usefully viewed as a programming model for such applications. Section 3 presents three di#erent voice-based interaction technologies and analyzes them in terms of their native support for mixing initiative. Finally, Section 4 discusses other facets of mixed-initiative and mentions other software models to which our approach can be extended. 2. PROGRAMMINGAMIXED-INITIATIVE APPLICATION Before we outline the design of a system such as Joe's Pizza (ref. Figs. 1 and 2), we introduce a notation [7, 11] that captures basic elements of initiative and response in an interaction sequence. The notation expresses the local organization of a dialog [14] as adjacency pairs; for instance, the dialog in Fig. 1 is represented as: The line numbers given below the interaction sequence refer to the utterance numbers in Fig. 1. The letter I denotes who has the initiative - caller (c) or the system and the letter R denotes who provides the response. It is easy to see from this notation that the dialog in Fig. 1 consists of five turns and that the system has the initiative for the last four turns. The initial turn is modeled as the caller having the initiative because he or she chose to place the phone call in the first place. The system quickly takes the initiative after playing a greeting to the caller (which is modeled here as the response to the caller's call). The subsequent four interactions then address three questions and a confirmation, all involving the system retaining the initiative (Is) and the caller in the responding mode (Rc). (with apologies to O. Henry) Husband: Della, Something interesting happened today that I want to tell you. Wife: I too have something exciting to tell you, Jim. Husband: Do you want to go first or shall I tell you my story? Figure 3: Example of a mixed-initiative conversation where initiative is negotiated. Caller: #calls Joe's Pizza on the phone# calling Joe's pizza ordering system. What size pizza would you like? 3 Caller: I'd like a sausage pizza, medium, and deep-dish. 4 System: So that is a medium sausage pizza with deep-dish crust. Is this correct? 5 Caller: Yes. (conversation continues to get delivery and payment information) Figure 4: Example of a mixed-initiative conversation with a frequent customer. Likewise, the mixed-initiative interaction in Fig. 2 is represented as: In this case, the system takes the initiative in utterance 2 but instead of responding to the question of size in utterance 3, the caller takes the initiative, causing an 'insertion' to occur in the interaction sequence (dialog) [11]. The system responds with an acknowledgement ('Okay, sausage.') utterance 4. This is represented as the nested pair (Ic Rs) above. The system then re-focusses the dialog on the question of pizza size in utterance 5 (thus retaking the ini- tiative). In utterance 6 the caller responds with the desired size (medium) and the interaction proceeds as before, from this point. There are various other possibilities for mixing initiative. For instance, if a user is a frequent customer of Joe's Pizza, he might take the initiative and specify all three pizza attributes on the first available prompt, as shown in Fig. 4. Such an interaction would be represented as: Notice that even though this dialog consists of only three turns it constitutes a complete instance of interaction with the pizza ordering service. The notation aids in understanding the structure and staging of interaction in a dialog. By a post-analysis of all interaction sequences described in this manner, we find that utterances 0 and 1 have to proceed in order. Utterances dealing with selection of {size,topping,crust} can then be nested in any order and provide interesting opportunities for mixing initiative. Finally, the utterances dealing with confirmation of the user's request can proceed only after choices of all three pizza attributes have been made. While the notation doesn't directly reflect the computational processing necessary to achieve the indicated struc- ture, it can be used to express a set of requirements for a dialog system design. There are 13 possible interaction sequences (discounting permutations of attributes specified in a given utterance): 1 possibility of specifying everything in one utterance, 6 possibilities of specification in two ut- terances, and 6 possibilities of specification in three utter- ances. If we include permutations, there are 24 possibilities (our calculations do not consider situations where, for in- stance, the system doesn't recognize the user's input and reprompts for information). Of these possibilities, all but one are mixed-initiative sequences. Many programming models view mixed initiative sequences as requiring some special attention to be accommodated. In particular, they rely on recognizing when a user has provided unsolicited input 1 and qualify a shift-in-initiative as a 'transfer of control.' This implies that the mechanisms that handle out-of-turn interactions are often di#erent from those that realize purely system-directed interactions. Fig. 5 (left) describes a typical software design. A dialog manager is in charge of prompting the user for input, queueing messages onto an output medium, event processing, and managing the overall flow of interaction. One of its inputs is a dialog script that contains a specification of interaction and a set of slots that are to be filled. In our pizza example, slots correspond to placeholders for values of size, topping, and crust. An interpreter determines the first unfilled slot to be visited and presents any prompts for soliciting user in- put. A responsive input requires mere slot filling whereas unsolicited inputs would require out-of-turn processing (in- volving a combination of slot filling and simplification). In turn, this causes a revision of the dialog script. The interpreter terminates when there is nothing left to process in the script. While typical dialog managers perform miscellaneous functions such as error control, transferring to other scripts, and accessing scripts from a database, the architecture in Fig. 5 (left) focusses on the aspects most relevant to our presentation. Our approach, on the other hand, is to think of a mixed-initiative dialog as a program, all of whose arguments are passed by reference and which correspond to the slots comprising information assessment. As usual, an interpreter in the dialog manager queues up the first applicable prompt to an output medium. Both responsive and unsolicited inputs by a user now correspond (uniformly) to values for arguments; they are processed by partially evaluating the program with respect to these inputs. The result of partial evaluation is another dialog (simplified as a result of user input) which is handed back to the interpreter. This novel We use the term 'unsolicited input' here to refer to expected but out-of-turn inputs as opposed to completely unexpected (or out-of-vocabulary) inputs. out-of-turn processing slot filling Interpreter user input? Dialog Manager user input Dialog script unsolicited responsive nothing to process STOP Interpreter user input? partial evaluator Dialog Manager user input Dialog script to process yes nothing Figure 5: Designs of software systems for mixed-initiative interaction. (left) Traditional system architecture, distinguishing between responsive and unsolicited inputs. (right) Using partial evaluation to handle inputs uniformly. { if (unfilled(size)){ /* prompt for size */ if (unfilled(topping)){ /* prompt for topping */ if (unfilled(crust)){ /* prompt for crust */ Figure Modeling a dialog script as a program parameterized by slot variables that are passed by reference. design is depicted in Fig. 5 (right) and a dialog script represented in a programmatic notation is given in Fig. 6. Partial evaluation of Fig. 6 with respect to user input will remove the conditionals for all slots that are filled by the utterance (global variables are assumed to be under the purview of the interpreter). The reader can verify that a sequence of such partial evaluations will indeed mimic the interaction sequence depicted in Fig. 2 (and any of the other mixed-initiative sequences). Partial evaluation serves two critical uses in our design. The first is obvious, namely the processing of out-of-turn interactions (and any appropriate simplifications to the dialog script). The more subtle advantage of partial evaluation is its support for staging mixed-initiative interactions. The mix-equation [9] holds for every possible way of splitting inputs into two categories, without enumerating and 'trap- ping' the ways in which the computations can be staged. For instance, the nested pair in Fig. 2 is supported as a natural consequence of our design, not by anticipating and reacting to an out-of-turn input. Another way to characterize the system designs in Fig. 5 is to say that Fig. 5 (left) makes a distinction of responsive versus unsolicited inputs, whereas Fig. 5 (right) makes a more fundamental distinction of fixed-initiative (interpre- tation) versus mixed-initiative (partial evaluation). In other words, Fig. 5 (right) carves up an interaction sequence into (i) turns that are to be handled in the order they are modeled (by an interpreter), and (ii) turns that can involve mixing of initiative (handled by a partial evaluator). In the latter case, the computations are actually used as a representation of interactions. Since only mixed-initiative interactions involve multiple staging options and since these are handled by the partial evaluator, our design requires the least amount of specification (to support all interaction sequences). For instance, the script in Fig. 6 models the parts that involve mixing of initiative and helps realize all of the 13 possible interaction sequences. At the same time it does not model the confirmation sequence of Fig. 2 because the caller cannot confirm his order before selecting the three pizza attributes! This turn must be handled by straightforward interpretation To the best of our knowledge, such a model of partial evaluation for mixed-initiative interaction has not been proposed before. While computational models for mixed-initiative interaction remain an active area of research [8], such work is characterized by keywords such as 'controlling mixed-initiative 'knowledge representation and reasoning strategies,' and `multi-agent co-ordination.' There are even projects that talk about 'integrating' mixed initiative interaction and partial evaluation to realize an architecture for planning and learning [17]. We are optimistic that our work has the same historical significance as the relation between explanation-based generalization (a learning technique in AI) and partial evaluation established by van Haremelen and Bundy in 1988 [16]. 2.1 Preliminary Observations It is instructive to examine the situations under which a concept studied in a completely di#erent domain is likened to partial evaluation. Establishing a resemblance to partial evaluation is usually done by mapping from an underlying idea of specialization or customization in the original do- main. This is the basis in [16] where specialization of domain theories is equated to partial evaluation of programs. The motivating theme is one of re-expressing the given program (domain theory) in an e#cient but less general form, by recognizing that parameters have di#erent rates of variation [9]. This theme has also been the primary demonstrator in the partial evaluation literature, where inner loops call-in confirm size d2 d1 d3 topping crust Figure 7: Modeling requirements for unsolicited reporting as traversals of a graph. interpretation layer call-in d1 confirm pizza size d2 d3 topping crust partial evaluation layer Figure 8: Layering interaction sequences for unsolicited reporting. in heavily parameterized programs are optimized by partial evaluation. Our model is grounded on the (more basic) view of parameters as involving di#erent times of arrival. By capturing the essence of unsolicited reporting as merely di#erences in arrival time (specification time) of aspects, we are able to use partial evaluation for mixed-initiative interaction. The property of partial evaluation that is pertinent here is not just that it is a specialization technique, but also that a sequence of such specializations corresponds to a valid instance of interaction with the dialog script. Moreover, the set of valid sequences of specializations is exactly the set of interactions to be supported. The program of Fig. 6 can thus be thought of as a compaction of all interaction sequences that involve mixing initiative. 2.1.1 Decomposing Interaction Sequences An important issue to be addressed is the decomposition of interaction sequences into parts that should be handled by interpretation and parts that can benefit from partial evaluation. We state only a few general guidelines here. Fig. 7 describes the set of valid interaction sequences for the pizza example as traversals of a graph. Nodes in the graph correspond to stages of specification of aspects. Thus, taking the outgoing edge from the call-in node implies that this stage of the interaction has been completed (the need for the dummy nodes such as d2 and d3 will become clear in a moment). The rules of the traversal are to find paths such that all nodes are visited and every node is visited exactly once. It is easy to verify that with these rules, Fig. 7 models all possibilities of mixing initiative (turns where the user specifies multiple pizza aspects can be modeled as a sequence of graph moves). Expressing our requirements in a graph such as Fig. 7 reveals that the signature bushy nature of mixing initiative is restricted to only a subgraph of the original graph. We can demarcate the starting (d2) and ending points (d3) of bakery item coffee eggs d2 d3 Figure 9: Modeling subdialog invocation. call-in d1 confirm breakfast d2 d3 eggs coffee bakery item layer layer first interpretation second interpretation partial evaluation layer Figure 10: Layering interaction sequences for sub-dialog invocation. the subgraph and layer it in the context of a larger inter-action sequence as shown in Fig. 8. The nodes in the top layer dictate a strict sequential structure, thus they should be modeled by interpretation. The nodes in the bottom layer encapsulate and exhibit the bushy characteristic; they are hence candidates for partial evaluation. The key lesson to be drawn from Fig. 8 is that partial evaluation e#ectively supports the mixed-initiative subgraph without maintaining any additional state (for instance after a node has been vis- ited, this information is not stored anywhere to ensure that it is not visited again). In contrast, the interpretation layer has an implicit notion of state (namely, the part of the interaction sequence that is currently being interpreted). The layered design can be implemented as two alternating pieces of code (for interpretation and partial evaluation, respec- tively) where the interpreter passes control to the partial evaluator for the segments involving pizza attribute specification and resumes after these segments have been evaluated For some applications, the alternating layer concept might need to be extended to more than two layers. Consider a hotel telephone service for ordering breakfast. Assume that ordering breakfast involves specifications of {eggs, co#ee, bakery item} tuples. The user can specify these items in any order, but each item involves a second clarification aspect. After the user has specified his choice of eggs, a clarification of 'how do you like your eggs?' might be needed. Similarly, when the user is talking about co#ee, a clarification of 'do take cream and sugar?' might be required, and so on. This form of mixed-initiative was introduced in Section 1.1 as subdialog invocation. The set of interaction sequences that address this requirement can be represented as shown in Fig. 9 (only the mixed-initiative parts are shown). In call-in confirm size d2 d1 d3 topping crust Figure 11: A requirement for mixing initiative that cannot be captured by partial evaluation. this case, it is not possible to achieve a clean separation of subgraphs into interpretation and partial evaluation in just two layers. One solution is to use three layers as shown in Fig. 10. If we implement both interpretation layers of this design by the same code, some form of scope maintenance (e.g., a stack) will be necessary when moving across the bottom two layers. Pushing onto the stack preserves the context of the original interaction sequence and is e#ected for each step of partial evaluation. Popping restores this context when control returns to the partial evaluator. The semantics of graph traversal remain the same. Once the nodes in the second and third layers are traversed, interpretation is resumed at the top layer to confirm the breakfast order. It is important to note that once again, the semantics of transfer of control between the interpreter and partial evaluator are unambiguous and occur at well defined branch points. The above examples highlight the all-or-nothing role played by partial evaluation. For a dialog script parameterized in terms of slot variables, partial evaluation can be used to support all valid possibilities for mixing initiative, but it cannot restrict the scope of mixing initiative in any way. In particular this means that, unlike interpretation, partial evaluation cannot enforce any individual interaction sequence! Even a slight deviation in requirements that removes some of the walks in the bushy subgraph will render partial evaluation inapplicable. For instance, consider the graph in Fig. 11 that is the same as Fig. 7 with some edges removed. Mixing initiative is restricted to visiting the size, topping, and crust nodes in a strict forward or reverse order. Partial evaluation cannot be used to restrict the scope of mixing initiative to just these two possibilities of specifying the pizza attributes. We can model this by interpretation but this requires anticipation, akin to the design of Fig. 5 (left). This is the essence of partial evaluation as a programming model; it makes some tasks extremely easy but like every other programming methodology, it is not a silver bullet. In a realistic implementation for mixed-initiative interaction, partial evaluation will need to be used in conjunction with other paradigms (e.g., interpretation) to realize the desired objectives. In this paper, we concentrate on only the unsolicited reporting of mixed initiative for which the decomposition illustrated by Fig. 8 is adequate. In other words, these are the applications where the all-or-nothing role of partial evaluation is su#cient to realize mixed-initiative interaction. Our modeling methodology is concerned only with the interaction staging aspect of dialog management, namely determining what the next step(s) in the dialog can or should be. We have not focused on aspects such as response gener- ation. In the pizza example, responses to successful partial evaluation (or interpretation) can be easily modeled as side- Traditional Browser partial input specification window Figure 12: Sketch of an interface for mixed-initiative interaction with web sites. e#ects. In other cases, we would need to take into account the context of the particular interaction sequence. The same argument holds for what is known as tapered prompting [4]. If the prompt for size needs to be replayed (because the user took the initiative and specified a topping), we might want to choose a di#erent prompt for a more natural dialog (i.e., instead of repeating 'What size pizza would you like?,' we might prompt as 'You still haven't specified a size. Please choose among small, medium, or large. We do not discuss these aspects further except to say that they are an important part of a production implementation of dialog based systems. 2.1.2 Implementation Technologies We now turn our attention to implementing our partial evaluation model for existing information access and delivery technologies. As stated earlier, our model is applicable to voice-based interaction technologies as well as web access via hyperlinks. In [15], we study the design and implementation of web site personalization systems that allow the mixing of initiative. In contrast to a voice-based delivery mechanism, (most) interactions with web sites proceed by clicking on hyperlinks. For instance, a pizza ordering web service might provide hyperlinks for choices of size so that clicking on a link implies a selection of size. This might refresh the browser to a new page that presents choices for topping, and so on. Since clicking on links implies a response to the initiative taken by the web page author, a di#erent communication mechanism needs to be provided to enable the user to take the initiative. Our suggested interface design is shown in Fig. 12. An extra window is provided for the user to type in his specification aspects. This is not fundamentally di#erent from a location toolbar in current browsers that supports the specification of URL. Consider that a web page presents hyper-links for choices of pizza size. Using the interface in Fig. 12, the user can either click on her choice of size attribute in the top window (e#ectively, responding to the initiative), or can use the bottom window to specify, say, a topping out- of-turn. To force an interpretation mode for segments of an interaction sequence, the bottom window can be made inactive. Modeling decisions, implementation details, and experimental results for two web applications developed in this manner are described in [15]; we refer the reader to this reference for details. Our original goal for this paper was to study these concepts for voice-based interaction technologies and to see if our model can be used for the implementation of a voice- Microphone extraction Feature Digital Converter Analog to Speech Recognizer Language model HMM accoustic models Language Processing Natural Management Dialog Dialog model analog signal digital signal feature vector response result(s) results+ Figure 13: Basic components of a spoken language processing system. Internet Server .html pages http protocol PC with web browser Voice Browser Platform Telephone Network Internet Server http protocol .vxml pages files Figure 14: (left) Accessing HTML documents via a HTTP web server. (right) Accessing VoiceXML documents via a HTTP web server. based mixed-initiative application. A variety of commercial technologies are available for developing voice-based appli- cations; as a first choice of a representational formalism, we proceeded to use the specification language of the VoiceXML dialog management architecture [4]. The idea was to describe dialogs in VoiceXML notation and use partial evaluation to realize mixed-initiative interaction. After some initial experimentation we realized that VoiceXML's form interpretation algorithm (FIA), which processes the dialogs, provides mixed-initiative interaction using a script very similar to the one we presented for use with a partial evaluator (see Fig. 6)! In other these, there is no real advantage to partially evaluating a VoiceXML specification! This pointed us to the possibility that perhaps we can identify an instantiation of our model in VoiceXML's dialog management architecture and especially, the FIA. The rest of the paper takes this approach and shows that this is indeed true. We also identify other implementation technologies where we can usefully implement a voice-based mixed-initiative system using our model. We merely identify the opportunities here and hope to elaborate on a complete implementation of our model in a future paper. 3. SOFTWARE TECHNOLOGIES FOR VOICE-BASED MIXED-INITIATIVE APPLICATIONS Before we can study the programming of mixed-initiative in a voice-based application, it will be helpful to understand the basic architecture (see Fig. 13) of a spoken language processing system. As a user speaks into the sys- tem, the sounds produced are captured by a microphone and converted into a digital signal by an analog-to-digital converter. In telephone-based systems (the VoiceXML architecture covered later in the paper is geared toward this mode), the microphone is part of the telephone handset and the analog-to-digital conversion is typically done by equipment in the telephone network (in some cellular telephony models, the conversion would be performed in the handset itself). The next stage (feature extraction) prepares the digital speech signal to be processed by the speech recognizer. Features of the signal important for speech recognition are extracted from the original signal, organized as an attribute vector, and passed to the recognizer. Most modern speech recognizers use Hidden Markov Models (HMMs) and associated algorithms to represent, train, and recognize speech. HMMs are probabilistic models that must be trained on an input set of data. A common technique is to create sets of acoustic HMMs that model phonetic units of speech in context. These models are created from a training set of speech data that is (hopefully) representative of the population of users who will use the system. A language model is also created prior to performing recognition. The language model is typically used to specify valid combinations of the HMMs at a word- or sentence-level. In this way, the language model specifies the words, phrases, and sentences that the recognizer can attempt to recognize. The process of recognizing a new input speech signal is then accomplished using e#cient search algorithms (such as Viterbi decoding) to find the best matching HMMs, given the constraints of the language model. The output of the speech recognizer can take several di#erent forms, but the basic result is a text string that is the recognizer's best guess of what the user said. Many recognizers can provide additional information such as a lattice of results, or an N-best ranked list of results (in case the later stages of processing wish to reject the recognizer's top choice). A good introduction to speech recognition is available in [10]. The stages after speech recognition vary depending on the application and the types of processing required. Fig. 13 presents two additional phases that are commonly included in spoken language processing systems in one form or an- other. We will broadly refer to the first post-recognition stage as natural language processing (NLP). NLP is a large field in its own right and includes many sub-areas such as parsing, semantic interpretation, knowledge representation, and speech acts; an excellent introduction is available in Allen's classic [1]. Our presentation in this paper has assumed NLP support for slot-filling (i.e., determining values for slot variables from user input). Slot-filling is commonly achieved by defining parts of a language model and associating them with slots. The language model could be specified as a context-free grammar or as a statistically-based model such as n-grams. Here we focus on the former: in this approach, slots can be specified within the productions of a context-free grammar (akin to a attribute grammar) or they can be associated with the non-terminals in the grammar. We will refer to the next phase of processing as simply 'dialog management' (see Fig. 13). In this phase, augmented results from the NLP stage are incorporated into the dialog and any associated logic of the application is executed. The job of the dialog manager is to track the proceedings of the dialog and to generate appropriate responses. This is often done within some logical processing framework and a dialog model (in our case, a dialog script) is supplied as input that is specific to the particular application being designed. The execution of the logic on the dialog model (script) results in a response that can be presented back to the user. Sometimes response generation is separated out into a subsequent stage. 3.1 The VoiceXML Dialog Management Architecture There are many technologies and delivery mechanisms available for implementing Fig. 13's basic components. A popular implementation can be seen in the VoiceXML dialog management architecture. VoiceXML is a markup language designed to simplify the construction of voice-response applications [4]. Initiated by a committee comprising AT&T, IBM, Lucent Technologies, and Motorola, it has emerged as a standard in telephone-based voice user interfaces and in delivering web content by voice. We will hence cover this architecture in detail. The basic idea is to describe interaction sequences using a markup notation in a VoiceXML document. As the VoiceXML specification [4] indicates, a VoiceXML document constitutes a conversational finite state machine and describes a sequence of interactions (both fixed- and mixed-initiative are supported). A web server can serve VoiceXML documents using the HTTP protocol (Fig. 14, right), just as easily as HTML documents are currently served over the Internet (Fig. 14, left). In addition, voice-based applications require a suitable delivery platform, illustrated by a telephone in Fig. 14 (right). The voice-browser platform in Fig. 14 (right) includes the VoiceXML interpreter which processes the documents, monitors user inputs, streams mes- sages, and performs other functions expected of a dialog management system. Besides the VoiceXML interpreter, the voice-browser platform typically includes a speech rec- ognizer, a speech synthesizer, and telephony interfaces to help realize these aspects of voice-based interaction. Dialog specification in a VoiceXML document involves organizing a sequence of forms and menus. Forms specify a set of slots (called field item variables) that are to be filled by user input. Menus are syntactic shorthands (much like a case construct); since they involve only one field item variable (argument), there are no opportunities for mixing initiative. We do not discuss menus further in this paper. An example VoiceXML document for our pizza application is given in Fig. 15. As shown in Fig. 15, the pizza dialog consists of two forms. The first form (welcome) merely welcomes the user and transitions to the second. The place order form involves four fields (slot variables) - the first three cover the pizza attributes and the fourth models the confirmation variable (re- call the dialogs in Section 1). In particular, prompts for soliciting user input in each of the fields are specified in Fig. 15. Interactions in a VoiceXML application proceed just like a web application except that instead of clicking on a hyper-link (to goto a new state), the user talks into a microphone. The VoiceXML interpreter then determines the next state to move to. Any appropriate responses (to user input) and prompts are delivered over a speaker. The core of the interpreter is a so-called form interpretation algorithm that drives the interaction. In Fig. 15, the fields provide for a fixed-initiative, system-directed interaction. The FIA simply visits all fields in the order they are presented in the document. Once all fields are filled, a check is made to ensure that the confirmation was successful; if not, the fields are cleared (notice the clear namelist tag) and the FIA will proceed to prompt for the inputs again, starting from the first unfilled field - size. The form in Fig. 15 is referred to as a directed one since the computer has the initiative at all times and the fields are filled in a strictly sequential order. To make the interaction mixed-initiative (with respect to size, crust, and topping), the programmer merely has to specify a form-level grammar that describes possibilities for slot-filling from a user utterance. An example form-level grammar file (size toppingcrust.gram) is given in Fig. 16. The productions for sizetoppingcrust cover all possibilities of filling slot variables from user input, including multiple slots filled by a given utterance, and various permutations of specifying pizza attributes. The grammar is associated with the dialog script by including the line: just before the definition of the first field (size) in Fig. 15. The form-level grammar contains productions for the various choices available for size, topping, and crust and also qualifies all possible parses for a given utterance (modeled by the non-terminal sizetoppingcrust). Any valid combination of the three pizza aspects uttered by the user (in any order) is recognized and the appropriate slot variables are instantiated. To see why this also achieves mixed-initiative, let us consider the FIA in more detail. Fig. 17 reproduces the salient aspects of the FIA relevant for our discussion. Compare the basic elements of the FIA to the stages in Fig. 5 (right). The Select phase corresponds to the interpreter, the Collect phase gathers the user input, and actions taken in the Process phase mimic the partial evaluator. Recall that 'programs' (scripts) in VoiceXML can be modeled by finite-state machines, hence the mechanics of partial evaluation are considerably simplified and just amount to filling the slot and tagging it as filled. Since the <?xml version="1.0"?> <vxml version="1.0"> <!- pizza.vxml A simple pizza ordering demo to illustrate some basic elements of VoiceXML. Several details have been omitted from this demo to help make the basic ideas stand out. -> <form id="welcome"> <block name="block1"> Thank you for calling Joe's pizza ordering system. <goto next="#place_order" /> <form id="place_order"> <field name="size"> What size pizza would you like? <field name="topping"> What topping would you like on your pizza? <field name="crust"> What type of crust do you want? <field name="verify"> So that is a <value expr="size"/> <value expr="topping"/> pizza with <value expr="crust"/> crust. Is this correct? yes | no <if cond="verify=='no'"> <clear namelist="size topping verify crust"/> Sorry. Your order has been canceled. <else/> Thank you for ordering from Joe's pizza. Figure 15: Modeling the pizza ordering dialog in a VoiceXML document. public {this.size=$} {this.crust=$} {this.topping=$} | {this.topping=$} {this.size=$} {this.crust=$} | {this.crust=$} {this.topping=$} {this.size=$}; small | medium | large; sausage | pepperoni | onions | green peppers; regular | deep dish | thin; Figure form-level grammar to be used in conjunction with the script in Fig. 15 to realize mixed-initiative interaction. While (true) { Select the first form item with an unsatisfied guard condition (e.g., unfilled) If no such form item, exit // COLLECT PHASE Queue up any prompts for the form item Get an utterance from the user // PROCESS PHASE foreach (slot in user's utterance) { if (slot corresponds to a field item) { copy slot values into field item variables set field item's `just_filled' flag some code for executing any 'filled' actions triggered Figure 17: Outline of the form interpretation algorithm (FIA) in the VoiceXML dialog management archi- tecture. Adapted from [4]. public {this.crust=$} | {this.topping=$}; small | medium | large; sausage | pepperoni | onions | green peppers; regular | deep dish | thin; Figure alternative form-level grammar to realize mixed-initiative interaction with the script in Fig. 15. FIA repeatedly executes while there are unfilled form items remaining, the processing phase (Process) is e#ectively parameterized by the form-level grammar file. In other words, the form-level grammar file not only enables slot filling, it also implicitly directs the staging of interactions for mixed- initiative. When the user specifies 'peperroni medium' in an utterance, not only does the grammar file enable the recognition of the slots they correspond to (topping and size), it also enables the FIA to simplify these slots (and mark them as 'filled' for subsequent interactions). The form-level grammar file shown in Fig. 16 (which is also a specification of interaction staging) may make Voice- XML's design appear overly complex. In reality, however, we could have used the vanilla form-level grammar file in Fig. 18. While helping to realize mixed-initiative, the new form-level file (as does our model) also allows the possibility of utterances such as 'pepperoni pepperoni,' or even, 'pepperoni sausage!' Suitable semantics for such situations (including the role of side-e#ects) can be defined and accommodated in both the VoiceXML model and ours. It should thus be obvious to the reader that VoiceXML's dialog management architecture is actually implementing a mixed evaluation model (for conversational finite state machines), comprising interpretation and partial evaluation. The VoiceXML specification [4] refers to the form-level file as a 'grammar file,' when it is actually also a specification of staging. Even though the grammar file serves the role of a language model in a voice application, we believe that recognizing its two functionalities is important in understanding mixed-initiative system design. A case in point is our study of personalizing interaction with web sites [15] (see also Fig. 12). There is no requirement for a 'grammar file,' as there is usually no ambiguity about user clicks and typed-in keywords. Specifications in this application thus serve to associate values with program variables and do not explicitly capture the staging of interactions. The advantageous of partial evaluation for interaction staging are thus obvious. 3.2 Other Implementation Technologies VoiceXML's FIA thus includes native support for slot fill- ing, slot simplification, and interaction staging. All of these are functions enabled by partial evaluation in our model. Table contrasts two other implementation approaches in terms of these aspects. In a purely slot-filling system, native support is provided for simplifying slots from user utterances but extra code needs to be written to model the control logic (for instance, 'the user still didn't specify his choice of size, so the question for size should be repeated. Several commercial speech recognition vendors provide APIs that operate at this level. In addition, many vendors support low-level APIs that provide basic access to recognition results (i.e., text strings) but do not perform any additional processing. We refer to these as recognizer-only APIs. They serve more as raw speech recognition engines and require significant programming to first implement a slot-filling engine and, later, control logic to mimic all possible opportunities for staging. Examples of the two latter technologies can be seen in the commercial telephone-based speech recognition market (from companies such as Nuance, SpeechWorks, and IBM). The study presented in this paper suggests a systematic way by which their capabilities for mixed-initiative interaction can be assessed. Table 1 also shows that in the latter two software technologies, our partial evaluation model can be implemented to achieve mixed-initiative interaction. 4. DISCUSSION Our work makes contributions to both partial evaluation and mixed-initiative interaction. For the partial evaluation community, we have identified a novel application where the motivation is the staging of interaction (rather than speedup). Since programs (dialogs) are used as specifications of interaction, they are written to be partially eval- uated; partial evaluation is hence not an 'afterthought' or an optimization. An interesting research issue is: Given (i) a set of interaction sequences, and (ii) addressable information (such as arguments and slot variables), determine (iii) the smallest program so that every interaction sequence can be staged in a model such as Fig. 5 (right). As stated earlier, this requires algorithms to automatically decompose and 'layer' interaction sequences into those that are best addressed in the interpreter and those that can benefit from representation and specialization by the partial evaluator. For mixed-initiative interaction, we have presented a programming model that accommodates all possibilities of stag- ing, without explicit enumeration. The model makes a distinction between fixed-initiative (which has to be explicitly programmed) and mixed-initiative (specifications of which can be compacted for subsequent partial evaluation). We have identified instantiations of this model in VoiceXML and slot-filling APIs. We hope this observation will help system designers gain additional insight into voice application design strategies. It should be recalled that there are various facets of mixed-initiative that are not addressed in this paper. Besides sub-dialog invocations, VoiceXML's design can support dialogs such as shown in Fig. 19. Caller 1's request, while demonstrating initiative, implies a dialog with an optional stage (which cannot be modeled by partial evaluation). Such a situation has to be trapped by the interpreter, not by partial evaluation. Caller 2 does specify a staging, but his staging poses constraints on the computer's initiative, not his own. Such a 'meta-dialog' facet [5] requires the ability to jump out of the current dialog; VoiceXML provides many elements for describing such transitions. Extending our programming model to cover these facets is an immediate direction of future research. VoiceXML also provides certain 'impure' features and side- e#ects in its programming model. For instance, after selecting a size (say, medium), the caller could retake the initiative in a di#erent part of the dialog and select a size again (this time, large). This will cause the new value to override any existing value in the size slot. In our model, this implies the dynamic substitution of an earlier, 'evaluated out,' stage with a functional equivalent. Obviously, the dialog manager has to maintain some state (across partial evaluations) to accomplish this feature or support a notion of despecializa- tion. This suggests new directions for research in program transformation. It is equally possible to present the above feature of VoiceXML as a shortcoming of its implementation of mixed ini- tiative. Consider that after selection of a size, the scope of any future mixing of initiative should be restricted to the remaining slots (topping and crust). The semantics of graph traversal presented earlier capture this requirement. Such an e#ect is cumbersome to achieve in VoiceXML and would Software Support for Support for Technology Slot Simplification Interaction Staging Slot Filling Systems # - Recognizer-Only APIs - Table 1: Comparison of software technologies for voice-based mixed-initiative applications. calling Joe's pizza ordering system. What size pizza would you like? Caller 1: What sizes do you have? 3 Caller 2: Err. Why don't you ask me the questions in topping-crust-size order? Figure 19: Other mixed-initiative conversations that are supported by VoiceXML. probably require transitioning to progressively smaller forms (with correspondingly restrictive form-level grammars). Our model provides this feature naturally; after size has been partially evaluated 'out,' the scope of future partial evaluations is automatically restricted to involve only topping and crust. Our long-term goal is to characterize mixed initiative facets, not in terms of initiative, interaction, or task models but in terms of the opportunities for staging and the program transformation techniques that can support such staging. This means that we can establish a taxonomy of mixed-initiative facets based on the transformation techniques (e.g., partial evaluation, slicing) needed to realize them. Such a taxonomy would also help connect the facets to design models for interactive software systems. We also plan to extend our software model beyond slot-and-filler structures, to include reasoning and exploiting context. 5. NOTES The work presented in this paper is supported in part by US National Science Foundation grants DGE-9553458 and IIS-9876167. After this paper was submitted, a new version (version 2.00) of the VoiceXML specification was released [12]. Our observations about the instantiation of our model in the VoiceXML dialog management architecture also apply to the new specification. 6. --R Natural Language Understanding. Towards Conversational Human-Computer Interaction Voice eXtensible Markup Language: VoiceXML. An Assessment of Written/Interaction Dialogue for Information Retrieval Applications. An Introduction to Discourse Analysis. Computational Models for Mixed Initiative Interaction (Papers from the Partial Evaluation and Automatic Program Generation. An Introduction to Natural Language Processing Cambridge University Press Voice eXtensible Markup Language: VoiceXML. What is Mixed-Initiative Interaction? In The Partial Evaluation Approach to Information Personalization. Integrating Planning and Learning: The PRODIGY Architecture. --TR Explanation-based generalisation = partial evaluation Partial evaluation and automatic program generation Natural language understanding (2nd ed.) A collaborative model of feedback in human-computer interaction Speech and Language Processing Toward conversational human-computer interaction Mixed-Initiative Interaction --CTR J. Wilkie , M. A. Jack , P. J. Littlewood, System-initiated digressive proposals in automated human-computer telephone dialogues: the use of contrasting politeness strategies, International Journal of Human-Computer Studies, v.62 n.1, p.41-71, January 2005

mixed-initiative interaction;VoiceXML;interaction sequences;dialog management;partial evaluation

503068

Timing verification of dynamically reconfigurable logic for the xilinx virtex FPGA series.

This paper reports on a method for extending existing VHDL design and verification software available for the Xilinx Virtex series of FPGAs. It allows the designer to apply standard hardware design and verification tools to the design of dynamically reconfigurable logic (DRL). The technique involves the conversion of a dynamic design into multiple static designs, suitable for input to standard synthesis and APR tools. For timing and functional verification after APR, the sections of the design can then be recombined into a single dynamic system. The technique has been automated by extending an existing DRL design tool named DCSTech, which is part of the Dynamic Circuit Switching (DCS) CAD framework. The principles behind the tools are generic and should be readily extensible to other architectures and CAD toolsets. Implementation of the dynamic system involves the production of partial configuration bitstreams to load sections of circuitry. The process of creating such bitstreams, the final stage of our design flow, is summarized.

INTRODUCTION In dynamically reconfigurable logic (DRL), a circuit or system is adapted over time. This presents additional design and verification problems to those of conventional hardware design [1] that standard tools cannot cope with directly. For this reason, DRL design methods typically involve the use of a mixture of industry standard tools, along with custom tools and some handcrafting to cover the conventional tools inadequacies. This paper introduces extensions to a previously reported CAD tool named DCSTech [2] which was created to automate the process of translating dynamic designs from VHDL into placed and routed circuits. The original version of the tool supported the Xilinx XC6200 family of FPGAs, and concentrated on the timing verification aspects of the problem. This paper reports on the extensions made to DCSTech to target the Xilinx Virtex family, and to enhance its capabilities. As a mainstream commercial FPGA, the design tool capabilities available with this family exceed those of the XC6200, allowing the designer to work more productively at a higher level of abstraction. By combining the Virtex platform's capabilities with those of the extended DCSTech, the designer has the ability to specify designs in RTL/behavioural VHDL, place and route them and verify their timing. DCSTech's back-annotation support has been extended to produce VITAL VHDL models suitable for DRL in addition to processing SDF timing information. This enables back-annotated timing analysis regardless of the level of abstraction at which the original design was produced. The original DCSTech tool was written to be extensible to other architectures. This work verifies the validity of its extensibility hooks. The extensibility of DCSTech to other architectures relies on the architecture's CAD tools supporting a select set of capabilities. Most modern CAD tools meet the majority of these requirements (with the exception of configuration bitstream access), although some weaknesses, particularly in the control of routing, are apparent. Therefore, the design techniques presented here should be readily extensible to other dynamically reconfigurable FPGAs. The paper begins by reviewing existing work in section 2 before presenting the challenges of DRL design in section 3. In section 4 we provide an overview of the principles behind DCSTech while section 5 describes how they are applied to the Virtex. Section 6 discusses the enhanced back annotation capabilities necessary for design at the RTL and behavioral abstraction levels. The tools are designed to be as architecture independent as possible and section 7 describes how the tool may be extended *Now at Xilinx Inc. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. FPGA'02, February 24-26, 2002, Monterey, California, USA. to support other dynamically reconfigurable FPGAs. In section 8, we describe how partial configuration bitstreams may be obtained from the placed-and-routed subsections of the dynamic design, an area of current research. The design flow is illustrated with an example in section 9 before the paper concludes with remarks on future research into the use of other modern CAD techniques such as Static Timing Analysis (STA) within the DRL design flow. 2. EXISTING WORK Over the last six years, researchers have developed a number of tools and techniques, supporting different target DRL systems. The target systems can be characterized by their component set, the set of resources that make up the system. Custom Computing Machines (CCMs), for example, include processors, FPGAs and memory in their component set. Tools ranging from high-level language compilers to structural-level hardware/software co-design environments have been designed for such target systems. CCM compilers include tools such as Nimble [3], and compilers for the GARP chip [4], which compile ANSI-C. In addition to standard C compilation, CCM compilers partition the application into a software executable and a set of hardware modules that can be loaded onto the reconfigurable datapath or FPGA. As these tools are aimed at achieving a rapid design flow, similar to conventional computer programming, they do not usually achieve optimum results. Tools such as JHDL [5][6], a structural/RT level hardware/software codesign environment based on Java, allow the designer to customize his circuitry and specify its placement. This allows designers to use their own expertise to optimize the layout and composition of their circuits to achieve better results (e.g. faster circuits and smaller reconfiguration bitstreams if partial reconfiguration is used) as well as designing the associated software in one environment. Another design challenge is found when the component set is a single FPGA device or when dynamic reconfiguration is applied within individual devices. This sort of design throws up many situations that most industry standard tools cannot handle at all, such as verification, partial bitstream generation and automatic configuration controller production. Many of the solutions developed for this type of design also apply to CCM design. In [7], Luk et al described a CAD framework for DRL design targeted at the Xilinx XC6200 FPGA. A library based design approach was used to encourage design reuse and control circuit placement. This increases the similarity between successive configurations and reduces the size of the partial configuration files required. However, such a structural design approach limits the portability of the tools, since new libraries targeted to each device are required. Vasilko's DYNASTY [8] CAD framework uses a designer driven temporal floorplanning approach, in which the designer can visualise the layout of tasks on the FPGA over time. It acts as a DRL-aware replacement to a place and route (PAR) tool and operates on synthesised gate-level designs. This has a number of advantages, such as ease of area estimation and the ability to control routing and component placement exactly. The designer therefore has the ability to generate exactly the required layouts. However, as the tools are closely associated with the XC6200 architecture considerable effort would be required to port them to operate with other devices. Research has also taken place into the use of alternative languages that have useful properties in expressing aspects of a DRL design. Ruby [9], Pebble [10] and Lava [11] allow the designer to specify component placement using more convenient methods than the usual use of attributes associated with standard HDL designs. Pebble also includes a reconfigure-if statement, which builds in support for DRL. Recent work with Lava has seen it used with the Xilinx Virtex FPGA. The DCS CAD framework provides simulation (DCSim) [1], technology mapping and back annotation (DCSTech) [2] and configuration controller synthesis (DCSConfig) [12]. Although DYNASTY uses the same ideas as DCSTech, DCSTech partitions the design at a higher level of abstraction. This gives two advantages in the form of portability and circuit specialisation by the synthesis tool. Since the design is partitioned at an abstract level, DCSTech requires only a little device specific knowledge. The majority of the partitioning process is platform independent, as is the resulting circuit description. The tool is therefore easily ported to support different architectures. As the designs are synthesised after partitioning any optimisations such as constant propagation can be performed by the synthesis tools. If the design is partitioned after synthesis, a further optimization stage may be required to obtain the best results. At this level of abstraction the area requirements of the circuit are more difficult to estimate, so some iteration may be required to obtain the optimal layout. Other researchers have concentrated on design at lower levels of abstraction, allowing the designer absolute control over component placement and routing. Such tools include CHASTE [13], which provides access to the XC6200 configuration file and the JBits SDK [14][15], which provides a variety of tools to access, modify and verify Virtex configurations. In addition, it allows the designer to produce new whole or partial configurations. This approach could also be valuable as a method of performing final optimizations at the end of a higher-level design flow. 3. IMPLEMENTATION CHALLENGES DRL is based on a many-to-one temporal logic mapping. This means that different logic functions occupy the same area of the logic array at different points in time. Tasks that share physical resources cannot be active at the same time; they are mutually exclusive. Tasks can also be mutually exclusive for algorithmic reasons. A set of mutually exclusive tasks is called a mutex set and the swappable tasks are termed dynamic tasks. Tasks that are not altered in any way over time are described as static tasks. In designing a dynamic system, the various tasks must be placed in such a way as to ensure that no task is accidentally overwritten while it is active. The consequences of such an error range from subtle errors in operation to damage to the FPGA itself. Dynamic tasks are added to and removed from the array by loading partial configuration files to alter logic and routing. The designer has to guarantee that all necessary connections between the dynamic task and the surrounding environment will be made. The routing paths configured onto the array with the dynamic task must meet the routing coming from the surrounding array to which they are intended to connect. The bitstreams must not cause contention, for example by configuring a second driver onto a bidirectional routing resource. The final problem the designer faces is that standard CAD tools, which are intended for the design of static circuits, will not accept the mapping of more than one function to a particular logic resource. Similarly, multiple drivers for a particular signal would be treated as an error, since no mechanism exists to indicate that the drivers are scheduled to operate at different times. 4. AUTOMATING DYNAMIC DESIGN PROCESSING WITH DCSTech 4.1 Overview DCSTech was designed as a tool to help the designer to overcome these problems. It can be thought of as a domain converter between the static and dynamic domains, fig. 1. The input dynamic system is split into a series of static designs on which conventional design tools (synthesis and APR) can be used. After the required implementation steps have been performed on these static sub-designs, a number of files are produced. VITAL compliant VHDL files that describe the systems functionality are created, along with SDF files specifying the circuit's timing and configuration bitstreams. These files all require further processing before they are useful. To verify the designs functionality and timing, the SDF and VHDL files must be converted back to the dynamic domain, in order to simulate them in the context of the overall system. To implement the dynamic system, the configuration bitstreams must also be converted into valid partial reconfigurations. The original version of DCSTech supported the domain conversion of timing information. This was because designs were specified at the netlist level, and therefore a VITAL compliant simulation model could be produced from the original design. The use of higher design abstractions such as behavioural code combined with synthesis means that no such netlist exists before synthesis. The current version has therefore added netlist conversion to the original SDF conversion, leaving bitstream conversion as a manual process. The progress so far is illustrated in fig. 2. 4.2 Design and File Flow The dynamic design input to DCSTech consists of VHDL files describing the systems functionality. Each dynamic task is represented as a component instantiation. Hence, the top-level of the design is structural. Within each component, any synthesisable level of abstraction can be used. The designer assigns each dynamic task to a mutex set. This mutex set is assigned a zone on the logic array and all dynamic tasks within that set must reside within the corresponding zone. Thus, tasks within a mutex set can overwrite each other, but static logic and tasks in other mutex sets are unaffected. The correct system operation is then assured so long as an appropriate reconfiguration schedule is used (it is possible that the configuration control mechanism used to activate and deactivate tasks could cause problems if it is incorrectly designed). Clearly, the zone of each mutex set must be large enough to accommodate its largest task. DCSTech Bitstream Original Version This Version Future work DCSTech Bitstream Original Version This Version Future work The dynamic intent of the system is captured in a Reconfiguration Information Format (RIF) file. This file describes the conditions under which tasks activate (are configured onto the FPGA) and deactivate (are removed or stopped), the mutex set to which they belong and their placement. Information on the RIF file was published in [2]. In the static domain, one of the sub-designs deals with all the static tasks in the design while each dynamic task is placed into a Figure 1: between static and dynamic domains Figure 2: Domain transforms performed by different DCSTech versions Reconfiguration Information Conventional CAD Tools DCSTech Conventional CAD Tools Timing Model Timing Model Timing Model DCSTech Dynamic Dynamic static designs static results dynamic timing results Reconfiguration Information Reconfiguration Information Conventional CAD Tools DCSTechDCSTech Conventional CAD Tools Timing Model Timing Model Timing Model Timing Model Timing Model Timing Model DCSTechDCSTech Dynamic Dynamic static designs static results dynamic timing results sub-design of its own. The concept of terminals is used to ensure the correct routing connectivity with the dynamic tasks surrounding environment. These are special components used to lock the end of hanging signals to a particular location on the logic array. By locating both hanging ends of the signal at the same place, the connection can be easily produced. One reserved area is added to the static sub-design for each mutex set in the original design. Similarly, the dynamic task components are surrounded by a bounding-box that ensures that they will be placed within the reserved area for their mutex set, fig. 3. After the sub-designs have been placed and routed by standard back-end tools, accurate estimates of their timing can be made. These estimates are typically written out into an SDF file. To allow evaluation of the performance of the system, this information must be applied to the overall dynamic system. DCSTech is capable of mapping the SDF information into the dynamic design simulation model that DCSim creates, allowing timing simulation. To apply the SDF file to the dynamic domain, the cells must each be changed to match the hierarchy of the dynamic system simulation to which it is applied. In addition, the timing entries for the terminals are removed and their relevant timing information mapped to isolation switches (simulation artefacts added by DCSim to mimic the design's dynamic behavior in a conventional simulator). Although the system hierarchy is altered during this domain conversion process, the actual timing information is unaltered, providing an accurate timing model. Further details of the process can be found in [2]. 5. CHANGES MADE TO DCSTech TO A number of changes were required in order to retarget the static design representations to Virtex synthesis and APR tools, as summarized in table 1. These changes allow us to replicate the capabilities DCSTech made available for the XC6200 on the Virtex. Table 1. Methods of implementing DCSTech requirements on XC6200 and Virtex Problem XC6200 Solution Virtex Solution Reserving areas of the array Reserve constraint Prohibit constraint Locating dynamic tasks within a zone bbox attribute assigns a bounding box loc constraint allows ranges to be assigned Preventing partial circuits from being removed Use register as terminal component on hanging signals Changes to design representation and software settings Lock hanging signals to fixed array locations Terminal components with rloc constraints Terminal components with loc constraints Reserving areas on the logic array is a simple change of attribute from a RESERVE constraint which prevents XACT6000 from placing logic in the specified zone to specifying a PROHIBIT constraint which does the same task in the Xilinx CAD tools. This is added to the User Constraints Format (UCF) file. Dynamic task locations can be set using a combination of an rloc and a bbox attribute in XACT6000. The Virtex tools allow location ranges to be specified with the loc attribute. Because registers could be read from and written to through the configuration interface, any line connected to and from registers was considered a valid connection, even when the A A and D are static tasks and C are mutually exclusive dynamic tasks floorplan of dynamic design floorplan of static-task design floorplan of a dynamic task B floorplan of a dynamic task C A A DCSTech Special terminals used to lock the ends of cut signals to a fixed position in the static-task design and in relevant dynamic task designs Dynamic Static Reserved A A and D are static tasks and C are mutually exclusive dynamic tasks floorplan of dynamic design floorplan of static-task design floorplan of a dynamic task B floorplan of a dynamic task C A A DCSTech Special terminals used to lock the ends of cut signals to a fixed position in the static-task design and in relevant dynamic task designs Dynamic Static Reserved Figure 3: Floorplan of a DRL circuit containing two dynamic tasks before and after processing by DCSTech register had incomplete connectivity, such as no output connection. Using registers to terminate hanging nets therefore prevented partial circuits from being removed. This technique does not work with the Virtex synthesis tools, making two changes necessary in the way that dynamic designs were represented in the static domain. Firstly, the VHDL entity of each dynamic task must have ports in it to describe its connectivity, whereas before terminal components were all that was required. In addition, to prevent large areas of the static design being optimised away, the connectivity between the inputs and outputs of the reserved area should be indicated. Instantiating a black-box "mutex set" component, encapsulating the inputs and outputs of all the dynamic tasks in the mutex set solves this problem. The Xilinx Foundation tools support an option not to remove unconnected logic, which suffices for the placement and routing stage. The terminal component used to terminate hanging nets has been changed to a wire or buffer mapped to a look-up-table. This component replaces the RPFDs and FDCs used on the XC6200 and has an advantage in that it does not contribute any functionality, while accepting location constraints. This simplifies the changes required in the final bitstream generation stage and the netlist conversion process. The changes described above allow most of the basic requirements outlined in section 3 to be met by the standard Virtex tools. However, one area of weakness is constraining the placement of routing. The constraints described above only apply to logic placement, and therefore the routing from circuits can exceed their bounding boxes and invade reserved zones, although the Xilinx modular design tools [16] can help alleviate this problem. These are factors that the designer must take account of when configuration bitstreams are being produced, either by re-routing the offending lines, or by including the routes in the appropriate configurations. In effect, the dynamic task bounding-box should be increased in size to accommodate any wayward routing. 6. ENHANCED BACKANNOTATED The original static-to-dynamic domain conversion support for SDF files has been enhanced in the new revision of DCSTech. SDF information can only be applied to gate-level VITAL compliant designs. If a design is produced at an abstract level, then SDF information cannot be applied to it. As with most modern APR tools, the Virtex tools are capable of writing out a VITAL VHDL netlist that matches their SDF files. The netlists are typically flat "seas of gates" with no hierarchy (although many tools allow control over hierarchy flattening). These files must be included in the domain conversion process in order to allow timing analysis to be performed when design abstractions above the structural level are used. DCSTech handles this domain conversion process by instantiating the dynamic tasks into the VHDL netlist for the static design. The resulting dynamic circuit is, in effect a gate-level version of the original RTL design, such as a DRL aware synthesis tool might produce. DCSim is used to simulate the circuit. Since the hierarchy of the system often changes if synthesis and APR tools flatten the design, it may not match the hierarchy entries in the original RIF file. Therefore, a new RIF file is written as part of the domain conversion process. The domain conversion therefore produces a complete new dynamic design representation that DCSim can use to build a simulation model. As reported in section 4, the relevant timing information associated with the terminal components is usually applied to DCSim's isolation switches while the references to the terminal components are removed from the design. However, the Virtex terminal components have no functionality and therefore do not interfere with the simulation of the system. As a result, those components that contribute timing data do not need to be removed during the static to dynamic domain conversion; hence, there is no need to retarget the timing data to the isolation switches (although the isolation switches are still introduced as they are needed to simulate the circuit). This simplifies the conversion process thereby reducing the runtime of the DCSTech tool. 7. THE EXTENDED DCSTech TOOL DCSTech now provides multi-architecture support and interfaces with several third party CAD tools. It was originally designed to be extensible with as much of the technique as generic and device independent as possible. Obviously, changes in the CAD environment and device architectures will mean that parts of the technique will need to be changed, either to take advantage of device features or to coexist with its supporting CAD DCSTech dependent functions RIF VHDL Dynamic design domain Virtex dependent functions Other device dependent functions Static design domain Log file CRF file Options DCSim's CRF file DCSTech dependent functions RIF VHDL Dynamic design domain Virtex dependent functions Other device dependent functions Static design domain Log file CRF file Options DCSim's CRF file Figure 4: File flow for the extended DCSTech tool framework. The major changes came on the back annotation side, where support for VHDL domain conversion was added. However, this is not something that is specific to the Virtex device, but a necessary step to enable designers to work at higher levels of abstraction. Therefore, the concepts behind the tool remain generic and architecture independent and the design methodology, outlined in section 4, remains unchanged in this revision. To facilitate this extensibility, the device dependent functions are stored in dynamic link libraries. New devices can therefore be supported with the addition of a DLL. The file flow for DCSTech is shown in fig. 4. The shaded files represent non-design files used as part of DCSTech's operation. The CRF files are cross-reference files used to store information such as terminal component connectivity and isolation switch locations (DCSim). The log file contains reports for the user. The options file can be used as an alternative to typing in command line switches. The design philosophy described in this paper will be able to provide DRL design support for any FPGA and CAD tool set provided it complies with the following requirements: . The FPGA is dynamically reconfigurable . Synthesis or design translation from VHDL is available . A suitable component can be found to lock the ends of hanging nets to a particular location on the logic array . A method is available to prevent unconnected circuits being removed from the design . Components can be assigned a bounding-box constraining them to a location on the array . Areas of the array can be reserved, prohibiting other logic from being placed within that area . The APR tools produce back annotated VITAL VHDL and SDF files . The names of elements instantiated into the design in a structural manner are predictable within the SDF and VITAL VHDL models. Components generated by synthesis tools generally have unpredictable names, but structural components are usually named after their instantiation label and the hierarchy above them in the original design. DCSTech has to be able to find the terminal components that are added to the design in the dynamic-to-static conversion as part of the static-to- dynamic conversion after APR . The configuration file is open to modification, via an open file format or APIs such as JBits. This is not necessary for DCSTech itself, but would be necessary to modify the bitstreams in order to actually implement the system Since most modern CAD packages fulfil these requirements, with the exception of bitstream access, support for the majority of modern dynamically reconfigurable FPGAs should be possible with only minor alterations in addition to those described in sections 5 and 6. 8. BITSTREAM GENERATION Conventional CAD tools can provide a configuration bitstream for each of the partial circuits produced by DCSTech's dynamic- to-static conversion process. As shown in fig. 3, the partial circuits consist of one configuration representing all the static circuits and a configuration for each dynamic circuit. The static circuits are connected to terminal components that lock the ends of floating connections to dynamic circuits in place. Similarly, floating connections to the static circuits within each dynamic task are locked in place by identically located terminals. These overlying terminal components must be converted to a connection between the two routes, by altering the configuration bitstream. Unless the tools are capable of producing partial configuration files, their output files represent a configuration of each partial circuit on an otherwise unconfigured FPGA. If these files were applied to the FPGA, they would blank out all the existing circuitry. For the system to operate correctly, however, only circuitry that shares resources with the partial circuit to be loaded should be disrupted when it is activated. The partial circuit configurations need to be converted to partial configurations, which reconfigure only the area occupied by a dynamic task within its mutex set zone. A further complication is caused by the lack of control over routing placement noted in section 5. It is possible that routing in a dynamic task will use the same line as routing in a static task. If the dynamic task is then configured onto the array, the routing conflict will cause errors in operation and possibly device damage. The designer must ensure that the routing resources used by each dynamic task are not shared by static tasks or dynamic tasks in other mutex sets. The target device configuration mechanism is another factor in the strategy used to produce partial configurations. The XC6200 allows individual parts of the logic array to be altered; therefore, only parts of the array in the dynamic task bounding-box need be considered. In the Virtex, however, reconfiguration takes place in columns. The smallest unit of configuration data that can be applied is a frame, which configures a subset of the resources in a column. Forty-eight frames are required to completely configure a column [17]. As a result, all the logic and routing in any column which makes up part of a dynamic task bounding-box must be included in the partial reconfiguration bitstreams. Therefore, any static logic or routing that overlaps these columns, must be included in the partial configuration bitstream of that dynamic task otherwise it could be overwritten. For devices that contain bidirectional routing resources, care must be taken not to configure a second driver onto a line during the course of a partial reconfiguration otherwise device damage may occur. One possible solution to this problem is to apply a deactivate configuration, which blanks out existing circuitry on part of the array, prior to loading a new dynamic task, but this would increase the reconfiguration interval. To prevent static circuit disruption, the deactivate configuration needs to contain any static logic within the reconfiguration zone. The generation of partial bitstreams for the Virtex device therefore consists of several steps. Firstly, all the routing resources used by each partial circuit must be evaluated. JRoute [18], part of the JBits SDK includes functions that perform this step. The routing should then be checked for conflicts between circuits that can reside on the array concurrently. The physical bounding-box for each dynamic task (which includes both logic and routing) should then be determined and, from this, the area occupied by each mutex set. The circuitry to be reconfigured for each dynamic task therefore includes all logic and routing within all the columns occupied by the mutex set area. In the Virtex FPGA, the terminal components can be converted to connections, simply by connecting the routes to both sides of the LUT (i.e. merging the routing to and from the two overlapping terminals). This is because the LUT is configured to behave like a wire. Once these processes have been completed, partial bitstreams for the affected FPGA areas can be generated (possibly including deactivate configurations). JBits includes support for this process via JRTR [15]. 9. EXAMPLE COMPLEX NUMBER MULTIPLIER As a simple example to demonstrate the operation of DCSTech, a dynamically reconfigurable constant complex number multiplier is presented. Complex numbers consist of two parts: the real part and the imaginary part, which is a coefficient of j (the square root of -1). The product of two complex numbers is calculated as follows: imag imag a real real a real _ _ _ _ real imag a imag real a imag _ _ _ _ where p_real and p_imag are the real and imaginary parts of the product, p, of complex numbers a and b. The operation therefore requires four multipliers, an adder and a subtractor. In the example, the complex product is formed by multiplying the input complex number, x, by a constant complex coefficient. The constant coefficient values can be hardwired into constant coefficient multipliers potentially saving area and improving performance. A diagram of the system, with a coefficient of 10 j12, is presented in fig. 5. The constant complex coefficient is dependent on the multiplication factors of the four multiplier circuits. Therefore, to support a different coefficient, the four constant coefficient multipliers need to be changed. p_real p_imag x_real x_imag p_real p_imag x_real x_imag Figure 5. Circuit to multiply by 10+j12 The multipliers can be reconfigured to alter their multiplication factor and thus allow the system to support other coefficients. The remaining circuitry does not require alteration in any way. The set of four multipliers therefore forms a dynamic task. One dynamic task is required for each coefficient supported. As the different coefficients are mutually exclusive, the dynamic tasks are all members of the same mutex set and can be assigned the same area of the logic array. Since the registers and adders surrounding the dynamic multipliers are not altered during the reconfigurations, they constitute its static circuitry. Based on these assignments, DCSTech can partition the dynamic design into multiple static designs that can be placed and routed as shown in fig. 1. Figure 7. Layout of the complex multiplier's static circuitry. This consists of the registers, adder and subtractor in fig. 5, with terminal components locking the ends of connections to and from the multipliers in place. A complex multiplier with two dynamic tasks allowing multiplication by the coefficients (10 created. The layout of the (10 task and the static circuitry after APR on a XCV50 is shown in figs. 6 and 7. Fig. 6 Figure 6. Post-APR layout of the 10+j12 dynamic task. This comprises the four multipliers shown in fig. 5, surrounded by terminal components. The areas highlighted in gray indicate terminal components, while the area highlighted in white indicates the dynamic task bounding-box shows evidence of routing exceeding the dynamic task's bounding-box. Similarly, fig. 7 shows that some of the static circuit's routing has been placed within the bounding-box. When implemented, the partial configuration bitstreams should include such stray routing as discussed in section 8. After APR, the circuits timing can be verified. DCSTech is used to reassemble the static parts of the system into a VITAL compliant gate-level model of the dynamic system and create a matching SDF file. A new RIF file is written as part of this process, to match any design hierarchy changes which occurred during synthesis and APR. This model is then further processed by DCSim to produce a dynamic simulation model, making use of the new RIF file. A waveform for the timing simulation of the system is shown in fig. 8. The input number is represented by x_real and x_imag and is set to ns, the n_Rst (reset) input is de-asserted, allowing the multiplier to begin operation. The two status signals at the bottom of fig. 8 indicate the configuration status of the two dynamic tasks. Initially, task activates. The first multiplication is therefore: which matching the result displayed on the outputs p_real and p_imag after 200 ns. At 240 ns task activates. For simplicity, a time of 50 ns is assumed for the reconfiguration. Two clock edges occur during the reconfiguration interval. The exact configuration of the mutex set zone is uncertain during this time. The simulation model therefore puts 'X' on all the dynamic task outputs during this period. These can be seen emerging from the pipeline between 290 and 355 ns. Thereafter, the result of multiplication by (10 j12), which is (-42 + j340), is displayed on the p output. 10. FUTURE WORK For large systems, the use of timing simulation to verify timing is a slow process. Not only does the simulator require long run- times, but also a lot of effort is required to generate a testbench with sufficient test-vectors and coverage. Static Timing Analysis (STA) is a timing verification approach that evaluates the timing of each path through the circuit without test-vectors. These tools can read a variety of file formats including VHDL and SDF. Since the new version of DCSTech produces both these files, it therefore may enable the application of STA to the dynamic design. While this would not take into account the time consumed by reconfigurations, it would allow the verification of all the timing issues that affect circuit performance, such as maximum clock speed, critical path and set-up and hold times. In the DRL design flow presented in this paper, the designer is faced with the problem of partitioning the design at the RT level, rather than a lower level of abstraction. At this level, the exact area occupied by each block is unknown, although it can be estimated approximately. Therefore, some iteration and refinement may be required to obtain a suitable partitioning. A design management tool could simplify this process, by estimating area requirements for each task in the application and presenting the information graphically. Temporal floorplanners for netlists have already been developed. This would be a similar idea but at a higher level of abstraction. Most of the bitstream generation steps outlined in section 8 are currently carried out manually. As the APIs in JBits carry out many of the more complex functions associated with Virtex partial bitstream generation, it is possible to automate the process and this is the focus of future work. 11. CONCLUSIONS This paper shows how the major similarities between the standard CAD tools available for different FPGA architectures can be exploited to implement an easily portable CAD framework for DRL design. The technique relies on a select set of capabilities, supported by most CAD toolsets, within the underlying FPGA platform's supporting tools. From this, automated support for the main stages of the DRL design flow can be provided, including design specification, simulation, synthesis, APR and timing extraction. The final stage of the design flow is partial bitstream generation. The ideas behind partial bitstream generation, which are common across different FPGA families, were outlined. However, the exact method used to produce these bitstreams depends on both the capabilities of the standard CAD tools and the FPGA's configuration interface. The broad similarities Figure 8. A backannotated timing simulation waveform for the dynamically reconfigurable complex multiplier n_Rst x_real x_imag p_imag p_real 10+j12Status 15+j14Status n_Rst x_real x_imag p_imag p_real 10+j12Status 15+j14Status evident in the standard CAD tool support for most platforms are not replicated at this level. Indeed, most vendors provide no mechanism for accessing configuration bitstreams at all, since this compromises design security. As a result, bitstream generation techniques will not port well between families. For the Virtex, however, the availability of the JBits SDK provides convenient access to its bitstream along with a number of functions useful in bitstream generation. 12. --R "Verification of Dynamically Reconfigurable Logic" "Methods of Exploiting Simulation Technology for Simulating the Timing of Dynamically Reconfigurable Logic" "Hardware-Software CoDesign of Embedded Reconfigurable Architectures" "The GARP Architecture and C Compiler" "JHDL-An HDL for Reconfigurable Systems" "Synthesizing RTL Hardware from Java Byte Codes" "Compilation Tools for Run-Time Reconfigurable Designs" "DYNASTY: A Temporal Floorplanning Based CAD Framework for Dynamically Reconfigurable Logic Systems" "New HDL Research Challenges posed by Dynamically Reprogrammable Hardware" "Pebble: A Language for Parameterised and Reconfigurable Hardware Design" "Lava and JBits: From HDL to Bitstreams in Seconds" "Modelling and Synthesis of Configuration Controllers for Dynamically Reconfigurable Logic Systems Using the DCS CAD Framework" "CHASTE: a Hardware/Software Co-design Testbed for the Xilinx XC6200" "JBits: Java based interface for reconfigurable computing" "Partial Run-Time Reconfiguration Using JRTR" "Xilinx Alliance 3.1i Modular Design" "Virtex Series Configuration Architecture User Guide" "JRoute: A Run-Time Routing API for FPGA Hardware" --TR Hardware-software co-design of embedded reconfigurable architectures The Garp Architecture and C Compiler JRoute Pebble Modelling and Synthesis of Configuration Controllers for Dynamically Reconfigurable Logic Systems Using the DCS CAD Framework Partial Run-Time Reconfiguration Using JRTR Verification of Dynamically Reconfigurable Logic Synthesizing RTL Hardware from Java Byte Codes Compilation tools for run-time reconfigurable designs JHDL - An HDL for Reconfigurable Systems --CTR Mahmoud Meribout , Masato Motomura, New design methodology with efficient prediction of quality metrics for logic level design towards dynamic reconfigurable logic, Journal of Systems Architecture: the EUROMICRO Journal, v.48 n.8-10, p.285-310, March Mahmoud Meribout , Masato Motomura, Efficient metrics and high-level synthesis for dynamically reconfigurable logic, IEEE Transactions on Very Large Scale Integration (VLSI) Systems, v.12 n.6, p.603-621, June 2004 Ian Robertson , James Irvine, A design flow for partially reconfigurable hardware, ACM Transactions on Embedded Computing Systems (TECS), v.3 n.2, p.257-283, May 2004

run-time reconfiguration;verification;dynamic reconfiguration;FPGA

503224

The structure and value of modularity in software design.

The concept of information hiding modularity is a cornerstone of modern software design thought, but its formulation remains casual and its emphasis on changeability is imperfectly related to the goal of creating added value in a given context. We need better explanatory and prescriptive models of the nature and value of information hiding. We evaluate the potential of a new theory---developed to account for the influence of modularity on the evolution of the computer industry---to inform software design. The theory uses design structure matrices to model designs and real options techniques to value them. To test the potential utility of the theory for software we apply it to Parnas's KWIC designs. We contribute an extension to design structure matrices, and we show that the options results are consistent with Parnas's conclusions. Our results suggest that such a theory does have potential to help inform software design.

Figure 1: An elementary DSM with three design parameters. Groups of interdependent design parameters are clustered into a proto-module to show that the decisions are managed collectively as a single design task (See Figure 2; the dark lines denote the desired proto-module clusters). In essence, such a proto-module is a composite design parameter. To be a true module, in the lexicon of Baldwin and Clark, there can be no marks in the rows or columns outside the bounding box of its cluster connecting it to other modules or proto-modules in the system. A . X X . Figure 2: A clustered ( prot o-modular ) design. Merely clustering cannot convert a monolithic design comprising one large, tightly coupled proto-module into a modular design. Interdependent parameter cycles must be broken to define modules of reasonable size and complexity. Breaking a cycle between two interdependent parameters like B and C requires an additional step called splitting. The first step in splitting identifies the cause of the cycle-say, a shared data structure definition-and splits it out as its own design parameter (D). B and C no longer cyclically depend on each other, instead taking on a simple hierarchical dependence on D. However, B and C must still wait for the completion of D's design process in order to undertake their own. To counter this, a design as represented by a DSM can be further modularized during the process of splitting by the introduction of design rules. Design rules are additional design parameters that decouple otherwise linked parameters by asserting "global" rules that the rest of the design parameters must follow. Thus, design rules are de facto hierarchical parameters with respect to the other parameters in a DSM. The most prevalent kind of design rule in software is a module interface. For example, for the DSM in Figure 2, an "A interface" rule could be added that asserts that the implementation of B can only access the implementation of A through an interface defined for A. Thus, A could change details of its implementation freely without affecting B, as long as A's interface did not have to be changed as well. The effects of splitting B and C and adding design rules to break the non-modular dependencies is shown in Figure 3. eport CS -2001-13. Submitted for Publication to ESEC/FSE 2001. A-I B-C D-IA B-C D B C A interface B-C D interface A B-C data struct. Figure 3: A modular DSM resulting from splitting, adding design rules, and clustering. In Baldwin and Clark's terminology, a design rule is a (or part of a) visible module, and any module that depends only on design rules is a hidden module. A hidden module can be adapted or improved without affecting other modules by the application of a second operator called substitution. The splitting and substitution operations are examples of six atomic modular operators that Baldwin and Clark introduced to parsimoniously and intuitively describe the operations by which modular designs evolve. The others are augmentation, which adds a module to a system, exclusion, which removes a module, which standardizes a common design element, and porting, which transports a module for use in another system. We do not address these other operators any further in this paper. 3.3 Net Options Value of a Modular Design Not all modularizations are equally good. Thus, in evolving a design, it is useful to be able to evaluate alternative paths based on quantitative models of value. Such models need not be perfect. What is essential is that they capture the most important terms and that their assumptions and operation be known and understood so that analysts can evaluate their predictions. 3.3.1 Introduction to Real Options Concepts Baldwin and Clark's theory is based on the idea that modularity in design multiplies and decentralizes real options that increase the value of a design. A monolithic system can be replaced only as a whole. There is only one option to replace, and exercising it requires that both the good and the bad parts of the new system be accepted. In a sense, the designer has one option on a portfolio of assets. A system that has two modules, by contrast, can be kept as is, or either or both of the new modules can be accepted, for a total of four options. The designer can accept only the good new modules. By contrast, this designer has a portfolio of options on the modules of the system. A key result in modern finance shows that all else remaining equal, a portfolio of options is worth more than an option on a portfolio. Baldwin and Clark's theory defines a model for reasoning about the value added to a base system by modularity. They formalize the options value of each modular operator: How much is it worth to be able to substitute modules, augment, etc. 3.3.2 The Net Options Value of a Modular Design In this paper, we address only substitution options. Splitting a design into n modules increases its base value S0 by a fraction that is obtained by summing the net option values (NOVi) of the resulting options: R NOV is the benefit gained by exercising an option optimally accounting for both payoffs and exercise costs. Baldwin and Clark present a model for calculating NOV. A module creates an opportunity to invest in k experiments to (a) create candidate replacements, (b) each at a cost related to the complexity of the module, and, (c) if any of the results are better than the existing choice, to substitute in the best of them, (d) at a cost that related to the visibility of the module to other modules in the system: First, for module i, sini1/2Q(ki) is the expected benefit by the best of ki independently developed candidate replacements under certain assumptions about the distribution of such values. Ci(ni)ki is the cost to run ki experiments as a function Ci of the module complexity ni. is the cost to replace the module given the number of other modules in the system that directly depend on it, the complexity ni of each, and the cost to redesign each of its parameters. The max picks the number of experiments ki that maximizes the gain from module i. Figure 4 presents a typical scenario: module value-added increases in the number of experiments (better candidates found) until experiment costs meet diminishing returns. The max is the peak. In this case, six experiments maximizes the net gain and is expected to add about 41% value over the existing module.0.40.30.20.10 Number of Experiments Figure 4. The value added by k experiments. The NOVi formula assumes that the value added by a candidate replacement is a random variable normally distributed about the value of the existing module choice (normalized to zero), with a variance si2ni that reflects the technical potential si of the module (the standard deviation on its returns) and the complexity ni of the module. The assumption of a normal distribution is consistent with the empirical observation that high and low outcomes are rare, with middle outcomes more common. The Q(k) represents the expected value of the best of k independent draws from a standard normal distribution, assuming they are positive, and is the maximum order statistic of a sample of size k. 4. OVERVIEW OF ANALYSIS APPROACH As an initial test of the potential for DSM's and NOV to improve software design, we apply the ideas to a reformulation of Parnas's comparative analysis of modularizations of KWIC (a program to compute permuted indices) [1]. The use of KWICas a eport CS -2001-13. Submitted for Publication to ESEC/FSE 2001. benchmark for assessing concepts in software design is well established [7][9]. Parnas presents two modularizations: a traditional strawman based on the sequence of abstract steps in converting the input to the output, and a new one based on information hiding. The new design used abstract data type interfaces to decouple key design decisions involving data structure and algorithm choices so that they could be changed without unduly expensive ripple effects. Parnas then presents a comparative analysis of the changeability of the two designs. He postulates changes and assesses how well each modularization can accommodate them, measured by the number of modules that would have to be redesigned for each change. He finds that the information-hiding modularization is better. He concludes that designers should prefer to use an information hiding design process: begin the design by identifying decisions that are likely to change; then define a module to hide each such decision. Our reformulation of Parnas's example is given in two basic steps. First, we develop DSM's for his two modularizations in order to answer several questions. Do DSM's, as presented by Baldwin and Clark (as well as in the works of Eppinger and Steward [5][10], who invented DSMs), have the expressive capacity to capture the relevant information in the Parnas examples? Do the DSM's reveal key aspects of the designs? Do we learn anything about how to use DSM's to model software? Second, we apply Baldwin and Clark's substitution NOV model to compute quantitative values of the two modularizations, using parameter values derived from information in the DSM's combined with the judgments of a designer. The results are back-of-the-envelope predictions, not precise market valuations; yet they are useful and revealing. We answer two questions. Do the DSM's contain all of the information that we need to justify estimates of values of the NOV parameters? Do the results comport with the accepted conclusions of Parnas? Our evaluation revealed one shortcoming in the DSM framework relative to our needs. DSM's as used by Baldwin and Clark and in earlier work do not appear to model the environment in which a design is embedded. Consequently, we were unable to model the forces that drove the design changes that Parnas hypothesized for KWIC. Thus, DSM's, as defined, did not permit sufficiently rich reasoning about change and did not provide enough information to justify estimates of the environment-dependent technical potential parameters of the NOV model. We thus extended the DSM modeling framework to model what we call environment parameters (EP). We call such models environment and design structure matrices (EDSM). DP's are under the control of the designer. Even design rules can be changed, albeit possibly at great cost. However, the designer does not control EP's. Our extension to the EDSM framework appears to be both novel and useful. In particular, it captures a number of important issues in software design and, at least in the case of the Parnas modularization, it allows us to infer some of Parnas's tacit assumptions about change drivers. The next section presents our DSM's for Parnas's KWIC. Next we present our NOV results. Finally, we close with a discussion. 5. DSM-BASED ANALYSIS OF KWIC For the first modularization, Parnas describes five modules: Input, Circular Shift, Alphabetizing, Output, and Master Control. He concludes, The defining documents would include a number of pictures showing core format, pointer conventions, calling conventions, etc. All of the interfaces between the four modules must be specified before work could begin.This is a modularization in the sense meant by all proponents of modular programming. The system is divided into a number of modules with well-defined interfaces; each one is small enough and simple enough to be thoroughly understood and well programmed [8]. 5.1 A DSM Model of the Strawman Design We surmise Parnas viewed each module interface as comprising two parts: an exported data structure and a procedure invoked by Master Control. We thus took the choice of data structures, procedure declarations, and algorithms as the DP's of this design. The resulting DSM is presented in Figure 5. DP's A, D, G, and J model the procedure interfaces, as design rules, for running the input, shift, sort and output algorithms. B, E, H, and K model the data structure choices as design rules. Parnas states that agreement on them has to occur before independent module implementation can begin. C, F, I, L, and M model the remaining unbound parameters: the choices of algorithms to manipulate the fixed data structures. The DP dependencies are derived directly from Parnas's definitions. A D G J A - Input Type G - Alph Type In Data Alph Data . X X I - Alph Alg Figure 5: DSM for strawman modularization The DSM immediately reveals key properties of the design. First, the design is a modularization, as Parnas claims: designers develop their parts independently as revealed by the absence of unboxed marks in the lower right quadrant of the DSM. Second, only a small part-the algorithms-is hidden and independently changeable. Third, the algorithms are tightly constrained by the data structure design rules. Moreover, the data structures are an interdependent knot (in the upper left quadrant). The shift data structure points into the line data structure; the alphabetized structure is identical to the shifted structure; etc. Change is thus doubly constrained: Not only are the algorithms constrained by the data structure rules, but these rules themselves would be hard to change because of their tight interdependence. eport CS -2001-13. Submitted for Publication to ESEC/FSE 2001. 5.2 A DSM Model of a Pre -Modular Design By declaring the data structures to be design rules, the designer asserts that there is little to gain by letting them change. Parnas's analysis reflects the costly problems that arise when the designer makes a mistake in prematurely accepting such a conclusion and basing a modularization on it. Furthermore, we can see that the design is also flawed because most of the design parameters are off limits to valuable innovation. The designer has cut off potentially valuable parts of the design space. One insight emerging from this work is that there can be value in declining to modularize until the topography of the value landscape is understood. This conclusion is consistent with Baldwin and Clark's view: .designers must know about parameter interdependencies to formulate sensible design rules. If the requisite knowledge isn't there, and designers attempt to modularize anyway, the resulting systems will miss the 'high peaks of value,' and, in the end, may not work at all [p. 260]. Letting the design rules revert to normal design parameters and clustering the data structures with their respective algorithms (because they are interdependent) produces the DSM of Figure 6. This DSM displays the typical diagonal symmetry of outlying marks indicating a non-modular design. We have not necessarily changed any code, but the design (and the design process) is fundamentally different. Rather than a design overconstrained by Draconian design rules, the sense of a potentially complex design process with meetings among many designers is apparent. Innovative or adaptive changes to the circular shifter might have upstream impacts on the Line Store, for example-a kind of change that Parnas did not consider in his analysis. A - In Type In Data In Alg G - Alph Type Alph Data I - Alph Alg Figure 5.3 DSM for the Information -Hiding Design The Line Store that is implicitly bundled with the Input Data is a proto-module that is a prime target for modularization: many other parameters depend on it and vice versa. Splitting the Line Store from the Input and giving each its own interface as a design rule is a typical design step for resolving such a problem. An alternative might be to merely put an interface on the pair and keep them as a single module. However, this DSM does not show that the Line Store is doing double-duty as a buffer for the Input Algorithm as well as serving downstream clients. Thus, it R is more appropriate to split the two. The other proto-modules are modularized by establishing interface design rules for them. The resulting DSM is shown in Figure 7 . It is notable that this design has more hidden information (parameters O down to L in the figure) than the earlier designs. We will see that under our model, this permits more complex innovation on each of the major system components, increasing the net options value of the design. A - In Type G - Alph Type . X X . . X X . . X X . . X X . O - Line Data In Data In Alg Alph Data I - Alph Alg . X X . Figure 7: DSM for information hiding modularization 5.4 Introducing Environment Parameters We can now evaluate the adequacy of DSM's to represent the information needed to reason about modular design in the style of Parnas. We find the DSM to be in part incomplete. In particular, to make informed decisions about the choice of design rules and clustering of design parameters, we found we needed to know how changes in the environment would affect them. For example, we can perceive the value of splitting apart the Line Store and the Input design parameters by perceiving how they are independently affected by different parameters in the environment. For instance, Input is affected by the operating system, but the line store is affected by the size of the corpus. Indeed, the fitness functions found in evolutionary theories of complex adaptive systems, of which Baldwin and Clark's theory is an instance, are parameterized by the environment. Not surprisingly perhaps, we were also finding it difficult to estimate Baldwin and Clark's technical potential term in the NOV formula, which models the likelihood that changing a module will generate value. This, too, is dependent on environmental conditions (e.g., might a change be required). In this paper we address this lack with an extension to the DSM framework. We introduce environment parameters (EP) to model environments. The key property of an EP as distinct from a DP is that the designer does not control the EP. (Designers might be able influence EP's, however.) We call our extended models environment and design structure matrices (EDSM's). Figure 8 presents an EDSM for the strawman KWIC design. eport CS -2001-13. Submitted for Publication to ESEC/FSE 2001. X . A - In Type G - Alph Type -In Data Alph Data In Alg . X X I - Alph Alg . X - Computer . Figure 8: EDSM for strawman modularization The rows and columns of an EDSM are indexed by both EP's and DP's, with the EP's first by convention. The upper left block of an EDSM thus models interactions among EP's; the middle left block, the impact of EP's on the design rules; the lower left block, their impact on the hidden DPs. The lower right block is the basic DSM, partitioned as before to highlight DR's; and the upper right block models the feedback influence of design decisions (DP's) on the environment (EP's). Applying the EDSM concept to Parnas's example reveals that the EDSM provides a clear visual representation of genuine information hiding. In particular, the sub-block of an EDSM where the EP's intersect with the DR's should be blank, indicating that the design rules are invariant with respect to changes in the environment: only the decisions hidden within modules have to change when EP's change, not the design rules-the load bearing walls of the system. We can now make these ideas more concrete in the context of the KWIC case study. Parnas implicitly valued his KWIC designs in an environment that made it likely that certain design changes would be needed. He noted several decisions are questionable and likely to change under many circumstances [p. 305] such as input format, character representation, whether the circular shifter should precompute shifts or compute them on the fly, and similar considerations for alphabetization. Most of these changes are said to depend on a dramatic change in the input size or a dramatic change in the amount of memory. What remains unclear in Parnas's analysis is what forces would lead to such changes in use or the computing infrastructure. We also do not know what other possible changes were ruled out as likely or why. At the time, these programs were written in assembler. Should Parnas have been concerned that a new computer with a new instruction set would render his program inoperable? A dramatic change in input size or memory size could certainly be accompanied by such a change. Y Z X . A - Input Type G - Alph Type Alph Data I - Alph Alg Figure 9: EDSM for inferred proto -modular design X . A - In Type G - Alph Type O - Line Data Alph Data I - Alph Alg . X X . . X X . . X X . . X X . . X X . . X - Computer Figure 10: EDSM for information hiding Modularization By focusing on whether internal design decisions are questionable rather than on the external forces that would bring them into question, the scope of considerations is kept artificially narrow. Not long ago, using ASCII for text would be unquestionable. Today internationalization makes that not so. By turning from design decisions to explicit EP's, such issues can perhaps be discovered and accounted for to produce more effective information-hiding designs. To make this idea concrete, we illustrate it by extending our DSM's for KWIC. We begin by hypothesizing three EP's that Parnas might have selected, and which appear to be implied in his analysis: computer configuration (e.g., device capacity, eport CS -2001-13. Submitted for Publication to ESEC/FSE 2001. corpus properties (input size, language-e.g., Japanese); and user profile (e.g., computer savvy or not, interactive or offline). Figure 8, 9, and 10 are EDSM's for the strawman, pre- modular, and information hiding designs, respectively. The key characteristic of the strawman EDSM is that the DR's are not invariant under the EP's. We now make a key observation: The strawman is an information-hiding modularization in the sense of Baldwin and Clark: designers can change non-DR DP's (algorithms) independently; but it is not an information-hiding design in the sense of Parnas. Basic DSM's alone are insufficient to represent Parnas's idea. We could have annotated the DP's with change probabilities, but we would still miss the essence: the load-bearing walls of an information hiding design (DR's) should be invariant with respect to changes in the environment. Our EDSM notation expresses this idea clearly. Figure 9 is the EDSM for the pre-modular design in which the data structures are not locked down as DR's. The remaining DR's (the procedure type signatures) are invariant with the EP's, but the extensive dependencies between proto-module DR's suggest that changes in EP's will have costly ripple effects. The design evolution challenge that this EDSM presents is to split the proto-modules in a way that does not create new EP-dependent DR's. Figure models the result: Parnas's information hiding design. The EDSM highlights the invariance of the DRs under the EPs in the sector where the EPs meet the DRs. 6. NOV-BASED ANALYSIS OF KWIC We can now apply the NOV model to model how much the flexibility is worth in both Parnas designs as a fraction of the value of the base system, taking Parnas's notion of information hiding into account. This analysis is illustrative, of course, and the outputs are a function of the inputs. We justify our estimates of the model parameters using the EDSM's and reasonable back- of-the-envelope assumptions. A benefit of the mathematical model is that it supports rigorous sensitivity analysis. Such an analysis is beyond the scope of this paper; but we will pursue this issue in the future. We make the following assumptions and use the following notations in our analysis: N is the number of design parameters in a given design. For the proto-modular and strawman modularizations, 13. In the information hiding design Given a module of p parameters, its complexity is p/N. The value of one experiment on an unmodularized design, sN1/2Q(1)= 1, is the value of the original system. The design cost c=1/N of each design parameter is the same, and the cost to redesign the whole system is cN 1. The visibility cost of a module i of size n is icn. One experiment on an unmodularized system breaks even: Balwin and Clark make the break-even assumption for an example in their book [1]. For a given system size, it implies a choice of technical potential for an unmodularized design: in our R case, 2.5. We take this as the maximum technical potential of any module in a modularized version. This assumption for the unmodularized KWIC is a modeling assumption, not a precisely justified estimate. In practice, a designer would have to justify the choices of parameter values. The model of Baldwin and Clark in quite sensitive to technical potential, but they give little guidance in how to estimate it. We have observed that the environment is what determines whether variants on a design are likely to have added value. If there is little added value to be gained by replacing a module in a given environment, no matter how complex it is, that means the module has low technical potential. We chose to estimate the technical potential of each module as the system technical potential scaled by the fraction of the EP's relevant to the module. We further scaled the technical potential of the modules in the strawman design by 0.5, for two reasons. First, about half of the interactions of the EPs with the strawman design are with the design rules (but as we will see, their visibility makes the cost to change them prohibitive). Second- and more of a judgment call-the hidden modules in this design (algorithms) are tightly constrained by the design rules (data structures that are assumed not to change). There would appear to be little to be gained by varying the algorithm implementations, alone. Figure 11 shows our assumptions about the technical potential of the modules in the strawman and information-hiding designs. Strawman Info Hiding Module Name sigma Z sigma Z Design Rules 2.5 1 Line Storage NA NA 1.6 0 Input 1.25 0 2.5 0 Circular Shift 1.25 0 2.5 0 Alphabetizing 1.25 0 2.5 0 Output Figure 11. Assumed Technical Potential and Visibility Figures 12 present the NOV data per module and Figure 13 the corresponding plots for the information hiding design. Figure 14 presents the plots for the strawman. The option value of each module is the value at the peak. We omit this disaggregated data for the strawman design. What matters is the bottom line: Summing the module NOV's gives that the system NOV is 0.26 for the strawman design but 1.56 for the information-hiding design. These numbers are percentages of the value of the non- modularized system, which has base value 1. Thus the value of the system with the information-hiding design is 2.6 times that of the system with the unmodularized design, and the strawman's is worth only 1.26 times as much. Thus, the information-hiding version of the system is twice as valuable as the strawman. Ignoring the base value and focusing just on modularity, we observe that the information-hiding design provides 6 times more value in the form of modularitythan the strawman's design. eport CS -2001-13. Submitted for Publication to ESEC/FSE 2001. Baldwin and Clark acknowledge that designing modularizations is not free; but, once done, the costs are amortized over future evolution; so the NOV model ignores those costs. Accounting for them is important, but not included in our model. It is doubtful they are anywhere near 150% of the system value. On the other hand, they would come much closer to 26%, which would tend to further reduce the value added by the strawman modularization. Design Rules 0 -1.3 -1.6 -1.9 -2.25 -2.56 -2.88 -3.2 -3.5 -3.81 -4.1 0 Alpha MsCon. Fig 12. Option Values for Information Hiding Design0.40 -0.2 -0.4 Number of Experiments Line Store Input CirShift Alpha Output Master Cont. Figure 13: Options Values for Information Hiding Design0.1-0.1 -0.2 -0.3 -0.4 Number of Experiments Input CirShift Alpha Output MsControl Figure 14: Options Values for Strawman Design 7. DISCUSSION AND CONCLUSION Parnas's information-hiding criterion for modularity has been enormously influential in computer science. Because it is a qualitative method lacking an independent evaluation criterion, it is not possible to perform a precise comparison of differing designs deriving from the same desiderata. This paper is a novel application of Baldwin and Clark's options- theoretic method of modular design and valuation to the subject of information-hiding modularity. Our goal was to lend insight into both information-hiding modularity and the ability of options theory to capture Parnas's intent of designing for change. We have provided an early validation of the application of their method to software design by reformulating Parnas's KWIC modularizations in the Baldwin and Clark framework. Baldwin and Clark's method has two main components, the Design Structure Matrix (DSM) and the Net Option Value formula (NOV). DSM's provide an intuitive, qualitative framework for design. NOV quantifies the consequences of a particular design, thus permitting a precise comparison of differing designs of the same system. We have shown that these tools provide significant insight into the modularity in the design of software. Yet, precisely modeling Parnas's information-hiding criterion requires explicitly modeling the environment-the context in which the software is intended to be used-in order to capture the notion of design stability in the face of change. We model the environment by extending DSM's to include environment parameters alongside the traditional design parameters. The environment parameters then inform the estimation of the technical potential in the NOV computation. In the process, we learned that Parnas had largely conceived of change in terms of intrinsic properties of the design, rather than in terms of the properties of the environment in which the software is embedded. With these extensions to the Baldwin and Clark model, we were able to both model the Parnas designs and quantitatively show- under a set of assumptions-that the information-hiding design is indeed superior, consistent with the accepted results of Parnas. This result has value in at least three dimensions. First, it provides a quantitative account of the benefits of good design. Second, it provides limited but significant evidence that such models have the potential to aid technical decision-making in design with value added as an explicit objective function. This paper is thus an early result in the emerging area of strategic software design [4], which aims for a descriptive and prescriptive theoretical account of software design as a value-maximizing investment activity. Third, the result supports further investigation of implications that follow from acceptance of such a model. For example, because the value of an option increases with technical potential (risk), modularity creates seemingly paradoxical incentives to seek risks in software design, provided they can be managed by active creation and exploitation of options. The paradox is resolved in large part by the options model, which clarifies that one has the right, but not a requirement, to exercise an option, thus the downside risk (cost) is largely limited to the purchase of the option itself. In the introduction, we also raised the question of when is the right time to modularize or commit to a software architecture? eport CS -2001-13. Submitted for Publication to ESEC/FSE 2001. Parnas's method says to write down the design decisions that are likely to change and then design modules to hide them. This implicitly encourages programmers to modularize at the early stages of design. The NOV calculations of the two KWIC modularizations make the possible consequences clear: without knowledge of the environment parameters, a designer might rush in to implement the strawman design, effectively sacrificing the opportunity to profit from the superior modularization. Yet, designers often do not have the luxury to wait until there is sufficient information to choose the optimal modularization. It may be difficult to precisely estimate how the environment is going to change-innovation and competitive marketplaces are hard to predict. Moreover, many of the best ideas come from the users of the software, so uncertainty is almost certain until the product is released. New design techniques that create options to delay modularizing until sufficient information is available might be explored as a possible solution to this conundrum. The inclusion of environment parameters in the design process has additional implications. For example, making the most of these parameters requires being able to sense when they are changing and to influence them (slow their change) when possible. Careful design of the coupling between the development process and the environment is critical in strategic software design. For example, for parameters whose values are subject to change, sensor technologies-perhaps as simple as being on the mailing list of a standards-setting committee-can help to detect changes and report them to the designers in a timely fashion. Conversely, lobbying a standards-setting organization to, say, deprecate interfaces rather than change them outright can slow environmental change. Thus, accommodating environmental change is not limited to just anticipating change, as originally stated by Parnas, but includes more generally both responsiveness to change and manipulation of change. This paper represents a first step in the validation of Baldwin and Clark's option-theoretic approach for quantifying the value of modularity in software. Additional studies are required to adequately validate the theory and provide insight into its practical application. Also, in our paper study, we found it difficult to estimate the technical potentials of the modules, despite the added resolution provided by the environment parameters. Validation in an industrial project would not only provide realistic scale, but it would also have considerable historical data to draw upon for the computation of NOV. Such studies would help move the field of software design further down the path to having powerful quantitative models for design. ACKNOWL EDGMENTS This work was supported in part by the National Science Foundation under grants CCR-9804078, CCR-9970985, and ITR- 0086003. Our discussions with graduate students at the University of Virginia in CS 851, Spring 2001, at the have been very helpful. --R Design Rules: The Power of Modularity Using Tools to Compose Systems. On the Criteria to be Used in Decomposing System into Modules Candidate Model Problems in Software Architecture. --TR Using Tool Abstraction to Compose Systems Software design Extreme programming explained Software economics On the criteria to be used in decomposing systems into modules Design Rules Value based software reuse investment Software Design Decisions As Real Options --CTR Verifying design modularity, hierarchy, and interaction locality using data clustering techniques, Proceedings of the 45th annual southeast regional conference, March 23-24, 2007, Winston-Salem, North Carolina Yuanyuan Song, Adaptation Hiding Modularity for Self-Adaptive Systems, Companion to the proceedings of the 29th International Conference on Software Engineering, p.87-88, May 20-26, 2007 Neeraj Sangal , Ev Jordan , Vineet Sinha , Daniel Jackson, Using dependency models to manage complex software architecture, ACM SIGPLAN Notices, v.40 n.10, October 2005 Mikio Aoyama , Sanjiva Weerawarana , Hiroshi Maruyama , Clemens Szyperski , Kevin Sullivan , Doug Lea, Web services engineering: promises and challenges, Proceedings of the 24th International Conference on Software Engineering, May 19-25, 2002, Orlando, Florida Barry Boehm , Li Guo Huang, Value-Based Software Engineering: A Case Study, Computer, v.36 n.3, p.33-41, March Sushil Krishna Bajracharya , Trung Chi Ngo , Cristina Videira Lopes, On using Net Options Value as a value based design framework, ACM SIGSOFT Software Engineering Notes, v.30 n.4, July 2005 John M. Hunt , John D. McGregor, A series of choices variability in the development process, Proceedings of the 44th annual southeast regional conference, March 10-12, 2006, Melbourne, Florida Rami Bahsoon , Wolfgang Emmerich, Economics-Driven Software Mining, Proceedings of the First International Workshop on The Economics of Software and Computation, p.3, May 20-26, 2007 Yuanfang Cai , Kevin J. Sullivan, A value-oriented theory of modularity in design, ACM SIGSOFT Software Engineering Notes, v.30 n.4, July 2005 Sunny Huynh , Yuanfang Cai, An Evolutionary Approach to Software Modularity Analysis, Proceedings of the First International Workshop on Assessment of Contemporary Modularization Techniques, p.6, May 20-26, 2007 Yuanfang Cai , Kevin J. Sullivan, Simon: modeling and analysis of design space structures, Proceedings of the 20th IEEE/ACM international Conference on Automated software engineering, November 07-11, 2005, Long Beach, CA, USA Yuangfang Cai , Sunny Huynh, An Evolution Model for Software Modularity Assessment, Proceedings of the 5th International Workshop on Software Quality, p.3, May 20-26, 2007 Kevin Sullivan , William G. Griswold , Yuanyuan Song , Yuanfang Cai , Macneil Shonle , Nishit Tewari , Hridesh Rajan, Information hiding interfaces for aspect-oriented design, ACM SIGSOFT Software Engineering Notes, v.30 n.5, September 2005 Barry Boehm, Value-based software engineering, ACM SIGSOFT Software Engineering Notes, v.28 n.2, March Cristina Videira Lopes , Sushil Krishna Bajracharya, An analysis of modularity in aspect oriented design, Proceedings of the 4th international conference on Aspect-oriented software development, p.15-26, March 14-18, 2005, Chicago, Illinois Barry Boehm, Value-based software engineering: reinventing, ACM SIGSOFT Software Engineering Notes, v.28 n.2, March

modularity;design structure matrix;real options;software

503244

Coverage criteria for GUI testing.

A widespread recognition of the usefulness of graphical user interfaces (GUIs) has established their importance as critical components of today's software. GUIs have characteristics different from traditional software, and conventional testing techniques do not directly apply to GUIs. This paper's focus is on coverage critieria for GUIs, important rules that provide an objective measure of test quality. We present new coverage criteria to help determine whether a GUI has been adequately tested. These coverage criteria use events and event sequences to specify a measure of test adequacy. Since the total number of permutations of event sequences in any non-trivial GUI is extremely large, the GUI's hierarchical structure is exploited to identify the important event sequences to be tested. A GUI is decomposed into GUI components, each of which is used as a basic unit of testing. A representation of a GUI component, called an event-flow graph, identifies the interaction of events within a component and intra-component criteria are used to evaluate the adequacy of tests on these events. The hierarchical relationship among components is represented by an integration tree, and inter-component coverage criteria are used to evaluate the adequacy of test sequences that cross components. Algorithms are given to construct event-flow graphs and an integration tree for a given GUI, and to evaluate the coverage of a given test suite with respect to the new coverage criteria. A case study illustrates the usefulness of the coverage report to guide further testing and an important correlation between event-based coverage of a GUI and statement coverage of its software's underlying code.

INTRODUCTION The importance of graphical user interfaces (GUIs) as critical components of today's software is increasing with the recognition of their usefulness. The widespread use of GUIs has led to the construction of more and more complex GUIs. Although the use of GUIs continues to grow, GUI testing has, until recently, remained a neglected research area. Because GUIs have characteristics different from conventional software, techniques developed to test conventional software cannot be directly applied to GUI testing. Recent advances in GUI testing have focused on the development of test case generators [8, 11, 12, 14, 18, 6] and test oracles [9] for GUIs. However, development of coverage criteria for GUIs has not been addressed. Coverage criteria are sets of rules used to help determine whether a test suite has adequately tested a program and to guide the testing process. The most well-known coverage criteria are statement coverage, branch coverage, and path coverage, which require that every statement, branch and path in the program's code be executed by the test suite respectively. However such criteria do not address the adequacy of GUI test cases for a number of reasons. First, GUIs are typically developed using instances of precompiled elements stored in a library. The source code of these elements may not always be available to be used for coverage evalu- ation. Second, the input to a GUI consists of a sequence of events. The number of possible permutations of the events may lead to a large number of GUI states and for adequate testing, a GUI event may need to be tested in a large number of these states. Moreover, the event sequences that the GUI must be tested for are conceptually at a much higher level of abstraction than the code and hence cannot be obtained from the code. For the same reason, the code cannot be used to determine whether an adequate number of these sequences have been tested on the GUI. The above challenges suggest the need to develop coverage criteria based on events in a GUI. The development of such coverage criteria has certain requirements. First, since there are a large number of possible permutations of GUI events, the GUI must be decomposed into manageable parts. GUIs, by their very nature, are hierarchical and this hierarchy may be exploited to identify groups of GUI events that can be tested in isolation. Hence, each group forms a unit of test- ing. Such a decomposition also allows coverage criteria to be developed for events within a unit. Intuitively, a unit of testing has a well-defined interface to the other parts of the software. It may be invoked by other units when needed and then terminated. For example, when performing code-based testing, a unit of testing may be a basic block, procedure, an object, or a class, consisting of statements, branches, etc. Next, interactions among units must be identified and coverage developed to determine the adequacy of tested interac- tions. Second, it should be possible to satisfy the coverage criterion by a finite-sized test suite. The finite applicability [20] requirement holds if a coverage criterion can always be satisfied by a finite-sized test suite. Finally, the test designer should recognize whether a coverage criterion can ever be fully satisfied [16, 17]. For example, it may not always be possible to satisfy path coverage because of the presence of infeasible paths, which are not executable because of the context of some instructions. Detecting infeasible paths in general is a NP complete problem. No test case can execute along an infeasible path, perhaps resulting in loss of coverage. Infeasibility can also occur in GUIs. Similar to infeasible paths in code, static analysis of the GUI may not reveal infeasible sequences of events. For example, by performing static analysis of the menu structure of MS Word- pad, one may construct a test case with Paste as the first event. However, experience of using the software shows that such a test case will not execute since Paste is highlighted only after a Cut or Copy. 1 In this paper, we define a new class of coverage criteria called event-based coverage criteria to determine the adequacy of tested event sequences, focusing on GUIs. The key idea is to define the coverage of a test suite in terms of GUI events and their interactions. Since the total number permutations of event sequences in any non-trivial GUI is extremely large, the GUI's hierarchical structure is exploited to identify the important event sequences to be tested. The GUI is decomposed into GUI components, 2 each of which is a unit of testing. Events within a component do not interleave with events in other components without explicit invocation or termination events. Because of this well-defined behavior, a component may be tested in isolation. Two kinds of coverage criteria are developed from the decomposition intra-component coverage criteria for events within a component and inter-component coverage criteria for events among components. Intra-component criteria include event, event-selection, and length-n event-sequence coverage. Inter-component criteria include invocation, invocation-termination and length-n event-sequence coverage. A GUI component is represented by a new structure called an event-flow graph that identifies events within a component. The interactions among GUI components are captured by a representation called the integration tree. We present algorithms to automatically construct event-flow graphs and the integra- 1 Note that Paste will be available if the ClipBoard is not empty, perhaps because of an external software. External software is ignored in this simplified example. components should not be confused with GUI elements that are used as building blocks during GUI develop- ment. We later provide a formal definition of a GUI component tion tree for a given GUI and to evaluate intra- and inter-component coverage for a given test suite. We present a case study to demonstrate the correlation between event-based coverage of our version of WordPad's GUI and the statement coverage of its underlying code for a test suite. The important contributions of the coverage method presented in this paper include: 1. a class of coverage criteria for GUI testing in terms of GUI events. 2. the identification of a GUI component, useful for GUI testing. 3. a representation of a GUI component called an event- flow graph that captures the flow of events within a component and a representation called the integration tree to identify interactions between components. 4. an automated technique to decompose the GUI into interacting components and coverage criteria for intra-component and inter-component testing. 5. a technique to compute the coverage of a given test suite. 6. a case study demonstrating the correlation between coverage in terms of events and code. In the next section we present a classification of GUI events and use the classification to identify GUI components. In Section 3 we present coverage criteria for event interactions within a component and between components. Section 4 presents algorithms to construct event-flow graphs and an integration tree for a given GUI and then evaluate intra- and inter-component coverage of the GUI for a given test suite. In Section 5, we present details of a case study conducted on our version of the WordPad software. Lastly, Section 6 presents related work and in Section 7 we conclude with a discussion of ongoing and future work. 2. STRUCTURE OF GUIS A GUI uses one or more metaphors for objects familiar in real life, such as buttons, menus, a desktop, and the view through a window. The software user performs events to interact with the GUI, manipulating GUI objects as one would real objects. These events may cause deterministic changes to the state of the software that may be reflected by a change in the appearance of one or more GUI objects. Moreover, GUIs, by their very nature, are hierarchical. This hierarchy is reflected in the grouping of events in windows, dialogs, and hierarchical menus. For example, all the "options" in MS Internet Explorer can be set by interacting with events in one window of the software's GUI. The important characteristics of GUIs include their graphical orientation, event-driven input, hierarchical structure, the objects they contain, and the properties (attributes) of those objects. Formally, a GUI may be defined as follows: Graphical User Interface (GUI) is a hierar- chical, graphical front-end to a software that accepts user-generated and system-generated events, from a fixed set of events, as input and produces deterministic graphical output. A GUI contains graphical objects; each object has a fixed set of properties. At any time during the execution of the GUI, these properties have discrete values, the set of which constitutes the state of the GUI. 2 The above definition specifies a class of GUIs that have a fixed set of events with deterministic outcome that can be performed on objects with discrete valued properties. This paper develops coverage criteria for the class of GUIs defined above. In this section, a new unit of testing called a GUI component is defined. It consists of a number of events, selections, invocations, and terminations that restrict the focus of a GUI user. The user interacts with a component by explicitly invoking it, performing events, and then terminating the component. Note that since events within a component cannot be interleaved with events in other components, the interaction among events within a component may be tested independently of other components. A classification of GUI events is used to identify GUI components. GUI events may be classified as: Menu-open events open menus, i.e., they expand the set of GUI events available to the user. By definition, menu-open events do not interact with the underlying software. The most common example of menu-open events are generated by buttons that open pull-down menus, e.g., File and Edit. Restricted-focus events open modal windows, i.e., windows that have the special property that once invoked, they monopolize the GUI interaction, restricting the focus of the user to a specific range of events within the window until the window is explicitly terminated by a termination event. Preference setting is an example of restricted-focus events in many GUI systems; the user clicks on Edit and Preferences, a window opens and the user then spends time modifying the prefer- ences, and finally explicitly terminates the interaction by either clicking OK or Cancel. Unrestricted-focus events open modeless windows that do not restrict the user's focus; they merely expand the set of GUI events available to the user. Note that the only difference between menu-open events and unrestricted-focus events is that the latter open windows that have to be explicitly terminated. For exam- ple, in the MS PowerPoint software, the Basic Shapes are displayed in an unrestricted-focus window. System-interaction events interact with the underlying software to perform some action; common examples include cutting and pasting text, and opening object windows. Termination events close modal windows; common examples include Ok and Cancel. At all times during interaction with the GUI, the user interacts with events within a limited focus. This limited focus consists of a restricted-focus window X and a set of unrestricted-focus windows that have been invoked, either directly or indirectly by X. The limited focus remains in place until X is explicitly terminated using a termination event such as OK, Cancel, or Close. Intuitively, the events within the limited focus form a GUI component. component C is an ordered pair (RF, UF), where RF represents a modal window in terms File Edit Help Open Save Cut Copy Paste and Help Figure 1: An Event-flow Graph for a Part of MS WordPad. of its events and UF is a set whose elements represent modeless windows also in terms of their events. Each element of UF is invoked either by an event in UF or A common example of a GUI component is the FileOpen modal window (and its associated modeless windows) found in most of today's software. The user interacts with events within this component, selects a file and terminates the component by performing the Open event (or sometimes the Cancel event). Formally, a GUI component can be represented as a flow graph. An event-flow graph for a GUI component C is a 4-tuple !V, E, B, I? where: 1. V is a set of vertices representing all the events in the component. Each v 2V represents an event in C. 2. is a set of directed edges between vertices. We say that event e i follows e j iff e j may be performed immediately after e i . An edge the event represented by vy follows the event represented by vx . 3. is a set of vertices representing those events of C that are available to the user when the component is first invoked. 4. I ' V is the set of restricted-focus events of the component.An example of an event-flow graph for a part of the Maincomponent of MS WordPad is shown in Figure 1. At the top are three vertices (File, Edit, and Help) that represent part of the pull-down menu of MS WordPad. They are menu-open events that are available when the Main component is first invoked. Hence they form the set B. Once File has been performed in WordPad any of Edit, Help, Open, and Save may be performed. Hence there are edges in the event-flow graph from File to each of these events. Note that Open is shown with a dashed oval. We use this representation for restricted-focus events, i.e., events that invoke components. Similarly, About and Contents are also 3 We assume that all GUIs have a Main component, i.e., the component that is presented to the user when the GUI is first invoked. Main FileNew FileOpen Print FormatFont Properties FileSave PageSetup ViewOptions Figure 2: An Integration Tree for a Part of MS WordPad. restricted-focus events, i.e., for this component, I = fOpen, About, Contentsg. Other events (i.e., Save, Cut, Copy, and Paste) are all system-interaction events. After any of these is performed in MS WordPad, the user may perform File, Edit, or Help, shown as edges in the event-flow graph. Once all the components of the GUI have been represented as event-flow graphs, the remaining step is to identify their interactions. Testing interactions among components is also an area of research in object-oriented software testing [5] and inter-procedural data-flow testing [4]. The identification of interactions among objects and procedures is aided by structures such as function-decomposition trees and call-graphs [4]. Similarly, we develop a structure to identify interactions among components. We call this structure an integration tree because it shows how the GUI components are integrated to form the GUI. Formally, an integration tree is defined as: An integration tree is a 3-tuple where N is the set of components in the GUI, R 2 N is a designated component called the Main component. We say that a component Cx invokes component Cy if Cx contains a restricted-focus event ex that invokes Cy . B is the set of directed edges showing the invokes relation between components, i.e., (Cx ; Cy invokes Cy . 2 Figure 2 shows an example of an integration tree representing a part of the MS WordPad's GUI. The nodes represent the components of the MS WordPad GUI and the edges represent the invokes relationship between the components. Main is the top-level component that is available when Word- Pad is invoked. Other components' names indicate their functionality. For example, FileOpen is the component of WordPad used to open files. The tree in Figure 2 has an edge from Main to FileOpen showing that Main contains an event, namely Open (see Figure 1) that invokes FileOpen. 3. COVERAGE CRITERIA Having created representations for GUI components and events among components, we are ready to define the coverage criteria. We will first define coverage criteria for events within a component, i.e., intra-component coverage criteria and then for events among components, i.e., inter-component criteria. 3.1 Intra-component Coverage In this section, we define several coverage criteria for events and their interactions within a component. We first formally define an event sequence. An event-sequence is en ? where All the new coverage criteria that we define next are based on event-sequences. 3.1.1 Event Coverage Intuitively, event coverage requires each event in the component to be performed at least once. Such a requirement is necessary to check whether each event executes as expected. set P of event-sequences satisfies the event coverage criterion if and only if for all events v 2 V, there is at least one event-sequence event v is in p. 2 3.1.2 Event-selection Coverage Another important aspect of GUI testing is to check the interactions among all possible pairs of events in the com- ponent. However, we want to restrict the checks to pairs of events that may be performed in a sequence. We focus on the possible implicit selection of events that the user may encounter during interaction with the GUI. Definition: The event-selections for an event e is the set In this criterion, we require that after an event e has been performed, all event-selections of e should be executed at least once. Note that this requirement is equivalent to requiring that each element in E be covered by at least one test case. set P of event-sequences satisfies the event- selection coverage criterion if and only if for all elements there is at least one event-sequence contains 3.1.3 Length-n Event-sequence Coverage In certain cases, the behavior of events may change when performed in different contexts. In such cases event coverage and event-selection coverage on their own are weak requirements for sufficient testing. We now define a criterion that captures the contextual impact. We first formally define a context. Definition: The context of an event en in the event-sequence Intuitively, the context for an event e is the sequence of events performed before e. An event may be performed in an infinite number of contexts. For finite applicability, we define a limit on the length of the event-sequence. Hence, we define the length-n event-sequence criterion. set P of event-sequences satisfies the length-n event-sequence coverage criterion if and only if P contains all event-sequences of length equal to n. 2 Note the similarity of this criterion to the length-n path coverage criterion defined by Gourlay for conventional software [2], which requires coverage of all subpaths in the program's flow-graph of length less than or equal to n. As the length of the event-sequence increases, the number of possible contexts also increases. 3.2 Subsumption A coverage criterion C1 subsumes criterion C2 if every test suite that satisfies C1 also satisfies C2 [13]. Since event coverage and event-selection coverage are special cases of length-n event-sequence coverage, i.e., length 1 event-sequence and length 2 event-sequence coverage respectively, it follows that length-n event-sequence coverage subsumes event and event-selection coverage. Moreover, if a test suite satisfies event-selection coverage, it must also satisfy event coverage. Hence, event-selection subsumes event coverage. 3.3 Inter-component Criteria The goal of inter-component coverage criteria is to ensure that all interactions among components are tested. In GUIs, the interactions take the form of invocation of components, termination of components, and event-sequences that start with an event in one component and end with an event in another component. 3.3.1 Invocation Coverage Intuitively, invocation coverage requires that each restricted- focus event in the GUI be performed at least once. Such a requirement is necessary to check whether each component can be invoked. set P of event-sequences satisfies the invocation coverage criterion if and only if for all restricted- focus events i 2 I, where I is the set of all restricted- focus events in the GUI, there is at least one event- sequence event i is in p. 2 Note that event coverage subsumes invocation coverage since it requires that all events be performed at least once, including restricted-focus events. 3.3.2 Invocation-termination Coverage It is important to check whether a component can be invoked and terminated. Definition: The invocation-termination set IT of a GUI is the set of all possible length 2 event sequences ! component Cx and e j terminates component Cx , for all components Cx 2 N .Intuitively, the invocation-termination coverage requires that all length 2 event sequences consisting of a restricted-focus event followed by the invoked component's termination events be tested. set P of event-sequences satisfies the invocation- termination coverage criterion if and only if for all i 2 IT , there is at least one event-sequence that i is in p. 2 Satisfying the invocation-termination coverage criterion assures that each component is invoked at least once and then terminated immediately, if allowed by the GUI's specifica- tions. For example, in WordPad, the component FileOpen is invoked by the event Open and terminated by either Open or Cancel. Note that WordPad's specification do not allow Open to terminate the component unless a file has been se- lected. On the other hand, Cancel can always be used to terminate the component. v: Vertex or Event)f 1 system-interaction 7 return(B of Invoking Invoked component); 12 return(B of Invoked component); 14 Figure 3: Computing follow set(v) for a Vertex v. 3.3.3 Inter-component Length-n Event-sequence Cov- erage Finally, the inter-component length-n event-sequence coverage criterion requires testing all event-sequences that start with an event in one component and end with an event in another component. Note that such an event-sequence may use events from a number of components. A criterion is defined to cover all such interactions. set P of event-sequences satisfies the inter-component length-n event-sequence coverage criterion for components C1 and C2 if and only if P contains all length-n event-sequences that may belong to C1 or C2 or any other component C i . 2 Note that the inter-component length-n event-sequence coverage subsumes invocation-termination coverage since length-n event sequences also include length 2 sequences. 4. EVALUATING COVERAGE Having formally presented intra- and inter-component coverage criteria, we now present algorithms to evaluate the coverage of a test suite using these criteria. In this section, we present algorithms to evaluate the coverage of the GUI for a given test suite. We show how to construct an event- flow graph and use it to evaluate intra-component coverage. Then we show how to construct an integration tree and use it to evaluate inter-component coverage. 4.1 Construction of Event-flow Graphs The construction of event-flow graphs is based on the structure of the GUI. The classification of events in the previous section aids the automated construction of the event-flow graphs, which we describe next. For each v 2 V, we define follow set(v) as the set of all events vx such that vx follows v. Note that follow set(v) is the set of outgoing edges in the event-flow graph. We determine follow set(v) using the algorithm in Figure 3 for each vertex v. The recursive algorithm contains a switch structure that assigns follow set(v) according to the type of each event. If the type of the event v is a menu-open event (line represents events that are available when a component is invoked) then the user may either perform v again, its sub-menu choices, or any event in B (line 4). However, if v 62 B then the user may either perform all sub-menu choices of v, v itself, or all events in follow set(parent(v)) (line 6). We define parent(v) as any event that makes v available. If v is a system-interaction event, then after performing v, the GUI reverts back to the events in B (line 8). If v is an exit event, i.e., an event that terminates a component, then follow set(v) consists of all the top-level events of the invoking component (line 10). If the event type of v is an unrestricted-focus event then the available events are all top-level events of the invoked component available as well as all events of the invoking component (line 12). Lastly, if v is a restricted-focus event, then only the events of the invoked component are available. 4.2 Evaluating Intra-component Coverage Having constructed an event-flow graph, we are now ready to evaluate the intra-component coverage of any given test suite using the elements of this graph. Figure 4 shows a dynamic programming algorithm to compute the percentage of length-n event-sequences tested. The final result of the algorithm is Matrix, where Matrix i;j is the percentage of length-j event-sequences tested on component i. The main algorithm is ComputePercentageTested. In this algorithm, two matrices are computed (line 6,7). Count i;j is the number of length-j event-sequences in component i that have been covered by the test suite T (line 6). Total i;j is the total number of all possible length-j event-sequences in component i (line 7). The subroutine ComputeCounts calculates the elements in count matrix. For each test case in T, ComputeCounts finds all possible event-sequences of different lengths (line 19.21). The number of event- sequences of each length are counted in (lines 22, 23). In- tuitively, the ComputeTotals subroutine starts with single-length event-sequences, i.e., individual events in the GUI (lines 31.33). Using follow set (line 38), the event- sequences are lengthened one event at each step. A counter keeps track of the number of event-sequences created (line 39). Note that since ComputeCounts takes a union of the event sequences, there is no danger of counting the same event sequence twice. The result of the algorithm is Matrix, the entries of which can be interpreted as follows: Event Coverage requires that individual events in the GUI be exercised. These individual events correspond to length 1 event-sequences in the GUI. Matrix j;1 j 2 S represents the percentage of individual events covered in each component. Event-selection Coverage requires that all the edges of the event-flow graph be covered by at least one test case. Each edge is effectively captured as a length-2 event-sequence. Matrix j;2 j 2 S represents the percentage of branches covered in each component j. Length-n Event-sequence Coverage is available directly from Matrix. Each column i of Matrix represents the number of length-i event-sequences in the GUI. S: Set of Components; 2 T: Test M: Maximum Event-sequence Length) 4 fcount /\Gamma ComputeCounts(T, S, M); 5,6 count i;j is the tested number of length-j event-sequences in component i */ total /* total i;j is the total number of length-j event-sequences in component i */ Matrix i;j /\Gamma (count i;j /total i;j ) \Theta 100; 10 return(Matrix)g 11 T: Test M: Maximum Event-sequence Length) 14 A /\Gamma fg; /* Empty Set */ 17 A /* count number of sets of length j */ count i;j /\Gamma NumberOfSetsOfLength(S, j); 23 return(count)g S: Set of Components; 26 M: Maximum Event-sequence Length) 27 28 x total j;k /\Gamma total j;k newfreq i /\Gamma 0; 44 return(total)g Figure 4: Computing Percentage of Tested Length-n Event-sequences of All Components. 4.3 Evaluating Inter-component Coverage Once all the components in the GUI have been identified, the integration tree may be constructed by adding, for each restricted-focus event ex , the element (Cx ; Cy ) to B where Cx is the component that contains ex and Cy is the component that it invokes. The integration tree may be used in several ways to identify interactions among components. For example, in Figure 2 a subset of all possible pairs of components that interact would be f (Main, FileNew), (Main, FileOpen), (Main, Print), (Main, FormatFont), and (Print, g. To identify sequences such as the ones from Main to Properties, we traverse the integration tree in a bottom-up manner, identifying interactions among Print and Properties. We then merge Print and Properties to form a super-component called PrintProperties. We then check interactions among Main and PrintProperties. This process continues until all components have been merged into a single super-component. Evaluating the inter-component coverage of a given test suite requires computing the (1) invocation coverage, (2) invocation- termination coverage, and (3) length-n event sequence cover- age. The total number of length 1 event sequences required to satisfy the invocation coverage criterion is equal to the number of restricted-focus events available in the GUI. The percentage of restricted-focus events actually covered by the test cases is (x=I) \Theta 100, where x is the number of restricted- focus events in the test cases, and I is the total number of restricted-focus events available in the GUI. Similarly, the total number of length 2 event sequences required to satisfy the invocation-termination criterion is and T i are the number of restricted-focus and termination events that invoke and terminate component C i respectively. The percentage of invocation-termination pairs actually covered by the test cases is (x= where x is the number of invocation-termination pairs in the test cases. Computing the percentage of length-n event sequences is slightly more involved. The algorithm shown in Figure 5 computes the percentage of length-n event sequences tested among GUI components. Intuitively, the algorithm obtains the number of event sequences that end at a certain restricted- focus event. It then counts the number of event sequences that can be extended from these sequences into the invoked component. The main algorithm called Integrate is recursive and performs a bottom-up traversal of the integration tree T (line 2). Other than the recursive call (line 8), Integrate makes a call to ComputeTotalInteractions that takes two components as parameters (lines 13,14). It initializes the vector Total for all path lengths i (1 i M) (line 16,17). We assume that a freq matrix has been stored for each component. The freq matrix is similar to the freq vector already computed in the algorithm in Figure 4. freq i;j is the number of event-sequences that start with event i and end with event j. After obtaining both frequency matrices for both C1 and C2 , for all path lengths (lines 21,26), the new vector Total is obtained by adding the frequency entries from F1 and F2 (lines 28.30). A new frequency matrix is computed for the super-component This new frequency matrix will be utilized by the same algorithm to integrate "C1C2 " to other components. The results of the above algorithm are summarized in Ma- trix. Matrix i;j is the percentage of length-j event-sequences that have been tested in the super-component represented by the label i. 5. CASE STUDY We performed a case study on our version of WordPad to determine the (1) total number of event sequences required T: Integration Tree) 2 ComputeTotalInteractions(newT, c); 9 MatrixnewT+c Component Component Total x /* get freq table of C1 for event-seq of length i */ 20 /* Add all values in column x */ 22 /* get freq table of C2 for event-seq of length j */ 25 28 Total i+j /\Gamma Total i+j ComputeFreqMatrix(C1 , C2 ); 31 Figure 5: Computing Percentage of Tested Length-n Event-sequences of All Components. to test the GUI and hence enable a test designer to compute the percentage of event sequences tested, (2) correlation between event-based coverage of the GUI and statement coverage of the underlying code, and (3) time taken to evaluate the coverage of a given test suite and usefulness of the coverage report to guide further testing. In the case study, we employed our own specifications and implementation of the WordPad software. The software consists of 36 modal windows, and 362 events (not counting short-cuts). Our implementation of WordPad is similar to Microsoft's WordPad except for the Help menu, which we did not model. 5.1 ComputingTotalNumber of Event-sequences for WordPad In this case study, we wanted to determine the total number of event sequences that our new criteria specify to test parts of WordPad. We performed the following steps: Components and Events: Individual Word- Pad components and events within each component were identified. Table 1 shows some of the components of WordPad that we used in our case study. Each row represents a component and each column shows the Component Name Open System Interaction Restricted Focus Unrestricted Focus Termination Sum Main 7 FileOpen FileSave Properties Sum 7 78 Event Type Table 1: Types of Events in Some Components of MS WordPad. different types of events available within each component Creating Event-flow Graphs: The next step was to construct an event-flow graph for each component. In Figure 1 we showed a part of the event-flow graph of the most important component, i.e., Main. Recall that each node in the event-flow graph represents an event. Computing Event-sequences: Once the event-flow graphs were available, we computed the total number of possible event-sequences of different lengths in each component by using the computeTotals subroutine in Figure 4. Note that these event-sequences may also include infeasible event-sequences. The total number of event-sequences is shown in Table 2. The rows represent the components and the shaded rows represent the inter-component interactions. The columns represent different event-sequence lengths. Recall that an event-sequence of length 1 represents event coverage whereas an event-sequence of length 2 represents event-selection coverage. The columns 1' and 2' represent invocation and invocation-termination coverage respectively. The results of this case study show that the total number of event sequences grows with increasing length. Note that longer sequences subsume shorter sequences; e.g., if all event sequences of length 5 are tested, then so are all sequences of length-i, where 4. It is difficult to determine the maximum length of event sequences needed to test a GUI. The large number of event sequences show that it is impractical to test a GUI for all possible event sequences. Rather, depending on the resources, a subset of "important" event sequences should be identified, generated and executed. Identifying such important sequences requires that they be ordered by assigning a priority to each event sequence. For example, event sequences that are performed in the Main component may be given higher priority since they will be used more frequently; all the users start interacting with the GUI using the Main component. The components that are deepest in the integration tree may be used the least. This observation leads to a heuristic for ordering the testing of event sequences within components of the GUI. The structure of the integration tree may be used to assign priorities to components; Main will have the highest priority, decreasing for components at the second level, with the deepest components having the lowest priority. A large number Component Name 1' Main 56 791 14354 255720 4490626 78385288 FileOpen FileSave Print 12 108 972 8748 78732 708588 Properties 13 143 1573 17303 190333 2093663 PageSetup 11 88 704 5632 45056 360448 FormatFont 9 63 441 3087 21609 151263 Print+Properties Main+FileSave Main+FormatFont Main+Print+Properties 12 145 1930 28987 466578 Event-sequence Length Table 2: Total Number of Event-sequences for Selected Components of WordPad. Shaded Rows Show Number of Interactions Among Components. of event sequences in the high priority components may be tested first; the number will decrease for low priority components 5.2 Correlation Between Event-based Coverage and Statement Coverage In this case study, we wanted to determine exactly which percentage of the underlying code is executed when event- sequences of increasing length are executed on the GUI. We wanted to see whether testing longer sequences adds to the coverage of the underlying code. We performed the following steps: Code Instrumentation: We instrumented the underlying code of WordPad to produce a statement trace, i.e., a sequence of statements in the order in which they are executed. Examining such a trace allowed us to determine which statements are executed by a test case. Event-sequence Generation: We wanted to generate all event-sequences up to a specific length. We modified ComputeTotals in Figure 4 resulting in an event- sequence generation algorithm that constructs event sequences of increasing length. The dynamic programming algorithm constructs all event sequences of length 1. It then uses follow set to extend each event sequence by one event, hence creating all length 2 event- sequences. We generated all event-sequences up to length 3. In all we obtained 21659 event-sequences. Controlling GUI's State: Bringing a software to a state S i in which a test case T i may be executed on it is traditionally known as the controllability problem [1]. This problem also occurs in GUIs and for each test case, appropriate events may need to be performed on the GUI to bring it to the state S i . We call this sequence of events the prefix, P i , of the test case. Although generating the prefix in general may require the development of expensive solutions, we used a heuristic for this study. We executed each test case in a fixed state S0 in which WordPad contains text, part of the text was highlighted, the clipboard contains a text object, and the file system contains two text files. We traversed the event-flow graphs and the integration tree to produce the prefix of each test case. We do, however, note that using this heuristic may render Event-sequence Length Percentage of Statements Executed Figure The Correlation Between Event-based Coverage and Statement Coverage of WordPad. some of the event sequences non-executable because of infeasibility. We will later see that such sequences do exist but are of no consequence to the results of this study. We have modified WordPad so that no statement trace is produced for P i . Test-case Execution: After all event-sequences up to length 3 were obtained, we executed them on the GUI using our automated test executor [10] and obtained all the execution traces. The test case executor executed without any intervention for hours. We note that (or 19.3%) of the test cases could not be executed because of infeasibility. Analysis: In analyzing the traces for our study, we determined the new statements executed by event-sequences of length 1, i.e., individual events. The graph in Figure 6 shows that almost 92% of the statements were executed by just these individual events. As the length of the event sequences increases, very few new statements are executed (5%). Hence, a high statement coverage of the underlying code may be obtained by executing short event sequences. The results of this case study can be explained in terms of the design of the WordPad GUI. Since the GUI is an event-driven software, a method called an event handler is implemented for each event. Executing an event caused the execution of its corresponding event handler. Code inspection of the WordPad implementation revealed that there were few or no branch statements in the code of the event han- dler. Consequently, when an event was performed, most of the statements in the event-handler were executed. Hence high statement coverage was obtained by just performing individual events. Whether other GUIs exhibit similar behavior requires a detailed analysis of a number of GUIs and their underlying code. The result shows that statement coverage of the underlying code can be a misleading coverage criterion for GUI test- ing. A test designer who relies on statement coverage of the underlying code for GUI testing may test only short event sequences. However, testing only short sequences is not enough. Longer event sequences lead to different states of the GUI and that testing these sequences may help detect a larger number of faults than short event sequences. Component Name 1' Main FileOpen 9 FileSave 9 33 132 Print 11 37 313 787 3085 1314 Properties 12 Print+Properties Main+FileSave Main+Print+Properties 6 56 123 189 423 Event-sequence Length Table 3: The Number of Event-sequences for Selected Components of WordPad Covered by the Test Cases. For example, in WordPad, the event Find Next (obtained by clicking on the Edit menu) can only be executed after at least 6 events have been performed; the shortest sequence of events needed to execute Find Next is !Edit, Find, TypeInText, FindNext2, OK, Edit, Find Next?, which has 7 events. If only short sequences (! are executed on the GUI, a bug in Find Next may not be detected. Extensive studies of the fault-detection capabilities of executing short and long event sequences for GUI testing are needed, and is targeted for future work. Another possible extension to this experiment is to determine the correlation between event-based coverage and other code-based coverage, e.g., branch coverage. 5.3 Evaluating the Coverage of a Test Suite We wanted to determine the time taken to evaluate the coverage of a given test suite and how the resulting coverage report could guide further testing. We used our earlier developed planning-based test case generation system called Planning Assisted Tester for grapHical user interface Systems(PATHS) to generate test cases [8]. We performed the following steps: Tasks: In PATHS, commonly used tasks were identified. A task is an activity to be performed by using the events in the GUI. In PATHS, the test designer inputs tasks as pairs (I, G), where I is the initial GUI state before the task is performed and G is the final GUI state after the task has been performed. We carefully identified 72 different tasks, making sure that each task exercised at least one unique feature of WordPad. For example, in one task we modified the font of text, in another we printed the document on size paper. Generating Test Cases: Multiple test cases were generated using a plan generation system to achieve these tasks. In this manner, we generated 500 test cases (multiple cases for each task). Coverage Evaluation: After the test cases were available, we executed the algorithms of Figures 4 and 5. The algorithms were implemented using Perl and Mathematica [19] and were executed on a Sun UltraSPARC workstation (Sparc Ultra running SunOS 5.5.1. Even Component Name 1' Main FileOpen 90 56 17.50 0.72 0.06 FileSave 90 41 20.63 1.27 0.47 0.02 Print Properties PageSetup 91 FormatFont Print+Properties 100 0 46 51.15 8.18 3.87 Main+FileOpen Main+FileSave 100 0 20 13.00 8.64 1.26 0.28 Main+FormatFont 100 0 33 28.40 5.17 0.97 0.10 Main+Print+Properties 50 38.62 6.37 0.65 0.09 Event-sequence Length Table 4: The Percentage of Total Event-sequences for Selected Components of WordPad Covered by the Test Cases. with the inefficiencies inherent in the Perl and Mathematica implementation, we could process the 500 test cases in 47 minutes (clock time). The results of applying the algorithms are summarized as coverage reports in Tables and 4. Table 3 shows the actual number of event-sequences that the test cases covered. Table 4 presents the same data, but as a percentage of the total number of event-sequences. Column 1 in Table 4 shows close to 90% event coverage. The remaining 10% of the events (such as Cancel) were never used by the planner since they did not contribute to a goal. Column 2 shows the event-selection coverage and the test cases achieved 40-55% coverage. Note that since all the components were invoked at least once, 100% invocation coverage (column 1') was obtained. However, none of the components were terminated immediately after being invoked. Hence, no invocation-termination coverage (column 2') was obtained. This result shows that the coverage of a large test suite can be obtained in a reasonable amount of time. Looking at columns 4, 5, and 6 of Table 4, we note that only a small percentage of length 4, 5, and 6 event sequences were tested. The test designer can evaluate the importance of testing these longer sequences and perform additional testing. Also, the two-dimensional structure of Table 4 helps target specific components and component-interactions. For example, 60% of of length 2 interactions among Main and PageSetup have been tested whereas only 11% of the interactions among Main and FileOpen have been tested. Depending on the relative importance of these components and their interactions, the test designer can focus on testing these specific parts of the GUI. 6. RELATED WORK Very little research has been reported on developing coverage criteria for GUI. The only exception is the work by Ostrand et al. who briefly indicate that a model-based method may be useful for improving the coverage of a test suite [12]. However, they have deferred a detailed study of the coverage of the generated test cases using this type of GUI model to future work. There is a close relationship between test-case generation techniques and the underlying coverage criteria used. Much of the literature on GUI test case generation focuses on describing the algorithms used to generate the test cases [14, 6]. Little or no discussion about the underlying coverage criteria is presented. In the next few paragraphs, we present a discussion of some of the methods used to develop test cases for GUIs and their underlying coverage criteria. We also present a discussion of automated test case generation techniques that offer a unique perspective of GUI coverage. The most commonly available tools to aid the test designer in the GUI testing process include record/playback tools [15, 3]. These tools record the user events and GUI screens during an interactive session. The recorded sessions are later played back whenever it is necessary to generate the same GUI events. Record/playback tools provide no functionality to evaluate the coverage of a test suite. The primary reason for no coverage support is that these tools lack a global view of the GUI. The test cases are constructed individually with a local perspective. Several attempts have been made to provide more sophisticated tools for GUI testing. One popular technique is programming the test case generator [7]. The test designer develops programs to generate test cases for a GUI. The use of loops, conditionals, and data selection switches in the test case generation program gives the test designer a broader view of the generated test cases' coverage. Several finite-state machine (FSM) models have also been proposed to generate test cases [14]. Once an FSM is built, coverage of a test suite is evaluated by the number of states visited by the test case. This method of evaluating coverage of a test suite needs to be studied further as an accurate representation of the GUI's navigation results in an infinite number of states. White et al. presents a new test case generation technique for GUIs [18]. The test designer/expert manually identifies a responsibility, i.e., a GUI activity. For each responsibility, a machine model called the complete interaction sequence (CIS) is identified manually. To reduce the size of the test suite, the CIS is reduced using constructs/patterns in the CIS. The example presented therein showed that testing could be performed by using 8 test cases instead of 48. How- ever, there is no discussion of why no loss of coverage will occur during this reduction. Moreover, further loss of coverage may occur in identifying responsibilities and creating the CIS. The merit of the technique will perhaps be clearer when interactions between the CIS are investigated. 7. CONCLUSION In this paper, we presented new coverage criteria for GUI testing based on GUI events and their interactions. A unit of testing called a GUI component was defined. We identified the events within each component and represented them as an event-flow graph. Three new coverage criteria were de- fined: event, event-selection, and length-n event-sequence coverage. We defined an integration tree to identify events among components and defined three inter-component coverage criteria: invocation, invocation-termination and inter-component length-n event-sequence coverage. In the future we plan to examine the effects of the GUI's structure on its testability. As GUIs become more struc- tured, the integration tree becomes more complex and inter-component testing becomes more important. We also plan to explore the possibility of using the event-based coverage criteria for software other than GUIs. We foresee the use of these criteria for (1) object-oriented soft- ware, which use messages/events for communication among objects, (2) networking software, which use messages for communication, and (3) the broader class of reactive soft- ware, which responds to events. 8. --R A framework for testing database applications. A mathematical framework for the investigation of testing. Integrated data capture and analysis tools for research and testing an graphical user interfaces. Interprocedual data flow testing. Toward automatic generation of novice user test scripts. The black art of GUI testing. Using a goal-driven approach to generate test cases for GUIs Automated test oracles for GUIs. A planning-based approach to GUI testing Hierarchical GUI test case generation using automated planning. A visual test development environment for GUI systems. Selecting software test data using data flow information. A method to automate user interface testing using variable finite state machines. The applicability of program schema results to programs. Translatability and decidability questions for restricted classes of program schemas. Generating test cases for GUI responsibilities using complete interaction sequences. A System for Doing Mathematics by Computer. Test data adequacy measurements. --TR Selecting software test data using data flow information Mathematica: a system for doing mathematics by computer Integrated data capture and analysis tools for research and testing on graphical user interfaces Test data adequacy measurement Object-oriented integration testing Toward automatic generation of novice user test scripts A visual test development environment for GUI systems Using a goal-driven approach to generate test cases for GUIs A framework for testing database applications Automated test oracles for GUIs Hierarchical GUI Test Case Generation Using Automated Planning A Method to Automate User Interface Testing Using Variable Finite State Machines Generating Test Cases for GUI Responsibilities Using Complete Interaction Sequences --CTR Yanhong Sun , Edward L. Jones, Specification-driven automated testing of GUI-based Java programs, Proceedings of the 42nd annual Southeast regional conference, April 02-03, 2004, Huntsville, Alabama Aine Mitchell , James F. Power, An approach to quantifying the run-time behaviour of Java GUI applications, Proceedings of the winter international synposium on Information and communication technologies, January 05-08, 2004, Cancun, Mexico Philippe Palanque , Regina Bernhaupt , Ronald Boring , Chris Johnson, Testing Interactive Software: A Challenge for Usability and Reliability, CHI '06 extended abstracts on Human factors in computing systems, April 22-27, 2006, Montral, Qubec, Canada Ping Li , Toan Huynh , Marek Reformat , James Miller, A practical approach to testing GUI systems, Empirical Software Engineering, v.12 n.4, p.331-357, August 2007 Christopher J. Howell , Gregory M. Kapfhammer , Robert S. Roos, An examination of the run-time performance of GUI creation frameworks, Proceedings of the 2nd international conference on Principles and practice of programming in Java, June 16-18, 2003, Kilkenny City, Ireland Geoffrey R. Gray , Colin A. Higgins, An introspective approach to marking graphical user interfaces, ACM SIGCSE Bulletin, v.38 n.3, September 2006 Atif M. Memon , Mary Lou Soffa, Regression testing of GUIs, ACM SIGSOFT Software Engineering Notes, v.28 n.5, September Mikael Lindvall , Ioana Rus , Paolo Donzelli , Atif Memon , Marvin Zelkowitz , Aysu Betin-Can , Tevfik Bultan , Chris Ackermann , Bettina Anders , Sima Asgari , Victor Basili , Lorin Hochstein , Jrg Fellmann , Forrest Shull , Roseanne Tvedt , Daniel Pech , Daniel Hirschbach, Experimenting with software testbeds for evaluating new technologies, Empirical Software Engineering, v.12 n.4, p.417-444, August 2007 Qing Xie , Atif M. Memon, Designing and comparing automated test oracles for GUI-based software applications, ACM Transactions on Software Engineering and Methodology (TOSEM), v.16 n.1, p.4-es, February 2007 Atif Memon , Adithya Nagarajan , Qing Xie, Automating regression testing for evolving GUI software: Research Articles, Journal of Software Maintenance and Evolution: Research and Practice, v.17 n.1, p.27-64, January 2005 Atif Memon , Adithya Nagarajan , Qing Xie, Automating regression testing for evolving GUI software, Journal of Software Maintenance: Research and Practice, v.17 n.1, p.27-64, January 2005

component testing;integration tree;event-flow graph;event-based coverage;GUI testing;GUI test coverage

503246

An empirical study on the utility of formal routines to transfer knowledge and experience.

Most quality and software process improvement frameworks emphasize written (i.e. formal) documentation to convey recommended work practices. However, there is considerable skepticism among developers to learn from and adhere to prescribed process models. The latter are often perceived as overly "structured" or implying too much "control". Further, what is relevant knowledge has often been decided by "others"---often the quality manager. The study was carried out in the context of a national software process improvement program in Norway for small- and medium-sized companies to assess the attitude to formalized knowledge and experience sources. The results show that developers are rather skeptical at using written routines, while quality and technical managers are taking this for granted. This is an explosive combination. The conclusion is that formal routines must be supplemented by collaborative, social processes to promote effective dissemination and organizational learning. Trying to force a (well-intended) quality system down the developers' throats is both futile and demoralizing. The wider implications for quality and improvement work is that we must strike a balance between the "disciplined" or "rational" and the "creative" way of working.

Figure 1. A model of knowledge conversion between tacit and explicit knowledge [27]. Figure 1 expresses that practitioners first internalize new knowledge (i.e. individual learning). The new knowledge is then socialized into revised work processes and changed behavior (group learning). The new work processes and the changed behavior are then observed and abstracted, i.e. externalized. This new knowledge is then combined to refine and extend the existing knowledge (organizational learning). This process continues in new cycles etc. To enable learning is the crucial issue, both at the individual, group, and organizational level. The latter means creating and sustaining a learning organization that constantly improves its work, by letting employees share experience with each other. Around the underlying experience bases, there may be special (sub-)organizations to manage and disseminate the stored experience and knowledge, as exemplified by the Experience Factory [7]. We also refer to the workshop series of Learning Software Organizations [4] [9]. Other fields have introduced the term organizational or corporate memory to characterize an organization's strategic assets, although not only from a learning point of view [1]. The knowledge engineering community has also worked on experience bases, often with emphasis on effective knowledge representations, deduction techniques etc., and towards a wide range of applications. The subfield of Case-Based Reasoning [3] has sprung up from this work, enabling reuse of similar, past information (cases) to better master new situations. We will also mention the subfield of Data Mining [20]. Social anthropologists and psychologists have studied how organizations learn, and how their employees make use of information sources in their daily work. Much R&D effort has been spent on the externalizing" flow, looking for valid experience that can be analyzed, generalized, synthesized, packaged and disseminated in the form of improved models or concepts. For instance, to make, calibrate and improve an estimation model based on the performance of previous software projects. Explicit knowledge may nevertheless be misunderstood due to lack of context and nuances, e.g. how to understand the context of post-mortems? However, the hard part is the internalizing" flow. That is, how to make an impact on current practice, even if updated knowledge may be convincingly available? See for instance the ethnographic study on the use of quality systems in [40]. Typical inhibitors are not-invented-here", mistrust (been-burned-before"), lack of extra time/resources (not-getting started), or plain unwillingness to try something new or different (like adhering to formal procedures in a quality system). A study of maintenance technicians for copy machines indicated that such experts were most likely to ask their colleagues for advice, rather than to look it up in or even to follow the book [10]. Indeed, how many times have not computer scientists asked their office mates about commands in Word or NT-Windows, instead of directly consulting relevant documentation - although a query into the latter can be hard to formulate. Furthermore, the existence of software quality manuals, either on paper in thick binders (sometimes 1-2 m in the shelves) or in web documents on an Intranet, is no guarantee for their use. In fact, since manuals may dictate people on how to perform their job, traditional quality departments in many software organizations are not looked upon with high esteem by developers. For instance, there are over 250 proposed software standards [32], many of them recommending standard process models, but how many of these are in practical use? So, if we are to succeed with formal routines and explicit knowledge in a quality system or a SEB to achieve learning, we must not carry the traditional QA hat of control. This does not mean that all, formal knowledge in the form of books, reports etc. (like this article) has to be discarded. The lesson is just that formal routines must be formulated and introduced with proper participation from the people involved, in order to have the intended effect on practice. Lastly, many of the ideas and techniques on quality improvement (TQM and similar) come from manufacturing, with rather stable products, processes and organizations. Information technology, on the other hand, is characterized by rapid product innovation, not gradual process refinement [33]. One IT year is like a dog year (7 years) in other disciplines, and time-to-market seems sacred (i.e. schedule pressure). The strength of many software SMEs (Small and Medium-sized Enterprises) lies in their very ability to turn around fast and to convert next week's technologies into radically new products and services. Barrett [6] has used the term improvisation, a jazz metaphor, to characterize performers that execute evolving activities while employing a large competence base. With reference to our software development context, we must carefully adopt a set of quality and improvement technologies that can function in a very dynamic environment - so how to manage constant change [17]? Since SPI assumes that there is something stable that can be improved, we must pick our learning focus accordingly. For instance, the Norwegian Computer Society (www.dnd.no) is now offering a course in chaos and complexity theory as an alternative to manage highly evolving projects. However, it is fair to say that especially TQM is aware of the cultural and social dimensions of quality work. TQM has a strong emphasis on creating a learning organization, and having all employees participate and involve themselves in order to satisfy their customers. So, how can software organizations best systematize, organize and exploit previous experience in order to improve their work? 3. CONTEXT, QUESTIONS, AND 3.1 The SPIQ Project The SPIQ project [12] was run for three years in 1997-99, after a half-year pre-study in 1996. SPIQ stands for SPI for better Quality. The project, which was funded in part by the Research Council of norway, involved three research institutions and 12 IT companies, mostly SMEs. More than 20 SPI pilot projects were run in these companies. A follow-up project called PROFIT is now carried out in 2000-2002. The main result of the SPIQ project was a pragmatic method handbook [16], with the following components: A dual, top-down/bottom-up approach, using TQM [15] and Quality Improvement Paradigm [7] ideas. An adapted process for ESSI-type [19] Process Improvement Experiments (PIEs). The Goal-Question-Metric (GQM) method [8], and e.g. GQM feedback sessions. The Experience Factory concept [7], to refine and disseminate project experiences. An incremental approach, relying on action research [22]. Reported empirical studies from five SPIQ companies. Typical work in the 12 SPIQ companies included pilot projects to test out a certain improvement technology, like novel inspection techniques, incremental development, or use of measurement and software experience bases. For further results from comparative studies of SPI success in the SPIQ companies and in several other Scandinavian PIEs, all emphasizing organizational and cultural factors see e.g. [13], [14], [18], and [37]. 3.2 How the Study Was Performed organization: The actual study was carried out between NTNU/SINTEF and five companies participating in the SPIQ project. Data collection was carried out by two NTNU students in the last year of their M.Sc. study, as part of a pre-thesis project [11]. The two students were advised by the two authors of this paper, the former being their teacher and also a SPIQ researcher, the latter being a researcher and Ph.D. student attached to the project. Preparation: First, the student read and learnt about the project and relevant literature. Then we tried an initial formulation of some important issues in the form of research questions, and discussed these. At the same time, we contacted potential companies and checked their willingness to participate and their availability. Then a more detailed interview guide (see 3.3) was designed, in dialogue between the students and their advisors. The companies had been briefed about the questions, and when and how the interviews were going to be run. Research questions - important issues to address are: Q1: What is the knowledge of the routines being used? Q2: How are these routines being used? Q3: How are they updated? Q4: How effective are they as a medium for transfer of knowledge and experience? And, furthermore, are there important differences between developers and managers, and how much cooperation is involved in making and updating the routines? Subjects: Initially, we tried to have a dozen companies involved, but the time frame of the students' availability (three months in spring of 1999) only allowed five companies. One of these was in Trondheim, the rest in Oslo. Three of the companies were ISO- 9001 certified. Two of the companies were IT/telecom companies, the rest were software houses. A convenience sample (i.e. available volunteers) of 23 persons were interviewed based on their experience with SPI, whereof 13 developers and 10 managers. The latter group included one quality manager and one software manager (e.g. division or project manager) from each of the five companies. Data collection: After finishing the interview guide, this was sent by email to the respondents. A few days later, the two students visited the companies, and spent a full day at each company. At each place they spent up to one hour with each respondent in semi-structured interviews. In each interview, the four questions were treated one after another. One of the students was asking the questions, and both students made notes (interview records) during the interview. The other student served as a scribe, and wrote down a structured summary immediately after the interview. The first student then checked this against his own notes. Data analysis: The ensuing categorization and data analysis was done by the two students, in cooperation with the authors, and reported in the students' pre-diploma thesis. 3.3 The Interview Guide As mentioned, we formulated four main research questions, with a total of 14 sub-questions: Q1. Knowledge of the routines. 1.1 Describe the (possible) contents in routines being used for software development. 1.2 How were these routines introduced in the company? 1.3 What was the purpose of these routines? 1.4 What is you personal impression of the routines? Q2. Use of routines. 2.1 What status does a given routine have among developers? 2.2 To what degree are the routines actually being used? 2.3 Who are the most active/passive users? 2.4 What is the availability of the routines? 2.5 How is follow-up and control of usage done? Q3. Updating of routines. 3.1 What procedures are used to update the routines? 3.2 Who participates in the update activities? Q4. Routines as a medium for transfer of knowledge and experience. 4.1 Do you regard written routines as an efficient medium for transfer of knowledge and experience? 4.2 What alternatives to written routines do you think are useful or in use in the company? 4.3 What barriers against transfer of experiences do you think are most important? The interview guide contained advice on how to deal with structured questions, usually with three answer categories such as yes-maybe-no or little-some-much. We allowed more unstructured commentaries in the form of prose answers to solicit more direct and commentary opinions. 4. RESULTS In this section, we present the results of our study regarding the utility of formal routines as a medium for transfer of knowledge and experience. The focus is on participation and potential differences in opinion between developers and managers regarding the utility of the routines. 4.1 Knowledge of the Routines All respondents had a fairly good knowledge of the routines that were in place in their respective companies. In fact, two thirds of the respondents showed good knowledge about the content of the routines. Table 1 illustrates this, and shows how well developers and managers were able to describe the specific contents of the routines in their company. Table 1. Knowledge of company routines. Software developers Managers Frequency Percent Frequency Percent Little - Some 646220 Much 754880 However, when it came to knowledge about how the routines were introduced, 50% of the developers did not know anything about this process. On the other hand, only one manager did not know about the introduction process. All in all, it turned out that about 30% of the developers and 70% of the managers had actively participated in the introduction of routines (Table 2). Table 2. Degree of involvement during introduction of routines. Degree of involvement Low High Freq. Percent Freq. Percent Developers 969431 Managers 330770 Furthermore, it seemed to be a common understanding regarding the objective of having formal routines. Most respondents said that such routines were useful with respect to quality assurance. Other respondents said that they would enable a more unified way of working. However they emphasized that: Routines should not be formalistic, but rather useful and necessary. Respondents in the three ISO-9001 certified companies claimed that their routines were first and foremost established to get the certificate on the wall, and that the quality of their software processes had gained little or nothing from the ISO certification. One of the respondents expressed his views on this by the following example: You might be ISO certified to produce lifebelts in concrete, as long as you put the exact same amount of concrete in each lifebelt. Although some of the respondents were critical to the routines, stating that: 10% of the routines are useful, while the remaining 90% is nonsense Most respondents, nevertheless, had a good impression of the routines, typically stating that: Routines are a prerequisite for internal collaboration. Routines are a reassurance and of great help. 4.2 Use of Routines Software developers and managers agreed on the degree to which the routines were used. In general, they answered that about 50% of the routines were in use, and that the more experienced developers used the routines to a lesser extent than the more inexperienced developers do. Furthermore, it was a common agreement that: There is no point in having routines that are not considered useful. However, the status of the routines among the software developers was highly divergent, as seen from the following statements: The routines are generally good and useful, but some developers are frustrated regarding their use. The system is bureaucratic - it was better before, when we had more freedom to decide for ourselves what should best be done. The routines are easy to use. Routines are uninteresting and revision meetings are boring. 4.3 Updating of Routines None of the companies had scheduled revisions as part of the process for updating their routines. Most answers to this issue were rather vague. Some respondents explained that such revisions were informally triggered, while other respondents did not know how to propose and implement changes to existing routines. However, respondents from all of the companies, both managers and software developers, said that all employees in their respective companies could participate in the revision activities if they wanted to. 4.4 Routines as a Medium for Transfer of Knowledge and Experience The answers to this issue varied a lot, and indicated highly different attitudes regarding the effectiveness of formal routines for knowledge and experience transfer. Particularly, it seemed to be a clear difference in judgment between software developers and managers. While seven of the ten managers regarded written routines as an efficient medium for knowledge transfer, none of the developers did! Furthermore, half of the developers considered such routines to be inefficient for knowledge transfer, while only one of the managers shared this view. Typically, managers said that written routines were important as means for replacing the knowledge of the people that had left the company. Software developers, on the other hand, did not make such a clear connection between experience, knowledge transfer and formal routines. One software developer said that different groups within the company never read each other's reports, while another developer maintained that it would take too much time to learn about the experience of the other groups. Several of the developers explained their views by stating that the documentation was not good enough, it was hard to find, boring to use and that it takes too much time. When asked about useful alternatives to written routines, the respondents answered that they regarded some kind of Experience base or Newsgroup as the highest ranked alternative. Other high-ranked alternatives were Socialization, Discussion groups, Experience reports, Group meetings, and On-the-job training. Table 3 shows these alternatives in rank order (1 is best) for software developers and managers respectively. We also asked the respondents about what they regarded as the most important barriers against transfer of knowledge and experience. Nearly all of them said that such transfer, first and foremost, is a personal matter depending on how much each individual whishes to teach their lessons-learned to others. Furthermore, the willingness to share depends on available time, personality, self-interest, and company culture. Table 3. Alternative media for knowledge transfer. Rank Medium Developers Managers Experience base/newsgroups 1 1 Socialization Discussion groups 3 2 Experience reports 4 3 On-the-job-training 5 6 Work with ext. consultants 6 - Group meetings 7 5 As shown in Table 4, seven (six developers and one manager) of the 23 respondents answered a definite No to the question regarding the efficiency of written routines as a medium for transfer of knowledge and experience. Likewise, seven respondents (managers only) answered an equally clear Yes to the same question. The last nine respondents answered somewhere in between, and said that in some cases written routines could be effective, while in other situations they would rather be a barrier. Table 4. Do you regard written routines as an efficient medium for transfer of knowledge and experience? Software developers Managers Frequency Percent Frequency Percent Both 754220 Due to the rather small sample size in this study, and the low expected frequency in several of the cells in Table 4, we compared the respondents' assessments of the routines and their job function using Fisher's exact probability test. With this test, the exact probability (or significance level) that the obtained result is purely a product of chance is calculated [23]. The test statistic of 13.02 was highly significant (p=0.002, two-tailed). Thus, we rejected the hypothesis of independence and concluded that there is a difference in the distribution of assessment of the usefulness of formal routines as an efficient medium for transfer of knowledge and experience between software developers and managers. Since software developers had been involved in the process of introducing the routines to a much lesser extent than the managers, we compared the respondent's assessment of the routines with the level of involvement using Fisher's exact test Table 5). The test statistic of 14.71 was highly significant (p<0.0005, two-tailed). Thus, we concluded that there is a difference in the assessment of the usefulness of formal routines as an efficient medium for transfer of knowledge and experience with respect to the degree of involvement in the introduction process. Table 5. Degree of involvement vs. assessment of formal routinized solutions that often ignore the social realities of the routines as an efficient medium for transfer of knowledge and workplace, see work by Kunda [24] and Thomas [38]. experience. Against this background, we can more easily understand the Degree of involvement preference of formal routines within the SPI community as espoused by quality managers or members of Software Efficient medium? Low High Engineering Process Groups (SEPGs). Likewise, managers will -7rather put emphasis on rules, procedures, and instructions than on dialog, discussion and employee participation. Both 54 software development is radically different from manufacturing. The former is not a mechanical process with strong causal models, where we just need to establish the right formal routines. Rather, the developers view software 5. DISCUSSION development largely as an intellectual and social activity. In this section, we restrict the discussion to possible explanations Therefore, we cannot apply a rationalistic, linear model to of why none of the software developers in our study regarded software engineering. We should admit that reality for most formal routines as an efficient medium for transfer of knowledge software organizations is a non-deterministic, multi-directional and experience. The reason for this is that we regard formalization flux that involves constant negotiation and renegotiation among and participation as important issues for the relevance of much of and between the social groups shaping the software [17]. the SPI work done today by both researchers and practitioners. This does not mean that we should discard discipline and The respondents in the study were software engineers and formalization altogether. What is needed, is to balance the odd managers with an engineering background. Furthermore, software couple of discipline and creativity in software development [21]. and quality managers with an engineering background wrote most This balance can be challenging, since losing sight of the creative, of the routines. Thus, the routines were for a large part written by design-intense nature of software work leads to stifling rigidity, engineers - for engineers. Still, there was a highly significant while losing sight of the need for discipline leads to chaos. difference in attitudes regarding the usefulness of the routines for This leads us to the second possible reason for the divergent transferring knowledge and experience between software attitudes between developer and managers; that of employee engineers and managers. participation around formal routines. As seen from our point of view, there are three main reasons for the observed diversity regarding the assessment of the efficiency 5.2 Participation of routines. One is the potential conflict between the occupational Employee participation, and the way people are treated, has been cultures of software developers and managers. The second reason noted as a crucial factor in organizational management and has to do with the degree of developer participation in developing development ever since the famous productivity studies at and introducing the routines. The third explanation has to do with Western Electric's Hawthorne plant in the 1920s. The results of the views of working and learning and thus, the ability of written these studies started a revolution in management thinking, routines in general to transfer human knowledge and experience. showing that even routine jobs can be improved if the workers are These reasons are discussed in 5.1-5.3 below. treated with respect. 5.1 Occupational Culture Interestingly our study shows, that not only did managers participate significantly more during the introduction of routines, There was a general agreement among all respondents that the but also during the actual development of the routines. However, intention of introducing formal routines was to contribute to an no one is more expert in the realities of a software company's efficient process of developing quality software. In other words, business than the software developers themselves. They are not the intention behind the formal routines was to provide only experts on how to do the work - they are also the experts on appropriate methods and techniques, and standardize work how to improve it. Thus, the developers are a software processes needed to solve the problems at hand. organization's most important source of productivity and profits - The differences we observed in attitude to the efficiency of formal the human capital view. It is therefore important to involve all routines between software developers and managers has close the people that take part in a problem or its solution, and have resemblance to the lack of alignment among executives, decisions made by these. In this respect, all of the companies engineers, and operators described by Schein [34]. He explained violated one of the most important aspects of employee these differences from a cultural perspective, defining culture as involvement on their own work environment. They may even a set of basic tacit assumptions about how the world is and ought have violated the Norwegian work environment legislation! to be, that a group of people share and that determines their Formalization is a central feature of Weber's [39] bureaucratic perceptions, thoughts, feelings, and, to some degree, their overt ideal type. Viewed in the light of our results, it is not surprising behavior (ibid., p. 11). Schein claimed that major occupational that research on formalization often presents conflicting empirical communities do not really understand each other, and that this findings regarding its efficiency. Adler and Borys [2] explained leads to failures in organizational learning. According to Schein, this divergence by saying that prior research has focused on the engineering culture and the executive culture has a common different degrees of formalization, and has paid insufficient preference to see people as impersonal resources that generate attention to different types of formalization. They emphasize an problems rather than solutions. Furthermore, the engineers' need enabling type of formalization, where procedures provide to do real engineering will drive them toward simplicity and organizational memory as a resource to capture lessons-learned or best practice. The opposite is the coercive type of formalization, where procedures are presented without a motivating rationale and thus tend to be disobeyed, resulting in a non-compliant process. Our results regarding the developers' assessment of the routines closely resemble the coercive type of formalization. The developers are clearly not against formal routines. In fact, they expressed views in favor of such routines, especially those that captured prior project experience. Contrary to the existing routines, which they deemed coercive, they wanted routines of the enabling type of formalization. Thus, the highest ranked alternative to formal routines was some sort of experience base or newsgroup. 5.3 Working and Learning Another aspect of our results is that they support Brown and Duguid's [10] perspective on learning-in-working. That is, we should emphasize informal, as opposed to formal learning. The same authors referred to these learning modes, respectively, as non-canonical and canonical practices. They suggested that training and socialization processes are likely to be ineffective if based on canonical practice, instead of the more realistic non-canonical practice: People are typically viewed as performing their jobs according to formal job descriptions, despite the fact that daily evidence points to the contrary. They are held accountable to the map, not to road conditions. (ibid., p. 42) Thus, formal routines alone are inadequate, and might very well demand more improvisational skills among developers. This is because of the rigidities of the routines, and the fact that they do not reflect actual experience [17]. Although many routines are prescriptive and simple, they are still hard to change, and they cannot help in all the complex situations of actual practice from which they are abstracted. It is not surprising, therefore, that socialization and discussion groups were among the highest ranked alternatives to formal routines. This is also in agreement with Brown and Duguid's finding that story-telling is of utmost importance for dealing with the complexities of day-to-day practice. Furthermore these authors highlighted story telling as a means of diagnosing problems and as shared repositories of accumulated wisdom. This is similar to Zuboff's [40] emphasis on story-telling to deal with smart machines, and to Greenwood and Levin's [22] use of narratives in action research. Thus, contrary to the rigidities of formal routines, stories and the tacit social activities are seen as more flexible, adaptable and relevant by the software developers in our study. Furthermore, our results support the assertion that significant learning should not be divorced from its specific context - so-called situated learning. Therefore, any routines, generalizations or other means that strip away context should be examined with caution. Indeed, it seems that learning could be regarded as a product of a community i.e. organizational learning, rather than of the individual in it. Thus lessons-learned cannot easily be transferred from one setting to another, see Lave and Wenger [25]. 5.4 Implications Although the study is limited, the discussion above suggests several implications. First, studies of the effects of formalization, whether they are enabling or coercive, should focus on the features of the actual routines as well as their implementation. In addition, we should pay attention to the process of designing the features and the goals that govern this process. Second, we must recognize and confront the implications of the deeply embedded and tacit assumptions of the different occupational cultures. And, furthermore, learn how to establish better cross-cultural dialogues in order to enable organizational learning and SPI. Third, a major practical implication is that managers should recognize the needs of balancing discipline and creativity, in order to supplement formal routines with collaborative, social processes. Only by a deep and honest appreciation of this, can managers expect effective dissemination of knowledge and experience within their organization. Based on the findings of this study, we conclude that both software managers and developers must maintain an open dialogue regarding the utility of formal routines. Such a dialogue will open the way for empirically based learning and SPI, and thus attain the rewards of an enabling type of formalization. 5.5 Limitations and Recommendations for Future Research This study focused on the utility of formal routines to transfer knowledge and experience. Although it can provide valuable insights for introduction of formal routines in the software industry, our study is not without limitations. First, the small sample and lack of randomness in the choice of respondents may be a threat to external validity. In general, most work on SPI suffers from non-representative participation, since companies that voluntarily engage in systematic improvement activities must be assumed to be better-than-average. Second, a major threat to internal validity is that we have not assessed the reliability of our measures. Variables such as degree of involvement and efficiency of routines are measured on a subjective ordinal scale. An important issue for future studies is therefore to ensure reliability and validity of all measures used, see [18]. We may also ask if the respondents were truthful in their answers. For instance, they may have sensed we were looking for trouble, and thus giving us what we wanted - i.e. exaggerating possible problems. However, their answers to the four main questions and their added qualitative comments show a consistent picture of skepticism and lack of participation concerning formal routines. We therefore choose to generally believe their answers. Despite the mentioned limitations and lack of cross-checks, we feel that this study makes an important contribution to the understanding of formal routines and their role in organizational learning and SPI. Future studies should examine the enabling features of formal routines in much more detail. The features could be refined and operationalized and used for cross-sectional and longitudinal studies of a much larger number of companies. Furthermore, such studies should include a multiple respondent approach to cover all major occupational cultures. They should also perform supplementary, ethnographic studies on how developers really work and how their work relate to formal routines - see [31] on observational studies of developers at ATT. 6. CONCLUSION Results from the survey reported in this paper show that software developers do not perceive formal routines alone as an efficient way to transfer knowledge and experience. Furthermore, the study confirms our suspicions about large differences in perception of the utility of formal routines to transfer experiences and knowledge. That is, developers are skeptical to adopt formal routines found in traditional quality systems. They also want that such routines are introduced and updated in a more cooperative manner. These results are not revolutionary and in line with many other investigations on similar themes [2], [24], [34]. See also Parnas and Clements' article [29] on how to fake a rational design process. So in spite of a small sample, we think that the results are representative for a large class of software companies. The remedy seems to create a more cooperative and open work atmosphere, with strong developer participation in designing and promoting future quality systems. The developers also seem open to start exploiting new electronic media as a means for collaboration and linking to newer SEB technologies - see also our previous studies [13] on this. However, the major and most difficult work remains non-technical, that is, to build a learning organization. Lastly, we were not able to make precise hypotheses on our four issues beforehand, so the study has a character of a preliminary investigation. Later studies may be undertaken with more precise hypotheses and on a larger sample. 7. ACKNOWLEDGEMENTS Thanks to colleagues in the SPIQ project, to colleagues at NTNU and SINTEF, and not at least to the students Jon E. Carlsen and Marius Fornss that did the fieldwork. 8. --R Wolfgang M Carlsen and Marius Forn "Software Experience Bases: A Consolidated Evaluation and Status Report" Out of the crisis Tore Dyb Tore Dyb Tore Dyb ESSI project office Chapter on Introduction to Action Research: Social Research for Social Change Control and Commitment in a High-Tech Corporation Legitimate Peripheral Participation Marciniak, editor, Encyclopedia of Software Engineering - 2 Volume <Volume>Set</Volume> The Knowledge-Creating Company The Capability Maturity Model for Software: Guidelines for Improving the Software Process "Discipline of Market Leaders and Other Accelerators to Measurement" The Fifth Discipline: The Art and Practice of the Learning Organization What Machines Can't Do Makt og byr In the Age of the Smart Machine --TR A rational design process: How and why to fake it In the age of the smart machine: the future of work and power Encyclopedia of software engineering People, Organizations, and Process Improvement Software creativity The capability maturity model An ISO 9000 approach to building quality software From data mining to knowledge discovery Reexamining organizational memory An Instrument for Measuring the Key Factors of Success in Software Process Improvement Evaluating software engineering standards Improvisation in Small Software Organizations Software Experience Bases <I>Coda</I> --CTR Tore Dyb, Enabling Software Process Improvement: An Investigation of the Importance of Organizational Issues, Empirical Software Engineering, v.7 n.4, p.387-390, December 2002 Reidar Conradi , Alfonso Fuggetta, Improving Software Process Improvement, IEEE Software, v.19 n.4, p.92-99, July 2002 Tore Dyba, An Empirical Investigation of the Key Factors for Success in Software Process Improvement, IEEE Transactions on Software Engineering, v.31 n.5, p.410-424, May 2005

knowledge transfer;software process improvement;knowledge management;formal routines;developer attitudes

503505

Featherweight Java.

Several recent studies have introduced lightweight versions of Java: reduced languages in which complex features like threads and reflection are dropped to enable rigorous arguments about key properties such as type safety. We carry this process a step further, omitting almost all features of the full language (including interfaces and even assignment) to obtain a small calculus, Featherweight Java, for which rigorous proofs are not only possible but easy. Featherweight Java bears a similar relation to Java as the lambda-calculus does to languages such as ML and Haskell. It offers a similar computational "feel," providing classes, methods, fields, inheritance, and dynamic typecasts with a semantics closely following Java's. A proof of type safety for Featherweight Java thus illustrates many of the interesting features of a safety proof for the full language, while remaining pleasingly compact. The minimal syntax, typing rules, and operational semantics of Featherweight Java make it a handy tool for studying the consequences of extensions and variations. As an illustration of its utility in this regard, we extend Featherweight Java with generic classes in the style of GJ (Bracha, Odersky, Stoutamire, and Wadler) and give a detailed proof of type safety. The extended system formalizes for the first time some of the key features of GJ.

Introduction "Inside every large language is a small language struggling to get out." Formal modeling can offer a significant boost to the design of complex real-world artifacts such as programming languages. A formal model may be used to describe some aspect of a design precisely, to state and prove its properties, and to direct attention to issues that might otherwise be overlooked. In formulating a model, however, there is a tension between completeness and compactness: the more aspects the model addresses at the same time, the more unwieldy it becomes. Often it is sensible to choose a model that is less complete but more compact, offering maximum insight for minimum investment. This strategy may be seen in a flurry of recent papers on the formal properties of Java, which omit advanced features such as concurrency and reflection and concentrate on fragments of the full language to which well-understood theory can be applied. We propose Featherweight Java, or FJ, as a new contender for a minimal core calculus for modeling Java's type system. The design of FJ favors compactness over completeness almost obsessively, having just five forms of expression: object creation, method invocation, field access, casting, and variables. Its syntax, typing rules, and operational semantics fit comfortably on a single page. Indeed, our aim has been to omit as many features as possible - even assignment - while retaining the core features of Java typing. There is a direct correspondence between FJ and a purely functional core of Java, in the sense that every FJ program is literally an executable Java program. FJ is only a little larger than Church's lambda calculus [3] or Abadi and Cardelli's object calculus [1], and is significantly smaller than previous formal models of class-based languages like Java, including those put forth by Drossopoulou, Eisenbach, and Khurshid [?], Syme [20], Nipkow and Oheimb [17], and Flatt, Krish- namurthi, and Felleisen [14]. Being smaller, FJ lets us focus on just a few key issues. For example, we have discovered that capturing the behavior of Java's cast construct in a traditional "small-step" operational semantics is trickier than we would have expected, a point that has been overlooked or underemphasized in other models. One use of FJ is as a starting point for modeling languages that extend Java. Because FJ is so compact, we can focus attention on essential aspects of the exten- sion. Moreover, because the proof of soundness for pure FJ is very simple, a rigorous soundness proof for even a significant extension may remain manageable. The second part of the paper illustrates this utility by enriching FJ with generic classes and methods 'a la GJ [7]. Al- though, the model omits a few important aspects of GJ (such as "raw types" and type argument inference for generic method calls), it has already revealed portions of the design that were underspecified and bugs in the GJ compiler. Our main goal in designing FJ was to make a proof of type soundness ("well-typed programs don't get stuck'') as concise as possible, while still capturing the essence of the soundness argument for the full Java language. Any language feature that made the soundness proof longer without making it significantly different was a candidate for omission. As in previous studies of type soundness in Java, we don't treat advanced features such as concurrency, inner classes, and reflection. Other Java features omitted from FJ include assignment, in- terfaces, overloading, messages to super, null pointers, base types (int, bool, etc.), abstract method declara- tions, shadowing of superclass fields by subclass fields, access control (public, private, etc.), and exceptions. The features of Java that we do model include mutually recursive class definitions, object creation, field access, method invocation, method override, method recursion through this, subtyping, and casting. One key simplification in FJ is the omission of as- signment. We assume that an object's fields are initialized by its constructor and never changed afterwards. This restricts FJ to a "functional" fragment of Java, in which many common Java idioms, such as use of enumerations, cannot be represented. Nonetheless, this fragment is computationally complete (it is easy to encode the lambda calculus into it), and is large enough to include many useful programs (many of the programs in Felleisen and Friedman's Java text [12] use a purely functional style). Moreover, most of the tricky typing issues in both Java and GJ are independent of assign- ment. An important exception is that the type inference algorithm for generic method invocation in GJ has some twists imposed on it by the need to maintain soundness in the presence of assignment. This paper treats a simplified version of GJ without type inference. The remainder of this paper is organized as follows. Section 2 introduces the main ideas of Featherweight Java, presents its syntax, type rules, and reduction rules, and sketches a type soundness proof. Section 3 extends Featherweight Java to Featherweight GJ, which includes generic classes and methods. Section 4 presents an erasure map from FGJ to FJ, modeling the techniques used to compile GJ into Java. Section 5 discusses related work, and Section 6 concludes. In FJ, a program consists of a collection of class definitions plus an expression to be evaluated. (This expression corresponds to the body of the main method in Java.) Here are some typical class definitions in FJ. class Pair extends Object - Object fst; Object snd; Pair(Object fst, Object snd) - Pair setfst(Object newfst) - return new Pair(newfst, this.snd); class A extends Object - class B extends Object - For the sake of syntactic regularity, we always include the supertype (even when it is Object), we always write out the constructor (even for the trivial classes A and B), and we always write the receiver for a field access (as in this.snd) or a method invocation. Constructors always take the same stylized form: there is one parameter for each field, with the same name as the field; the super constructor is invoked on the fields of the supertype; and the remaining fields are initialized to the corresponding parameters. Here the supertype is always Object, which has no fields, so the invocations of super have no arguments. Constructors are the only place where super or = appears in an FJ program. Since FJ provides no side-effecting operations, a method body always consists of return followed by an expression, as in the body of setfst(). In the context of the above definitions, the expres- sion new Pair(new A(), new B()).setfst(new B()) evaluates to the expression new Pair(new B(), new B()). There are five forms of expression in FJ. Here, new A(), new B(), and new Pair(e1,e2) are object constructors, and e3.setfst(e4) is a method invocation. In the body of setfst, the expression this.snd is a field access, and the occurrences of newfst and this are variables. FJ differs from Java in that this is an ordinary variable rather than a special keyword. The remaining form of expression is a cast. The expression ((Pair)new Pair(new Pair(new A(), new B()), new A()).fst).snd evaluates to the expression new B(). Here, ((Pair)e7), where e7 is new Pair(.fst, is a cast. The cast is required, because e7 is a field access to fst, which is declared to contain an Object, whereas the next field access, to snd, is only valid on a Pair. At run time, it is checked whether the Object stored in the fst field is a Pair (and in this case the check succeeds). In Java, one may prefix a field or parameter declaration with the keyword final to indicate that it may not be assigned to, and all parameters accessed from an inner class must be declared final. Since FJ contains no assignment and no inner classes, it matters little whether or not final appears, so we omit it for brevity. Dropping side effects has a pleasant side effect: evaluation can be easily formalized entirely within the syntax of FJ, with no additional mechanisms for modeling the heap. Moreover, in the absence of side effects, the order in which expressions are evaluated does not affect the final outcome, so we can define the operational semantics of FJ straightforwardly using a nondeterministic small-step reduction relation, following long-standing tradition in the lambda calculus. Of course, Java's call-by-value evaluation strategy is subsumed by this more general relation, so the soundness properties we prove for reduction will hold for Java's evaluation strategy as a special case. There are three basic computation rules: one for field access, one for method invocation, and one for casts. Recall that, in the lambda calculus, the beta-reduction rule for applications assumes that the function is first simplified to a lambda abstraction. Similarly, in FJ the reduction rules assume the object operated upon is first simplified to a new expression. Thus, just as the slogan for the lambda calculus is "everything is a function," here the slogan is "everything is an object." Here is the rule for field access in action: new Pair(new A(), new B()).snd \Gamma! new B() Because of the stylized form for object constructors, we know that the constructor has one parameter for each field, in the same order that the fields are declared. Here the fields are fst and snd, and an access to the snd field selects the second parameter. Here is the rule for method invocation in action (= denotes substitution): new Pair(new A(), new B()).setfst(new B()) \Gamma! new B()=newfst; new Pair(new A(),new B())=this new Pair(newfst, this.snd) i.e., new Pair(new B(), new Pair(new A(), new B()).snd) The receiver of the invocation is the object new Pair(new A(), new B()), so we look up the setfst method in the Pair class, where we find that it has formal parameter newfst and body new Pair(newfst, this.snd). The invocation reduces to the body with the formal parameter replaced by the actual, and the special variable this replaced by the receiver object. This is similar to the beta rule of the lambda calculus, (x.e0)e1 \Gamma! [e1=x]e0. The key differences are the fact that the class of the receiver determines where to look for the body (supporting method override), and the substitution of the receiver for this (supporting "recursion through self"). Readers familiar with Abadi and Cardelli's Object Calculus will see a strong similarity to their & reduction rule [1]. In FJ, as in the lambda calculus and the pure Abadi-Cardelli calculus, if a formal parameter appears more than once in the body this may lead duplication of the actual, but since there are no side effects this causes no problems. Here is the rule for a cast in action: (Pair)new Pair(new A(), new B()) \Gamma! new Pair(new A(), new B()) Once the subject of the cast is reduced to an object, it is easy to check that the class of the constructor is a subclass of the target of the cast. If so, as is the case here, then the reduction removes the cast. If not, as in the expression (A)new B(), then no rule applies and the computation is stuck, denoting a run-time error. There are three ways in which a computation may get stuck: an attempt to access a field not declared for the class, an attempt to invoke a method not declared for the class ("message not understood"), or an attempt to cast to something other than a superclass of the class. We will prove that the first two of these never happen in well-typed programs, and the third never happens in well-typed programs that contain no downcasts (or "stupid casts"-a technicality explained below). As usual, we allow reductions to apply to any subexpression of an expression. Here is a computation for the second example expression, where the next subexpression to be reduced is underlined at each step. ((Pair)new Pair(new Pair(new A(), new B()), new A()).fst).snd \Gamma! ((Pair)new Pair(new A(),new B())).snd \Gamma! new Pair(new A(), new B()).snd \Gamma! new B() We will prove a type soundness result for FJ: if an expression e reduces to expression e 0 , and if e is well typed, then e 0 is also well typed and its type is a subtype of the type of e. With this informal introduction in mind, we may now proceed to a formal definition of FJ. 2.1 Syntax The syntax, typing rules, and computation rules for FJ are given in Figure 1, with a few auxiliary functions in Figure 2. The metavariables A, B, C, D, and E range over class names; f and g range over field names; m ranges over method names; x ranges over parameter names; d and e range over expressions; CL ranges over class decla- rations; K ranges over constructor declarations; and M ranges over method declarations. We write f as short-hand for f1,. ,fn (and similarly for C, x, e, etc.) and write M as shorthand for M1 . Mn (with no commas). We write the empty sequence as ffl and denote concatenation of sequences using a comma. The length of a sequence x is written #(x). We abbreviate operations on pairs of sequences in the obvious way, writing "C f" as shorthand for "C 1 f1,. ,Cn fn ", and similarly "C f;" as short-hand for "C 1 f1;. Cn fn;", and "this.f=f;" as short-hand for "this.f1=f1;. ;this.fn=fn;". Sequences of field declarations, parameter names, and method declarations are assumed to contain no duplicate names. A class table CT is a mapping from class names C to class declarations CL. A program is a pair (CT ; e) of a class table and an expression. To lighten the notation in what follows, we always assume a fixed class table CT . The abstract syntax of FJ class declarations, constructor declarations, method declarations, and expressions is given at the top left of Figure 1. As in Java, we assume that casts bind less tightly than other forms of expression. We assume that the set of variables includes the special variable this, but that this is never used as the name of an argument to a method. Every class has a superclass, declared with extends. This raises a question: what is the superclass of the Syntax: CL ::= class C extends C -C f; K M e ::= x new C(e) Subtyping: class C extends D -. Computation: (new C(e)).f i \Gamma! e i (R-Field) (new C(e)).m(d) \Gamma! [x 7! d; this 7! new C(e)]e0 (D)(new C(e)) \Gamma! new C(e) (R-Cast) Expression typing: (T-Field) stupid warning Method typing: class C extends D -. Class typing: class C extends D -C f; K M OK Figure 1: FJ: Main definitions Field lookup: class C extends D -C f; K M class C extends D -C f; K M class C extends D -C f; K M m is not defined in M Method body lookup: class C extends D -C f; K M class C extends D -C f; K M m is not defined in M Valid method overriding: Figure 2: FJ: Auxiliary definitions Object class? There are various ways to deal with this issue; the simplest one that we have found is to take Object as a distinguished class name whose definition does not appear in the class table. The auxiliary functions that look up fields and method declarations in the class table are equipped with special cases for Object that return the empty sequence of fields and the empty set of methods. (In full Java, the class Object does have several methods. We ignore these in FJ.) By looking at the class table, we can read off the sub-type relation between classes. We write when C is a subtype of D - i.e., subtyping is the reflexive and transitive closure of the immediate subclass relation given by the extends clauses in CT . Formally, it is defined in the middle of the left column of Figure 1. The given class table is assumed to satisfy some sanity conditions: (1) CT class C. for every dom(CT ); (3) for every class name C (except Object) appearing anywhere in CT , we have C 2 dom(CT ); and (4) there are no cycles in the subtype relation induced by CT - that is, the relation is antisymmetric. For the typing and reduction rules, we need a few auxiliary definitions, given in Figure 2. The fields of a class C, written fields(C), is a sequence C f pairing the class of a field with its name, for all the fields declared in class C and all of its superclasses. The type of the method m in class C, written mtype(m; C), is a pair, written B!B, of a sequence of argument types B and a result type B. Similarly, the body of the method m in class C, written mbody(m; C), is a pair, written (x,e), of a sequence of parameters x and an expression e. The predicate override(C0!C; m; D) judges if a method m with argument types C and a result type C0 may be defined in a subclass of D. In case of overriding, if a method with the same name is declared in the superclass then it must have the same type. 2.2 Typing The typing rules for expressions, method declarations, and class declarations are in the right column of Figure 1. An environment \Gamma is a finite mapping from variables to types, written x:C. The typing judgment for expressions has the form "in the environment \Gamma, expression e has type C." The typing rules are syntax directed, with one rule for each form of expression, save that there are three rules for casts. The typing rules for constructors and method invocations check that each actual parameter has a type that is a subtype of the corresponding formal. We abbreviate typing judgments on sequences in the obvious way, writing C1 , . , \Gamma ' en 2 Cn and writing for One technical innovation in FJ is the introduction of "stupid" casts. There are three rules for type casts: in an upcast the subject is a subclass of the target, in a downcast the target is a subclass of the subject, and in a stupid cast the target is unrelated to the subject. The Java compiler rejects as ill typed an expression containing a stupid cast, but we must allow stupid casts in FJ if we are to formulate type soundness as a subject reduction theorem for a small-step semantics. This is because a sensible expression may be reduced to one containing a stupid cast. For example, consider the fol- lowing, which uses classes A and B as defined as in the previous section: (A)(Object)new B() \Gamma! (A)new B() We indicate the special nature of stupid casts by including the hypothesis stupid warning in the type rule for stupid casts (T-SCast); an FJ typing corresponds to a legal Java typing only if it does not contain this rule. (Stupid casts were omitted from Classic Java [14], causing its published proof of type soundness to be incorrect; this error was discovered independently by ourselves and the Classic Java authors.) The typing judgment for method declarations has the form M OK IN C, read "method declaration M is ok if it occurs in class C." It uses the expression typing judgment on the body of the method, where the free variables are the parameters of the method with their declared types, plus the special variable this with type C. The typing judgment for class declarations has the form CL OK, read "class declaration CL is ok." It checks that the constructor applies super to the fields of the superclass and initializes the fields declared in this class, and that each method declaration in the class is ok. The type of an expression may depend on the type of any methods it invokes, and the type of a method depends on the type of an expression (its body), so it behooves us to check that there is no ill-defined circularity here. Indeed there is none: the circle is broken because the type of each method is explicitly declared. It is possible to load and use the class table before all the classes in it are checked, so long as each class is eventually checked. 2.3 Computation The reduction relation is of the form e \Gamma! e 0 , read "expression e reduces to expression e 0 in one step." We write \Gamma! for the reflexive and transitive closure of \Gamma!. The reduction rules are given in the bottom left column of Figure 1. There are three reduction rules, one for field access, one for method invocation, and one for casting. These were already explained in the introduction to this section. We write [d=x; e=y]e0 for the result of replacing x1 by d1 , . , xn by dn , and y by e in expression e0 . The reduction rules may be applied at any point in an expression, so we also need the obvious congruence rules (if e \Gamma! e .f, and the like), which we omit here. 2.4 Properties Formal definitions are fun, but the proof of the pudding is in. well, the proof. If our definitions are sensible, we should be able to prove a type soundness result, which relates typing to computation. Indeed we can prove such a result: if a term is well typed and it reduces to a second term, then the second term is well typed, and furthermore its type is a subtype of the type of the first term. 2.4.1 Theorem [Subject Reduction]: If \Gamma ' e 2 C and e \Gamma! e Proof sketch: The main property required in the proof is the following term-substitution lemma: If This is proved by induction on the derivation of \Gamma; x : interesting case is when the final rule used in the derivation is T-DCast. Suppose the type of e0 is C0 and . By the induction since D0 and C may or may not be in the subtype rela- tion, the derivation of \Gamma ' (C)[d=x]e 2 C may involve a stupid warning. On the other hand, if (C)e0 is derived using T-UCast, then (C)[d=x]e will also be an upcast. The theorem itself is now proved by induction on the derivation of e \Gamma! e , with a case analysis on the last rule used. The case for R-Invk is easy, using the lemma above. Other base cases are also straightforward, as are most of the induction steps. The only interesting case is the congruence rule for casting-that is, the case where (C)e \Gamma! (C)e 0 is derived using e \Gamma! e . Using an argument similar to the term substitution lemma above, we see that a downcast expression may be reduced to a stupid cast and an upcast expression will be always reduced to an upcast. \Xi We can also show that if a program is well typed, then the only way it can get stuck is if it reaches a point where it cannot perform a downcast. 2.4.2 Theorem [Progress]: Suppose e is a well-typed expression. (1) If e includes new C0(e).f as a subexpression, then (2) If e includes new C0(e).m(d) as a subexpression, then mbody(m; To state a similar property for casts, we say that an expression e is safe in \Gamma if the type derivations of the underlying CT and \Gamma ' e 2 C contain no downcasts or stupid casts (uses of rules T-DCast or T-SCast). In other words, a safe program includes only upcasts. Then we see that a safe expression always reduces to another safe expression, and, moreover, typecasts in a safe expression will never fail, as shown in the following pair of theorems. 2.4.3 Theorem: [Reduction preserves safety] If e is safe in \Gamma and e\Gamma!e 0 , then e 0 is safe in \Gamma. 2.4.4 Theorem [Progress of safe programs]: Suppose e is safe in \Gamma. If e has (C)new C0(e) as a subexpression, then 3 Featherweight GJ Just as GJ adds generic types to Java, Featherweight GJ (or FGJ, for short) adds generic types to FJ. Here is the class definition for pairs in FJ, rewritten with type parameters in FGJ. class Pair!X extends Object, Y extends Object? extends Object - Y snd; !Z extends Object? Pair!Z,Y? setfst(Z newfst) - return new Pair!Z,Y?(newfst, this.snd); class A extends Object - class B extends Object - Both classes and methods may have generic type pa- rameters. Here X and Y are parameters of the class, and Z is a parameter of the setfst method. Each type parameter has a bound ; here X, Y, and Z are each bounded by Object. In the context of the above definitions, the expres- sion new Pair!A,B?(new A(), new B()).setfst!B?(new B()) evaluates to the expression new Pair!B,B?(new B(), new B()) If we were being extraordinarily pedantic, we would write A!? and B!? instead of A and B, but we allow the latter as an abbreviation for the former in order that FJ is a proper subset of FGJ. In GJ, type parameters to generic method invocations are inferred. Thus, in GJ the expression above would be written new Pair!A,B?(new A(), new B()).setfst(new B()) with no !B? in the invocation of setfst. So while FJ is a subset of Java, FGJ is not quite a subset of GJ. We regard FGJ as an intermediate language - the form that would result after type parameters have been inferred. While parameter inference is an important aspect of GJ, we chose in FGJ to concentrate on modeling other aspects of GJ. The bound of a type variable may not be a type variable, but may be a type expression involving type variables, and may be recursive (or even, if there are several bounds, mutually recursive). For example, if C!X? and D!Y? are classes with one parameter each, one may have bounds such as !X extends C!X?? or even !X extends C!Y?, Y extends D!X??. For more on bounds, including examples of the utility of recursive bounds, see the GJ paper [7]. GJ and FGJ are intended to support either of two implementation styles. They may be implemented di- rectly, augmenting the run-time system to carry information about type parameters, or they may be implemented by erasure, removing all information about type parameters at run-time. This section explores the first style, giving a direct semantics for FGJ that maintains type parameters, and proving a type soundness theo- rem. Section 4 explores the second style, giving an erasure mapping from FGJ into FJ and showing a correspondence between reductions on FGJ expressions and reductions on FJ expressions. The second style corresponds to the current implementation of GJ, which compiles GJ into the Java Virtual Machine (JVM), which of course maintains no information about type parameters at run-time; the first style would correspond to using an augmented JVM that maintains information about type parameters. 3.1 In what follows, for the sake of conciseness we abbreviate the keyword extends to the symbol / . The syntax, typing rules, and computation rules for FGJ are given in Figure 3, with a few auxiliary functions in Figure 4. The metavariables X, Y, and Z range over type variables; T, U, and V range over types; and N and O range over nonvariable types (types other than type variables). We write X as shorthand for X1,. ,Xn (and similarly for T, N, etc.), and assume sequences of type variables contain no duplicate names. The abstract syntax of FGJ is given at the top left of Figure 3. We allow C!? and m!? to be abbreviated as C and m, respectively. As before, we assume a fixed class table CT , which is a mapping from class names C to class declarations CL, obeying the same sanity conditions as given previously. 3.2 Typing environment \Delta is a finite mapping from type variables to nonvariable types, written takes each type variable to its bound. Bounds of types We write bound \Delta (T) for the upper bound of T in \Delta, as defined Figure 4. Unlike calculi such as F [9], this promotion relation does not need to be defined recur- sively: the bound of a type variable is always a nonva- riable type. Subtyping The subtyping relation is defined in the left column of Figure 3. As before, subtyping is the reflexive and transitive closure of the / relation. Type parameters are invariant with regard to subtyping (for reasons explained in the GJ paper), so does not imply Well-formed types If the declaration of a class C begins class C!X / N?, then a type like C!T? is well formed only if substituting respects the bounds N, that is if We write well-formed in context \Delta. The rules for well-formed types appear in Figure 3. Note that we perform a simultaneous substitution, so any variable in X may appear in N, permitting recursion and mutual recursion between variables and bounds. A type environment \Delta is well formed if \Delta ' \Delta(X) ok for all X in dom (\Delta). We also say that an environment \Gamma is well formed with respect to \Delta, written if \Delta ' \Gamma(x) ok for all x in dom (\Gamma). Field and method lookup For the typing and reduction rules, we need a few auxiliary definitions, given in Figure 4; these are fairly straightforward adaptations of the lookup rules given previously. The fields of a nonvariable type N, written fields(N), are a sequence of corresponding types and field names, T f. The type of the method invocation m at nonvariable type N, written mtype(m; N), is a type of the form !X / N?U!U. Similarly, the body of the method invocation m at nonvariable type N with type parameters V, written mbody(m!V?; N), is a pair, written (x,e), of a sequence of parameters x and an expression e. Syntax: CL ::= class C!X / N? / N -T f; K M e ::= x new N(e) Subtyping: class C!X / N? / N -. Well-formed types: class C!X / N? / N -. Computation: (new N(e)).f i \Gamma! e i (new N(e)).m!V?(d) \Gamma! [d=x; new N(e)=this]e0 (O)(new N(e)) \Gamma! new N(e) Expression typing: stupid warning Method typing: class C!X / N? / N -. Class typing: class C!X / N? / N -T f; K M OK Figure 3: FGJ: Main definitions Bound of type: bound bound Field lookup: class C!X / N? / N -S f; K M class C!X / N? / N -S f; K M class C!X / N? / N -S f; K M m is not defined in M Method body lookup: class C!X / N? / N -S f; K M class C!X / N? / N -S f; K M m is not defined in M Valid method overriding: implies Figure 4: FGJ: Auxiliary definitions Typing rules Typing rules for expressions, methods, and classes appear in Figure 3. The typing judgment for expressions is of form read as "in the type environment \Delta and the environment \Gamma, e has type T." Most of the subtleties are in the field and method lookup relations that we have already seen; the typing rules themselves are straightforward. In the rule GT-DCast, the last premise ensures that the result of the cast will be the same at run time, no matter whether we use the high-level (type-passing) reduction rules defined later in this section or the erasure semantics considered in Section 4. For example, suppose we have defined: is, the result type of a method may be a subtype of the result type of the corresponding method in the su- perclass, although the bounds of type variables and the argument types must be identical (modulo renaming of type variables). As before, a class table is ok if all its class definitions are ok. 3.3 Reduction The operational semantics of FGJ programs is only a little more complicated than what we had in FJ. The rules appear in Figure 3. 3.4 Properties FGJ programs enjoy subject reduction and progress properties exactly like programs in FJ (2.4.1 and 2.4.2). The basic structures of the proofs are similar to those of Theorem 2.4.1 and 2.4.2. For subject reduction, how- ever, since we now have parametric polymorphism combined with subtyping, we need a few more lemmas. The main lemmas required are a term substitution lemma as before, plus similar lemmas about the preservation of subtyping and typing under type substitution. (Read- ers familiar with proofs of subject reduction for typed lambda-calculi like F [9] will notice many similarities). We begin with the three substitution lemmas, which are proved by straightforward induction on a derivation of 3.4.1 Lemma: [Type substitution preserves sub- with and none of X appearing in \Delta 1 , then 3.4.2 Lemma: [Type substitution preserves typ- and none of X appears in \Delta 1 , then such that 3.4.3 Lemma: [Term substitution preserves typ- such that 3.4.4 Theorem [Subject reduction]: If \Delta; \Gamma ' e 2 that Proof By induction on the derivation of e \Gamma! e 0 with a case analysis on the reduction rule used. We show in detail just the base case where e is a method invocation. From the premises of the rule GR-Invk, we have e By the rule GT-Invk and GT-New, we also have By examining the derivation of mtype(m; bound \Delta (N)), we can find a supertype C!T? of N where and none of the Y appear in T. Now, by Lemma 3.4.2, From this, a straightforward weakening lemma (not shown here), plus Lemma 3.4.3 and Lemma 3.4.1, gives Letting T finishes the case, since \Delta ' S by S-Trans. \Xi Theorem [Progress]: Suppose e is a well-typed expression. (1) If e includes new N0(e).f as a subexpression, then (2) If e includes new N0(e).m!V?(d) as a subexpres- sion, then mbody(m!V?; #(d). FGJ is backward compatible with FJ. Intuitively, this means that an implementation of FGJ can be used to typecheck and execute FJ programs without changing their meaning. We can show that a well-typed FJ program is always a well-typed FGJ program and that FJ and FGJ reduction correspond. (Note that it isn't quite the case that the well-typedness of an FJ program under the FGJ rules implies its well-typedness in FJ, because FGJ allows covariant overriding and FJ does not.) In the statement of the theorem, we use \Gamma! FJ and \Gamma! FGJ to show which set of reduction rules is used. 3.4.6 Theorem [Backward compatibility]: If an FJ program (e; CT ) is well typed under the typing rules of FJ, then it is also well-typed under the rules of FGJ. Moreover, for all FJ programs e and e (whether well typed or not), e \Gamma! FJ e Proof: The first half is shown by straightforward induction on the derivation of \Gamma ' e 2 C (using FJ typing rules), followed by an analysis of the rules GT-Method and GT-Class. In the second half, both directions are shown by induction on a derivation of the reduction re- lation, with a case analysis on the last rule used. \Xi Compiling FGJ to FJ We now explore the second implementation style for GJ and FGJ. The current GJ compiler works by translation into the standard JVM, which maintains no information about type parameters at run-time. We model this compilation in our framework by an erasure translation from FGJ into FJ. We show that this translation maps well-typed FGJ programs into well-typed FJ programs, and that the behavior of a program in FGJ matches (in a suitable sense) the behavior of its erasure under the FJ reduction rules. A program is erased by replacing types with their erasures, inserting downcasts where required. A type is erased by removing type parameters, and replacing type variables with the erasure of their bounds. For example, the class Pair!X,Y? in the previous section erases to the class Pair extends Object - Object fst; Object snd; Pair(Object fst, Object snd) - Pair setfst(Object newfst) - return new Pair(newfst, this.snd); Similarly, the field selection new Pair!A,B?(new A(), new B()).snd erases to (B)new Pair(new A(), new B()).snd where the added downcast (B) recovers type information of the original program. We call such downcasts inserted by erasure synthetic. 4.1 Erasure of Types To erase a type, we remove any type parameters and replace type variables with the erasure of their bounds. Write jTj \Delta for the erasure of type T with respect to type environment \Delta 4.2 Field and Method Lookup In FGJ (and GJ), a subclass may extend an instantiated superclass. This means that, unlike in FJ (and Java), the types of the fields and the methods in the subclass may not be identical to the types in the superclass. In order to specify a type-preserving erasure from FGJ to FJ, it is necessary to define additional auxiliary functions that look up the type of a field or method in the highest superclass in which it is defined. For example, we previously defined the generic class Pair!X,Y?. We may declare a specialized subclass PairOfA as a subclass of the instantiation Pair!A,A?, which instantiates both X and Y to a given class A. class PairOfA extends Pair!A,A? - super(fst, snd); PairOfA return new PairOfA(newfst, this.fst); Note that, in the setfst method, the argument type A matches the argument type of setfst in Pair!A,A?, while the result type PairOfA is a subtype of the result type in Pair!A,A?; this is permitted by FGJ's covariant subtyping, as discussed in the previous section. Erasing the class PairOfA yields the following: class PairOfA extends Pair - PairOfA(Object fst, Object snd) - super(fst, snd); Pair setfst(Object newfst) - return new PairOfA(newfst, this.fst); Here arguments to the constructor and the method are given type Object, even though the erasure of A is itself; and the result of the method is given type Pair, even though the erasure of PairOfA is itself. In both cases, the types are chosen to correspond to types in Pair, the highest superclass in which the fields and method are defined. We define variants of the auxiliary functions that find the types of fields and methods in the highest superclass in which they are defined. The maximum field types of a class C, written fieldsmax (C), is the sequence of pairs of a type and a field name defined as follows: class C!X / N? / D!U? -T f; . The maximum method type of m in C, written (m, C), is defined as follows: class C!X / N? / D!U? -. class C!X / N? / D!U? -. undefined We also need a way to look up the maximum type of a given field. If fieldsmax then we set 4.3 Erasure of Expressions The erasure of an expression depends on the typing of that expression, since the types are used to determine which downcasts to insert. The erasure rules are optimized to omit casts when it is trivially safe to do so; this happens when the maximum type is equal to the erased type. Write jej \Delta;\Gamma for the erasure of a well-typed expression e with respect to environment \Gamma and type environment \Delta: jnew N(e)j (Strictly speaking, one should think of the erasure operation as acting on typing derivations rather than expressions. Since well-typed expressions are in 1-1 correspondence with their typing derivations, the abuse of notation creates no confusion.) 4.4 Erasure of Methods and Classes The erasure of a method m with respect to type environment \Delta in class C, written jMj \Delta;C , is defined as follows: e (In GJ, the actual erasure is somewhat more complex, involving the introduction of bridge methods, so that one ends up with two overloaded methods: one with the maximum type, and one with the instantiated type. We don't model that extra complexity here, because it depends on overloading of method names, which is not modeled in FJ.) The erasure of constructors and classes is: jclass C!X / N? / N -T f; K Mj class C / jNj 4.5 Properties of Erasure Having defined erasure, we may investigate some of its properties. First, a well-typed FGJ program erases to a well-typed FJ program, as expected: 4.5.1 Theorem [Erasure preserves typing]: If an FGJ class table CT is ok and \Delta; \Gamma ' e 2 T, then using FJ rules. Proof By induction on the derivation of using the following lemmas: (1) if if \Delta ' N ok and methodtype FGJ (m; then mtypemax (m; jNj some well-formed type environment \Delta, then More interestingly, we would intuitively expect that erasure from FGJ to FJ should also preserve the reduction behavior of FGJ programs: e reduce erase eerase reduce Unfortunately, this is not quite true. For example, consider the FGJ expression where a and b are expressions of type A and B, respec- tively, and its erasure: (A)new Pair(jaj \Delta;\Gamma ,jbj \Delta;\Gamma ).fst In FGJ, e reduces to jaj \Delta;\Gamma , while the erasure jej \Delta;\Gamma reduces to (A)jaj \Delta;\Gamma in FJ; it does not reduce to jaj \Delta;\Gamma when a is not a new expression. (Note that it is not an artifact of our nondeterministic reduction strategy: it happens even if we adopt a call-by-value reduction strategy, since, after method invocation, we may obtain an expression like (A)e where e is not a new expres- sion.) Thus, the above diagram does not commute even if one-step reduction (\Gamma!) at the bottom is replaced with many-step reduction (\Gamma! ). In general, synthetic casts can persist for a while in the FJ expression, although we expect those casts will eventually turn out to be upcasts when a reduces to a new expression. In the example above, an FJ expression d reduced from jej \Delta;\Gamma had more synthetic casts than je ever, this is not always the case: d may have less casts than je when the reduction step involves method invocation. Consider the following class and its erasure: class C!X extends Object? extends Object - return new C!X?(this.f); class C extends Object - Object f; C(Object f) - return new C(this.f); Now consider the FGJ expression new C!A?(new A()).m() and its erasure new C(new A()).m(): In FGJ, e \Gamma! FGJ new C!A?(new C!A?(new A()).f): In FJ, on the other hand, jej \Delta;\Gamma reduces to new C(new C(new A()).f), which has fewer synthetic casts than new C((A)new C(new A()).f), which is the erasure of the reduced expression in FGJ. The subtlety we observe here is that, when the erased term is re- duced, synthetic casts may become "coarser" than the casts inserted when the reduced term is erased, or may be removed entirely as in this example. (Removal of downcasts can be considered as a combination of two operations: replacement of (A) with the coarser cast (Object) and removal of the upcast (Object), which does not affect the result of computation.) To formalize both of these observations, we define an auxiliary relation that relates FJ expressions differing only by the addition and replacement of some synthetic casts. Let us call a well-typed expression d an expansion of a well-typed expression e, written e =) d, if d is obtained from e by some combination of (1) addition of zero or more synthetic upcasts, (2) replacement of some synthetic casts (D) with (C), where C is a supertype of D, or (3) removal of some synthetic casts. 4.5.2 Theorem: [Erasure preserves reduction modulo expansion] If \Delta; \Gamma ' e 2 T and e \Gamma! FGJ e , then there exists some FJ expression d 0 such that 0 and jej \Delta;\Gamma \Gamma! FJ d 0 . In other words, the following diagram commutes. e reduce erase erase reduce d As easy corollaries of this theorem, it can be shown that, if an FGJ expression e reduces to a "fully-evaluated expression," then the erasure of e reduces to exactly its erasure, and that if FGJ reduction gets stuck at a stupid cast, then FJ reduction also gets stuck because of the same typecast. We use the metavariable v for fully evaluated expressions, defined as follows: 4.5.3 Corollary: [Erasure preserves execution results] If \Delta; \Gamma ' e 2 T and e \Gamma! FGJ v, then jvj \Delta;\Gamma . 4.5.4 Corollary: [Erasure preserves typecast er- e ehas a stuck subexpression (C!S?)new D!T?(e), then d 0 such that d 0 has a stuck subexpression (C)new D(d), where d are expansions of the erasures of e, in the same position (modulo synthetic casts) as the erasure of e 5 Related Work Core calculi for Java. There are several known proofs in the literature of type soundness for subsets of Java. In the earliest, Drossopoulou and Eisenbach [11] (using a technique later mechanically checked by Syme [20]) prove soundness for a fairly large subset of sequential Java. Like us, they use a small-step operational semantics, but they avoid the subtleties of "stupid casts" by omitting casting entirely. Nipkow and Oheimb [17] give a mechanically checked proof of soundness for a somewhat larger core language. Their language does include casts, but it is formulated using a "big- step" operational semantics, which sidesteps the stupid cast problem. Flatt, Krishnamurthi, and Felleisen [14] use a small-step semantics and formalize a language with both assignment and casting. Their system is somewhat larger than ours (the syntax, typing, and operational semantics rules take perhaps three times the space), and the soundness proof, though correspondingly longer, is of similar complexity. Their published proof of subject reduction is slightly flawed - the case that motivated our introduction of stupid casts is not handled properly - but the problem can be repaired by applying the same refinement we have used here. Of these three studies, that of Flatt, Krishnamurthi, and Felleisen is closest to ours in an important sense: the goal there, as here, is to choose a core calculus that is as small as possible, capturing just the features of Java that are relevant to some particular task. In their case, the task is analyzing an extension of Java with Common Lisp style mixins - in ours, extensions of the core type system. The goal of the other two systems, on the other hand, is to include as large a subset of Java as possible, since their primary interest is proving the soundness of Java itself. Other class-based object calculi. The literature on foundations of object-oriented languages contains many papers formalizing class-based object-oriented languages, either taking classes as primitive (e.g., [21, 8, 6, 5]) or translating classes into lower-level mechanisms (e.g., [13, 4, 1, 19]. Some of these systems (e.g. [19, 8]) include generic classes and methods, but only in fairly simple forms. Generic extensions of Java. A number of extensions of Java with generic classes and methods have been proposed by various groups, including the language of Agesen, Freund, and Mitchell [2]; PolyJ, by Myers, Bank, and Liskov [16]; Pizza, by Odersky and Wadler [18]; GJ, by Bracha, Oderksy, Stoutamire, and Wadler [7]; and NextGen, by Cartwright and Steele [10]. While all these languages are believed to be typesafe, our study of FGJ is the first to give rigorous proof of soundness for a generic extension of Java. We have used GJ as the basis for our generic extension, but similar techniques should apply to the forms of genericity found in the rest of these languages. 6 Discussion We have presented Featherweight Java, a core language for Java modeled closely on the lambda-calculus and embodying many of the key features of Java's type sys- tem. FJ's definition and proof of soundness are both concise and straightforward, making it a suitable arena for the study of ambitious extensions to the type sys- tem, such as the generic types of GJ. We have developed this extension in detail, stated some of its fundamental properties, and sketched their proofs. FJ itself is not quite complete enough to model some of the interesting subtleties found in GJ. In particular, the full GJ language allows some parameters to be instantiated by a special "bottom type" *, using a slightly delicate rule to avoid unsoundness in the presence of as- signment. Capturing the relevant issues in FGJ requires extending it with assignment and null values (both of these extensions seem straightforward, but cost us some of the pleasing compactness of FJ as it stands). The other somewhat subtle aspect of GJ that is not accurately modeled in FGJ is the use of bridge methods in the compilation from GJ to JVM bytecodes. To treat this compilation exactly as GJ does, we would need to extend FJ with overloading. Our formalization of GJ also does not include raw types, a unique aspect of the GJ design that supports compatibility between old, unparameterized code and new, parameterized code. We are currently experimenting with an extension of FGJ with raw types. Formalizing generics has proven to be a useful application domain for FJ, but there are other areas where its extreme simplicity may yield significant leverage. For example, work is under way on formalizing Java 1.1's inner classes using FJ [15]. Acknowledgments This work was supported by the University of Pennsylvania and the National Science Foundation under grant CCR-9701826, Principled Foundations for Programming with Objects. Igarashi is a research fellow of the Japan Society for the Promotion of Science. --R A Theory of Ob- jects Adding type parameterization to the Java language. The Lambda Calculus. An imperative first-order calculus with object extension A core calculus of classes and mix- ins A core calculus of classes and objects. Making the future safe for the past: Adding genericity to the Java programming language. "Safe type checking in a statically typed object-oriented programming language" An extension of system F with subtyping. Steele Jr. Is the Java Type System Sound? A little Java On the relationship between classes Classes and mixins. On inner classes Parameterized types for Java. Java light is type-safe - definitely Pizza into Java: Translating theory into practice. Simple type-theoretic foundations for object-oriented pro- gramming Proving Java type soundness. Type inference for objects with instance variables and inheritance. --TR An extension of system <italic>F</italic> with subtyping A syntactic approach to type soundness Parameterized types for Java Pizza into Java Adding type parameterization to the Java language Java<i>_light</i> is type-safeMYAMPERSANDmdash;definitely Classes and mixins A little Java, a few patterns Making the future safe for the past Compatible genericity with run-time types for the Java programming language On the relationship between classes, objects, and data abstraction Is the Java type system sound? Modular type-based reverse engineering of parameterized types in Java code Parametric polymorphism in Java Types and programming languages A Theory of Objects Partial Evaluation for Class-Based Object-Oriented Languages An Imperative, First-Order Calculus with Object Extension A Core Calculus of Classes and Mixins On Inner Classes True Modules for Java-like Languages --CTR Maurizio Cimadamore , Mirko Viroli, Reifying wildcards in Java using the EGO approach, Proceedings of the 2007 ACM symposium on Applied computing, March 11-15, 2007, Seoul, Korea Peter Hui , James Riely, Typing for a minimal aspect language: preliminary report, Proceedings of the 6th workshop on Foundations of aspect-oriented languages, p.15-22, March 13-13, 2007, Vancouver, British Columbia, Canada Giovanni Rimassa , Mirko Viroli, Understanding access restriction of variant parametric types and Java wildcards, Proceedings of the 2005 ACM symposium on Applied computing, March 13-17, 2005, Santa Fe, New Mexico Atsushi Igarashi , Hideshi Nagira, Union types for object-oriented programming, Proceedings of the 2006 ACM symposium on Applied computing, April 23-27, 2006, Dijon, France Marco Bellia , M. Eugenia Occhiuto, Higher order Programming in Java: Introspection, Subsumption and Extraction, Fundamenta Informaticae, v.67 n.1-3, p.29-44, January 2005 Tomoyuki Aotani , Hidehiko Masuhara, Towards a type system for detecting never-matching pointcut compositions, Proceedings of the 6th workshop on Foundations of aspect-oriented languages, p.23-26, March 13-13, 2007, Vancouver, British Columbia, Canada Jeffrey Fischer , Rupak Majumdar , Todd Millstein, Tasks: language support for event-driven programming, Proceedings of the 2007 ACM SIGPLAN symposium on Partial evaluation and semantics-based program manipulation, January 15-16, 2007, Nice, France Jaakko Jrvi , Jeremiah Willcock , Andrew Lumsdaine, Associated types and constraint propagation for mainstream object-oriented generics, ACM SIGPLAN Notices, v.40 n.10, October 2005 Gustavo Bobeff , Jacques Noy, Component specialization, Proceedings of the 2004 ACM SIGPLAN symposium on Partial evaluation and semantics-based program manipulation, p.39-50, August 24-25, 2004, Verona, Italy Alex Potanin , James Noble , Robert Biddle, Generic ownership: practical ownership control in programming languages, Companion to the 19th annual ACM SIGPLAN conference on Object-oriented programming systems, languages, and applications, October 24-28, 2004, Vancouver, BC, CANADA Neal Glew , Jens Palsberg, Type-safe method inlining, Science of Computer Programming, v.52 n.1-3, p.281-306, August 2004 Tian Zhao , Jens Palsber , Jan Vite, Lightweight confinement for featherweight java, ACM SIGPLAN Notices, v.38 n.11, November Vijay S. Menon , Neal Glew , Brian R. Murphy , Andrew McCreight , Tatiana Shpeisman , Ali-Reza Adl-Tabatabai , Leaf Petersen, A verifiable SSA program representation for aggressive compiler optimization, ACM SIGPLAN Notices, v.41 n.1, p.397-408, January 2006 Alex Potanin , James Noble , Dave Clarke , Robert Biddle, Featherweight generic confinement, Journal of Functional Programming, v.16 n.6, p.793-811, November 2006 Nathaniel Nystrom , Stephen Chong , Andrew C. Myers, Scalable extensibility via nested inheritance, ACM SIGPLAN Notices, v.39 n.10, October 2004 G. M. Bierman, Formal semantics and analysis of object queries, Proceedings of the ACM SIGMOD international conference on Management of data, June 09-12, 2003, San Diego, California Gurevich , Benjamin Rossman , Wolfram Schulte, Semantic essence of AsmL, Theoretical Computer Science, v.343 n.3, p.370-412, 17 October 2005 Ernst , Klaus Ostermann , William R. Cook, A virtual class calculus, ACM SIGPLAN Notices, v.41 n.1, p.270-282, January 2006 Todd Millstein , Colin Bleckner , Craig Chambers, Modular typechecking for hierarchically extensible datatypes and functions, ACM SIGPLAN Notices, v.37 n.9, p.110-122, September 2002 Suresh Jagannathan , Jan Vitek , Adam Welc , Antony Hosking, A transactional object calculus, Science of Computer Programming, v.57 n.2, p.164-186, August 2005 Lorenzo Bettini , Sara Capecchi , Elena Giachino, Featherweight wrap Java, Proceedings of the 2007 ACM symposium on Applied computing, March 11-15, 2007, Seoul, Korea Adriaan Moors , Frank Piessens , Wouter Joosen, An object-oriented approach to datatype-generic programming, Proceedings of the 2006 ACM SIGPLAN workshop on Generic programming, September 16-16, 2006, Portland, Oregon, USA Christian Skalka, Trace effects and object orientation, Proceedings of the 7th ACM SIGPLAN international conference on Principles and practice of declarative programming, p.139-150, July 11-13, 2005, Lisbon, Portugal Alessandro Warth , Milan Stanojevi , Todd Millstein, Statically scoped object adaptation with expanders, ACM SIGPLAN Notices, v.41 n.10, October 2006 Dan Grossman , Jeremy Manson , William Pugh, What do high-level memory models mean for transactions?, Proceedings of the 2006 workshop on Memory system performance and correctness, October 22-22, 2006, San Jose, California Matthew S. Tschantz , Michael D. Ernst, Javari: adding reference immutability to Java, ACM SIGPLAN Notices, v.40 n.10, October 2005 Juan Chen , David Tarditi, A simple typed intermediate language for object-oriented languages, ACM SIGPLAN Notices, v.40 n.1, p.38-49, January 2005 Polyvios Pratikakis , Jaime Spacco , Michael Hicks, Transparent proxies for java futures, ACM SIGPLAN Notices, v.39 n.10, October 2004 Marko van Dooren , Eric Steegmans, Combining the robustness of checked exceptions with the flexibility of unchecked exceptions using anchored exception declarations, ACM SIGPLAN Notices, v.40 n.10, October 2005 Franz Achermann , Oscar Nierstrasz, A calculus for reasoning about software composition, Theoretical Computer Science, v.331 n.2-3, p.367-396, 25 February 2005 Alexander Ahern , Nobuko Yoshida, Formalising Java RMI with explicit code mobility, ACM SIGPLAN Notices, v.40 n.10, October 2005 Gerwin Klein , Tobias Nipkow, A machine-checked model for a Java-like language, virtual machine, and compiler, ACM Transactions on Programming Languages and Systems (TOPLAS), v.28 n.4, p.619-695, July 2006 Philip W. L. Fong, Reasoning about safety properties in a JVM-like environment, Science of Computer Programming, v.67 n.2-3, p.278-300, July, 2007 Chris Andreae , Yvonne Coady , Celina Gibbs , James Noble , Jan Vitek , Tian Zhao, Scoped types and aspects for real-time Java memory management, Real-Time Systems, v.37 n.1, p.1-44, October 2007 dependent types for higher-order mobile processes, ACM SIGPLAN Notices, v.39 n.1, p.147-160, January 2004 Atsushi Igarashi , Mirko Viroli, Variant parametric types: A flexible subtyping scheme for generics, ACM Transactions on Programming Languages and Systems (TOPLAS), v.28 n.5, p.795-847, September 2006 Alex Potanin , James Noble , Dave Clarke , Robert Biddle, Generic ownership for generic Java, ACM SIGPLAN Notices, v.41 n.10, October 2006 Daniel J. Dougherty , Pierre Lescanne , Luigi Liquori, Addressed term rewriting systems: application to a typed object calculus, Mathematical Structures in Computer Science, v.16 n.4, p.667-709, August 2006 Tian Zhao , Jens Palsberg , Jan Vitek, Type-based confinement, Journal of Functional Programming, v.16 n.1, p.83-128, January 2006 Todd Millstein , Colin Bleckner , Craig Chambers, Modular typechecking for hierarchically extensible datatypes and functions, ACM Transactions on Programming Languages and Systems (TOPLAS), v.26 n.5, p.836-889, September 2004 Chris Andreae , James Noble , Shane Markstrum , Todd Millstein, A framework for implementing pluggable type systems, ACM SIGPLAN Notices, v.41 n.10, October 2006 Radha Jagadeesan , Alan Jeffrey , James Riely, Typed parametric polymorphism for aspects, Science of Computer Programming, v.63 n.3, p.267-296, 15 December 2006 Martin Abadi , Cormac Flanagan , Stephen N. Freund, Types for safe locking: Static race detection for Java, ACM Transactions on Programming Languages and Systems (TOPLAS), v.28 n.2, p.207-255, March 2006 Einar Broch Johnsen , Olaf Owe , Ingrid Chieh Yu, Creol: a type-safe object-oriented model for distributed concurrent systems, Theoretical Computer Science, v.365 n.1, p.23-66, 10 November 2006 Anindya Banerjee , David A. Naumann, Ownership confinement ensures representation independence for object-oriented programs, Journal of the ACM (JACM), v.52 n.6, p.894-960, November 2005

language design;compilation;generic classes;language semantics

504212

Analysis and comparison of two general sparse solvers for distributed memory computers.

This paper provides a comprehensive study and comparison of two state-of-the-art direct solvers for large sparse sets of linear equations on large-scale distributed-memory computers. One is a multifrontal solver called MUMPS, the other is a supernodal solver called superLU. We describe the main algorithmic features of the two solvers and compare their performance characteristics with respect to uniprocessor speed, interprocessor communication, and memory requirements. For both solvers, preorderings for numerical stability and sparsity play an important role in achieving high parallel efficiency. We analyse the results with various ordering algorithms. Our performance analysis is based on data obtained from runs on a 512-processor Cray T3E using a set of matrices from real applications. We also use regular 3D grid problems to study the scalability of the two solvers.

Introduction We consider the direct solution of sparse linear equations on distributed memory computers where communication is by message passing, normally using MPI. We study in detail two state-of-the-art solvers, MUMPS (Amestoy, Duff, L'Excellent and Koster 1999, Amestoy, Duff and L'Excellent 2000) and SuperLU (Li and Demmel 1999). The first uses a multifrontal approach with dynamic pivoting for stability while the second is based on a supernodal technique with static pivoting and iterative refinement. We discuss the detailed algorithms used in these two codes in Section 3. Two very important factors affecting the performance of both codes are the use of preprocessing to preorder the matrix so that the diagonal entries are large relative to the off-diagonals and the strategy used to compute an ordering for the rows and columns of the matrix to preserve sparsity. We discuss these aspects in detail in Section 4. We compare the performance of the two codes in Section 5, where we show that such a comparison can be fraught with difficulties even though the authors of both codes are involved in the study. In Section 6, regular grids problems are used to further illustrate and analyse the difference between the two approaches. We had originally planned a comparison of more sparse codes but, given the difficulties we have found in assessing codes that we know well, we have for the moment shelved this more ambitious project. However, we feel that the lessons that we have learned in this present exercise are both invaluable to us in our future wider study and have given us some insight into the behaviour of sparse direct codes which we feel is useful to share with a wider audience at this stage. In addition to valuable information on the comparative merits of multifrontal versus supernodal approaches, we have examined the parameter space for such a comparison exercise and have identified several key parameters that influence to a differing degree the two approaches. Test environment Throughout this paper, we will use a set of test problems to evaluate the performance of our algorithms. Our test matrices come from the forthcoming Rutherford-Boeing Sparse Matrix Collection (Duff, Grimes and Lewis 1997) 1 , the industrial partners of the Project 2 , Tim Davis' collection 3 , SPARSEKIT2 4 and the EECS Department of UC Berkeley 5 . The PARASOL test matrices are available from Parallab, Bergen, Norway 6 . Two smaller matrices (garon2 and lnsp3937) are included in our set of matrices but will be used only in Section 4.1 to illustrate differences in the numerical behaviour of the two solvers. Note that, for most of our experiments, we do not consider symmetric matrices in our test set because SuperLU cannot exploit the symmetry and is unable to produce an LDL T factorization. However, since our test examples in Section 6 are symmetric, we do Web page http://www.cse.clrc.ac.uk/Activity/SparseMatrices/ Project 20160 3 Web page http://www.cise.ufl.edu/-davis/sparse/ 4 Web page http://math.nist.gov/MatrixMarket/data/SPARSKIT/ 5 Matrix ecl32 is included in the Rutherford-Boeing Collection 6 Web page http://www.parallab.uib.no/parasol/ Real Unsymmetric Assembled (rua) Matrix name Order No. of entries bbmat 38744 1771722 0.54 Rutherford-Boeing (CFD) Department of UC Berkeley garon2 13535 390607 1.00 Davis collection (CFD) lhr71c 70304 1528092 0.00 Davis collection (Chem Eng) Rutherford-Boeing (CFD) mixtank 29957 1995041 1.00 PARASOL (Polyflow S.A.) collection (CFD) twotone 120750 1224224 0.28 Rutherford-Boeing (circuit sim) wang4 26068 177196 1.00 Rutherford-Boeing (semiconductor) Table 2.1: Test matrices. ( ) StrSym is the number of nonzeros matched by nonzeros in symmetric locations divided by the total number of entries (so that a structurally symmetric matrix has value 1.0). show results with both the symmetric and unsymmetric factorization versions of MUMPS. Matrices mixtank and invextr1 have been modified because of out-of-range (underflow) values in matrix files. To keep the same sparsity pattern, we did not want to replace those underflow values by zeros. Instead, we have replaced all entries with an exponent smaller than -300 to numbers with the same mantissa but with an exponent of -300. For each linear system, the right-hand side vector is generated so that the true solution is a vector of all ones. All results presented in this paper have been obtained on the Cray T3E-900 (512 DEC EV-5 processors, 256 Mbytes of memory per processor, 900 peak Megaflop rate per processor) from NERSC at Lawrence Berkeley National Laboratory. We will also refer to experiments on a 35 processor IBM SP2 (66.5 MHertz processor with 128 Mbytes of physical memory and 512 Mbytes of virtual memory and 266 peak Megaflop rate per processor) at GMD in Bonn, Germany, used during the PARASOL Project. The performance characteristics of the two machines are listed in Table 2.2. Computer CRAY T3E-900 IBM SP2 Frequency of the processor 450 MHertz 66 MHertz Peak uniproc. performance 900 Mflops 264 Mflops Effective uniproc. performance 340 Mflops 150 Mflops Peak communication bandwidth 300 Mbytes/sec 36 Mbytes/sec Latency Bandwidth/Effective performance 0.88 0.24 Table 2.2: Characteristics of the CRAY T3E-900 and the IBM SP2. The factorization of matrix wang4 using MUMPS was used to estimate the effective uniprocessor performance of the computers. 3 Description of the algorithms used In this section, we briefly describe the main characteristics of the algorithms used in the solvers and highlight the major differences between them. For a complete description of the algorithms, the reader should consult previous papers by the authors of these algorithms (Amestoy et al. 1999, Amestoy et al. 2000, Li and Demmel 1998, Li and Demmel 1999). Both algorithms can be described by a computational tree whose nodes represent computations and whose edges represent transfer of data. In the case of the multifrontal method, MUMPS, some steps of Gaussian elimination are performed on a dense frontal matrix at each node and the Schur complement (or contribution block) that remains is passed for assembly at the parent node. In the case of the supernodal code, SuperLU, the distributed memory version uses a right-looking formulation which, having computed the factorization of a block of columns corresponding to a node of the tree, then immediately sends the data to update the block columns corresponding to ancestors in the tree. Both codes can accept any pivotal ordering and both have a built-in capability to generate an ordering based on an analysis of the pattern of A + A T , where the summation is performed symbolically. However, for the present version of MUMPS, the symbolic factorization is markedly less efficient if an input ordering is given since different logic is used in this case. The default ordering used by MUMPS is approximate minimum degree (AMD) (Amestoy, Davis and Duff 1996a) while the default for SuperLU is multiple minimum degree (MMD) (Liu 1985). However, in our experiments using a minimum degree ordering, we considered only the AMD ordering since both codes can generate this using the subroutine MC47 from HSL (2000). It is usually far quicker than MMD and produces a symbolic factorization close to that produced by MMD. We also use nested dissection orderings (ND). Sometimes we use the ON-MeTiS ordering from MeTiS (Karypis and Kumar 1998), and sometimes the nested dissection/haloamd ordering from SCOTCH (Pellegrini, Roman and Amestoy 1999) depending on which performs better on each particular problem. In addition, it is sometimes very beneficial to precede the ordering by performing an unsymmetric permutation to place large entries on the diagonal and then scaling the matrix so that the diagonals are all of modulus one and the off-diagonals have modulus less than or equal to one. We use the MC64 code of HSL to perform this preordering and scaling (Duff and Koster 1999) and indicate clearly when this is done. The effect of using this preordering of the matrix is discussed in detail in Section 4.1. Finally, when MC64 is not used, our matrices are always scaled. In both approaches, a pivot order is defined by the analysis and symbolic factorization stages. In MUMPS, the modulus of the prospective pivot is compared with the largest modulus of an entry in the row and is only accepted if this is greater than a threshold value, typically between 0.001 and 0.1 (our default value is 0.01). Note that, although MUMPS can choose pivots from off the diagonal, the largest entry in the column might be unavailable for pivoting at this stage if all entries in its row are not fully summed. This threshold pivoting strategy is common in sparse Gaussian elimination and helps to avoid excessive growth in the size of entries during the matrix factorization and so directly reduces the bound on the backward error. If a prospective pivot fails the test, all that happens is that it is kept in the Schur complement and is passed to the parent node. Eventually all rows with entries in the column will be available for pivoting, at the root if not before, so that a pivot can be chosen from the column. Thus the numerical factorization can respect the threshold criterion but at the cost of increasing the size of the frontal matrices and causing more work and fill-in than were forecast. For the SuperLU approach, a static pivoting strategy is used and we keep to the pivotal sequence chosen in the analysis. The magnitude of the potential pivot is tested against a threshold of ffl 1=2 jjAjj, where ffl is the machine precision and jjAjj is the norm of A. If it is less than this value it is immediately set to this value (with the same sign) and the modified entry is used as pivot. This corresponds to a half-precision perturbation to the original matrix entry. The result is that the factorization is not exact and iterative refinement may be needed. Note that, after iterative refinement, we obtained an accurate solution in all the cases that we tested. If problems were still to occur then extended precision BLAS (Li, Demmel, Bailey, Henry, Hida, Iskandar, Kahan, Kapur, Martin, Tung and Yoo 2000) could be used. 3.1 MUMPS main parallel features The parallelism within MUMPS is at two levels. The first uses the structure of the assembly tree, exploiting the fact that computations at nodes that are not ancestors or descendents are independent. The initial parallelism from this source (tree parallelism) is the number of leaf nodes but this reduces to one at the root. The second level is in the subdivision of the elimination operations through blocking of the frontal matrix. This blocking gives rise to node parallelism and is either by rows (referred to as 1D-node parallelism) or by rows and columns (at the root and referred to as 2D-node parallelism). Node parallelism depends on the size of the frontal matrix which, because of delayed pivots, is only known at factorization time. Therefore, this is determined dynamically. Each tree node is assigned a processor a priori, but the subassignment of blocks of the frontal matrix is done dynamically. Most of the machine dependent parameters in MUMPS that control the efficiency of the code are designed to take into account both the uniprocessor and multiprocessor characteristics of the computers. Because of the dynamic distributed scheduling approach, we do not need as precise a description of the performance characteristics of the computer as for approaches based on static scheduling such as PaStiX (Henon, Ramet and Roman 1999). Most of the machine dependent parameters in MUMPS are associated with the block sizes involved in the parallel blocked factorization algorithms of the dense frontal matrices. Our main objective is to maintain a minimum granularity to efficiently exploit the potential of the processor while providing sufficient tasks to exploit the available parallelism. Our target machines differ in several respects. The most important ones are illustrated in Table 2.2. We found that smaller granularity tasks could be used on the CRAY T3E than on the IBM SP2 because of the relatively faster rate of communication to Megaflop rate on the CRAY T3E than on the IBM SP2 (see Table 2.2). That is to say that the communication is relatively more efficient on the CRAY T3E. Dynamic scheduling is a major and original feature of the approach used in MUMPS. A critical part of this algorithm is when a process associated with a tree node decides to reassign some of its work, corresponding to a partitioning of the rows, to a set of so-called worker processes. We call such a node a one-dimensional parallel node. In earlier versions of MUMPS, a fixed block size is used to partition the rows and work is distributed to processes starting with the least loaded process. (The load of a process is determined by the amount of work (number of operations) allocated to it and not yet processed, which can be determined very cheaply.) Since the block size is fixed, it is possible for a process in charge of a one-dimensional parallel node to give additional work to processes that are already more loaded than itself. This can happen near the leaf nodes of the tree where sparsity provides enough parallelism to keep all processes busy. On the other hand, insufficient tasks might be created to provide work to all idle processes. This situation is more likely to occur close to the root of the tree. In the new algorithm (available since Version 4.1 of MUMPS), the block size for the one-dimensional partitioning can be dynamically adjusted by the process in charge of the node. Early in the processing of the tree (that is, near the leaves) this gives a relatively bigger block size so reducing the number of worker processes; whereas close to the root of the tree the block size will be automatically reduced to compensate for the lack of parallelism in the assembly tree. We bound the block size for partitioning a one-dimensional parallel node by an interval. The lower bound is needed to maintain a minimum task granularity and control the volume of messages. The upper bound of the interval is less critical (it is by default chosen to be about eight times the lower bound) but it is used in estimating the maximum size of the communication buffers and of the factors and so should not be too large. This "all dynamic" strategy of both partitioning and distributing work onto the processors could cause some trouble on a large number of processors (more than 128). In that case, it can be quite beneficial to take into account some "global" information to help the local decisions. For example one could restrict the choice of worker processes to a set of candidate processors determined statically during the analysis phase. This notion, commonly used in the design of static scheduling algorithms such as that in Henon et al. (1999), could reduce the overhead of the dynamic scheduling algorithm, reduce the increase in the communication volume when increasing the number of processors, and improve the local decision. The tuning of the parameters controlling the block size for 1D partitioning would then be easier and the estimation of the memory required during factorization would be more accurate. On a large number of processors, both performance and software improvements could thus be expected. This feature is not available in the current Version 4.1 of MUMPS but will be implemented in a future release. We will see that by adding this feature, one could address some of the current limitations of the MUMPS approach, see Section 5.2. The solution phase is also performed in parallel and uses asynchronous communications both for the forward elimination and the back substitution. In the case of the forward elimination, the tree is processed from the leaves to the root, in a similar way to the factorization, while the back substitution requires a different algorithm that processes the tree from the root to the leaves. A pool of ready-to-be-activated tasks is used. We do not change the distribution of the factors as generated in the factorization phase. Hence, type 2 and 3 node parallelism are also used in the solution phase. 3.2 SUPERLU main parallel features SuperLU also uses two levels of parallelism although more advantage is taken of the node parallelism through blocking of the supernodes. Because the pivotal order is fully determined at the analysis phase, the assignment of blocks to processors can be done statically a priori before the factorization commences. A 2D block-cyclic layout is used and the execution can be pipelined since the sequence is predetermined. The matrix partitioning is based on the notion of an unsymmetric supernode introduced in Demmel, Eisenstat, Gilbert, Li and Liu (1999). The supernode is defined over the matrix factor L. A supernode is a range of columns of L with the triangular block just below the diagonal being full, and the same nonzero structure elsewhere (this is either full or zero). This supernode partition is used as the block partition in both row and column dimensions, that is the diagonal blocks are square. If there are N supernodes in an n-by-n matrix, there will be N 2 blocks of non-uniform size. Figure 3.1 illustrates such a block partition. The off-diagonal blocks may be rectangular and need not be full. Furthermore, the columns in a block of U do not necessarily have the same row structure. We call a dense subvector in a block of U a segment. The P processes are also arranged as a 2D mesh of dimension . By block-cyclic layout, we mean block (I; J) (of L or U) is mapped onto the process at coordinate of the process mesh. During the factorization, block L(I; J) is only needed by the processes on the process row Similarly, block U(I; J) is only needed by the processes on the process column This partitioning and mapping can be controlled by the user. First, the user can set the maximum block size parameter. The symbolic factorization algorithm identifies supernodes, and chops the large supernodes into smaller ones if their sizes exceed this parameter. The supernodes may be smaller than this parameter due to sparsity and the blocks are then defined by the supernode boundaries. (That is, supernodes can be smaller than the maximum block size but never larger.) Our experience has shown that a good value for this parameter on the IBM SP2 is around 40, while on the Cray T3E it is around 24. Second, the user can set the shape of the process grid, such as 2 \Theta 3 or 3 \Theta 2. The more square the grid, the better the performance expected. This rule of thumb was used on the Cray T3E to define the grid shapes.4301204441 00130143 1Global Matrix5Process Mesh5 Figure 3.1: The 2D block-cyclic layout used in SuperLU. In this 2D mapping, each block column of L resides on more than one process, namely, a column of processes. For example in Figure 3.1, the second block column of L resides on the column processes f1, 4g. Process 1 only owns two nonzero blocks, which are not contiguous in the global matrix. The main numerical kernel involved during numerical factorization is a block update corresponding to the rank-k update to the Schur complement: see Figure 3.2. In the earlier versions of SuperLU, this computation was based on Level 2.5 BLAS. That is, we call the Level 2 BLAS routine GEMV (matrix-vector product) but with multiple vectors (segments), and the matrix L(I; K) is kept in cache across these multiple calls. This to some extent mimics the Level 3 BLAS GEMM (matrix-matrix product) performance. However, the difference between Level 2.5 and Level 3 is still quite large on many machines, e.g. the IBM SP2. This motivated us to modify the kernel in the following way in order to use Level 3 BLAS. For best performance, we distinguish two cases corresponding to the two shapes of a U(K;J) block. ffl The segments in U(K;J) are of same height, as shown in Figure 3.2 (a). Since the nonzero segments are stored contiguously in memory, we can call GEMM directly, without performing operations on any zeros. ffl The segments in U(K;J) are of different heights, as shown in Figure 3.2 (b). In this case, we first copy the segments into a temporary working array T , with some leading zeros padded if necessary. We then call GEMM using L(I; K) and T (instead of U(K;J)). We perform some extra floating-point operations for those padding zeros. The copying itself does not incur a run time cost, because the data must be loaded in the cache anyway. The working storage T is bounded by the maximum block size, which is a tunable parameter. For example, we usually use \Theta 40 on the IBM SP2 and 24 \Theta 24 on the Cray T3E. Depending on the matrix, this Level 3 BLAS kernel improved the uniprocessor factorization time by about 20% to 40% on the IBM SP2. A performance gain was also observed on the Cray T3E. It is clear that the extra operations are well offset by the benefit of the more efficient Level 3 BLAS routines. (b) U(K, A(I, J) L(I, K) U(K, J) (a) U(K, Figure 3.2: Illustration of the numerical kernels used in SuperLU. The current factorization algorithm has two limitations to parallelism. Here we explain, by examples, what the problems are and speculate how the algorithm may be improved in the future. In the following matrix notation, the zero blocks are left blank. For each nonzero block we mark in box the process which owns the block. ffl Parallelism from the sparsity. Consider a matrix with 4-by-4 blocks mapped onto a 2-by-2 process mesh660 1 Although node 2 is the parent of node 1 in the elimination tree (associated with A T +A), not all processes in column 2 depend on column 1. Only process 1 depends on the L block on process 0. Process 3 could start factorizing column 2 at the same time as process 0 is factorizing column 1, before process 1 starts factorizing column 2. But the current algorithm requires all the column processes to factorize the column synchronously, thereby introducing idle time for process 3. We can relax this constraint by allowing the diagonal process (3 in this case) to factorize the diagonal block and then send the factored block down to the off-diagonal processes (using mpi isend), even before the off-diagonal processes are ready for this column. This would eliminate some artificial interprocess dependencies and potentially reduce the length of the critical path. Note that this kind of independence comes from not only the sparsity but also the 2D process-to-matrix mapping. An even more interesting study would be to formalize these 2D task dependencies into a task graph, and perform some optimal scheduling on it. ffl Parallelism from the directed acyclic elimination graphs (Gilbert and Liu 1993) often referred to as elimination dags or edags. Consider another matrix with 6-by-6 blocks mapped onto a 2-by-3 process mesh66664 Columns 1 and 3 are independent in the elimination dags. The column process sets f0, 3g and f2, 5g could start factorizing columns 1 and 3 simultaneously. However, since process 2 is also involved in the update task of block (5; associated with Step 1 and our algorithm gives precedence to all the tasks in Step 1 over any task in Step 3, process 2 does not factorize column 3 immediately. We may change this task precedence by giving the factorization task of a later step higher priority than the update tasks of the previous steps, because the former is more likely to be on the critical path. This would exploit better the task independence coming from the elimination dags. We expect the above improvements will have a large impact for very sparse and/or very unsymmetric matrices, and for the orderings that give wide and bushy elimination trees, such as nested dissection. The triangular solution algorithm is also designed around the same distributed 2D data structure. The forward substitution proceeds from the bottom of the elimination tree to the root, whereas the back substitution proceeds from the root to the bottom. The algorithm is based on a sequential variant called "inner product" formulation. The execution of the program is completely message-driven. Each process is in a self-scheduling loop, performing appropriate local computation depending on the type of the message received. The entirely asynchronous approach enables large overlap between communication and computation and helps to overcome the much higher communication to computation ratio in this phase. 3.3 First comments on the algorithmic differences Both approaches use Level 3 BLAS to perform the elimination operations. However, in MUMPS the frontal matrices are always square. It is possible that there are zeros in the frontal matrix especially if there are delayed pivots or the matrix structure is markedly asymmetric but the present implementation takes no advantage of this sparsity and all the counts measured assume the frontal matrix is dense. It is shown in Amestoy and Puglisi (2000) that one can detect and exploit the structural asymmetry of the frontal matrices. With this new algorithm, significant gains both in memory and in time to perform the factorization can be obtained. For example, using MUMPS with the new algorithm, the number of operations to factorize matrices lhr71c and twotone would be reduced by 30% and 37%, respectively. The approach, tested on a shared memory multifrontal code (Amestoy and Duff 1993) from HSL (2000), is however not yet available in the current version of MUMPS. In SuperLU, advantage is taken of sparsity in the blocks and usually the dense matrix blocks are smaller than those used in MUMPS. In addition, SuperLU uses a more sophisticated data structure to keep track of the irregularity in sparsity. Thus, the uniprocessor Megaflop rate of SuperLU is much worse than that of MUMPS. This performance penalty is to some extent alleviated by the reduction in floating-point operations because of the better exploitation of sparsity. As a rule of thumb, MUMPS will tend to perform particularly well when the matrix structure is close to symmetric while SuperLU can better exploit asymmetry. We note that, even if the same ordering is input to the two codes, the computational tree generated in each case will be different. In the case of MUMPS, the assembly tree generated by MC47 is used to drive the MUMPS factorization phase, while, for SuperLU, the directed acyclic computational graphs (dags) are built implicitly. In Figures 3.3 and 3.4, we use a vampir trace (Nagel, Arnold, Weber, Hoppe and Solchenbach 1996) to illustrate the typical parallel behaviour of both approaches. These traces correspond to a zoom in the middle of the factorization phase of matrix bbmat on 8 processors of the CRAY T3E. Black areas correspond to time spent in communications and related MPI calls. Each line between two processes corresponds to one message transfer. From the plots we can see that SuperLU has distinct phases for local computation and interprocess communication, whereas for MUMPS, it is hard to distinguish when the process performs computation and when it transfers a message. This is due to the asynchronous scheduling algorithm used in MUMPS which may have a better chance of overlapping communication with computation. 4 Impact of preprocessing and numerical issues In this section, we first study the impact on both solvers of the preprocessing of the matrix. In this preprocessing, we first use row or column permutations to permute large entries onto the diagonal. In Section 4.1, we report and compare both the structural and the numerical impact of this preprocessing phase on the performance and accuracy of our solvers. After this phase, a symmetric ordering (minimum degree or nested dissection) is used and we study the relative influence of these orderings on the performance of the solvers in Section 4.2. We also comment on the relative cost of the analysis phase of the two solvers. 4.1 Use of a preordering to place large entries onto the diagonal and the cost of the analysis phase Duff and Koster (1999) developed an algorithm for permuting a sparse matrix so that the diagonal entries are large relative to the off-diagonal entries. They have also written a computer code, MC64 (available from HSL (2000)), to implement this algorithm. Here, we use option 5 of MC64 which maximizes the product of the modulus of the diagonal entries and then scales the permuted matrix so that it has diagonal entries of modulus one and all off-diagonals of modulus less than or equal to one. The importance of this preordering and scaling is clear. For MUMPS it should limit the amount of numerical pivoting during the factorization, which increases the overall cost of the factorization. For SuperLU, we expect such a permutation to be even more crucial, reducing the amount of small pivots that are modified and set to " 1=2 jjAjj. The MC64 code of Duff and Koster (1999) is quite efficient and so should normally require little time relative to the matrix factorization even if the latter is executed on many processors while MC64 runs on only one processor. Results in this section will show that it is not always the case. Moreover, matrices which are unsymmetric but have a symmetric or nearly symmetric structure are a very common problem class. The problem with these is that MC64 performs an unsymmetric permutation and will tend to destroy the symmetry of the pattern. Since both codes use a symmetrized pattern for the sparsity ordering (see Section 4.2) and MUMPS uses one also for the symbolic and numerical factorization, the overheads in having a markedly unsymmetric pattern can be high. Conversely, when the initial matrix is very unsymmetric (as for example lhr71c) the unsymmetric permutation may actually help to increase structural symmetry thus giving a second benefit to the subsequent matrix factorization. We show the effects of using MC64 on some examples in Table 4.1. In Table 4.4, we illustrate the relative cost of the main steps of the analysis phase when MC64 is used to preprocess the matrix. We see in Table 4.1 that, for very unsymmetric matrices (lhr71c and twotone), MC64 is really needed by MUMPS and SuperLU to factorize these matrices efficiently. Both matrices have zeros on the diagonal. Because of the static pivoting approach used by SuperLU, unless these zeros are made nonzero by fill-in and are then large enough, they will be perturbed Process Process Process Process Process Process Process 6 108 108 5 Process 7 108 MPI Application 9.05s 9.0s 8.95s 8.9s Figure 3.3: Illustration of the asynchronous behaviour of the MUMPS factorization phase. Process 0 Process 1 Process Process 3 Process 4 Process 5 Process 6 Process 7 MPI VT_API Comm 9.32s 9.3s 9.28s Figure 3.4: Illustration of the relatively more synchronous behaviour of the SuperLU factorization phase. Matrix Solver Ordering StrSym Nonzeros Flops in factors bbmat MUMPS AMD 0.54 46.1 41.5 fidapm11 MUMPS AMD 1.00 16.1 9.7 garon2 MUMPS AMD 1.00 2.4 0.3 mixtank MUMPS AMD 1.00 39.1 64.4 twotone MUMPS AMD 0.28 235.0 1221.1 wang4 MUMPS AMD 1.00 11.6 10.5 Table 4.1: Impact of permuting large entries onto the diagonal (using MC64) on the size of the factors and the number of operations. ( ) estimation given by the analysis (not enough memory to perform factorization). StrSym denotes the structural symmetry after ordering. during factorization and a factorization of a nearby matrix is obtained. In the case of MUMPS, the dramatically higher fill-in obtained without MC64 makes it also necessary to use MC64. For MUMPS, the main benefit from using MC64 is more structural than numerical. The permuted matrix has in fact a larger structural symmetry (see column 4 of Table 4.1) so that a symmetric permutation can be obtained on the permuted matrix that is more efficient in preserving sparsity. SuperLU benefits in a similar way from symmetrization because the computation of the symmetric permutation is based on the same assumption even if SuperLU preserves better the asymmetric structure of the factors by performing a symbolic analysis on a directed acyclic graph and exploiting asymmetry in the factorization phase (compare, for example, results with MUMPS and SuperLU on matrices lhr71c, mixtank and twotone). Matrix Iter. No bbmat Table 4.2: Illustration of the convergence of iterative refinement. The use of MC64 can also improve the quality of the factors and the numerical behaviour of the factorization phase, and can reduce the number of steps of iterative refinement required to reduce the backward error to machine precision. This is illustrated in Table 4.2 where we show the number of steps of iterative refinement required to reduce the componentwise relative backward error, (Arioli, Demmel and Duff 1989), to machine precision (" - 2:2 \Theta 10 \Gamma16 on the CRAY T3E). Iterative refinement will stop when either the required accuracy is reached or the convergence rate is too slow (Berr does not decrease by at least a factor of two). The true error is reported as jjx true jj . This table illustrates the impact of the use of MC64 on the quality of Matrix Solver WITHOUT Iter. Ref. WITH Iterative Refinement Berr bbmat MUMPS 7.4e-11 1.3e-06 2 3.2e-16 3.0e-09 lhr71c MUMPS Not enough memory SuperLU Not enough memory mixtank MUMPS 1.9e-12 4.8e-09 2 5.9e-16 1.4e-11 twotone MUMPS 5.0e-07 1.3e-05 3 1.3e-15 2.1e-11 Matrix Solver WITHOUT Iter. Ref. WITH Iterative Refinement Berr bbmat MUMPS 1.2e-11 6.5e-08 2 2.7e-16 3.5e-09 lhr71c MUMPS 1.1e-05 9.9e+00 3 3.2e-13 1.0e+00 mixtank MUMPS 4.8e-12 2.3e-08 2 4.2e-16 4.0e-11 twotone MUMPS 3.2e-13 1.6e-10 2 1.6e-15 2.3e-11 Table 4.3: Comparison of the numerical behaviour, backward error (Berr) and forward error (Err), of the solvers. Nb indicates the number of steps of iterative refinement. the initial solution obtained with both solvers prior to iterative refinement. Additionally, it shows that, thanks to numerical partial pivoting, the initial solution is almost always more accurate with MUMPS than with SuperLU and is usually markedly so. These observations are further confirmed on a larger number of test matrices in Table 4.3. The same stopping criterion was applied for these runs as for the runs in Table 4.2. In the case of MUMPS, MC64 can also result in a reduction in the number of off-diagonal pivots and in the number of delayed pivots. For example on the matrix invextr1 the number of off-diagonal pivots drops from 1520 to 109 and the number of delayed pivots drops from 2555 to 42. One can also see in Table 4.2 (e.g., bbmat) that MC64 does not always improve the numerical accuracy of the solution obtained with SuperLU. As expected, we see that, for matrices with a fairly symmetric pattern (e.g., matrix fidapm11 in Table 4.1), the use of MC64 leads to a significant decrease in symmetry which, for both solvers, results in a significant increase in the number of operations during factorization. We additionally recollect that the time spent in MC64 can dominate the analysis time of either solver (see Table 4.4), even for matrices such as fidapm11 and invextr1 for which it does not provide any gain for the subsequent steps. Thus, for both solvers, the default should be to not use MC64 on fairly symmetric matrices. In practice, the default option of the MUMPS package is such that MC64 is automatically invoked when the structural symmetry is found to be less than 0:5. For SuperLU, zeros on the diagonal and numerical issues must also be considered so that an automatic decision during the analysis phase is more difficult. We finally compare, in Figure 4.1, the time spent by the two solvers during the analysis phase when reordering is based only on AMD (MC64 is not invoked). Since the time spent bbmat ecl32 invextr1 fidapm11 mixtank rma10 wang42610Seconds MUMPS Figure 4.1: Time comparison of the analysis phases of MUMPS and SuperLU. MC64 preprocessing is NOT used and AMD ordering is used. in AMD is very similar in both cases, this gives a good estimation of the cost difference Matrix Solver Preprocess. Total MC64 AMD bbmat MUMPS AMD 4.7 - 3.0 - MC64+AMD 7.2 2.1 3.1 mixtank MUMPS AMD 3.2 - 0.8 twotone MUMPS AMD 12.7 - 8.7 Table 4.4: Influence of permuting large entries onto the diagonal (using MC64) on the time (in seconds) for the analysis phase of MUMPS and SuperLU. of the analysis phase of the two solvers. Note that SuperLU is not currently tied to any specific ordering code and does not take advantage of all the information available from an ordering algorithm. A tighter coupling with an ordering, as is the case with MUMPS and AMD, should reduce the analysis time for SuperLU. However, during the analysis phase of SuperLU, all the asymmetric structures needed for the factorization are computed and the directed acyclic graph (Gilbert and Liu 1993) of the unsymmetric matrix must be built and mapped onto the processors. With MUMPS, the main data structure handled during analysis is the assembly tree which is produced directly as a by-product of the ordering phase. No further data structures are introduced during this phase. Dynamic scheduling will be used during factorization so that only a simple massage of the tree and a partial mapping of the computational tasks onto the processors are performed during analysis. 4.2 Use of orderings to preserve sparsity On matrices for which MC64 is not used we show, in Table 4.5, the impact of the choice of the symmetric permutation on the fill-in and floating-point operations for the factorization. As was observed in Amestoy et al. (1999), the use of nested dissection can significantly improve the performance of MUMPS. We see here that SuperLU will also, although to a lesser extent, benefit from the use of a nested dissection ordering. We examine the influence of the ordering on the performance further in Section 5. We also notice that, for both orderings, SuperLU exploits the asymmetry of the matrix somewhat better than MUMPS (see bbmat with structural symmetry 0:53). We expect the asymmetry of the problem to be better exploited by MUMPS when the approach described in Amestoy and Puglisi (2000) is implemented. Matrix Ordering Solver NZ in LU Flops bbmat AMD MUMPS 46.1 41.5 41.2 34.0 ND MUMPS 35.8 25.7 ND MUMPS 24.8 20.9 ND MUMPS 16.2 8.1 mixtank AMD MUMPS 39.1 64.4 ND MUMPS 19.6 13.2 Table 4.5: Influence of the symmetric sparsity orderings on the fill-in and floating-point operations on the factorization of unsymmetric matrices. (MC64 is not used.) 5 Performance analysis on general matrices 5.1 Performance of the numerical phases In this section, we compare the performance and study the behaviour of the numerical phases (factorization and solve) of the two solvers. For the sake of clarity, we will only report results with the best (in terms of factorization time) sparsity ordering for each approach. When the best ordering for MUMPS is different from that for SuperLU, results with both orderings will be provided. This means that results with both nested dissection and minimum degree orderings are given that illustrate the different sensitivity of the codes to the choice of the ordering. We note that, even when the same ordering is given to each solver, they will not usually perform the same number of operations. In general, SuperLU performs fewer operations than MUMPS because it exploits better the asymmetry of the matrix although the execution time is less for MUMPS because of the Level 3 BLAS effect. Although results are very often matrix dependent, we will try, as much as possible, to identify some general properties of the two solvers. We should point out that the maximum dimension of our unsymmetric test matrices is only 120750 (see Table 2.1). 5.1.1 Study of the factorization phase We show in Table 5.1 the factorization time of both solvers. On the smaller matrices, we only report in Table 5.2 results with up to 64 processors. We observe that MUMPS is usually faster than SuperLU and is significantly so on a small number of processors. We believe there are two reasons. First, MUMPS handles symmetric and more regular data structures better than SuperLU, because MUMPS uses Level 3 BLAS kernels on bigger blocks than those used within SuperLU. As a result, the Megaflop rate of MUMPS on one processor is on average about twice that of the SuperLU factorization. This is also evident in the results on smaller test problems in Table 5.2 and from the results on 3D grid problems in Section 6. Note that, even on the matrix twotone, for which performs three times fewer operations than MUMPS, MUMPS is over 2.5 times faster than SuperLU on four processors. On a small number of processors, we also notice that SuperLU does not always fully benefit from the reduction in the number of operations due to the use of a nested dissection ordering (see bbmat with SuperLU using 4 processors). Furthermore, one should notice that, with matrices that are structurally very asymmetric, SuperLU can be much less scalable than MUMPS. For example, on matrix lhr71c in Table 5.2, speedups of 2.5 and 8.3 are obtained with SuperLU and MUMPS, respectively. This is due to the two parallel limitations of the current SuperLU algorithm described in Section 3.2. First, SuperLU does not fully exploit the parallelism of the elimination dags. Second, the pipelining mechanism does not fully benefit from the sparsity of the factors (a blocked column factorization should be implemented). This also explains why SuperLU does not fully benefit, as in the case for MUMPS, from the better balanced tree generated by a nested dissection ordering. We see that the ordering very significantly influences the performance of the codes (see results with matrices bbmat and ecl32) and, in particular, MUMPS generally outperforms SuperLU, even on a large number of processors, when a nested dissection ordering is used. On the other hand, if we use the minimum degree ordering, SuperLU can be faster than MUMPS on a large number of processors. We also see that, on most of our unsymmetric problems, neither solver provides enough parallelism to benefit from using more than 128 processors. The only exception is matrix ecl32 using the AMD ordering (requiring flops for the factorization), for which only SuperLU continues to decrease the factorization time up to 512 processors. Our lack of other large unsymmetric systems gives us few data points in this regime but one might expect that, independently of the ordering, the 2D distribution used in SuperLU should provide better scalability (and hence eventually better performance) on a large number of processors than the mixed 1D and 2D distribution used in MUMPS. To further analyse the scalability of our solvers, we consider three dimensional regular grid problems in Section 6. Matrix Ord. Solver Number of processors bbmat AMD MUMPS - 45.7 24.0 16.5 13.7 11.9 11.2 9.1 12.6 SuperLU 68.2 23.1 13.3 9.1 6.7 5.7 4.7 6.1 5.8 mixtank ND MUMPS 40.8 13.0 7.8 5.6 4.4 3.9 4.2 4.2 5.4 twotone MC64 MUMPS - 40.3 22.6 18.6 14.7 14.4 14.3 14.0 14.3 Table 5.1: Factorization time (in seconds) of large test matrices on the CRAY T3E. "-" indicates not enough memory. Matrix Ordering Solver Number of processors fidapm11 AMD MUMPS 31.6 11.7 8.4 6.5 5.7 5.7 lhr71c MC64+AMD MUMPS 13.3 4.3 2.9 1.7 1.5 1.6 rma10 AMD MUMPS 8.1 3.1 2.2 2.1 2.0 2.1 wang4 AMD MUMPS 30.6 11.1 7.0 5.2 4.3 3.9 56.3 19.4 13.9 7.9 5.8 5.6 Table 5.2: Factorization time (in seconds) of small test matrices on the CRAY T3E. "-" indicates not enough memory. To better understand the performance differences observed in Tables 5.1 and 5.2 and to identify the main characteristics of our solvers we show, in Table 5.3, the average communication volume. The speed of communication can depend very much on the number and the size of the messages and we also indicate the maximum size of the messages and the average number of messages. To overlap communication by computation, MUMPS uses fully asynchronous communications (during both sends and receives). The use of non-blocking sends during the more synchronous scheduled approach used by SuperLU also enables overlapping between communication and computation. Matrix Ord Solver Number of processors Max Vol. #Mess Max Vol. #Mess Max Vol. #Mess bbmat AMD MUMPS 4.9 44 3240 3.3 63 1700 2.9 20 2257 ND MUMPS 2.2 7 2214 2.8 43 1441 1.5 48 3228 fidapm11 AMD MUMPS 2.5 28 3000 2.4 22 1471 2.4 6 1323 mixtank ND MUMPS 3.5 twotone MC64 MUMPS 8.8 61 5076 2.9 139 4144 2.1 Table 5.3: Maximum size of the messages (Max in Mbytes), average volume of communication (Vol. in Mbytes) and number of messages per processor (#Mess) for large matrices during factorization. From the results in Table 5.3, it is difficult to make any definitive comment on the average volume of communication. Overall it is broadly comparable with sometimes MUMPS and sometimes SuperLU having lower volume, occasionally by a significant amount. However, although the average volume of messages with 64 processors can be comparable with both solvers, there is between one and two orders of magnitude difference in the average number of messages and therefore in the average size of the messages. This is due to the much larger number of messages involved in a fan-out approach (SuperLU) compared to a multifrontal approach (MUMPS). Note that, with MUMPS, the number of messages includes the messages (one integer) required by the dynamic scheduling algorithm to update the load on the processes. The average volume of communication per processor of each solver depends very much on the number of processors. While, with SuperLU, increasing the number of processors will generally decrease the communication volume per processor it is not always the case with MUMPS. Note that adding some global information to the local dynamic scheduling algorithm of MUMPS will help to increase the granularity of the level 2 node subtasks without losing parallelism (see Section 3.1) and thus can result in a decrease in the average volume of communication on a large number of processors. 5.1.2 Study of the solve phase We already discussed in Section 4.1 the difference in the numerical behaviour of the two solvers, showing that, in general, SuperLU will involve more steps of iterative refinement than MUMPS to obtain the same accuracy in the solution. In this section, we focus on the time spent to obtain the solution. We apply enough steps of iterative refinement to ensure that the componentwise relative backward error (Berr) is less than p \Gamma8 . Each step of iterative refinement involves not only a forward and a backward solve but also a matrix-vector product with the original matrix. With MUMPS, the user can provide the input matrix in a very general distributed format (Amestoy et al. 1999). This functionality was used to parallelize the matrix-vector products. With SuperLU, the parallelization of the matrix-vector product was easier because the input matrix is duplicated on all the processors. In Table 5.4, we report both the time to perform one solution step (using the factorized matrix to solve when necessary (Berr greater than p ") the time to improve the solution using iterative refinement (lines with "+ IR"). With SuperLU, except on ecl32 and mixtank which did not require any iterative refinement, one step of iterative refinement was required and was always enough to reduce the backward error to ". With MUMPS, iterative refinement was only required on the matrix invextr1 and the backward error was already so close to " (on one processor that on 4 and 8 processors no step of iterative refinement was required (Berr for the initial solution was already equal to In this case, the time reported in the row "+ IR" corresponds to the time to perform the computation of the backward error. We first observe (compare, for example, Tables 5.1 and 5.4) that, on a small number of processors (less than 8), the solve phase is almost two orders of magnitude less costly than the factorization. On a large number of processors, because our solve phases are relatively less scalable than the factorization phases, the difference drops to one order of magnitude. On applications for which a large number of solves might be required per factorization this could become critical for the performance and might have to be addressed in the future. We show solution times for our smaller matrices in Table 5.5 where we have not run iterative refinement. The performance reported in Tables 5.4 and 5.5 would appear to suggest that the regularity in the structure of the matrix factors generated by the factorization phase of MUMPS is responsible for a faster solve phase than that of SuperLU for up to 256 processors. On 512 processors, the solve phase of SuperLU is sometimes faster than that of MUMPS although in all cases the fastest solve time is recorded by MUMPS usually on a fewer number of processors. The cost of iterative refinement can significantly increase the cost of obtaining a solution. With SuperLU, because of static pivoting, it is more likely that iterative refinement will be required to obtain an accurate solution on numerically difficult matrices (see bbmat, and twotone). With MUMPS, the use of partial pivoting during the factorization will reduce the number of matrices for which iterative refinement is required. (In our set, only invextr1 requires iterative refinement.) For both solvers, the use of MC64 to preprocess the matrix can also be considered to reduce the number of steps of iterative refinement and even avoid the need to use it in some cases (see Section 4.1). Matrix Order. Solver Number of processors bbmat AMD MUMPS - 0.53 0.38 0.31 0.32 0.32 0.36 0.40 0.56 twotone MC64 MUMPS - 1.03 0.92 0.97 0.98 0.98 1.03 1.13 1.41 Table 5.4: Solve time (in seconds) for large matrices on the CRAY T3E. " shows the time spent improving the initial solution using iterative refinement. "-" indicates not enough memory. Matrix Ord. Solver Number of processors Table 5.5: Solve time (in seconds) for small matrices on the CRAY T3E. 5.2 Memory usage In this section, we study the memory used during factorization as a function of both the solver used and the number of processors, see Table 5.6. We want first to point out that, because of the dynamic scheduling approach and the threshold pivoting used in MUMPS, the analysis phase cannot fully predict the space that will be required on each processor and an upper bound is therefore used for the memory allocation. With the static task mapping approach used in SuperLU, the memory used can be predicted during the analysis phase. In this section, we only compare the memory actually used by the solvers during the factorization phase. This includes reals, integers and communication buffers. Storage for the initial matrix is, however, not included but we have seen, in Amestoy et al. (1999), that the input matrix can also be provided in a general distributed format and can be handled very efficiently by the solver. This option is available in MUMPS. In SuperLU the initial matrix is currently duplicated on all processors 7 . Matrix Ordering Solver Number of processors Avg. Max. Avg. Max. Avg. Max. bbmat AMD MUMPS 147 176 52 ND MUMPS 114 118 44 53 28 35 43 44 ND MUMPS 132 139 39 44 25 28 28 17 22 mixtank ND MUMPS 84 87 29 31 19 21 twotone MC64 MUMPS 167 180 57 67 42 Table Memory used during factorization (in Megabytes, per processor). We notice, in Table 5.6, a significant reduction in the memory required when increasing the number of processors. We also see that, in general, SuperLU usually requires less memory than MUMPS although this is less apparent when many processors are used showing the better memory scalability of MUMPS. One can observe that there is little difference 7 For MUMPS, note that the storage reported still includes another internal copy of the initial matrix in a distributed arrowhead form, necessary for the assembly operations during the multifrontal algorithm. between the average and maximum memory usage showing both algorithms are well balanced, with SuperLU the better of the two. Note that memory scalability can be critical on globally addressable platforms where parallelism increases the total memory used. On purely distributed machines such as the T3E, the main factor remains the memory used per processor which should allow large problems to be solved when enough processors are available. 6 Performance analysis on 3-D grid problems To further analyse and understand the scalability of our solvers, we report in this section on results obtained for the 11-point discretization of the Laplacian operator on three-dimensional (NX, NY, NZ) grid problems. We consider a set of 3D cubic (NX=NY=NZ) and rectangular (NX, NX/4, NX/8) grids on which a nested dissection ordering is used. The size of the grids used, the number of operations and the timings are reported in Table 6.1. When increasing the number of processors, we have tried as much as possible to maintain a constant number of operations per processor while keeping as much as possible the same shape of grids. It was not possible to satisfy all these constraints, thus the number of operations per processor is not completely constant. Nprocs Grid size LDL T factorization LU factorization flops time flops time flops time Cubic grids (nested dissection) 4 36 13.4 19.9 26.8 28.1 26.8 53.3 Rectangular grids (nested dissection) 128 208 52 26 243.1 27.4 485.8 53.6 485.6 60.7 Table 6.1: Factorization time (in seconds) on Cray T3E. LU factorization is performed for MUMPS-UNS and SuperLU, LDL T for MUMPS-SYM. Since all our test matrices are symmetric, we can use MUMPS to compute either an LDL T factorization, referred to as MUMPS-SYM, or an LU factorization, referred to as MUMPS-UNS. will compute an LU factorization. Note that, for a given matrix, the unsymmetric solvers (SuperLU and MUMPS-UNS) perform roughly twice as many operations as MUMPS-SYM. To overcome the problem of the number of operations per processor being non-constant, we first report in Figures 6.1 and 6.2 the Megaflop rate per processor for our three approaches on cubic and rectangular grids, respectively. In our context, the Megaflop rate is meaningful because on those grid problems the number of operations is almost identical for MUMPS-UNS and SuperLU (see Table 6.1), thus it corresponds to the absolute performance of the approach used for a given problem. We first notice that on up to 8 processors, and independently of the grid shape, MUMPS-UNS is about twice as fast as SuperLU and also has a much higher Megaflop rate than MUMPS-SYM. On 128 processors on both rectangular and cubic grids, all three solvers have similar Megaflop rates per processor. In Figures 6.3 and 6.4, we show the parallel efficiency on cubic and rectangular grids respectively. The efficiency of a solver on p processors is computed as the ratio of its Megaflop rate per processor on p processors over its Megaflop rate on 1 processor. In terms of efficiency, SuperLU is generally more efficient on cubic grids than MUMPS-UNS even on a relatively small number of processors. MUMPS-SYM is relatively more efficient than MUMPS-UNS and the MUMPS-SYM efficiency is very comparable to that of SuperLU. On a large number of processors SuperLU is significantly more efficient than MUMPS-UNS. The peak ratio between the methods is reached on cubic grids (128 processors) for which SuperLU is about three and two times more efficient than MUMPS-UNS and MUMPS-SYM, respectively. Finally, we report in Table 6.2 a quantitative evaluation of the overhead due to parallelism on cubic grids, using the analysis tool vampir (Nagel et al. 1996). In the rows "computation", we report the percentage of the time spent doing numerical factorization. MPI calls and idle time due to communications or synchronization are reported in rows "overhead" of the table. Nprocs Grid size MUMPS-SYM MUMPS-UNS SuperLU(NX=36) computation 69% 76% 87% overhead 31% 24% 13%(NX=46) computation 67% 69% 75% overhead 33% 31% 25%(NX=57) computation 50% 36% 56% overhead 50% 64% 44% Table 6.2: Percentage of the factorization time (cubic grids) spent in computation and in overhead due to communication and synchronization. Table 6.2 shows that SuperLU has less overhead than either version of MUMPS. We also observe a better parallel behaviour of MUMPS-SYM with respect to MUMPS-UNS, as analysed in Processors rate MUMPS-SYM MUMPS-UNS Figure 6.1: Megaflop rate per processor (cubic grids, nested dissection). Processors rate MUMPS-SYM MUMPS-UNS Figure 6.2: Megaflop rate per processor (rectangular grids, nested dissection). 0.40.81.2 Processors Efficiency MUMPS-SYM MUMPS-UNS Figure 6.3: Parallel efficiency (cubic grids, nested dissection). 1280.20.61Processors Efficiency MUMPS-SYM MUMPS-UNS Figure 6.4: Parallel efficiency (rectangular grids, nested dissection). Amestoy et al. (2000), which is mainly due to the fact that node level parallelism provides relatively more parallelism in a symmetric context. 7 Concluding remarks In this paper, we have presented a detailed analysis and comparison of two state-of-the- art parallel sparse direct solvers-a multifrontal solver MUMPS and a supernodal solver SuperLU. Our analysis is based on experiments using a massively parallel distributed-memory machine-the Cray T3E, and a dozen matrices from different applications. Our analysis addresses the efficiency of the solvers in many respects, including the role of preordering steps and their costs, the accuracy of the solution, sparsity preservation, the total memory required, the amount of interprocessor communication, the times for factorization and triangular solves, and scalability. We found that both solvers have strengths and weaknesses. We summarize our observations as follows. ffl Both solvers can benefit from a numerical preordering scheme implemented in MC64, although SuperLU benefits to a greater extent than MUMPS. For MUMPS, this helps reduce the number of off-diagonal pivots and the number of delayed pivots. For SuperLU, this may reduce the need for small diagonal perturbations and the number of iterative refinements. However, since this permutation is asymmetric, it may destroy the structural symmetry of the original matrix, and cause more fill-in and operations. This tends to introduce a greater performance penalty for MUMPS than for SuperLU although recent work by Amestoy and Puglisi (2000) might affect this conclusion. This is why by default, MUMPS does not use MC64 on fairly symmetric matrices. ffl MUMPS usually provides a better initial solution; this is due to the effect of dynamic versus static pivoting. With one step of iterative refinement, SuperLU usually obtains a solution with about the same level of accuracy. ffl Both solvers can accept as input any fill-in reducing ordering, which is applied symmetrically to both the rows and columns. MUMPS performs better with nested dissection than minimum degree, because it can exploit the better tree parallelism provided by a nested dissection ordering, whereas SuperLU does not exploit this level of parallelism and its parallel efficiency is less sensitive to different orderings. ffl Given the same ordering, SuperLU preserves the sparsity and the asymmetry of the L and U factors better. SuperLU usually requires less memory than MUMPS, and more so with smaller numbers of processors. On 64 processors, MUMPS requires 25-30% more memory on average. ffl Although the total volume of communication is comparable for both solvers. MUMPS requires many fewer messages, especially with large numbers of processors. The difference can be up to two orders of magnitude. This is partly intrinsic to the algorithms (multifrontal versus fan-out), and partly due to the 1D (MUMPS) versus 2D (SuperLU) matrix partitioning. ffl MUMPS is usually faster in both factorization and solve phases. The speed penalty for partly comes from the code complexity in order to preserve the irregular sparsity pattern, and is partly due to more communication messages. With more processors, SuperLU shows better scalability, because its 2D partitioning scheme does a better job in keeping all the processors busy despite the fact that it introduces more messages. As we said in the introduction, we started this exercise with the intention of comparing a wider range of sparse codes. However, as we have demonstrated in the preceding sections, the task of conducting such a comparison is very complex. We do feel though that the experience we have gained in this task will be useful in extending the comparisons in the future. In the following tables, we summarize the major characteristics of the parallel sparse direct codes of which we are aware. A clear description of the terms used in the tables is given by Heath, Ng and Peyton (1991). Code Technique Scope Availability Reference CAPSS Multifrontal SPD www.netlib.org/scalapack (Heath and Raghavan 1997) MUMPS Multifrontal SYM/UNS www.enseeiht.fr/apo/MUMPS (Amestoy et al. 1999) PaStiX Fan-in SPD see caption x (Henon et al. 1999) PSPASES Multifrontal SPD www.cs.umn.edu/-mjoshi/pspases (Gupta, Karypis and Kumar 1997) SPOOLES Fan-in SYM/UNS www.netlib.org/linalg/spooles (Ashcraft and Grimes 1999) SuperLU Fan-out UNS www.nersc.gov/-xiaoye/SuperLU (Li and Demmel 1999) S+ Fan-out y UNS www.cs.ucsb.edu/research/S+ (Fu, Jiao and Yang 1998) WSMP z Multifrontal SYM IBM product (Gupta 2000) Table 7.1: Distributed memory codes. x www.dept-info.labri.u-bordeaux.fr/-ramet/pastix y Uses QR storage to statically accommodate any LU fill-in z Only object code for IBM is available. No numerical pivoting performed. Code Technique Scope Availability Reference GSPAR Interpretative UNS Grund (Borchardt, Grund and Horn 1997) Multifrontal UNS www.cse.clrc.ac.uk/Activity/HSL (Amestoy and Duff 1993) Multifrontal QR RECT www.cse.clrc.ac.uk/Activity/HSL (Amestoy, Duff and Puglisi 1996b) PanelLLT Left-looking SPD Ng (Ng and Peyton 1993) PARDISO Left-right looking UNS Schenk (Schenk, G-artner and Fichtner 2000) PSLDLT y Left-looking SPD SGI product (Rothberg 1994) PSLDU y Left-looking UNS SGI product (Rothberg 1994) Left-looking UNS www.nersc.gov/-xiaoye/SuperLU (Demmel et al. 1999) Table 7.2: Shared memory codes y Only object code for SGI is available Acknowledgments We want to thank James Demmel, Jacko Koster and Rich Vuduc for very helpful discussions. We are grateful to Chiara Puglisi for her comments on an early version of this article and her help with the presentation. We also want to thank John Reid for his comments on the first version of this paper. --R An unsymmetrized multifrontal LU factorization A fully asynchronous multifrontal solver using distributed dynamic scheduling SPOOLES: An object-oriented sparse matrix library Parallel numerical methods for large systems of differential-algebraic equations in industrial applications On algorithms for permuting large entries to the diagonal of a sparse matrix To appear in SIAM Journal on Matrix Analysis and Applications. WSMP: Watson Sparse Matrix Package Part I - direct solution of symmetric sparse systems Version 1.0 http://www. A mapping and scheduling algorithm for parallel sparse fan-in numerical factorization http://www. Making sparse Gaussian elimination scalable by static pivoting A scalable sparse direct solver using static pivoting Hybridizing nested dissection and halo approximate minimum degree for efficient sparse matrix ordering Efficient sparse Cholesky factorization on distributed-memory multiprocessors --TR Parallel algorithms for sparse linear systems Elimination structures for unsymmetric sparse <italic>LU</italic> factors A supernodal Cholesky factorization algorithm for shared-memory multiprocessors Modification of the minimum-degree algorithm by multiple elimination An Approximate Minimum Degree Ordering Algorithm Highly Scalable Parallel Algorithms for Sparse Matrix Factorization Efficient Sparse LU Factorization with Partial Pivoting on Distributed Memory Architectures A Supernodal Approach to Sparse Partial Pivoting The Design and Use of Algorithms for Permuting Large Entries to the Diagonal of Sparse Matrices An Asynchronous Parallel Supernodal Algorithm for Sparse Gaussian Elimination Making sparse Gaussian elimination scalable by static pivoting Preconditioning Highly Indefinite and Nonsymmetric Matrices On Algorithms For Permuting Large Entries to the Diagonal of a Sparse Matrix A Fully Asynchronous Multifrontal Solver Using Distributed Dynamic Scheduling A Mapping and Scheduling Algorithm for Parallel Sparse Fan-In Numerical Factorization --CTR Laura Grigori , Xiaoye S. Li, A new scheduling algorithm for parallel sparse LU factorization with static pivoting, Proceedings of the 2002 ACM/IEEE conference on Supercomputing, p.1-18, November 16, 2002, Baltimore, Maryland Mark Baertschy , Xiaoye Li, Solution of a three-body problem in quantum mechanics using sparse linear algebra on parallel computers, Proceedings of the 2001 ACM/IEEE conference on Supercomputing (CDROM), p.47-47, November 10-16, 2001, Denver, Colorado Olaf Schenk , Klaus Grtner, Solving unsymmetric sparse systems of linear equations with PARDISO, Future Generation Computer Systems, v.20 n.3, p.475-487, April 2004 Patrick R. Amestoy , Iain S. Duff , Jean-Yves L'Excellent , Xiaoye S. Li, Impact of the implementation of MPI point-to-point communications on the performance of two general sparse solvers, Parallel Computing, v.29 n.7, p.833-849, July Xiaoye S. Li , James W. Demmel, SuperLU_DIST: A scalable distributed-memory sparse direct solver for unsymmetric linear systems, ACM Transactions on Mathematical Software (TOMS), v.29 n.2, p.110-140, June Anshul Gupta, Recent advances in direct methods for solving unsymmetric sparse systems of linear equations, ACM Transactions on Mathematical Software (TOMS), v.28 n.3, p.301-324, September 2002 Timothy A. Davis, A column pre-ordering strategy for the unsymmetric-pattern multifrontal method, ACM Transactions on Mathematical Software (TOMS), v.30 n.2, p.165-195, June 2004 Patrick R. Amestoy , Iain S. Duff , Stphane Pralet , Christof Vmel, Adapting a parallel sparse direct solver to architectures with clusters of SMPs, Parallel Computing, v.29 n.11-12, p.1645-1668, November/December

sparse direct solvers;multifrontal and supernodal factorizations;parallelism;distributed-memory computers

504225

The look of the link - concepts for the user interface of extended hyperlinks.

The design of hypertext systems has been subject to intense research. Apparently, one topic was mostly neglected: how to visualize and interact with link markers. This paper presents an overview of pragmatic historical approaches, and discusses problems evolving from sophisticated hypertext linking features. Blending the potential of an XLink-enhanced Web with old ideas and recent GUI techniques, a vision for browser link interfaces of the future is being developed. We hope to stimulate the development of a standard for hyperlink marker interfaces, which is easy-to-use, feasible for extended linking features, and more consistent than current approaches.

Figure 1: HyperTIES used cyan to highlight links [from Shneiderman & Kearsley 1989] HyperTIES (Fig. 1) avoided this problem by using a distinct text color for link markers, similar to Hyper-G's browser Harmony (Fig. 5) which utilized background colors [Shneiderman & Kearsley 1989; Andrews 1996]. This has the advantage that the typeface and style of the text can be chosen freely. Figure 2: Intermedia's link marker arrows IRIS' "Intermedia" marked hyperlinks with little arrow icons between lines of text (Fig. 2), showing the start of the link span but not its endpoint [Yankelovich et al. 1988]. "Emacs Info" brackets link markers with asterisk symbols. These methods occupy extra screen space and change the layout of the text by inserting additional elements. Figure 3: Link markup by the Neptune Document Viewer [from Delisle & Schwartz 1987] Bernstein's Hypergate, the Neptune hypertext system (Fig. and some early Web browsers UdiWWW (Fig. boxes around the link marker text. This works with the layout but is quite obtrusive and distracting. An improvement to this technique could be found in HyperCard and Storyspace. They drew the boxes only when the reader pressed particular keys, making links evident on request thus keeping the text pristine the rest of the time. Figure 4: Anchor highlighting in UdiWWW In fact, this was the consensus solution after the Hypertext '87 demo sessions, when hypertext designers could first compare all existing systems side by side [Bernstein 1996]. Hiding links has many advantages: pages stay uncluttered, text stays readable, and page design is less influenced by link appearance. However, when using "links on demand" the interface designer must be aware of a potential dis- advantage: since links are not always visible, possibly distracting mode switches have to be applied. Therefore this link trigger has to be seamlessly integrated into the interface. A less desirable variant of this method was used by Symbolics Document Examiner [Walker 1987]: link markers were hidden so well that they were only highlighted when the mouse passed over them. This forces a "hunt and peck" search for active regions. In Microcosm and DLS the user can query the system for invisible links by marking a word or some text and then issue a search for matching links [Carr et al. 1995]. Though anchors are not marked, this model is well applicable to generic links, given that most words can be selected as anchors. Ignoring previous experiences, Mosaic returned to colored (blue) and underlined text to indicate link markers. This is not the optimal solution as it emphasizes the link marker text permanently. The underlined markers stick out from the surrounding text and decrease its readability as underlining interferes with descenders, letters that drop below the line like p, q and j. The blue color is also an imperfect choice, since especially elderly people have problems to perceive it; the human eye is less sensitive to the color blue than to other colors [Lythgoe, 1979]. The reasons for Mosaic's link marker appearance were obviously of technical nature: it was quite simple to implement, and at that time most computers had only 16 colors or a black and white display. Blue was the darkest of the available colors, closest to the black text; for monochrome displays, the text was underlined [NCSA 1996]. The pervasiveness of the Web has led us to accept underlined colored text as the de facto marker standard for all mainstream hypertext systems today. It can be found in help systems on various platforms and even in operating system components like the Microsoft Windows Explorer?. Though the use of a standard is desirable from a consistency point of view, user interfaces have otherwise improved, and we are still wedged with this historical hack. More recent technologies like Cascading Style Sheets [Bos et al. 1998] allow to define the appearance of text links in various ways. Also, the look of links can be configured somehow in current browsers, however, the standard setting is still blue and underlined, and links on demand are still not possible. Even worse, for the visualization of link maps in graphics no standard method can be found nor is implemented in Web browsers. This shows how important it is to design user interfaces well, considering earlier experiences, and that even an interface with obvious design mistakes can become a standard. ENTER XML LINKING While the concept of links made it possible to create non-sequential texts, rich hypertext systems offer a more sophisticated functionality, i.e., they support comprehensive structuring, editing and navigation features [Yankelovich et al. 1988; Bieber et al. 1997]. However, to date the Web itself only supports embedded one-way links. This limitation made the authoring of Web pages and the development of Web servers and browsers simple, enabling the Web to grow extremely fast. On the other hand, all approaches that try to integrate some extended functionality into the Web have to utilize workarounds to overcome the weaknesses of this simple approach. It was also hard to compete with the well-known standard browsers, as these often were better suited to display existing stylishly designed Web pages. With the advent of XML linking, the Web will be able to offer many of the features hypertext experts are missing today [Vitali & Bieber 1999]. Standard browsers already migrate to XML and XSLT [Clark 1999] and hopefully linking will become widely available soon. The linking potential of XML linking1 is based on two key standards which are necessary to create and describe links and link anchors: XLink itself defines links as relation between resources or portions thereof. Syntactically, a link consists of an arbitrary number of resources (local or remote) and arcs. A resource is any addressable unit of information or service, while arcs create directed connections between two resources each [DeRose, Maler, Orchard 2000]. XPointer allows to address different kinds of spans in XML documents [see DeRose, Maler, Daniel 2001]. They can vary from points to complex regions and can even be distributed over the document, e.g. a XPointer could be used to address the string "Goethe" in all citations in an XML file: xpointer(string-range(//cite, "Goethe")). Two types of semantic attributes are defined: machine-readable information is stored in the role and arcrole attributes and the corresponding human-readable textual information is kept in title attributes. This type information can be specified for the link as a whole, each endpoint of a link and for every arc. linking also presents a solution for linking into other authors' read-only material, by addressing parts of the documents' structure. There is no need for tailored target anchors, which are embedded in the target document, any more. The importance of this can be seen from printed media, i.e., referring to distinct pages or paragraphs. linking may not only be used for hypertext links, but for any kind of application describing relations, associations or compositions of XML documents. However, this paper focuses on its use for hypertext. To summarize, XML linking will allow a multitude of new hyperlink features, among them: ? structure and contents may be separated; links may be bi-directional; links may be typed; links may have multiple endpoints; anchors may be complex or overlap. While the syntax of XLink has been elaborately defined, most presentation and behavioral aspects of links have been deliberately excluded from the model. Only few hints on how to implement and interact with these features can be found in the XLink definition, and few ideas exist to enable the user to cope with this extended functionality. There is a vague notion of displaying title attributes to enrich link anchors semantically and computing role attributes to realize typed links [DeRose, Maler, Orchard 2000]. Pop-up windows are suggested for links with multiple endpoints [DeRose 1989; Bosak & Bray 1999]. Though this might suggest that not much has changed, for the user interface, these new features create a lot of new questions for link visualization. The chances seem to be good that XLink will succeed. The changes in browser capabilities since Mosaic suggest that more extensive hypertext features will eventually be ac- cepted: Forms, JavaScript, Java Applets and Flash animations are now widely used. Some extended link functionality is already being simulated with DHTML, showing pop-up menus with multiple destinations or extra information. VISUALIZING EXTENDED HYPERTEXT FEATURES The two well-known hypertext models Dexter [Halasz 1990] and HyTime [Derose & Durand 1994] had features almost matching, and sometimes exceeding XLink, but no system ever fully implemented them [Gr?nb?k & Trigg 1999, p. 42]. Nonetheless, many systems existed that were far ahead of their time and offered functionality that is not available in the Web today. This enables us to find ideas and detect problems by looking at all these systems and how they implemented and visualized hyperlinks. We map the user interfaces of these programs to the linking features that are made available with the introduction of XLink. We can thus discuss what is needed by the user to profit from the extended functionality Separation of Structure and Content Several former Open Hypermedia Systems like Microcosm/ DLS [Carr et al. 1995] or the Devise Hypermedia System permitted to store links separately from documents in dedicated linkbases. Likewise, XLink will allow the separation of structure and content for the Web. The external storage of links permits multiple linkbases to be used for single Web pages. These links may originate from the original author but also from other authors without write access to the original document, like a single user or members of a group2. The use of several different linkbases can result in an unintentionally great number of links3. Therefore, the user must be enabled to select the employed linkbases. An example for such a method can be found in ?Third Voice?, a browser plug-in that adds annotations to Web pages4 (Fig. 5). A part of its functionality is a service that adds links from an external linkbase to keywords. These annotation link markers are distinguished by orange underlines. Third Voice offers an extra tool bar in the head of the browser where the presentation of its links can easily be toggled. Figure 5: Third Voice adds additional links to an existing Web page and offers a choice of targets. Microcosms/DLS already permitted the use of several linkbases. It offered a configuration screen to select the utilized link database from a given set [Carr et al. 1996]. Unfortunately this menu was not directly integrated into the browser interface and the addition of new linkbases was quite complicated. An XLink browser will also need the potential to find new linkbases5 and add them to a personal list. So far, no standard means has been established to do so. Bi-directional Links From the technical point of view bi-directional links help to links consistent and to avoid broken links. From the usability viewpoint they also permit to follow links backwards as opposed to the uni-directional ?goto? links of the Web. A user could use this feature to find e.g. more recent information which is referring to an old but valuable document. To benefit from bi-directional linking, the user interface has to support the backward traversal of links. Most hyper-text systems with bi-directional links like Sepia, MacWeb or Hyper-G offered a "local map", showing nodes and connecting links. This visualizes the topology and permits the user to select source objects directly on the map. These additional links can be used to annotate and supplement the existing information with other information of personal importance. 3 These links may also overlap (see following sections). Third Voice is available at http://www.thirdvoice.com 5 Furthermore, the primary linkbase will frequently change when browsing the Web, as it usually will be provided by the server hosting the current document. For the Web, the retrieval of links that refer to the current document poses a serious problem. A prototype Web browser tool described in [Chakrabarti et al. 1999] gathered this information from search engines. They alternatively proposed to extend the HTTP protocol to send backlink information gathered from the referrer URLs in the server log. The prototype offered a list of titles of Web documents that were linking to the current document. Both approaches have their limitations if the number of links is high. Especially graphical maps use a lot of screen space if dozens of nodes and links are displayed. Thus, the number of objects has to be limited, e.g. by filtering the most appropriate ones. Typed Links A link type describes the relationship between source and destination of a link, often derived from semantic categories like "explanation" or "example" [Trigg 1983]. They were introduced to help users to get a better idea of a link target. Streitz et al. list semantic link information as their first principle of useful hypermedia system design [Streitz et al. 1992]. However, typed links are only helpful if the user can distinguish the different types. Tim Berners-Lee's WWW proposal [Berners-Lee 1989] included typed links, and HTML allows Web authors to set the link type attributes rel and rev. Though, this feature is not supported by any current Web browser. Sepia [Streitz et al. 1992] and MacWeb [Nanard & Nanard 1993] displayed the link type in an overview map close to the arrow visualizing the link. Once more, this link information is only available to the user if he considers two areas at the same time: a document view and a link map. He has to join these two information segments cognitively. Other systems use text style to distinguish different link anchor types: the current Microsoft help system displays explanatory pop-up links in green with a dotted underline and uses icons to indicate specific actions as the execution of a program. However, the potentials of Text style are quite limited, and inline icons can be distracting and create problems with the layout. Figure Different mouse pointers utilized by the Guide system. The Guide system utilized different mouse pointers to make link characteristics apparent [Brown 1987]. The pointer changed according to the link type if it hovered over a link (Fig. 6). Since mouse pointers are independent from screen and text layout, this may be an interesting option for Web clients, too. Standard software, like word processors and graphics programs, and also operating systems, commonly employ these differently shaped mouse pointers as it is possible to indicate many different actions in a non-obtrusive, yet immediately visible manner. Multiple endpoints Links with multiple endpoints do not connect only two, but a set of related nodes. Thus different alternative destinations can be provided. When a user initiates the traversal of a link with multiple endpoints, he can be requested to choose between the available options. This solution was preferred by most former hypertext systems. Microcosm and DLS presented a list of generated link targets on an intermediary page as the result of a user query [Hall, Davis, Hutchings 1996; Carr et al. 1995]. Intermedia displayed a dialogue box with a list of link titles. Likewise, the preferred idea for XLink seems to be a pop-up menu [Halsey, Anderson 2000; DeRose 2000]. Though lists of targets are probably the most straightforward approach, they may slow down Web navigation. A user has to make an additional selection from the pop-up list each time he follows a link. Multiple links can also be used to automatically select the most decent destination by applying a filter. Already the father of hypertext Vannevar Bush suggested filters for links. If the user follows a Guided Tour, links of the displayed documents should be hidden [Bieber 97; Bush 45]. could filter links by link attributes and Hyper-G by user rights. It would be even more desirable to filter by semantic criteria like a user's task or profile. Complex Link Anchors Many Web usability guidelines confine the setting and the length of link markers, e.g. Nielsen recommends that link markers should be about 5 words long [Nielsen 2000]. This restriction is a concession to the limited link visualization potentials of current Web browsers, where extended link spans result in hardly readable underlined text regions. Hypertext systems that displayed links only on demand avoided these readability problems. The XML linking standard allows arbitrary complex link anchors. As explained before, it is even possible to create discontinuous anchors, i.e., anchors that consist of several distinct regions. To the user this may appear like multiple anchors that share the same destination, which can be irritating. In Web system evaluations, already links that are displayed in more than one line have been found confusing, as the beginning and end of the anchor were not indicated by the browsers used [Spool et al. 1999]. Consequently, the extent of a link marker should be visuali- zed. This is possible in recent Web browsers: the link marker can be highlighted if the mouse hovers over the link. However, the browser configuration has to be changed or an appropriate CSS must be defined. Overlapping Link Markers Link markers may overlap, either because an author creates two anchors at two intersecting text sections which are related to different destinations6, or because other authors create anchors overlapping with the link spans of the original author. Hardly any current Web user will be familiar with the idea of overlapping link markers as they cannot be found on the Web or any popular hypertext system. Currently, it is not possible to create such constructs in HTML, since there is no way to distinguish different opening and closing anchor tags. This technological problem can easily be solved even with embedded links as Hyper-G's markup language HTF demonstrated. It used link identifiers to associate opening and closing link tags [Maurer 1996]. Nonetheless it is much harder to find a usable solution for the visualization of overlapping link spans. Harmony, Hyper-G's browser, used overlapping colored background boxes to mark the beginning and end of up to six overlapping markers (Fig. 5). But even two overlapping links are hardly readable and this method will finally fail if a larger number of anchors intersects: the increasing number of boxes will shrink to pixel height before they finally disappear. Figure 7: Link Overlap in Harmony. The user must also be able to choose a desired link in the overlapping section. Third-Voice (Fig. 5) displays a pop-up window where the user can pick the link to follow. Harmony lets the user first select an overlapping link by single- left clicks and then follow it by a double-left click [Andrews 1996, p. 54]. Both solutions are not optimal, as the first one needs always two and the second one may even need an uncertain number of clicks to follow a link. The current version of Hyper-G does not support overlapping links any more7. A VISION FOR IMPROVED HYPERTEXT USER INTERFACES The Web, undoubtedly the most successful distributed hypertext system ever, has despite its simplicity already serious usability problems. It must be prevented that this becomes worse when extended linking features are introduced We would like to revive a discussion by presenting ideas of an user interface strategy for extended links. To accomplish this we consolidated experiences of earlier hypertext research with established and innovative GUI techniques to create a consistent vision. These thoughts are widely based 6 Example: the phrase "psycholinguistics department" might be a link to the department home page, while another link explains the meaning of "psycholinguistics". 7 HyperWave Information Server Version 5.5 uses HTML as markup language. Figure 8: Mockup: Outgoing XLinks can overlap and are marked by transparency. Note the marked scrollbar. on the analogy of the hypertext reader as a traveler, introduced in Landow's authoritative "Rhetoric of Hypertext" paper. He divides the interaction with links into two key parts: departure and arrival [Landow 1987]. Landow's ideas were based on his experiences with Inter- media. On the background of later hypertext research and the enriched linking capabilities of XLink a further discrimination is possible. The action of departure can be split into two sub-actions: first, the problem of locating the point of departure (identifying link markers) and second, the problem of getting sufficient information about the destination of the journey (understanding the link relationship). Considering the arrival procedure, the reader must get a reception at the destination to understand the extent and the context of the referenced material. The direction he came from, i.e., the origin of the journey, is the last page he visited and therefore known. Finally, XLink does not only allow for links that connect just two endpoints ? it is also possible to build XLinks that represent whole paths or structures. Thus, XLink at last embodies a standard Web storage format for structural information, e.g. for guided paths or for hierarchical site maps. We will discuss the uses of these hidden links (hid- den in the sense that they are not originating from rendered page content) in a separate section. Point of Departure Current methods of Web authors ? emphasizing text anchors by using color and style and using specially tailored graphics to mark graphical link anchors ? are already so common that they will probably continue to exist when XLink is introduced. However, as illustrated above, these methods do not have the potential to identify extended or externally defined XLinks. Furthermore, no prevalent standard visualization method can be found to identify graphical or image map links. Consequently, new schemata are needed to display supplementary links, e.g. from an external linkbase provided by an XLink service. From the usability point of view, a consistent and uniform technique is desirable, that does not distract from reading and does not interfere with text and graphical layout tech- niques, but enables the user to identify even complex anchors clearly. We think that an appropriate way to accomplish this might be the use of transparent areas overlaid on the hypertext document. Overlays have the advantage to be feasible with text and graphics, indicating active areas directly by masking them. They can be applied also at places where the document author did not plan a link. A possible distraction can be reduced by using soft and light colors for bright background and shady colors for dark background. User tests with more sophisticated transparent user interfaces showed promising results [Harrison 1995; Cox 1998]. An important factor that has to be considered is link den- sity. If the ratio of marker area to unlinked area is high, the distinctive anchor appearance may overwhelm the "normal" text. Since an arbitrary number of "alien" links can relate to an XML Web page, a selection mechanism will have to prevent a phenomenon we would call "link overload", similar to information overload, which could overshadow the interaction potentials of the approaching XLink Web. Therefore, the user interface must provide means to select which links will be put on view. Once more, this calls for links-on-demand display techni- ques. The selection mechanism may be provided by an additional tool bar or window. A "link database browser" could be displayed at the left side of the window like the history list of Microsoft Internet Explorer or the Sidebar of Netscape 6. The tool would not only allow to select new link databases8, it would also permit to enable and disable 8 XLink offers a standard storage mechanism for external links. This permits the construction of hyperbase systems that offer compiled collections of links, e.g. as the result of a query [Gr?nbaek & Trigg 1999, p. 167]. These services could be provided just like today 's search engines or Web catalogues. linkbases, making them appear or disappear. Colors may be used to associate listed linkbases to the anchors on the screen (Fig. 8). When a link starts from an anchor longer than a few words, the overall readability of the text rapidly decreases, at least as long as persistent highlighting is used. As with the introduction of XLink source anchors can become arbitrarily long, this question becomes increasingly important. We suggest a simple method to reduce the impact on read- ability: a narrow bar on the right side of the anchored paragraph. The use of different techniques for short and long anchors is suggested by looking at the use of conventional paper: markup on paper consists of highlighting words by coloring them with a transparent marker and, when longer passages need to be distinguished, marking whole paragraphs by using vertical lines on the page border. Sometimes a title is given to help recognize the underlying concepts of the passage. Markup of this style has the advantage to be apparent but not as distracting as long underlined text. It uses only little screen space and is goes along well with most layouts. Since this simple technique does not show the exact location of link marker start and end, it should be supplemented by a rollover effect. The scrollbar, or optionally a small-scale overview window could be used to show the location of link anchors that are outside the currently visible page section. Using the scrollbar to locate particular areas on long Web pages was already suggested by [Laakso, Laakso Overlapping links of several linkbases could be visualized by transparent overlays in a defined neutral color, like bright gray. If the user clicks on such an area, a transparent pop-up appears, showing a list of the available link titles in the color of their associated database. Moving the mouse over these transparent items will highlight the related link markers in the document. Destinations At first sight the more or less uniform looking links of the Web are not typed. Apart from the marked text, the only preview a user can always get in a Web browser is the destination URI [Spool et al 1999]. Even this scant information is frequently utilized by Web users. Sometimes link titles or alternative descriptions to graphics are provided to hint at the content of the target document. Looking closer the current Web could already provide much richer information. Link targets can differ in type ("mailto:" links, downloads), availability (broken links), size, and connection speed (affecting download time). Further information of semantic nature like title, author and language of the target document or structural hints like indicating out-of-site links could be used to automatically enhance link preview. In a paper on our project HyperScout we already suggested techniques to display such information in pop-ups [Weinreich 2000]. Figure 9: A pop-up menu that renders both XLink-specific, and other automatically gathered information. While XLink's title information can also be straight-forwardly displayed in such a pop-up window, the machine-readable information is provided to compute type information. This can be used to induce alternative traversal behavior, or to get advance information about file types of target documents. It could also be used to filter links according to a specified user profile. If this leads to alternative browser behavior, this must not be hidden from the user. In addition to pop-ups, we suggest the use of different mouse pointers to immediately indicate link actions, comparable to the method of the Guide system. If an XLink offers several destinations, the problem of selection occurs. Pop-up menus with a list of available links are suggested in the XLink definition and some publications [DeRose 2000; Bosak and Bray1999]. They have, however, the disadvantage to require additional user action: the user has first to choose a link, click and then he has to choose a target anchor and click again. We would suggest to use the role attributes to allow filtering, thus displaying only part of the link targets available. In certain cases it might even be desirable that a default destrination is automatically selected when the left mouse button is clicked. Indicated by the mouse pointer, a pop-up appears only on right mouse click, presenting a choice of complementary link targets. Arrival The rhetoric of arrival in the sense of Landow requires that the reader gets the feeling of welcome at the destination document: ?One must employ devices that enforce hyper-text capacity to establish intellectual relations.? [Landow 1987]. Establishing such an intellectual relation requires the user to determine the target of a link and its context. The method of today's Web browsers to present the target is simple: the whole document is shown, or, if a fragment identifier was specified, the browser tries to scroll to the position of the fragment anchor. In fact this is a known usability problem of the Web: as the span of the link target is not visualized the user cannot identify the extent of the destination. If the fragment locator is near the end of a page, the browser often cannot scroll sufficiently down to display the link target at the top of the window. Tim Berners-Lee's first graphical Web browser on the NextStep system and later versions of Mosaic did already highlight the target anchor [Caillau & Ashman 1999], a feature lost in current mainstream browsers. Since we already suggested transparent highlighting areas to identify starting anchors, a different method would be advisable to prevent misconceptions. Figure 10: A mock-up showing the incoming link target in its context and information about co-references. We again suggest a technique long before known to work on paper: lateral marker lines. A narrow bar on the far left side of the window is used to indicate the target span. The chance of confusing incoming and outgoing links is thus (incoming links: left vs. outgoing links: right) kept to a minimum. The Devise Hypertext System utilized a similar technique to indicate target anchors [Gr?nb?k & Trigg 1999, p. 314]. When a more precise visualization of the target becomes necessary (e.g. for tables), an on-demand method may be used: moving the mouse over the marker bar will shade the rest of the document except for the target area; a method already used in Harmony's PostScript viewer [Andrews 1996]. If the target section is larger than the visible window, clicking on the bar will ?pin? the shading and the user may scroll the page. Additionally, the scrollbar may be used to show the extent of the target span compared to the whole page, especially useful if it does not fit into the window. Although a very precise notion of the target anchor can be specified with XML linking, a weakness of XLink emerges: what the standard lacks, is a definition of the link context. Nanard and Nanard argue for a distinction of link anchor (as trigger) and link context (as minimum reading context) at both ends of the link. The link anchor is usually quite short and focused, as the link context is embedding the anchor, enabling the reader to understand the relationship of a link better [Nanard & Nanard 1993; Hardman et al. 1994]. The meaning of a sequence of words can change severely when torn out of its context, i.e., the surrounding sentence, paragraph or chapter. XLink misses an explicit definition of such context spans. This is a serious dis- advantage, that could be easily fixed by an additional attribute for the resource tags. Then, if an anchor is selected, the context should be made visible. It is also sometimes useful to supplement other links pointing to the target anchor, called co-references. These links could have been collected in earlier sessions, retrieved from search engines or compiled by linkbase systems. They can stem from material that was not visited in the course of the current search, and, if followed in the reverse direction, they can provide material related to the current target anchor. We suggest to make the most appropriate co-references of the last navigated link available by right clicking on the lateral marker bar, just as the right click opens a pop-up on link anchors. A double click can be used to open a larger list in the left side of the browser with more references pointing to that document. Finally, we think it would be feasible to apply filtering mechanisms utilizing the arc roles, e.g. to display only links that use the current anchor as an ?example?. The Use of Hidden Links So far we have tried to optimize the visualization of hyper-links with markers in documents. However, XLink can also be used to describe relations without markers, e.g. non-associative structural XLinks. These links can either be supplied by the author of a site or by an external source, e.g. a guided tour or a trail [Hammond & Allinson 1987; Bush 1945]. Automatically generated link overview maps (local map, fisheye view, 3D landscapes) often seem to be more confusing than helpful, when used in large hyperspaces [Utting & Yankelovich 1989]. Because of the immense size and the distribution of the Web, structural information has to be provided for an overview that truly can help to find semantically related content. A special link type introduced by HTML 2.0 (<LINK REL>) to distinguish between structural and associative links [Berners-Lee 1995]. Though this made it possible to separate structure-related and content-related navigation, it is poorly supported by current Web browsers. Only some less widely spread browsers like Lynx and iCab (Fig. 10) support the use of structural links. Thus, so far there are only a handful of Web sites that offer structural links. Yet, this information could often be easily provided, especially for generated Web content or sites created with an authoring tool. Figure 11: iCab's structural link navigation toolbar. To support structural information in XLinks special link roles and arc roles would have to be defined. This, how- ever, would use the role attributes not for semantic but rather for syntactical information. Then again, it would be possible to provide complete structures, e.g. Guided Tours or Site Maps, in a single link, something not possible with the LINK element. As for link markers, a consistent interface is needed for structural navigation: we suggest that XLink-aware browsers should provide an iCab-like toolbar for basic structural navigation. Furthermore a hierarchical view, like Hyper- G's collection tree browser, can be provided on demand. This additional navigation tool should be displayed in the same browser window, e.g. in place of the sidebar. The interface should also provide a standard interface for Guided Tours or other meta-structures, thereby eliminating the need for workarounds. We can also imagine hybrid XLinks which bear structure, and have link markers in the Web page9. This implicit structure could be extracted and displayed in the standard- 9 Such structural links include: links on a homepage pointing into the site, site logos pointing to the homepage, arrows for next and last pages, etc. ized user interface. The original embedded links should not be hidden: the user can thus either use the consistent standard interface (without having to search for navigation or follow the rendered structural links (without having to leave the page context). CONCLUSION Usability has become a key factor for the success of soft- ware. Despite the intensive research on hypertext systems, no standard hyperlink user interface has been agreed on. We are thus bound to the de-facto standard of the Web, a design with many inherent weaknesses that does not agree with extended linking features. Experiences from software engineering have shown how to do better: the initial design of a system has to include its user interface as well as its functionality [Nielsen 1993]. The representation of data is only of secondary importance. This demonstrates the need of reconsidering currently developed standards: the XLink standard does hardly mention the user interface. The same lack of consideration of link interfaces is apparent in other W3C activities: Neither HTML nor its present descendants nor other standards like SMIL or the Semantic Web Initiative mention design issues regarding the user interface. In this paper we try to stimulate a discussion on the visualization of and the interaction with extended hyperlink features. We believe that this is necessary to prevent an impairment of Web usability when new linking features are introduced. Experiences from historical systems can help to avoid mistakes and to provide solutions that are still topical. This paper presents problems and solutions for the presentation of and the interaction with extended hyperlink features. Though we are aware that the developed vision can still be enhanced, we gathered well-tried methods to create a consistent and easy-to-use interface. In this process, design issues for XLink arose: we found some open issues, i.e., the missing definition of contexts, default arcs, syntax attributes or attributes needed to carry preview information (like the size of a target document). Some issues were completely left out, like the distribution of links via linkbases or an exact specification for the use of the semantic attributes. Nonetheless, XLink can be used even today: when XLinks are used on Web servers, the centralized storage makes link management much easier [Markos 2000]. Using XSL Transformations, XML or XHTML documents and XLink linkbases can be converted to HTML and be accessed by conventional browsers ? right now. In the long run, however, this functionality should be moved to the client ? only then the browser will be able to exploit the full power of XLink. The success of XLink or a similar standard will eventually depend on two factors: decent tools for authors and readers. --R Information Management: A Proposal. Hypertext Markup Language - 2.0 HypertextNow: Showing Links Forth Generation Hypermedia: Some Missing Links for the World Wide Web. Cascading Style Sheets Level 2 Specification. XML and the Second-Generation Web Turning Products: The Guide System As We May Think Hypertext in the Web The Distributed Link Service: A Tool For Publishers Web Links as User Artefacts Surfing the Web Backwards XSL Transformations (XSLT) Version 1.0 XML Path Language (XPath) Version 1.0 Hypertext: An Introduction and Sur- The Usability of Transparent Overview Layers. Neptune: A Hypertext System for CAD Applications XML Pointer Language (XPointer) Version 1.0 Linking Language (XLink) Making Hyper-media Work: A User's Guide to HyTime The Dexter Hypertext Reference XLink and open hypermedia systems: a preliminary investigation The Travel Metaphor as Design Principle and Training Aid for Navigating around Complex Systems. Adding Time and Context to the Dexter Model An Experimental Evaluation of Transparent User Interface Tools and Information Content Relationally Encoded Links and 46. The Ecology of Vision --TR Neptune: a hypertext system for CAD applications Hypertext: an introduction and survey The travel metaphor as design principle and training aid for navigating around complex systems Hypertext hands-onMYAMPERSANDmdash;an introduction to a new way of organizing and accessing information Context and orientation in hypermedia networks SEPIA Should anchors be typed too? The Amsterdam hypermedia model An experimental evaluation of transparent user interface tools and information content As we may think Fourth generation hypermedia Fluid links for informed and incremental link transitions The usability of transparent overview layers Web site usability Surfing the Web backwards Turning ideas into products Document Examiner Relationally encoded links and the rhetoric of hypertext XLink and open hypermedia systems Hypermedia on the Web Hypertext in the Web MYAMPERSANDmdash; a history Concepts for improved visualization of Web link attributes Designing Web Usability Hyperwave Usability Engineering Rethinking Hypermedia Making Hypermedia Work A network-based approach to text handling for the on-line scientific community --CTR Duncan Martin , Helen Ashman, Goate: XLink and beyond, Proceedings of the thirteenth ACM conference on Hypertext and hypermedia, June 11-15, 2002, College Park, Maryland, USA Duncan Martin , Mark Truran , Helen Ashman, The end-point is not enough, Proceedings of the fifteenth ACM conference on Hypertext and hypermedia, August 09-13, 2004, Santa Cruz, CA, USA Paolo Ciancarini , Federico Folli , Davide Rossi , Fabio Vitali, XLinkProxy: external linkbases with XLink, Proceedings of the 2002 ACM symposium on Document engineering, November 08-09, 2002, McLean, Virginia, USA Hartmut Obendorf , Harald Weinreich, Comparing link marker visualization techniques: changes in reading behavior, Proceedings of the 12th international conference on World Wide Web, May 20-24, 2003, Budapest, Hungary Delfina Malandrino , Vittorio Scarano, Tackling web dynamics by programmable proxies, Computer Networks: The International Journal of Computer and Telecommunications Networking, v.50 n.10, p.1564-1580, 14 July 2006 Niels Olof Bouvin, Augmenting the web through open hypermedia, The New Review of Hypermedia and Multimedia, v.8 n.1, p.3-25, January Whitehead, As we do write: hyper-terms for hypertext, ACM SIGWEB Newsletter, v.9 n.2-3, p.8-18, June-October 2000

XLink;user interface;distributed hypertext;link marker

504364

Class-is-type is inadequate for object reuse.

It is well known that class and type are two different concepts in object-oriented programming languages (OOPLs). However, in many popular OOPLs, classes are used as types. In this paper, we argue that the class-is-type principle is a major obstacle to software reuse, especially to object reuse. The concepts of the basic entities, i.e., objects, object classes, object types, and object kinds, in the type hierarchy of OOPLs are revisited. The notion of object reuse is defined and elaborated. In addition, we show that parameterized types and generic functions are better served by used kind-bounded quantification than universal quantification and other mechanisms.

Introduction Object-oriented languages and technology have propelled software reuse to an unprecedented level. However, software reuse at the current stage is still mostly (1) within a single software project and (2) in the form of source-code reuse and as libraries. The dream that software construction would be similar to assembling Integrated Circuits (ICs) in hardware is still not realized. Programming in many essential aspects is the same as twenty years ago. Software construction is still basically at the stage of program writing, and has not advanced to the stage of software integration (SI). We consider that object reuse (rather than class reuse) is a key concept in software integration. Objects are the basic building blocks of an object-oriented system. They are first-class citizens. Programming via software integration means that we can reuse and integrate objects, which are developed independently (by other This work is supported by the Natural Sciences and Engineering Research Council of Canada grant OGP0041630. projects or object vendors) and in an executable form, into our program conveniently without knowing their implementation details. We consider that object reuse is a much more important concept than class reuse in software integration. Consider an example in the physical world: When we purchase an air-conditioner for a house, we do not need or want to know the design and technical details of the air-conditioner. It is only necessary for us to know whether the air-conditioner satisfies a few specifications, e.g., cooling capacity, physical dimensions, and voltage. In this case, the object (the air-conditioner) and its type (a few specifications) are our concern and interest. We do not need or even want to know the class of the object (the technical details of the air-conditioner). We consider that one major cause of the stagnation in reuse is that, in most major (class-based) object-oriented programming languages, object classes are directly used as object types. For example, when we specify the object type of a function parameter, we use a class in place of a type. This approach may be due to the fact that the semantics of an object type is easy to define by a specific implementation. It is important to notice that classes are implementation-dependent entities in general and also objects cannot exist autonomously apart from their class definitions in current class-based languages. In contrast, object types are implementation-independent entities. The class-is-type principle used in many object-oriented languages restricts an object type to a specific implementation. We argue in this paper that the class-is-type principle, not the classes themselves, is an obstacle to object reuse. Another issue that concerns the type hierachy of object-oriented languages is about parameterized types. Note that a C++ class template is a parameterized type under the class-is-type principle. We consider that a parameterized type is just a type function which maps a given type (or types) to a type. The domain of such a type function, i.e., the domain of its type parameter(s), is commonly represented by T (T n ), where T is the set of all types. However, in many cases, not all types can be used to replace the type parameter of a parameterized type. There are restrictions which are implicitly imposed by the parameterized type definition. We consider that object kinds, which are higher-level entities than object types [31, 32, 26, 25, 34], are best suited for defining the domains of type parameters of parameterized types as well as for those of generic functions. Objects, object classes, object types and object kinds are basic entities in the type hierarchies of various object-oriented languages [31]. In the next section, we review the basic concepts of those entities as well as the relationships among them. We also discuss the similarities and differences between abstract classes, interfaces, and object types. We focus our attention to the class-is-type principle, giving an initial analysis on its advantages as well as its potential problems. We also distinguish object kinds from supertypes. Object reuse is defined and analyzed in Section 3. We consider that object reuse concerns mainly the following five issues: (1) object creation, (2) object autonomy, (3) object application, (4) object integration, and (5) object modification. We describe each of the five issues as well as its role in object reuse. In Section 4, we discuss what roles classes and the class-is-type principle play in each of the five issues of object reuse. We find that while classes can be apt at object creation and object autonomy, the class-is-type principle is inadequate for object application, object integration, and object modification, i.e., all the last three issues of object reuse. We also give our insight analysis on why this is the case. In Section 5, we argue that parameterized types and generic functions can be best served by kind-bounded qualification, i.e., each type parameter is qualified by a higher-level entity "object kind". We conclude our paper in Section 6 by summarizing our suggestions on the further development of object-oriented languages. Type hierarchy of object-oriented systems In this section, we give our view on the basic type entities in object-oriented systems. We will consider objects, object classes, object types, as well as the concept of object kinds. We will be brief on the commonly accepted concepts and pay more attention on the points that are related to the topics of the subsequent sections of the paper. Note that object classes and object types are very different entities; however, the approach that object classes are used in place of object types, i.e., the class-is-type principle, is used in many object-oriented languages, including C++ and Java. We will also describe the concept of object kind and the fundamental differences between supertypes and object kinds. The three seemingly equivalent terms, abstract class, interface, and object type, are also compared. 2.1 Objects An object is an integrated unit of data and methods, which is an abstraction of an application domain or an implementation domain entity, which has crisp boundaries and meaning for the problem concerned. Each object has an identity, data fields, and method fields. The data fields store the internal state of the object and act as data storage, while the method fields are functions that modify the internal state of the object, access the data storage, and interface the object with the outside world. The internal state of an object is modified dynamically during execution. Normally, this change can be made only through its own methods. Not all methods may be accessed from the outside of the object. Some may only be used by other methods of the object. We call those methods that can be directly accessed from the outside the properties of the object. Clearly, objects are first-class entities of object-oriented systems. 2.2 Object classes serve as the descriptions of objects as well as a mechanism to create objects. Objects are instances of classes. In class-based languages, the only way to describe an object is through its class. A class is also considered a set of objects which share the same description. Thus, the relation between an object O and its class C can be denoted O 2 C. In general, object classes are implementation-dependent entities. For example, a stack class with an array implementation and a stack class with a linked-list implementation are considered as different classes. Different objects of the same class have the same implementation, but different identities and maybe different internal states. An abstract class may be independent of any implementation. However, an abstract class cannot be the description of an object directly. An object can only be instantiated directly from a concrete class but not from an abstract class. So, an abstract class as a set has no direct object members. Object classes are static entities. Unlike objects, classes cannot be changed at runtime once they are defined. The relationship between classes are static, too. We can modify the relationship between objects but not their classes at runtime. We call a property of a class is a method of the class that can be accessed directly from the outside. All objects of a class (direct members) share exactly the same set of methods and, thus, they have exactly the same properties. 2.3 Object types Intuitively speaking, an object type is a description of a collection of properties without giving any implemen- tation. Thus, an object type is also considered a set of objects that have the same properties, but possibly different implementations. Clearly, object types are implementation independent entities. All objects of the same class are also of the same object type. However, all objects of the same object type may not belong to the same class. For example, two integer-stack classes S 1 and S 2 are implemented with a linked list and an array, respectively. They have the same set of operations: push(int), pop(), top(), empty(). S 1 and S 2 are two different classes, but of the same type, say T S . In this case, we also say that is an implementation of T S . A stack object O S of type T S is considered a member of the set T S , denoted O The fact that S 1 An object type may be defined in many ways. For example, an object type can be define by listing all its properties (the names and signatures of all its methods) [18], by extracting a type from a class using an operator [1], or by implicitly deleting the implementation details of a class when a type is needed. Similar to classes, subtypes can be defined in terms of their supertypes. We say that an object O is of type T if O possesses the properties of T . Note that O can be of type T in spite of that the creation of O may be independent of the definition of T . In some cases, an object O actually possesses the properties of an object type T , but O and T use different names for their properties. For example, O has the two properties below: void insert-top( Content ); Content delete-top(); and T is defined as: Type T - Content pop(); void push( Content ); and we know that insert top is the same property as push and delete top corresponds to pop. Then it would be convenient to have a "is of " language construct like the following: Object O is-of Type T - insert-top() is push(); delete-top() is pop(); which allows the programmer to claim that O is an object of type T and to specify the correspondence between the properties of O and those of T . Type checking and type equivalence are usually quite complex. The above construct would be helpful in reducing the work of the compiler. More formally, the "is of " is a binary operator which links an object to an object type. The "is" operator inside the "is of " construct is a propterty matching operator, which links a property of the object to a property of the object type. A simple interpretaion of the "is of " construct is that we require the two properties on the both side of the "is" operator to have exactly the same signature although they may have different names. This interpretaion is in general good enough for practical purposes. A more flexible and complex interpretaion is to require that the object type of the property on the lefthand side, which may be a function type, is a subtype of that of the property on the righthand side. This interpretation would involve contravariance and covariance rules. The relationship between classes and object types is an interesting one. They are both collections or sets of objects. But, object types are implementation independent entities while classes are not. We say that an object type defines a set of objects that have the same external behaviors. However, the precise meaning of behaviors would become extremely complicated to define if the behaviors were not restricted to a specific implementation. We may use, e.g., axiomatic, denotational, or logical way to specify the behaviors of an object type, but all those methods are too complex to be practical at the moment. The easiest way to precisely define the semantics of the properties of a type is, perhaps, to give a specific implementation of the operations. In other words, it is easy and safe to define an object type by an object class. This comes what we see in many popular object-oriented languages the class-is-type principle: using object classes as object types. The main advantage of using the class-is-type principle is that object types defined this way are rigorous, unambiguous, precise, and easy to define, although not necessarily easy to comprehend. The main disadvantage is that each object type is restricted to only one particular implementation and, thus, becomes implementation- dependent. We will argue later that this is a major obstacle to object reuse and software integration. In some object-oriented languages, e.g., Theta [18], object types are separate entities from classes. In Theta, a new object type, or simply a new type, is defined by the names and the signatures of its methods (properties). The semantics of the methods (properties) is not formally defined and is left for the programmer to interpret from their names and possibly informal definitions in the comments. For example, a stack type and its operations push, pop, top, etc. are well understood simply from their names. A graph type can be easily understood with a few lines of explanations in the comments. This informal approach can be dangerous if without care. However, this is perhaps the only practical way of defining an object type in a program without restricted it to a specific implementation, and this way works well in the physical world. Abstract classes are available in many major object-oriented languages. Abstract classes can play the role of object types in certain extent. However, abstract classes are not object types. The main differences between an abstract class and an object type are the following: (1) Abstract classes may or may not be implementation independent while object types are implementation independent with no exception. There is no consistency in this aspect on abstract classes. (2) For an object O and an abstract class C, we say that O 2 C if there exists a concrete class C 0 that is a descendant of C such that O is an instance of C 0 . In other words, there has to exist a declaration link or family relation between O and C, that is, we have to declare that C 1 is a subclass of subclass of C n , and O is an instance of C 0 . In contrast, the fact that O is of type T can be simply implied by the fact that O has all the properties of T . The relationship can be direct. There does not necessarily exist a family relation here. Interfaces in Java are one-step closer to object types than abstract classes are. Interfaces are implementation-independent entities. Thus, the above (1) does not apply on them, but (2) still does. 2.4 Object kinds All object types that share a certain set of common properties form an object kind and it is specified by those properties. Hence, object kinds are at a higher-level than object types. The members of an object kind, if any, are object types. Object kinds have been introduced and studied in [31, 32, 26, 25, 34]. The name kind has also appeared in [16, 21, 8, 9, 10], however, the connotations of this word are not the same. We will show later in Section 5 that object kinds are useful in defining parameterized types and generic functions. The relation between an object kind and its members, i.e., object types, cannot be represented as a super-type relation. Consider two object types: word (or character string) and acyclic graph, where a word can be implemented by an array or a linked list and an acyclic graph by an adjacency matrix or an adjacency list, etc. Here, we name them type W and type G, respectively. These two different types have a common property: the distance between two words, respectively two acyclic graphs, can be measured, i.e., each of the two object types has a distance function. Naturally, we can represent the set of types that have a set of common properties by a higher-level entity. In this case, we define an object kind K to be a set of all types that have a distance function. Clearly, both W and G are types in K. It is important to notice that K is not a supertype of W and G, and K cannot be replaced by a supertype of W and G in this case. Assume that we define a supertype T (instead of a kind K) such that T has a distance function which measures the distance between each pair of objects of type T . This would imply that there is a distance function between a word and an acyclic graph by the subtyping principle. This is clearly not what we intend to define. 3 Object reuse In this section, we consider the issues that concern object reuse. We classify objects into the following two categories: internal objects and external objects. In a program, we call the objects created within the program internal objects and those created elsewhere but used in the program external objects. Since the reuse of internal objects is relatively straightforward and involved in fewer issues than that of external objects, our discussion is focused on the reuse of external objects. The sources of the external objects may include other projects and object vendors. Object reuse is different from class reuse or source code reuse. By object reuse, we mean that we reuse objects that are already created and in an executable form, especially those that are developed independently, in a larger programming environment. Object reuse may involve the following five interrelated issues: (1) object creation: how an object is created, (2) object autonomy: whether an object can survive autonomously, (3) object application: how an object is used, object integration: how two or more objects are integrated into larger objects, object modification: whether an individual object can modified and what are the side effects. An object has to be created first before any application. Object can be created in many ways. In class-based languages like C++ [28], Java [3], Eiffle [22], and Modula-3 [23], objects are instantiated from classes. In object-based languages, e.g., Cecil [11], Self [30], and Omega [6], objects are created by cloning and extension of prototypical objects without the existence of classes. Apparently, how an object is created does not necessarily affect how an object is used internally or externally. We consider that object creation is not an essential issue for object reuse. In order for an object to be used not only internally but also externally in terms of the environment where it is created, it is essential that the object should be able to survive autonomously. This means that the object is executable without its class or other objects of its creation environment. Many issues may involve object autonomy, which include embedding versus delegation, global variables and class variables, and visibility of attributes. However, we feel that it is not difficult for any object-oriented languages to be modified to safely export objects, which can live autonomously and be reused safely in other runtime environments. For example, objects that are going to be exported can be specially marked as "exporting" objects, and the compiler will check whether they satisfy certain conditions for autonomy, e.g., no global variables, and if they do, then generate the autonomous objects using embedding instead of delegation, etc. Object application by definition is clearly the most important issue in object reuse. What conditions are necessary for an external object to be used in (or linked to) a programming environment? Let us again consider the air-conditioner example. When we purchase an air-conditioner for a house, it is necessary for us to know whether the air-conditioner satisfies a few specifications, e.g., cooling capacity, physical dimensions, and voltage. However, we do not need or want to know the design and technical details of the air-conditioner. Also, any air-conditioner that satisfies our requirements, whatever its internal implementation is, would do the job. The internal implementation is not part of our requirement. In this case, what we need are the object and its object type, and whether the object type conforms the required object type, and we do not need and want to know the object class. Thus, an object type system which is separate from the class system need to set up for object reuse. In this type system, for example, an object type may be explicitly declared, or obtained by using an operator which, given a class C, removes all the implementation details of C and returns an object type. Object types may have their own inheritance hierarchies. Most importantly, the system should be able to check, for a given object and a given object type, whether the object is of the give object type. When an external object is used in a programming environment, it may not just act as a server providing one-way services, it may also call the methods of (or send messages to) some internal objects. In other words, the object may interact with internal objects. Also consider the situation when we use two or more external objects in our programming environment. Those objects may require to interact with each other. In a strong- typed object-oriented system (as we have assumed), each object involved in a multi-object interaction needs to know the type information of the other objects. However, we may be at a situation that external objects are developed independently by different vendors. Thus, in general, the developer of each object does not know how the types of other objects are defined. There are many different ways to solve this problem. Example solutions are (1) standardization of object types for object interfaces, (2) writing wrapper or adapter programs in order for independently developed objects to interact, and (3) using the "is of " construct suggested in Section 2 to match the equivalent types that have different appearances. Solution (1) above may be a long-term solution. However, even in the long term, it is difficult to standardize all the types for reusable objects. Other mechanisms have to be in place. The wrapper and adapter programs are complicated and the programmer has to know the type declarations of each object involved (and have to have their class declarations under the current class-is-type principle). It would be much more convenient to use the "is of " construct together with the standardization. It would also be convenient if objects can be modified after creation. In several object-based languages, both data fields and methods of an object can be modified in a uniform way without any side-effect. Object modification and object autonomy are two closely related issues. However, we consider this issue as a convenience issue. The notion of object reuse appears to be similar to that of component-based computing [29] in many ways. Both a component and an object (in object reuse) are being independently developed and in an executable form. We consider that a component is just an object or a collection of closely related objects in our term. However, component-based computing mechanisms in general are not part of an object-oriented language; they impose an external structure on the programs written in current programming languages rather than change them. IDL can be consider a standard way to specify object types. In some sense, the component-based computing mechanisms are meant to alleviate the problems of current object-oriented languages in object reuse from the outside rather than from the inside. 4 Class-is-type is inadequate for object reuse play two major roles in class-based languages: (1) as an object description and creation mechanism, and (2) as object types. As an object creation mechanism, classes can serve the first two issues of object reuse well. With some minor modifications, the current popular class-based languages can be used to develop autonomous objects. Objects can be also created in many other ways in object-based [1] or prototype-based languages [27, 24, 13], e.g., by direct declaration and by cloning and extension from prototypical objects. We consider that classes as an object creation mechanism is as adequate as any other mechanism whether it is for creating a single object or a stock of objects. We now consider the class-is-type principle, i.e., classes plying the role of object types, in relation to object reuse. Let us assume that we are going to use a "registrar" object, from an external source, in our program. In our program, we instantiate many student objects from the "STUDENT" class (type) in our program, each of which has a method for registration. When a STUDENT s1 registers, it calls the registration method and uses the external object registrar for registration: s1:register( registrar ); This implies that we have the following definition of a method for registration in STUDENT class: Boolean register( REGISTRAR ); where REGISTRAR is the class, as well as the object type, of registrar. The class REGISTRAR has to be declared in our program. Since the class-is-type principle is used, we have the following problems: ffl The class REGISTRAR (and possibly its descendants), which describes all the implementation details of the object registrar, can be very complicated and large (this is why we reuse it instead of developing it by ourselves). Since we have to include the class REGISTRAR (and its descendants) in our program, it would be more economical just to instantiate a registrar object in our program rather than to use the external object registrar. Then object reuse degenerates into class reuse (or source code reuse). ffl If both registrar and student s1 are external objects and are created independently, then the register method of the s1 object may not be defined as above since the student object does not have the class definition of a registrar when it is created. (Also, the registrar object does know the class definition of s1 objects.) This seriously restricts the way the registrar and s1 objects are defined and makes integration of those objects extremely complex. ffl Neither the registrar object nor the s1 object can be modified to have additional features or more efficient implementations once it is created. From the above example, we can observe that there are the following three general problems of the class-is- type principle for object reuse: (1) Classes are implementation dependent entities. The class-is-type principle unnecessarily restricts an object type to a specific implementation. It is not compatible with the general idea of reuse. (2) Under the class-is-type principle, an object O can be used where object type (class) T is specified if O is an instance of T or an instance of a descendant class of T . In other words, O has to have a declaration link or a family relation with T . Since external objects are developed (programmed) independently, this necessary link makes the reuse of external objects complicated and difficult. (3) Classes are detailed description of objects. When the information about the type of an object is re- quired, the class-is-type principle makes the object type information too cumbersome and even sometime redundant to the object itself. The class-is-type principle makes the realization of object application, object integration, and object modification difficult and even impossible in some cases. 5 Parameterized types and generic functions parameterized type is a function which maps a type (or several types) into another type. Let T denote the set of all types. Then a parameterized type P is a function For example, a C++ Sorted Vector template declaration [28] is in the following: template!class T? class Sorted-Vector - T* v; int sz; // size of the vector int n; // current number of elements public: explicit Vector( int ); T& operator[](int i); if (n == zs) return false; for (int return true; bool delete( int i ); // delete ith element where T in the above is the type parameter of the template. The template maps the int type to an integer Sorted Vector type, maps the char type to the character Sorted Vector type, etc. It appears that given any type the template would map it to a Sorted Vector of that type. However, this is not totally true. Notice that there is a comparison v[i] ! a between two objects of type T in the function insert. This implies that T cannot be any type but only types that have a '!' operator. So, the domain of this type function (the template) is not T (the set of all types) but a subset of T . We call this subset a KIND K. Then the Sorted Vector template is a type function: K ! T . (In general, a parameterized type is a function: K 1 are KINDs.) The KIND K may be defined as follows: KIND K: for T in K - bool operator! (T, T); Then the template may be written as follows: template!K T? class Sorted-Vector - where !K T? denotes that type T is bound to KIND K. Similarly, a generic function is a function that maps from a type to a function and, thus, kinds can be used to define the domain of such type functions, too. We call such type of polymorphism KIND-bounded polymorphism. Note that we have discussed in Section 2 that KINDs cannot be replaced by supertypes. KIND and supertype are clearly two different concepts. Consider this situation: a type T has a comparison operator, which is essentially the same as ! but uses a different name, say "less than". A language construct like the following to explicitly link the corresponding names may be introduced: Type T is-in KIND K - less-than is operator!; By using KIND-bounded polymorphism, there are at least the following advantages: (1) The restrictions, on type variables, that are implicitly imposed by the definition of a parameterized type or a generic function are explicitly and clearly stated by the definition of a KIND. The user of the parameterized type or generic function need not read the detailed definition of the parameterized type or generic function to find out all the buried constraints. (2) KINDs are at a higher level than types. A KIND does not affiliate to a specific parameterized type or a generic function. KINDs are natural entities to represent constraints on types, which make generic constructs conceptually more transparent. Notions similar to kinds have been studied in [16, 21, 20, 8, 9, 10]. The where clause of parameterized procedures and parameterized types in CLU [19] and Theta [18], as well as the with statement of generic procedures in Ada [2] are all similar to the definition of KINDs. However, unlike kinds, the where clause and the with statement are not independent entities. They are an inseparable part of a specific parameterized type etc. In contrast, kinds are treated uniformly as other entities in our type hierarchy. Several parameterized types and genric functions that have the same restrictions on their type parameters can be expressed by the same KIND. Comparisons between KIND and opaque types in Modula-3, signatures in G++ [5], "kinds" in Quest [10], etc. can be found in [31]. The concept of object kinds has shown to be useful in realizing algorithmic abstraction [31] in programming languages. We conjecture that object kinds will also be useful in formal or semi-formal descriptions of design patterns as well as in their implementation in object-oriented languages. 6 Conclusion In the past half century, programming languages have been continuously evolving to higher levels. There exist two opposite directions of the development: programming languages are getting more and more complicated conceptually, but easier and easier technically. It takes a much longer time to learn a programming language now than forty years ago, but programming for the same task takes much less time. In this paper, we have made several suggestions to the development of object-oriented programming lan- guages, which are summarized in the following: 1. Separation of object types from object classes. 2. Exporting autonomous objects with their type information. 3. Introducing a type-matching construct like the "is of " construct in Section 2, which matches the type of a given object to a given object type. 4. Introducing object kinds, kind-bounded polymorphism, and the "is in" construct for fitting a object type into an object kind. An object type system that is separated from classes would be much more complicated to implement. However, the system would make an object-oriented language much more flexible and feasible for object reuse and software integration. --R A Theory of Objects Ada 9X mapping/Revision Team The Java Programming Language Programming in Ada "Type-Safe OOP with Prototypes: The Concept of Omega" Algebraic Specification Techniques in Object Oriented Programming Environments "A Modest Model of Records, Inheritance, and Bounded Quantification" "The Semantics of Second-order Lambda Calculus" Formal Description of Programming Concepts - IFIP State-of-art Report "The Cecil Language Specification and Rationale" "F-Bounded Polymorphism for Object-Oriented Programming" The Interpretation of Object-Oriented Programming Languages Data Types Are Values Principles of OBJ2 Interpretation Fonctionelle et 'elimination des coupures de l'arithm'etique d'ordre sup'erieur The Algebraic Specification of Abstract Data Types Theta Reference Manual - Preliminary Version Abstraction and Specification in Program Development A Semantic Model of Types For applicative Languages An Investigation of a Programming Language with a Polymorphic Type Structure Eiffle: The Language Systems Programming with Modula-3 Introducing KINDS to C "A shared view of sharing: The treaty of Orlando" Component Software - Beyond Object-Oriented Programming "Self: The Power of Simplicity" "Algorithmic Abstraction in Object-Oriented Languages" "Software Reuse via Algorithm Abstraction" On Parametric Polymorphism in Object-Oriented Languages Algorithm Abstraction via Polymorphism In Object-Oriented Languages --TR Data types are values Abstraction and specification in program development A shared view of sharing: the treaty of Orlando The semantics of second-order lambda calculus A modest model of records, inheritance, and unbounded quantification Inheritance is not subtyping F-bounded polymorphism for object-oriented programming Systems programming with Modula-3 Eiffel: the language Component software Principles of OBJ2 The C++ Programming Language The Integration of Object-Oriented Programming Languages Programming ADA A Theory of Objects A semantic model of types for applicative languages An investigation of a programming language with a polymorphic type structure. Algorithmic abstraction via polymorphism in object-oriented programming languages --CTR Chitra Babu , D. Janakiram, Method driven model: a unified model for an object composition language, ACM SIGPLAN Notices, v.39 n.8, August 2004

kinds;object reuse;classes;kind-bounded polymorphism;parameterized types;generic functions;types;objects;class-is-type principle

504535

Minimal cover-automata for finite languages.

A cover-automaton A of a finite language L &Sgr; is a finite deterministic automaton (DFA) that accepts all words in L and possibly other words that are longer than any word in L. A minimal deterministic finite cover automaton (DFCA) of a finite language L usually has a smaller size than a minimal DFA that accepts L. Thus, cover automata can be used to reduce the size of the representations of finite languages in practice. In this paper, we describe an efficient algorithm that, for a given DFA accepting a finite language, constructs a minimal deterministic finite cover-automaton of the language. We also give algorithms for the boolean operations on deterministic cover automata, i.e., on the finite languages they represent.

Introduction Regular languages and finite automata are widely used in many areas such as lexical analysis, string matching, circuit testing, image compression, and parallel processing. However, many applications of regular languages use actually only finite languages. The number of states of a finite automaton that accepts a finite language is at least one more than the length of the longest word in the language, and can even be in the order of exponential to that number. If we do not restrict an automaton to accept the exact given finite language but allow it to accept extra words that are longer than the longest word in the language, we may obtain an automaton such that the number of states This research is supported by the Natural Sciences and Engineering Research Council of Canada grants OGP0041630. Preprint submitted to Elsevier Preprint is significantly reduced. In most applications, we know what is the maximum length of the words in the language, and the systems usually keep track of the length of an input word anyway. So, for a finite language, we can use such an automaton plus an integer to check the membership of the language. This is the basic idea behind cover automata for finite languages. Informally, a cover-automaton A of a finite language L ' \Sigma is a finite automaton that accepts all words in L and possibly other words that are longer than any word in L. In many cases, a minimal deterministic cover automaton of a finite language L has a much smaller size than a minimal DFA that accept L. Thus, cover automata can be used to reduce the size of automata for finite languages in practice. Intuitively, a finite automaton that accepts a finite language (exactly) can be viewed as having structures for the following two functionalities: (1) checking the patterns of the words in the language, and (2) controlling the lengths of the words. In a high-level programming language environment, the length-control function is much easier to implement by counting with an integer than by using the structures of an automaton. Furthermore, the system usually does the length-counting anyway. Therefore, a DFA accepting a finite language may leave out the structures for the length-control function and, thus, reduce its complexity. The concept of cover automata is not totally new. Similar concepts have been studied in different contexts and for different purposes. See, for exam- ple, [1,7,4,10]. Most of previous work has been in the study of a descriptive complexity measure of arbitrary languages, which is called "automaticity" by Shallit et al. [10]. In our study, we consider cover automata as an implementing method that may reduce the size of the automata that represent finite languages. In this paper, as our main result, we give an efficient algorithm that, for a given finite language (given as a deterministic finite automaton or a cover automaton), constructs a minimal cover automaton for the language. Note that for a given finite language, there might be several minimal cover automata that are not equivalent under a morphism. We will show that, however, they all have the same number of states. Preliminaries Let T be a set. Then by #T we mean the cardinality of T . The elements of are called strings or words. The empty string is denoted by -. If w 2 T then jwj is the length of x. We define T l is an ordered set, k ? 0, the quasi-lexicographical order on denoted OE, is defined by: x OE y iff jxj ! jyj or or y. We say that x is a prefix of y, denoted x - p y, if A deterministic finite automaton (DFA) is a quintuple where \Sigma and Q are finite nonempty sets, q is the transition function. We can extend ffi from Q \Theta \Sigma to Q \Theta \Sigma by We usually denote ffi by ffi. The language recognized by the automaton A is Fg. For simplicity, we assume that In what follows we assume that ffi is a total function, i.e., the automaton is complete. Let l be the length of the longest word(s) in the finite language L. A DFA A such that called a deterministic finite cover-automaton (DFCA) of L. Let A = (Q; \Sigma; ffi; 0; F ) be a DFCA of a finite language L. We say that A is a minimal DFCA of L if for every DFCA of L we have #Q - #Q 0 . a) q 2 Q is said to be accessible if there exists w 2 \Sigma such that ffi(0; b) q is said to be useful (coaccessible) if there exists w 2 \Sigma such that ffi(q; w) 2 F . It is clear that for every DFA A there exists an automaton A 0 such that all the states of A 0 are accessible and at most one of the states is not useful (the sink state). The DFA A 0 is called a reduced DFA. 3 Similarity sequences and similarity sets In this section, we describe the L-similarity relation on \Sigma , which is a generalization of the equivalence relation jL (x jL y: xz 2 L iff yz 2 L for all z 2 \Sigma ). The notion of L-similarity was introduced in [7] and studied in [4] etc. In this paper, L-similarity is used to establish our algorithms. Let \Sigma be an alphabet, L ' \Sigma a finite language, and l the length of the longest in L. Let x; y 2 \Sigma . We define the following relations: (1) x -L y if for all z 2 \Sigma such that jxzj - l and jyzj - l, xz 2 L iff yz 2 L; (2) x 6- L y if x -L y does not hold. The relation -L is called similarity relation with respect to L. Note that the relation -L is reflexive, symmetric, but not transitive. For exam- ple, let aabbg. It is clear that aab -L aabb (since but aab 6- L baa (since for 2 L and The following lemma is proved in [4]: be a finite language and x; The following statements hold: (1) If x -L y, x -L z, then y -L z. (2) If x -L y, y -L z, then x -L z. (3) If x -L y, y6- L z, then x6- L z. If x 6- L y and y -L z, we cannot say anything about the similarity relation between x and z. Example 2 Let x; We may have y, y -L z and x -L z, or y, y -L z and x6- L z. Indeed, if we choose be a finite language. (1) A set S ' \Sigma is called an L-similarity set if x -L y for every pair x; y 2 S. (2) A sequence of words [x is called a dissimilar sequence of for each pair (3) A dissimilar sequence [x called a canonical dissimilar sequence of L if there exists a partition of \Sigma such that for each is a L-similarity set. (4) A dissimilar sequence [x of L is called a maximal dissimilar sequence of L if for any dissimilar sequence [y Theorem 4 A dissimilar sequence of L is a canonical dissimilar sequence of L if and only if it is a maximal dissimilar sequence of L. PROOF. Let L be a finite language. Let [x be a canonical dissimilar sequence of L and the corresponding partition of \Sigma such that for each is an L-similarity set. Let [y be an arbitrary dissimilar sequence of L. Assume that m ? n. Then there are y i and is a L-similarity set, y i -L y j . This is a contradiction. Then, the assumption that m ? n is false, and we conclude that [x is a maximal dissimilar sequence. Conversely, let [x dissimilar sequence of L. Without loss of generality we can suppose that jx 1 j - jx n j. For Note that for each y 2 \Sigma , y -L x i for at least one is a maximal dissimilar sequence. Thus, is a partition of \Sigma . The remaining task of the proof is to show that each X i , set. We assume the contrary, i.e., for some i, 1 - i - n, there exist such that y6- L z. We know that x i -L y and x i -L z by the definition of We have the following three cases: (1) jx (or or (2), then y -L z by Lemma 1. This would contradict our assumption. If (3), then it is easy to prove that y 6- x j and z 6- x j , for all j 6= i, using Lemma 1 and the definition of X i . Then we can replace x i by both y and z to obtain a longer dissimilar sequence This contradicts the fact that is a maximal dissimilar sequence of L. Hence, y - z and X i is a similarity set. Corollary 5 For each finite language L, there is a unique number N(L) which is the number of elements in any canonical dissimilar sequence of L. Theorem 6 Let S 1 and S 2 be two L-similarity sets and x 1 and x 2 the shortest words in S 1 and S 2 , respectively. If x 1 -L x 2 then set. PROOF. It suffices to prove that for an arbitrary word y 1 and an arbitrary word y 2 holds. Without loss of generality, we assume that jx 1 j - jx 2 j. We know that jx 1 we have y 1 -L y 2 (Lemma 1 (1)). 4 Similarity relations on states it is clear that if y. Therefore, we can also define similarity as well as equivalence relations on states. be a DFA. We define, for each state i.e., level(q) is the length of the shortest path from the initial state to q. If for each q 2 Q, we denote xA qg, where the minimum is taken according to the quasi- lexicographical order, and LA Fg. When the automaton A is understood, we write x q instead of xA (q) and L q instead LA (q). The length of x q is equal to level(q), therefore level(q) is defined for each We say that is equivalent to q in A) if for every w 2 \Sigma , ffi(s; w) 2 F iff be a DFCA of a finite language L. Let jg. We say that p -A q (state p is L-similar to q in A) if for every w 2 \Sigma -l\Gammam , ffi(p; w) 2 F iff ffi(q; w) 2 F . be a DFCA of a finite language L. Let such that ffi(0; y. PROOF. Let Choose an arbitrary w 2 \Sigma such that jxwj - l and jywj - l. Because i - jxj and j - jyj it follows that jwj - l \Gamma m. Since p -A q we have that ffi(p; w) 2 F which means that xw 2 L(A) Hence x -L y. Lemma be DFCA of a finite language L. Let that ffi(0; PROOF. Let x -L y and w 2 \Sigma -l\Gammam . If ffi(p; w) 2 F , then ffi(0; xw) 2 F . Because x -L y, it follows that ffi(0; yw) 2 F , so ffi(q; w) 2 F . Using the symmetry we get that p -A q. Corollary 12 Let A = (Q; \Sigma; ffi; 0; F ) be a DFCA of a finite language L. Let \Sigma , such that ffi(0; x 1 then x 2 -L y 2 . Example 13 If x 1 and y 1 are not minimal, i.e. jx then the conclusion of Corollary 12 is not true. 3. The following is a DFCA of L and we \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi* a -a ja b' -@ @ @ @ @R HY a HY Fig. 1. If x -L y then we do not have always that ffi(0; x) -A ffi (0; y) have that b -L bab, but b6- L a (ba 2 L, aa 2 L and Corollary 14 Let A = (Q; \Sigma; ffi; 0; F ) be a DFCA of a finite language L and If p -A q, and level(p) - level(q) and q 2 F then p 2 F . Lemma be a DFCA of a finite language L. Let s; m. The following statements are true: (1) If s -A p, s -A q, then p -A q. (3) If s -A p, p6- A q, then s6- A q. PROOF. We apply Lemma 1 and Corollary 14. Lemma be a DFCA of a finite language L. Let is a L-similarity set. Therefore ffi(p; w) 2 F , and jwj - l \Gamma m. Hence, because p -A q, ffi(p; w) 2 F , so w 2 L q " \Sigma -l\Gammam . Lemma be a DFCA of a finite language L. If -A q for some Then we can construct a DFCA A of L such that ffi(s; a) if ffi(s; a) 6= q; for each s 2 Q 0 and a 2 \Sigma. Thus, A is not a minimal DFCA of L. PROOF. It suffices to prove that A 0 is a DFCA of L. Let l be the length of the longest word(s) in L and assume that Consider a word w 2 \Sigma -L . We now prove that w 2 L iff If there is no prefix w 1 of w such that ffi(0; w is the shortest prefix of w such that In the remaining, it suffices to prove that ffi 0 (p; w We prove this by induction on the length of w 2 . First consider the case by the construction of A 0 . Thus, . Suppose that the statement holds for jw Consider the case that jw 2 . If there does not exist u such that and u be the shortest nonempty prefix of w 2 such that ffi(p; By induction hypothesis, Lemma A be a DFCA of L and L x -L y. PROOF. Let l be the length of the longest word(s) in L. Let x jL 0 y. So, for each z 2 \Sigma ; xz 2 L 0 iff yz 2 L 0 . We now consider all words z 2 \Sigma , such that j xz j- l and j yz j- l. Since have xz 2 L iff yz 2 L. Therefore, x -L y by the definition of -L . Corollary 19 Let A = (Q; \Sigma; ffi; 0; F ) be a DFCA of a finite language L, L L(A). Then p jA q implies p -A q. Corollary 20 A minimal DFCA of L is a minimal DFA. PROOF. Let A = (Q; \Sigma; ffi; 0; F ) be a minimal DFCA of a finite language L. Suppose that A is not minimal as a DFA for L(A), then there exists such that p jL 0 q, then p -A q. By Lemma 17 it follows that A is not a minimal DFCA, contradiction. Remark 21 Let A be a DFCA of L and A a minimal DFA. Then A may not be a minimal DFCA of L. Example 22 We take the DFA's: a a a @ \Gamma\Psi a Automaton A @ \Gamma\Psi a Fig. 2. Minimal DFA is not always a minimal DFCA. The DFA A in Figure 2 is a minimal DFA and a DFCA of aag but not a minimal DFCA of L, since the DFA B in Figure 2 is a minimal DFCA of L. Theorem 23 Any minimal DFCA of L has exactly N(L) states. PROOF. Let A = (Q; \Sigma; ffi; 0; F ) be DFCA of a finite language L, and n. Suppose that n ? N(L). Then there exist q, such that x p -L x q (because of the definition of N(L)). Then p -A q by Lemma 14. Thus, A is not minimal. A contradiction. Suppose that N(L) ? n. Let [y be a canonical dissimilar sequence of L. Then there exist Again a contradiction. Therefore, we have 5 The construction of minimal DFCA The first part of this section describe an algorithm that determines the similarity relations between states. The second part is to construct a minimal DFCA assuming that the similarity relation between states is known. An ordered DFA is a DFA where ffi(i; implies that i - j, for all states letters a. Obviously for such a DFA is the sink state. 5.1 Determining similarity relation between states The aim is to present an algorithm which determines the similarity relations between states. of a finite language L. Define D Fg, and D is taken according to the quasi-lexicographical order. If the automaton A is understood then we write D i and fl s instead of D i (A) and respectively fl s (A). Lemma of a finite language L, and p 2 PROOF. We can assume that i ! j. Then obviously ffi(p; j. So, we have that jfl Lemma accepting L, . If for all a 2 \Sigma, ffi(p; a) -A ffi(q; a), then p-A q. PROOF. Let a 2 \Sigma and ffi(p; a) = r and ffi(q; a) = s. If r -A s then for all jwj, also have: xA (q)aw 2 L iff xA (s)w 2 L for all w 2 \Sigma , jwj - l \Gamma jx A (s)j and Hence xA (p)aw 2 L iff xA (q)aw 2 L, for all w 2 \Sigma , jwj - l \Gamma maxfjxA (r)j; jx A (s)jg. Because jx A (r)j - jx A Since a 2 \Sigma is chosen arbitrary, we conclude that xA (p)w 2 L iff xA (q)w 2 L, for all w 2 \Sigma , jwj - l \Gamma maxfjxA (p)j; jx A (q)jg, i.e. xA (p) -A xA (q). Therefore, by using Lemma 11, we get that p -A q. Lemma 26 Let accepting L such that . If there exists a 2 \Sigma such that ffi(p; a)6- A ffi(q; a), then p6- A q. PROOF. Suppose that p -A q. Then for all aw 2 \Sigma l\Gammam , ffi(p; aw) 2 F iff by definition that ffi(p; a) -A ffi(q; a). This is a contradiction. Our algorithm for determining the similarity relation between the states of a DFA (DFCA) of a finite language is based on Lemmas 25 and 26. However, most of DFA (DFCA) do not satisfy the condition of Lemma 26. So, we shall first transform the given DFA (DFCA) into one that does. be a DFCA of L. We construct the minimal DFA for the language \Sigma -l , The DFA B will have exact l states. Now we use the standard Cartesian product construction (see, e.g., [3], for details) for the DFA (taking the automata in this order) and we eliminate all inaccessible states. Obviously, satisfies the condition of Lemma 26. Lemma 27 For the DFA C constructed above, PROOF. We have Lemma 28 For the DFA C constructed above we have (p; q) -C (p; r). PROOF. If p 2 D \Gamma1 (A), the lemma is obvious. Suppose now that and q - r. Then r - l so It follows that Lemma 29 For the DFA C constructed above we have that (#Q \Gamma PROOF. We have that ffi C 2 FC and 2 FC it follows the conclusion. Now we are able to present an algorithm, which determines the similarity relation between the states of C. Note that QC is ordered by that (p A ; pB . Attaching to each state of C is a list of similar states. For ff; fi 2 QC , if ff -C fi and ff ! fi, then fi is stored on the list of similar states for ff. We assume that reduced (so is the sink state of A). (1) Compute (2) Initialize the similarity relation by specifying: (a) For all (n \Gamma (b) For all (n \Gamma (3) For each D i (C), create a list List i , which is initialized to ;. (4) For each ff 2 following the reversed order of QC , do the following: Assuming ff 2 D i (C). (a) For each fi 2 List i , if ffi C (ff; a) -C ffi C (fi; a) for all a 2 \Sigma, then ff -C fi. (b) Put ff on the list List i . By Lemma 24 we need to determine only the similarity relations between states of the same D i (C) set. The Step 2(a) follows from Lemma 28, 2(b) from Lemma 29 and Step 4 from Lemma 15. Remark 30 The above algorithm has complexity O((n \Theta l) 2 ), where n is the number of states of the initial DFA (DFCA) and l is the maximum accepted length for the finite language L. 5.2 The construction of a minimal DFCA As input we have the above DFA C and, with each ff 2 QC , a set S fig. The output is DFCA for L. We define the following: while (T 6= ;) do the following: and ;g. Note that the constructions of x i above are useful for the proofs in the following only, where the min (minimum) operator for x i is taken according to the lexicographical order. According to the algorithm we have a total ordering of the states QC : (p; q) - Also, using the construction (i.e. the total order on QC ) it follows that Lemma 31 The sequence [x constructed above is a canonical L-dissimilar sequence. PROOF. We construct the sets X g. Obviously it follows that X i is a L-similarity set for all Let w 2 \Sigma . Because (S i ) 1-i-m\Gamma1 is a partition of Q, w is a partition of \Sigma and therefore a canonical L-dissimilar sequence. Corollary 32 The automaton D constructed above is a minimal DFCA for L. PROOF. Since the number of states is equal to the number of elements of a canonical L-dissimilar sequence, we only have to prove that D is a cover automaton for L. Let w 2 \Sigma -l . We have that ffi D (0; w) 2 FD iff ffi C ((0; 0); w) 2 (because C is a DFCA for L). 6 Boolean operations We shall use similar constructions as in [3] for constructing DFCA of languages which are a result of boolean operations between finite languages. The modifications are suggested by the previous algorithm. We first construct the DFCA which satisfies hypothesis of Lemma 26 and afterwards we can minimize it using the general algorithm. Since the minimization will follow in a natural way we shall present only the construction of the necessarily DFCA. Let A two DFCA of the finite languages L i , l 2. 6.1 Intersection We construct the following DFA: lg. Theorem 33 The automaton A constructed above is a DFA for PROOF. We have the following relations: The rest of the proof is obvious. 6.2 Union We construct the following DFA: where r is such that l r = l. Theorem 34 The automaton A constructed above is a DFA for PROOF. We have the following relations: The rest of the proof is obvious. 6.3 Symmetric difference We construct the following DFA: f(s; r is such that l r = l. Theorem 35 The automaton A constructed above is a DFA for PROOF. We have the following relations: or exclusive w or exclusive w 2 The rest of the proof is obvious. 6.4 Difference We construct the following DFA: Theorem 36 The automaton A constructed above is a DFA for PROOF. We have the following relations: and The rest of the proof is obvious. Open Problems 1) Try to find a better algorithm for minimization or prove that any minimization algorithm has complexity Find a better algorithm for determining similar states in any DFCA of L. Find better algorithms for boolean operations on DFCA. --R Uniform characterisations of non-uniform complexity measures "zone" Regular languages and programming languages time complexity gap for two-way probabilistic finite-state automata Two memory bounds for the recognition of primes by automata Introduction to Automata Theory Minimal Nontrivial Space Space Complexity of Probabilistic One-Way Turing Machines Running time to recognise non-regular languages by 2- way probabilistic automata A class of measures on formal languages Properties of a Measure of Descriptional Complexity Theory of Automata The state complexities of some basic operations on regular languages Finite Automata: Behaviour and Synthesis On the State Complexity of Intersection of Handbook of Formal Languages --TR Uniform characterizations of non-uniform complexity measures Minimal nontrivial space complexity of probabilistic one-way turing machines time complexity gap for two-way probabilistic finite-state automata Running time to recognize nonregular languages by 2-way probabilistic automata On the state complexity of intersection of regular languages The state complexities of some basic operations on regular languages Automaticity I Regular languages Introduction To Automata Theory, Languages, And Computation Theory of Automata --CTR Martin Kappes , Frank Niener, Succinct representations of languages by DFA with different levels of reliability, Theoretical Computer Science, v.330 n.2, p.299-310, 2 February 2005 Martin Kappes , Chandra M. R. Kintala, Tradeoffs between reliability and conciseness of deterministic finite automata, Journal of Automata, Languages and Combinatorics, v.9 n.2-3, p.281-292, September 2004

finite languages;deterministic cover automata;deterministic finite automata;cover language;finite automata

504537

Normal form algorithms for extended context-free grammars.

We investigate the complexity of a variety of normal-form transformations for extended context-free grammars, where by extended we mean that the set of right-hand sides for each nonterminal in such a grammar is a regular set. The study is motivated by the implementation project GraMa which will provide a C++ toolkit for the symbolic manipulation of context-free objects just as Grail does for regular objects. Our results generalize known complexity bounds for context-free grammars but do so in nontrivial ways. Specifically, we introduce a new representation scheme for extended context-free grammars (the symbol-threaded expression forest), a new normal form for these grammars (dot normal form) and new regular expression algorithms. Copyright 2001 Elsevier Science B.V.

Introduction In the 1960's, extended context-free grammars were introduced, based on Backus-Naur form, as a useful abbreviatory notation that made context-free grammars easier to write. More recently, the Standardized General Markup Language (SGML) [16] used a similar abbrevia- tory notation to define extended context-free grammars for documents. Currently, Extensible Markup Language (XML) [6], which is a simplified version of SGML, is being promoted as the markup language for the web, instead of HTML (a specific grammar or document type definition (DTD) specified using SGML). These developments led to the investigation of how notions applicable to context-free grammars could be carried over to extended context-free grammars. There does not appear to have been any consolidated effort to study extended context-free grammars in their own right. We begin such an investigation with the most basic This research was supported under a grant from the Research Grants Council of Hong Kong SAR. It was carried out while the first and second authors were visiting HKUST. y Lerhstuhl fur informatik II, Universitat Wurzburg, Am Hubland, D-97074 Wurzburg, Germany. E-mail: albert@informatik.uni-wuerzburg.de. z Dipartimento di Matematica Applicata e Informatica, Universit'a Ca' Foscari di Venezia, via Torino 155, 30173 Venezia Mestre, Italy. E-mail: dora@dsi.unive.it. x Department of Computer Science, Hong Kong University of Science & Technology, Clear Water Bay, Kowloon, Hong Kong SAR. E-mail: dwood@cs.ust.hk. problems for extended context-free grammars: reduction and normal-form transformations. There has been some related work that is more directly motivated by SGML issues; see the proof of decidability of structural equivalence for extended context-free grammars [4] and the demonstration that SGML exceptions do not add expressive power to extended context-free grammars [17]. We are currently designing a manipulation system toolkit GraMa for extended context-free grammars, pushdown machines and context-free expressions. It is an extension of Grail [20, 19], a similar toolkit for regular expressions and finite-state machines. As a re- sult, we need to choose appropriate representations of grammars and machines that admit efficient transformation algorithms (as well as other algorithms of interest). Earlier results on context-free grammars were obtained by Harrison and Yehudai [12, 13, 26] and by Hunt et al. [15] among others. Harrison's chapter on normal form transformations [12] provides an excellent survey of early results. Cohen and Gotlieb [5] suggested a specific representation for context-free grammars and demonstrated how it aided the programming of various operations on them. We first define extended context-free grammars using the notion of production schemas that are based on regular expressions. In a separate paper [9], we discuss the algorithmic impact of basing the schemas on finite-state machines. Since finite-state machines and regular expressions are both first-class objects in Grail, they can be used interchangeably as we expect they will be in GraMa. We next describe algorithms for the fundamental normal-form transformations in Section 3. Before doing so, we propose a representation for extended context-free grammars as regular expression forests with symbol threads. We then discuss some algorithmic problems for regular expressions before tackling the various normal forms. We close the presentation, in Section 4, with a brief discussion of our ongoing investigations. Notation and terminology We treat extended context-free grammars as context-free grammars in which the right-hand sides of productions are regular expressions. Let V be an alphabet. Then, we define a regular expression over V and its language in the usual way [1, 25] with the Kleene plus as an additional operator. The symbol denotes the null string. An extended context-free grammar G is specified by a tuple (N; \Sigma; and \Sigma are disjoint finite alphabets of nonterminal symbols and terminal symbols, re- spectively, P is a finite set of production schemas, and the nonterminal S is the sentence symbol. Each production schema has the form A!EA , where A is a nonterminal and EA is a regular expression over that does not contain the empty-set symbol. When the string fi 1 fffi 2 can be derived from the string fi and we denote this fact by writing fi)fi 1 fffi 2 . The language L(G) of an extended context-free grammar G is the set of terminal strings derivable from the sentence symbol of G. Formally, denotes the transitive closure of the derivability relation. Even though a production schema may correspond to an infinite number of ordinary context-free productions, it is known that extended and standard context-free grammars describe exactly the same languages; for example, see the texts of Salomaa [23] and of Wood [25]. We denote the size of a regular expression E by jEj and define it as the number of symbols and operators in E. We denote the size of a set A also by jAj. To measure the complexity of any grammatical transformation we need to define the size of a grammar. There are two traditional measures of the size of a context-free grammar that we generalize to extended context-free grammars as follows. Given an extended context-free grammar we define the size jGj of G to be X and we define the norm k G k of G to be Clearly, the norm is a more realistic measure of a grammar's size as it takes into account the size of the encoding of the symbols of the grammar. We use only the size measure however, since the extension of our results to the norm measure is straightforward. 3 Normal-form transformations We introduce the notion of an expression forest that is a tree-based representation for the set of regular expressions that appear as right-hand sides of production schemas. Each production schema's right-hand side is represented as an expression tree in the usual way, internal nodes are labeled with operators and external nodes are labeled with symbols. In addition, we represent the nonterminal left-hand side of a production schema with a single node labeled with that nonterminal. The node also references the root of the expression tree of its corresponding right-hand side. In Fig. 3, we give an example forest of two regular expressions. Since an extended context-free grammar has a number of production schemas that are regular expressions, we represent such grammars as an expression forest, where each tree in the forest corresponds to one production schema and each tree is named by its corresponding nonterminal. (The naming avoids the tree repetition problem.) We now add threads such that the thread for symbol X connects all appearances of the symbol X in the expression forest. 3.1 Reachability and usefulness Recall that a symbol X is reachable if it appears in some string derived from the sentence symbol; that is, if there is a derivation S) ffXfi where ff and fi are (possibly null) strings As in standard context-free grammars, reachable symbols can be easily identified by means of a digraph traversal. More precisely, we construct a digraph whose vertices are symbols in N [ \Sigma and there is an edge from A to B if and only if B labels an external node of the expression tree named A. (We assume that the production schemas do not contain a \Theta \Theta A AA \Theta \Theta a \Theta \Theta A \Theta \Theta \Omega \Omega \Omega \Omega J JJ A \Theta \Theta \Theta \Theta a Figure 1: An expression forest for the extended context-free grammar with production schemas: We have omitted the symbol threads for clarity. the empty-set symbol.) Then, a depth-first traversal of this digraph starting from S gives all reachable symbols of the grammar. The times taken by the digraph construction and traversal are both linear in the size of the grammar. A nonterminal symbol A is useful if there is a derivation ff is a terminal string. The set of useful symbols can be computed recursively as follows. Compute B such that L(EB ) contains a string of terminal symbols (possibly the null string). All such are useful symbols. Then, a symbol A is useful if L(EA ) contains a string of terminals and the currently detected reachable symbols, and so on until no newly useful symbols are identified. We can formalize this inductive process with a marking algorithm such as described by Wood [24] for context-free grammars. The major difference between previous work and the approach taken here is that we want to obtain an efficient algorithm. Yehudai [26] designed an efficient algorithm for determining usefulness for context-free grammars; our approach can be viewed as a generalization of his algorithm. To explain the marking algorithm, we assume that we have one bit available at each node of the expression forest to indicate the marking. We initialize these bits in a preorder traversal of the forest as follows: The bits of all nodes are set to zero (unmarked) except for nodes that are labeled with a Kleene star symbol, a terminal symbol or a null-string symbol-the bits of these nodes are set to one (marked). In the algorithm, whenever a node u is marked, it is useful and it satisfies the condition: The language of the subtree rooted at u contains a string that is completely marked. A Kleene-star node is marked since its subtree's language contains the null string; that is, a Kleene-star node is always useful. After completing the initial marking, we bubble markings up the trees in a propagation phase as follows: Repeatedly examine newly marked nodes as follows until no newly marked nodes are obtained. For each newly marked node u, where p(u) is u's parent if u is not the root, perform one of the following actions: if p(u) is a plus node and p(u) is not marked, then mark p(u). if p(u) is a dot node, p(u) is not marked and u's sibling is marked, then mark p(u). if p(u) is a Kleene-plus node, then mark p(u). if p(u) is a Kleene-star node, it is already marked. if u is a root node and the expression tree's nonterminal symbol is not marked, then mark the expression tree's nonterminal symbol. If there are newly marked nonterminals after this initial round, then we mark all their appearances in the expression forest and repeat the propagation phase which bubbles the markings of newly marked symbols up the trees. If there are no newly marked nonterminals, then the algorithm terminates. The algorithm has, therefore, a number of rounds and at the beginning of each round it marks all appearances of newly discovered useful nonterminals (discovered in the previous round) and then bubbles the newly marked nonterminals up the trees. As long as a round marks new nodes, the propagation process is repeated. To implement this process efficiently, we construct, at the beginning of each round, a queue of newly marked nodes. Note that the queue is a catenation of appearance lists after the first round. The algorithm then repeatedly deletes a newly marked node from the queue and, using the preceding propagation rules, it may also add newly marked nodes to the queue. A round terminates when the queue is empty. Observe that each node of the expression forest is visited at most twice because a dot node can be visited twice. Thus, the marking algorithm runs in O(jGj) time and space. Recall that a grammar G is reduced if all its symbols are both useful and reachable. As for standard context-free grammars, to reduce a grammar we first identify all useful symbols and then select (together with the corresponding schemas) those that are also reachable. More formally, after identifying the useless nonterminals (terminals are always useful), we first remove their production schemas from G. Second, we remove all productions (not schemas) that contain a useless nonterminal in their right-hand sides. In both steps we have to ensure that the threads are maintained correctly. In the first step, we need not only to remove the production schemas, but also to reconnect the threads of of all symbol appearances that are removed. We can use a traversal of each schema to identify the symbols in it and remove their appearances from the appropriate threads. In the second step, we use the threads of useless symbols to remove the corresponding productions. We simply replace each useless symbol with the empty-set symbol and remove it from its thread, and then apply the empty-set removal algorithm for regular expressions to each production schema. Thus, we obtain the equivalent grammar We next identify the unreachable symbols of G and then remove the production schemas of the unreachable nonterminals and, once more, maintain the threads correctly. Observe that an unreachable terminal symbol can only appear in production schemas of unreachable nonterminals and that reachable symbols can only appear in production schemas of reachable nonterminals. Thus, we obtain G 0 from G in linear time. We summarize the result of this section as follows. Theorem 1 Let be an extended context-free grammar represented as an expression forest. Then, an equivalent, reduced extended context-free grammar G can be constructed from G in time O(jGj) such that jG represented as an expression forest. 3.2 Null-free form Given a reduced extended context-free grammar S), we can determine the nullable nonterminals (the ones that derive the null string) using a similar algorithm to the one we used for usefulness in Section 3.1. This algorithm takes O(jGj) time. Given the nullability information we can then make the given grammar null free in two steps. First, we replace all appearances of each nullable symbol A with the regular expression This step takes time proportional to the total number of appearances of nullable symbols in G-we use the symbol threads for fast access to them. Second, we transform each production schema A!EA , where A is nullable, into a null-free production schema A , where 62 L(E 0 A ). Unfortunately, this step can take time O(2 jGj ) in the worst case when we use the typical textbook algorithm and each production schema has nested dotted subexpressions in which each operand of the dot can produce the null string. We replace each dotted subexpression F \Delta G with is the transformed version of F that does not produce the null string. Note that we at least double the length of the dotted subexpressions. Because similar doubling can occur in the subexpressions of F and G and of their subexpressions, we obtain the exponential worst-case bound. (Note that this is the same case that occurs with a standard context-free grammar in which every nonterminal is nullable.) We want, however, to obtain at most a linear blowup in the size of the resulting grammar. Since nested dot expressions cause the nonlinearity, we modify the grammar to remove nested dot expressions. This approach was first suggested by Yehudai [13, 26] for standard context-free grammars-he converted a given grammar into Chomsky normal form to avoid the dot problem. We take a similar approach by removing nested dot, Kleene-plus and Kleene-star subexpressions from production schemas. The removal generates new nonterminals and their production schemas; however, the size of the resulting grammar is only linearly larger than the original grammar. We replace each dot, Kleene-plus and Kleene-star node of an expression tree that has a dot, Kleene-plus or Kleene-star ancestor with a new nonterminal and add a new production schema to the grammar. We repeat this local modification until no such nested nodes exist. For example, given the production schema we can replace it with the new production schemas: and Repeating the transformation for B, we obtain Repeating the transformation for A, we obtain A!D and We say that the resulting grammar is in dot normal form. Its size is of the same order as the original size and the number of nonterminals is increased by at most the size of the original grammar. be a reduced, extended context-free grammar represented as an expression forest. Then, an equivalent, reduced extended context-free grammar G in dot normal form can be constructed from G in time O(jGj) such that jG 0 j is O(jGj), jN 0 j is O(jGj) and jP 0 j is O(jGj). Moreover, G 0 is also represented as an expression forest. We now apply the simple null-removal construction on a grammar G in dot normal form to produce a new grammar that has size O(jGj). The algorithm runs in time O(jGj). Theorem 3 Let be a reduced, extended context-free grammar in dot normal form represented as an expression forest. Then, an equivalent, reduced, null-free extended context-free in dot normal form can be constructed from G in time O(jGj) such that jG 0 j is O(jGj), jN 0 j is O(jN j) and jP 0 j is O(jP j). Moreover, G 0 is also represented as an expression forest. 3.3 Unit-free form A unit production is a production of the form A!B. We transform an extended context-free grammar into unit-free form in three steps. First, we identify all instances of unit productions. Second, we remove each unit-production instance from its schema. Third and last, for each modified schema, we add the unit-free schemas of the unit-production instances to the modified schema. We now discuss these three steps in more detail. We assume that each reduced, null- free, extended context-free grammar G, is also in dot normal form. To identify instances of unit productions observe that, for each schema EA , each root-to-frontier path contains at most one dot or Kleene-plus node, and no Kleene-star nodes. Now, assume that there is a unit-production instance of B in EA (that is, A!B is in A!EA ). Immediately, none of B's ancestors can be dot nodes; an ancestor can be a plus node and at most one ancestor can be a Kleene-plus node. To identify unit productions, we carry out a preorder traversal of EA and identify root-to-frontier paths that satisfy the necessary conditions for unit-production instances and also have a nonterminal at their frontier nodes. This step takes O(jEA Whenever the traversal meets a Kleene-plus node or a plus node it continues the traversal in the corresponding subtrees. When it meets a dot node it terminates that part of the traversal. When eventually the traversal reaches a node labeled with a nonterminal B, then that occurrence of B corresponds to a unit production for A. The overall running time for the first step is O(jGj). Second, we remove the instances of unit productions from their schemas. That is, we transform each production schema A!EA into a production schema A!E 0 A such that We define a new threading, which we refer to as the unit thread that connects all occurrences of nonterminals that correspond to unit productions in the schemas. The threading can be constructed during the identification step but it is used in the second step. Furthermore, while identifying unit productions, we determine, for each nonterminal A, the set UA of nonterminals that are unit reachable from A. (Note that UA may contain A.) We use these sets to modify the production schemas in the third step. We traverse the expression trees from their frontiers to their roots and, in particular, we follow the paths that start from the nodes labeled with nonterminals that correspond to unit productions (we access them by following the unit threads). Assume that we are removing an instance of B. Then, its ancestors are only plus nodes with the possible exception that one ancestor is a Kleene-plus node. To remove unit appearances from Kleene-plus subtrees, we globally transform all Kleene- plus subexpressions of the form F + in the expression forest into )). The idea behind this global transformation is that we have separated the unit appearances in F from the non-unit appearances of the same symbols in F + , since the unit appearances now occur only in the first F and the non-unit appearances of the same symbols appear in the subexpression node u is a Kleene-plus node in some expression tree, then we make two copies of u's only subtree R (we call them S and T ) and ensure we maintain all threads in S and T except for the unit threads. We then remove the Kleene-plus node and reconnect R, S and T as (R The removal of all unit appearances of each nonterminal B is now straightforward. We arrive at a node labeled B by following the unit thread and we replace B and B's parent with B's sibling and terminate the process for this occurrence of B. The only case we have not covered is when A!B is the only production for A. In this case, B has no parent; therefore, we are left, temporarily, with an empty expression tree for A. (Note that B 6= A since A is useful.) The time complexity of this second step is the same as that of the first step. Third and last, we modify the production schemas such that, for each nonterminal A, if are the nonterminal symbols that are unit reachable from A that do not include A, then the new production schema for A is The resulting grammar has size O(jGj 2 ), a quadratic blow up, since we must make copies of the subtrees to give an expression tree for A. The algorithm takes, therefore, O(jGj 2 ) time. Theorem 4 Let be a reduced, null-free extended context-free grammar in dot normal form that is represented as an expression forest. Then, an equivalent, reduced, dot-normal-form, null-free, unit-free extended context-free grammar G be constructed from G in time O(jGj 2 ) such that jG 0 j is O(jGj 2 O(jGj). Moreover, G 0 is also represented as an expression forest. Note that we can ensure that the blow up is linear, if we do not make multiple copies of the various subtrees, but we merely provide links to one copy of each distinct subtree. This approach takes O(jN space to the grammar G. 3.4 Greibach form This normal form result for context-free grammars was established by Sheila Greibach [10] in the 1960's; it was a key result in the use of the multiple-path syntactic analyzer developed at Harvard University at that time. An extended context-free grammar is in Greibach normal form if its productions are of only the following form: where a is a terminal symbol and ff is a possibly empty string of nonterminal symbols. The transformation of an extended context-free grammar into Greibach normal form requires two giant steps: left-recursion removal and back left substitution. Recall that a grammar is left recursive if there is a nonterminal A such that in the grammar, for some string ff. We consider the second step first. Assume that the given extended context-free grammar factored if, for each nonterminal A, a string x in L(EA ) is either completely nonterminal or it is a single terminal symbol. It is straightforward to factor a grammar and if we do it before we make the grammar null free, we avoid the possible introduction of unit productions.) In addition, for the second step we also assume that the grammar is non-left recursive. Since the grammar is non-left recursive there is a partial order on the nonterminals, left reachability, that is defined by A B if there is a leftmost derivation As usual, we can consider the nonterminals to be enumerated as A such that whenever A i A j , then i j. Observe that A n is already in Greibach normal form since it has only productions of the form A n !a, where a 2 \Sigma. We now convert the nonterminals one at a time from A n\Gamma1 down to A 1 . The conversion is conceptually simple, yet computational expensive. When converting A i , we replace all nonterminals that can appear in the first positions in the strings in L(EA i schemas. Thus, the resulting schema A i !E 0 is now in Greibach normal form. To be able to carry out this substitution efficiently we first convert each schema EA i into first normal form; that is, we express each schema as the sum of regular expressions each of which begins with a unique symbol. More precisely, letting and using the notation E i instead of EA i , for simplicity, we replace which is defined as follows: n+m \Delta a n+m ; The conversion must into an equivalent regular expression in Greibach normal form, we need only replace the first A k of each term A k k . If the grammar is left recursive, we first need to make it non-left recursive. We use a technique introduced by Greibach [11], investigated in detail by Hotz and his co-workers [14, 21, 22] and rediscovered by others [7]. It involves producing, for each nonterminal, a distinct subgrammar of G that is left linear; hence, it can be converted into an equivalent right linear grammar. This conversion changes left recursion into right recursion and does not introduce any new left recursion. For more details, see Wood's text [25]. The important property of the left-linear subgrammars is that every sentential leftmost derivation sequence in G can be mimicked by a sequence of leftmost derivation sequences, each of which is a sentential leftmost derivation sequence in one of the left-linear subgrammars. Once we convert the left-linear grammars into right-linear grammars this property is weakened in that we mimic the original derivation sequence with a sequence of sentential rightmost derivation sequences in the right-linear grammars. The new grammar that is equivalent to G is the collection of the distinct right-linear grammars, one for each nonterminal in G. As the modified grammar is no longer left recursive, we can now apply back left substitution to obtain a final grammar in Greibach normal form. How well does this algorithm perform? Left recursion removal causes a blow up of jN jjGj at worst. Converting the production schemas into first normal form causes an additional blow up of jN jjGj. We use the derivative dE dX of a regular expression E by a symbol X to give a new expression F such that L(F L(E). The derivative of a regular expression was introduced by Brzozowski [3] who defined it inductively. Now, given a schema EA , we obtain its derivatives for each symbol X 2 N [ \Sigma. When we catenate X with its derivative we obtain one of the terms in the first normal form. Since G is null free, the only derivative that can cause exponential blow up is dF dX dX We transform G such that no Kleene-plus subexpression is nested within any other Kleene- plus expression-a similar transformation to the one we used for conversion to dot normal form. This modification ensures that exponential blow up does not occur. The back left substitution causes, in the worst case, an additional blow up of jN jjGj in the size of the Greibach normal form grammar. As all three operations take time proportional to the sizes of their output grammars essentially, the transformation to Greibach normal form takes O(jN in the worst case. The reason for the jN j 5 term is that we first remove left recursion which not only increases the size of the grammar but also squares the number of nonterminals from jN j to . The number of nonterminals is crucial in the size bound for the grammar obtained by first normal form conversion and by back left substitution. We can however, reduce the worst-case time and space by using indirection as we did for unit-production removal. Rather than performing the back left substitution for a specific nonterminal, we use a reference to its schema. This technique gives a blowup of only jGj+jN j 2 at most; thus, it reduces the complete conversion time and size to O(jN j 3 jGj) in the worst case. We may also apply the technique that Koch and Blum [18] suggested; namely, leave unit- production removal until after we have obtained a Greibach-like normal form. Moreover, transforming an extended context-free grammar into dot normal form appears to be a very useful technique to avoid undesirable blow up in grammar size. We are currently investigating this and other approaches. The results that we have presented are truly a generalization of similar results for context-free grammars. The time and space bounds are similar when relativized to the grammar sizes. The novelty of the algorithms is four-fold. First, we have introduced the regular expression forest with symbol threads as an efficient data representation for context-free grammars and extended context-free grammars. We believe that this representation is new. The only previously documented representations are those of Cohen and Gotlieb [5] and of Barnes [2] and they are more simplistic. Second, we have demonstrated how indirection using referencing can save time and space in the null-production removal and back left substitution algorithms. Although the use of the technique is novel in this context, it is well known technique in other applications. It is an application of lazy evaluation or evaluation on demand. Third, we have introduced dot normal form for extended context-free grammars that plays a role similar to normal form for standard context-free grammars. Fourth, we have generalized the left-linear to right-linear grammatical conversion for extended grammars. We are currently investigating whether we can obtain Greibach normal form more efficiently and whether we can improve the performance of unit-production removal. We would like to mention some other applications of the regular expression forest with threads. First, we can reduce usefulness determination to nullability determination. Given an extended context-free grammar S), we can replace every appearance of every terminal symbol with the null string to give G nonterminal A in G is useful if and only if it is nullable in G 0 . Second, we can use the same algorithm to determine the length of the shortest terminal strings generated by each nonterminal symbol. The idea is that we replace each appearance of a terminal symbol with the integer 1 and each appearance of the null string with 0. We then repeatedly replace: each node labeled "+" that has two integer children with the minimum of the two integers; each node labeled "\Delta" that has two integer children with the sum of the two integers; and each node labeled " " with 0. The root value is the required length. We can use the same "generic" algorithm to compute the smallest terminal alphabet for the terminal strings derived from a nonterminal, the LL(1) first and follow sets, and so on. Last, the careful reader will have observed that the space-efficient algorithms for unit freeness and Greibach normal form produce output grammars that are not represented as expression forests-they are represented as a set of expression dags (directed acyclic graphs). The dags have as many roots as there are nonterminals. Not surprisingly, each root-to-frontier traversal yields a tree since we have reduced space by sharing common subtrees among trees in the underlying expression forest. Clearly, we may also share common subtrees within the original trees in the expression forest, although we do not know of any "practical" computation that would benefit from such sharing. We are currently investigating the complexity of the transformations we have discussed when we are given a collection of expressions dags as the representation of an extended grammar. Although, a collection of dags is a dag, the dags we are dealing with have three properties. First, a traversal from any root node yields a tree that corresponds to a production schema; second, there are as many roots as there are nonterminals; and, third, the dags are threaded. For this reason, we call such a collection of expression dags, a dagwood 1 . --R The Theory of Parsing Exploratory steps towards a grammatical manipulation package (GRAMPA). Derivatives of regular expressions. Structural equivalence of extended context-free and extended E0L grammars A list structure form of grammars for syntactic analysis. W3C web page on XML. An easy proof of Greibach normal form. A system for manipulating polynomials given by straight-line programs Transition diagram systems and normal form trans- formations A new normal form theorem for context-free phrase structure grammars A simple proof of the standard-form theorem for context-free grammars Introduction to Formal Language Theory. Eliminating null rules in linear time. ISO 8879: Information processing-Text and office systems-Standard Generalized Markup Language (SGML) SGML and exceptions. Greibach normal form transformation revisited. Grail: Engineering automata in C Grammar Transformations Based on Regular Decompositions of Context-Free Derivations A general Greibach normal form transforma- tion Formal Languages. Theory of Computation. Theory of Computation. On the Complexity of Grammar and Language Problems. --TR An easy proof of Greibach normal form Formal languages <italic>Grail</italic>: a C++ library for automata and expressions Dagwood Derivatives of Regular Expressions A New Normal-Form Theorem for Context-Free Phrase Structure Grammars A List Structure Form of Grammars for Syntactic Analysis Theory of Computation Introduction to Formal Language Theory The Theory of Parsing, Translation, and Compiling SGML and Exceptions Greibach Normal Form Transformation, Revisited On the complexity of grammar and language problems. Grammar transformations based on regular decompositions of context-free derivations. --CTR Marcelo Arenas , Leonid Libkin, A normal form for XML documents, Proceedings of the twenty-first ACM SIGMOD-SIGACT-SIGART symposium on Principles of database systems, June 03-05, 2002, Madison, Wisconsin Frank Neven, Attribute grammars for unranked trees as a query language for structured documents, Journal of Computer and System Sciences, v.70 n.2, p.221-257, March 2005 Marcelo Arenas , Leonid Libkin, A normal form for XML documents, ACM Transactions on Database Systems (TODS), v.29 n.1, p.195-232, March 2004

grammatical representations;complexity;efficient algorithms;normal forms;symbolic manipulation;extended context-free grammars

504540

Using acceptors as transducers.

We wish to use a given nondeterministic two-way multi-tape acceptor as a transducer by supplying the contents for only some of its input tapes, and asking it to generate the missing contents for the other tapes. We provide here an algorithm for assuring beforehand that this transduction always results in a finite set of answers. We also develop an algorithm for evaluating these answers whenever the previous algorithm indicated their finiteness. Furthermore, our algorithms can be used for speeding up the simulation of these acceptors even when not used as transducers. Copyright 2001 Elsevier Science B.V.

Introduction In this paper we study the following problem: assume that we are given a nondeterministic two-way multi-tape acceptor A and a subset X of its tapes. We would like to use A no longer as an acceptor which receives input on all its tapes, but instead as a kind of transducer [15, Chapter 2.7] which receives input on tapes X only and generates as output the set of missing inputs onto the other tapes. We then face the following two problems: Problem 1 Can it be guaranteed that given any choice of input strings for tapes X the set of corresponding outputs of A will always remain finite? Problem 2 In those cases where Problem 1 can be solved positively, how can the actual set of outputs corresponding to a given choice of input strings be Supported by the Academy of Finland, grant number 42977. To appear in the Theoretical Computer Science special issue on the Third International Workshop on Implementing Automata (WIA'98). Preprint submitted to Elsevier Preprint 18th January 2000 The motivation for studying these two problems came from string databases [3,7,11] which manipulate strings instead of indivisible atomic entities. Such databases are of interest for example in bioinformatics, because they allow the direct representation and manipulation of the stored nucleotide (DNA or sequences. While one can base a query language for these databases on a fixed set of sequence comparison predicates, such as in for example the PROXIMAL system [5], it would be more flexible to allow user-defined string processing predicates as well. If we assume an SQL-like notation [1, Chapter 7.1] for the query language, then one possible query for such a string database might be stated as follows. Here # rev user-defined expression which compares the strings w 1 and w 2 denoted by the variables x 1 and x 2 , say "w 2 is the reversal of w 1 ". Then this query requests every string w 2 that is the reversal of some string w 1 currently stored in the database table R. Note in particular that these strings w 2 need (and in general can) not be stored anywhere in the database; the query evaluation machinery must generate them instead as needed. We have developed elsewhere [10,11,17] a logical framework for such a query language. This framework accommodates expressions like # rev multidimensional extension of the modal Extended Temporal Logic suggested by Wolper [29]. The multi-tape acceptors studied here are exactly the computational counterparts to these logical expressions. A given query to a database is considered to be "safe" for execution if there is a way to evaluate its answer finitely [1, pages 75-77]. One safe plan for evaluating the aforementioned query would be as follows, where is the string relation accepted by A rev , a multi-tape acceptor corresponding to the expression such acceptor is shown as Figure 1 below.) for all strings w 1 in table R do output every string in V end for Our two problems stem from these safe evaluation plans. Problem 1 is "How could we infer from # rev that the set V is always going to be finite for every string w 1 that could possibly appear in R?" Problem 2 is in turn "Once the finiteness of every possible V has been ensured, how can we simulate this A rev (e#ciently) for each w 1 to generate the V corresponding to this We have studied elsewhere [9][17, Chapter 4.4] how solutions for Problem 1 can be used to guide the selection of appropriate safe execution plans. To this end, Section 1.2 presents the problem in not only automata but also database theoretic terms. One possible solution would have been to restrict beforehand the language for the string handling expressions such as # rev into one which ensures this finiteness by definition, say by fixing x 1 to be the input variable, which is mapped into the output variable x 2 as a kind of transduction [3,7]. However, in logic-based data models [1], the use of transducers seems less natural than acceptors, because the concept of information flow from input to output is alien to the logical level, and of interest only in the query evaluation level. But we must eventually also evaluate our string database queries, and then we must infer which of our acceptors can be used as transducers, and how to perform these inferred transductions, and thus we face the aforementioned problems. The rest of this paper is organized as follows. Section 1.1 presents the acceptors we wish to use as transducers, while Section 1.2 formalizes Problem 1. Section 2 first reviews what is already known about its decidability, and then presents our algorithms, which give su#cient conditions for answering Problem 1 in the a#rmative. Section 3 presents then an explicit evaluation method for those acceptors that these algorithms accept, answering Problem 2 in turn. Finally, Section 4 concludes our presentation. 1.1 The Automaton Model Let the alphabet # be a finite set of characters fixed in advance, let # denote the set of all finite sequences of these characters, let # denote the empty sequence of this kind, and let w t denote concatenating t # N copies of w # , as usual. We shall study relations on these sequences, or strings drawn from # in what follows. On the other hand, database theory often studies sequence database models where # is taken to be conceptually infinite instead, as in for example [7,19,22,24,25]. Then the emphasis is on data given as lists of data items provided by the user. Conversely, our emphasis is on data given as strings in an alphabet fixed beforehand by the database designer. In other words, our approach fixes an appropriate alphabet # as part of the database schema, while the list approach considers # as part of the data instead. However, our approach has been employed even for managing list data [2]. We furthermore assume left and right tape end-markers '[' and `]' not in #. Then we define the n th character of a given string w # with length as | {z } Intuitively our automaton model is a "two-way multi-tape nondeterministic finite state automaton with end-markers"; similar devices have been studied by for example Harrison and Ibarra [14] and Rajlich [20]. Formally, a k-tape Finite State Automaton (k-FSA) [11, Section 3][17, Chapter 3.1] is a tuple (1) # is the finite alphabet as explained above; is the number of tapes; (3) QA is a finite set of states; QA is a distinguished start state; QA is the set of final states; and is a set of transitions of the form p c1 ,. ,c k q, where p, q # QA , each and each d i # {-1, 0, +1}. We moreover require that d ensures that the heads do indeed stay within the tape area limited by these end-markers. A configuration of A on input # is of the form corresponds intuitively to the situation, where A is in state p, and each head scanning square number n i of the tape containing string w i . Hence we say that is a possible next configuration of C if and only if p w1 [n1 ],. ,w k [n k ] -# d1 ,. ,d k Now +1 can be interpreted as "read forward", while -1 means "rewind the tape to read the preceding square again", and 0 "stand still". We call tape i of A unidirectional if no transition in specifies direction -1 for it; otherwise tape i is called bidirectional instead. A computation of A on input # w is a sequence . of these config- urations, which starts with the initial configuration C each C j+1 is a possible next configuration of the preceding configuration C j . This computation C is accepting if and only if it is finite, its last configuration C f has no possible next configurations, and the state of this C f belongs to FA . The language L(A) accepted by A consists of those inputs # w, for which there exists an accepting computation C. Note that this language is a k-fold relation on strings in the general case. Because A is nondeterministic, we can without loss of generality assume that no transitions leave the final states FA . We can for example introduce a new III a, a Figure 1. A 2-FSA for recognizing strings and their reversals. state f A into QA , and set for every state p previously in FA , and every character combination c 1 , . , c k # {[, ]}, on which there is no transition leaving p, we add the transition p c 1 ,. ,c k -# 0,. ,0 f A . In this way, whenever a computation of A would halt in state p, it performs instead one extra transition into the (now unique) new final state f A , and halts there instead. We often view A as a transition graph GA with nodes QA and edges In particular, a computation of A can be considered to trace a path P within GA starting from node s A . It is furthermore expedient to restrict attention to non-redundant A where each state is either s A itself or on some path P from it into some state in FA . Figure 1 presents a 2-FSA A rev in this form b}. The language accepted by it consists of the pairs string v is the reversal of string u: looping in state II finds the right end of the bidirectional tape 1 without moving the unidirectional tape 2, while looping in state III compares the contents of these two tapes in opposite directions. Another often useful simplification is the following way to detect mutually incompatible transition pairs. of k-FSA A is locally consistent if and only if every consecutive pair ,. ,c # k r of transitions in satisfies the condition This ensures that there are configurations, in which this pair can indeed be taken; whether these configurations do ever occur in any computation is quite another matter. For example, both tapes in Figure 1 are locally consistent. If in particular tape i is both unidirectional and locally consistent, then given any path c (1,1) ,. ,c (k,1) -# d (1,1) ,. ,d (k,1) c (1,2) ,. ,c (k,2) -# d (1,2) ,. ,d (k,2) c (1,3) ,. ,c (k,3) -# d (1,3) ,. ,d (k,3) in GA we can construct an input string for tape i, which allows P to be followed, if we choose c (i,j) if (d # otherwise. For example in Figure 1 the w 2 from Eq. (2) spells out the sequence of transitions taken when looping in state III. Harrison and Ibarra provide a related construction for deleting unidirectional input tapes from multi-tape pushdown automata [14, Theorem 2.2], while Rajlich [20, Definition 1.1] allows the reading head to scan two adjacent squares at the same time for similar purposes. Again the nondeterminism of A allows us to enforce Definition 1 for tape i at the cost of expanding the size of A by a factor of (|#| a k-FSA B with state space Q which remembers the character under tape head i. Add for each transition p c1 ,. ,c k -# d1 ,. ,d k q the transitions -# d1 ,. ,d k to complete our construction. 1.2 The Limitation Problem This section introduces our limitation problem [11, Definition 3.1][17, Definition 3.3] concerning the automata defined in Section 1.1. determine if there exists a limitation # N with the following property: if #u 1 , . , L(A), then If this is the case, then we say that A satisfies the finiteness dependency [21] These dependencies are a special case of functional dependencies in database theory [1, Section 8.2]. Intuitively, A is a finite representation of the conceptually infinite database table assures that if we select rows from this table by supplying values for the columns 1, . , k, we do always receive a finite answer. In this way A can be used both safely and declaratively as a string processing tool within our string database model. Thus our goal is to treat the (user-defined) string processing operation A as just another relation as far as the database query language is concerned; such transparency is in fact being advocated for the forthcoming object/relational database proposal [4, pages 49-55]. We discuss elsewhere [12] how this overall string processing mechanism of ours relates to this proposal and how it could be incorporated into such database management systems. In terms of automata theory we require that for any input #u 1 , . , the possible outputs #v 1 , . , v l # ) l must remain a finite set. This is what is meant by "using acceptors as transducers": we supply strings for only some tapes (here 1, . , k) of the acceptor A, and ask it to produce us all those contents for the missing tapes (here k would have accepted given the known tape contents. The limitation problem is then to determine beforehand whether this computation will always return a finite result or not. Weber [27,28] has studied a related question whether the set of all possible outputs on any inputs of a given transducer remains finite, and if so, what is the maximal output length. 2 Solving the Limitation Problem The hardness of the limitation problem has been shown to depend crucially on the amount of bidirectional tapes in A. The problem has been shown elsewhere to be undecidable for FSAs with two bidirectional tapes [11, Theorem 5.1][17, Chapter 4.1]: given a Turing machine [15, Chapter 7] M one can write a corresponding 3-FSA AM with two bidirectional tapes, which accepts exactly the tuples #u, v, w#, where v and w together encode a sequence of computation steps taken by M on input u. Here v and w must be read twice, requiring bidirectionality. Then asking whether AM satisfies {1} # {2, 3} amounts to asking whether M is total. This read-twice construction is reminiscent of representing the valid computations of a given Turing machine as an intersection of two Context-Free Languages [15, Chapter 8.6], and shows that it is also undecidable to determine whether a given finiteness dependency is satisfied by the intersection of the relations denoted by two given FSAs, even when these FSAs have no bidirectional tapes at all [17, Corollary 6.1]. On the other hand, the limitation problem becomes decidable if we restrict attention to those FSAs with at most one bidirectional tape [11, Theorem 5.2][17, Chapter 4.2]. Intuitively, all the unidirectional tapes are first made locally con- sistent, after which Eq. (2) allows us to construct their contents at will, so that we can concentrate on the sole bidirectional tape. This tape can in turn be studied by using an extension of the well-known crossing sequence construction [15, Chapter 2.6] for converting two-way finite automata into classical one-way iii iii iii a, a Figure 2. A crossing behavior of the 2-FSA in Figure 1. finite automata. This method is clearly impractical, however. Therefore this paper presents in Section 2.1 a practical partial solution, which furthermore applies even in some cases involving multiple bidirectional tapes. Section 2.2 then develops this solution further to yield yet more explicit limitation information Example 3 The 2-FSA A rev in Figure 1 satisfies both {1} # {2} and {2} # {1} with the same limitation function the reversal of a string is no longer than the string itself. This is moreover decidable, because only tape 1 is bidirectional in A rev . To see how limitation inference proceeds consider Figure 2, which exhibits the crossing behavior of A rev when tape 1 contains the string ab. For example, determining {2} # {1} involves checking that every character written onto the bidirectional output tape 1 is "paid for" by reading something from the unidirectional input tape 2 as well, although this payment may occur much later during the computation; here it occurs when tape 1 is reread in reverse. This can in turn be seen from the automaton B produced by the crossing sequence construction by noting that the loops of B around the repeating crossing sequence indicated in Figure 2 consume tape 2 as well. The 2-FSA A rev is also considered to satisfy the trivial finiteness dependency by definition. On the other hand, A rev does not satisfy # {1, 2}, because 2.1 An Algorithm for Determining Limitation Our technique for solving the limitation problem given in Definition 2 is based on the following two observations. Let A be the l} the finiteness dependency in question. Observation 1 If A accepts some input #w 1 , . , w k+l # with some computation never visits the corresponding right end-marker ']', then A also accepts all the su#xed inputs with the same C. Hence, A cannot satisfy # in this case. Observation 2 If on the other hand every accepting computation of A visits the right end-marker ']' on all output tapes, then the only way A can violate # is by looping while generating output onto some output tape but without "consuming" any of the inputs at the same time - that is, by returning again and again to read the same squares of the input tapes. However, the (un)decidability results mentioned in the beginning of this section indicate that reasoning about actual computations is infeasible. Thus we reason instead about the structure of the transition graph GA . Therefore, instead of Observation 1, the algorithm in Figure 3 merely tests that there is no path P from the start state s A into a final state, which never requires ']' to appear on some output tape, whereas it would have su#ced to show that no such P is ever traversed during any accepting computation. (B denotes the Boolean type with values 0 as 'false' and 1 as `true'.) Similarly, the algorithm in Figure 4 enforces a more stringent condition than Observation 2: every cycle L in GA , during which some output tape is advanced to direction +1, must also move some input tape i into direction 1 but not back into the opposite direction #1. Then this tape i acts as a clock, which eventually terminates the potentially dangerous repetitive traversals of L. Again, if A violates #, then some L # failing this condition must exist in GA , but the converse need not hold, because repetitions of L # need not necessarily occur during any accepting computation of A; Figure 6 presents a loop which seems at first glance to generate arbitrary many copies of character 'a' onto its output tape 2, because it seems to move back and forth on its input tape 1. However, closer scrutiny reveals that this behavior is in fact impossible because the same square on tape 1 must first contain character 'a' in order to get into state q, but later character 'b' in order to get back into state p. Making the tapes locally consistent as in Definition 1 will catch some of these impossible transition sequences, including all cases that arise due to the demands on the contents of the unidirectional tapes. On the other hand, Figure 6 presents a bidirectional tape 1 which is already locally consistent but still im- possible. If there is just one bidirectional tape altogether, then the crossing sequence construction alluded to above in Example 3 can be seen as a method for detecting these impossibilities and eliminating them from further consid- eration. Unfortunately, we have no method of this kind for the general case. The more stringent condition given above is enforced by repeatedly deleting those transitions, which can justifiably be argued not to take part in any loops of the kind mentioned in Observation 2. This technique is related to analyzing the input-output behavior of logic programs [16,26], which analyze the call graph of the given program component by component. However, our technique remains simpler, because our automata are more restricted than general logic programs. function halting( G :transition graph GA of a k-FSA A; X:subset of {1, . , k}):B; 2: for all i # {1, . , k} \ X do 3: H # G without transitions that specify ']' for tape 4: b # b # (H contains no path from s A into any state in FA ) 5: end for return b Figure 3. An algorithm for testing Observation 1. function looping( G:subgraph of GA for a k-FSA A; X:subset of {1, . , k}):B; 2: H 1 , . , Hm # the maximal strongly connected components of G; 3: Delete from G all transitions between di#erent components (a.); 4: for all i # 1, . , m do 5: if some tape j # X winds to direction 1 in H i but not to #1 then Delete from H i all transitions that wind this tape j (b.); 7: d # looping(H i , X) 8: else 9: d # no tape in {1, . , k} \ X winds into direction +1 in H i 12: end for 13: return b Figure 4. An algorithm for testing Observation 2. function limited( A :k-FSA; X:subset of {1, . , k}):B; 1: return halting(G A , X) # looping(G A , X) Figure 5. An algorithm for determining limitation. a, a a, a a, a b, a Figure 6. A loop that cannot be traversed repeatedly. More precisely, the edge deletions made by the algorithm in Figure 4 can be justified as follows. Consider the first call made by the main algorithm in Figure 5. Every loop mentioned in Observation 2 must clearly be contained in some component H i of the entire transition graph of the k-FSA A. (a) A transition between two di#erent strongly connected components cannot then surely belong to any loop of this kind. The deletions in step 3 are therefore warranted. (b) Any transition # that winds the clock tape j selected for the current component H i cannot belong to any loop of this kind either, because # cannot be traversed indefinitely often. These traversals will namely wind the input tape j eventually onto either end-marker, because tape j is not wound into the opposite direction by any other transition # in H i . The deletions in step 6 are therefore warranted as well. This reasoning can then be applied in the subsequent recursive calls on the reduced components H i as well, because we can then assume inductively that the loops broken during the earlier calls could not have been ones mentioned in Observation 2. Formalizing this reasoning shows that the algorithm in Figure 5 is indeed correct as follows. Theorem 4 Let A be a (p and let the algorithm in Figure 5 return 1 on A and {1, . , satisfies {1, . , with the function where Y PROOF. Let us assume that C is an arbitrary computation of A on some input We begin by proving the following two claims about this C which correspond to Observations 1 and 2. If C is accepting, then for every tape p r the computation C takes some transition which requires ']' on tape j. Let otherwise h be a tape which violates this Claim 5. C traces a path through GA from s A into some state in FA . Then step 4 of the algorithm in Figure 3 sets which violates our assumption that the algorithm in Figure returns 1, thus proving this Claim 5. moves to direction +1 more than times during the computation C. Assume to the contrary that some h violates this Claim 6. Post a fence between two adjacent configurations C g and C g+1 in C whenever tape to direction +1. By our contrary assumption, at least l + 1 of these fences are posted. Consider on the other hand two configurations C x and C y of C to have the same color if and only if they share the same state and the same head positions for tapes 1, . , most l of these colors are available, recalling the assumption that tapes 1, . , p are unidirectional. Therefore C must contain two configurations C x and C y which have the same color but are separated by an intervening fence. Consider then the sequence of transitions which transform C x into C y , as a path L within GA . This L forms a closed cycle, and tape heads 1, . , are on the same squares both before and after traversing L, because C x and C y shared a common color. Let us then see which of the steps 3 or 6 of the algorithm in Figure 4 will first delete some transition that belongs to L. It cannot be step 3, because all of L belongs initially to the same maximal strongly connected component. But it cannot be step 6 either, because if L ever moves a tape j # X into some direction 1, it must also move tape j into the opposite direction #1 as well, in order to return its head onto the same square both before and after L. Hence L persists untouched to the very end of the recursion on step 9, and there the presence of the transition of L that crosses the fence between C x and C y yields which subsequently violates our assumption that the algorithm in Figure 5 returns 1, thus proving this Claim 6. Claims 5 and 6 are combined into a proof of the theorem as follows. Assume that #z # L(A); that is, A has some accepting computation C on input #z. It su#ces to show that |w h | # l - 1 for every 1 # h # r. Tape head must cross every border between two adjacent tape squares from left to right, because otherwise C would not meet Claim 5. Claim 6 states in turn that C performs at most l crossings of this kind. This means that tape p contains at most l squares, of which the first and the last are reserved for the end-markers, leaving at most l - 1 squares for the characters of the input string w h . # Example 7 Consider the 2-FSA A rev in Figure 1. The algorithm in Figure 5 can detect that it satisfies {1} # {2} as follows. The algorithm in Figure 3 returns 1, because every path into the final state IV must contain III [,] -# 0,0 IV. Evaluating the algorithm in Figure 4 proceeds in turn as follows. First, all transitions from one state into another are deleted in step 3, leaving only the loops ii iii a, a Figure 7. The division of the 2-FSA in Figure 1 into components. around states II and III. This is depicted in Figure 7, where the components themselves are dotted, and the transitions between them (and thus deleted in step are dashed. These loops are in turn deleted in step 6 when processing the corresponding components, and therefore this function eventually returns 1 as well. However, Theorem 4 provides a rather imprecise limitation function compared to the one given in Example 3. On the other hand, the algorithm in Figure 5 fails to detect {2} # {1}, which was detected in Example 3: looping in state II advances tape 1 without moving tape 2, and seems therefore dangerous to the algorithm in Figure 4. Intuitively, A rev first guesses nondeterministically some string, and only later verifies its guess against the input. In Example 3, crossing sequences were examined to see that this later checking in state III indeed reduces the acceptable outputs into only finitely many (here just one). Essentially the same limitation function as in Theorem 4 su#ces whenever all of the output tapes r to be limited are unidirectional, even if {1, . , cannot be verified by the algorithm in Figure 5 [9, Theorem 2.1]. This is natural, because the algorithm in Figure 5 ignored the e#ects of moving any output tape p+q+1, . , p+q+r into direction -1. Theorem 8 Let p+1, . , p+q be all the bidirectional tapes in the (p+q+r)- FSA A, and let A satisfy Then is a corresponding limitation function, where g A is as in Theorem 4. PROOF. Consider the proof of Theorem 4, and assume further in Claim 6 that all the output tapes p+q+1, . , p+q+r are made locally consistent as in Definition new assumption introduces the factor (|#|+2) r into W # . With this modification, the original fencing-coloring construction shows that if some accepting computation C on input #z advances some output tape p+q+h more than then the path of transitions taken by this C can be partitioned into three sub-paths KLM where L begins in C x and ends in C y which share the same color and contains a transition # that crosses some fence between C x and C y . However, now A must also accept all the pumped inputs where t # N, and each w (k,J ) denotes the string of characters in those squares of output tape which the head lands during J (']' excluded). This is in e#ect an application of Eq. (1) to the output tapes p+q+1, . , p+ r. The presence of # within L shows that w (h,L) #, and hence # fails by observation 2, thereby proving this modified Claim 6. 5 continues to hold, as reasoned in Observation 1, and the theorem follows again as before. # Turning now to assessing when the algorithm in Figure 5 does detect finiteness dependencies we see that it is successful at least when all tapes are unidirectional Theorem 9 Let A be a non-redundant l)-FSA with all tapes unidirectional and locally consistent, and let the algorithm in Figure 5 return 0 on A and {1, . , k}. Then A does not satisfy PROOF. The non-redundancy of A and the unidirectionality and local consistency of all its tapes imply by Eq. (2) that for every path P in GA we can always find an accepting computation C on some input # traversing P. Letting P then be any subgraph of GA which caused the algorithm in Figure 5 to return 0 yields some C whose existence violates # along the lines of Theorem 8. # 2.2 Two Variants of the Limitation Algorithm This section explores two possible directions into which the algorithm given in Section 2.1 could be developed further. They both alter the non-recursive step 9 of the algorithm in Figure 4, which tests some strongly connected component H i in the transition graph GA of A. Moreover, this H i is known not to be shrinkable further by the algorithm. The first direction enlarges the set of FSAs A and finiteness dependencies # that can be verified to hold by relaxing this test as follows. Suppose H i contained some output tape j # {1, . , k} \ X that is wound into the reverse if some tape j # X winds to direction 1 in H i but not in direction #1 then Delete from H i all transitions that wind this tape j (b.); else if some tape j # {1, . , k} \ X winds to direction -1 in H i but not in direction +1 then Delete from H i all transitions that wind this tape j; else d # no tape in {1, . , k} \ X winds into direction +1 in H i Figure 8. The enlarging additions to the algorithm in Figure 4. direction -1 but not into the forward direction +1. Then this tape j can again be used as a clock for shrinking H i further, similarly to steps 5-7, because the head on tape j cannot travel backwards forever, but must stop at least once the left end-marker '[' is reached. Algorithmically this direction leads to replacing steps 5-10 of the algorithm in Figure 4 with the steps given in Figure 8. The other direction into which the algorithm given in Section 2.1 can be developed is to constrain further the set of FSAs A and finiteness dependencies # that can be verified to hold by restricting the test performed on step 9 of the algorithm in Figure 4 to require that the component H i being tested must not have any transitions left. Call the correspondingly modified limitation algorithm of Section 2.1 fastidious; it thus requires that all the transitions of A are deleted in order to verify #. Example 10 The calculations in Example 7 show that the 2-FSA A rev in Figure does actually satisfy the finiteness dependency {1} # {2} even fas- tidiously. One advantage of fastidious verification is more e#cient simulation of A, as will be explained in Section 3.1. The rest of this section explains another ad- vantage, namely that it enables straightforward construction of better explicit limitation functions than the generic one provided by Theorem 4. This construction is similar in spirit to van Gelder's analysis of logic programs with systems of equations [26], except that recurrences are used instead. Let therefore A be a satisfying the finiteness dependency suitable limitation function would evidently be if each auxiliary function f p limit to the character position on output tape k+j where the right end-marker ']' is encountered. Here A is assumed to be in state p # QA , each of its input tapes of an unspecified input string with length m i # N and its designated output tape k + j on character h of some unspecified output string. These auxiliary functions can in turn be obtained by labeling the transition graph GA of A with suitable expressions as follows: . The expression for a state p # QA is the maximum of the expressions for those transitions # that leave from p: leaves from p Graphically speaking, one can consider each node p of graph GA to become labeled with the operator 'max' applied to the arrows that exit from p. . The expression for a transition # is the expression for the state that # enters, adjusted with the e#ects of # on the tapes in question: f where Here the "otherwise" branch of Eq. (5) denotes the case when the transition cannot apply by virtue of some input tape head position n i violating its corresponding restriction # i . Then the value 1 is warranted, because it is the earliest possible position in which ']' can possibly appear. Graphically speaking, this shows how to construct the expressions for the arrows maximized by node p of graph GA in Eq. (4) above. Fastidiousness guarantees that this labeling does indeed yield a function: otherwise the expression for some f p refers (indirectly) back to itself with no change to its arguments n 1 , . , n k . This, however, implies a cycle C in GA from p back to itself for which none of the input tapes 1, . , k moves in exactly one direction. This in turn means that C could not be deleted by the fastidious limitation algorithm after all. Accordingly, if a fastidious version of the algorithm in Figure 8 is used instead, the labeling must then include all of h 1 , . , h l instead of just h, because then also the output tapes k+1, . , k+l may have been used in deleting transitions from the (k in question, and thus these output tape head positions might be the ones that cannot repeat as arguments for some expression. On the other hand, dropping the fastidiousness restriction does not altogether invalidate this approach either, it just makes it more di#cult to provide an explicit counterpart for Eq. (4). It is namely now possible that the expression for some f p does indeed refer back to itself without changing its arguments n 1 , . , the expressions f # j of Eq. (5) for the transitions # that still remain in the component containing state p even after the successful execution of the algorithm in Figure 4. However, these expressions are also guaranteed not to increase h, for otherwise the execution of the algorithm in Figure 4 would have been unsuccessful. Thus a function for f p does still exist, even though giving an explicit expression for it is di#cult. A similar labeling technique su#ces even for the crossing sequence construction mentioned in Example 3 instead of a fastidious algorithm from Section 2.1, provided that the labeling is performed relative to the resulting crossing sequence automaton instead of the original [17, Chapter 5.2]. Example 11 The labeling to generate Eq. (3) for Example 10 proceeds as follows. Applying Eq. (5) to the transitions leaving state III leads to f III -# 0,0 IV f III a,a III III b,b III f III which appear in Eq. (4) for state III, namely f III f III -# 0,0 IV III a,a III f III b,b III f III a, a a, a a, a Figure 9. A bidirectional loop that eventually ends. where the last simplification solves the recurrence found above for f III 1 . Continuing similarly for state II leads eventually to the tight limitation function mentioned in Example 3. A di#erent possibility for improving on the limitation algorithm given in Section 2.1 would be to take into account not only the directions but also the total amount of tape movement. For instance, the current algorithms will not break a loop like in Figure 9, because the input tape 1 in question moves in both directions, even though the overall net e#ect of these movements is +1, or to move one square forward, and therefore the loop cannot execute indefi- nitely. Calculating such net e#ects have been recently studied in [18], but the resulting algorithms di#er significantly from the approach presented here. 3 Evaluation of the Limited Answers After inferring that the given (k does indeed satisfy the given finiteness dependency want to generate the (finite) set of outputs #, or to solve Problem 2. This problem is known to be di#cult in the general case: let B be a 2-FSA with an unidirectional input tape 1 and a bidirectional output tape 2, and ask if a given input u can produce any output v. This problem is equivalent to whether B, considered as a checking stack automaton, accepts u [20, Theorem 5.1] which is known to be either PSPACE- or NP-complete, depending on whether B # is a part of the instance or not [6, Problem AL5]. However, the additional information # provides certain optimization possibilities. A straightforward way to obtain an evaluation algorithm is to convert the output tapes from read-only into write-once, and perform these writing operations concurrently with the simulation of the nondeterministic control. Figure shows the resulting algorithm where the simulations of all the possible computations are performed in a depth-first order using a stack S. The algorithm maintains for each 1 # j # l an extensible character array W j [0 . L j which holds the contents for the tape squares the output head k + j has al- procedure simulate( A :(k 2: 3: 4: Initialize stack S to contain #0; L 1 , . , L l # 5: while S is nonempty do l # be top element in S, and let # t 7: if q # FA then 8: output(v 1 , . , v l ) where each 10: else if t > |T A | then 13: Pop o# the top element from 15: else if # -# d1 ,. ,d k ,e 1 ,. ,e l l on which 19: 21: else 23: end if 24: end while Figure 10. An algorithm for using acceptors as transducers. ready examined during the computation C of A currently being simulated. Figure 11 shows the indices during one simulation of the 2-FSA A rev from Figure 1, where the input tape 1 contains the string ab whose reversal is being generated onto the output tape 2. In Figure 10, is enumerated as # 1 , . , # |T A | , and # 0 is a new starting pseudo- transition -# 0,. ,0 s A . We also assume as in Section 1.1 that no final state in FA has any outgoing transitions. Note that # alone does not guarantee that the algorithm in Figure 10 halts, it just guarantees that only finitely many di#erent outputs are ever generated. Consider namely the situation in Figure 12, where the 2-FSA A rev in Figure 1 output tape 2: [ b iii state: Figure 11. Simulating the 2-FSA in Figure 1 as a transducer. input tape 2: [ b a ] output tape 1: [ a a a a ii state: Figure 12. Generating an indefinitely long output with the 2-FSA in Figure 1. is being used as a transducer in the opposite direction to Figure 11: input is read from tape 2 and written onto tape 1. As explained in Example 7, A rev is now guessing nondeterministically a possible output for later verification against the input. But how long guesses should A rev be allowed to make? This question can be answered by adding the extra conditions for each 1 # j # l into branch 15 of the algorithm in Figure 10 where W is a limitation function corresponding to #. This addition is warranted, because if during the currently simulated computation C some j violates Eq. (6) then more than W(|u 1 | , . , |u k |) characters from # have been output onto tape k+ j. In this case C must be eventually rejecting and can hence be discarded at once without further ado. Now that the L j have been bounded by W the stack S will always contain only finitely many di#erent configurations C of the transducer being simulated on input #u. (These transducer configurations C can be defined in a straightforward way by extending the acceptor configurations defined in Section 1.1 with write-once output tapes.) Although stack S represents these configurations C only implicitly, they can be reconstructed as in branch 10 of the algorithm in Figure 10. However, some of these configurations C can still repeat, because the transducer being simulated can also loop on the already known parts of its tapes without generating new output. Fortunately this looping can be detected and eliminated simply by testing in branch 15 of the algorithm in Figure 10 that the new configuration C new being pushed into stack S does not yet occur in stack S. This is a standard way to avoid repetition during a depth-first search [23, Chapter 3.6]. We have also experimented with comparing C new against all the configurations C encountered so far in the entire search conducted by the algorithm in Figure 10 on the current input #u, but this proved to be extremely ine#cient in practice. Now we have solved Problem 2 by developing a halting variant of the evaluation algorithm in Figure 10. However, this solution su#ers from two drawbacks. Drawback 1 The value of a limitation function is needed in Eq. (6) to estimate - and hopefully tightly - the depth at which ultimately rejecting output- generating computations can be pruned. This could be termed the "compile-time" drawback: the limitation function must be formed when the acceptor is proposed as a possible transducer, while its value is required before each invocation of the simulation algorithm in Figure 10. Drawback 2 The whole stack S must be scanned against repeating configurations C when pushing each new configuration C new in the algorithm in Figure 10. This could in turn be termed the "run-time" drawback, because it adds loop checking overhead the execution of the simulation algorithm in Figure 10. Fortunately both of these drawbacks can be alleviated by considering how # was inferred to hold, as discussed in the remainder of this section. 3.1 When the Limitation Algorithm Succeeds Consider first the case where # was inferred to hold by having the algorithm in Figure return 1 on A and #. Claim 6 in the proof of Theorem 4 shows that every computation C of A is "self-limiting" in the sense that no L j can grow indefinitely. Thus Eq. (6) is not needed after all, thereby alleviating Drawback 1. alleviates also Drawback 2. Two occurrences C x and C y of the same configuration during C have the same color by definition. The proof of the claim shows that C x and C y can only arise by traversing a closed loop L, which is not deleted during the algorithm in Figure 4. We therefore modify this algorithm to mark in A the transitions it considers deleted. Then the algorithm in Figure 10 can stop scanning its stack S as soon as the most recent marked transition is seen. This holds even when the marking has been performed by the enlarged algorithm in Figure 8. This reasoning shows another benefit of the fastidious variant of the algorithm in Figure 4: then every transition gets marked, and therefore scanning the stack S is no longer required at all. That is, the algorithm in Figure 10 su#ces unmodified in this case, and all run-time loop checking overhead has been eliminated. Example 12 Applying this modified marking algorithm in Figure 4 into the 2-FSA A rev in Figure 1 and even fastidiously by Example 10. Then the algorithm in Figure 10 and A rev can generate the reversal of any given string in linear time with respect to its length. In other words, choosing this evaluation strategy leads into an optimal way to perform string reversals. Note finally that this marking technique can also speed up the simulation of those m-FSAs B which are still used as acceptors and not transducers: just compute the marking given by the modified algorithm in Figure 4 with B and {1, . , m} (which yields 1), and use the resulting stack scanning optimization strategy during the simulation of B on any given input #u 1 , . , um #. Again the marks identify transitions under which it is not necessary to look when scanning the stack for repeating configurations during the simulation. 3.2 When All the Outputs are Unidirectional Another strategy related to the one developed in Section 3.1 works when all the output tapes of A are unidirectional and the finiteness dependency # still holds but this fact can no longer be inferred by the algorithm developed in Section 2.1. In this case the proof of Theorem 8 shows that a halting but still correct variant of the simulation algorithm in Figure 10 can be obtained by adding into its branch 15 the extra condition that configurations C x and C y of the same color - in the sense of that proof - may not repeat within any computation C: if the path L of transitions from C x into C y advances any of the unidirectional output tapes r, then this C must be rejecting because it would violate # via Observation 2, while otherwise taking L during C was unnecessary because then C x and C y are the same configuration. This reasoning provides the loop checking discipline which guarantees the halting of the simulation algorithm in this case. Furthermore, this simulation is also amenable to the stack scanning optimization technique developed in Section 3.1: a variant of the algorithm in Figure 4 which merely attempts to mark every transition it possibly can - instead of trying to test for # and fail - identifies some of those cycles L that can cause some color to repeat. These marks can then again be used for limiting stack scanning during simulation. Conclusions We studied the problem of using a given nondeterministic two-way multi-tape acceptor as a transducer by supplying inputs onto only some of its tapes, and asking it to generate the rest. We developed a family of algorithms for ensuring that this transduction does always yield finite answers, and another family of algorithms for actually computing these answers when they are guaranteed to exist. In addition, these two families of algorithms provided a way to execute and optimize the simulation of nondeterministic two-way multi-tape acceptors by restricting the amount of work that must be performed during run-time loop checking. These algorithms have been implemented in the prototype string database management system being developed at the Department of Computer Science in the University of Helsinki [8,12,13]. --R Foundations of Databases. Datalog and transducers. Foundation for Object/Relational Databases - The Third Manifesto PROXIMAL: a database system for the e Computers and Intractability: A Guide to the Theory of NP-Completeness Regular sequence operations and their use in database queries. AQL: An alignment based query language for querying string databases. translation and evaluation of Alignment Calculus. Reasoning about strings in databases. Reasoning about strings in databases. A declarative programming system for manipulating strings. Implementing a declarative string query language with string restructuring. Introduction to Automata Theory A framework for testing safety and e Finding paths with the right cost. Absolutely parallel grammars and two-way finite-state transducers Safety of recursive Horn clauses with infinite relations (extended abstract). Supporting lists in a data model (a timely approach). Artificial Intelligence: a Modern Approach. SEQ: A model for sequence databases. The AQUA approach to querying lists and trees in object-oriented databases Deriving constraints among argument sizes in logic programs. On the valuedness of finite transducers. On the lengths of values in a finite transducer. Temporal logic can be more expressive. --TR A database language for sets, lists and tables Safety of recursive Horn clauses with infinite relations On the valuedness of finite transducers On the lengths of values in a finite transducer Reasoning about strings in databases Artificial intelligence A framework for testing safety and effective computability Foundation for object/relational databases Regular sequence operations and their use in database queries Sequences, datalog, transducers Reasoning about strings in databases Foundations of Databases Introduction To Automata Theory, Languages, And Computation Computers and Intractability The AQUA Approach to Querying Lists and Trees in Object-Oriented Databases Implementing a Declarative String Query Language with String Restructuring Supporting Lists in a Data Model Timely Approach) --CTR Matti Nyknen , Esko Ukkonen, The exact path length problem, Journal of Algorithms, v.42 n.1, p.41-53, January 2002 Gsta Grahne , Raul Hakli , Matti Nyknen , Hellis Tamm , Esko Ukkonen, Design and implementation of a string database query language, Information Systems, v.28 n.4, p.311-337, June

transducers;finiteness dependencies;multi-tape automata

504552

A new lower bound for the list update problem in the partial cost model.

The optimal competitive ratio for a randomized online list update algorithm is known to be at least 1.5 and at most 1.6, but the remaining gap is not yet closed. We present a new lower bound of 1.50084 for the partial cost model. The construction is based on game trees with incomplete information, which seem to be generally useful for the competitive analysis of online algorithms. 2001 Elsevier Science B.V.

Introduction The list update problem is a classical online problem in the area of self-organizing data structures [4]. Requests to items in an unsorted linear list must be served while maintaining the list so that access costs remain small. We assume the partial cost model where accessing the ith item in the list incurs a cost of units. This is simpler to analyze than the original full cost model [14] where that cost is i. After an item has been requested, it may be moved free of charge closer to the front of the list. This is called Any other exchange of two consecutive items in the list incurs cost one and is called a paid exchange. An online algorithm must serve the sequence oe of requests one item at a time, without knowledge of future requests. An optimum offline algorithm knows the entire sequence oe in advance and can serve it with minimum cost OFF (oe). If the online algorithm serves oe with cost ON (oe), then it is called c-competitive if for a suitable constant b, ON for all request sequences oe. The competitive ratio c in this inequality is the standard yardstick for measuring the performance of the online algorithm. The well-known move- to-front rule MTF , for example, which moves each item to the front of the list after it has been requested, is 2-competitive [14, 15]. This is also the best possible competitiveness for any deterministic online algorithm for the list update problem [10]. 1 Institute for Theoretical Computer Science, ETH Z-urich, 8092 Z-urich, Switzerland. Email: ambuehl@inf.ethz.ch, gaertner@inf.ethz.ch Mathematics Department, London School of Economics, London WC2A 2AE, Great Britain. Email: stengel@maths.lse.ac.uk. Support by a Heisenberg grant from the Deutsche Forschungsgemeinschaft and the hospitality of the ETH Z-urich for this research are gratefully acknowledged. Randomized algorithms can perform better on average [9]. Such an algorithm is called c-competitive if \Theta ON (oe) where the expectation is taken over the randomized choices of the online algorithm. Fur- thermore, we call the algorithm strictly c-competitive if (2) holds with The best randomized list update algorithm known to date is 1.6-competitive. This algorithm COMB [2] serves the request sequence with probability 4/5 using the algorithm BIT [14], which alternately moves a requested item to the front or leaves it in place. With probability 1/5, COMB treats the request sequence using a deterministic TIMESTAMP algorithm [1], where a requested item x is moved in front of the first item in the list that has been requested at most once since the last request to x. Randomization is useful only against the oblivious adversary [5] that generates request sequences without observing the randomized choices of the online algorithm. If the adversary can observe those choices, it can generate requests as if the algorithm was deter- ministic, which is then at best 2-competitive. We therefore consider only the interesting situation of the oblivious adversary. Lower bounds for the competitive ratio can be proved using Yao's theorem [18]: If there is a probability distribution on request sequences so that the resulting expected competitive ratio for any deterministic online algorithm is d or higher, then every deterministic or randomized online algorithm has competitive ratio d or higher [8]. Teia [16] described a simple distribution on request sequences that, adapted to the partial cost model, shows a lower bound of 1.5. The optimal competitive ratio for the list update problem is therefore between 1.5 and 1.6, but the true value is as yet unknown. For lists with up to four items, it is possible to construct an online list update algorithm that is 1.5-competitive [3] and therefore optimal. In this paper, we show a lower bound that is greater than 1.5 when the list has at least five items. We will prove this bound for the standard assumption that algorithms may use paid exchanges. One can also prove a lower bound above 1.5 for the variant of the list update problem where only free exchanges are allowed. For that purpose, we have to modify and extend our method in certain ways, as mentioned at the end of this paper. Our construction uses a game tree where alternately the adversary generates a request and the online algorithm serves it. The adversary is not informed about the action of the online algorithm, so the game tree has imperfect information [12]. We consider a finite tree where after some requests, the ratio of online versus optimal offline cost is the payoff to the adversary. This defines a zero-sum game, which we solve by linear pro- gramming. For a game tree that is sufficiently deep, and restricted to a suitable subset of requests so that the tree is not too large in order to stay solvable, this game has a value of more than 1.50084. This shows that any strictly c-competitive online algorithm fulfills c - 1:50084. In order to derive from this a new lower bound for the competitive ratio c according to (1) with a nonzero constant b, one has to generate arbitrarily long request sequences. This can be achieved by composing the game trees repetitively, as we will show. A drawback is our assumption of the partial instead of the full cost model. In the latter model, where a request to the ith item in the list incurs cost i, the known lower bound is 1:5 \Gamma 5=(n+5) for a list with n items. This result by Teia [16] yields a lower bound for the competitive ratio much below 1.5 when the list is short. In fact, the algorithm COMB [2] is 1.5-competitive when n ! 9. To prove a lower bound above 1.5 for the full cost model we would have to extend our construction to longer lists. Unfortunately, a straightforward extension cannot compensate for the reduction of the competitive ratio by 5=(n (or any term proportional to 1=n) when considering the full instead of the partial cost model, so this case remains open. Nevertheless, we think a result for the partial cost model is still interesting since that model is more canonical when one looks at the analysis, and it is still close to the original problem formulation. 2. Pairwise analysis and partial orders The analysis of a list update algorithm is greatly simplified by observing separately the relative movement of any pair of items in the list. Let oe be a sequence of requests. Consider any deterministic algorithm A that processes oe. For any two items x and y in the list, let A xy (oe) be the number of times where y is requested and is behind x in the list, or vice versa. Then it is easy to show [6, 9, 2] that A xy (oe) ; where L is the set of items in the list. In that way, A xy (oe) represents the cost of the online algorithm projected to the unordered pair fx; yg of items. Let oe xy be the request sequence oe with all items other than x or y deleted. Many list update algorithms, like MTF , BIT , and TIMESTAMP, are projective in the sense that at any time the relative order of two items x and y in the list depends only on the projected request sequence oe xy and the initial order of x and y, which we denote by [xy] if x precedes y. (In general, we list items between square brackets to denote their current order in the list maintained by the algorithm.) For the optimal offline algorithm OFF , the projected cost OFF xy (oe) is clearly at least as high as the cost of serving oe xy optimally on the two-element list consisting of x and y. The latter cost is easy to determine since, for example, it is always optimal to move an item to the front at the first of two or more successive requests. In fact, the item must be moved to the front at the first of three or more successive requests. On the other hand, it is usually not optimal to move an item that is requested only once. Hence, for any two items x and y where x precedes y, an online algorithm serving a request to y can either leave y behind x or move y in front of x, which, either way, is a "mistake" depending on whether y will be requested again before x or not. Based on this observation, Teia [16] has constructed a lower bound of 1.5 for the competitive ratio. The requests are generated in runs which are repeated indefinitely. At the start of each run, the list currently maintained by the offline algorithm has a particular this list is traversed from front to back, requesting each item with equal probability either once or three times. If an item is requested three times, then it is moved to the front at the first request, otherwise left in place, which is an optimal offline treatment. This results in a new offline list, which determines the next run. The following table (3) lists the resulting costs for the possible actions of the online algorithm, projected on items x and y. In that table, WAIT refers to an online algorithm that moves the requested item only at the second request in succession (if it is not moved then, the online costs are even higher), and MTF moves the item to the front at the first request. In (3), item x is assumed to precede y in the lists maintained by both offline and online algorithm. The four request sequences each have probability 1/4. For each of the four possible combinations of WAIT and MTF , the column ON denotes the online cost and I after denotes the number of inversions in the online list after the requests have been served. an inversion is a transposition of two items relative to their position in the offline list. OFF ON with [xy]; I before oe xy with x WAIT x MTF x WAIT x MTF ON I after ON I after ON I after ON I after Without inversions, the MTF algorithm, for example, would incur the same cost as the optimal offline cost. However, the inversion increases the online cost by a full unit in the next run, where [xy] is the order for the offline algorithm but [yx] is the order of x and y in the list used by the online algorithm. The following table shows these online costs when the algorithm starts with such an inversion, denoted by I before = 1. OFF ON with inversion [yx]; I before oe xy with x WAIT x MTF x WAIT x MTF ON I after ON I after ON I after ON I after Tables (3) and (4) list all possible online actions, except for those that perform even worse (leaving a triply requested item in place, for example). Note that it does not matter if the online algorithm conditions its action on the presence of inversions or not. Let T be the distribution on request sequences generated by the described method of Teia. Then the expected online costs together with the change in the number of inversions fulfill the inequality This follows from (3) and (4) by considering the projected sequences and telescoping the sum for the inversion counts from one run to the next (where I before for that run is equal to I after for the previous run and cancels). Inequality (5) shows that the number of inversions can serve as a potential function [7]. The variation \DeltaI = I after \Gamma I before of this potential function is bounded, so that (5) implies that any online algorithm is at most 1.5-competitive for the distribution T on request sequences. We will extend Teia's method in our lower bound construction. Using partial orders, one can construct a 1.5-competitive list update algorithm for lists with up to four items [3]. The partial order is initially equal to the linear order of the items in the list. After each request, the partial order is modified as follows, where xky means that x and y are incomparable: partial order after request to before z 62 fx; yg x y That is, a request only affects the requested item y in relation to the remaining items. Then y is in front of all items x except if x ! y held before, which is changed to xky. The initial order in the list and the request sequence determine the resulting partial order. One can generate an arbitrary partial order in this way [3]. The partial order defines a position for each item x. If the online algorithm can maintain a distribution on lists so that the expected cost of accessing an item x is equal to p(x), then this algorithm is 1.5-competitive [3]. One can show that then x is with probability one behind all items y so that y ! x, and precedes with probability 1/2 those items y where xky. Incomparable elements reflect the possibility of a "mistake" of not transposing these items, which should have probability 1/2. For lists with up to four items, one can maintain such a distribution using two lists only. That is, the partial order is represented as the intersection of two lists, where each list is updated by moving the requested item suitably to the front, using only free exchanges. The algorithm works by choosing one of these lists at the beginning with probability 1/2 as the actual list and serving it so as to maintain the partial order (with the aid of the separately stored second list). The partial order approach is very natural for the projection on pairs and when the online algorithm can only use free exchanges. A lower bound above 1.5 must exploit a failure of this algorithm. This is already possible for lists with five items, despite the fact that all five-element partial orders are two-dimensional (representable as the intersection of two linear orders). Namely, let the items be integers and let the initial list be [12345], and consider the request sequences After the first request to 4, the partial order states 4k1, 4k2, 4k3, and 4 ! 5, and otherwise 5. Using a free exchange, 4 can only be moved forward and has to precede each with probability 1/2. This is achieved uniquely with the uniform distribution on the two lists [12345] and [41235] (this, as well as the following, holds even though distributions on more than two lists are allowed). The next request to 2 induces 2 ! 4, so 2 must be moved in front of 4 in the list [41235], where 2 already passes 1, which yields the unique uniform distribution on [12345] and [24135]. The next request to 5 entails that 5 is incomparable with all other items. It can be handled deterministically in exactly two ways (or by a random choice between these two ways): Either 5 is moved to the front in [24135], yielding the two lists [12345] and [52413] with equal probability, or 5 is moved to the front in [12345], yielding the two lists [51234] and [24135] with equal probability. If the two lists are [12345] and [52413], the algorithm must disagree with the partial order after the request to 4 as in oe 1 , since then 4 must precede both 1 and 5 in both lists (so 4 is moved to the front in both lists) but then incorrectly passes 2 where only 2k4 should hold. Similarly, for the two lists [51234] and [24135] the request to 3 as in oe 2 moves 3 in front of 5 and 4 in both lists, so that it passes 1, violating 1k3. Thus, either oe 1 or oe 2 in (6) causes the poset-based algorithm to fail, which otherwise achieves a competitive ratio of 1.5. These sequences will be used with certain probabilities in our lower bound construction. 3. Game trees with imperfect information Competitive analysis can be phrased as a zero-sum game between two players, the adversary and the online algorithm (or online player ). In order to deal with finite games, we assume a finite set S of request sequences oe (of a given bounded length, for example), which represent the pure strategies of the adversary. These can be mixed by random- ization. The online player has a finite number N of possible ways of deterministically serving these request sequences. These deterministic online algorithms can also be chosen randomly by suitable probabilities . In this context of finitely many request sequences, an arbitrary constant b in (2) is not reasonable, so we look at strict competitiveness. The randomized online algorithm is strictly c-competitive if for all oe in S, where ON j (oe) is the cost incurred by the jth online algorithm and OFF (oe) is the optimal offline cost for serving oe. We can disregard the trivial sequences oe with OFF consist only of requests to the first item in the list. In this case (7) is equivalent to ON j (oe) OFF (oe) The terms ON j (oe)=OFF (oe) in (8), for 1 - j - N and oe 2 S, can be treated as a payoff to the adversary in a zero-sum game matrix with rows oe and columns j. Correspondingly, a lower bound d for the strict competitive ratio is an expected competitive ratio [8] resulting from a distribution on request sequences. This distribution is a mixed strategy of the adversary with probabilities q oe for oe in S so that for all online strategies ON j (oe) The minimax theorem for zero-sum games [18] asserts that there are mixed strategies for both players and reals c and d so that (8) and (9) hold with c is the "value" of the game and the optimal strict competitive ratio for the chosen finite approximation of the list update problem. Note that it depends on the admitted length of request sequences. Due to the complicated implicit definition and large size of the game matrix, we only know bounds c and d in (8) and (9) that hold irrespective of the length of the request sequences where d ! c. The number of request sequences is exponential in the length of the sequences. The online player has an even larger number of strategies since that player's actions are conditional on the observed requests. This is best described by a game tree. At each nonterminal node of the tree, a player makes a move corresponding to an outgoing edge. The game starts at the root of the tree where the adversary chooses the first request. Then, the online player moves with actions corresponding to the possible reorderings of the list after the request. There are n! actions corresponding to all possible reorderings. (Later, we will see that most of them need not be considered.) The players continue to move alternatingly until the last request and the reaction by the online player. Each leaf of the tree defines a sequence oe and an online cost ON (oe) (depending on the online actions leading to that leaf), with payoff ON (oe)=OFF (oe) to the adversary. The restricted information of the adversary in this game tree is modeled by information sets [12]. Here, an information set is a set of nodes where the adversary is to move and which are preceded by the same previous moves of the adversary himself. Hence, the nodes in the set differ only by the preceding moves of the online player, which the adversary cannot observe. an action of the adversary is assigned to each information set (rather than an individual node) and is by the definition the same action for every node in that set. On the other hand, the online player is fully informed about past requests, so his information sets are singletons. Figure 1 shows the initial part of the game tree for a list with three items for the first and second request by the adversary, and the first online response, here restricted to free exchanges only. A A A A A A @ @ @ @ @ @ [123] [123] [213] [123] [213] [312] ffl \Theta \Theta \Theta \Theta \Theta \Theta ffl \Theta \Theta \Theta \Theta \Theta \Theta \Theta \Theta \Theta \Theta \Theta \Theta \Theta \Theta \Theta \Theta \Theta \Theta ffl \Theta \Theta \Theta \Theta \Theta \Theta ffl \Theta \Theta \Theta \Theta \Theta \Theta adv ON ON ON adv adv adv Figure 1. Game tree with information sets. A pure strategy in a game tree assigns a move to every information set of a player, except for those that are unreachable due to an earlier choice of that player. Here, the online player has information sets (like in Figure 1) where each combination of moves defines a different strategy. This induces an exponential growth of the number of strategies in the size of the tree. The strategic approach using a game matrix as in (8) and becomes therefore computationally intractable even if the game tree is still of moderate size. Instead, we have used a recent method [17, 11] which allows to solve a game tree with a "sequence form" game matrix and corresponding linear program that has the same size as the game tree. Using game trees, a first approach to finding a randomized strategy for the adversary is the following. Consider a list with five items, the minimumnumber where a competitive ratio above 1.5 is possible. Fix a maximum length m of request sequences, and generate the game tree for requests up to that length. At each leaf, the payoff to the adversary is the quotient of online and offline cost for serving that sequence. Then convert the game tree to a linear program, and compute optimal strategies with an LP solver (we used CPLEX). However, this straightforward method does not lead to a strict competitiveness above 1.5, for two reasons. First, "mistakes" of an algorithm, like in the last column in (3), manifest themselves only later as actual costs, so there is little hope for an improved lower bound using short request sequences. Secondly, even if only short sequences are considered, the online player has n! responses to every move of the adversary, so that the game tree grows so fast that the LP becomes computationally infeasible already for very small values of m. The first problem is overcome by adding the number of inversions of the online list, denoted by I after in (3) and (4) above, to the payoff at each leaf. This yields a strict competitive ratio greater than 1.5 for rather short sequences. The inversions are converted into actual costs by attaching a "gadget" to each leaf of the game tree that generates requests similar to Teia's lower bound construction. The next section describes the details. The second problem, the extremely rapid growth of the game tree, is avoided as follows. First, we limit the possible moves of the online player by allowing only paid exchanges of a special form, so-called subset transfers [13]. A subset transfer chooses some items in front of the requested item x and puts them in the same order directly behind x (e.g. [12345x67] ! [13x24567]). Afterwards, the adversary's strategy computed against this "weak" online player is tested against all deterministic strategies of the online player, which can be done quickly by dynamic programming. Then the lower bound still holds, that is, the "strong" online player who may use arbitrary paid exchanges cannot profit from its additional power. 4. The game tree gadgets We compose a game tree from two types of trees or "gadgets". The first gadget called FLUP (for "finite list update problem") has a small, irregular structure. The second gadget called IC (for "inversion converter") is regularly structured. Both gadgets come with a randomized strategy for the adversary, which has been computed by linear programming for FLUP. an instance of IC is appended to each leaf of FLUP. The resulting tree with the specified strategy of the adversary defines a one-player decision problem for the online player that has an expected strict competitive ratio of at least 1:5 about 1.50084, for the simplest version of FLUP that we found; larger versions of FLUP give higher lower bounds. Both gadgets assume a particular state of the offline list, which is a parameter that determines the adversary strategy. Furthermore, at the root of FLUP (which is the beginning of the entire game), it is assumed that both online and offline list are in the same state, say [12345]. Then the adversary strategy for FLUP generates only the request sequences 4, 425, 4253, and 4254 with positive probability, which are the sequences in (6) or a prefix thereof. After the responses of the online player to one of these request sequences, the FLUP tree terminates in a leaf with a particular status of the online list and of the offline list, where the latter is also chosen by the adversary, independently of the online list. For the request sequence 4, that offline list is [41235], that is, the offline algorithm has moved 4 to the front. If the FLUP game terminates after the request sequence 425, the adversary makes an additional internal choice, unobserved by the online player, between the offline lists [51234] and [52134]. In the first case, the offline player brought 5 to the front but left 4 and 2 in their place, in the second, 2 was also moved to the front. Similar choices are performed between the offline lists for the request sequences 4253 and 4254. requests offline list probability OFF MTF The specific probabilities for these choices of the adversary in FLUP are shown in (10). The last three columns denote the cost for the offline algorithm and, as an example, the two online algorithms MTF and WAIT , where WAIT moves an item to the front at the second request. The FLUP tree starts with 4 as the first request, followed by the possible responses of the online player. Next, the adversary exits with probability 396/1184, without a request, to the leaf with offline list [41235], and with complementary probability requests item 2, which is followed by the online move, and so on. Each leaf of the FLUP tree is the root of an IC gadget which generates requests (similar to the runs in Teia's construction, see below), depending on the offline list. The number of inversions of the online list relative to this offline list is denoted by I in (10). The purpose of the IC gadget is to convert these inversions into actual costs. Any request sequence generated by the IC gadget can be treated with the same offline cost v, here Thereby, the online algorithm makes mistakes relative to the offline algorithm, so that the additional online cost in IC is 1:5v. Let ON be the online cost incurred inside FLUP. Then FLUP can be represented as a game tree where the IC gadget at each leaf is replaced by the payoff D to the adversary, Using these payoffs, the probabilities in (10) have been computed by linear programming. One can show that any online strategy, as represented in the FLUP tree, has an expected strict competitive ratio of at least 1:5 + 1, or about 1.50084. Two optimal online strategies are MTF and WAIT , where the values of ON I as used in (11) are also shown in (10). Here, WAIT moves only item 4 to the front at the end of the request sequence 4254. The well-defined behavior of the IC gadget allows to replace it by a single payoff D as in (11). Furthermore, the online player knows that the FLUP gadget has been left and the IC gadget has been entered, since IC starts with a request to the first item in the offline list, like 4 when that list is [41235] as in the first row of (10). In this context, we make a certain assumption about the internal choice of the adversary between different offline lists. Namely, at the start of the IC gadgets for the offline lists [12345] and [13524] which follow the request sequence 4253, the first request is 1 and then the online player cannot yet tell which IC gadget has been entered. Strictly speaking, the two gadgets have to be bridged by appropriate information sets for the online player. However, we assume instead that the internal choice of the adversary between the two lists is revealed to the online player at the beginning of IC, which is implicit in replacing IC by a single payoff. This is allowed since it merely weakens the position of the adversary: Any online strategy without this extra information can also be used when the online player is informed about the adversary's internal choice, so then the online payoff cannot be worse. The offline list assigned to a leaf of the FLUP gadget is part of an optimal offline treatment (computed similar to [13]) for the entire request sequence. However, that list may even be part of a suboptimal offline treatment, which suffices for showing a lower bound since it merely increases the denominator in (9). Some of the offline costs in (10) can only be realized with paid exchanges by the offline algorithm. For example, the requests 4253 are served with cost 10 yielding the offline list [23451] by initial paid exchanges that move 1 to the end of the list. With free exchanges, this can only be achieved by moving every requested item in front of 1, which would result in higher costs. In the remainder of this section, we describe the IC gadget. Its purpose is to convert the inversions at the end of the FLUP game to real costs while maintaining the lower bound of at least 1.5. At the same time, these inversions are destroyed so that both the online list and the offline list are in the same order after serving the IC. The IC extends the construction by Teia [16] described in Section 2 above. Let T k be the sequence that requests the first k items of the current offline list in ascending order, requesting each item with probability 1/2 either once or three times. Assume that the offline algorithm treats T k by moving an item that is requested three times to the front at the first request, leaving any other item in place, which is optimal. The triply requested items, in reverse order, are then the first items of the new offline list, followed by the remaining items in the order they had before. Then T n is a run as used in Teia's construction for a list with n items. The random request sequence generated there can be written as T w n , that is, a w-fold repetition of T n , where w goes to infinity. Note that the offline list and hence the order of the requests changes from one run T n to the next, so T 2 for example, is not a repetition of two identical sequences. The optimal offline treatment of T k costs units. The difference between our construction and Teia's is the use of only a prefix of the elements in the offline list. We show which for has already been proved above, see (5). To see (12) also for k ! n, we use projection on pairs and consider the case of only two items. If none of the two items occur in T k , both sides of (12) are zero, because, by definition, projection on pairs ignores items that are not projected. If only one item occurs in T k , only the first one in the offline list was requested, so OFF (T k because the online algorithm incurs cost at least one if there is an inversion. This shows (12). As in (5) above, (12) can be extended to concatenations T of sequences T k . We let IC be the randomly generated sequence defined by which by the preceding considerations fulfills If E[I after in (13), that is, there are no inversions left after serving IC, then inversions would indeed have been converted to actual costs. Otherwise, suppose that after serving IC, there is an inversion between two items x and y, say, with x in front of y in the final offline list. Then by the definition of IC, item x was requested at least three more times after the last request to y. So the online player could have saved a cost unit by moving x in front of y in his list after the second request to x. To summarize, the sequence IC produces an additional online cost unit for every inversion that holds between the online and the offline list at the end of IC. Then, however, we can assume without loss of generality that both lists are in the same state after IC. Namely, if they are not, the online player could as well have served IC as intended (leaving no inversions) and invested the saved cost units in creating the inversions at the beginning of the next FLUP gadget, where their costs are taken into account. Thus, indeed, (13) holds with E[I after inversions have become actual costs as stated in (11). The offline costs there are Since the online and offline list are identical at the end of the IC, a new FLUP game can be started. This generates request sequences of arbitrary length. In that way, we obtain a lower bound above 1.5 for the competitive ratio c in (2) for any additive constant b. In the above construction, the value of the lower bound does not depend on whether the online player may use paid exchanges or not, but the adversary's strategy does use paid exchanges. So it seems that the online player cannot gain additional power from paid exchanges. This raises the conjecture that by restricting both players to free exchanges only, the list update problem might still have an optimal competitive ratio of 1.5. However, this is false. There is a randomized adversary strategy where the offline algorithm uses only free exchanges which cannot be served better than with a competitive ratio of 1:5+1=5048. Because of the length of the sequences used in the corresponding FLUP game, this result is more difficult to obtain. First of all, the sequences used in that game are not found by brute force any more, but by dynamic game tree search with alpha-beta pruning in an approximate game. In that approximate game, the online player is restricted to a small set of random moves, similar to the poset algorithm. Secondly, the above argument about the order of the elements in the online list after leaving the IC gadget no longer holds. This can be resolved by a further elaboration of our method. The details are beyond the scope of this paper. Extending our result to the full cost model requires a systematic treatment of lists of arbitrary length n. This is easy for the IC gadget but obviously difficult for the FLUP gadget. We hope to clarify the connection of FLUP with the sequences in (6) that beat the partial order approach to make progress in this direction. --R Improved randomized on-line algorithms for the list update prob- lem A combined BIT and TIME-STAMP algorithm for the list update problem List update posets. A survey of self-organizing data structures On the power of randomization in on-line algorithms Amortized analyses of self-organizing sequential search heuristics Online Computation and Competitive Analysis. an optimal online algorithm for metrical task systems. Two results on the list update problem Fast algorithms for finding randomized strategies in game trees. Extensive games and the problem of information. Optimum off-line algorithms for the list update problem Randomized competitive algorithms for the list update problem Amortized efficiency of list update and paging rules A lower bound for randomized list update algorithms Efficient computation of behavior strategies. Probabilistic computations: Towards a unified measure of com- plexity --TR Amortized efficiency of list update and paging rules Amortized analyses of self-organizing sequential search heuristics On the power of randomization in online algorithms Two results on the list update problem An optimal on-line algorithm for metrical task system A lower bound for randomized list update algorithms Fast algorithms for finding randomized strategies in game trees A combined BIT and TIMESTAMP algorithm for the list update problem Online computation and competitive analysis Improved randomized on-line algorithms for the list update problem Self-Organizing Data Structures

list-update;analysis of algorithms;competitive analysis;on-line algorithms

504560

Online request server matching.

In the following paper an alternative online variant of the matching problem in bipartite graphs is presented. It is triggered by a scheduling problem. There, a task is unknown up to its disclosure. However, when a task is revealed, it is not necessary to take a decision on the service of that particular task. On the contrary, an online scheduler has to decide on how to use the current resources. Therefore, the problem is called online request server matching (ORSM). It differs substantially from the online bipartite matching problem of Karp et al. (Proc. 22nd Annual ACM Symp. on Theory of Computing, Baltimore, Maryland, May 14-16, 1990, ACM Press, New York, 1990). Hence, the analysis of an optimal, deterministic online algorithm for the ORSM problem results in a smaller competitive ratio of 1.5. An extension to a weighted bipartite matching problem is also introduced, and results of its investigation are presented. Additional concepts for the ORSM model (e.g. lookahead, parallel resources, deadlines) are studied. All of these modifications are realized by restrictions on the input structure. Decreased competitive ratios are presented for some of these modified models. Copyright 2001 Elsevier Science B.V.

Introduction In this report we study a model which was developed for a rather simple scheduling problem. Consider a single resource and a discrete time model. The resource is available for one unit in every time step. In the following this resource is called server. Every time step a task can occur that has a demand of one unit of the server. This task is called request. Such a request includes a set of time steps which specifies acceptable times for serving. These times must not be situated in the past and neither need to be consecutive nor include the time step of the arrival of the request. It is obvious that we can model this problem by a bipartite graph (R E). The two disjoint partitions represent the server resource at every time step (S) and the set of requests (R). Whenever a request r 2 R can be served at time j, there is an edge fr; sg between request vertex r and the vertex s 2 S which represents the j-th time step. Then the problem to decide when a request should be served is nothing else than to construct a matching in G. Our scheduling problem is online, i. e. the requests appear through time while the current usage of the server has to be determined. These decisions have to be made without knowledge of further requests and cannot be taken back. The definition of our problem implies that in general it is impossible to serve all requests 1 . Additionally, the lack of information about the future prevents an optimal solution to be found which maximizes the number of successful requests. This online matching problem in a bipartite graph is called online request server matching (ORSM). We make use of competitive analysis to investigate the described problem. Here the worst case factor between the quality of an optimal solution and a solution calculated by an online algorithm on the same input is determined. This factor is called competitive ratio. For our matching problem the size of a maximum matching in G is the value of the optimal solution. It is well-known how to calculate it in O( (see [ET75] or any comprehensive textbook on algorithms). In this report we present an optimal online algorithm which constructs a matching of at least 2of the maximal size. We also investigate a weighted variant of this online matching problem. Then all edges have a weight and the objective is the construction of a maximal weighted matching. For this problem our analysis shows a lower bound of and an upper bound of 2 for the competitive ratio. The material of this report is organized as follows. The next subsection presents a short introduction to competitive analysis. Thereafter, an overview of related work is given. Section 2 defines our model formally. It will be compared precisely with two models that have been studied in literature. Additionally, a few definitions and notations are given in the end. The simple, unweighted variant of our model is analysed in Sect. 3. It includes a general lower bound for the competitive ratio, a deterministic online algorithm, and a matching upper bound. Section 4 investigates the weighted version of this problem. The proof of a general lower bound is followed by the presentation of an algorithm. Its tight Imagine two requests to exactly the same, single server resource. analysis establishes a gap to the previous lower bound. The section ends with some remarks and a suggestion for a more sophisticated online algorithm. This report is completed by a description of a few open problems. 1.1 Competitive Analysis Algorithms are often based on the principle of the following sequence: input, computation, and output. After reading the complete input, which is a description of a finite problem, it computes a solution and presents it at the end. Such a behaviour is called offline. In real world applications like e. g. embedded systems we find different requirements. Here, parts of the input appear by and by for an indefinite period of time and decisions of the solution have to be made immediately without knowing the future. These types of problems, including their solving algorithms, are called online. Different methods were suggested to analyse such online algorithms. The most popular methods of the past, which are still in use, assume a known input distribution (a stochastic model for the input). Of course, the expressiveness of results of these studies are highly dependent on the correct choice of the distribution. In contrast 1984 Daniel D. Sleator and Robert E. Tarjan introduced a new method for analysing online algorithms (journal version is [ST85]). It is a method of worst case analysis which avoids these difficulties and has been becoming more and more popular in the last decade. The basic idea is a comparison of the quality of the solution computed by the online algorithm and the quality of an optimal, offline solution for the same input. Let A be an online algorithm and let OPT be an optimal offline algorithm for a payoff maximization problem 2 . Then, for an input sequence oe we denote with perf A (oe) the performance of the algorithm A and perf OPT (oe) is the performance of OPT respectively. A is called c-competitive if there is a constant ff such that The constant ff has to be independent of the input and can compensate irregularities right at the beginning. Then the infimum over all values c such that A is c-competitive is called the competitive ratio of A. Whenever the constant alternatively define the competitive ratio as sup oe perf OPT (oe) perf A (oe) 2 In this report we do not consider problems which have to minimize costs. The definitions would slightly differ to ensure that a competitive ratio is always at least 1. We immediately realize that this analysis is independent of a stochastic model for the inputs and gives performance guarantees. On the other hand this kind of analysis is sometimes unrealistically pessimistic. Indeed, an online algorithm approximates an optimal solution while working under the restriction of incomplete knowledge. So the distinguished name competitive ratio for the approximation factor seems to be adequate. An excellent, introductory article on competitive analysis is [Kar92]. A comprehensive introduction is the textbook by Allan Borodin and Ran El-Yaniv [BEY98] which includes an extensive bibliography. Two essential techniques to prove competitive ratios have been established. A description of these methods illustrates the character of this type of analysis. Lower bounds for the competitive ratio are typically shown by an adversary strategy. We can interpret competitive analysis as a game between an online algorithm A and an omniscient adversary. The adversary creates the input sequence oe with knowledge of A. So it can in advance calculate every decision of A on oe, perf A (oe), and perf OPT (oe). Due to the unlimited computational power the malicious adversary is able to construct a worst case input sequence. When we derive general lower bounds for the competitive ratio of a problem we slightly change our view. Then a strategy is given which tells the adversary how to construct the next part of the input sequence dependent on the decisions an online algorithm was able to make. The performance analysis of these sequences gives a lower bound for the competitive ratio when the strategy is able to generate infinite inputs. Otherwise the loss of performance of an online algorithm could be compensated by the constant ff in the above definition. To prove upper bounds for an online problem the performance of a designed algorithm A is investigated. Sometimes, ad hoc arguments can be applied to show how A can bound the competitive ratio. However, it is more common to apply a potential function argument for an analysis of amortized performance. Hence, every pair of a state of A and a state of OPT is mapped to a real value by a potential function \Phi. Whenever it can be shown that for every step i with input oe i holds and \Phi is bounded above, then A is c-competitive. It was shown that there exists such a potential function for every online algorithm (see the description in [IK96], page 534). Nevertheless the variant usage of potential functions and combinations with additional arguments are nowadays commonly used in proofs of the performance of online algorithms. Several extensions of the competitive analysis were suggested. A major influence had the seminal work [BDBK + 94] which introduces randomized online algorithms and adapted methods for their analysis. However, in this report we limit our study to deterministic online algorithms. 1.2 Previous Work We introduced our online problem in terms of a scheduling problem. However, there is a vast literature on the subject of online scheduling and related problems like load balancing and routing. We will not review these works here. The reader may consult the survey by instead. The publications of online matching problems are more relevant to the studies of this report. In the following, we discuss these papers. The first article about an online version of a bipartite matching problem is by Richard M. Karp, Umesh V. Vazirani and Vijay V. Vazirani [KVV90]. The partition U of the graph E) is known in advance and the vertices of V including their edges arrive through time. Whenever a vertex v 2 V is revealed, an online algorithm has to decide which edge incident to v is added to the online matching M . The objective is the maximization of the size of M . For deterministic online algorithms the competitive ratio is exactly 2. An adversary can present an input with a vertex being adjacent to two vertices After the decision of the online algorithm, the adversary presents vertex which is adjacent to the previous matched vertex u or u 0 , only. The online algorithm is not able to match v 1 and v 2 . However, there is such an offline solution. The infinite repetition of this strategy results in the lower bound of 2. The greedy algorithm which adds the first possible edge of an incoming vertex to the matching M achieves this competitive ratio. From graph theory we know that this greedy algorithm constructs a maximal matching (every edge of the graph is either itself a matching edge or is adjacent to one) and every maximal matching has at least half the size of an optimal, maximum matching. The key contribution of the discussed paper is the analysis of a randomized online algorithm for the described problem. Ming-Yang Kao and Stephen R. Tate investigated the above model with blocked inputs [KT91]. When an input occurs k vertices were revealed 'in a block' instead of one at a time. Let be the number of vertices in the partition V of the bipartite graph. Of course with the problem is the same as above and for we have the offline version of the matching problem. For deterministic online algorithms, no improvements are possible as long as k - n. For randomized algorithms the result of [KVV90] cannot be improved apart from low order terms as long as The online b-matching was analysed by Bala Kalyanasundaram and Kirk Pruhs in [KP95]. In this problem the vertices of the partition V can be incident to at most b edges out of the matching. Again, the size of the matching has to be maximized. The authors give an optimal deterministic algorithm with competitive ratio A weighted variant of the online bipartite matching was studied by the same authors in [KP93]. When allowing an arbitrary weight function for the edges, no competitive algorithm can exist. To see this consider that an online algorithm has chosen a matching edge of weight one. As a next step an adjacent edge of arbitrary weight is revealed. Then no online algorithm can bound the competitive ratio. Hence, the model was restricted to a weighted, complete bipartite graph of partitions of equal which is a metric space (especially the triangle inequation is fulfilled). When asked for a maximal weighted matching, the greedy algorithm is optimal and 3-competitive. On the other hand in the minimum weighted matching problem, a perfect matching of minimal weight has to be determined in a bipartite graph defined as above. Then an optimal and (2n \Gamma 1)-competitive algorithm is shown. The same result was independently discovered by Samir Khuller, Stephen G. Mitchell and Vijay V. Vazirani in [KMV94] 3 . The last article also contains a study of an online version of the stable marriage problem. In [ANR95] Yossi Azar, Joseph Naor and Raphael Rom studied a different model based on bipartite graphs. They called it online assignment. The partition U has a fixed size and vertices out of V are adjacent to a subset of U . For each v 2 V one of its weighted edge must be selected immediately after its arrival with the objective to minimize the maximal weight of all selected edges incident to a vertex in U . Indeed, this is a load balancing problem. For deterministic online algorithms a lower bound of dlog 2 1)e and an upper bound of dlog 2 ne shown for the competitive ratio. The technical report by Ethan Bernstein and Sridhar Rajagopalan [BR93] is of major importance for our following studies. An online matching problem in general graphs called roommates problem has been introduced. The graph is undirected, simple and weighted. An unweighted version of this model has also been investigated. The input sequence consists of the vertices of V . Whenever a vertex is revealed, all of its adjacent vertices and the weighted edges in between become known. We want to emphasize that this process includes adjacent vertices never seen before in the input sequence. Then, an edge of the current vertex to a non-matched vertex (that has been revealed previously) can be added to the online matching. An edge to a non-revealed and incompletely known vertex can be selected later when this adjacent vertex is the current one in the input sequence. For the roommates problem the unweighted model is interpreted as follows: People arrive one at a time to a hotel where a conference takes place. The hotel consists of double rooms, only. Every person gives a list of possible roommates independently of whether they have arrived yet. The model assumes that these lists are symmetric, i. e., every potential roommate will accept to share this room. The hotel manager has to decide in which room the person will stay. The objective is to minimize the allocated rooms, i. e. to maximize the matching in the implicitly defined graph. In the weighted version the aim is an online construction of a weighted matching of maximal size in G. In the paper a tight analysis of the unweighted model is given. Therefore the competitive ratio is 1:5 : For the weighted model a lower bound of 3 is proven. A suggested online algorithm is shown to be 4-competitive. 3 Both articles ([KP93] and [KMV94]) were previously published on conferences in 1991. At the end of the next section we will revisit the roommates problem and we will compare it to our model. Indeed, our model is a special case of the roommates problem. For our investigations we were able to adapt proofs taken from the discussed paper. 2 The Model In the beginning of the introduction we described an online matching problem. Now we will present a formal definition for the online request server matching problem (ORSM). The underlying structure of the problem is a bipartite graph G := (R : E). Both partitions R and S are totally ordered. We denote the vertices by r and the indices indicate the position within the order. We interpret this order as a discrete time model. The vertices of partition S represent a simple resource called server, which is available for one unit each time step. Partition R is interpreted as a set of tasks. Such a task has a demand of one server unit to be completed. We call them requests and every time step one of them might occur. An edge fr between a request vertex r i and a server vertex s j means that the request can be served in time step j. The set of edges E ae R \Theta S is constructed with a restriction: This means that a request that occurs at time step i must not specify a possible service time in the past. Without this restriction the modelled scheduling problem does not make sense and no competitive online algorithm would be possible. Now we have to specify how this model works online: When the system starts the partition R is completely unknown 4 . In the beginning of a time step i the request r i is revealed as input, i. e. the vertex r i and all of its edges are inserted into the previous known part of G. If no request appears, vertex r i is isolated. After this input, an algorithm has to decide how to use the server in the current time step i. Indeed, it can add an edge incident to s i to the online matching M . It is worth noting that due to the restriction (1) all edges incident to s i are known when this decision has to be made. The online algorithm has the objective to maximize the cardinality of the matching M , i. e. to serve as much requests as it can. Up to now the graph G was unweighted. We also study the weighted variant. Then the graph is defined as G := (R w). The weight function defines a positive real value for every edge. The objective is the construction of a weighted matching of maximal overall weight. Otherwise, the problem and its online fashion is completely the same as in the ORSM problem. This version is named online request server weighted matching problem or in short wORSM problem. 4 When we take a close look this is not the truth. We know that every time step i a new vertex r i is inserted but its set of incident edges is unknown before. For our convenience we will interpret the input process in the introduced way. 2.1 Our Model in Comparison with Models out of Literature At first we want to make a clear distinction between the ORSM model and the online bipartite matching problem in [KVV90]. In the later one the vertices from the unknown partition and their edges appear in the same way as in the ORSM problem. However, in the online bipartite matching problem no order on the vertices of the already known partition is given and no restrictions on the set of edges. After a new vertex v is inserted the online algorithm has to decide which edge incident to v should be added to the matching. In contrast, in the ORSM model, one asks for an edge incident to the current server vertex to add into the matching. This server vertex is situated in the known partition of the bipartite graph. The restriction (1) of the set of edges guarantees that all edges of this server vertex are known at that time. Due to the specified input process (revealing a request vertex and all of its edges), all adjacent edges are known as well. There is therefore the advantage to have some extra knowledge of the graph structure whenever a decision has to be taken. That is also the reason why it is possible to achieve a better competitive ratio for the ORSM problem. Additionally, in the weighted version, a restriction on a metric space, or the other restrictions described in Sect. 1.2 are not necessary. Again, we want to emphasize the difference: In the online bipartite matching problem decisions about adding an edge to the online matching are made with respect to a set of edges incident to a just revealed vertex. In the ORSM problem such a decision is made with respect to a set of edges incident to a vertex of the other (known) partition. In the previous section it was claimed that the ORSM problem is a special case of the roommates problem. Now we are able to give a transformation of an instance of the wORSM problem with its bipartite graph (R to an instance of the roommates problem with the underlying graph its total order OE on the set VR . This transformation defines: VR and the order OE on VR is defined by the use of the orders on R and S such that: Whenever a vertex r i is inserted, the roommates problem on GR is not able to add any edge to M because no adjacent s-vertices are revealed (remember restriction (1) on E). Then vertex s i is inserted. All of its incident edges and adjacent vertices respectively, are known at that time and so every edge of s i can be selected as a matching edge in the roommates problem. No edge of an unrevealed vertex was given simultaneously. That means s i will never be a candidate for the matching again. We conclude that the roommates problem can simulate the wORSM problem using the given transformation. Nevertheless, the two models are not able to simulate each other. Hence, we have to prove the lower bound of the competitive ratio for our more restrictive model. Additional we present an online algorithm and an analysis applying simplified arguments which are tailor-made to the ORSM problem. In the weighted model, the increase in knowledge of the graph structure, compared to the roommates problem, results in a lower value for the competitive ratio. Before starting our investigations a few definitions and notations are presented. 2.2 Definitions and Notations Although we assume that the reader is familiar with basic concepts in graph theory, and although we already used some graph theoretical notations, a few standard definitions are included in the following list. (weighted graph) is called a weighted graph iff V is a finite set of vertices, is a set of edges, and is a weight function. When using the unit weight function an unweighted graph E) is derived. As you can see, the weight function is omitted in the notation. This conversation from a weighted to an unweighted graph applies to the next definitions of graphs and matchings. Bipartite graphs consist of two disjoint sets of vertices without edges inside the sets: Definition 3 (bipartite weighted graph) is called a bipartite weighted graph iff U and V are finite sets of vertices, is a set of edges, and is a weight function. Definition 4 (vertex induced subgraph) be a weighted graph and V S ae V . Then := is the subgraph of G induced by vertex set be a weighted graph. M is called a matching in G iff Definition 6 (weight of a matching, jM be a weighted graph and let M be a matching in G. w(fu; vg) is called the weight of matching It is obvious that jM j counts the number of edges in M when it is applied to an unweighted graph. In that case jM j is called the cardinality or size of the matching M : E) be a graph and let M ae E be a matching. For Definition 8 (maximum weighted matching, M(G)) be a weighted graph. Then M(G) denotes a maximum weighted matching We will frequently denote a specially defined or calculated maximum matching by Definition 9 (symmetric difference, \Phi) Let A and B be two sets. Then A \Phi B denotes the symmetric difference: To illustrate structures of graphs and matchings, we use a graphical notation. It is fully interchangeable with the set theoretical representation. Vertices of the two sets in bipartite graphs were depicted by small, filled circles v and squares respectively. The circles represent vertices from the request partition R, and squares represent vertices out of the server partition S. Additionally, the label of such a vertex is written right next to its symbol. When we sketch an instance of the wORSM problem, the vertices of the request partition are drawn in their order along a horizontal line from left to right. The server vertices are drawn in the same way in a distance below. Indeed, a request vertex is situated precisely on top of the server vertex of the same time step: Edges are depicted by a line between two vertices. Of course in bipartite graphs such a line has a circle and a square on its ends: To mark the edges of a matching they are symbolized by a double line: Whenever edges cannot be selected for the matching anymore, due to previous decisions made by the online algorithm, they were depicted in grey: 3 The Online Request Server Matching Problem This section starts with a general lower bound for the competitive ratio of the ORSM problem for deterministic online algorithms. Then the optimal 1.5-competitive algorithm LMM is presented and analysed in the next subsections. 3.1 The Lower Bound By applying the standard argument of an adversary strategy, we will show the following general lower bound: Theorem 1 Every deterministic online algorithm A for the ORSM problem has a competitive ratio of at least 1:5. Proof. The adversary strategy starts with the following input structure: @ @ @ @ @ Figure 1: Situation at time 2. A can react to this input at time in three different ways: Case 1: A puts the edge to the online matching MA . In the next step the adversary presents edge g. @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ Figure 2: situation at time A is not able to use the serving interval s 3 . Therefore, jMA j - 2 whereas the optimal solution gives jM OPT Case 2: A puts the edge to the online matching MA . In the next step the adversary presents edge g. @ @ @ @ @ Figure 3: situation at time A cannot use s 4 . Again jMA j - 2 and the optimal matching results in jM OPT Case 3: A decides not to match s 2 . The adversary will present the input of Case 1 (the input of Case 2 would work as well) and jMA @ @ @ @ @ @ @ @ @ @ Figure 4: situation at time This strategy can be infinitely repeated every four time steps and this fact shows the ratio 3: Theorem 1 3.2 The Algorithm LMM At a time step i the graph G representing the input of an online algorithm is known up to request vertex r i . More precisely, we know the subgraph of G induced by the set [S. Due to the irreversible decisions of the former time steps all previous server vertices matched request vertices r k cannot be rearranged anymore (M i is the online matching up to time have a vertex induced bipartite subgraph of G: ~ Our online algorithm is called 'Local Maximum Matching' (LMM) because it constructs a maximum matching on every local subgraph B i (denoted by M(B i )). The exact function of LMM follows: 1: 2: loop 3: 4: read input of time i and build up B i 5: construct a maximum matching M(B i ) on start with all matching edges of M(B i\Gamma1 ) which are edges in B look for an augmenting path which starts at vertex r i and do the augmentation when found 5 7: if s i 8: add the matching edge of s i to the online matching MLMM 9: else if s i is not isolated in B i then fall neighbours of s i are matched in M(B i )g 10: add an arbitrary edge fs to the matching MLMM and delete the matching edge of r in M(B i ) 12: end loop Line 10 of this algorithm is essential and prefers a matching which includes the current vertex s i . Now we will analyse the performance of LMM. 5 due to the maximum cardinality of M(B augmenting path must have r i on one end 3.3 The Upper Bound At first, two invariants of LMM will be formalized in lemmata. After a request vertex r i has been matched in B i for the first time, it is in all following maximum matchings r i to the time step where its current matching edge is added to MLMM and so r i holds in the end. Proof. In line 5 of LMM at time j the matching edge of a previous matched request vertex copied to M(B j ) and an augmentation in line 6 can change r's matching edge but r ffi In line 10 the matching edge of a vertex but after this step r ffi Lemma 3 If s i is not isolated in B i , then s i holds. The lemma follows directly from line 7 to 11 of LMM. Applying these two invariants we can prove the upper bound now. Theorem 4 LMM is 1.5-competitive. Proof. We will show that no online matching MLMM can be increased by augmenting paths of length one or three. Therefore, shortest augmenting paths for MLMM must have a length of five. This fact immediately results in the theorem because by performing the augmentation, two matching edges out of MLMM become three edges in M OPT . Longer augmenting paths have a lower ratio and matching edges outside such paths decrease the overall ratio as well. Let us fix an arbitrary input graph G I and a maximum matching M can compare M OPT with the matching MLMM constructed by LMM on G I . The symmetric difference M OPT \Phi MLMM defines a set of disjoint augmenting paths for MLMM (for more details see [Edm65]). The augmentation of all of these paths transforms MLMM to M OPT . By contradiction, we will prove the non-existence of paths of length one and three in this set. Figure 5: Structure of augmenting paths of length one and three. Case 1: augmenting path of length one fs Vertex r a was never matched because r a (reverse application of Lemma 2). a g is in B i and this fact contradicts Lemma 3. Case 2: augmenting path of length three and r a was matched at time j only, which implies the edge fs was in B i . Then Case 3: augmenting path of length three and i ? j: At time j, request vertices r a ; r b 2 MLMM and so the whole path ffs gg is in B j . The case s i contradicts the optimality of M(B j ) because the path is an augmenting one. Therefore, at time j s i must hold, i. e. there exists a request vertex r c with fr c ; s i Later at time k (j deletes the matching edge fr c ; s i g and adds fs to MLMM . Due to the definition of the ORSM problem, both edges are known at time j. Now the above argument about s i can be recursively applied to s k and due to the finite structure of B i , this fact contradicts the existence of an augmenting path of length three in MLMM . Theorem 4 4 The Online Request Server Weighted Matching Problem Like in section 3, we present a general lower bound for the wORSM problem first. Then the algorithm wLMM is given and analysed. Unfortunately this algorithm cannot achieve the lower bound and so we suggest the algorithm PHI at the end of this section. 4.1 The Lower Bound Let OE := 5+1- 1:618034 be the golden ratio. Theorem 5 Every deterministic online algorithm A for the wORSM problem has a competitive ratio of at least Proof. The adversary strategy starts with input edges and their weights are as you can see in Fig. 6. @ @ @ @ @ Figure Situation at time A can react to this input in two ways: Case 1: A adds edge to the weighted online matching MA . Then the adversary does not presents any new edge incident to s 2 . So jMA holds. Case 2: A does not change the online matching MA . Then the adversary presents edge with weight w(fr @ @ @ @ @ Figure 7: Situation at time Now A is able to construct a matching with weight jMA j - OE only, whereas it holds OE. The ratio of these two values is: OE Every two time steps the adversary can repeat this strategy up to infinity and this fact shows the lower bound of the competitive ratio of 5+1. \Xi Theorem 5 4.2 The Algorithm wLMM The algorithm wLMM works very similarly to LMM. Of course, wLMM determines a maximum weighted matching on the local bipartite graph B i . Furthermore, the algorithm works without the special preference of the vertex s i . Later, on page 30, we will explain why a special treatment of s i , in the way LMM does, cannot increase the performance of wLMM. Our investigation will indicate the problems arising from this fact. A formal description of wLMM follows: 1: 2: loop 3: 4: read input of time step i and build up B i 5: construct a maximum weighted matching M(B i 7: add the matching edge of s i to the online weighted matching MLMM 8: end if 9: end loop 4.3 The Performance of wLMM In this section we present a tight analysis of wLMM being 2-competitive. At first we exhibit a lower bound of wLMM which establishes the fact that wLMM is not able to achieve the lower bound of theorem 5. The Lower Bound Theorem 6 The online algorithm wLMM has a competitive ratio of at least 2. Proof. An adversary presents the input w(fr and in the next time step w(fr @ @ @ @ @ Figure 8: Input structure that is used by the adversary The algorithm wLMM determines the online matching which has a weight of jMwLMM ". The optimal solution is with weight jM OPT 2. Consequently, and the limit for Theorem 6 The Upper Bound Firstly, we repeat a few definitions and give new ones which will be used extensively in the following proofs. Most of the notations are similar and comparable to [BR93], whereas one different notation is introduced. By the use of this notation, we aim to make the formulae more accessible. Definitions and Notations is the bipartite, weighted graph of the wORSM problem with the weight function w M i is the online matching which is calculated up to step i by wLMM. M(B) is a maximum weighted matching of a bipartite graph B. is the overall weight of a maximum weighted matching of B. is the local bipartite graph of step i which consists of all known and not matched request vertices R i ae R with non-passed server vertices edges between this two sets (R i Both notations above can be combined. We use indexed r's and s's to denote vertices of R and S, so the partition where these vertices come from is unambiguous. Additionally, we employ a list of inserted or deleted vertices when necessary. A typical usage of this notation is B i which describes the vertex induced subgraph Gj R i Let M(B i ) be a maximum weighted matching of B i and let M(B i !s ) be such a matching of B i after removing the vertex s. The symmetric difference of these two sets results in a path P := !s ). It is an augmenting path 6 in B i with respect to the matching M(B i !s ). We will use notation P for the set of edges as defined above and 6 this path can be empty for a graph is the defined set of edges, and set V P gathers all the incident vertices. With this definition, we define the value of the corresponding server vertex as: It is obvious that fi i !s ) as above. Sometimes we call fi i (s) the weight of path P because this value is added to the overall matching when augmenting P . Next we define the global potential function: At last the following abbreviated notations will be used for an expression f : f jv is the value of f in step i with \Deltaf jv is the difference of the value of expression f before and after processing vertex v. Now some important properties of weighted augmenting paths which corresponds to fi(v) are listed. In this context we can relax the restriction to bipartite graphs. All above mentioned definitions applies to simple, undirected, weighted graphs. Lemma 7 be a simple, undirected, weighted graph. Then it holds: Proof. G !v is a subgraph of G and both M(G) and M(G !v ) are optimal. So it holds Lemma 8 Every subpath P b which arises from P by deleting an even number of edges from the start vertex v and which starts itself by vertex b is optimally matched by the matching of P , i. e. M(P b Every subpath P c which arises from P by deleting an odd number of edges from the start vertex v and which starts itself by vertex c is optimally matched by the matching of P !v , i. e. M(P c v a b c a b c s q q q c Figure 9: Sketch of the paths described in Lemma 8 including their optimal matchings. Proof. We prove statement (I) by contradiction. Assume M(P b is not optimal: The matching edge of b is in P b , fb; cg 2 M(P ) and fb; cg 2 P b because fv; ag 2 M(P ) and b is in a distance of even edges away from v. Path P can be divided into two subpaths at vertex b and it holds: Then the above assumption implies: which contradicts the optimality of M(P In the same way statement (II) is proven. The line of argumentation is about P !v and instead. Notice that in M(P !v ) the matching and non-matching edges are exchanged in comparison to M(P ) by definition of P . \Xi Lemma 8 Lemma 9 path P and subpaths P b and P c be defined as in Lemma 8. The value fi(v) can be expressed from vertex b or c onwards by fi(b) or fi(c) in the following way: Matching M(P ) can be divided into M(P nP b ) and M(P b ) by Lemma 8. In the same way M(P !v ) can be divided into M((P n P b ) !v Matching M(P ) can be divided into M((P n P c !c ) by Lemma 8. In the same way M(P !v ) can be divided into M((P n P c it holds: fi(c) Proof. The Lemma follows from Lemma 9 part (II) with the subpath P c starts with vertex u. Then m((P n P u is the empty graph with and the application of Lemma 7 (fi(u) - 0) completes the proof. Lemma 11 be a simple, undirected, weighted graph, v 2 V , and let fi(v) correspond to a path P of even length. When a manipulation of G increases the length of P , the value of fi(v) will never decrease, i. e. Proof. Suppose b to be the last vertex of path P and P b be the extension of P . We can use Lemma 9 part (I) because P has an even length. We then achieve: Then and by applying of Lemma 7: Lemma 12 be a simple, undirected, weighted graph, v 2 V , and let fi(v) correspond to a path P . When a manipulation on G decreases the length of P such that fi + (v) corresponds to a path of odd length, then the value of fi(v) will never decrease, i. e. Proof. Suppose c to be the last vertex of the reduced path (corresponding to fi and P c !c be the shortening itself. We can use Lemma 9 part (II) because P + has an odd length, and we achieve: Then and by applying of Lemma 7: Now we are able to give two key lemmata, which are needed to prove the upper bound. The statements of these lemmata are the same as in [BR93] and there are called identically. The line of argumentation in our proof follows E. Bernstein's and S. Rajagopalan's proof, too. The Key Lemmata and the Proof Lemma 13 (Stability Lemma) The value of a server vertex never decreases, i. e. \Deltafi Proof. When wLMM is running the following changes on the local graph B can happen: Case 1 'matching': A matching edge added to the online matching, This edge including its both vertices is removed from Case 2 'non-matching': The current server vertex s i is not matched (s i is removed from This can happen when the weights of edges incident to s i are too small. Case 3 'input': A new request vertex r i is added to We will show that in all three cases the value fi(s) of an arbitrary server vertex s 2 S i will not decrease. Let Q be the path corresponding to fi(s). Whenever Q is not affected by one of the above cases, it holds: fi Henceforth, it is sufficient to assume a modification in the structure of Q. Case 1 shortened by removing one of its matching edges. This removal includes all adjacent non-matching edges. Hence, the shortened path Q has a matching edge on both of its ends and therefore it is of odd length. An application of Lemma 12 gives \Deltafi Case path Q is shortened by removing its last vertex s i (all other vertices of Q are matched in M(B i )). Again the removing of s i removes all incident edges has two matching edges on its ends. So it is of odd length, and by Lemma 12 it holds \Deltafi first we focus on the possibility that the new request vertex is not matched, However, such a situation does not change any path Q and therefore it holds The reason is that r i could be the last, non-matched vertex of Q only. Then Q has an even length. By definition, Q starts with a server vertex s 2 S i which implies that all request vertices are situated in an odd distance to s. This fact implies a contradiction. Henceforth, we assume that r i is matched to a server vertex s j in the enlarged graph was augmented which is described exactly by M(B /r i We define the weight of path P just as the value of a server vertex Now we have to distinguish whether s j was matched in M(B) before. Case 3.a: s j M(B). No path Q of odd length could become extended because these paths have two matched end vertices. This is a contradiction to the assumption of this case. If Q is of even length and has been extended, then Lemma 11 can be applied and this results in \Deltafi Case 3.b: s j M(B). The following statements are implicated by the preconditions of this case: ffl P is an augmenting path of length ? 1 because s j was matched before. ffl There exists a vertex v which is the first common vertex of path P and Q. At that point, the paths meet first time with respect to their start vertices. From this facts we can conclude: would be matched to a vertex of Q which is not in P and to another vertex of P . This is a contradiction. 7 Note: the definition of fi(s) was made for server vertices s 2 S only s Figure 10: Situation just before P is augmented. ffl From vertex v onwards, paths P and Q are identical. Lemma 9 part (I) gives the reason for this statement 8 , since fi(v) is the same for both paths. Lemma 9 part (I) allows us to express fi(s) by fi(v). Due to the fact that no matching edge is changed between s and v on path Q (see also Fig. 10) it holds: Hence, it is sufficient to investigate \Deltafi (v). We can observe that v is matched to the farther request vertex with respect to r i , before path P is augmented. This augmentation exchanges all matching and non-matching edges on P . Then, v is matched to the request vertex that is closer to r i and the path corresponding to fi + (v) has the 'opposite direction', i. e. it starts at vertex v and ends at vertex u where u is situated between v and r i . \Gamma! fi(v) This vertex u has to be a request vertex, u 2 R i . Otherwise it would be matched in M(P ) outside the path described by fi + (v) which is a contradiction to the definition of fi Vertex u and r i can be identical. 8 Note: A very careful inspection shows the possibility that this statement does not necessarily hold. On the one hand fi(v) is an optimized value and it is equal in P and Q. On the other hand, there can exist two different paths P tail and Q tail which start at vertex v and have the same value fi(v). Without loss of generality, select the maximum matchings which define P such that Q tail is the part of P form v onward. It is obvious that these matchings exist and that we get proper definitions for P , fi(s j ), and The definition of \Deltafi (v) gives: Substituting this claim in equation (2): and by Lemma 7 fi 0 (u) - 0, which completes the proof. Lemma 7 can be applied here because it holds for the value fi(\Delta) of a vertex in a general, simple, undirected graph. Proof of the claim: On the left hand side, the vertex v is removed out of path P in both terms. Hence, P is divided into two subpaths is situated in P 1 . The maximum weighted matching of P 2 is the same in M(P !v ) and M(P !r i ;v ), and it has no influence on the difference m(P !v Let P r be a subpath between r i and u, and let P u be the subpath of P from u onward. z - By definition of fi 0 (r i ) and Lemma 8: To establish the statement m(P r sufficient to have a look at P 1 . When P was augmented (from M(P !r i ) to M(P matching and non-matching edges were exchanged. After the removal of vertex v, path P 1 is augmented, and again all matching and non-matching edges are exchanged in the subpath between v and u which is described by fi (v). In this subpath we have the same situation as before when r i had been inserted. Then the difference between M(P 1 ) and M(P 1 be found in P r (to be more precise: M(P 1 From the above equation, we get: Lemma 14 (Pay-Off - Lemma) During a run of wLMM holds: Before proving the lemma we would like to present an interpretation of the formulae. Statement (I) establishes the fact that the potential function \Phi increases by at least the value of fi(s) when the server vertex s is processed. Statement (II) ensures the choice of the heaviest augmenting path when a request vertex r is inserted and matched. The left hand side of this inequation is the weight of the selected augmenting path whereas the right hand side describes the weight of all possible augmenting paths that start in vertex r (see Lemma 9 part (II)). Proof. For simplicity of notation the index i is omitted. and Substituting the above equations into the definition of \Delta\Phi js : \Delta\Phi and by Lemma 10: \Delta\Phi Case 2: 'non-matching', s and Substituting the above equations into the definition of \Delta\Phi js : \Delta\Phi and together we have: \Delta\Phi (proof by contradiction) Let r 2 R and m(B /r ) be the overall weight of an optimal matching of B including vertex r. Assumption: 9 s 2 S i such that The matching M is not changed by the insertion of r and we have: \Delta\Phi Substituting the assumption by this equation: \Delta\Phi The term is the weight of a matching in graph B /r because both vertices r and s are not in B !s . This contradicts the assumed optimality of m(B /r ) and the lemma is proven. \Xi Lemma 14 Theorem 15 The deterministic online algorithm wLMM for the wORSM problem is 2-competitive. Proof. For all request vertices r 2 R we get from Lemma 14 part (II): while the definition of the wORSM problem ensures that r is processed before s : We get from Lemma 14 part (I) for all server vertices s 2 S: and by the use of Lemma 13, since s is processed after r : The sum of equations (4), (5) and (6) is Then, for an arbitrary matching MG of G holds: \Delta\Phi js - Hence, we get for final - jM OPT j and from the definition of \Phi, and the fact that B final = ?, it follows: which shows the 2-competitiveness of wLMM. \Xi Theorem 15 A Comment on the wLMM-Algorithm Why does wLMM not prefer the current server vertex s i in the online matching, like LMM does? The answer to that question is a very short one: It does not help anyway. Suppose an algorithm wLMM which prefers the current server vertex in the way like LMM does. Furthermore, assume an input graph such that MwLMM 6= MwLMM . Let s i be the first vertex where wLMM makes a different decision than wLMM (s i and s i decreasing all weights of edges incident to s i by ", wLMM will behave like wLMM on s i . With the help of this trick, we can construct an input graph calculates the same matching as wLMM does on input G. The difference in the weights of the resulting online matchings jMwLMM j and jMwLMM j is small and disappears for " Possibly, the next section shows a way out. 4.4 The Algorithm PHI The algorithm wLMM implements a special greedy strategy. From its current point of view, it takes the maximal, additional weight for the online matching. Theorem 6 shows a situation where this strategy is not very clever. With regard to theorem 5 (construction of the general lower bound) and theorem 6 we suppose that there is an algorithmic improvement. Using the current vertex s i in the online matching M , and therefore removing a vertex r out of B i simultaneously, may be more valuable than the originated loss in M(B i ). This observation leads us to suggest another online algorithm for the wORSM problem. It works similarly to wLMM but after computing the local maximum weighted matching M(B i ), it checks the vertex s i . Whenever s i holds, the weights of all edges of B i incident to s i will be increased by the factor 5+1. This new local bipartite graph is called Now the algorithm determines M(B OE matching edge is added to the online matching M . The new online algorithm is called PHI and a formal description follows: 1: 2: loop 3: 4: read input of time i and build up B i 5: construct a maximum weighted matching M(B i 7: construct B OE fincrease the weight of every edge of B i incident to s i by factor 8: calculate the maximum weighted matching M(B OE 9: end if 11: add the matching edge of s i to the online weighted matching M PHI 12: end if 13: end loop A problem still unsolved is the performance analysis of PHI. It seems as if a modification of the technique, that is used in the proof of theorem 15 does not work. It is likely that ), then the algorithm does not need to calculate B OE ) and to test s i these facts (the necessity of a modified algorithm like PHI and a new technique to analyse it) also explain the gap in the analysis of the roommates problem in [BR93]. It is worth noting that the algorithm LMM is a special implementation of PHI for unit edge weights. The increasing of unit weights of s i -edges by a factor of OE is equivalent to the simple preference of an edge incident to s i at time step i. 5 Open Problems In the end of the last subsection we presented the most urgent research task: To close the gap between the lower and upper bound of the wORSM problem (Theorem 5 and Theorem 15). If we achieve this aim, the roommates problem (weighted version) of [BR93] should be revisited. A completely different research task deals with the ORSM problem (unweighted version). Here, the performance on inputs of strong, pre-determined structures is of interest. Finally, we would like to model a set of parallel, homogeneous resources and requests with deadlines. First results have been already proven. It turned out that dependent on the concrete model: ffl lower competitive ratios are achievable, ffl the algorithm LMM is not optimal anymore, ffl different arguments to prove upper bounds are necessary. These investigations are not completed by now but they will be presented in a later publication --R The competitiveness of on-line assignments On the power of randomization in on-line algorithms Online Computation and Competitive Analysis. The roommates problem: Online matching on general graphs. Network flow and testing graph connectivity. Online computation. Online weighted matching. An optimal deterministic algorithm for online b-matching Online matching with blocked input. An optimal algorithm for on-line bipartite matching Amortized efficiency of list update and paging rules. --TR An optimal algorithm for on-line bipartite matching Online computation Online computation and competitive analysis Simple competitive request scheduling strategies Online Scheduling of Continuous Media Streams On-line Network Optimization Problems On-line Scheduling The Roommates Problem

competitive analysis;online scheduling;online bipontite matching

504566

Reductions for non-clausal theorem proving.

This paper presents the TAS methodology as a new framework for generating non-clausal Automated Theorem Provers. We present a complete description of the ATP for Classical Propositional Logic, named TAS-D, but the ideas, which make use of implicants and implicates can be extended in a natural manner to first-order logic, and non-classical logics. The method is based on the application of a number of reduction strategies on subformulas, in a rewrite-system style, in order to reduce the complexity of the formula as much as possible before branching. Specifically, we introduce the concept of complete reduction, and extensions of the pure literal rule and ofthe collapsibility theorems; these strategies allow to limit the size ofthe search space. In addition, TAS-D is a syntactical countermodel construction. As an example of the power of TAS-D we study a class of formulas which has linear proofs (in the number of branchings) when either resolution or dissolution with factoring is applied. When applying our method to these formulas we get proofs without branching. In addition, some experimental results are reported. Copyright 2001 Elsevier Science B.V.

Introduction Much research in automated theorem proving has been focused on developing satisfiability testers for sets of clauses. However, experience has pointed out a number of disadvantages of this it is not natural to specify a real-world problem in clause form, the translation into clause form is not easy to handle and, although there a number of e#cient translation methods, models usually are not preserved under the translation. In addition, clausal methods are not easy to extend to non-classical logics, partially because no standard clause form can be defined in this wider setting. Non-clausal theorem proving research has been mainly focused on either tableaux methods or matrix-based methods; also some ideas based on the data structure of BDDs have been Partially supported by CICYT project number TIC97-0579-C02-02. stands for Transformaciones de - Arboles Sint-acticos, Spanish translation of Syntactic Trees Transformations used in this context. Recently, path dissolution [6] has been introduced as a generalisation of the analytic tableaux, allowing tableaux deductions to be substantially speeded up. The central point for e#ciency of any satisfiability tester is the control over the branching, and our approach is focussed on the previous reduction of the formula to be branched as much as possible before actually branching. Specifically, we introduce the concept of complete reduction, and extensions of the pure literal rule and of the collapsibility theorems. On the other hand, another interesting point in the design of ATPs is the capability of building models provided the input formula is satisfiable. A non-clausal algorithm for satisfiability testing in the classical propositional calculus, named TAS-D, is described. The input to the algorithm need not be in conjunctive normal form or any normal form. The output is either "Unsatisfiable", or "Satisfiable" and in the latter case also a model of the formula is given. To determine satisfiability for a given formula, firstly we reduce the size of the formula by applying satisfiability-preserving transformations, then choose a variable to branch and recursively repeat the process on each generated task. This feature allows: . to obtain useful information from the original structure of the formula. . to make clearer proofs. . to extend the method to non-classical logics which do not have a widely accepted normal form. Although our intention in this paper is to introduce the required metatheory, TAS-D is currently being tested and we are obtaining very promising results. In our opinion, the results of these tests allow to consider the TAS framework as a reliable approach to Automated Theorem Proving. TAS ideas are widely applicable, because apply to di#erent types of logics; flexible, because provide a uniform way to prove soundness and completeness; and, in addition, easily adaptable, because switching to a di#erent logic is possible without having to redesign the whole prover. In fact, it has been already extended to Classical First Order Logic [4], Temporal Logic [5] and Multiple-Valued Logic [1, 2]. The structure of the paper is as follows: 1. Firstly, the necessary definitions and theorems which support the reduction strategy are introduced in Section 2. 2. Later, the algorithm TAS-D is described in Section 3. 3. Finally, a comparative example is included in Section 4, which shows a class of formulas which has linear proofs (in the number of branchings) when either resolution or dissolution with factoring is applied [3, 7]. When applying TAS-D to these formulas we get proofs without branching. 1.1 Overview of TAS-D TAS-D is a satisfiability tester for classical propositional logic; therefore it can be used as a refutational ATP method and, like tableaux methods, it is a syntactical model construction. The reduction strategies are the main novelty of our method with respect to other non-clausal ATPs; like these methods, TAS-D is based on the disjunctive normal form. Its power is based not only on the intrinsically parallel design of the involved transformations, but also on the fact that these transformations are not just applied one after the other, but guided by some syntax-directed criteria, described in Sections 2.2 and 2.4, whose complexity is polynomial. These criteria allow us: 1. To detect subformulas which are either valid, or unsatisfiable, or equivalent to literals. 2. To detect literals # such that it is possible to obtain a equisatisfiable formula in which # appears at most once. Therefore, we can decrease the size of the problem as much as possible before branching. By checking these criteria we give ourselves the opportunity to reduce the size of the problem while creating just one subproblem; in addition, such reductions do not contribute to exponential growth. However, the most important feature of the reductions is that they enable the exponential growth rate to be limited. As an ATP, TAS-D is sound and complete and, furthermore, as a model building method, it generates countermodels in a natural manner. 1.2 Preliminary Concepts and Definitions Throughout the rest of the paper, we will work with a classical propositional language with connectives {-, #} and their standard semantics; V denotes the set of propositional denotes the set of literals; if # is a literal then # is its opposite literal; we will also use the usual notions of clause, cube, implicant, implicate and negation normal form (nnf). S # A denotes that S is a subformula of A, and S # A denotes that S is a proper subformula of A. An assignment I is an application from the set of propositional variables V to {0, 1}; the domain of an assignment can be uniquely extended, preserving the standard semantics, to the whole language. A formula A is said to be satisfiable if there exists an assignment I such that 1, in this case I is said to be a model for A; formulas A and B are said to be equisatisfiable if A is satisfiable i# B is satisfiable; formulas A and B are said to be equivalent, denoted A # B, if assignment I ; |= denotes the logical consequence; finally, the symbols # and # mean truth and falsity. The transformation of a formula into nnf is linear (by repeated application of De Morgan rules, the double negation rule and the equivalence A # B # -A#B), so in the following we will only consider formulas in nnf. In addition, by using the associative laws we can consider connectives # and # to have flexible arity, and expressions like A 1 # A n or A 1 #A n to be well formed formulas. We will use the standard notion of tree and address of a node in a tree. An address # in the syntactic tree of a formula A will mean, when no confusion arises, the subformula of A corresponding to the node of address # in will denote the address of the root node. Similarly, when we say a subformula B of A we mean an occurrence of B in A; if B # A, denotes the address of the node corresponding to B in is a set of literals, then # 1 , # 2 , . , # n }. is a set of literals in A and #}, then the expression A[#] denotes the formula obtained after substituting in A, for all #, every occurrence of # by #, and # by #. and C are formulas and B # A, then A[B/C] denotes the result of substituting in A any occurrence of B by C. If {# 1 , . , # n } is a set of literals in A and C i are formulas, then the expression A[# 1 /C 1 , . , # n /C n ] denotes the formula obtained after substituting in A, for all i, every occurrence of # i by C i . If # is an address in A and C is a nnf, then the expression A[#/C] is the formula obtained after substituting in A the subtree rooted in # by C. Adding Information to the Tree: #-lists and #-sets The idea underlying the reduction strategy is the use of information given by partial assignments (extensively used in Quine's method [8]) just for unitary assignments but, as we will show, in a powerful manner. We associate to each nnf A two lists 2 of literals denoted (the associated #-lists of A) and two sets, denoted c c whose elements are obtained out of the associated #-lists of the subformulas of A. The #-lists and the #-sets are the key tools of our method to reduce the size of the formula being analysed for satisfiability. 2.1 The #-lists In a nutshell, are, respectively, lists of implicates and implicants of A. The purpose of these lists is two-fold: firstly, to transform the formula A into an equivalent and smaller-sized one (Section 2.2), and secondly, by means of the c # b sets (Sections 2.3 and 2.4), to get an equisatisfiable and smaller-sized one. Their formal definition is the following: 1: Given an nnf A, are recursively defined as follows: In addition, elements in a # 0 -list are considered to be conjunctively connected and elements in a # 0 -list are considered to be disjunctively connected, so that some simplifications are applied. Namely, if {# 0 (A), then # 0 (A) is simplified to #, and if {# 1 (A), then is simplified to #. We use lists in lexicographic order just to facilitate the presentation of the examples. The reader can interpret them as sets. The intuition behind the definition is easy to explain, since in # 0 ( we intend to calculate implicates (for it is # 0 ), and since the union of the implicates of each conjunct is a set of implicates of the conjunction, then we use S , and so on. Example 1: 2. 3. ps # st 2.2 Information in the #-lists In this section we study the information contained in the #-lists of a given formula. Our first theorem states that elements of # 0 (A) are implicates of A, and elements of # 1 (A) are implicants of A, and follows easily by structural induction from the definition of # b -lists. Theorem 1 Let A be a nnf and # be a literal in A then: 1. If # 0 (A), then A |= # and, equivalently, A # A. 2. If # 1 (A), then # |= A and, equivalently, A # A. As an immediate corollary of the previous theorem we have the following result on the structure of the #-lists: Corollary 1 For every nnf A we have one and only one of the following possibilities: . There is b # {0, 1} such that # b . #, and then A #. The following corollary states a condition on the # 1 -lists which directly implies the satisfiability of a formula. then A is satisfiable, and if # 1 (A), then any assignment I such that is a model for A. On the other hand, the following result states conditions on the #-lists assuring the validity or unsatisfiability of a formula. Corollary 3 Let A be a nnf, then 1. If in which a conjunct A i 0 is a clause such that # 1 A #. 2. If A = W n in which a disjunct A i 0 is a cube such that # 0 A #. Proof: 1. Let using the results of Theorem 1, if A i 0 Therefore: _ _ 2. It is similar to the previous one. Definition 2: If A is a nnf, to #-label A means to associate to each node # in A the ordered Let us name those formulas whose #-lists allow to determine either its validity or its (un)satisfiability. Definition 3: A nnf A is said to be . finalizable if one of the following conditions holds: 1. 2. This definition will be applicable to the current formula when it is detected to be (un)satisfiable. The following three definitions are referred to subformulas of the current formula which are detected to be either valid, or unsatisfiable, or equivalent to a literal. . # 1 -conclusive if one of the following conditions holds: 1. 2. and a disjunct A i 0 is a cube such that # 0 . # 0 -conclusive if one of the following conditions holds: 1. 2. and a conjunct A i 0 is a clause such that # 1 . #-simple if A is not a literal and # #(rs, nil ) #(nil, qst ) #(qrs, nil ) s Figure 1: The tree The previous results state the amount of information in the #-lists which is enough to detect (un)satisfiability; when all these results are applied to a given formula, the resulting one is said to be #-restricted, and its formal definition is the following: Definition 4: Let A be an nnf, then it is said that A is #-restricted if it satisfies the following conditions: . it is not finalizable, . it has no subtree which is either # 0 -conclusive, or # 1 -conclusive, or #-simple, . it has neither # nor # leaves. 3 From the previous results we can state that if A is a nnf, then by repeatedly applying the following sequence of steps we get a #-restricted formula: 1. #-label. 2. Substitute subformulas B # A by either # (if B is # 1 -conclusive), or # (if B is # 0 - conclusive), or a literal # (if B is #-simple). 3. Simplify logical constants (# or #), as soon as introduced, by using the 0-1-laws. 4. Check for (un)satisfiability of A (namely, chech whether A is finalizable). Example 2: Given the formula a linear transformation allows to get a nnf which is equivalent to its negation, A, depicted in Fig. 1 (for readability reasons, leaves are not labelled in the figures). When #-labelling A, the method finds that node 6 (the right-most branch) is s-simple. The s-simple subtree is substituted by s and then formula B in Fig. 2 is obtained. New applications of the #-lists to get information (up to equivalence) of a formula A are given by the following theorem and its corollary. 3 Although the input formula is supposed not to contain occurrences of logical constants, they can be introduced by the reductions as we will see. #(rs, nil ) #(nil, qst ) r s Figure 2: The tree TB . Theorem 2 Let A be a nnf and # be a literal in A, then: 1. If # 0 (A), then A # A[#] 2. If # 1 (A), then A # A[#] Proof: 1. Let I be an assignment; we have to prove that . If Theorem 1 (item 1) since # 0 (A), we have that A # A, therefore Now the result is obvious. . If The second item is proved similarly. As an immediate consequence of the previous theorem, the following satisfiability-preserving result can be stated, which will be used later: Corollary 4 Let A be a nnf. If # 0 (A), then A and A[#] are equisatisfiable. Furthermore, if I is a model of A[#], then the extension I # of I such that I #) = 1 is a model of A. The following theorem allows to substitute a whole subformula C of A (not just literals as in Theorem 2) by a logical constant. Theorem 3 Let A be a nnf, C # A then: 1. If 2. If 3. If # 0 (A) and # 0 (C), then A # A[C/#] 4. If # 0 (A) and # 1 (C), then A # A[C/#] Proof: 1. By Theorem 1, we have A # A and C # C. Let I be an interpretation, then The rest of the items are proved similarly. 2.3 The #-sets In the previous section, the information in the #-lists has been used locally, that is, the information in # b (#) has been used to reduce node #, by using Theorem 1. In this section, the purpose of defining a new structure, the #-sets, is to allow the globalisation of the information, in that the information in # b (#) can be refined by the information in its ancestors. Given a #-restricted nnf A, we define the sets c whose elements are pairs (#) where # is a filtered # b -list associated to a subformula B of A, and # is the address of B in A. In Section 2.4 we will see how to transform the formula A into an equisatisfiable and smaller sized one by using these sets. The definition of the #-sets is based on the Filter operator which filters information in the #-lists according to Theorems 2 and 3. Specifically, some literals in the #-lists can allow to substitute a subformula by either # or # as a consequence of Theorem 3; on the other hand, when this theorem is not applicable, it is still possible to delete the rest of which are dominated, as an application of Theorem 2. In fact, the dominated literals will not be deleted, but framed, because they will be used in the extension of the mixed collapsibility theorem. Given a #-restricted nnf A and B # A we have: . Filter(# 0 (B)) is: 1. #, if there is a literal A. This is a consequence of Theorem 3, items 1 and 3. 2. The result of framing any literal A. This is a consequence of Theorem 2, items 1 and 2 resp. . Filter(# 1 (B)) is 1. #, if there is a literal A. This is a consequence of Theorem 3, items 2 and 4. 2. The result of framing any literal A. This is a consequence of Theorem 2, items 2 and 1 resp. Definition 5: Let A be a #-restricted formula. For b # {0, 1}, the set c recursively defined as follows: . If # is a literal, then c c . Otherwise, c a subformula of A and # b (B) #= nil} In the following example we present a step-by-step calculation of c -sets. Example 3: Consider the formula A whose #-labelled tree appears below, #(nil, rs ) r s # (nil, ps ) r #(nil, q ) s q #(nil, qrs ) s For this tree, we have c Literal p in nodes 3111, 3112 and 311 is framed because of its occurrence in # 0 literal - s in node 3112 is framed because of the occurrence of s in # 1 On the other hand, the c for A is the following: c q-s, 22122), q-r, 221), ( - qrs, 222), ( - q, 22), (-p-q, 2), (s, 3)} Node 211 is substituted by # because of the occurrences of - q in # 1 (2) and q in # 1 (211); and node 22121 is substituted by # because of the occurrences of p in # 1 (2) and - Finally, the occurrences of - are framed because of the occurrence of - p in # 1 (2), and the occurrences of - q are framed because of the occurrence of - q in # 1 (2). #b #-sets and meaning-preserving results In this section we study the information which can be extracted from the #-sets. This is stated in Theorem 4, and in its proof we will use the following facts about the #-sets of a given #-restricted nnf A: . No element in c is (#), since A is a #-restricted nnf and cannot be finalizable. . If (# c # b (A), then # is not the address of a leaf of A (since c c all literal #). . If # 0 (A), then (# c and the list # does not have framed literals. (Just note that a literal # is framed in (#) from the information in the #-lists of its ancestors). The following theorem states that, as the #-labels, the #-labels also allow substituting subformulas in A by either # or #. Theorem 4 Let A be a #-restricted nnf then 1. If (# c 2. If (# c Proof: 1. Suppose (# c let C be the subformula at address #. By the definition of c there exist a formula B such that C # B # A and a literal # satisfying By Corollary 1, using that # 0 (C), that the address # cannot correspond to a leaf, and that A is a #-restricted nnf (specifically, A does not have #-simple subformulas), we get that # 1 Note that, clearly, it is enough to prove that B # B[C/#]. Firstly, we will prove that, under these hypotheses: (a) If # 1 (B), then # 1 (B[C/#]). (b) If # 0 (B), then # 0 (B[C/#]). Proof of (a): By induction on the depth of # in B, denoted dB (#). (i) If dB to commutativity and associativity). It cannot be the case that (D), we would have # 1 (C), which contradicts the fact that # 1 Therefore, we must have consequently, B[C/# D. Now, using again # 1 we have that # 1 (D) and, therefore, # 1 (ii) Assume the result for dX us prove it for dB then the result is obvious. by the induction hypothesis we have that # 1 (B 1 [C/# 1 (B[C/#]). by the induction hypothesis, # 1 - The cases C # B 2 are similar. The proof of (b) is obtained by duality. Finally, we prove B # B[C/#] by considering the two possibilities in (1) above: . If # 0 (C) and # 1 (B), then B[C/#] by Theorem 3 # B[C/#] by (a) and Theorem 1 . If # 0 (C) and # 0 (B), then B[C/#] by Theorem 3 # B[C/#] by (b) and Theorem 1 2. The proof is similar. Note that this theorem introduces a meaning-preserving transformation which allows substituting a subformula by a constant. The information given by the #-lists substitutes subformulas which are equivalent to either # 1 -conclusive) or # 0 -conclusive); however, under the hypotheses of this theorem, it need not be true that # is equivalent to either # or #. A be an nnf then it is said that A is restricted if it is #-restricted and satisfies the following: . There are not elements (#) in c . There are not elements (#) in c If A is a nnf, to label A means #-label and associate to the root of A the ordered pair c c Note that given a #-restricted nnf, A, after calculating ( c c the (un)satisfiability of A or an equivalent and restricted nnf by means of the substitutions determined by Theorem 4, and the 0-1-laws.b #-sets and satisfiability-preserving results The following results will allow, by using the information in the #-sets, to substitute a nnf A by an equisatisfiable and smaller sized A # with no occurrences of some literals occurring in A. A complete reduction theorem To begin with, Corollary 4 can be stated in terms of the #-sets as follows: Theorem 5 (Complete reduction) Let A be a nnf such that (# c then A is satisfiable if and only if A[#] is satisfiable. Furthermore, if I is a model of A[#], then the extension I # of I such that I #) = 1 for all # is a model of A. Note that this result allows to eliminate all the occurrences of all the literals appearing in #, that is why it is named complete reduction. Its usefulness will be shown in the examples. Generalised pure literal rule The introduction of the #-sets allows a generalisation of the well-known pure literal rule for sets of clauses. Firstly, recall the standard definition and result for a formula in nnf: Definition 7: Let # be a literal occurring in a nnf A. Literal # is said to be pure in A if # does not occur in A. A be a nnf and # a pure literal in A then A is satisfiable i# A[#] is satisfiable. Furthermore, if I is a model of A[#], then the extension I # of I such that I #) = 1 is a model of A. Our #-sets allow to generalise the definition of pure literal and, as a consequence, to get an extension of the lemma above. Definition 8: Let A be a nnf. A literal # is said to be #-pure in A if it satisfies the following conditions: 1. # occurs in c 2. All the occurrences of # in c are framed. Next theorem is a proper extension of Lemma 1 (for it can be applied even when # and # occur in A). Theorem 6 (Generalised pure literal rule) Let A be a nnf, let # be a #-pure literal in A and let B be the formula obtained from A by the following substitutions: (i) If (# c with #, then node # in A is substituted by #]. (ii) If (# c with #, then node # in A is substituted by #. Then A is satisfiable if and only if B is satisfiable. Furthermore, if I is a model of B, then the extension I # of I such that I #) = 1 is a model of A. Proof: By Theorem 2, and the definition of the #-sets, we have that (a) If (# c in addition, we have #, then A # A[#])] (b) If (# c and if, in addition, we have #, then A # A[#])] Therefore, if we consider the formula A # , obtained when applying the equivalences of items (a) and (b), we get that literal # is pure in A # . Now, by an application of Lemma 1 to A # we get the formula B, which completes the proof. In the rest of the section we introduce the necessary definitions to extend the collapsibility results introduced in [9]. Collapsibility theorems Definition 9: Let A be a nnf and # 1 and # 2 literals in A. Literals # 1 and # 2 are 0-1-bound if the following conditions are satisfied: 1. There are no occurrences 4 of either # 1 or # 2 in c have that # 1 # i# 2 #. 2. There are no occurrences of either # 1 or # 2 in c we have that From the definition of #-sets we have that: Remark 1: If # 1 and # 2 are 0-1-bound in A, then every leaf in A with a literal in {# 1 has an ancestor # in A which is maximal in the sense that its associated #-lists satisfy one of the following conditions: is an ancestor of #, then none of the literals # 1 , # 2 , # 1 , # 2 occur in the #-lists associated to # . is an ascendant of #, then none of the literals # 1 , # 2 , # 1 , # 2 occur in the #-lists associated to # . We will use the following notation in the proof of the collapsibility results, where # i are literals, and b # {0, 1}: Theorem 7 (Collapsibility) Let A be a nnf and let # 1 and # 2 be literals in A. If # 1 and # 2 are 0-1-bound, then A is satisfiable if and only if A[# 1 /# 1 /#] is satisfiable. Furthermore, if I is a model of A[# 1 /# 1 /#], then any extension I # of I such that I # 1 is a model of A. Proof: The if part is immediate. For the only if part, let us suppose A is satisfiable. Let I be a satisfying assignment for A. If there is nothing to prove; so, let us consider prove that A is also satisfied by an assignment I # such that I # 1 From Remark 1, A can be considered as a formula in the language with the following set of atoms that is, in A every leaf is either a formula in S # or a literal # 1 Note that if S 1 then we have that I(S 1 Let I # be the assignment obtained from I by changing the values on # i as follows: I 4 In this section, when we say an occurrence of #, we mean an unframed occurrence of #. This assignment satisfies I # (S 1 coincides with I in the rest of leaves. Therefore I # is a satisfying assignment for A with I # 1 This result is a generalisation of van Gelder's collapsibility lemma, which treats the case in which all the occurrences of # 1 are bound as # 1 # 2 and # 1 # 2 can be represented by a single literal with # 1 # 2 , see [9] for the details. Our result drops the requirement that all the occurrences in the defining subset of # 1 and # 2 have to be children of a # node and the occurrences in the defining subset of {# 1 , # 2 } have to be children of a # node. Obviously, the previous result can be straightforwardly extended to the case of n literals which can be collapsed into one. Definition 10: Let A be a nnf and let # 1 , . , # n be literals in A, literals # 1 , . , # n are 0-1-bound if the following conditions are satisfied: 1. In c there are no occurrences of # 1 , . , # n and if (# c {1, . , n}, then we have that # i # i# j #. 2. In c there are no occurrences of # 1 , . , # n and if (# c {1, . , n}, then we have that # i # i# j #. Corollary 5 (Generalised collapsibility) Let A be a nnf, and let # 1 , . , # n be literals 0- 1-bound in A then A is satisfiable i# A[# 1 /# n-1 /# 1 /# n-1 /#] is satisfiable. Furthermore, if I is a model of A[# 1 /# n-1 /# 1 /# n-1 /#], then any extension I # of I such that I # j is a model of A. Example 4: van Gelder's reduction lemmas cannot be applied to the formula in Example 3, but it is collapsible in the sense of Theorem 7. We had the following #-sets: c c q-s, 22122), q-r, 221), ( - qrs, 222), ( - q, 22), (-p-q, 2), (s, 3)} Specifically, p and q are 0-1-bound. In order to state the generalisation of mixed collapsibility we need the following definition: Definition 11: Let A be a nnf, b # {0, 1} and let # 1 and # 2 be literals in A. Literal # 1 is b-bounded to # 2 if the following conditions are satisfied 1. In c (A) there are no occurrences (neither framed nor unframed) of either # 1 or # 1 . 2. If (# c (A), then we have that . If # 1 #, then # 2 #. . If # 1 # then # 2 #. By this definition if # 1 is b-bound to # 2 in a formula A, then every leaf of A belonging to has an ascendant # in S # (A). Theorem 8 (Mixed collapsibility) Let A be a nnf and # 1 , # 2 literals in A, 1. If # 1 is 0-bound to # 2 , and A # is the formula obtained from A by applying the following substitutions . If (# c is substituted in A by # 1 /# 1 /#]. . If (# c is substituted in A by # 1 /# 1 /#]. then A is satisfiable if and only if A # is satisfiable. In addition, if I is a satisfying assignment of A # , then any extension I # of I such that I(# 1 assignment for A. 2. If # 1 is 1-bound to # 2 , and A # is the formula obtained from A by applying the following substitutions . If (# c is substituted in A by #. then A is satisfiable if and only if A # is satisfiable. In addition, if I is a satisfying assignment of A # , then any extension I # of I such that I(# 1 assignment for A. Proof: 1. Note that A can be considered as a formula in the language with set of atoms that is, in A every leaf is either a formula in S # (A) or is a literal # 1 , # 1 }. Let I be a satisfying assignment for A: 1. if we have, by Theorem 2, that and for every leaf S in S # 1 # 2 ,0 (A) we have, again by Theorem 2, 2. If (A) we have, by Theorem 2, Consider the assignment I # , obtained from I by changing only its value on obviously, we have I(S) # I # (S). By monotonicity of boolean conjunction and disjunction, we have I # Conversely, let I be a satisfying assignment for A # and let I # any extension of I such that I Theorem 2, for every leaf S in S # 1 # 2 ,0 (A) we have and for every leaf S in S # 1 # 2 ,0 (A) we have 2. The proof is similar. Example 5: Following with the formula in Example 2, for the formula B in figure 2, we had c c therefore the first subtree can be pruned, obtaining the tree in Fig. 3. #(rs, nil ) # (nil, qst ) r s Figure 3: The tree TC . variables r and s can be deleted by Theorem 5 of complete reduction (for (rs, # c storing the information to be able to generate a model (if it exists) of the input formula. In addition, p is 1-bounded to t; therefore, by Theorem 8 of mixed collapsibility, (1) and (3) can be substituted by # and the information (p = t) is stored. The resulting formula is q # t, which is finalizable (for # 1 (q # Specifically, it is satisfiable and a model is We can deduce that the input formula in Example 2 is non-valid (for its negation is by collecting the stored information we get the following countermodel two possibilities: the information Splitting a formula We finish the section by introducing a satisfiability-preserving result which prevents a branching when suitable hypotheses hold. The splitting, as we call it, results as a consequence of the following well-known theorem. Theorem 9 (Quine) A is satisfiable if and only if A[p/# A[p/#] is satisfiable. Further- more, if I is a model of A[p/#], the extension of I with the assignment is a model of A; similarly, if I is a model of A[p/#], the extension of I with the assignment a model of A If no satisfiable-preserving reduction can be applied to a restricted conjunctive nnf, then we would have to branch. The following definition states a situation in which the formula has not to be branched but split. be a restricted nnf; A is said to be p-splittable if J p #J where Corollary 6 Let be a restricted and p-splittable nnf. Then A is satisfiable if and only if V i#Jp A i [p/# V i#J p A i [p/#] is satisfiable. Furthermore, if I is a model of A i [p/#], then the extension of I with the assignment is a model of A; similarly, if I is a model of then the extension of I with the assignment model of A. This result can be seen as a generalisation of the Davis-Putnam rule with the following advantages: . It is applicable to nnf, not only cnf. . It can be shifted to non-classical logics. . Its interactions with the reduction strategies turn out to be extremely e#cient. The advantage in the use of this transformation is that the problem is not branched but split in two subproblems where the occurrences of p are substituted by logical constants. Now, we can describe the algorithm of the prover following the steps we have applied in the previous examples. 3 The TAS-D algorithm In this section we describe the algorithm TAS-D and its soundness and completeness are proved. The flowchart of the algorithm appears in Figure 4; we have to keep in mind that: . TAS-D determines the (un)satisfiability of the input formula. Therefore, it can be viewed as a refutational ATP. . The data flow of the algorithm is a pair (B, M) where B is a nnf, and M is a set of expressions is a literal not occurring in B. . The elements in M define a partial interpretation for the input formula, which is used by CollectInfo, if necessary. This interpretation is defined as follows: In general, due to the second condition, M might define more than one interpretation, depending on the choosing of I(# ). The operators involved in the algorithm are described below, the soundness of each one follows from the results in previous sections: The initialisation stage: NNF The user's input A is translated into nnf by the operator NNF; specifically, where B is a nnf which is equivalent to A. A (B, M) Reduce Finalizable? Root #? Parallel Update SPReduce SubReduce QBranch Unsat. Model Satisfiable #-restrict -restrict Figure 4: Flowchart of the TAS-D algorithm. The Update module The di#erent stages in the algorithm transform subtrees of the input tree; in each trans- formation, the labels of the ascendant nodes (of the transformed node) are deleted; Update processes these trees by recalculating the missing labels and giving as output either a restricted nnf or a finalizable formula. From another point of view, this stage updates the formula in the sense that it prunes those subtrees that can be directly deduced to be equivalent either to #, or #, or a literal. The #-restrict operator The input of #-restrict is a pair (B, M), where B is a partially labelled formula, possibly with logical constants. Given a nnf B we have that #-restrict(B, M) = (C, M) where C is the #-restricted formula obtained from B as indicated in Definition 4. The #-restrict operator The input of #-restrict is a pair (B, M) where B is a #-restricted formula. We have where C is the restricted formula obtained from B as indicated in Definition 6. Parallelization The input of Parallel is a pair (B, M), where B is a restricted formula and We have Parallel( _ Since a disjunction is satisfiable i# some disjunct is satisfiable, each pair independently passed to Reduce, the following module in the algorithm. The Reduce module The input of Reduce is the labelled syntactic tree of a restricted nnf . In this stage we decrease, if possible, the size of B before branching, by using the information provided by the #-labels and the #-labels. Specifically, . the #-labels of the root node allow, using the SPReduce operator, to substitute B by an equisatisfiable formula in which some propositional variables have been eliminated; . the #-labels of a proper subtree X allow, using the SubReduce operator, to substitute the subformula X by an equivalent formula in which the symbols in its #-lists occur exactly once. The SPReduce operator A restricted nnf B is said to be SP-reducible if either it is completely reducible (i.e. there is an element (# c or it has #-pure literals, or it has a pair of 0-1 bound literals, or it has a literal b-bound to other literal; for these formulas we have obtained by applying the following items: 1. If (# c Theorem 5. 2. If # is #-pure in B, then C is the obtained formula after applying in B the substitutions in Theorem 6, and M 3. If # and # are 0-1-bound, then C is the obtained formula after applying in B the substitutions in Theorem 7, and M 4. If # is b-bound to # , then C is the obtained formula after applying in B the substitutions in Theorem 8, and The SubReduce operator The input of SubReduce is a restricted, not SP-reducible nnf A; its e#ect can be described as an application of Theorem 2 up to associativity and commutativity. The formal definition needs some extra terminology, included below: Definition 13: Let a such that is not SP-reducible, and consider W . The integers denoted by m(# j ), defined below, are associated to A: where | - | denotes the cardinality of a finite set. It is said that A is #-reducible if m(#) > 1 and associated with A } Let A be #-reducible and consider is defined as follows, by application of Theorem 2 A is subreducible if it has a subformula B such that one of the following conditions holds: . . . B is #-reducible for some literal #. By Theorem 2 we have that the subreduction preserves meaning, therefore where C is obtained by traversing the tree A depth-first in order to find the first subtree B indicated above, and 1. Apply Theorem 2, if either 2. Substitute B by B # , otherwise. The interest of using sub-reductions is that they can make possible further reductions. It is this use of reductions before branching one of the main novelties of this method with respect to others; specifically, the unit clause rule of the Davis-Putnam procedure is a special case of SP reduction; also [9] uses a weak version of our sub-reductions in his dominance lemma, but he only applies the substitutions to the first level of depth of each subformula. The Split operator The input of Split is a pair (B, M) where is a restricted and p-splittable nnf which is neither SP-reducible nor subreducible; we have These two tasks are treated independently by the Update process. Branching: the QBranch operator The input of QBranch is a pair (B, M) where B is a restricted nnf which is neither SP- reducible, nor splittable, nor sub-reducible, nnf. We have: These two tasks are treated independently by the Update process. Our experimental tests show that the best results are obtained when choosing p as the propositional variable with more occurrences in the formula being analysed (this information can be easily obtained from the #-sets). Collecting partial results: CollectInfo The CollectInfo operator collects the outputs of Update for each subproblem generated by either Parallel, or Split, or QBranch, and finishes the execution of the algorithm: . If all the outputs of the subproblems are #, then CollectInfo ends the algorithm with the output Unsatisfiable. . If some of the subproblems outputs (#, M), then CollectInfo ends the algorithm with output Satisfiable and a model, which is built from M. . If some of the subproblems outputs (A, M) satisfying ends the algorithm with output Satisfiable and a model is built from where # is the first element of # 1 (A). 3.1 Soundness and completeness of TAS-D The termination of the algorithm just described is obvious, for each applied process reduces the size and/or the number of propositional variables of the formula. Specifically, in the worst case, in which no reduction can be applied, the only applicable process is QBranch which decreases by one the number of propositional variables in the formula. Now, we can prove the soundness and completeness of TAS-D. Theorem 10 TAS-D(A)=Satisfiable if and only if A is satisfiable. Proof: It su#ces to show that all the processes in the algorithm preserve satisfiability. Process NNF clearly preserves the meaning, for it is the translation into nnf; all processes in the modules Update and Reduce preserve either meaning or satisfiability, by the results in Section 2. To finish the proof, one only has to keep in mind that the subproblem generating processes (Parallel, Split, QBranch) are based in the following fact: a disjunction is satisfiable if and only if a disjunct is satisfiable. So, the process CollectInfo preserves satisfiability as well. 3.2 Some complete examples Example Consider the formula The result of Update(NNF(-A)) is = (B, ?), where #({(pr, #)}, {( pq r , 1), ( pq, 2)} ) Now, as we have c reduction can be applied wrt p and as a consequence we get Update(SPReduce(B, and then the output is "-A is Unsatisfiable", therefore A is valid. Example 7: Consider the formula we have #(ps, nil ) s q # (nil, rs ) r s with c c the reduction module does not apply to this tree, that is, (B, ?) is the input of QBranch. We apply QBranch wrt variable p and obtain, The subproblem C 1 is studied below: After #-restrict we get the following tree # (nil, rs ) r s as For the second subproblem C 2 we have #(qs, nil ) rs ) r s for which c c #-restrict's output is fed into SPReduce; the formula can be completely reduced, for (qs, # c therefore, by applying substitutions [q/#] and [s/#] and simplifying the logical constants we get, which is # 0 -conclusive. Therefore, Update(C 2 As all the subproblems generated by QBranch output #, then CollectInfo produces the output "-A is Unsatisfiable", therefore A is valid. Example 8: Let us study the satisfiability of the formula in Example 4: # (nil, rs ) r s #(nil, ps ) s q #(ps, nil ) s The c -sets for the previous formula are the following: c c An application of #-restrict substitutes (211) and (22121) by #, the result is an equivalent formula B: #(nil, rs ) r s r s q #(nil, qrs ) #(ps, nil ) s c c Once again, the # in c allows to substitute (232) by #, obtaining the equivalent formula C: #(nil, rs ) r s s c c Therefore, Update(A, Now SPReduce can be applied, for literals p and q are 0-1-bound, we substitute all the occurrences of p by #, i.e. #-restrict(SPReduce(C, (D, # (nil, rs ) r s # (nil, rs ) r s s c c After substituting (311) by # we get : # (?,{(rs, 1), (qrs, 2), (qs, 3)} ) #(nil, rs ) r s #(nil, qrs ) In this formula q is 1-bounded to s and, SPReduce substitutes branches at addresses 2 and 3 by #; then As r # s is finalizable, for its # 1 #= nil, the stage CollectInfo ends the algorithm with output "A is Satisfiable" and the model determined by any interpretation I such that is a model of A. Note that I # defined as I # is also a model of A. 4 A comparative example To put our method in connection with other existent approaches in the literature, we will study the collection {T n } of clausal forms taken from [3], we also use their notation for the propositional variables. Consider, for instance, T 3 below: each clause contains atoms of the form p i , where # is a string of +'s and -'s. The superscripts in each clause always form the sequence 1,2, . , n. The subscript of each literal is exactly the sequence of signs of the preceding literals in its clause. When T n is built from T n-1 , each is added both positively and negatively. It is easy to see that T n has 2 n propositional variables, 2 n clauses, each of which contains n literals. In [3], Cook and Reckhow described the family {T n , n # 1} and showed that the study of its satisfiability is intractable for analytic tableaux but can be handled in linear time by resolution. In [7], Murray and Rosenthal showed that dissolution with factoring provides proofs for this class that are linear in the number of input clauses, |T n |. When we apply TAS-D to test the satisfiability of T n we get that it is subreducible. For instance, formula Reduce(T 3 ) can be expressed equivalently as the formula Thus, #-restrict reduces the previous tree, for there are four # 0 -conclusive subtrees (namely, the conjunctions p 3 simplifying the four # leaves, we get #. Therefore, when using TAS-D we can detect the unsatisfiability of the formulas T n with no branching at all. Conclusions We have presented a non-clausal satisfiability tester, named TAS-D, for Classical Propositional Logic. The main novelty of the method, in di#erence to other approaches, is that the reductions applied on each formula are dynamically selected, and applied to subformulas like in a rewrite system, following syntax-directed criteria. Specifically, we have introduced extensions of the pure literal rule and of the collapsibility theorems. This fact increases the e#ciency, for it decreases branching. As an example of the power of TAS-D we have studied a class of formulas which has linear proofs (in the number of branchings) when either resolution or dissolution with factoring is applied; on the other hand, when applying our method to these formulas we get proofs without branching. Acknowledgments The authors would like to thank Jos-e Meseguer and Daniele Mundici for their valuable comments on earlier drafts of this work. --R A reduction-based theorem prover for 3-valued logic Reducing signed propositional logics. The relative e Reduction techniques for translating into clause form by using prime implicates. Implicates and reduction techniques for temporal logics. Dissolution: Making paths vanish. On the relative meris of path dissolution and the method of analytic tableaux. Methods of logic. A satisfiability tester for non-clausal propositional calculus --TR A satisfiability tester for non-clausal propositional calculus Dissolution On the relative merits of path dissolution and the method of analytic tableaux Implicates and Reduction Techniques for Temporal Logics --CTR Jun Ma , Wenjiang Li , Da Ruan , Yang Xu, Filter-based resolution principle for lattice-valued propositional logic LP(X), Information Sciences: an International Journal, v.177 n.4, p.1046-1062, February, 2007 P. Cordero , G. Gutirrez , J. Martnez , I. P. De Guzmn, A New Algebraic Tool for Automatic Theorem Provers, Annals of Mathematics and Artificial Intelligence, v.42 n.4, p.369-398, December 2004 P. Cordero , G. Gutirrez , J. Martnez , I. P. De Guzmn, A New Algebraic Tool for Automatic Theorem Provers, Annals of Mathematics and Artificial Intelligence, v.42 n.4, p.369-398, December 2004

SAT problem;prime implicates/implicants;non-clausal theorem proving

504571

A typed context calculus.

This paper develops a typed calculus for contexts i.e., lambda terms with "holes". In addition to ordinary lambda terms, the calculus contains labeled holes, hole abstraction and context application for manipulating first-class contexts. The primary operation for contexts is hole-filling, which captures free variables. This operation conflicts with substitution of the lambda calculus, and a straightforward mixture of the two results in an inconsistent system. We solve this problem by defining a type system that precisely specifies the variable-capturing nature of contexts and that keeps track of bound variable renaming. These mechanisms enable us to define a reduction system that properly integrates&brg;-reduction and hole-filling. The resulting calculus is Church-Rosser and the type system has the subject reduction property. We believe that the context calculus will serve as a basis for developing a programming language with advanced features that call for manipulation of open terms. Copyright 2001 Elsevier Science B.V.

Introduction A context in the lambda calculus is a term with a \hole" in it. The operation for contexts is to ll the hole of a context with a term. For the purpose of 1 This is the authors' version of the article to appear in Theoretical Computer Science. current a-liation: Department of Information Science, University of Tokyo, Bunkyo-ku, Tokyo 113-0033, Japan. 3 Atsushi Ohori's work was partly supported by the Japanese Ministry of Education Grant-in-Aid for Scientic Research on Priority Area no. 275: \Advanced databases," and by the Parallel and Distributed Processing Research Consortium, Japan. Preprint submitted to Elsevier Preprint 23 January 2001 explanation in this section, we write C[] for a context containing the hole indicated by [], and write C[M ] for the term obtained from C[] by lling its hole with M . For example, if C[] (x:[]+y) 3 then C[x+z] (x:x+z+y) 3. As seen from this simple example, the feature that distinguishes this operation from substitution of the lambda calculus is that it captures free variables. In the above example, x in x z becomes bound when it is lled in the context One motivation behind using contexts in the theory of lambda calculus is to study properties of open terms. Since the behavior of an open term depends on bindings of their free variables, in order to analyze its behavior, it is essential to consider possible contexts in which the open term occurs. Study of program analyses based on contexts such as observational equivalence [13,11] yields important results in analysis of programming languages. In these and most of other usages, context is a meta-level notion and its applicability to programming languages has largely been limited to meta-level manipulation of programs. We believe that if a programming language is extended with rst-class contexts, then the extended language will provide various advanced features that call for manipulation of open terms. Let us brie y mention a few of them. Programming environment. In conventional programming environments, programs must rst be compiled into \object modules", and they must then be linked together to form an executable program. Moreover, an executable program must be a closed term. If a programming environment can be extended with the ability to link various software components dynamically, then its exibility will signicantly increase. Since the mechanism of contexts we are advocating oers a way of performing linking at runtime, it would provide a basis for developing such an environment in a theoretically sound way. Distributed programming. In distributed programming, one often wants to send a piece of code to a remote site and execute it there. As witnessed by recently emerging Internet programming languages such as Java [4], this feature will greatly enhance the expressive power of distributed programming. One naive approach to send a program is to pack all the necessary resources as a closure and send the entire closure to the remote site. An obvious drawback to this approach is ine-ciency. Since in most cases, communicating sites share common resources such as standard runtime libraries, a better approach would be to send an open term and to make the necessary binding at the remote site. A typed calculus with rst-class contexts would provide a clean and type safe mechanism for manipulating open terms. First-class modules. A program using a module can naturally be regarded as an open term containing free variables whose values will be supplied by the module. One way of modeling a module exporting a set of functions F through identiers f would therefore be regarding it as a context that captures variables f bind them to F respectively. Using (or \opening") a module then corresponds to lling the hole of the context with the variables. This approach can provide a new foundation for exible module systems. In conventional languages with modules such as Modula-2 [18] and Standard ML [12], there is rigid separation between the type system for modules and that of terms, and allowable operations on modules are rather limited. Signicant potential advantage of the \modules-as-contexts" approach is that modules can be freely combined with any other constructions available in the language i.e. that modules are treated as rst-class citizens. Needless to say, an actual module system must account for various features such as type abstraction, type sharing and separate compilation, and the above simple view alone does not immediately provide a proper basis for module systems. We nonetheless believe that, when properly rened with various mechanisms for module systems studied in literature, the above approach will open up a new possibility for exible module systems. Indeed, a recent work by Wells and Vestergaad [17] shows a connection between their module language and our context calculus. The general motivation for this study is to develop a programming language with rst-class contexts that can represent those features in a clean way. Despite those and other potentially promising features of contexts, a language with rst-class contexts has not been well investigated. Lee and Friedman [10] proposed a calculus where contexts and lambda terms are two disjoint classes of objects: contexts are regarded as \source code" and lambda terms as \compiled code". This separation is done by assuming two disjoint variable name spaces: one for lambda terms and one for contexts. As a consequence, in their system, -reduction and ll-reduction are two disjoint relations without non-trivial interaction. Dami [2] also announced a system for dynamic binding similar to that of Lee and Friedman. While these approaches would be useful for representing source code as a data structure, they do not allow contexts of the language itself to be treated as rst-class values inside the language. Kahrs [9] have developed a combinatory term rewriting system that is compatible with contexts. However, contexts and hole-lling themselves are not represented as terms within the system of terms. Talcott [16] developed an algebraic system for manipulating binding structures. Her system includes suitable mechanisms for manipulating contexts. In particular, it contains holes and hole-lling which commutes with substitution. However, this is a meta-level system, and the issue of representing contexts and the associated hole-lling operation inside of the reduction system of lambda calculus is not addressed. One of the features of contexts is to bind variables through holes. In this sense, contexts are closely related to environments. Abadi et al. [1] developed the -calculus for explicit substitutions. Their motivation is similar in spirit to ours in that it internalizes a meta-level mechanism in the lambda calculus. However, they did not address the problem of rst-class treatment of substitu- tions. In revising the present article, the authors noticed that Sato et al. [14] recently developed an environment calculus where environments are rst-class values. In obtaining a con uent calculus, they also address the problem of variable binding in the presence of rst-class environments. Their solution to this problem has some similarity to ours, although more general mechanisms are needed for a calculus with rst-class contexts. We shall comment on this in some detail when we describe our approach in the next section. The goal of this paper is to establish a type theoretical basis for a programming language with rst-class contexts by developing a typed context calculus where lambda terms are simply a special case of contexts. In particular, contexts and lambda terms belong to the same syntactic category sharing the same set of variables, and substitution and hole-lling are dened on the same syntactic objects. This property is essential for achieving various features explained above. As observed in the literature [9,10], however, -reduction and ll-reduction for contexts do not mix well, and a (naive) integration of them yields an inconsistent system. The development of a meaningful calculus containing -reduction and ll-reduction both acting on the same set of terms constitutes a non-trivial technical challenge. Our main technical contribution is to establish that such a calculus is possible. We prove that the calculus is Church-Rosser and its type system has the subject reduction property. To obtain a con uent calculus, we have to overcome various delicate problems in dealing with variables, and to introduce several new mechanisms in the lambda calculus. Before giving the technical development, in the next section, we explain the problems and outline our solution. 2 The Problem and Our Solution It is not hard to extend the syntax of the (untyped) lambda calculus with constructors for contexts. In conventional study, holes in contexts are name- less. However, since our goal is to develop a calculus with rst-class contexts, we should be able to consider a context containing other contexts. This requires us to generalize contexts to contain multiple dierent holes, only one of which is lled by each hole-lling operation. One way to dene a uniform syntax for those contexts is to introduce labeled holes [9]. We use upper case letters labeled holes. To incorporate operations for contexts as terms in a lambda calculus, we introduce hole abstraction -X:M which abstracts hole X in term M and creates a term that acts as a context whose hole is X, and we introduce context application which denotes the operation to ll the abstracted hole in M 1 with term M 2 . For example, the context C[] (x:[] represented by the term and the context application term denotes the term obtained by lling the hole in the context with x call a subterm of the form (-X:M which contracts to the term obtained from M 1 by lling the X-hole in M 1 with M 2 . Dierent from the meta notation C[x application is a term constructor, which allows us to exploit the features of rst-class contexts by combining it with lambda abstraction and lambda application. For example, we can write a term like which is contracted to the above term. The goal of this paper is to develop a type system and a reduction system for the lambda calculus extended with the above three term constructors, i.e., labeled holes, hole abstraction and context application. The crucial step is the development of a proper mechanism for integrating variable-capturing hole- lling and capture-avoiding substitution in the lambda calculus. To see the problem, consider the term where we use dierent type faces (x, x and x) to distinguish dierent occurrences of variable x to which we should pay attention. The above term has two -redexes and one ll-redex. Our intention is that the inner x should be captured by the x when it is lled in hole X, while the outer x is free. The following reduction sequence produces the intended result. However, reducing any of the -redexes before the ll-redex will result in a dierent term. If we reduce the inner -redex before the ll-redex then the binding of inner x will be lost, yielding x y. If we reduce the outer -redex before the ll-redex, then the outer x is unintentionally captured by y depending on the order of the ll-redex and the other -redex. To avoid these inconsistencies, we should redene the scope of lambda binding to re ect the behavior of terms of the form (-X:M 1 )M 2 . Suppose there is a x in M 1 whose scope contains X. Since M 2 is lled in X, the scope of the x also extends to M 2 . This property implies the following two requirements. First, a -redex containing a hole X cannot be contracted. Secondly, when substituting a term containing x for a free variable in M 2 , the x in M 1 and the corresponding variables in M 1 and M 2 need to be renamed to avoid unwanted capture. In the above example, we should not contract the inner - redex before hole-lling, and when we contract the outer -redex before hole- lling, we should rename x and x before doing -substitution. The situation becomes more subtle when we consider a term like (w: ((z:w(x Since w is a variable, simple inspection of term w(x+z) no longer tells which variables in x+z should be regarded as bound. However, variable-capture will still occur when the hole abstraction is substituted for w. Our strategy to solve this problem is to dene a type system that tells exactly which variables should be considered bound, and to introduce a rened notion of -equivalence that reconciles hole-lling and -substitution. To tell which variables should be considered bound, we type a hole abstracted term -X:M with a context type of the form: where 1 is the type of the abstracted hole, 2 is the type of the term that will be produced by lling the hole in the context with a term, and describes the set of variables being captured when they are lled in the hole X. We call those variables interface variables. For example, would be typed as [fx : intg . int] ) int. However, if we simply list the set of actual bound variables surrounding X in -X:M as interface variables in its type, then we cannot rename those bound variables. Since -substitution can only be dened up to renaming of bound variables, this causes a problem in extending substitution to hole abstracted terms. For example, we cannot rename bound variable x in the term -X:(x:X It should be noted that the usual \bound variable convention" does not solve the problem. In the lambda calculus, we can simply assume that \all bound variables are dierent from the free variables" for each -redex. This is only possible when we can freely rename bound variables. As well known in the theory of lambda calculus, the above condition is not preserved by substitution. Even if we start with a term satisfying the bound variable condition, anomalous terms like the above may appear during -reduction. To avoid this problem, we separate actual bound variables in -X:M and the corresponding interface variables, and rene hole-lling to an operation that also performs variable renaming. For manipulation of binding structures, Talcott [16] developed a technique to pair a hole with a substitution. We use this approach and annotate a hole X with a variable renamer , which renames interface variables to the corresponding bound variables. We write X for the hole X annotated with . The above context can now be represented as the typed term where x is an interface variable and is renamed to a when it is lled in X. By this separation, bound variable a can be renamed without changing the type of this term. This allows us to achieve a proper integration of hole-lling and -substitution with terms of the form -X:M . The semantics of hole-lling is preserved by applying the renamer fa=xg to the term to be lled in X. For example, we have the following reduction for the example before. Yet another delicate problem arises when we consider the interaction between substitution and a term of the form MN . This construct may bind some variables in N . In order to determine those bound variables, we need to annotate this construct with the set of variables in N that will be bound by forming this term. Since this set must correspond to the set of interface variables of the context term M , a naive attempt would be to annotate the constructor MN with this set. A subterm of the example term might then be represented as the term (-X:(a:X fa=xg z). As we noted earlier, the variable x must be treated as bound variable. This implies that, when combining -substitution, this variable needs to be renamed. Unfortunately, this is impossible for terms of the form w fxg we cannot rename the corresponding interface variables of the hole abstracted term that will be substituted later for w. So, again we need to separate the set of interface variables in the type of hole abstracted term and the set of variables that will be captured when they are lled in the hole of the context. To achieve this, we annotate the constructor for context application with a renamer and write M N . The renamer renames variables in N that are to be bound by hole-lling to the corresponding interface variables in the hole abstracted term. Its eect is obtained by composing it with the renamer of the hole. Now the bound variables in N are independent of the corresponding interface variables, we can perform bound variable renaming. The above example can be correctly represented by the following term: (b In this term, both a and b are bound variables, which can be renamed without changing the typing of the term. Again, the semantics of hole-lling is preserved by applying the composition fa=xg ? fx=bg( fa=bg) of renamers fa=xg and fx=bg to the term to be lled in X. The following is an example of reduction involving renamer applications. (b 3: Another slightly more general alternative to M N is to make a renamer as a term constructor [:N ] and introduce a new type constructor n g . ] for this constructor. We believe that this is also possible. In our system, however, we shall not take this approach, since the only elimination operation would be (the modied version of) the hole-lling and therefore the additional exibility is not essential in achieving our goal of rst-class treatment of contexts. Based on the strategies outlined above, we have worked out the denition of the type system of the calculus, and its reduction system, and proved that the type system has the subject reduction property and that the reduction system is Church-Rosser. In the work by Sato et al. [14], a type-theoretical approach similar to ours was taken in order to identify the set of free and bound variables. However, their system does not fully address the problem of mixing such a construct with - substitution. Their calculus contains a term constructor e 1 intuitive meaning is to evaluate e 2 under the bindings provided by the environment e 1 . However, the reduction for nested application of this construction is restricted to variables, and does not act on general terms. Because of this restricted treatment, the subtle problem of -equivalence explained above does not arise in their system. The careful reader may have noticed that some aspects of contexts can already be represented in the lambda calculus. If one can predetermine the exact order of variables exported by a context and imported by a term to be lled in the context, then one can represent hole abstractions and context applications simply by functionals as seen in the following encoding scheme. A hole-lling of the form: can be represented as a lambda term of the form: However, such encoding eliminates the ability to bind variables through names, and it therefore signicantly reduces the benets of rst-class contexts we have advocated in the introduction. The rest of the paper is organized as follows. In Section 3 we dene the context calculus. Section 4 denes the reduction system and proves the subject reduction property and Church-Rosser property of the calculus. Section 5 concludes the paper with the discussion of further investigations. Appendix contains proofs of some of the lemmas. 3 The Calculus We use the following notation for functions. The domain and the codomain of a function f are written as dom(f) and cod(f) respectively. We sometimes regard a function as a set of pairs and write ; for the empty function. Let f; g be functions. We write f ; g for f [ g provided that dom(f) \ dom(g) = ;. We omit \;" if g is explicitly represented as a set, writing ff: : :g for f :g. The restriction of a function f to the domain D is written as f j D . The set of types (ranged over by ) of the calculus is given by the syntax: where b ranges over a given set of base types, and ranges over variable type assignments each of which is a function from a nite set of variables to types. We let x range over a countably innite set of variables; we let X range over a countably innite set of labeled holes; and we let range over variable renamers each of which is a function from a nite set of variables to variables denoted by fy 1 =x g. Let be a renamer. To avoid unnecessary complication, we assume that fy i ng \ fx i n). That is, a renamer changes each name in the domain of the renamer to a fresh name, if it is not an identity. A renamer is extended to the set of all variables by letting In what follows, we identify a renamer with its extension. However, we maintain that the domain dom() of a renamer always means the domain of the original nite function . The composition 1 ? 2 of two variable renamers 1 and 2 is the function such that The set of (unchecked) terms (ranged over by M) of the calculus is given by the syntax: A term -X:M binds the hole X in M . The denitions of bound holes and free holes are given similarly to the usual denition of bound variables and free variables in the ordinary lambda calculus. We write FH(M) for the set of Fig. 1. The sets of free and bound variables free holes in M . Since -X is the only binder for holes, this does not create any of the subtle problems we have explained for variables in our calculus, and therefore we can safely assume -renaming of bound holes just as in - congruence in the ordinary lambda calculus. In what follows, we regard terms as their -equivalence classes induced by bound holes renaming. The set of free variables, denoted by FV (M ), and that of bound variables of M , denoted by BV (M) are given in Figure 1. These denitions correctly model the eect of context application terms of the form M 1 M 2 which binds the variables in dom() in M 2 . In addition to the sets of free and bound variables, we need to distinguish three other classes of variables. Let M be a term containing a hole X . The variables in cod(), which we call free variable candidates, behave similarly to free variables if they are not abstracted in M ; The variables in the set dom(), which we call interface variable candidates, are the source of interface variables. To see the last one, consider a term which contains M 1 M 2 . The variables in cod(), which we call exported variables, are used to match the variables exported by the context M 1 with bound variables in M 2 . The formal denitions of the set FV C(M) of free variable candidates of M , and the set IV C(M) of interface variables candidates of M are given in Figure 2, and the denition of the set EV (M) of exported variables of M is given in Figure 3. Fig. 2. The sets of free variable candidates and interface variable candidates. Fig. 3. The set of exported variables We dene the set PFV (M) of potentially free variables of M as We are now in the position to dene the type system of the calculus. Since a term may contain free holes as well as free variables, its type depends not only on types of variables but also on types of free holes. A hole type is determined by a triple ([ . ]; ) consisting of type of a term to be lled, type assignment describing the set of interface variables and their types, and variable renamer which is used to keep track of the correspondence between bound variables and interface variables. While describes the set of all abstracted variables, describes the set of free variable candidates to be abstracted. We for the triple obtained from ([ . ]; ) by abstracting x, whose denition is given below. . ]; ) This operation is extended to type assignments as follows: A hole type assignment , ranged over by , is a nite function which assigns a hole to a triple ([ . ]; ) describing the type of the hole, and we call the variables in dom() interface variables. We write Clos(; ) for the hole type assignment dom()g. We write for the hole type assignment g. The type system of the calculus is dened as a proof system to derive a typing of the form: which indicates that term M has type under variable type assignment and hole type assignment . The set of typing rules is given in Figure 4. Some explanations are in order. Rule (hole). Since X is not surrounded by any at this moment, the associated type assignment in the hole type assignment is empty, and the set of variable candidates of X is specied by . They will be abstracted by the rule (abs) and (ll). Rule (abs). Lambda abstracting x not only discharges x from the type hypothesis for the term M , but also extends the set of interface variables for each hole in M with corresponding x 0 . The later eect is represented by the operation Clos(fx : g; ), which extends each appearing in . Rule (ll). By forming the term M 1 fx 0=x 1 ;:::;x 0 =xng M 2 , each x i in M 2 becomes bound, and the set of interface variables of each hole in M 2 is extended with it. This property is modeled by discharging each x i from the typing judgment for M 2 and abstracting it from 2 . This rule is similar to the one for a \closure" i.e., a term associated with an explicit substitution, in -calculus [1]. Figure 5 shows an example of typing derivation. (abs) (app) (ll) if dom() \ fx 0 Fig. 4. The Type System Fig. 5. Example of Typing Derivation In our calculus, each free hole occurs linearly in a well-typed term. If multiple occurrences of a hole are allowed, then they could have dierent interface variables. This would considerably complicate the conceptual understanding of contexts as well as the type system. The linearity condition is ensured by the rule (hole), the condition implied by the notation 1 ; 2 in rules (app), (ll), and the property that there is no rule for adding redundant hypothesis to . The following lemma is easily shown by induction on the typing derivations. dom(). Moreover, each free hole appears exactly once in M . The following standard properties also hold for this type system, and can be easily shown by induction on the typing derivations. Lemma Lemma Lemma 7 If X fx 0 =yng occurs in M , fz ng \ fx 0 . ]; fy 0 . ]; fy 0 M 0 is obtained from M by substituting X fx 0 =yng for =wng . 4 The Reduction System To dene the reduction relation, we need to dene substitution and hole-lling operations. In the ordinary lambda calculus, substitution can be dened modulo -congruence, which allows us to simply assume that unwanted variable capture will not happen. In our calculus, since we have not yet obtained - congruence, we need at rst to dene substitution as an operation on syntactic terms (not on equivalence class). We write fM 0 =xgM for the term obtained by substituting M 0 for any free occurrence of x in M . The following lemma shows that substitution preserves typing under a strong variable hygiene conditions. Lemma The proof is deferred to the Appendix. As in the standard denition of sub- stitution, we have the following composition lemma: y. As we have explained earlier, hole-lling involves application of the variable renamer associated with the hole to the term being lled. To dene hole-lling, we extend a variable renamer to a function on terms as follows: We have the following renaming lemma, whose proof is deferred to the appendix Hole-lling is dened as a combination of variable renamer and substitution. We write M [M 0 =X] for the term obtained from M by syntactically substituting the term (M 0 ) for X in M where is the variable renamer associated with X. Its denition is obtained by simply extending the following clauses according to the structure of M . From this denition and the property of typing, it is easily seen that if ; ' . The following lemma shows that hole-lling preserves the typing. Lemma The proof is deferred to the appendix. The following is the composition lemma for the hole-lling, where IV CX (M) denotes the domain of the variable renamer on the shoulder of hole X in M . Lemma 12 PROOF. If dom( 1 )\(PFV (M)ndom( 2 M)).The notion of -congruence in our calculus is now dened as the congruence relation on the set of well typed terms generated by the following two axioms: =yng fy 1 =x if each y i 62 BV The following lemma shows that -renaming preserves typing, which is proved by induction on the derivation of M using lemma 10. -congruence allows us to rename bound variables whenever it is necessary. In what follows, we assume the following variable convention for our calculus: bound variables are all distinct and the set of bound variables has no intersection with the set of interface variable candidates, the set of potentially and the set of exported variables. Under this variable convention, the reduction axioms of our calculus are given as follows: ll In the axiom (), the restriction FH(M 2 is needed to ensure that each linearly. The restriction FH(M 1 is needed to maintain the binding generated by x for the holes in M 1 . Since in our calculus contexts are represented not by terms with free holes but by hole abstracted terms, this does not restrict rst-class treatment of contexts. The one-step reduction relation M ! N is dened on the set of well typed terms as: is well typed and M 0 is obtained by applying one of the two reduction axioms to some subterm of M . We write M the re exive, transitive closure of !. For this reduction, we have the following desired results. Theorem 14 (Subject Reduction) If ; (app) (ll) Fig. 6. Denition of the Parallel Reduction PROOF. This is a direct consequence of lemmas 7,8, 11 and 13. 2 Theorem 15 (Con uence) For any well typed term M , if M then there is some M 3 such that M 1 The proof is by using the technique of parallel reduction due to Tait and Martin-Lof. The parallel reduction relation of our calculus, written ! !, is given in Figure 6. From this denition, it is easily seen that the transitive closure of the parallel reduction coincides with the reduction relation of the calculus ( !). To prove the theorem, it is therefore su-cient to prove the diamond property of ! !. To show this, we follow Takahashi [15] and prove the following stronger property. Lemma 16 For any well typed term M , there exists a term M such that if In the lemma above, M denotes the term obtained from M by parallel reducing all the possible redexes of M , whose denition is given Figure 7. The proof of lemma 16 is by induction on the derivation of ; using the following lemmas: 1 is not a lambda abstraction 1 is not a hole abstraction Fig. 7. Denition of M Lemma terms M;M 0 such that fx : PROOF. We proceed by induction on the derivation of M ! !M 0 . Here we only show the cases (betred) and(lred). Case (betred) By the induction hypothesis, . Therefore by the rule(betred), 1 . The rest of this case is by lemma 9. Case (lred) By the induction hypothesis, . Therefore by the rule(lred), can assume x 62 dom() by the hygiene condition. Then, Lemma PROOF. We proceed by induction on the derivation of M ! !M 0 . We only show the crucial case (lred). Case (lred) By the induction hypothesis, . 2 =X] by the rule (lred). Since fx; x 0 g\BV Therefore ((fx 0 =xgM 0 for any terms M;M 0 such that n . PROOF. We proceed by induction on the derivation of M ! !M 0 . Here we only show the cases (hole) and (lred). . By repeated application of lemma 18. Case (lred) (Y 6 X). Suppose . By the induction hypothesis, M 1 [M Therefore by the rule(lred), The rest of this sub-case is by (M 0 . By the induction hypothesis, Therefore by the rule(lred), Therefore by lemma 12, This completes the proof of theorem 15. Conclusions We have developed a typed calculus for contexts. In this calculus, contexts and lambda terms share the same set of variables and can be freely mixed (as far as they type-check). This allows us to treat contexts truly as rst-class values. However, a straightforward mixture of -reduction and ll-reduction results in an inconsistent system. We have solved the problem by developing a type system that precisely species the variable-capturing nature of contexts. The resulting typed calculus enjoys the subject reduction property and Church-Rosser property. We believe that the typed context calculus presented here will serve as a type theoretical basis for developing a programming language with advanced features for manipulation of open terms. There are a number of interesting topics that merit further investigation. We brie y discuss some of them below. Integration with Explicit Substitution. In our calculus, -contraction is restricted to those redexes that do not contain free holes. While this does not restrict rst-class treatment of contexts, removing this restriction will make the reduction system slightly more general. As we have noted earlier, one reason for this restriction is that if we contract a -redex containing a free hole, then the binding through the hole will be lost. One way of solving this problem would be to integrate our calculus with -calculus of Abadi et al. [1], and to generalize variable renamers to explicit substitutions. Dowek et al. [3] considered a calculus containing holes and grafting, which roughly corresponds to hole-lling, and developed a technique to mingle capture-avoiding substitution with grafting by encoding them in a calculus of explicit substitution using de Bruijn notation. Although their calculus does not contain a term constructor for context application and therefore their technique is not directly applicable to our calculus, we believe that it is possible to extend their technique for our calculus by translating all the machinery we have developed for our calculus into de Bruijn notation. However, such translation would signicantly decrease the exibility of access to exported variables by names. It should also be noted that the notion of de Bruijn indexes presupposes -equivalence on terms, and therefore dening the context calculus using de Bruijn notation requires the mechanisms (or something similar to those) for obtaining -equivalence we have developed in this paper. Programming Languages with Contexts. Our motivation is to provide a basis for developing a programming language with the feature of rst-class contexts. The context calculus we have worked out in this article guarantees that we can have such a typed language with rst-class contexts. In order to develop an actual programming language, however, we need to develop a realistic evaluation strategy for the calculus. Our preliminary investigation shows that the usual call-by-value evaluation strategy using closures can be extended to our calculus. A more challenging topic is to develop a polymorphic type system and a type inference algorithm for our calculus, which will enable us to develop an ML-style programming language with the feature of contexts we have advocated. One crucial issue is the exible treatment of context types. In the current denition, the constructor fx1=x 0;:::;x n=x 0 n g is annotated with a variable renamer. This reduces the exibility of the calculus. A better approach would be to rene the type system so that if 0 then a context of type can be used whenever a context of type [ . One of the authors has recently developed an ML-style language with rst- class contexts [5] where an ML-style polymorphic type system, a call-by-value operational semantics and a type inference algorithm are given. Relationship with formula-as-type notion. It is intuitively clear that a context represented as a term in our calculus has constructive meaning. An important question is to characterize this intuition formally in the sense of Curry-Howard isomorphism [7]. This would lead us to a new form of proof normalization process corresponding to our ll-reduction. Since the context calculus is Church- Rosser, it should be possible to develop a proof system that is conservative over the conventional intuitionistic logic and supports a proof normalization process corresponding to ll-reduction. The authors recently noticed that there is an intriguing similarity between the proof system of typings in the context calculus and Joshi and Kulick's partial proof manipulation system [8] which is used to represent linguistic information. Another relevant system is Herbelin's lambda calculus isomorphic to a variant of sequent calculus, where proofs of certain sequents are interpreted by applicative contexts [6]. These results suggest some interesting connections between context calculus and proof systems. Acknowledgements The authors thank Pierre-Louis Curien, Laurent Dami, Yasuhiko Minamide, Didier Remy, Masahiko Sato and anonymous referees for their careful reading of a draft of this paper and numerous useful comments. The second author also thanks Shinn-Der Lee, and Dan Friedman for insightful discussions on contexts. --R Explicit substitutions. A lambda-calculus for dynamic binding The feel of Java. A lambda-calculus structure isomorphic to sequent calculus structure The formulae-as-types notion of construction Partial proof trees as building blocks for a categorized grammars. Context rewriting. Enriching the Lambda Calculus with Contexts: Towards A Theory of Incremental Program Construction. Fully abstract models of typed LCF considered as a programming language. Explicit Environments. Parallel reductions in A theory of binding structures and applications to rewriting. Con uent Equational Reasoning for Linking with First-Class Primitive Modules Programming in Modula-2 --TR Programming in MODULA-2 (3rd corrected ed.) A theory of binding structures and applications to rewriting Parallel reductions in MYAMPERSANDlgr;-calculus Enriching the lambda calculus with contexts A lambda-calculus for dynamic binding The Definition of Standard ML The Feel of Java Explicit Environments First-Class Contexts in ML A Lambda-Calculus Structure Isomorphic to Gentzen-Style Sequent Calculus Structure Context Rewriting Higher-order Unification via Explicit Substitutions --CTR Brigitte Pientka, Functional Programming With Higher-order Abstract Syntax and Explicit Substitutions, Electronic Notes in Theoretical Computer Science (ENTCS), v.174 n.7, p.41-60, June, 2007 Roger Keays , Andry Rakotonirainy, Context-oriented programming, Proceedings of the 3rd ACM international workshop on Data engineering for wireless and mobile access, September 19-19, 2003, San Diego, CA, USA Yosihiro Yuse , Atsushi Igarashi, A modal type system for multi-level generating extensions with persistent code, Proceedings of the 8th ACM SIGPLAN symposium on Principles and practice of declarative programming, July 10-12, 2006, Venice, Italy Christian Urban , Andrew M. Pitts , Murdoch J. Gabbay, Nominal unification, Theoretical Computer Science, v.323 n.1-3, p.473-497, 14 September 2004 Murdoch J. Gabbay, A new calculus of contexts, Proceedings of the 7th ACM SIGPLAN international conference on Principles and practice of declarative programming, p.94-105, July 11-13, 2005, Lisbon, Portugal Makoto Hamana, Term rewriting with variable binding: an initial algebra approach, Proceedings of the 5th ACM SIGPLAN international conference on Principles and practice of declaritive programming, p.148-159, August 27-29, 2003, Uppsala, Sweden Steven E. Ganz , Amr Sabry , Walid Taha, Macros as multi-stage computations: type-safe, generative, binding macros in MacroML, ACM SIGPLAN Notices, v.36 n.10, October 2001 Gavin Bierman , Michael Hicks , Peter Sewell , Gareth Stoyle , Keith Wansbrough, Dynamic rebinding for marshalling and update, with destruct-time ?, ACM SIGPLAN Notices, v.38 n.9, p.99-110, September Makoto Hamana, An initial algebra approach to term rewriting systems with variable binders, Higher-Order and Symbolic Computation, v.19 n.2-3, p.231-262, September 2006

alpha-renaming;lambda-calculus;type system;context

504573

Tractable disjunctions of linear constraints.

We study the problems of deciding consistency and performing variable elimination for disjunctions of linear inequalities and disequations with at most one inequality per disjunction. This new class of constraints extends the class of generalized linear constraints originally studied by Lassez and McAloon. We show that deciding consistency of a set of constraints in this class can be done in polynomial time. We also present a variable elimination algorithm which is similar to Fourier's algorithm for linear inequalities. Finally, we use these results to provide new temporal reasoning algorithms for the Ord-Horn subclass of Allen's interval formalism. We also show that there is no low level of local consistency that can guarantee global consistency for the Ord-Horn subclass. This property distinguishes the Ord-Horn subclass from the pointizable subclass (for which strong 5-consistency is sufficient to guarantee global consistency), and the continuous endpoint subclass (for which strong 3-consistency is sufficient to guarantee global consistency). Copyright 2001. Elsevier Science B.V.

Introduction Linear constraints over the reals have recently been studied in depth by researchers in constraint logic programming (CLP) and constraint databases (CDB) [JM94, KKR95, Kou94c]. Two very important operations in CLP and CDB systems are deciding consistency of a set of constraints, and performing variable elimination. Subclasses of linear constraints over the reals have also been considered in temporal reasoning [DMP91, Kou92, Kou94a, Kou95, NB95]. Important operations in temporal reasoning applications are the deciding consistency of a set of binary temporal constraints, (ii) performing variable elimination, and (iii) computing the strongest feasible constraints between every pair of variables. Disjunctions of linear constraints over the reals are important in many applications [JM94, DMP91, Kou92, Kou94a, Kou94b, Kou95, NB95]. The problem of deciding consistency for an arbitrary set of disjunctions of linear constraints is NP-complete [Son85]. It is therefore interesting to discover classes of disjunctions of linear constraints for which consistency can be decided in PTIME. In [LM89a], Lassez and McAloon have studied the class of generalized linear constraints which includes linear inequalities (e.g., 2x 1 This is a longer version of a paper with the same title which appears in the Proceedings of the 2nd International Conference on Principles and Practice of Constraint Programming (CP96), Cambridge, MA, August 19-22, 1996. Lecture Notes in Computer Science, Vol. 1118, pages 297-307. and disjunctions of linear inequations 1 (e.g., Among other things, they have shown that the consistency problem for this class can be solved in PTIME. [Kou92, IvH93, Imb93, Imb94] have studied the problem of variable elimination for generalized linear constraints. The basic algorithm for variable elimination has been discovered independently in [Kou92] and [Imb93], but [Kou92] has used the result only in the context of temporal constraints. The basic algorithm is essentially an extension of Fourier's elimination algorithm [Sch86] to deal with disjunctions of inequations. If S is a set of constraints, let jSj denote its cardinality. Let be a set of generalized linear constraints, where I is a set of inequalities and D n is a set of disjunctions of inequations. If we eliminate m variables from C using the basic algorithm proposed by Koubarakis and Imbert then the resulting set contains O(jI inequalities and O(jD n j jI j 2 m+1 disjunctions of inequations. A lot of these constraints are redundant. Imbert has studied this problem in more detail and presented more advanced algorithms that eliminate redundant constraints [Imb93, Imb94]. In this paper we go beyond the above work on generalized linear constraints. Our contributions can be summarized as follows: ffl We extend the class of generalized linear constraints to include disjunctions with an unlimited number of inequations and at most one inequality per disjunction. For example: The resulting class will be called the class of Horn constraints since there seems to be some analogy with Horn clauses. We show that deciding consistency can still be done in PTIME for this class (Theorem 3.4). This result has also been obtained independently by Jonsson and B-ackstr-om [Jon96]. We also extend the basic variable elimination algorithm of [Kou92, Imb93] for this new class of constraints. ffl We study a special class of Horn constraints, called Ord-Horn constraints, originally introduced in [NB95]. This class is important for temporal reasoning based on the Ord-Horn class of interval relations expressible in Allen's formalism [All83, NB95]. Our results allow us to improve the best known algorithms for consistency checking and computing the strongest feasible constraints for this class. This answers an open problem of [NB95]. The paper is organized as follows. Section 2 introduces the basic concepts needed for the developments of this paper. Section 3 presents the algorithm for deciding consistency. Section 4 presents the algorithm for variable elimination. Section 5 presents our results for the class of Ord-Horn constraints. Finally, Section 6 discusses future research. Preliminaries We consider the n-dimensional Euclidean space R n . A linear constraint over R n is an expression a are integers, x are variables ranging over the real numbers, and - is 6=. Depending on what - is, we will 1 Some people prefer the term disequations [Imb94]. distinguish linear constraints into inequalities (e.g. equations (e.g., inequations (e.g., Let us now define the class of constraints that we will consider. Definition 2.1 A Horn constraint is a disjunction d is a weak linear inequality or a linear inequation, and the number of inequalities among does not exceed one. If there are no inequalities then a Horn constraint will be called negative. Otherwise it will be called positive. Horn constraints of the form will be called disjunctive. Example 2.1 The following are examples of Horn constraints: The first and the third constraint are positive while the second and the fourth are negative. The third and fourth constraint are disjunctive. According to the above definition weak inequalities are positive Horn constraints. Sometimes we will find it more convenient to consider inequalities separately from positive disjunctive Horn constraints. If d is a positive disjunctive Horn constraint then where E is a conjunction of equations and i is an inequality. We will often use this notation for positive Horn constraints. Notice that we do not need to introduce strict inequalities in the above definition. A strict inequality like x 1 can be equivalently written as follows: Similarly, the constraint x 1 a disjunction of inequations can be equivalently written as the conjunction of the following constraints: A similar observation is made in [NB95]. Negative Horn constraints have been considered before in [LM89a, LM89b, Kou92, IvH93, Imb93, Imb94, Kou95]. Nebel and B-urckert have studied the class of Ord-Horn constraints in the context of qualitative interval reasoning [NB95]. Ord-Horn constraints form a proper subclass of Horn constraints, and will be considered in detail in Section 5. Horn constraints have also been studied by Jonsson and B-ackstr-om [Jon96] who discovered independently the result discussed in Section 3. We will now present some standard definitions. Definition 2.2 Let C be a set of constraints in variables x . The solution set of C, denoted by Sol(C), is: Each member of Sol(C) is called a solution of C. A set of constraints is called consistent if its solution set is non-empty. We will use the same notation, Sol(\Delta), for the solution set of a single constraint. Remark 2.1 In the rest of the paper we will usually consider one or more sets of constraints e.g., . In this case we will always regard a subset of R n even though C i might contain less than n variables. For example, if we happen to deal with we may write Similarly, we may write We will also use the alternative notation Sol (\Delta). If C is a set of constraints, Sol (C) will always be regarded a subset of R k where k is the number of variables of C (indepen- dently of any other constraint set considered at the same time). This notation will come handy in Section 4 where we study variable elimination. Definition 2.3 Let C 1 and C 2 be sets of constraints in the same set of variables. C 1 will be called equivalent to C 2 if logically follows from a set of constraints C, denoted by C every solution of C satisfies c. We will now present some concepts of convex geometry [Sch86, Gru67] that will enable us to study the geometric aspects of the constraints considered. We will take our definitions from [LM89a]. If V is a subspace of the n-dimensional Euclidean space R n and p a vector in R n then the translation called an affine space. The intersection of all affine spaces that contain a set S is an affine space, called the affine closure of S and denoted by Aff(S). If e is a linear equation then the solutions set of e is called a hyperplane. In R 3 the hyperplanes are the planes. In R 2 the hyperplanes are the straight lines. A hyperplane is an affine space and every affine space is the intersection of a finite number of hyperplanes. If E is a set of equalities then Sol(E) is an affine space. If i is a linear inequality then the solution set of i is called a half-space. If I is a set of inequalities then Sol(I) is the intersection of a finite number of half-spaces, and is called a polyhedral set. A set S ' R n is called convex if the line segment joining any pair of points in S is included in S. Affine subspaces of R n are convex. Half-spaces are convex. Also, polyhedral sets are convex. If d is a negative Horn constraint then the solution set of d is The constraint :d is a conjunction of equations thus Sol(:d) is an affine space. If :d is inconsistent then d is equivalent to true (e.g., x In the rest of the paper we will ignore negative Horn constraints that are equivalent to true. If d is a positive disjunctive Horn constraint of the form :(E - i) then R n n Sol(:d). The constraint :d is a conjunction E - i where E is a conjunction of equations and i is a strict inequality. If E j true then d is essentially a weak inequality inconsistent then its corresponding Horn constraint d is equivalent to true (e.g., x and Sol(i) ' Sol(E) then d j :E, so d is actually a negative Horn constraint (e.g., consistent and Sol(i) 6' Sol(E) then its solution set will be called a half affine space. In R 3 the half affine spaces are half-lines or half-planes. For example, plane. In the rest of the paper we will ignore positive disjunctive Horn constraints equivalent to a weak inequality, a negative Horn constraint or true. 3 Deciding Consistency [LM89a] showed that negative Horn constraints can be treated independently of one another for the purposes of deciding consistency. The following is one of their basic results. Theorem 3.1 Let be a set of constraints where I is a set of linear inequalities and D n is a set of negative Horn constraints. Then C is consistent if and only if I is consistent, and for each d 2 D n the set I [ fdg is consistent. Whether a set of inequalities is consistent or not, can be decided in PTIME using Kachian's linear programming algorithm [Sch86]. We can also detect in PTIME whether I [ fdg is consistent by simply running Kachian's algorithm 2n times to decide whether I implies every equality e in the conjunction of n equalities :d. In other words, deciding consistency in the presence of negative Horn contraints can be done in PTIME. 2 Is it possible to extend this result to the case of positive disjunctive Horn constraints? In what follows, we will answer this question affirmatively. Let us start by pointing out that the independence property of negative Horn constraints does not carry over to positive ones. Example 3.1 The constraint sets I are consistent. But the set I [ inconsistent. Fortunately, there is still enough structure available in our problem which we can exploit to come up with a PTIME consistecy checking algorithm. Let a set of constraints where I is a set of inequalities, D p is a set of positive disjunctive Horn constraints, and D n is a set of negative Horn constraints. Intuitively, the solution set of C is empty only if the polyhedral set defined by I is covered by the affine spaces and half affine spaces defined by the Horn constraints. The algorithm Consistency shown in Figure 1 proceeds as follows. Initially, we check whether I is consistent. If this is the case, then we proceed to examine whether Sol(I) can be covered by Sol(f:d To verify this, we make successive passes over In each pass, we carry out two checks. The first check discovers whether there is any positive Horn constraint d j :(E - i) such that Sol(I) is included in the affine space defined by E. If this is the case then d is discarded and I is updated to reflect the part possibly "cut off " by d. The resulting solution set Sol(I) is still a polyhedral set. An inconsistency can arise if Sol(I) is reduced to ; by successive "cuts". In each pass we also check whether there is an affine space (represented by the negation of a negative Horn constraint) which covers Sol(I). In this case there is an inconsistency as well. The algorithm stops when there are no more affine spaces or half affine spaces that pass the two checks. In this case C is consistent. Let us now prove the correctness of algorithm Consistency. First, we will need a few technical lemmas. The first two lemmas show that the sets resulting from successive "cuts" inflicted on Sol(I) by positive Horn constraints passing the first check of the algorithm are indeed polyhedral. The lemmas also give a way to compute the inequalities defining these sets. 2 The exact algorithm that Lassez and McAloon give in [LM89a] is different but this is not significant for the purposes of this paper. Algorithm Consistency Input: A set of constraints Output: "consistent" if C is consistent. Otherwise "inconsistent". If I is inconsistent then return "inconsistent" Repeat Done / true For each d 2 D p [ D n do I / I - :i If I is inconsistent then return "inconsistent" Done / false Remove d from D p Else If d 2 D n and Sol(I) ' Sol(:d) then Return "inconsistent" End If End For Until Done Return "consistent" Figure 1: Deciding consistency of a set of Horn constraints Lemma 3.1 Let I be a set of inequalities and :(E - i) be a positive disjunctive Horn constraint such that Sol(I) ' Sol(E). Then Sol(I - :(E - The other direction of the proof is trivial. Lemma 3.2 Let I be a set of inequalities and d k j be a set of positive disjunctive Horn constraints such that Sol(I) ' l Then Proof: The proof is by induction on m. The base case 3.1. For the inductive step, let us assume that the lemma holds for Then using the inductive hypothesis. The assumptions of this lemma and Lemma 3.1 imply that Thus The following lemmas show that if there are Horn constraints that do not pass the two checks of algorithm Consistency then the affine spaces or half affine spaces corresponding to their negations cannot cover the polyhedral set defined by the inequalities. Lemma 3.3 Let S be a convex set of dimension d and suppose that S are convex sets of dimension d Proof: See Lemma 2 of [LM89a]. Lemma 3.4 Let I be a consistent set of inequalities and d k j be a set of Horn constraints such that Sol(I) 6' Proof: The proof is very similar to the proof of Theorem 1 of [LM89a]. This means that Aff(Sol(I)) is an affine space of strictly lower dimension than Aff(Sol(I)). Then is of strictly lower dimension than Sol(I) since the dimension of Sol(I) is equal to that of Aff(Sol(I)). Thus from Lemma 3.3, Sol(I) 6' S m now that We can now conclude that Sol(I) 6' The following theorems demonstrate that the algorithm Consistency is correct and can be implemented in PTIME. Theorem 3.2 If algorithm Consistency returns "inconsistent" then its input C is inconsistent Proof: If the algorithm returns "inconsistent" in its first line then I , and therefore C, is inconsistent. If the algorithm returns "inconsistent" in the third if-statement then there are positive Horn constraints d k j such that the assumptions of Lemma 3:2 hold for I and d Therefore Consequently, If the algorithm returns "inconsistent" in the fourth if-statement then there are positive Horn constraints d negative constraint dm+1 2 D n such that the assumptions of Lemma 2 hold for I and d But then Theorem 3.3 If algorithm Consistency returns "inconsistent" then its input C is inconsistent Proof: If the algorithm returns "consistent" then I is consistent. Let d the positive Horn constraints removed from D p [ D n by the algorithm, and be the remaining Horn constraints. Then Notice that Sol(I - otherwise the algorithm outputs "inconsistent" in Step 2. Also, Sol(I - otherwise the algorithm would have removed d k from D p [ D n . any loss of generality we can also assume that for all this does not hold for constraint d k , this constraint can be discarded without changing Sol(C)). From Lemma 3.4 we can now conclude that Theorem 3.4 The algorithm Consistency can be implemented in PTIME. Proof: It is not difficult to see that the algorithm can be implemented in PTIME. The consistency of I can be checked in PTIME using Kachian's algorithm for linear programming [Sch86]. The test Sol(I) ' Sol(E) can be verified by checking whether every equation e in the conjunction E is implied by I . This can be done in PTIME using Kachian's algorithm 2n times where n is the number of equations in E. In a similar way one can implement the test Sol(I) ' Sol(:d) in PTIME when d is a negative Horn constraint. We have just showed that the consistency of a set of Horn constraints can be determined in PTIME. This is an important result with potential applications in any CLP or CDB system dealing with linear constraints [JM94, KKR95, Kou94c]. We will now turn our attention to the problem of eliminating one or more variables from a given set of Horn constraints. Algorithm VarElimination Input: A set of Horn constraints C in variables X , and a variable to be eliminated from C. Output: A set of Horn constraints C 0 in variables X n fxg such that Projection Xnfxg (Sol (C)). Rewrite each constraint containing x as x - U - OE or L - x - OE or x 6= A - OE where OE is a disjunction of inequations and x does not appear in OE. For each pair of positive Horn constraints x - U - OE 1 and L - x - OE 2 do End For For each pair of positive Horn constraints x - U - OE 1 and L - x - OE 2 do For each negative Horn constraint x 6= A - OE do Add A 6= L - A 6= U - OE - OE 1 - OE 2 to C 0 End For End For Add each constraint not containing x to C 0 Return C 0 Figure 2: A variable elimination algorithm 4 Variable Elimination In this section we study the problem of variable elimination for sets of Horn constraints. The algorithm VarElimination, shown in Figure 2, eliminates a given variable x from a set of Horn constraints C. More variables can be eliminated by successive applications of VarElimination. This algorithm does not consider inequalities separately from positive disjunctive Horn constraints (as algorithm Consistency did in Section 3). The algorithm VarElimination is similar to the one studied in [Kou92, Imb93] for the case of negative Horn constraints. Theorem 4.1 The algorithm VarElimination is correct. Proof: Let the variables of C be g. If is an element of Sol (C) then it can be easily seen that it is also an element of Sol (C 0 ). Conversely, take consider the set C(x; x 0 If this set is simplified by removing constraints equivalent to true, disjunctions equivalent to false, and redundant constraints then Let us now assume (by contradiction) that there is no value x 0 2 R n such that (C). This can happen only under the following cases: 1. come from positive Horn constraints otherwise these constraints would have been discarded from C(x; x 0 during its simplification. But because l - then l 0 - u 0 . Contradiction! 2. l reasoning similar to the above, we can show that this case is also impossible. Finally, we can conclude that there exists a value x 0 2 R such that be a set of Horn constraints. Eliminating m variables from C with repeated applications of the above algorithm will result in a set with O((jI positive Horn constraints and O(jD n j (jI negative Horn constraints. Many of these contraints will be redundant; it is therefore important to extend this work with efficient redundancy elimination algorithms that can be used together with VarElimina- tion. This section concludes our study of the basic reasoning problems concerning Horn constraints. We will now turn our attention to a suclass of Horn constraints with important applications to temporal reasoning. 5 Reasoning with Ord-Horn Constraints This section studies a special class of Horn constraints, called Ord-Horn constraints, originally introduced in [NB95]. This class is important in interval based temporal reasoning [All83] as we will immediately show below. Definition 5.1 An Ord-Horn constraint is a Horn constraint d each is an inequality x - y or an inequation x 6= y and x and y are variables ranging over the real numbers. Example 5.1 The following are examples of Ord-Horn constraints: The first and the last constraint are positive while the second and the third are negative. In [All83], Allen introduced a calculus for reasoning about intervals in time. An interval is an element of the following set I: If i is an interval variable, will denote the endpoints of i. Allen's calculus is based on thirteen mutually exclusive binary relations which can capture all the possible ways two time intervals can be related. These basic relations are and their inverses (equals is its own inverse). Figure 3 gives a pictorial representation of these relations. For reasons of brevity, we will use the symbols b; m; o; d; s; f and e to refer Basic Interval Symbol Pictorial Endpoint Relation Representation Constraints during j d iiiiiiiii includes i di Figure 3: The thirteen basic relations of Allen to the basic relations in Allen's formalism. The inverse of each relation will be denoted by the name of the relation with the suffix i (for example, the inverse of b will be denoted by bi). Allen's calculus has received a lot of attention and has been the formalism of choice for representing qualitative interval information. Whenever the interval information to be represented is indefinite, a disjunction of some of the thirteen basic relations can be used to represent what is known. There are 2 13 such disjunctions representing qualitative relations between two intervals. Each one of these relations will be denoted by the set of its constituent basic relations e.g., fb; bi; d; mg. The empty relation will be denoted by ;, and the universal relation will be denoted by ?. The set of all 2 13 relations expressible in Allen's formalism will be denoted by A [NB95]. The following definition will be useful below. Definition 5.2 Let S be a subset of A, i and j be variables representing intervals, and An S-constraint is any expression of the form i R j. Example 5.2 If interval i denotes the time that John reads his morning newspaper and denotes the time that he has breakfast, and we know that John never reads a newspaper while he is eating, then the A-constraint characterizes i and j according to the information given. Definition 5.3 Let C be a set of S-constraints. The solution set of C is: Unfortunately, all interesting reasoning problems associated with Allen's interval calculus are NP-hard [VKvB89] therefore it is interesting to consider subsets of Allen's for- malism, in the form of subsets of A, that have better computational properties. 3 Three such subsets of A have received more attention: ffl The set C which consists of all relations R 2 A which satisfy the following condition. If i and j are intervals, i R j can be equivalently expressed as a conjunction of inequalities are endpoints of i and j. The set C is called the continuous endpoint subclass of A [VKvB89]. ffl The set P which consists of all interval relations R 2 A which satisfy the following condition. If i and j are intervals, i R j can be equivalently expressed as a conjunction of inequalities are endpoints of i and j. The set C is called the pointisable subclass of A [VKvB89, vBC90, vB92]. Because ae P the pointisable subclass is more expressive than the continuous endpoint subclass. ffl The set H which consists of all interval relations R 2 A which satisfy the following condition. If i and j are intervals, i R j can be equivalently expressed as a conjunction of Ord-Horn constraints on the endpoints of i and j. The disjunctive Ord-Horn constraints involved in this equivalence are not arbitrary. There are at most three of them, and each one consists of two disjuncts of the form The set H was introduced by Nebel and B-urckert and named the Ord-Horn sub-class [NB95]. Because P ae H the Ord-Horn subclass is more expressive than the pointisable subclass. It consists of 868 relations i.e., it covers more than 10% of A. Example 5.3 The following are P-constraints: Their equivalent endpoint constraints are: second P-constraint is also a C-constraint while the first one is not. For an enumeration of C and P , see [vBC90]. Example 5.4 The A-constraint i fb; big j is not an H-constraint. The constraint is an H-constraint which is not a P-constraint. Its equivalent endpoint constraints are: enumeration of H together with several related C programs has been provided by Nebel and Burckert. See [NB95] for details. 3 At the expense of being less The following reasoning problems have been studied for the above subclasses [VKvB89, ffl Given a set C of S-constraints, decide whether C is consistent. ffl Given a set C of S-constraints, determine the strongest feasible relation between each pair of interval variables i and j. The strongest feasible relation between two interval variables i and j is the smallest set R such that C j. This is the same as computing the minimal network corresponding to the given set of constraints. 4 In this section we will show how our results can be used to improve the best known algorithms for the above reasoning problems in the case of the Ord-Horn subclass. We start with two theorems from [NB95]. Theorem 5.1 Let C be a set of H-constraints. Deciding whether C is consistent can be done in O(n 3 is the number of variables in C. Theorem 5.2 Let C be a set of H-constraints. Computing the feasible relations between all pairs of variables can be done in O(n 5 is the number of variables in C. We will now use the results of Section 3 to improve the complexity bounds of the above theorems. Theorem 5.3 Let C be a set of H-constraints. Let n be the number of variables in C, and h be the number of constraints (i R such that R 2 H n C. Deciding whether C is consistent can be done in O(max(n 2 ; hn)) time. Proof: First we translate C into a set of Ord-Horn constraints C 0 . Since this translation can be achieved in O(n 2 ) time. Let C I is a set of inequalities, D p is a set of positive disjunctive Horn constraints and D n a set of negative Horn constraints. The translation of Nebel and B-urckert shows that C 0 contains point variables and jD p [ We will use algorithm Consistency from Figure 1 to decide C 0 . In this case Consistency can be made to work in O(max(n 1 Checking the consistency of I can be done in O(n 1 constructing a directed graph G corresponding to I and examining its strongly connected components [vB92]. Now notice that the statement If-Else-End-If in algorithm Consistency is executed O(jD p [ D n times. Each execution of this statement takes O(n 1 ) time. Let us see why. If the examined constraint d is in D p , the test Sol(I) ' Sol(E) amounts to checking whether the single inequality E is implied by I . This can be done in O(n 1 examining the strongly connected components of G. Similarly, if d is in D n , the test Sol(I) ' Sol(:d) can be done in O(n 1 ) time. Therefore deciding whether C 0 is consistent can be done in deciding whether C is consistent can be done in Theorem 5.4 Let C be a set of H-constraints. Let n be the number of variables in C, and h be the number of constraints (i R such that R 2 H n C. Computing the feasible relations between all pairs of variables can be done in O(max(n 4 ; hn 3 4 We will not define minimal networks formally here. The interested reader can consult [vB92] (or many other temporal reasoning papers). Proof: As in [NB95], we will consider all O(n 2 ) pairs of variables in turn. For each pair we check whether each of the thirteen basic relations is consistent with the given constraints. The basic relations that satisfy this criterion form the strongest feasible relation between the pair. Using the algorithm of Theorem 5.3, each check can be done in O(max(n 2 ; hn)) time. The bound of the theorem follows immediately. In the worst case the parameter h, as specified in the above theorems, can be O(n 2 ). However in practical applications, we expect h to be significantly less than O(n 2 ) thus the above theorems are real improvements over [NB95]. 6 Future Research In future research we would like to study more advanced variable elimination algorithms for Horn constraints. The results of [Imb93, Imb94] that apply to negative Horn constraints only, should be a good starting point in this direction. Another interesting problem, which occupies us currently, is to extend the results of [Kou95] to the pointizable subclass P and the Ord-Horn subclass H of A. [Kou95] studied the problem of enforcing global consistency for sets of quantitative temporal constraints over the real numbers. In a globally consistent constraint set all interesting constraints are explicitly represented and the projection of the solution set on any subset of the variables can be computed by simply collecting the constraints involving these variables. An important consequence of this property is that a solution can be found by backtrack- Enforcing global consistency can take an exponential amount of time in the worst case [Fre78, Coo90]. As a result it is very important to identify cases in which local consistency, which presumably can be enforced in polynomial time, implies global consistency [Dec92]. The class of temporal constraints considered in [Kou95] includes equalities of the form inequalities of the form x \Gamma y - r and inequations of the form x \Gamma y 6= r where x; y are variables ranging over the real numbers and r is a real constant. [Kou95] shows that strong 5-consistency is necessary and sufficient for achieving global consistency for this class of constraints. It also gives an algorithm which achieves global consistency in is the number of variables and H is the number of inequations. The details of this algorithm demonstrate that there are situations where it is impossible to enforce global consistency without introducing disjunctions of inequations e.g., The results of [Kou95] can provide the basis for efficient global consistency algorithms for the pointizable subclass P . The open question is whether one can use the results of this paper and [Kou95] to find efficient algorithms for global consistency for the ORD-Horn subclass. --R Maintaining Knowledge about Temporal Intervals. An optimal k-consistency algorithm From local to global consistency. Temporal Constraint Networks. Synthesizing Constraint Expressions. A Sufficient Condition For Backtrack-Free Search Convex Polytopes. On the Handling of Disequations in CLP over Linear Rational Arithmetic. Constraint Logic Programming: A Survey. Constraint Query Languages. Dense Time and Temporal Constraints with 6 Complexity Results for First-Order Theories of Temporal Con- straints Database Models for Infinite and Indefinite Temporal Infor- mation Foundations of Indefinite Constraint Databases. From Local to Global Consistency in Temporal Constraint Networks. A Canonical Form for Generalized Linear Constraints. A Canonical Form for Generalized Linear Costraints. On binary constraint problems. Bernhard Nebel and Hans-J-urgen B-urckert Theory of Integer and Linear Programming. Real Addition and the Polynomial Time Hierarchy. Reasoning About Qualitative Temporal Information. Exact and Approximate Reasoning about Temporal Relations. Constraint Propagation Algorithms for Temporal Reasoning: A Revised Report. --TR Theory of linear and integer programming An optimal <italic>k</>-consistency algorithm Constraint propagation algorithms for temporal reasoning: a revised report Exact and approximate reasoning about temporal relations Temporal constraint networks From local to global consistency A canonical form for generalized linear constraints Reasoning about qualitative temporal information Variable elimination for generalized linear constraints On the handling of disequations in CLP over linear rational arithmetic On binary constraint problems Database models for infinite and indefinite temporal information Redundancy, variable elimination and linear disequations Reasoning about temporal relations Constraint query languages From local to global consistency in temporal constraint networks The complexity of query evaluation in indefinite temporal constraint databases A unifying approach to temporal constraint reasoning A Sufficient Condition for Backtrack-Free Search Maintaining knowledge about temporal intervals Synthesizing constraint expressions Foundations of Indefinite Constraint Databases From Local to Global Consistency in Temporal Constraint Networks --CTR Mathias Broxvall, A method for metric temporal reasoning, Eighteenth national conference on Artificial intelligence, p.513-518, July 28-August 01, 2002, Edmonton, Alberta, Canada Peter Jonsson , Andrei Krokhin, Complexity classification in qualitative temporal constraint reasoning, Artificial Intelligence, v.160 n.1, p.35-51, December 2004 Mathias Broxvall , Peter Jonsson, Point algebras for temporal reasoning: algorithms and complexity, Artificial Intelligence, v.149 n.2, p.179-220, October Andrei Krokhin , Peter Jeavons , Peter Jonsson, Reasoning about temporal relations: The tractable subalgebras of Allen's interval algebra, Journal of the ACM (JACM), v.50 n.5, p.591-640, September Manolis Koubarakis, Querying Temporal Constraint Networks: A Unifying Approach, Applied Intelligence, v.17 n.3, p.297-311, November-December 2002

global consistency;ORD-horn constraints;variable elimination;linear constraints;interval algebra

504577

Loop checks for logic programs with functions.

Two complete loop checking mechanisms have been presented in the literature for logic programs with functions: OS-check and EVA-check. OS-check is computationally efficient but quite unreliable in that it often misidentifies infinite loops, whereas EVA-check is reliable for a majority of cases but quite expensive. In this paper, we develop a series of new complete loop checking mechanisms, called VAF-checks. The key technique we introduce is the notion of expanded variants, which captures a key structural characteristic of in finite loops. We show that our approach is superior to both OS-check and EVA-check in that it is as efficient as OS-check and as reliable as EVA-check. Copyright 2001 Elsevier Science B.V.

Introduction The recursive nature of logic programs leads to possibilities of running into innite loops with top-down query evaluation. By an innite loop we refer to any innite SLD-derivation. An illustrative example is the evaluation of the goal p(a) against the logic program which leads to the innite loop Another very representative logic program is Currently on leave at Department of Computing Science, University of Alberta, Edmonton, Alberta, Canada T6G 2H1. Email: ydshen@cs.ualberta.ca. Fax: (780) 492-1071. against which evaluating the query p(g(a)) generates the innite loop Loop checking is a long recognized problem in logic programming. 1 Although many loop checking mechanisms have been proposed during the last decade (e.g. [1, 2, 6, 7, 9, 12, 14, 17, 19, 20, 22, 23]), a majority of them (e.g. [1, 2, 6, 7, 9, 12, 19, 20, 22, 23]) are suitable only for function-free logic programs because they determine innite loops by checking if there are variant goals/subgoals in SLD-derivations. Variant goals/subgoals are the same goals/subgoals up to variable renaming. Hence, an innite loop like L 2 can not be detected because no variant goals/subgoals occur in the derivation. An important fact is that for function-free logic programs, innite loops can be completely avoided by appealing to tabling techniques [4, 5, 18, 21, 23, 24]. However, innite loops with functions remain unresolved even in tabling systems [13]. To our best knowledge, among all existing loop checking mechanisms only two can deal with innite loops like L 2 . One is called OS-check (for OverSize loop check) [14] and the other EVA-check (for Extended Variant Atoms loop check) [17]. OS-check, rst introduced by Sahlin [14, 15] and further formalized by Bol [3], determines innite loops based on two parameters: a depth bound d and a size function size. Informally, OS- check says that an SLD-derivation may go into an innite loop if it generates an OverSized subgoal. A subgoal A is said to be OverSized if it has d ancestor subgoals in the SLD-derivation that have the same predicate symbol as A and whose size is smaller than or equal to A. For example, if we choose is an innite loop. It is proved that OS-check is complete in the sense that it cuts all innite loops. However, because it merely takes the number of repeated predicate symbols and the size of subgoals as its decision parameters, without referring to the informative internal structure of the subgoals, the underlying decision is fairly unreliable; i.e. many non-loop derivations may be pruned unless the depth bound d is set su-ciently large. EVA-check, proposed by Shen [17], determines innite loops based on a depth bound d and generalized variants. Informally, EVA-check says that an SLD-derivation may go into an innite loop if it generates a subgoal A 0 that is a generalized variant of all its d ancestor subgoals. A subgoal A 0 is said to be a generalized variant of a subgoal A if it is the same as A up to variable renaming except for some arguments whose size increases from A via a set of recursive clauses. Recursive clauses are of the form like C 21 in P 2 , one distinct property of which is that repeatedly applying them may lead to recursive increase in size of some subgoals. 1 There are two dierent topics on termination of logic programs. One is termination analysis (see [8] for a detailed survey), and the other is loop checking (see [1, 23]). In this paper, we study loop checking. Recursive increase in term size is a key feature of innite loops with functions. That is, any innite loops with innitely large subgoals are generated by repeatedly applying a set of recursive clauses. Due to this fact, EVA-check is complete and much more reliable than OS-check in the sense that it is less likely to mis-identify innite loops [17]. OS-check has the obvious advantage of simplicity, but it is unreliable. In contrast, EVA-check is reliable in a majority of cases, but it is computationally expensive. The main cost of EVA-check comes from the computation of recursive clauses. On the one hand, given a logic program we need to determine which clauses in it are recursive clauses. On the other hand, for any subgoals A and A 0 in an SLD-derivation, in order to determine if A 0 is a generalized variant of A, we need to check if A 0 is derived from A by applying some set of recursive clauses. Our observation shows that both processes are time-consuming. In this paper, we continue to explore complete loop checking mechanisms, which have proved quite useful as stopping criteria for partial deduction in logic programming [11] (see [3] for the relation between stopping criteria for partial deduction and loop checking). On the one hand, unlike OS-check, we will fully employ the structural characteristics of innite loops to design reliable loop checking mechanisms. On the other hand, instead of relying on the expensive recursive clauses, we extract structural information on innite loops directly from individual subgoals. We will introduce a new concept expanded variants, which captures a key structural characteristic of certain subgoals in an innite loop. Informally, a subgoal A 0 is an expanded variant of a subgoal A if it is a variant of A except for some terms (i.e. variables or constants or functions) in A each of which grows in A 0 into a function containing the term. The notion of expanded variants provides a very useful tool by which a series of complete loop checking mechanisms can be dened. In this paper, we develop four such VAF-checks (for Variant Atoms loop checks for logic programs with Functions) V AF 1 4 (d), where d is a depth bound. loops based on expanded variants. V AF 2 (d) enhances V AF 1 (d) by taking into account one (innitely) repeated clause. V AF 3 (d) enhances V AF 2 (d) with a constraint of a set of (innitely) repeated clauses. And V AF 4 (d) enhances V AF 3 (d) with a constraint of recursive clauses. The reliability increases from V AF 1 (d) to V AF 4 (d), but the computational overhead increases, too. By balancing between the two key factors, we choose V AF 2 (d) as the best for practical applications. V AF 2 (d) has the same complexity as OS-check, but is far more reliable than OS-check. When d 2, V AF 2 (d) is reliable for a vast majority of logic programs. Moreover, while no less reliable than EVA-check, V AF 2 (d) is much more e-cient than EVA-check (because like OS-check it does not compute recursive clauses). The plan of this paper is as follows. In Section 2, we review basic concepts concerning loop checking. In Section 3, we introduce expanded variants and examine their properties. In Section 4, we dene four VAF-checks and prove their completeness. In Section 5, we make a comparison of the VAF-checks with OS-check and EVA-check. Preliminaries In this section, we review some basic concepts concerning loop checking. We assume familiarity with the basic concepts of logic programming, as presented in [10]. Here and throughout, by a logic program we always mean a positive logic program. Variables begin with a capital letter, and predicate symbols, function symbols and constants with a lower case letter. Let A be an atom/function. The size of A, denoted jAj, is the count of function symbols, variables and constants in A. We use rel(A) to refer to the predicate/function symbol of A, and use A[i] to refer to the i-th argument of A, A[i][j] to refer to the j-th argument of the i-th argument, and A[i]:::[k] to refer to the k-th argument of . of the i-th argument. For example, let Denition 2.1 By a variant of an SLD-derivation (resp. a goal, subgoal, atom or function) D we mean a derivation (resp. a goal, subgoal, atom or function) D 0 that is the same as D up to variable renaming. Denition 2.2 ([1, 3]) Let P be a logic program, G 0 a top goal and S a computation rule. 1. Let L be a set of SLD-derivations of P [ fG 0 g under S. Dene that is a proper subderivation of Dg. L is subderivation free if 2. A (simple) loop check is a computable set L of nite SLD-derivations such that L is closed under variants and is subderivation free. Observe that a loop check L formally denes a certain type of innite loops generated from under S; i.e. an SLD-derivation G 0 is said to step into an innite loop at G k if G 0 ) is in L. Therefore, whenever such an innite loop is detected, we should cut it immediately below G k . This leads to the following denition. Denition 2.3 Let T be the SLD-tree of P [ fG 0 g under S and L a loop check. Let the SLD-derivation from the top goal G 0 to G 0 is in Lg. By applying L to T we obtain a new SLD-tree which consists of T with all the nodes (goals) in CUT pruned. By pruning a node from an SLD-tree we mean removing all its descendants. In order to justify a loop check, Bol et al. introduced the following criteria. Denition 2.4 ([1]) Let S be a computation rule. A loop check L is weakly sound if the following condition holds: for every logic program P , top goal G 0 and SLD-tree T of P [ fG 0 g under S, if T contains a successful branch, then contains a successful branch. A loop check L is complete if every innite SLD-derivation is pruned by L. (Put another way, a loop check L is complete if for any logic program P and top goal G 0 An ideal loop check would be both weakly sound and complete. Unfortunately, since logic programs have the full power of the recursive theory, there is no loop check that is both weakly sound and complete even for function-free logic programs [1]. As mentioned in the Introduction, in this paper we explore complete loop checking mechanisms. So in order to compare dierent complete loop checks, we introduce the following concept. Denition 2.5 A complete loop check L 1 is said to be more reliable 2 than a complete loop check logic program P and top goal G 0 , the successful SLD-derivations in TL1 are not less than those in TL2 , and not vice versa. It is proved that EVA-check is more reliable than OS-check [17]. In the Introduction, we mentioned a notion of ancestor subgoals. Denition 2.6 ([17]) For each subgoal A in an SLD-tree, its ancestor list ALA is dened recursively as follows: 1. If A is at the root, then fg. 2. Let Am be a node in the SLD-tree, with A 1 being selected to resolve against a clause A 0 1 . So M has a child node Let the ancestor list of each A i at M be ALA i . Then the ancestor list ALB i of each B i at N is ALA1 [ fA 1 g and the ancestor list ALA j of each A j is ALA j . Obviously, for any subgoals A and B, if A is in the ancestor list of B, i.e. A 2 ALB , the proof of A requires the proof of B. Denition 2.7 Let G i and G k be two nodes in an SLD-derivation and A and B the selected subgoals in G i and G k , respectively. We say A is an ancestor subgoal of B, denoted A ANC B, if The following result shows that the ancestor relation ANC is transitive. Theorem 2.1 If A 1 ANC A 2 and A 2 ANC A 3 , then A 1 ANC A 3 . Proof. By the denition of ancestor lists, for any subgoal A if A 2 ALA 0 , then ALA ALA 0 . So ALA 3 . Thus A 1 2 ALA 2 ALA3 implies A 1 2 ALA 3 . That is, A 1 ANC A 3 . 2 In [17], it is phrased as more sound. With no loss in generality, in the sequel we assume the leftmost computation rule. So the selected subgoal at each node is the leftmost subgoal. For convenience, for any node (goal) G i , unless otherwise specied we use A i to refer to the leftmost subgoal of G i . 3 Expanded Variants To design a complete and reliable loop check, we rst need to determine what principal characteristics that an innite loop possesses. Consider the innite loop L 2 (see the Introduction) again. We notice that for any i 0, the subgoal p(f(::f(f(g(a)))::)) at the (i + 1)-th node G i+1 is a variant of the subgoal p(f(::f(g(a))::)) at the i-th node G i except for the function g(a) at G i that grows into a function f(g(a)) at G i+1 . However, If we replace g(a) with a constant a in L 2 , then p(f(::f(f(a))::)) at G i+1 is a variant of p(f(::f(a)::)) at G i except for the constant a at G i that grows into a function f(a) at G i+1 . Furthermore, If we replace g(a) with a variable X in L 2 , then p(f(::f(f(X))::)) at G i+1 is a variant of p(f(::f(X)::)) at G i except for the variable X at G i that grows into a function f(X) at G i+1 . As another example, consider the program Let the top goal G Z). Then we will get an innite loop L 3 as depicted in Fig.1. Observe that for any i > 0, the subgoal at G 2(i+1) is a variant of that at G 2i except that the variable Y at G 2i grows into f(a; Y ) at G 2(i+1) . Fig.1 The innite loop L 3 . These observations reveal a key structural characteristic of some subgoals in an innite loop with functions, which can be formalized as follows. Denition 3.1 Let A and A 0 be two atoms/functions. A 0 is said to be an expanded variant of A, denoted A 0 wEV A, if A 0 is a variant of A except that there may be some terms at certain positions in A each A[i]:::[k] of which grows in A 0 into a function A 0 Such terms like A[i]:::[k] in A are then called growing terms w.r.t. A 0 . The following result is immediate. Theorem 3.1 If A is a variant of B, then A wEV B. Example 3.1 At each of the following lines, A 0 is an expanded variant of A because it is a variant of A except for the growing terms GT . However, at the following lines no A 0 is an expanded variant of A. c) /*c and b are not uniable is not in f(X) to the case that p(X; X) is not a variant of p(Y; X) In the above example, p(X; f(X)) is an expanded variant of p(X; X). It might be doubtful how that would happen in an innite loop. Here is an example. Example 3.2 Let P be a logic program and G a top goal. We have the following innite loop: Clearly, for any i 0, the subgoal A i+1 at G i+1 is the subgoal A i at G i with the second X growing to f(X). That is, A i+1 is a variant of A i except for A i+1 Any expanded variant has the following properties. 3 This example is suggested by an anonymous referee. Theorem 3.2 Let A 0 wEV A. (1) jAj jA 0 j. (2) For any i, ., k, jA[i]:::[k]j jA 0 [i]:::[k]j. (3) When are variants. When jAj 6= jA 0 j, there exists i such that jA[i]j < jA 0 [i]j. Proof: (1) and (2) are immediate from Denition 3.1. By (2), when That is, there is no growing term in A, so by Denition 3.1 A 0 is a variant of A. This proves (3). Finally, (4) is immediate from (2). 2 These properties are useful for the computation of expanded variants. That is, if jA 0 j < jAj, we conclude A 0 is not an expanded variant of A. Otherwise, if we determine if both are variants. Otherwise, we proceed to their arguments (recursively) to nd growing terms and check if they are variants except for the growing terms. The relation \variant of" dened in Denition 2.1 yields an equivalent relation; it is re exive (i.e., A is a variant of itself), symmetric (i.e., A being a variant of B implies B is a variant of A), and transitive (i.e., if A is a variant of B and B is a variant of C, then A is a variant of C). However, the relation wEV is not an equivalent relation. Theorem 3.3 (1) A wEV A. (2) A wEV B does not imply B wEV A. (3) A wEV B and B wEV C does not imply A wEV C. does not imply B wEV C. Proof. (1) Straightforward by Theorem 3.1. (2) Here is a counter-example: p(f(X)) wEV p(X), but p(X) 6w EV p(f(X)). (3) Immediate by letting and Immediate by letting Although wEV is not transitive, the sizes of a set of expanded variants are transitively decreas- ing. The following result is immediate from Theorem 3.2. Corollary 3.4 If A wEV B and B wEV C, then jAj jCj. The concept of expanded variants provides a basis for designing loop checking mechanisms for logic programs with functions. This claim is supported by the following theorem. Theorem 3.5 Let be an innite SLD-derivation with innitely large subgoals. Then there are innitely many goals G such that for any j 1, Proof. Since D is innite, by the justication given by Bol [3] (page 40) D has an innite subderivation D 0 of the form where for any j 1, A 0 . Since any logic program has only a nite number of clauses, there must be a set of clauses in the program that are invoked an innite number of times in D 0 . be the set of all dierent clauses that are used an innite number of times in must have an innite subderivation D 00 of the form where for any j 1, A 00 and any logic program has only a nite number of predicate/function/constant symbols and D contains innitely large subgoals, there must be an innite sequence of A 00 s in such that for any j 1, A i j ANC A i j+1 and A i j is a variant of A i j+1 except for a few terms in A i j+1 whose size increases. Note that such an innite increase in term size in D 00 must result from some clauses in S that cause some terms I to grow into functions of the form f(:::I:::) each cycle S is applied. This means that A i j is a variant of A i j+1 except for some terms I that grow in A i j+1 into f(:::I:::), i.e., A i j+1 wEV A i j with VAF-Checks Based on expanded variants, we can dene a series of loop checking mechanisms for logic programs with functions. In this section, we present four representative VAF-checks and prove their completeness. Denition 4.1 Let P be a logic program, G 0 a top goal, and d 1 a depth bound. Dene in which there are up to d goals that satisfy the following conditions: (1) For each j d, A i j ANC A i j+1 and A i j+1 wEV A i j . (2) For any j d, jA or for any j d, jA Theorem 4.1 (1) V AF 1 (d) is a (simple) loop check. (2) V AF 1 (d) is complete w.r.t. the leftmost computation rule. Proof. (1) Straightforward from Denition 2.2. (2) Let be an innite SLD-derivation. Since P has only a nite number of clauses, there must be a set of clauses in P that are invoked an innite 4 Note that (1) the order of clauses in fC j 1 is not necessarily the same as that in S, say fC may contain duplicated clauses, say fC C1g. number of times during the derivation. Let be the set of all distinct clauses that are applied an innite number of times in D. Then, by the proof of Theorem 3.5 D has an innite sub-derivation of the form where for any j 1, fC distinguish between two cases. (i) There is no subgoal in D whose size is innitely large. Because any logic program has only a nite number of predicate symbols, function symbols and constants, there must be innitely many atoms in T that are variants. Let fB 1 be the rst in T that are variants. Then, by Theorem 3.1, for each 1 j d B j+1 wEV B j with jB j+1 the conditions of V AF 1 (d) are satised, which leads to the derivation D being pruned at the node with the leftmost subgoal B d+1 . (ii) There is a subgoal in D with innitely large size. Then by Theorem 3.5, there must be innitely many atoms in T that are expanded variants with growing terms. Let fB 1 be the rst in T such that for each 1 j d, B j+1 wEV B j with jB j+1 j > jB j j. Again, the conditions of V AF 1 (d) are satised, so that the derivation D will be pruned. 2 is complete for any d 1, taking d ! 1 leads to the following immediate corollary to Theorem 4.1. Corollary 4.2 For any innite SLD-derivation, there is an innite sub-derivation of the form such that all A i j s satisfy the two conditions of V AF 1 (d) (d !1). Observe that V AF 1 (d) identies innite loops only based on expanded variants of selected subgoals. More reliable loop checks can be built by taking into account the clauses selected to generate those expanded variants. Denition 4.2 Let P be a logic program, G 0 a top goal, and d 1 a depth bound. Dene in which there are up to d goals that satisfy the following conditions: (1) For each j d, A i j ANC A i j+1 and A i j+1 wEV A i j . (2) For any j d, jA or for any j d, jA (3) For all j d, the clause selected to resolve with A i j s is the same. g) Theorem 4.3 (1) V AF 2 (d) is a (simple) loop check. (2) V AF 2 (d) is complete w.r.t. the leftmost computation rule. Proof. (1) Straightforward. (2) By Corollary 4.2, for any innite SLD-derivation D, there is an innite sub-derivation in D of the form such that all A 0 s satisfy the rst two conditions of V AF 2 (d). Since any logic program has only a nite number of clauses, there must be a clause C k that resolves with innitely many A 0 s in the sub-derivation. Let A d be the rst d A 0 s that resolve with C k . The third condition of (d) is then satised, so we conclude the proof. 2 Again, taking d !1 leads to the following immediate corollary to Theorem 4.3. Corollary 4.4 For any innite SLD-derivation, there is an innite sub-derivation of the form such that all A i j s satisfy the two conditions of V AF 1 (d) (d !1). (d) is a special case of V AF 1 (d), any SLD-derivation pruned by V AF 2 (d) must be pruned by V AF 1 (d), but the converse is not true. As an example, consider the SLD-derivation It will be cut by V AF 1 (2) but not by V AF 2 (2) because condition (3) is not satised. This leads to the following. Theorem 4.5 V AF 2 (d) is more reliable than V AF 1 (d). considers only the repetition of one clause in an innite SLD-derivation. More constrained loop checks can be developed by considering the repetition of a set of clauses. Denition 4.3 Let P be a logic program, G 0 a top goal, and d 1 a depth bound. Dene in which there are up to d goals that satisfy the following conditions: (1) For each j d, A i j ANC A i j+1 and A i j+1 wEV A i j . (2) For any j d, jA or for any j d, jA (3) For all j d, the clause selected to resolve with A i j s is the same. (4) For all j d, the set S of clauses used to derive A i j+1 from A i j is the same. g) Theorem 4.6 (1) V AF 3 (d) is a (simple) loop check. (2) V AF 3 (d) is complete w.r.t. the leftmost computation rule. Proof. (1) Straightforward. (2) By Corollary 4.4, for any innite SLD-derivation D, there is an innite sub-derivation in D of the form such that all A 0 s satisfy the rst two conditions of V AF 3 (d). Obviously, the third condition of is satised as well. Since any logic program has only a nite number of clauses, there must be an innite sequence, A 0 l 1 l j ; :::, of A 0 in the sub-derivation such that the set S of clauses used to derive A 0 l j+1 from A 0 l j is the same. Let A be the rst s. The fourth condition of V AF 3 (d) is then satised. 2 Taking d !1 leads to the following immediate corollary to Theorem 4.6. Corollary 4.7 For any innite SLD-derivation, there is an innite sub-derivation of the form such that all A i j s satisfy the three conditions of V AF 2 (d) (d ! 1) and that for any j 1, g. Obviously, any SLD-derivation pruned by V AF 3 (d) must be pruned by V AF 2 (d). But the converse is not true. Consider the SLD-derivation It will be cut by V AF 2 (2) but not by V AF 3 (2) because condition (4) is not satised. This leads to the following. Theorem 4.8 V AF 3 (d) is more reliable than V AF 2 (d). Before introducing another more constrained loop check, we recall a concept of recursive clauses, which was introduced in [16]. Denition 4.4 ([16]) A set of clauses, fR are called recursive clauses if they are of the form (or similar forms) where for any 0 < i < m, q i (:::X i 1 :::) in R i 1 is uniable with q i (:::X i :::) in R i with an mgu containing in Rm is uniable with q 0 (:::X 0 :::) in R 0 with an mgu containing f(:::X m :::)=X 0 . Put another way, fR is a set of recursive clauses if starting from the head of R 0 ( replacing X 0 with X) applying them successively leads to an inference chain of the form such that the last atom q 0 (:::f (:::X:::):::) is uniable with the head of R 0 with an mgu containing Example 4.1 The sets of clauses, fC 11 g in P 1 , fC 21 g in P 2 , fC in P 3 , and fC 41 g in P 4 , are all recursive clauses. Recursive clauses cause some subgoals to increase their size recursively; i.e., each cycle fR is applied, the size of q 0 (:) increases by a constant. If fR can be repeatedly applied an innite number of times, a subgoal q 0 (:) will be generated with innitely large size (note that not any recursive clauses can be repeatedly applied). Since any logic program has only a nite number of clauses, if there exist no recursive clauses in a program, there will be no innite SLD-derivations with innitely large subgoals, because no subgoal can increase its size recursively. This means that any innite SLD-derivation with innitely large subgoals is generated by repeatedly applying a certain set of recursive clauses. This leads to the following. Denition 4.5 Let P be a logic program, G 0 a top goal, and d 1 a depth bound. Dene in which there are up to d goals that satisfy the following conditions: (1) For each j d, A i j ANC A i j+1 and A i j+1 wEV A i j . (2) For any j d, jA or for any j d, jA (3) For all j d, the clause selected to resolve with A i j s is the same. (4) For all j d, the set S of clauses used to derive A i j+1 from A i j is the same. (5) If for any j d jA contains recursive clauses that lead to the size increase. g) Theorem 4.9 (1) V AF 4 (d) is a (simple) loop check. (2) V AF 4 (d) is complete w.r.t. the leftmost computation rule. Proof. (1) Straightforward. (2) By Corollary 4.7, for any innite SLD-derivation D, there is an innite sub-derivation E in D of the form such that all A 0 s satisfy the rst four conditions of V AF 4 (d) (d !1). Now assume that for any j. Then E contains A 0 innitely large size. Such innitely increase in term size in E must be generated by the repeated applications of some recursive clauses. This means that there must be an innite sequence, A 0 l 1 l j ; :::, of A 0 s in E such that the clauses used to derive A 0 l j+1 from A 0 l j contain recursive clauses that lead to the size increase from A 0 l j to l j+1 . Let A be the rst s. Then all A i j s satisfy the ve conditions of When d !1, we obtain the following corollary to Theorem 4.9. Corollary 4.10 For any innite SLD-derivation, there is an innite sub-derivation of the form such that for any j 1, A g, and for all or for all where the size increase results from the application of a set of recursive clauses in fC k ; :::; C n j g. is an enhancement of V AF 3 (d), any SLD-derivation pruned by V AF 4 (d) must be pruned by V AF 3 (d). But the converse is not true. Consider the program that consists of the clauses p(f(a)). The SLD-derivation p(a) )C1 p(f(a)) )C2 2 will be cut by V AF 3 (1) but not by V AF 4 (1) because there are no recursive clauses in the program. So we have the following result. Theorem 4.11 V AF 4 (d) is more reliable than V AF 3 (d). Example 4.2 Let us choose the depth bound d = 1. Then by applying any one of the four VAF- checks, all the four illustrating innite loops introduced earlier, will be cut at some node. That is, L 1 , L 2 and L 4 will be pruned at G 1 (the second node from the root), and pruned at G 4 . Example 4.3 Consider the following list-reversing program (borrowed from [3]) and the top goal G Z). Note that C 53 is a recursive clause. Again, let us choose d = 1. After successively applying the clauses C 52 , C 53 and C 53 , we get the following SLD-derivation: It is easy to check that there is no expanded variant, so we continue to expand G 3 . We rst apply C 51 to G 3 , generating a successful node 2; we then apply C 52 to G 3 , generating a node As A 3 ANC A 5 and A 5 wEV A 3 with jA 5 are satised, which stop expanding G 5 . We then apply C 53 to G 3 , generating a node Obviously, A 3 ANC A 6 and A 6 wEV A 3 with jA 6 j > jA 3 j where the size increase of A 6 is via the recursive clause C 53 , so V AF 1 4 (1) are satised again, which stop expanding G 6 . Since V AF 1 4 (1) cut all innite branches while retaining the (shortest) successful SLD-derivation they are weakly sound for g. Observe that each condition of the above VAF-checks captures one characteristic of an innite loop. Obviously, except (1) and (5), all the conditions (2) (4) make sense only when d > 1. Because expanded variants capture a key structural characteristic of subgoals in innite loops, all the VAF-checks with are weakly sound for a majority of representative logic programs (see the above examples). However, considering the undecidable nature of the loop checking problem, choosing d > 1 would be safer. 5 The following example, although quite articial, illustrates this point. Example 4.4 Consider the following logic program p(f(a)). C 62 and the following successful SLD-derivation D for the top goal G 5 As mentioned by Bol [3], the question of which depth bound is optimal remains open. However, our experiments show that V AF2(2) is weakly sound for a vast majority of logic programs. Obviously, p(a) ANC p(f(a)), p(f(a)) wEV p(a), and C 61 is a recursive clause. If we choose the derivation D will be pruned at G 1 by all the above four VAF-checks. That is, V AF 1 4 (1) are not weakly sound for this program. Apparently, V AF 1 4 (2) are weakly sound. Observe that from V AF 1 (d) to V AF 4 (d), the reliability increases, but the computational overhead increases as well. Therefore, we need to consider a trade-o in choosing among these VAF- checks. For practical applications, when d > 1 we suggest choosing the VAF-checks in the following (d). The basic reasons for such a preference are (i) our experience shows that V AF 2 (2) is weakly sound for a vast majority of logic programs, and (ii) the check of condition (3) of V AF 2 (d) takes little time, whereas the check of recursive clauses (condition (5) of V AF 4 (d)) is rather costly. 5 Comparison with OS-Check and EVA-Check Because OS-check, EVA-check and V AF 1 4 (d) are complete loop checks, we make the comparison based on the two key factors: reliability and computational overhead. 5.1 Comparison with OS-Check We begin by recalling the formal denition of OS-check. Denition 5.1 ([3, 14]) Let P be a logic program, G 0 a top goal, and d 1 a depth bound. Let size be a size-function on atoms. Dene in which there are up to d goals such that for any 1 j d There are three versions of OS-check, depending on how the size-function size is dened [14, 3]. In the rst version, atoms A and B, so condition (2) will always hold and thus can be ignored. In the second version, atom A. And in the third version, for any atoms A and B with the same arity n, jA[i]j jB[i]j. Obviously, the third version is more reliable than the rst two versions so we can focus on the third version for the comparison. OS-check is complete [3], but is too weak in that it identies innite loops mainly based on the size-function, regardless of what the internal structure of atoms is. Therefore, in order to increase its reliability, we have to choose the depth bound d as large as possible. For example, in [14] However, because the internal structure of atoms with functions may vary drastically in dierent application programs, using only a large depth bound together with the size-function as the loop checking criterion could in general be ineective/ine-cient. For example, when applying OSC(10; size) to the programs would generate a lot of redundant nodes. The following example further illustrates this fact. Example 5.1 Consider the following logic program and top goal: 100). C 7;100 The successful SLD-derivation for is as follows: | {z } It is easy to see that OSC(d; size) is not weakly sound for this program unless we choose d 100. In contrast, in our approach the common structural features of repeated subgoals in in- nite loops are characterized by expanded variants. Based on expanded variants the VAF-checks are weakly sound with small depth bounds (e.g. d 2) for a majority of logic programs. For instance, V AF 1 4 (1) are weakly sound for P 7 in the above example, which shows a dramatical dierence. The above discussion is summarized in the following results Theorem 5.1 Let size be the size-function of the third version of OS-check. For any atoms A and B, A wEV B implies size(B) size(A). Proof. Immediate from Theorem 3.2. 2 Theorem 5.2 For any 1 i 4, V AF i (d) is more reliable than OSC(d; size). Proof. By Theorem 5.1 and Corollary 3.4, OSC(d; size) will be satised whenever condition (1) of V AF i (d) holds. So any SLD-derivations pruned by V AF i (d) will be pruned by OSC(d; size) as well. But the reverse is not true. As a counter-example, when d < 100, the SLD-derivation in Example 5.1 will be pruned by OSC(d; size) but not by V AF i (d). 2 We now discuss computational overhead. First note that in both OS-check and the VAF- checks, the ancestor checking, A i j ANC A i j+1 , is required. Moreover, for each ancestor subgoal A i j of A k , in OSC(d; size) we compute Although the computation of expanded variants is a little more expensive than that of the size-function, both are processes of two strings (i.e. atoms). Since string processing is far faster than ancestor checking (which needs to scan the goal-stack), we can assume that the two kinds of string computations take constant time w.r.t. scanning the goal-stack. Under such an assumption, the complexity of OSC(d; size) and V AF 1 2 (d) is the same (note that the check of conditions (2) and (3) of the VAF-checks takes little time). Since the check of condition (4) of the VAF-checks requires scanning the goal-stack, V AF 3 (d) is more expensive than OSC(d; size). Furthermore, condition (5) of the VAF-checks, i.e. the computation of recursive clauses, is quite expensive because on the one hand, given a logic program we need to determine which clauses in it are recursive clauses, and on the other hand, for two subgoals A i j and A i j+1 with jA in an SLD-derivation, we need to nd if the size increase from A i j to A i j+1 results from some recursive clauses. This means that V AF 4 (d) could be much more expensive than OSC(d; size). The above discussion further suggests that V AF 2 (d) is the best choice (balanced between reliability and overhead) among OSC(d; size) and V AF 1 4 (d). 5.2 Comparison with EVA-Check We begin by reproducing the denition of EVA-check. Denition 5.2 ([17]) Let P be a logic program, G 0 a top goal, and d 1 a depth bound. Dene in which there are up to d goals such that for any 1 j d (2) A k is a generalized variant of A i j .g) Here, a subgoal A 0 is said to be a generalized variant of a subgoal A if it is a variant of A except that there may be some arguments whose size increases from A via a set of recursive clauses. The following characterization of generalized variants is immediate from the above denition and Denition 3.1. Theorem 5.3 For any subgoals A 0 and A in an SLD-derivation, A 0 is a generalized variant of A if and only if A 0 wEV A and if jA 0 j > jAj then the size increase is via a set of recursive clauses. EV A(d) relies heavily on recursive clauses, so its complexity is similar to V AF 4 (d). Since the computation of recursive clauses is too expensive, we will not choose EV A(d) in practical applications unless it is more reliable than some V AF i (d). However, the following example shows that EV A(d) can not be more reliable than any of the four VAF-checks. Example 5.2 Consider the following logic program and top goal: p(f(a)). C 83 A successful SLD-derivation for is as follows: It can be easily seen that fC 81 ; C 82 g and fC 82 g are two sets of recursive clauses. Let us choose 2. Then A 2 is a generalized variant of both A 0 and A 1 , so EV A(2) will cut the derivation at . However, this SLD-derivation will never be cut by any V AF i (2) because condition (2) of the VAF-checks is not satised (i.e. we have jA 6 Conclusions We have developed four VAF-checks for logic programs with functions based on the notion of expanded variants. We observe that the key structural feature of innite loops is repetition (of selected subgoals and clauses) and recursive increase (in term size). Repetition leads to variants (because a logic program has only a nite number of clauses and predicate/function/constant recursive increase introduces growing terms. The notion of expanded variants exactly catches such a structural characteristic of certain subgoals in innite loops. Due to this, the VAF-checks are much more reliable than OS-check and no less reliable than EVA-check even with small depth bounds (see Examples 5.1 and 5.2). On the other hand, since the structural information is extracted directly from individual subgoals, without appealing to recursive clauses, the VAF-checks (except V AF 4 (d)) are much more e-cient than EVA-check. In balancing between the reliability and computational overhead, we choose V AF 2 (d) as the best one for practical applications. Although V AF 2 (2) is reliable for a vast majority of logic programs, due to the undecidability of the loop checking problem, like any other complete loop checks, V AF 2 (d) in general cannot be weakly sound for any xed d. The only way to deal with this problem is by heuristically tuning the depth bound in practical applications. Methods of carrying out such a heuristic tuning then present an interesting open problem for further study. Acknowledgements We thank the anonymous referees for their constructive comments, which have greatly improved the presentation. The rst author is supported in part by Chinese National Natural Science Foundation and Trans-Century Training Programme Foundation for the Talents by the Chinese Ministry of Education. --R An analysis of loop checking mechanisms for logic programs Towards more e-cient loop checks Loop checking in partial deduction Tabulated resolution for the Well-Founded semantics Tabled evaluation with delaying for general logic programs Eliminating unwanted loops in Prolog A further note on loops in Prolog Termination of logic programs: the never-ending story Redundancy elimination and loop checks for logic pro- grams Foundations of Logic Programming Partial evaluation in logic programming On eliminating loops in Prolog The XSB Programmer's Manual (Version 1.8) The mixtus approach to automatic partial evaluation of full Prolog Mixtus: an automatic partial evaluator for full Prolog Verifying local strati An extended variant of atoms loop check for positive logic programs Linear tabulated resolution for the well founded semantics An abstract approach to some loop detection problems OLD resolution with tabulation the power of logic Memoing for logic programs --TR Controlling recursive inference OLD resolution with tabulation Efficient loop detection in Prolog using the tortoise-and-hare technique Foundations of logic programming; (2nd extended ed.) Recursive query processing: the power of logic An analysis of loop checking mechanisms for logic programs Partial evaluation in logic programming The Mixtus approach to automatic partial evaluation of full Prolog Towards more efficient loop checks Memoing for logic programs Mixtus Sound and complete partial deduction with unfolding based on well-founded measures Redundancy elimination and loop checks for logic programs Tabled evaluation with delaying for general logic programs An extended variant of atoms loop check for positive logic programs An abstract approach to some loop detection problems Linear Tabulated Resolutions for the Well-Founded Semantics --CTR Yi-Dong Shen , Jia-Huai You , Li-Yan Yuan , Samuel S. P. Shen , Qiang Yang, A dynamic approach to characterizing termination of general logic programs, ACM Transactions on Computational Logic (TOCL), v.4 n.4, p.417-430, October Etienne Payet , Fred Mesnard, Nontermination inference of logic programs, ACM Transactions on Programming Languages and Systems (TOPLAS), v.28 n.2, p.256-289, March 2006 Alexander Serebrenik , Danny De Schreye, Inference of termination conditions for numerical loops in Prolog, Theory and Practice of Logic Programming, v.4 n.5-6, p.719-751, September 2004

logic programming;loop checking

504580

Partial correctness for probabilistic demonic programs.

Recent work in sequential program semantics has produced both an operational (He et al., Sci. Comput. Programming 28(2, and an axiomatic (Morgan et al., ACM Trans. Programming Languages Systems 18(3) (1996) 325-353; Seidel et al., Tech Report PRG-TR-6-96, Programming Research group, February 1996) treatment of total correctness for probabilistic demonic programs, extending Kozen's original work (J. Comput. System Sci. 22 (1981) 328-350; Kozen, Proc. 15th ACM Symp. on Theory of Computing, ACM, New York, 1983) by adding demonic nondeterminism. For practical applications (e.g. combining loop invariants with termination constraints) it is important to retain the traditional distinction between partial and total correctness. Jones (Monograph ECS-LFCS-90-105, Ph.D. Thesis, Edinburgh University, Edinburgh, UK, 1990) defines probabilistic partial correctness for probabilistic, but again not demonic programs. In this paper we combine all the above, giving an operational and axiomatic framework for both partial and total correctness of probabilistic and demonic sequential programs; among other things, that provides the theory to support our earlier---and practical---publication on probabilistic demonic loops (Morgan, in: Jifeng et al. (Eds.), Proc. BCS-FACS Seventh Refinement Workshop, Workshops in Computing, Springer, Berlin, 1996. Copyright 2001 Elsevier Science B.V.

Introduction Deterministic computation over a state space S can be modelled as functions of type S ! S, from initial to final states. A 'powerdomain' construction extends that to nondeterminism, and although the traditional powerdomains - Smyth, Hoare and Plotkin - differ in their treatment of non-termination, they all agree that nondeterminism is 'demonic', resolved in some arbitrary way. The probabilistic powerdomain [7, 6] instead resolves nondeterminism according to some specified distribution over final states: demonic choice is removed, and replaced by probabilistic choice. He et al. [4] do not remove demonic choice; rather they model demonic and probabilistic nondeterminism jointly by combining a special case of the probabilistic powerdomain, for imperative programs [9], with the Smyth (de- monic) construction. Morgan et al. [13] then complement that with a programming logic of 'greatest pre-expectations' (extending Kozen's work [10]), resulting overall in a treatment of total correctness for probabilistic and demonic sequential programs. In this paper we extend the constructions of He and Morgan to partial correctness also. One important application for both forms of correctness is the justification of invariant/variant principles for probabilistic demonic loops: although published [11], those principles are based on (only) postulated connections between wp and wlp for probabilistic programs - here we provide the theory for their proof. Another application is in the abstraction from probabilistic to demonic choice; we discuss both issues in the conclusion. To model partial (Hoare-style) and total (Smyth-style) correctness within the same framework, we choose the Egli-Milner construction for demonic choice (rather than the Smyth); and its being based on distributions rather than simple states introduces some novel considerations (in particular the need for linear interpolation between distributions). For the logic we are then able to formulate both greatest- and greatest liberal pre-expectations, reflecting the same total/partial distinction as in standard programming logic [2]. In Sec. 2 we construct the probabilistic Plotkin-style powerdomain, the 'convex' powerdomain (which is a subset of the general representation in Abramsky and Jung [1]), and we show how to extract partial and total information from it; in Sec. 3 we link that to greatest-precondition-based probabilistic programming logic; and in Sec. 4 and Sec. 5 we specialise to both a liberal and 'ordinary' logic for a sequential programming language with both probabilistic and demonic nondeterminism. An extensive discussion of examples, and the general treatment of loops, is given elsewhere [11]. Throughout we use infix dot ':' for function application, associating to the left so that f:x:y means (f(x))(y); and we write ': =' for 'is defined to be equal to'. Quantifications (including set comprehensions) are written in the order quantifier (or 'f' for sets), bound variable with optional type, range and finally term - thus for example is the set of the first ten squares. 2 A convex powerdomain of distributions In program semantics, powerdomains are used to study nondeterminism, a phenomenon arising when a program outputs an arbitrary result (drawn from a set of possible results) rather than a single, determined function of its input. Here we consider powerdomains over a domain of probability distributions rather than of 'definite' single states. There are several ways of ordering sets of distributions (each resulting in a different powerdomain): the choice depends on criteria which can be explained in terms of the desired treatment of programs' possible non-terminating behaviour. The Smyth order 1 (Def. B2) treats non-termination as the worst behaviour and thus the Smyth powerdomain models total cor- rectness. Similarly the Hoare order (Def. B3) models partial correctness: non-termination is treated as the best outcome in that order. The Plotkin power- domain uses the Egli-Milner order (Def. B4) which combines both views, and is useful when both partial and total correctness are to be modelled within a single framework. In general the Plotkin powerdomain is not decomposable into the Smyth and Hoare powerdomains, but in some special cases it is: Abramsky and Jung [1] show that one such case is when the underlying domain is !-continuous (Def. B9). In this section we show how Abramsky and Jung's results apply to an appropriately-defined powerdomain of distributions, thus providing a single semantic space for both partial and total correctness of programs exhibiting probabilistic and demonic nondeterminism. 1 For this and other facts and definitions from domain theory, we follow the conventions set out in [1]. We summarise the details for this paper (often specialising them to our particular application) in Appendix B. We write S for the state space, and assume it is finite. The space of probability distributions over S is defined as follows. Definition 2.1 For state space S, the space of distributions 2 (S; v) over S is defined and for F; F 0 in S we define show first that (S; v) is an !-continuous complete partial order. Lemma 2.2 For S a finite state space, its distributions (S; v) form an !- continuous complete partial order. Proof: The completeness of (S; v) is trivial, given the completeness of the interval [0; 1] under - over the reals. To show that S is !-continuous we only need exhibit a countable basis (Def. B8). Since S is finite, we use the set of distributions contained in since any real is the least upper bound of the rationals way-below it (Def. B7).We now define a Plotkin-style powerdomain over S. For subset A of S we write "A for its up-closure and #A for its down-closure (Def. B1). We say that A is up-closed (Smyth-closed) if "A = A, that it is down-closed (Hoare-closed) if A and that it is Egli-Milner-closed 3 if A further closure condition is related to continuity. These special distributions are more precisely called discrete sub-probability measures [9]; they do not necessarily sum to 1, and the deficit gives the probability of nontermination. The 'everywhere zero' distribution for example, that assigns zero probability to all states, models nowhere-terminating behaviour. (An alternative though less convenient treatment would assign probability 1 to some special state ?.) 3 This is often called convex closed ; but we will need that term for another purpose. Definition 2.3 For subset A of S we define its limit closure lc:A to be the smallest set containing A itself, together with tA 0 and uA 00 for all up-directed (Def. B5) subsets A 0 and down-directed (Def. B6) subsets A 00 of A. 4 We say that A is limit-closed if Before defining our powerdomain we must introduce one further closure condition, specific to distributions. For distributions F; F 0 in S and p in [0; 1] we can form F p \Phi F 0 , the weighted average, defined pointwise over S as p \Theta F + (1\Gammap) \Theta F 0 (with usual scalar multiplication and addition). For sets of distributions we define p-averaging as follows. Definition 2.4 For p in [0; 1] and subsets A; A 0 of S we define We say that A is convex if A p \Phi We now can define our powerdomain over S; it is a subset of the Plotkin powerdomain. Definition 2.5 The convex powerdomain (CS; vEM ) over the space of distributions S comprises those subsets of S that are non-empty, Egli-Milner closed, limit closed and convex. Its order vEM is the usual Egli-Milner order (Def. B4). 2 The convex powerdomain is a subset of the Plotkin powerdomain because it includes only the convex subsets of the latter. Our aim for this section is to show that CS is itself limit complete, and that from its limits the Smyth and Hoare limits can be extracted - for that is what makes it suitable for our application of it to probabilistic program semantics. Thus we show that the least upper bound of an Egli-Milner-directed subset of CS lies within CS, and that it is a combination of the Hoare least upper bound (tH ) and the Smyth least upper bound (t S ). The next lemma is the specialisation of a general decomposition result of Plotkin powerdomains to (S; v). We write Lens(S) (Def. B10) for the set of non-empty, Egli-Milner closed, limit closed subsets of S. 5 4 Note that both tA and uA exist for all subsets A of S in particular, directed or not. 5 Thus CS comprises just the convex lenses of S. Lemma 2.6 For any Egli-Milner-directed subset A of Lens(S) the limit tEMA exists, and satisfies (Insisting on limit closure after tH can be seen here as a continuity 6 condi- tion.) Proof: The decomposition will be a consequence of the isomorphim between the abstract Plotkin powerdomain (Def. B12) and the space of lenses of an !-continuous domain with the topological Egli-Milner ordering (Def. B13). The isomorphism is given by Abramsky and Jung [1] and is reproduced here in Thm. B.1. The decomposition (1) holds for abstract Plotkin powerdomains in general [8]. To establish the isomorphism, note first that the closure conditions on Lens(S) imply that vEM between lenses reduces to the topological Egli- Milner ordering; then Lem. 2.2 ensures the conditions of Thm. B.1, namely that (S; v) is an !-continuous complete partial order. 2 Since CS is a subset of Lens(S), we have in the following corollary our closure under limits. Corollary 2.7 For any Egli-Milner-directed subset A of CS the limit tEM exists in CS, and satisfies Proof: Given Lem. 2.6, we need only show that t S A " lc:(t H if all the elements of A are, and that follows from these elementary facts: up- closing preserves convexity (v-monotonicity of p \Phi); the intersection of convex sets is convex; down-closing preserves convexity (similar to up-closing); the union of a '-directed set of convex sets is convex; and limit-closing preserves convexity (v-continuity of p \Phi). 2 6 Consider the Egli-Milner chain on sets of real intervals in [0; 1], which has limit f1g (the limit point of the underlying series). The union of the down sets is the half-closed interval [0; 1) however, and the intersection of the up sets is f1g. Failing to limit-close the Hoare limit would produce an empty result. Cor. 2.7 gives us our main result, that the Egli-Milner limit determines the Smyth limit (in the Smyth ordering) and the Hoare limit (in the Hoare ordering). We write respectively (Def. B14) for Smyth equivalence and Hoare equivalence between elements of CS: our theorem below shows in addition that the limits are indistinguishable relative to the appropriate equivalences. Theorem 2.8 For any vEM -directed subset A of CS, the following equivalences hold: Proof: This too is a property of abstract Plotkin powerdomains [8], and so follows from the isomorphism (Thm. B.1) used in the proof of Thm. 2.7.This section has defined the convex powerdomain, whose use for modelling probabilistic imperative programs now follows from the constructions for the Smyth-style domain [4]: for example sequential composition is a generalised functional composition; nondeterministic choice is union (then convex closure); and probabilistic choice is weighted average as defined above. In Sec. 4 we give further details. For recursion one takes limits of chains, and here is the significance of Thm. 2.8: we must be sure that taking the limit in the convex domain agrees with the more specialised limit in the Smyth domain and with the more specialised limit in the Hoare domain - for that is what allows us to use the more general convex domain for either. It is known that the equivalence holds for standard (nonprobabilistic) domains 7 ; Thm. 2.8 confirms the preservation of the property when probability is included. In the next section we link the convex powerdomain with its partial/total probabilistic programming logic. 7 In fact the same general argument of this section establishes that, since the standard approach takes a flat domain S for the state space, trivially !-continuous if S is finite, or even countable. 3 Probabilistic programs and logic In this section we use the convex powerdomain CS of Sec. 2 to construct a model for both total and partial correctness of probabilistic demonic programs. 8 We begin with a review of methods that treat the two aspects separately. Over standard (non-probabilistic) demonic programs, a popular model for total correctness is S ! SS? , where S? is the flat domain extending state space S with ? for non-termination, and S forms the Smyth powerdomain over that; Dijkstra's weakest `ordinary' preconditions PS ! PS [2] support a programming logic suitable for total correctness. For partial correctness one can use S ! HS? (Hoare) for the model and weakest 'liberal' preconditions for the logic. Finally, although partial and total correctness are available simultaneously via S ! KS? (Plotkin), for r in S ! KS? and postcondition Q in PS still it is more convenient to define separately weakest precondition weakest liberal precondition (2) to give the total (wp) and partial (wlp) programming logics. Note that the definitions (2) work together only over KS (the intersection of HS? and SS? ) - wp does not work over HS? and wlp does not work over SS? . (Nelson [14] gives a nice treatment of the issues.) For probabilistic programs, He et al. [4] propose S ! C S S for total correctness, where C S S is like CS of the previous section, but based on the Smyth order. Morgan et al. [13] provide a probabilistic 'greatest pre- expectation' logic for that, where expectations are non-negative real-valued functions over the state space (extending Kozen's treatment [10] for non- programs). To access total and partial correctness simultaneously, by analogy with the standard case we simply replace He's Smyth-based C S by our Egli-Milner- based C. Yet rather than define two forms of logic (as at (2) above) we generalise as do Morgan and McIver [12] by allowing expectations to take negative values: roughly speaking, for total correctness one uses non-negative post-expectations and for partial correctness one uses non-positive. Kozen [9] has modelled total correctness of probabilistic (but not demonic) programs, and Jones [6] extended that to partial correctness. He [4] models total (but not partial) correctness of probabilistic demonic programs. We begin the details with the construction of the probabilistic, demonic model of programs. Definition 3.1 For (finite) state space S the space of probabilistic, demonic programs (MS;vEM ) is given by with the order induced pointwise from CS, so that for in MS we have We occasionally use v S and vH over MS, analogously lifted from CS. 2 Thus our programs take initial states to sets of final distributions: the plurality of the sets represents demonic nondeterminism; the distributions they contain each represent probabilistic nondeterminism. The next task is to investigate the dual representation of programs as expectation transformers. We extend the expectations found in Morgan et al. [13, 11], where the topic was total correctness (the Smyth order and up-closed sets) and expectations were of type S ! [0; 1], by using [\Gamma1; 1] instead: we write ES for S ! [\Gamma1; 1], and use lower-case Greek letters for typical elements. Expectation transformers T S are thus functions of type ES ! ES. We R F ff for the expected value of ff in ES averaged over distribution F in S. As a special case of expectations, we interpret predicates as f0; 1g- valued functions of the state space, and for predicate A holding at state s we convenient. For a scalar c we write c for the constant expectation evaluating to c over all of S. With those conventions the predicates true and false correspond to the expectations 1 and 0 respectively. Finally, for relations between expectations we write everywhere no more than everywhere equal to so that we generalise respectively implication, equivalence and reverse implication on predicates. Our logic is based on the 'extended greatest pre-expectation transformer', defined as follows. Definition 3.2 Let r be a program in MS, taking initial states in S to sets of final distributions over S. Then the greatest pre-expectation at state s of program r, with respect to post-expectation ff in ES, is defined: 9 Z F ff) :The effect of the definition is to consider all possible post-distributions F in r:s, and then demonically to choose the one that gives the least (the 'worst') expectation for ff: thus nondeterminism is demonic in that it minimises the pre-expectation at each initial state, and Def. 3.2 is then the greatest expectation everywhere no more than those pointwise minima. For standard programs, if executing a program r from a state s is certain to establish a postcondition A then that state is contained in the associated weakest precondition; with our definition we would have ewp:r:A:s = 1. For probabilistic programs, if the standard postcondition A is established with only a probability at least p say, then the greatest pre-expectation on executing r from s initially is at least p and we have Thus as a special case, when A is a predicate we can interpret ewp:r:A:s as the greatest assured probability that A holds after execution of r from s. Now we discover the various refinement orders over T S that correspond via ewp with orders over MS. First, we generalise the observation from standard programming (eg. [14]) that the Smyth order on programs corresponds to the implication order lifted to predicate transformers and that the Hoare order similarly corresponds to (lifted) reverse implication. We use PS (typi- cal element -) to denote the set of non-negative valued expectations and NS (typical element -) for the non-positive valued expectations. They are both subsets of ES. Lemma 3.3 For r 9 This reduces to Kozen's definition [9] for deterministic programs. There programs are functions from initial state to distributions, so that the minimisation ranges over a singleton set and is thus superfluous. 10 The apparent confusion between expectations and probabilities is deliberate and harm- less: the probability of an event A over a distribution is equal to the expected value of (the characteristic function of) A over that same distribution. Proof: For in MS, any s in S and - in PS we reason as follows r implies "(r:s) ' ''(r 0 :s) definition v S implies R R F -) R R For the deferred justification we appeal to the monotonicity of the arithmetic over non-negative arguments without subtraction: r:s differs from "(r:s) only by the addition of 'larger elements' according to Def. 2.1, and so the minimum selection made in Def. 3.2 cannot be increased by making the selection over the up-closure instead. The result now follows by generalising on s, and a similar argument justifies the second statement (but note the reversal W). 2 Lem. 3.3 is the key to defining the expectation-transformer equivalents to the Smyth, Hoare and Egli-Milner orders where, as usual, the Egli-Milner order is the intersection of the Smyth and Hoare orders. Definition 3.4 For t; t 0 in T S we define :That the Egli-Milner order between programs is preserved under ewp now directly. Corollary 3.5 For Proof: Lem. 3.3 and Def. 3.4. 2 The corollary shows only that ewp is an order-preserving mapping between (MS;vEM ) and (T S; vEM ). The next result of this section is to show that it is also an injection, and therefore that programs can be modelled equivalently either as relations or as expectation transformers. To establish that ewp is an embedding we prove the converse to Cor. 3.5. Lemma 3.6 For Proof: Suppose for a contradiction that ewp:r vEM ewp:r 0 but r 6v EM r 0 , for some in MS. Without loss of generality assume r 6v S r 0 , so that for some distribution F and state s we have both F 62 "(r:s) (3) From (3), with the aid of the separating hyperplane lemma A.1, we have for some expectation - in PS that Z F Z and thus that R ewp:r:-:s. From (4) however we have ewp:r 0 :-:s - R directly, giving together Z F and contradicting the hypothesis (at the state s). 2 Since in the proof of Lem. 3.6 we have actually proved the converse to Lem. 3.3, we can now state the correspondence between the relational model and program logic for all three orders. Theorem 3.7 The following equivalences hold for all r r vEM r 0 iff ewp:r vEM ewp:r 0 :We have shown that ewp injects MS into T S - but there are many vEM - monotonic expectation transformers that are not ewp-images of MS. The final result of this section identifies the exact sub-space of expectation transformers that form the image of MS through ewp. We identify 'healthiness in the style of Dijkstra [2] - for standard programs 11 The relevance of the healthiness conditions applied to the programming logic are treated in Sec. 5. - and of Morgan et al. [13] - for probabilistic programs - that identify the (images of) programs of MS within it. The importance of that result is that theorems proved within T S about healthy expectation transformers correspond to theorems about programs in MS. The first healthiness condition is a slight generalisation of the sublinearity of Morgan [13]. To state it we define, for expectations ff; fi in ES and real non-negative scalar c, the expectations ff+fi and cff, where (as for p-averaging of distributions) we mean a pointwise lifting of standard addition and scalar multiplication. Definition 3.8 An expectation transformer, t: T S is sublinear iff for all ff; fi in ES, and a; b; c non-negative reals, note first that sublinearity is satisfied by all images of MS under ewp. Lemma 3.9 Any expectation transformer ewp:r, for r in MS, is sublinear. Proof: Def. 3.2 and properties of arithmetic. (Morgan [13] gives a more detailed proof.) 2 For total correctness (for the Smyth C S ), sublinearity tells the whole story [13, Thm. 8.7]; in our more general CS however, there are sublinear elements of MS that are not ewp-images: take S to be the two-element state space fx; yg, and consider the result set It is convex, but not Egli-Milner closed 12 ; its associated expectation transformer formed by ewp is sublinear, but it is not the ewp-image of any element of MS. The characterisation of Egli-Milner closure is captured by a second healthiness condition - 'partial linearity' - which states that t:ff depends only on the pre-expectations of t applied to expectations in PS [ NS. In fact its closure is fF which it is indistinguishable using ewp for any ff in PS [ NS. Definition 3.10 An expectation transformer, t in T S is said to be partially linear if for all states s in S, and all expectations ff in ES, there are expectations - in PS and - in NS such that :Note that the implicit existential quantification in Def. 3.10 means there may be many decompositions of ff as a sum -. 13 We complete the correspondence between healthy expectation transformers and MS with the next theorem, which we state only. The proof is omitted as it is overly technical and not necessary for the rest of the paper. Theorem 3.11 An expectation transformer t in T S is both sublinear and partially linear if and only if there is r in MS such that 4 Probabilistic Guarded Commands In this section we illustrate the constructions of the previous two sections by giving equivalent relational (MS) and expectation transformer (T S) semantics for a simple sequential programming language that extends Dijkstra's guarded commands [2] with a binary probabilistic choice operator p \Phi. 14 The relational semantics of Fig. 1 writes s for the point distribution 15 13 A perhaps more alluring healthiness condition would be that t:ff is determined by its positive part (ff t 0) and its negative part (ff u 0); but does not hold for general probabilistic programs, although it does in the restricted set of standard programs and f0; 1; \Gamma1g-valued expectations [12]. 14 He et al. [4] and Morgan et al. [13] give similar presentations of semantics, but for total rather than partial correctness, and more detailed motivation can be found there. Note that demonic choice is retained, not replaced. 15 For state s in S we define the point distribution s: S ! [0; 1] to be As a special case, we write ? for (-s:0), the distribution that evaluates to 0 everywhere on S. R MS !MS For p \Phi, u and sequential composition, the Egli-Milner closure should be taken of the right-hand side. Figure 1: Probabilistic relational semantics concentrated on a single state s, and u for the demonic combination of pro- grams. The symbol p \Phi is used both for probabilistic combination of programs and for the p\Gammaaveraging explained in Sec. 2 between sets of distributions. Because they contain no demonic nondeterminism, the three primitive commands all yield singleton result sets. Probabilistic choice p \Phi returns the weighted average of the result sets of its operands, a singleton set if its arguments are. 16 The result set of the demonic nondeterministic choice between two programs is the convex closure of the union of the results of the operands - the closure models operationally that a demon could resolve the choice by flipping a coin of arbitrary bias. In sequential composition, both r and r 0 can be considered as sets of purely probabilistic programs 'compatible' with them: such programs, say are thus of type S ! S, and we write r 3 f for compatibility, meaning (8s: S \Delta f:s 2 r:s). Since the effect of two purely probabilistic programs in sequence is just a single distribution over s 00 say, defined Z f:s ds 0 (5) by 'averaging' the effects of f 0 over the intermediate distribution produced by f:s, for the general case we vary (5) over all f; f 0 to give In fact p can depend on the state; but for simplicity we assume here that it's constant. F is the v T -monotonic function such that F Figure 2: Probabilistic ewp semantics, where oe is in PS [NS and s is in S. That simplifies to the definition in Fig. 1 if we regard f 0 in R parametrised by s 00 , abbreviating (-s 0 R ds 0 ). In the ewp-semantics of Fig. 2 the post-expectation oe varies only over PS [NS; 17 we use u also for the pointwise minimum between expectations. Here probabilistic choice p \Phi returns the weighted average of the results of its expectation operands; and u takes the pointwise minimum, reflecting the demon's striving for the worst (least) result. Sequential composition becomes simple functional composition (as for standard predicate transformers [2]). In both cases recursion is dealt with by least fixed points in the appropriate orders: Thm. 3.7 shows that the two orders correspond. Our concern in the next section will be to recover total and partial semantics separately from Fig. 2: we will show how to define probabilistic wp and wlp that generalise Dijkstra's operational interpretations and that satisfy probabilistic versions of the standard laws for combining partial and total correctness. 5 Partial and total correctness Athough ewp acts over all of ES, we can extract two more-specialised logics from it: each acts conventionally over just the non-negative expectations PS. 17 Because an Egli-Milner-closed set is determined by its Smyth- and Hoare closures, that is sufficient. For a total correctness logic we merely restrict to PS directly, and use the order V. Definition 5.1 Let r be a program in MS; then the greatest pre-expectation of program r with respect to post-expectation - in PS, associating 0 with non-termination, is defined: easily from sublinearity: if - is in PS then so that wp:r:- is in PS also. Moreover Lem. 3.3 shows that this wp semantics of programs corresponds to a relational model with the Smyth ordering [13] - non-termination is the worst outcome in both semantics. For partial correctness we define a probabilistic wlp, again we restrict to the subspace (PS; V). Definition 5.2 r be a program in MS; then the greatest liberal pre- expectation of program r with respect to the post-expectation - in PS, associating 1 with non-termination, is easily from sublinearity of ewp:r that for - in NS, 3.2 we can readily show Def. 5.2 to be identical to Z F which is a demonic generalisation of the probabilistic wlp defined only for nondemonic programs by Jones [6]. Morgan [12] shows that (6) also generalises standard wlp [2]. and thus since lies in NS, so does ewp:r:(- \Gamma 1) from which we deduce that wlp:r is a well-defined expectation transformer in PS ! PS. Also Lem. 3.3 implies that the wlp semantics corresponds to a relational model with the Hoare ordering - accordingly non-termination is the best outcome. For example, let S be some finite portion of N , and for natural number N for the assignment taking every initial state to the final state N . The program illustrates the difference between wp and wlp. Writing for the expectation that evaluates to 1 when s is N and to 0 otherwise, we have indicating that the greatest expectation of termination in state 0 is pq, for all initial states. The greatest expectation of either termination at 0 or nontermination is found with wlp; we have Thus the wp observation gives the greatest guaranteed probability 19 of termination at 0 - and non-termination guarantees nothing. The wlp observation on the other hand returns the probability that either 0 is reached or the program fails to terminate - the usual interpretation for partial correctness. 19 This follows from the usual interpretation of the expected value of a f0; 1g-valued random variable A with respect to a probability distribution: it is the chance that the random outcome will establish the event "A evaluates to 1". F is the v T -monotonic function such that F Figure 3: Probabilistic wlp semantics, where - is in PS and s is in S. The greatest fixed point - is used for recursion. Similarly the wlp semantics of a looping program is calculated as the greatest fixed point 20 of a monotonic function over PS (rather than the least, as for wp). It is easily checked that specialising Fig. 2 to wlp produces only the changes shown in Fig. 3. (Note that PS [NS is closed under \Sigma1.) An alternative view of wlp and wp often put forward is to regard them as functions on programs which satisfy certain properties - for example Nelson [14] defines (standard) wp and wlp axiomatically by enforcing both the coupling law (at (7) below) and conjunctivity properties (both wp and wlp induce conjunctive predicate transformers). Thus the efficacy of the probabilistic definitions lies in the generalisation of those properties: Lem. 5.3 and Thm. 5.4 (also below) generalise respectively conjunctivity and coupling (7), and as such they form the main results of this section. We state the coupling here for standard programs [5]: where r is a program and A a predicate and we are using (though only here) the original meanings for wp and wlp [2], with V for 'implies at all states'. Law (7) implies that wp:r and wlp:r agree on initial states from which termination is guaranteed, and thus it underlies the practical treatment of looping programs - to prove total correctness of an iteration the work is divided between ensuring partial correctness (with a loop invariant), and an independent termination argument (with a variant). The probabilistic coupling Thm. 5.4 will allow a similar treatment for probabilistic looping programs. This follows from Def. 5.2 since the least fixed point of a v T -monotonic function becomes specialised first to W on NS ! NS which order is then shifted to PS ! PS by applying "1+". We consider first the appropriate generalisation of conjunctivity: conjunction of predicates is replaced by probabilistic conjunction [17] defined for non-negative expectations - 0 where t is pointwise maximum between expectations. Probabilistic conjunction reduces to ordinary conjunction when specialised to predicates. 21 Its importance in probabilistic reasoning is that it subdis- tributes through both wp and wlp images of programs - another consequence of sublinearity. 22 Lemma 5.3 For r in MS and - 0 in PS, Proof: Sublinearity (with monotonicity of ewp:r, and Defs. 5.1, 5.2. 2 Next we deal with coupling - Thm. 5.4, generalising (7), is the main result of this section. Theorem 5.4 For r in MS and - 0 in PS, Proof: take only the extreme values 0 and 1, then 22 One might have guessed that u is the appropriate generalisation of - but u does not (even sub-) distribute [6, 17]. Having established wp:r:(- 0 taking t0 on both sides: since on the left it has no effect, we achieve our result. 2 As a corollary we recover the standard rule (generalised) for combining partial and total correctness. Corollary 5.5 For r in MS and - in PS, Proof: Thm. 5.4 and that - j - & 1 for - in PS. 2 As a special case note that the wlp result implies the wp result at those states from which termination occurs with probability 1 - where wp:r:1 because (&1) is the identity. 6 Conclusion The technical contribution of this paper is to extend the Plotkin construc- tion, with its Egli-Milner order, to act over probability distributions rather than points, then showing that the decomposition [15] into respectively the convex Hoare (partial correcness) and convex Smyth (total correctness) pow- erdomains continues to hold - even under taking limits. A key feature is the imposition of convexity (linear interpolation) between distributions. Jones [6] defines a partial correctness logic based on expectations, but only for non-demonic programs, and she does not discuss the healthiness conditions on which the applicability of such logics (as calculational tools) depends. It was the realisation [13] that adding nondeterminism to Kozen's model corresponds to a weakening of the additive property of his logic to sublinearity that makes proofs such as in Thm. 5.4 and those in [11] reduce to simple arithmetic arguments. The use of general expectations (thus superseding purely non-negative expectations [13]) leads to an even simpler presentation of sublinearity - the more useful of the two healthiness conditions described here. There are two immediate applications. The first is the discovery of proof rules for loops in which partial and total correctness are separated: with wp and wlp together it can be shown [11] that I preserved by loop body is sufficient for I V wp:(do G ! body od):(I u :G) provided I V T , where T gives for each initial state the probability of the loop's termination. The second application is abstraction. For some programs only termination is probabilistic, and (partial) correctness can be established informally by considering all probabilistic choices to be demonic: algorithms falling in to that category typically use randomization as a method for searching a large space of potential witnesses, and examples include finding perfect hash functions and finding irreducible polynomials (Mehlhorn, Rabin [3]). Suppose the loop do G ! body od contains probabilistic choices in body , and denote by body the result of replacing them all with demonic choice. If (the fairness of) probability was essential for the loop's termination, the resulting standard loop do though probability-free, would not terminate and so a wp analysis of it would be useless. With a theory of partial correctness however we can reason as follows: for a standard loop invariant I, G u I V wlp:body:I established by standard reasoning say hence I V wlp:loop:(I u :G) standard loop rule hence I V wlp:loop:(I u where loop and loop are the two loops containing body and body respectively. Thus from standard reasoning about body we reach probabilistic conlusions about loop. The last step relies only on loop v loop, which by wlp-monotonicity of do od follows from body v body ; that in turn is guaranteed by (u) v ( p \Phi) for all p. Note that probabilistic wlp is essential - the last step would not hold if we had written wp:body . One would conclude the argument by determining T j wp:loop:1 sepa- rately, then having I from Thm. 5.4. A Standard results from linear programming Lemma A.1 The separating hyperplane lemma. Let C be a convex and (limit-) closed subset of R N , and F a point in R N that does not lie in C. Then there is a separating hyperplane ff with F on one side of it and and all of C on the other. Proof: See for example Trustrum [18]. 2 Our use in Lem. 3.6 is based on interpreting distribution F as a point in R N , and an expectation ff as the collective normal of a family of parallel hyperplanes. The integral R F ff then gives the constant term for the ff-hyperplane that passes through the point F . More generally, for a convex region C in R N , the minimum Z F gives the constant term for the ff-hyperplane that touches C, with its normal pointing into C. Thus when specialised for the applications in this paper, the lemma implies that if F 62 C for some closed convex set of distributions C, then there is an expectation ff with Z F Z Moreover if C is up-closed (down-closed) then the range of ff above specialises to PS (NS). Facts and definitions from domain theory We summarise here Abramsky and Jung's presentation [1] of facts from domain theory, giving page numbers where appropriate. Assume (D; v) is a complete partial order - i.e. that every directed (defined below) set has a least upper bound. 1. up-, down-closure For subset A of D we define its up-closure "A to be the set Similarly we define its down-closure #A to be the set 2. Smyth order (p.97) The Smyth order v S on PD, for subsets A; A 0 of D, is given by A 3. Hoare order (p.97) The Hoare order vH on PD, for subsets A; A 0 of D, is given by A vH A 0 iff A ' #A 4. Egli-Milner order The Egli-Milner order vEM between subsets combines the Smyth (Def. B2) and Hoare (Def. B3) order. For subsets of D we define A vEM A 0 iff "A ' A 0 and A ' #A 5. up-directed (Def. 2.1.8, p.10) A subset A of D is up-directed (or simply directed) iff for any u; v in A there is a w also in A such that for the least upper bound of A (if it exists). 6. down-directed A subset A of D is down-directed iff for any u; v in A there is a w also in A such that w v u and w v v. We write uA for the greatest lower bound of A (if it exists). 7. way-below (Def. 2.2.1, p.15) The way-below relation - on D is defined as follows: for u; v in D we say u up-directed subsets A of D with v v tA there is some w in A with u v w. We also say u approximates v iff u - v. 8. basis (Def. 2.2.3, p.16) A basis for D is a subset B such that every element of D is the t-limit of the elements way below it in B. 9. !-continuity (Def. 2.2.6, p.17) D is !-continuous if it has a countable basis. 10. lens (Def. 6.2.15, p.100) We define the lenses of D, Lens(D), to be the set of non-empty, Egli-Milner closed, limit-closed subsets of D. 11. ideal (p.10) A subset I is an ideal if it is directed and down-closed. 12. Plotkin powerdomain (Theorem 6.2.3, p.95) The Plotkin powerdo- main of D with basis B is given by the ideal (Def. B11) completion of where F(B) denotes the set of finite, nonempty subsets of B. 13. topological Egli-Milner order (Def. 6.2.16, p.101) Define the topological Egli-Milner order v TEM on the set of Egli-Milner closed subsets, Lens(D), of D as follows: 14. Smyth-, Hoare-equivalence Two subsets A; A 0 of D are Smyth equivalent are Hoare-equivalent if Theorem B.1 (Theorem 6.2.19, p.101) If D is an !-continuous complete partial order, its Plotkin powerdomain is isomorphic to (Lens(D); v TEM ). 2 --R Domain theory. A Discipline of Programming. Probabilistic models for the guarded command language. Probabilistic nondeterminism. A probabilistic powerdomain of evaluations. Private communication. Semantics of probabilistic programs. A probabilistic PDL. Proof rules for probabilistic loops. Unifying wp and wlp. Probabilistic predicate transformers. A generalization of Dijkstra's calculus. "Pisa Notes" An introduction to probabilistic predicate transformers. Linear Programming. --TR Parallel program design: a foundation A generalization of Dijkstra''s calculus Probabilistic non-determinism Building on the unity experience On randomization in sequential and distributed algorithms Domain theory and integration Domain theory Probabilistic predicate transformers PCF extended with real numbers Modeling and verification of randomized distributed real-time systems Unifying <italic>wp</italic> and <italic>wlp</italic> Probabilistic models for the guarded command language A Discipline of Programming A probabilistic PDL --CTR Joe Hurd , Annabelle McIver , Carroll Morgan, Probabilistic guarded commands mechanized in HOL, Theoretical Computer Science, v.346 n.1, p.96-112, 23 November 2005 Yuxin Deng , Rob van Glabbeek , Matthew Hennessy , Carroll Morgan , Chenyi Zhang, Remarks on Testing Probabilistic Processes, Electronic Notes in Theoretical Computer Science (ENTCS), 172, p.359-397, April, 2007

program logic;partial correctness;probability;verification

504618

Loss probability calculations and asymptotic analysis for finite buffer multiplexers.

In this paper, we propose an approximation for the loss probability, PL (x), in a finite buffer system with buffer size x. Our study is motivated by the case of a high-speed network where a large number of sources are expected to be multiplexed. Hence, by appealing to Central Limit Theorem type of arguments, we model the input process as a general Gaussian process. Our result is obtained by making a simple mapping from the tail probability in an infinite buffer system to the loss probability in a finite buffer system. We also provide a strong asymptotic relationship between our approximation and the actual loss probability for a fairly large class of Gaussian input processes. We derive some interesting asymptotic properties of our approximation and illustrate its effectiveness via a detailed numerical investigation.

Introduction loss probability is an important QoS (Quality of Service) measure in communication networks. While the overflow probability, or the tail of the queue length dis- tribution, in an infinite bu#er system has been extensively studied [1], [2], [3], [4], [5], [6], [7], there have been relatively few studies on the loss probability in finite bu#er systems [8], [9], [10], [11]. In this paper, we propose a simple method to estimate the loss probability PL (x) in a finite bu#er system from the tail of the queue length distribution (or tail probabil- of an infinite bu#er system. We estimate PL (x) by making a simple mapping from P{Q > x}. Hence, we consider both a finite bu#er queueing system and an infinite bu#er queueing system. We model both systems by a discrete-time fluid queue consisting of a server with constant rate c and a fluid input #n . Both queues are fed with the same input. Let - Qn and Qn denote the queue length in the finite queue and in the infinite queue at time n, re- spectively. We assume that #n is stationary and ergodic and that the system is stable, i.e., E{#n } < c. Under this assumption, it has been shown that Qn converges to a stationary and ergodic process [12]. It has also been shown that - Qn converges to a stationary process when the system is a GI/GI/m/x type of queue [13], [14], and when the system is a G/M/m/x type of queue [15]. Since proving the convergence of - Qn is not the focus of this paper, and more- over, practical measurements of PL (x) and P{Q > x} are based on "time averaging" assuming ergodicity (see (1) and (2)), we assume that both - Qn and Qn started at and that they are ergodic and stationary. 1 The time index H.S.Kim and N.B.Shro# are with the School of Electrical and Computer Engineering, Purdue University, West Lafayette, Indiana We refer the interested reader to our technical report [17], where we have studied the relationship between finite and infinite bu#er queues without assuming ergodicity of - Qn and derived similar asymp- n is often omitted to represent the stationary distribution, The loss probability, PL (x), for a bu#er size x is defined as the long-term ratio of the amount of fluid lost to the amount of fluid fed. It is expressed as and where the second equality is due to the ergodicity assumption. The tail probability (or tail of the queue length distribution, also sometimes called the overflow probability) P{Q > x} is defined as the amount of time the fluid in the infinite bu#er system spends above level x divided by the total time. It is expressed as: now on, when we write "loss probability" it will only be in the context of a finite bu#er system, and when we write "tail probability" it will only be in the context of an infinite bu#er system. Note that since P{Q > x} is averaged by time, and PL (x) is averaged by the input, in general there is no relationship between these two quantities. However, PL (x) is often approximated as: This approximation usually provides an upper bound (sometimes a very poor bound) to the loss probability, although in general this cannot be proven, and in fact counter-examples can easily be constructed. What we have learned from simulation studies is that the curves PL (x) versus x and P{Q > x} versus x exhibit a similar shape (e.g., see Fig. 1), which motivates this work. Further, it has been shown in [16] that for M/Subexponential/1 and GI/Regularly-varying/1 with i.i.d. interarrival times and i.i.d. service times, P{Q > x}/PL (x) converges to a con- stant, as x #. Hence, it seems reasonable that if we have a good estimate of the tail probability P{Q > x} and a way to calculate PL (a), the loss probability for some bu#er size a, then totic results to Equation (22) in this paper. However, this involves mathematical technicalities that take away from the main message in this paper, i.e., developing a simple approximation for the loss probability. we can calculate the loss probability PL (x) as PL (a) In particular, we will choose a = 0 because this allows us to compute the loss probability (PL (0)) quite easily. This is the basic idea that drives this paper. In addition to developing a methodology to calculate the loss probability, we will also show that asymptotically the loss probability and the tail probability curves are quite similar, and if they diverge, they do so slowly, which is an interesting result by itself. For our study in this paper, we focus on the case when the aggregate tra#c can be characterized by a stationary Gaussian process. Recently, Gaussian processes have received significant attention as good models for the arrival process to a high-speed multiplexer [3], [18], [19], [20], [21], [22], [23]. There are many reasons for this. Due to the huge link capacity of high-speed networks, hundreds or even thousands of network applications are likely to be served by a network multiplexer. Also, when a large number of sources are multiplexed, characterizing the input process with traditional Markovian models results in computational infeasibility problems [24] that are not encountered for Gaussian processes. Finally, recent network tra#c studies suggest that certain types of network traffic may exhibit self-similar or more generally asymptotic self-similar type of long-range dependence [25], [26], and various Gaussian processes can be used to model this type of behavior. Hence, our motivation to study the case when the input process #n can be characterized by a Gaussian process. This paper is organized as follows. In Section II, we review the maximum variance asymptotic (MVA) results for the infinite bu#er queue, and then demonstrate how to obtain similar results for the loss probability. Then, we compare our approach to an approach based on the many-sources asymptotics. In Section III, we validate our result with several numerical examples, including those for self-similar/long-range dependent tra#c. In Section IV, we find the asymptotic relationship between the loss probability and our approximation. In Section V, we describe the applicability of our approximation for on-line tra#c mea- surements. We finally state the conclusions in Section VI. II. Maximum Variance Asymptotic (MVA) Approximation for Loss Remember that the first component in our development of an approximation for PL (x) is to find a good estimate of P{Q > x}. Fortunately, this part of the problem has already been solved in [20], [21], [27]. By developing results based on Extreme Value Theory, it has been found that the Maximum Variance Asymptotic (MVA) approach (first named in [20]) provides an accurate estimate of the tail probability. We briefly review it here. As mentioned before, we focus on the case when the aggregate tra#c can be characterized by a Gaussian process, hence #n , the input process to the queue is Gaussian. Let - The queue length Qn (or workload) at time n in the infinite bu#er system is expressed by Lindley's equation: We define a stochastic process Xn as We assume that #n is stationary and ergodic and that the system is stable, i.e., E{#n } < c. Then, it has been shown that the distribution of Qn converges to the steady state distribution as n # and that the supremum distribution of Xn is the steady state queue distribution [12]: Let C # (l) be the autocovariance function of #n . Then, the variance of Xn can be expressed in terms of C # (l). For each x > 0, define the normalized variance # 2 x,n of Xn as x,n := where # := c- #. Let m x be the reciprocal of the maximum of # 2 x,n for given x, i.e., x,n and we define n x to be the time n at which the normalized variance Although the estimate 2 called the Maximum Variance Asymptotic (MVA) approximation has been theoretically shown to be only an asymptotic upper bound, simulation studies in di#erent papers have shown that it is an accurate approximation even for small values of x [27], [18], [20], [28]. Now, for some a, we need to evaluate the ratio PL (a)/P{Q > a} given in (4). As mentioned earlier, it is easy to find PL (a) for a = 0, hence what we need to do is to first estimate P{Q > 0} from the MVA result. For a given x both n x and m x in the MVA approximation cannot generally be obtained in a simple closed form, hence search algorithms 2 are likely to be used to evaluate them. may not be unique especially for a small value of x. However, when 0, we can obtain them right away as demonstrated in the following proposition. Proposition 1: Let n x be the value of n at which # 2 x,n attains its maximum # 2 . (10) Simple local search algorithms starting at #x (2-# are good enough to find nx within a small number of iterations. Proof of Proposition 1: To prove the proposition, it su#ces to show that sup . (11) . (12) , we have (11). Now, we show how to calculate PL (0). Since #n is assumed Gaussian, the mean and the variance provide su#- cient information to calculate PL (0), i.e., c As long as the number of input sources is large enough for the aggregate tra#c to be characterized as a Gaussian process, (13) gives an accurate estimate (ex- act for a Gaussian input) and is often called the Gaussian approximation [29]. Note that C # in (10). From (4), (10), and (13), we have where exp (c - #) 2 c We call this above approximation the MVA approximation for loss. The MVA approach is based on the large bu#er asymptotics and it also applies in the context of the many- sources asymptotics [20], [28]. We next compare this approach with an approximation based on the many-sources asymptotics. The many-sources asymptotics have been widely studied and can be found in many papers on queueing analysis using large-deviation technique [5], [30], [31], [32]. Most of the papers deal with the tail distribution rather than the loss probability. In [9], the authors developed the first result on the loss probability based on the many-sources asymptotics. We call this the Likhanov-Mazumdar (L-M) approximation for loss. Since the L-M result was obtained for a fairly general class of arrival processes and is much stronger than typical large-deviation types of results, we feel that it is important to compare our result with the result. Consider N i.i.d. sources, each with input rate # (i) It is assumed that the moment generating function of # (1) exists, and that the input rate n is bounded. The L-M approximation has the following and is theoretically justified by where N is the number of sources, NC is the link ca- pacity, NB is the bu#er size, - # is a value of # such that # #n (#) log #n (# n ), and - n is a value of n that maximizes I n (C, B), for a given C and B. becomes exact as N #. Consider the numerical complexity of (16). Suppose that we calculate (16) for given N,C,B, and # (1) n . In general, since there are no closed-form solutions for # n and - n, we have to find them numerically. Two iteration loops are nested; The inner loop iterates over # to find # n for given n, and the outer loop iterates over n to find - n. Hence, it can take a long time to find a solution of (16) by numerical iteration. However, the MVA approximation requires only an one-dimensional iteration over n to find n x at which m x is minimized. There is another problem in applying the L-M approximation for control based on on-line measurements. When the distribution of a source is not known beforehand, in the L-M approach the moment generating function of a source should be evaluated for the two dimensional arguments, (#, n), whereas only the first two moments are evaluated for the one argument, n, in the MVA approach (see Section V). Note that one could avoid the above problems by making a Gaussian approximation on the aggregate source first, and then using the the L-M approximation given by (16). Specifically, if we assume that the input process is Gaussian, we have a closed-form solution for # n , i.e., as k } and k }, we have Cn +B -m(n) . (17) Hence for given C and B both I n (C, B) and # 2 NB,n (the normalized variance of Xn ) are expressed in terms of n, m(n), and v(n), we can avoid the two dimensional evaluation of the moment generating function. 3 This expression is just a rewriting of equation (2.6) in [9]. The only problem is that the theoretical result that says that the L-M approximation in (16) becomes exact as the number of sources N becomes large, is not proven for unbounded (e.g. Gaussian) inputs. Still, since making this approximation reduces the complexity of the search space, it would be instructive to also investigate the performance of such an approximation. In Section III, we will numerically investigate our MVA approximation for loss, the L-M approximation, and some other approximations developed in the literature. III. Numerical Validation of the MVA Approximation for Loss In this section, we investigate the accuracy of the proposed method by comparing our technique with simulation results. In all our simulations we have obtained 95% confidence intervals. However, to not clutter the figures, the error bars are only shown in the figures when they are larger than -20% of the estimated probability. To improve the reliability of the simulation, we use Importance Sampling [33] whenever applicable. 4 We have attempted to systematically study the MVA approximation for various representative scenarios. For example, we begin our investigation with Gaussian input processes. Here, we only check the performance of our approximation (not compare with other approximations in the literature), since other approximations are not developed for Gaussian inputs. We then consider non-Gaussian input sources and compare our MVA approximation for loss with other approximations in the literature. Specifically, we consider MMF sources which have been used as representative of voice tra#c in many different papers (e.g. [34], [35]) and also consider JPEG and MPEG video sources that have been used in other papers in the literature (e.g. [20], [36]). A. Gaussian Processes We begin by considering the simple case when the input is a Gaussian Autoregressive (AR) process with autocovariance (note that AR processes have been used to model VBR video [22]). In Fig. 2 one can see that the simulation and MVA-Loss result in a close match over the entire range of bu#ers tested. The next example, in Fig. 3, covers a scenario of multi- time scale correlated tra#c. Note that multiple-time scale correlated tra#c is expected to be generated in high-speed networks because of the superposition of di#erent types of sources [37]. In this case, the autocovariance function of the Gaussian input process is the weighted sum of three di#erent powers, i.e., C # . One can see from Fig. 3 that because of the multi-time scale correlated nature of the input, the loss probability converges to its asymptotic decay rate only at large bu#er sizes. This observation is consistent with observations made on the tail probability when fed with multi-time scale correlated tra#c [20]. Again, it can be 4 For interesting readers, the software used for the analysis and simulation will be available upon request. seen that the analytical result tracks the simulation results quite closely. The next example, deals with a well known input pro- cess, the fractional Brownian motion process, which is the classical example of a self-similar process [23]. 5 The results are shown in Figs. 4 and 5, demonstrating the accuracy of MVA-Loss, even for self-similar sources. Due to the di#- culty in applying importance sampling techniques to obtain loss probabilities for self-similar tra#c, in Figs. 4 and 5, we show probabilities only as low as 10 -6 . In Fig. 4, the input tra#c is characterized by a single Hurst parameter. How- ever, even if the tra#c itself is long-range dependent, due to the heterogeneity of sources that high-speed networks will carry, we expect that it will be di#cult to characterize the tra#c by simply one parameter, such as the Hurst param- eter. Hence, we also run an experiment for a more realistic scenario, i.e., the input process being the superposition of fractional Brownian motion processes with di#erent Hurst parameters. The numerical result is shown in Fig. 5. One can see from Figs. 4 and 5 that MVA-Loss works well for self-similar sources. B. Non-Gaussian Processes In this section we will compare the performance our MVA-loss approximation with simulations and also with other schemes in the literature. We call the Likhanov- Mazumdar technique described earlier "L-M," or "L- M:Gaussian" when further approximated by a Gaussian process, the Cherno# dominated eigenvalue technique in [38] "Cherno#-DE," the average/peak rate method in [39] "Ave/Peak," the analytical technique developed in [24] "Hybrid," and the famous e#ective bandwidth scheme "Ef- fective BW" [40]. We now consider the practically important case of multiplexed voice sources. The input MMF process, which has widely been used to model voice tra#c source [34], [35], has the following state transition matrix and rate vector: Input rate vector : 0 cells/slot cells/slot These values are chosen for a 45 Mbps ATM link with 10 time slot and 53 byte ATM cell. In this example, we assume that 2900 voice sources are multiplexed on a Mbps ATM link with 10 msec time slot and 53 byte ATM cell. As shown in Fig. 6, the MVA-Loss obtains the loss probability calculations accurately and better than the other techniques. 5 For computer simulations, since continuous-time Gaussian processes cannot be simulated, one typically uses a discrete-time version. In the case of fractional Brownian Motion, the discrete-time version is called fractional Gaussian noise and has autocovariance function given by: is the Hurst parameter. We next investigate the accuracy of our approximation when the sources to the queue are generated from actual MPEG video traces. The trace used to generate this simulation result comes from an MPEG-encoded action movie (007 series) which has been found to exhibit long-range dependence [36]. In Fig. 7, 240 MPEG sources are multiplexed and served at 3667 cells/slot (OC - 3 line), where we assume 25 frames/sec and a 10 msec slot size. The loss probability versus bu#er size result in this case is shown in Fig. 7. Again, it can be seen that the MVA-Loss approximation tracks the simulation results quite closely. C. Application to Admission Control The final numerical result is to demonstrate the utility of MVA-Loss as a tool for admission control. We assume that a new flow is admitted to a multiplexer with bu#er size x if the loss probability is less than the maximum tolerable loss probability #. In this example, we consider multiplexed voice sources on a 45Mbps link (Fig. 8(a)) or multiplexed video sources (Fig. 8(b)) for an admission control type of application. The QoS parameter # is set to 10 -6 . For each voice source in Fig. 8(a), we use the same MMF On-O# process that was used for Fig. 6. For each video source, we use the same MPEG trace that was used in Fig. 7 (with start times randomly shifted). Then, the admission policy using MVA- Loss is the following. Let - # and v(n) be the mean and the variance function of a single source, i.e., Let - # := E{# (1) and v(n) := Var{ are currently serviced, a new source is admitted if where # is defined as in (14). In Fig. 8(a) and (b), we provide a comparison of admissible regions using di#er- ent methods. It can be seen that MVA-Loss curve most closely approximates the simulation curve in both figures. In Fig. 8(a), the L-M approximation performs as well, and the Cherno# DE approximation only does slightly worse. In Fig. 8(b), however, the Cherno# DE approximation in this case is found to be quite conservative. This is because for sources that are correlated at multiple time-scales (such as the MPEG-video sources in Fig. 8(b) shown here), the loss probability does not converge to its asymptotic decay rate quickly (even if there exists an asymptotic decay rate), and hence approximations such as the Cherno# DE scheme (or the hybrid scheme shown earlier) perform quite poorly. Admission control by MVA-Loss can be extended to a case where heterogeneous flows are multiplexed. The link capacity is 622.02Mbps (OC - 12 line), the bu#er size x is fixed to 20000 cells, and the QoS parameter # is 10 -6 . In this system, the input sources are of two types; JPEG- video and voice. As a video source, we use a generic model that captures the multiple-time scale correlation observed in JPEG video traces. It is a superposition of an i.i.d. Gaussian process and 3 two-state MMF processes: State transition matrices : 0.999 0.001 0.9999 0.0001 Input rate vectors [cells/slot] : Mean of i.i.d. Gaussian : 82.42 Variance of i.i.d. Gaussian : 8.6336 Then, the admission policy is the following. Let - be the mean and the variance function of a single voice source. Let - # 2 and v 2 (n) be the mean and the variance function of a single video source. When (N 1 -1) voice and N 2 video flows are currently serviced, a new voice flow is admitted if The boundary of the admissible region is obtained by finding maximal N 1 satisfying (19) for each N 2 . As one can see in Fig. 9, the admissible region estimated by simulations and via MVA-Loss is virtually indistinguish- able. In fact, the di#erence between the two curves is less than 1% in terms of utilization. IV. Asymptotic Properties of the MVA Approximation for Loss We now find a strong asymptotic relationship between the loss probability and the tail probability. More specifi- cally, under some conditions (to be defined later in Theorem 5), we find that log means that lim sup |f/g| < #. Equation tells us that the divergence between the approximation #e -mx/2 given in (14), and the loss probability is slow if at all (this may be easier to see if we rewrite (20) in the form log PL (x) - log #e In [27] and [28], under a set of general conditions it has been shown for the continuous-time case that log P{Q > x} We will obtain (20) by finding a relationship between PL (x) and P{Q > x}, i.e., log P{Q > x} - log PL under the set of conditions given in Theorem 5 (PL (x) will be bounded from above and below by some expressions in terms of P{Q > x}), and then by applying (21) and some properties of m x . Note, that finding the asymptotic relationship (22) between P{Q > x} and PL (x) is by itself a valuable and new contribution. We first list a set of conditions for which (21) holds in the discrete-time case that are equivalent to the set of conditions in [27] defined for the continuous-time case. Let v n :=Var{Xn}, #(n) := log v n , and # := lim n#(n)/ log n (assuming that the limit exists). The notation f(n) n# # g(n) means that lim n# 1. The parameter # cannot be larger than 2 due to the stationarity of #n , and # (0, 2) covers the majority of non-trivial stationary Gaussian processes. The Hurst parameter H is related to # by We now state the following results that are the discrete-time versions of the results in [27], [28], [41]. The proofs for these results are identical to those given in [27], [28], [41], with trivial modifications accounting for the discrete-time version, and, hence, we omit them here. These results are stated as Lemmas here, since we will be using them to prove our main theorem. Lemma 2: Under hypotheses (H1) and (H2), x# x 2-# . Lemma 3: Under hypotheses (H1) and (H2), log P{Q > x} It is easier for us to work with conditions on the autocovariance function of the input process rather than conditions (H1) and (H2). Hence, we first define a condition on the autocovariance function C # (l) which guarantees (H1) and l=-n # S#n #-1 . Note that condition (C1) is quite general and is satisfied not only by short-range dependent processes but also by a large class of long-range dependent processes including second-order self-similar and asymptotic self-similar processes [42]. Lemma 4: If the autocovariance function C # (l) of #n satisfies (C1), then (H1) and (H2) hold. Proof of Lemma 4: Let h(n) := (l). Note that and that v n+1 - v First we show condition (H2). Since both v n and n # approach #, lim n# vn should be equal to lim n# vn+1-vn , if it exists (this is the discrete version of L'Hospital's rule). Hence, lim where lim n# vn we show that (H1) also follows from (C1). Since h(n) that Note that a function g(x) is o(x) if lim x# g(x)/x # 0. Now, vn vn (by Taylor Expansion) #S -S The loss probability is closely related to the shape of the sample path, or how long Qn stays in the overflow state. Before we give an illustrative example, we provide some notation. We define a cycle as this period, i.e., an interval between time instants when Qn becomes zero. We let S x denote the duration for which Qn stays above threshold x in a cycle to which n belongs. Formally, let: . Un := sup{k time of the current cycle to which n belongs). . Vn := inf{k time of the next cycle). . Wn := Vn -Un (Duration of a cycle to which n belongs). . Zn := Vn - n (Residual time to reach the end of cycle). . S x k=Un (Duration for which Q k > x in a cycle containing n). Note that if Qn > 0, Zn is equal to the elapsed time to return to the empty-bu#er (or zero) state. Since Qn is stationary and ergodic, so are the above. Hence, their expectations are equal to time averages. Consider two systems whose sample paths look like those in Fig. 10. The sample paths are obtained when the input is a deterministic three-state source which generates fluid at rate c and 0, at state 1, 2, and 3, respectively. The duration of each state is the same, say b. Use the superscript (1) and (2) to represent values for the upper and the lower sample path. Set a 2b (1) . Then, both cases have the same overflow probability. Now, consider a time interval from 0 to 3b (2) . The amount of fluid generated for that interval is clearly the same for both cases. But, the amount of loss in the upper case is exactly the twice of that in the lower case, hence, the upper case has the larger loss probability. We can infer from this that the loss probability is closely related to the length of n and the slope of the sample path. Since loss happens only when Qn is greater than the bu#er size x, we consider the condition that Qn > x. Since it is di#cult to know the distribution of S x n , and since S x n is determined by the sample path, we use a stochastic process defined as Here we have chosen 0 as the origin, but, because of sta- tionarity, the distribution of Yn does not depend on the origin. Note that if Q 0 > 0, Yn will be identical to Qn till the end of cycle. We want to know the distribution of Yn given is Gaussian, the distribution of Yn can be characterized by the mean and the variance of Yn . However, since Q 0 is the result of the entire history up to time 0 and the future is correlated with the past, it is di#cult to find an explicit expression of the mean and the variance of Yn given Q 0 > x. Hence, we introduce upper-bound types of conditions on the mean and the variance of Yn as (26) and (27). For notational simplicity, let be the expectation and the variance under P x , respectively. We now state our main theorem. Theorem 5: Assume condition (C1). Further assume that for any # > 0, there exist x 0 , K,M, and # such that for all x # x 0 and n # Mx # . Then, -# < lim inf Though the conditions of Theorem 5 look somewhat complex, they are expected to be satisfied by a large class of Gaussian processes. If the input process is i.i.d. with it can be easily checked that and (C1), (26), and (27) are satisfied with # , and 1. It has been shown that Gaussian processes represented by the form of finite-ordered Autoregressive Moving Average (ARMA) satisfy (26) and (27) [17]. Since the autocovariance function of a stable ARMA process is in the form of C # it satisfies (C1) with 1. So Theorem 5 is applicable to Gaussian ARMA processes. More generally, E{ P n # Sn # under (C1). Thus, for each x, #n and Var x {Yn } n# and we can find #, x) as small as possible. If sup x K(#, x), sup x M(#, x), and sup x #, x) are finite, then (26) and (27) hold. We conjecture that they are all finite for a large class of stationary Gaussian processes, and we are trying to show it. Note that the rightmost inequality (limsup part) in (28) holds without conditions (26) and (27), and it agrees with empirical observations that the tail probability curve provides an upper bound to the loss probability curve. Before we prove the theorem, we first define the derivative of m x with respect to x, m # x . Recall (9), or m vnx . Since n x is an integer value, m x is di#erentiable except for countably many x at which n x has a jump. Let x is not di#erentiable}. Note that D has measure zero, and that the left and right limits of m # x exist for all x # D. For simplicity, abuse notations by setting z and m # z for x # D. The reason we set the (right) limit is that we will find the similarity relation (29) in Lemma 6, which is useful in proving Theorem 5. In fact, we may take the left limit to have the same asymptotic behavior. By building m # x in this way, it directly follows from Lemma 2 that m # x # bx -# for some constants a > 0 and b. We now state three lemmas which are useful in proving the theorem (Their proofs are in Appendix). Lemma Under hypotheses (H1) and (H2), x y dy x# 2x K x where K is a constant. Lemma 7: If P{Q > x} > 0 and E{Z|Q > x} < # for all x, #PL (x). (30) Lemma 8: Under conditions (26) and (27), E{Z|Q > Now, we are ready to prove Theorem 5. Proof of Theorem 5: First of all, we find expressions in terms of P{Q > x} which are greater than or less than - #PL (x). If P{Q > would contradict the asymptotic relation in Lemma 3. Hence, P{Q > x} > 0 for all x. If E{Z|Q > would contradict the asymptotic relation in Lemma 8. Hence, E{Z|Q > x} < # for all x. Thus, by Lemma 7 we have (30). Now, since x By Lemma 4, (C1) implies (H1) and (H2). Hence, by Lemma 3, we have (21). Equation (21) means that there are x 0 , K 1 and K 2 such that Note that since E{Z|Q > we can choose K 3 > 0 such that E{Z|Q > x} # K 3 x # for all x # x 0 . Combining with (30) and (31), integrate all sides of (32) to get dy #PL (x) x with the constant a > 0, by Lemma 6, there exist x 1 # x 0 , K 4 > 0 and K 5 > 0 such that dy, and x From (33), (34) and (35), Take logs and rearrange to get log K 4 Divide by log x and take x #. Then, the theorem follows V. Applications to On-line Measurements In this section, we describe how to apply the MVA approach for the estimation of the loss probability, based on on-line measurements. In many practical situations, the characteristics of a flow may not be known beforehand or represented by a simple set of parameters. Hence, when we use a tool for the estimation of the loss probability, parameter values such as the moment generating function and the variance function should be evaluated from on-line measurements. Then, the question is what range of those parameters should be evaluated. If an estimation tool needs, for example, the evaluation of the moment generating function for the entire range of (#, n), the tool may not be useful. This is fortunately not the case for the MVA approximation for loss. Note that the MVA result has the form #e -mx/2 . The parameter m x is a function of c, - #, x, and v(n), where - and v(n) is the mean and the variance of the input, i.e., Hence, by measuring only the first two moments of the input we can estimate the loss probability. Recall that m that v(n) (c- is maximized at This means that the result only depends on the value of v(n) at This value of n x corresponds to the most likely time-scale over which loss occurs. This is called the dominant time scale (DTS) in the literature [43], [20]. Thus, the DTS provides us with a window over which to measure the variance func- tion. It appears at first, however, that this approach may not work because the DTS requires taking the maximum of the normalized variance over all n, which means that we would need to know v(n) for all n beforehand. Thus, we are faced with a chicken and an egg type of problem, i.e., which should we do first: measuring the variance function v(n) of the input, or estimating the measurement window n x . Fortunately, this type of cycle has recently been broken and a bound on the DTS can in fact be found through on-line measurements (see Theorem 1 and the algorithm in [44]). Thus, since our approximation is dependent on the DTS, we only need to estimate v(n), for values of n up to a bound on the DTS (given in [44]), thereby making it amenable for on-line measurements. VI. Concluding Remarks We have proposed an approximation for the loss probability in a finite queue by making a simple mapping from the MVA estimate of the tail probability in the corresponding infinite queue. We show first via simulation results that our approximation is accurate for di#erent input processes and a variety of bu#er sizes and utilization. Since the loss probability is an important QoS measure of network tra#c, this approximation will be useful in admission control and network design. Another feature of the approximation is that it is given in a single equation format and hence can easily be implemented in real-time. We have compared our approximation to existing methods including the effective bandwidth approximation, the Cherno# dominant eigenvalue approximation, and the many-sources asymptotic approximation of Likhanov and Mazumdar. In this paper we also study the theoretical aspects of our approximation. In particular, we provide a strong asymptotic result that relates our approximation to the actual loss probability. We show that if our approximation were to diverge (with increasing bu#er size) from the loss prob- ability, it would do so slowly. For future work we plan on simplifying the conditions given in Theorem 5 and to extend the approximation result to a network of queues. VII. Appendix Proof of Lemma x x# x /2. Hence, to prove the lemma, it su#ces to show that x e -f(y) dy x# e -f(x) . (36) x is not di#erentiable}. For x # D, d dy e -f(y) y=x e -f(x) . (37) Since D has measure zero, R [x,#)-D (-)dy and we may assign any values to f # (x) and f # (x) for all x # D. Recall z and m # z for x # D. Set let x be any value. Integrating both sides of (37) from x to #, we have e -f(x) x e -f(y) dy - Z # x e -f(y) dy. (38) Note that m # a > 0 and b. Since f # x # (0, 1). We can find x 0 such that f # (x) x e -f(y) dy # Z # x e -f(y) dy x e -f(y) dy x which means that1 x e -f(y) dy and the result follows. Proof of Lemma 7: Recall the notations: . Un := sup{k time of the current cycle to which n belongs). . Vn := inf{k time of the next cycle). . Wn := Vn -Un (Duration of a cycle to which n belongs). . Zn := Vn - n (Residual time to reach the end of cycle). . S x k=Un (Duration for which Q k > x in a cycle containing n). one more: . R x k=n 1 {Qk>x} . (Residual duration for which x in a cycle containing n) Since Qn is stationary and ergodic, so are the above. Hence, their expectations are equal to time averages. Since we are interested in the behavior of Qn after loss happens, we consider the conditional expectations: R x Clearly, E{R x And it can also be easily checked that 2E{R x where the inequality is due to that n is discrete. 6 Since < c, there are infinitely many cycles for a sample path. Index cycles in the following manner: . A (i) . S (i) . - x Now, we prove the lemma in two steps: . 1) Derive x- . The amount of loss in cycle i is greater than or equal to the di#erence between the maximum value of the queue level Qn in cycle i and the bu#er size x of the finite bu#er queue, i.e., k#A (i) x k#A (i) x I(S (i) y > 0)dy. Take summation over i and divide by the total time, denotes the number of elements 6 Since n is discrete, for given n such that Qn > x, R x n and S x (positive) integer values. If Sn is, for example 2, Rn can be either 1 or 2, and its expectation is 1.5 which is greater than 2/2. of A (i) . Then, x I(S (i) |A (i) | dy x y y dy x sup l#m y y dy. Recalling (1) and (2), -# PL (x), - #, y sup l#m y as m #. Since all components are nonnegative, by Fatou's Lemma, (44) becomes x- Step 2) For better understanding, we first show lim sup m#m x Note that all components are nonnegative. Let am :=2m+1 P m lim sup am , and b . For any # > 0, we can choose M such that aM - a # and |b # - b M | < #. Then, since x x x Since # is arbitrary, we have b # a # . Now, we will verify that x Construct a new sequence {T (i) x } by removing zero-valued elements of {S (i) x }. Then, as in 45, lim sup m#m x Note that lim sup m#m x x x for all x and |B (i) x , xk x x | x x x x Combining (47), (48) and (49), we have (46). At last, we have from which (30) follows. Proof of Lemma 8: . #(x, n) := P x {Yn > 0}, . V The proof will be done in two steps: . 1) Find x 1 > x 0 such that #(x, n) # n -2 for all (x, n) # . 2) Using 1), show that E{Z|Q 0 > x} is O(x # ). # be so small that -# < 0. Then, we choose x 0 , M , and # satisfying (26) and (27). Let m(n) := Then, the moment generating function of Gaussian Yn is given by e #m(n)+ 1 From (26) and (27), m(n) #)n and v(n) # Kn # for all (x, n) # V where #. Let for all (x, n) # V Kn 2#-(2-#) . (52) Note that # > 2# - (2 - #). Since the coe#cient of the leading term, -# , is negative and its order, #, is positive, we have for all (x, n) # V Kn 2#-(2-# 0. (53) Note that in (53) # , K, and # are fixed constants for all there exists x 2 such that Now, we choose x 1 > x 0 such that Mx # Step From the definition of Z, Z > n implies Yn > 0. Thus, Therefore, we have or Obviously as shown in Step 3), #(x, n) # n -2 for all n # Mx # . Applying this and (55), #(x, n) #(x, n) x# where #x# denotes the smallest integer which is greater than or equal to x. Since E x {Z} is nonnegative, E x --R "Stochastic Theory of a Data Handling System with Multiple Sources," "Asymptotics for steady-state tail probabilities in structured Markov queueing models," "An Approximation for Performance Evaluation of Stationary Single Server Queues," "Stability, Queue Length, and Delay of Deterministic and Stochastic Queueing Networks," "Large deviations and overflow probabilities for the general single server queue, with appli- cation," "Squeezing the Most Out of ATM," "Logarithmic asymptotics for steady-state tail probabilities in a single-server queue," "Loss Performance Analysis of an ATM Multiplexer loaded with High Speed ON-OFF Sources," "Cell-loss asymptotics in bu#ers fed with a large number of independent stationary sources," "Improved Loss Calculations at an ATM Multiplexer," "Investigation of Cell Scale and Burst Scale E#ects on the Cell Loss Probability using Large Deviations," "The Stability of a Queue with Non-independent Inter-arrival and Service Times," The Single Server Queue "Limits for Queues as the Waiting Room Grows," Stationary Stochastic Mod- els "A fluid queue with a finite bu#er and subexponential input," "On the Asymptotic Relationship between the Overflow Probability in an Infinite Queue and the Loss Probability in a Finite Queue," "A Central Limit Theorem Based Approach to Analyze Queue Behavior in ATM Networks," "A New Method to Determine the Queue Length Distribution at an ATM Multiplexer," "A Central Limit Theorem Based Approach for Analyzing Queue Behavior in High-Speed Networks," "On the supremum distribution of integrated stationary Gaussian processes with negative linear drift," "Performance Models of Statistical Multiplexing in Packet Video Communication," "On the Use of Fractal Brownian Motion in the Theory of Connectionless Networks," "Improved Loss Calculations at an ATM Multiplexer," "Long- range dependence in variable-bit-rate video tra#c," "On the Self-Similar Nature of Ethernet Tra#c (Extended Version)," "Queueing Analysis of High-Speed Multiplexers including Long-Range Dependent Arrival Processes," "Use of Supremum Distribution of Gaussian Processes in Queueing Analysis with Long-Range Dependence and Self-Similarity," "Multiplexing gains in bit stream mul- tiplexors," "Large deviations, the shape of the loss curve, and economies of scale in large multiplexers," "Economies of scale in queues with sources having power-law large deviation scaling," "Large deviations approximation for fluid queues fed by a large number of on-o# sources," "E#ective bandwidth and fast simulation of ATM intree net- works," "Models for Analysis of Packet Voice Communication Systems," "Characterizing Superposition Arrival Processes in Packet Multiplexer for Voice and Data," "Second Moment Resource Allocation in Multi-Service Networks," "Statistical Multiplexing of Multiple Time-scale Markov Streams," "Fundamental bounds and approximations for ATM multiplexers with applications to video teleconferencing," "Design of a real-time call admission controller for ATM networks," "E#ective Bandwidth of General Markovian Tra#c Sources and Admission Control of High Speed Networks," Queueing Analysis of High-Speed Networks with Gaussian Tra#c Models "Self-similar processes in communications networks," "On the Relevance of Time Scales in Performance Oriented Tra#c Characterization," "The measurement-analytic frame-work for QoS estimation based on the dominant time scale," --TR Limits for queues as the waiting room grows Effective bandwidth of general Markovian traffic sources and admission control of high speed networks On the self-similar nature of Ethernet traffic (extended version) Effective bandwidth and fast simulation of ATM intree networks Multiplexing gains in bit stream multiplexors Design of a real-time call admission controller for ATM networks Second moment resource allocation in multi-service networks Improved loss calculations at an ATM multiplexer A central-limit-theorem-based approach for analyzing queue behavior in high-speed networks --CTR Han S. Kim , Ness B. Shroff, The notion of end-to-end capacity and its application to the estimation of end-to-end network delays, Computer Networks: The International Journal of Computer and Telecommunications Networking, v.48 n.3, p.475-488, 21 June 2005 Mahmoud Elhaddad , Rami Melhem , Taieb Znati, Analysis of a transmission scheduling algorithm for supporting bandwidth guarantees in bufferless networks, ACM SIGMETRICS Performance Evaluation Review, v.34 n.3, p.48-63, December 2006 Aimin Sang , San-qi Li, Measurement-based virtual queue (VQ): to estimate the real-time bandwidth demand under loss constraint, Computer Networks: The International Journal of Computer and Telecommunications Networking, v.46 n.4, p.519-539, 15 November 2004 Ness B. Shroff, A predictive flow control scheme for efficient network utilization and QoS, IEEE/ACM Transactions on Networking (TON), v.12 n.1, p.161-172, February 2004 Jzsef Br, Loss ratio approximations in buffered systems with regulated inputs, Proceedings of the 1st international conference on Performance evaluation methodolgies and tools, October 11-13, 2006, Pisa, Italy

maximum variance asymptotic;asymptotic relationship;queue length distribution;loss probability

504633

QoS provisioning and tracking fluid policies in input queueing switches.

The concept of tracking fluid policies by packetized policies is extended to input queueing switches. It is considered that the speedup of the switch is one. One of the interesting applications of the tracking policy in TDMA satellite switches is elaborated. For the special case of 2 2 switches, it is shown that a tracking nonanticipative policy always exists. It is found that, in general, nonanticipative policies do not exist for switches with more than two input and output ports. For the general case of N N switches, a heuristic tracking policy is provided. The heuristic algorithm is based on two notions: port tracking and critical links. These notions can be employed in the derivation of other heuristic tracking policies as well. Simulation results show the usefulness of the heuristic algorithm and the two basic concepts it relies on.

INTRODUCTION One of the main issues in the design of integrated service networks is to provide the service performance requirements for a broad range of applications. Applications requirements are translated into network quantitative parameters. The most common performance measures are packet loss probability, delay, and jitter. The delay and jitter characteristics at each switch of the network is determined by the scheduling algorithm used in the switch and the incoming traffic pattern. On the other hand, the network should also be capable to analyze the amount of resources that each particular application requires. Based on this analysis a connection request is admitted or rejected. It is therefore, very important for the network designer to understand the effect a scheduling policy has on the connection performance and on the usage of network resources In many cases, it is easier to perform the analysis and design of scheduling policies under the modeling assumption that the traffic arrives and is treated as a fluid, i.e., the realistic case where information is organized into packets is not taken into account, [3],[4],[11],[12],[6]. Under the fluid policy, we assume that at every time instant arbitrary fractions of the link capacity can be shared among different applications. Although in most of the practical situations this is an idealistic assump- tion, it enables us to analyze the effect of scheduling policy on the network resources as well as the major performance pa- rameters, and therefore to design the scheduling policies more conveniently. One approach to the design of packetized policies is to first find an appropriate fluid policy, and then to derive a packetized policy that resembles or tracks the fluid policy in a certain sense. Existence of packetized tracking policies is a well established fact in the single link case. In fact, several tracking policies are suggested and their performance and efficiency are analyzed [11], [12], [6], [5], [9]. However, the existence of such policies in input queueing switches is still an open problem. This is the main subject of this paper. The research on scheduling N \Theta N switches is mainly concentrated on output queueing switches. In an N \Theta N switch, it is possible that all N inputs have packets for the same output at the same time. In order to accommodate such a scenario in an output queueing switch, the switch fabric should work N times faster than the line rates. This might be acceptable for moderate size switches working on moderate line rates, but as the capacity of the lines as well as the switch sizes increase, memories with sufficient bandwidth are not available and input queueing is becoming a more attractive alternative. One way to circumvent this problem is to have Combined Input Output Queueing (CIOQ) switches with limited speed up that matches the output sequence of a purely output queueing switch. In fact, it is shown in [2] that speed up of 2 is sufficient to resemble the output pattern of any output queueing switch. However, the scheduling algorithm proposed to do that is fairly complicated, and the arbiter still requires to receive information from the input ports of the switch with speed up of N . In this paper we consider an input queueing switch, where every input and output port can service 1 packet per time unit (All packets are considered to have equal size). In a fluid policy model, at every time slot every input (output) port can be connected to several output (input) ports, however the total service rate of any port should not exceed its capacity. Under a packetized policy, every input (output) port can at most be connected to one output (input) port at every time slot, i.e., there is no speed up in the switch fabric. Under these circum- stances, our objective is to find a packetized policy that tracks a given fluid policy in an appropriate manner. For the special case of 2 \Theta 2 switches the existence of tracking policies is proved and a non-anticipative tracking policy is provided. For the general case a heuristic algorithm with good, but not perfect tracking properties is proposed. In fact, in the simulations done, less than 1 percent of the packets lost track of the fluid policy, when the utility of the switch is around 92%. Another interesting application of the tracking policies is in the scheduling of TDMA Satellite Switches (TDMA-SS) with multi-periodic messages. In this problem the objective is to schedule a packet during every period of a connection stream, and before arrival of the next packet. Since it is not usually possible to queue the packets in the satellite switches, an input queueing model is more appropriate in this case. The fluid policy that accomplishes the specified TDMA objective is trivial. The original problem is then solved by specifying a packetized policy that tracts that fluid policy. The organization of paper is as follows. In the next sec- tion, we will review the concepts of fluid and tracking poli- cies, and provide the feasibility condition for both cases.The problem of scheduling multi-periodic messages in TDMA-SS is explained and elaborated in section III. We will show that this problem is essentially a special case of the posed input queueing scheduling problem. In section IV, we show that for the 2 \Theta 2 switches a tracking policy always exist, and we provide a non-anticipative algorithm to find the tracking policy. In this section, we also address the problem of providing a packetized policy that satisfies pre-specified packet deadlines. In section V, some useful ideas regarding the design of heuristic tracking policies are given. Based on these concepts a heuristic simulation algorithm is proposed. The heuristic algorithm is applied to the scheduling of a multi-periodic TDMA-SS and the simulation results are given. II. FLUID AND PACKETIZED TRACKING POLICIES We consider input queueing switches that serve fixed size packets. Each input and output port has the capacity of serving 1 packet per time unit. Since queues exist only at the input ports, the latter assumption implies that traffic of at most 1 packet per unit of time can be transferred from the input ports to a given output port. We assume that the time is slotted and the length of each slot is equal to the length of a packet. Slots are numbered starting from 1, 2, . Slot k is taking the time interval is the beginning (end) of time slot k: Packets arrive at the beginning of each time slot. Packets with origin input port i and destination output port j are served FCFS. Two broad classes of policies are considered, the fluid and the packetized policies. During time slot k a fluid policy trans- units of information from input port i to output port j. w ij (k) is a nonnegative real number and is measured in units of packets. Since at most one unit of work can be transferred from a given input port to the output ports, and since no queueing is permitted at the output ports, the w ij (k)'s must satisfy the following inequalities. (1) A packetized policy is based on the assumption that during a time slot an input port can transmit a single packet to any one of the output ports. Therefore, for a packetized policy we have that J ij (k), the number of packets transmitted from port i to port j during slot k, is either 0 (no packet transmission during slot single packet transmission during slot k). A packetized policy is feasible if at every time slot k we have, (2) Note that the conditions in (2) imply that for any k, there can be at most a single 1 in each column of row of the matrix That is, the matrix J ij (k) is a sub-permutation matrix Usually fluid policies cannot be applied directly in a net-work since mixing of traffic belonging to different packets is not allowed. However, they are considered in this paper, because the performance analysis and the scheduling policy design is often more convenient for fluid policies. An approach to the design of packetized policies is to first design and analyze a fluid policy, and then implement a packetized policy that resembles in a certain sense the departure process of the fluid policy. Such a packetized policy is called a tracking pol- icy. More precisely, for our purposes, we use the following definition: Definition I: Given a fluid policy , we say that a packetized policy is tracking if every packet departs under the packetized policy at the latest by the end of the time slot at which the same packet departs under the fluid policy. A basic question is if tracking policies exist for a given fluid policy. This question is answered positively for the single link case, where different sessions share a single link [11], [6]. In that case, perhaps the most well known fluid policies are GPS and rate controlled service disciplines. Several tracking policies are suggested for the single link case [11], [6], [9], [1]. The same concepts of GPS and rate controlled schedulers can be extended to the multi-input,multi-output input queueing switches. However, the existence of tracking policies for these switches is still an open question. Searching for a tracking policy can be converted to another scheduling problem, scheduling of packets with dead- lines. Suppose that a set of packets are given, every packet has two associated time stamps. The first time stamp is the eligible time, which is the time after which we can schedule the packet. For instance, this can be the arrival time of the packet to the switch. The second time stamp is the deadline of the packet. The objective is to schedule a packet inside the time frame of eligibility time and deadline time. Obviously, if a packetized scheduling policy satisfies all the deadlines induced by a fluid policy, it is a tracking packetized policy by our definition. In section IV, we study the special case of 2\Theta2 switches. We will prove that for the special case of 2 \Theta 2 switches, for every feasible fluid policy there exists a feasible packetized policy. In fact, our proof is constructive and will provide an algorithm to derive a tracking policy. We will also show that a natural extension of the earliest deadline first scheduling can be used to solve the problem of scheduling packets with deadline. The general case of an N \Theta N input queuing switch is currently under investigation. For the latter case, we provide in this paper a heuristic algorithm which shows very good performance under a number of simulation studies. III. MULTI-PERIODIC TDMA SATELLITE SWITCHES One of the potential applications of tracking policies is in the scheduling of TDMA Satellite Switches (TDMA-SS). The conventional method to do the scheduling is based on the Inukai method [10]. This method is based on the assumption that all messages have the same period. The scheduling is done for a frame length equal to the period of the messages and it is repeated periodically thereafter. Let L be equal to the maximum number of packets that can be serviced by an input/output port during one period. A set of messages is schedulable if for every port the total number of packets that should be serviced is not more than L. Inukai has provided a scheduling algorithm for any set of schedulable messages. The Inukai algorithm does not work appropriately when messages have different periods. Let message m from input port s m to output port dm has period pm . To apply the Inukai method the frame length should be set to the LCM of all message periods, say L. For each message m, L=pm unit length packets are scheduled in the frame. Each of these packets is associated to one period of the original message. Then, we can use the Inukai method to allocate these packets inside the frame length L. The problem is that there is no control over the place of packets inside the frame in the Inukai method. Thus, it is possible that all packets attributed to a single periodic message are placed next to each other. Such an assignment suffers from high jitter. Moreover, the delay of a packet can be equal to L, which can be very large. Suppose that the objective is to schedule every packet in the time frame of its period. Thus, every packet can tolerate a delay up to its period. The question then arises whether it is possible to provide a schedule under these constraints. A necessary condition for schedulability, is that the utilization of every input port i and output port j should not be greater than unity, i.e., If one considers fluid policies, then it is easy to provide a schedule provided that (3), is satisfied. Specifically, consider the fluid policy that assigns the fix service rate of 1=pm to every message m. Under this policy the switch starts servicing every packet immediately after its arrival, and it takes pm time units to complete its service. This means that the target deadlines are all accomplished. Therefore, if we can provide a packetized policy that tracks the fluid policy, then this packetized policy will satisfy the delay constraints as well, and (3) becomes the necessary and sufficient condition for the schedulability of packetized policies under the specified constraints. In [13] Philp and Liu conjectured that (3) is the necessary and sufficient condition for schedulability under the specified delay constraints. Giles and Hajek [8] have proved this conjecture for a special case. In their model the messages are sorted based on their period, such that, Moreover, for every two subsequent messages we have, where k is an integer. Unfortunately, their algorithm does not work well in the general case. As we saw, the correctness of the conjecture is trivial for the fluid policies, but it has not been proved for packetized policies. In the next section, we show the existence of tracking policies for the special case of 2 \Theta 2 switches. Thus, the conjecture is proved for the special case of 2 \Theta 2 switches. IV. In this section, we consider a 2 \Theta 2 input queueing switch and provide an algorithm for designing a packetized policy t that tracks a given fluid policy . be the total amount of traffic transferred under the fluid policy, from input port i to output port j by the end of slot k. Thus, we have, conditions (1). In the following, we address the following problem. Problem I: Find a sequence of sub-permutation matrices such that for all k we have where I ij If a solution to Problem I can be found, then the packetized policy that uses the sub-permutation matrix J(k) to schedule packets during slot k is tracking the given fluid policy. To see this, notice first that according to the leftmost inequality in (5), by the end of slot k, if a (integer) number of packets completes transmission between input port i and output port j under the fluid policy, then at least the same number of packets completes transmission between the same input and output port under the packetized policy: This and the fact that packets between ports i and j are served FCFS under both policies, imply that: If a packet completes transmission in slot l under the fluid pol- icy, it will also complete transmission at the latest by the end of slot l under the packetized policy. We note next that a realizable packetized policy should serve packets after their arrival to the switch. This is ensured by the rightmost inequality in (5). To see this, notice that according to this inequality, the number of packets that complete transmission between input i and output port j under the packetized policy cannot exceed the number of packets that complete transmission between i and j under the fluid policy by more than 1. Moreover, if W ij (k) is integer then necessarily, This and the fact that packets are served FCFS under both policies, imply that: A packet cannot complete transmission in slot k under the packetized policy unless part of this packet has already been transmitted up to the end of slot k by the fluid policy. Since the fluid policy is feasible and never begins transmission of a packet before its arrival, the same holds for the packetized policy. Note that if we solve Problem I, we in fact have a packetized policy that not only tracks the finishing times of the packets under the fluid policy, but also the times when the fluid policy starts transmission of these packets. Before we proceed we need another definition. Definition II: An integer valued matrix I is called a u- neighbor of a matrix W if I ij for all j; (7) I ij for all i: (8) We now address a stricter version of Problem I. Problem II: Find a sequence of sub-permutation matrices such that for all k, I(k) is a u-neighbor of I ij We now proceed to provide a solution for Problem II. The proof is by induction. At the beginning of slot 0 no traffic has been processed under either of the policies (fluid and packe- tized), and we set Assume now that we have found appropriate sub-permutation matrices until slot k. We show how to construct the sub- permutation matrix for slot k +1; J(k +1); based on I(k) and so that the matrix I(k) of J(k+1) is a sub-permutation matrix. To see this note first that The second inequality holds because or 1. Next we have to show that there can be no more than a single 1 in each column or row of To see this, note that if Hence, if there is more than a single 1 in, say, column 1 of I i1 (k) 2 \GammaX sinceX But the last inequality contradicts the fact I i1 We now show that I(k + 1) satisfies (6). From (9) we have that Also, 0: while It remains to prove (7) and (8) for the matrix I(k +1): Con- sider, say, column 1. We distinguish the following cases. Case 1. J i1 and henceX I The third relation above holds because of (7). The fourth relation is correct sincebffc integer, and the fifth relation follows from Case 2. J 0: Then by I i1 (k) bW 2. Since we also have I conclude that the difference I i1 only the values 0 and 1. We now need to distinguish the following sub-cases. Sub-case 2.1 I i1 I I i1 (k) The last relation is correct since 1 1. Sub-case 2.2 I i1 If I i1 (k) I The last relation follows since J i1 2. It remains to consider the case where W i1 non-integer for all i. We may then redefine J still have a sub-permutation matrix. To see this, consider the following two cases (a) In column 2 there is already J say a sub-permutation matrix, we know that J 22 (k Therefore, setting J 21 1 we still have a sub-permutation matrix J(k 1). (b) J k+1 Then we can set J k+1 i1 for one of the s and still have a sub-permutation matrix J(k 1). In order to show that (6) still holds with the modified J(k+1), note that since I i1 1)c and W i1 (k +1) is non-integer I and clearly, To show that (8) holds, note that since for the modified matrix we now have J i1 (k can apply the same argument as in case 1 above. We may continue in this fashion and examine the rest of the columns and rows, and update, if necessary, the matrix J We eventually will have that the matrix I is a u-neighbor of a k+1 . According to the procedure above, we have the following algorithm for creating the tracking policy, i.e., for generating the sub-permutation matrix J(k 1). ALGORITHM 1. At the beginning of slot create the matrix elements 2. If for some column or row, say column 1, it holds I i1 non-integer for all then redefine J i1 so that the modified matrix J(k+ 1) is still a sub-permutation matrix. Note that the policy obtained in this way is non-anticipative, i.e., it does not depend on future arrivals and therefore, can be implemented on-line. So far, we have shown that the tracking policy exists, and a procedure to convert a fluid policy to a packetized tracking policy is given. In the previous section, it was mentioned that to obtain a tracking policy, we can convert the problem to a deadline scheduling problem and solve that problem. In the following a simple extension of the EDF algorithm for a 2 \Theta 2 switch is given. This policy, which we call it the EDF2 always come up with an admissible schedule if such a schedule exists. A schedule is admissible if it satisfies all the deadlines. A. Per-flow Tracking So far, we have considered the per-link tracking policies, and proved their existence for 2 \Theta 2 switches. Basically, we showed that for every fluid policy there exists a packetized policy that tracks the aggregate traffic going from every input port i to every output port j. In many circumstances this is not quite enough. The aggregate traffic between any pair of input/output nodes consists of several distinct flows or sessions. Generally, to provide per-flow QoS, the fluid policy should work on the granularity of flows and guarantees the service rate given to every flow individually. This should be also reflected in the packetized tracking policy. More specifically, the corresponding packetized policy should track the service given to every flow under the fluid policy. Let W l ij (k) and I l ij (k) be the total amount of flow l traffic transferred from port i j up to time k under the fluid policy, and the packetized policy, p respectively. Then, the packetized policy p is a per-flow tracking policy of f if and only if, bW l ij (k)c I l Obviously, this implies a stricter definition for tracking poli- cies. Recently, we were able to prove this notion of tracking policy for 2 \Theta 2 switches as well. The result is given without proof in the following theorem. Theorem 1: Consider any arbitrary fluid policy f for input queueing 2 \Theta 2 switches. There exists a non-anticipative packetized policy p that tracks the individual services given to every flow under the fluid policy f . B. QoS provisioning in 2 \Theta 2 switches An alternative approach to the QoS provisioning problem is to view directly the requirements of the real-time traffic and to attempt to satisfy them. A natural framework for this approach is deadline satisfaction. Every packet presents upon arrival the maximum waiting time that may tolerate and a deadline is determined by which it should depart. The scheduler attempts to satisfy as many deadlines as possible. For a single link it is known that the Earliest Deadline First policy satisfies the packet deadlines if those are satisfiable. We show here that the same effect is achievable in 2 \Theta 2 switches if the same notion is generalized. For each link (i; j) at every slot t let D ij (t) be the earliest deadline among the backlogged packets. There are two service configurations 1)g. Let be the sorted deadline vectors associated to the two configurations, such that the first component always be the minimum deadline, i.e., We say that the deadline vector D i is lexicographically smaller than D j if and only if, (D i (D i 1 and D i Definition III: The scheduling policy EDF2 is the policy that at every time slot selects the configuration with minimum lexicographical deadline vector. EDF2 is the natural extension of the EDF to the 2 \Theta 2 case. Suppose that a sequence of packets with deadlines are given, and it is known that the deadlines are satisfiable, i.e., there is a feasible scheduling that meet all the deadlines. Then, the scheduling policy EDF2 also satisfies the deadlines. The proof of this claim is straightforward and is similar to the proof of same result for the EDF policy in the single link case. The EDF2 policy can be applied to obtain a tracking pol- icy. In that case, at every time slot k, deadline of packets are set to the end of the time slot that they depart the switch under the fluid policy. The departing times are calculated based on the assumption that there is no future arrivals. Note that the crucial information for scheduling is the relative order of packets departure times not the departing times themselves. If we assume that the future arrivals will not change the departure order of the packets, then the departure orders obtained under the no future arrival assumption will be correct, regardless of future arrivals. Therefore, the EDF2 policy would be a non-anticipative tracking policy. As we will illustrate in the next section, the basic assumption regarding the independence of future arrivals and the departure order of backlogged packets is not a reasonable assumption for the general case of N \Theta N switches. V. HEURISTIC ALGORITHMS Let F be a feasible fluid policy that at every time slot k specifies the appropriate fluid scheduling matrix w(k). The scheduling matrix w(k) is an N \Theta N matrix indicating the rate of transmission from every input port to every output port, and is a function of arrivals, (a k ) and back logged traffic (q k ) at time k. In general, the arrival process is non-anticipative, and therefore the scheduling function is not known in advance. Under these circumstances, it is impossible to track the fluid policy perfectly for N ? 2. We will illustrate this by an example. ffl Example: Consider a 4 \Theta 4 switch. Let 0:25 0:25 0:25 0:25 0:25 0:25 0:25 0:25 Without loss of generality, assume that the tracking policy selects the following two sub-permutation matrices for the two first time slots. Now assume that the fluid policy, because of arrival of some packets with higher priority, takes the following rate matrix for the next time slot, It is clear that the fluid policy finishes four packets on the links (1,2), (1,4), (2,1), (2,3) in the third time slot, and none of them is scheduled in the first two time slots by the packetized policy. Thus, at the end of third time slot, the tracking policy will, at least miss deadlines of 2 packets. It is worth mentioning that one of the crucial assumptions in the single link sharing case is that the departure order of the packets present in the node is independent of future arrivals. This is not a valid and justifiable assumption in the N \Theta N case. The properties of non-anticipative tracking algorithms for single link case is due to this assumption, so one should not expect that similar results regarding the non-anticipative tracking policies hold for N \Theta N switches. Recall that we have already provided a non-anticipative tracking policy for the 2 \Theta 2 case. The above example shows that result can not be generalized to the N \Theta N case. This fact motivates us to seek for heuristic algorithms with good but not perfect tracking properties. We are not intending to provide a complete efficient tracking policy here, but we want to illustrate what can be the appropriate approach to design the tracking policies. Basically, we introduce two main concepts that could be employed in the design of tracking poli- cies. We then illustrate the simulation results of a simple tracking policy implemented based on these two concepts. The two concepts that we elaborate on them in this section are port based tracking and critical links. A. Based Tracking One way to implement a packetized tracking policy can be based on the weighted matching in the bipartite graphs. In this method, the weights of the links are the difference between the fluid policy and the tracking policy total job done on that link. We call these weights the tracking weights, t ij (k): We add up all the positive weights of links associated with every input and output port and assign this weight to the corresponding vertex in the bipartite graph. We show the vertex weights by v i and v j , If we had added up weight of all links, regardless of their sign, the ones with negative weight would have lessen the total weight of the vertex and this will reduce the chance of those with positive weights (the lagging ones) to be scheduled. Next, we have to select a criterion function of weights of nodes in the matching, or perhaps, the ones left out of the matching that we attempt to optimize. We propose two possible candidates here. Summation Criterion: One possible candidate is to maximize the summation of all scheduled vertices weights. In this way, at every step we select the set of the nodes with over-all maximum weights, and therefore the overall lagging of the tracking policy is minimized. Prioritization Criterion: The other possible choice would be a prioritization criterion. Nodes are prioritized based on their weights, such that the nodes with greater weight have higher priority. Any node is excluded from the matching if and only if including it necessitates removing a node with higher priority from the matching. In this approach the absolute lagging value of a node is considered essential and the ones with greater lagging weight are scheduled. It is not hard to show that both of the above mentioned criteria are equivalent. This is due to special structure of bipartite graphs. In the next lemma we prove the equivalence of the two criteria. represents the optimal matching graph based on the summation (priority) criterion. Then, M 1 is optimal based on the priority (summation) criterion too. proof: Suppose that M 1 represents the optimal matching graph based on the summation criterion, but it is not optimal based on the priority criterion. Let M 2 be the optimal matching graph based on the priority criterion. Let i 1 be a node in 1 but not in M 2 : Consider the graph graph consists of links that are in one and only one of the two graphs M 1 and M 2 ). The maximum degree of a vertex in G is two, therefore it consists of a union of distinct paths and loops. should be the first vertex in a path in G, because its degree is one. We focus on this path. Two alternative cases are possible, number of nodes in the path are either even or odd. If it is even the last node in the path is also a member of M 1 , while all intermediate nodes belong to both matching. Thus, if we replace the alternative set of links in the path belonging to M 2 with those belonging to M 1 , two more vertices will be included to M 2 , and this contradicts with the optimality assumption of M 2 . If number of the vertices in the path is odd, the last vertex is a member of M 2 only. The last node should not have less weight than the first vertex i 1 , otherwise M 2 is not the optimal graph based on the priority criterion. It should not have greater weight, because this contradicts with the optimality of Thus, its weight should be equal to i 1 's. Therefore, for every node in M 1 there is a node in M 2 with equal weight and vice versa. This means that both graphs are optimal according to both criteria. The above argument also provides an algorithm to derive the optimal matching graph. The algorithm is based on exploring the augmenting path similar to the case of maximum matching algorithm. The only difference is that in the maximum matching case the search results a better matching, whenever a free node is detected. Here, the search results a better matching, if or a node with smaller weight than the first node of the augmenting path is detected. The following algorithm finds the maximum weighted vertex matching. Matching Algorithm: 1. Sort the nodes according to their weights. 2. Select the highest weight node not in the matching that is not selected yet. 3. Search for an augmenting path started from the selected node in step 2. Search ends successfully either if a free vertex is detected (an augmenting path is found), or when a node with smaller weight than the first node is found. Search ends unsuccessfully if all possible paths are searched and none of the above mentioned cases are occurred. 4. Repeat steps 2 and 3 until all nodes are selected. Although bipartite matching algorithms have been extensively used in scheduling of the switches, they are either based on the maximum link (edge) weighted matching or maximum matching algorithms. The problem with maximum link weighted matching algorithms are their complexity, and the problem with maximum matching algorithms are their poor performance. The vertex based matching can be considered as an intermediate solution. We can enhance the performance of a scheduling algorithm based on vertex maximum matching by selecting an appropriate initial set of eligible links, which will be provided to the ar- biter, and by modifying the weights of the nodes that are more urgent to be scheduled. Basically, we consider a two stage scheduling. At the first step a set of eligible links for scheduling and weights of vertices are derived. In the next step, the arbiter select the links in the matching based on the optimum vertex weighted matching algorithm described above. One way to select an eligible set of links in the tracking problem is based on the notion of critical links. The critical links are those that are urgent to be scheduled, and if they are not scheduled the tracking policy will loose the track of the fluid policy. If such a link is detected all non-critical links sharing same input or output port with the critical link will be excluded. The precise definition of critical nodes is given in the next section. B. Critical Ports and Links A critical port is a port that should be scheduled in the next time slot, in order not to miss a deadline in the future. As an example suppose that we are at the beginning of k \Gamma th time slot. Let say that there are two packets one from node i to j 1 and the other from node i to j 2 both with deadline k 1. Note that if we do not schedule any of these packets, no deadline will be missed in the k \Gamma th time slot. However, we will definitely miss a deadline at the subsequent time slot, 1. We say that node i is a critical node, and links (i; and are associated critical links. In general, sufficient condition for a port to be critical at time k is to have at least packets with deadlines less than or equal to k + p. Denote that we are stating a sufficient condition. In other words, there might be some critical ports that can not be detected well in advance using this criterion. In case of the tracking policy, the deadline of packets are implicitly given, and are equal to the end of the time slot that the packet departs the switch under fluid policy. We may not know the deadlines well in advance, since the future rate of every link under the fluid policy depends on the future arrivals, which are non-anticipative. Nevertheless, we may have an approximate deadline for every packet based on back-logged traffic or the average arrival rate of the links (for instance, based on a (oe; ae) flow model for ATM traffic). Also, in some approaches, a constant rate is assigned to a session, regardless of the arrival process, such as in multi-periodic messages in TDMA-SS or rate controlled service disciplines. In the case of networks if no packet from the assigned session is available at time of scheduling, packets from other kind of traffic (for instance, the best effort traffic) can use the available slot [14]. The next issue is to set an appropriate inspection horizon. The inspection horizon is the number of time slots in future that we inspect to detect the critical nodes. Based on our ex- periments, we found that inspection horizon around five is ade- quate. In fact, there is a trade-off involved here, increasing the inspection horizon helps us in detecting more critical links, but it increases the complexity of the algorithm as well. After detecting a critical port, we know that we have to schedule one of the critical links associated with that, otherwise we will miss the deadlines. Thus, we remove all links associated with critical port, which are not critical. In this way, the chance for the scheduler to arbitrate one of the critical links is increased. We also increase the weight of the critical ports with a constant, so that their weight become greater than all non-critical ports. Therefore, these nodes are prioritized by the scheduler. The critical nodes detecting algorithm can be described as Critical Node Detecting Algorithm: 1. For every node i and j do steps 2 and 3; 2. Calculate number of Packets that should be sent in the next time slots. 3. If the number of packets that should be sent in the next l time steps is equal to l; the corresponding node is critical. Moreover, the associated links that should at least send one packet in the next l time slots are critical links. The whole scheduling process of a switch can be divided into two stages. In the first stage, weight of the ports are calculated and the criticality of the ports are investigated. These computation can be done in parallel for different ports of the switch, and no interaction between them is necessary. In the next stage the computed weights for the nodes and the eligible links for every node are provided to the arbiter processor. One of the main concerns in high speed switches is the volume of information required to be exchanged between different cards of the switch, namely the signaling scalability of the algorithm. In the link weighted matching arbiters, weights associated to every link should be sent to the arbiter. Thus, the exchanging information is in the order of N 2 . In our approach the weights of the nodes and the eligible links (one bit information per link) should be sent to the arbiter, which is in the order of N . Finally we provide a simple scheduling algorithm based on the algorithms described above. Heuristic Scheduling Algorithm: 1. At every time slot k do the following steps. 2. Calculate weights of the nodes using (10),(11). 3. Insert all links with positive weights in eligible links set. 4. Check for critical nodes and their associated critical links. 5. Increase weight of every critical node, such that its weight exceed all non-critical node's weights. Moreover, remove all non-critical links of the critical nodes from the eligible links set. 6. Pass weights of the nodes and the critical links set to the matching algorithm. The result of the matching is the schedule for time slot k. This algorithm will be used in next section for scheduling in a TDMA-SS, and the simulation results will be illustrated. C. Simulation Results In this section we provide some primary results regarding the heuristic algorithm. The thorough investigation and analysis of the heuristic algorithms are still continuing. How- ever, the primary results appear promising. We have used the heuristic algorithm for scheduling of a TDMA-SS with multi- periodic messages. Suppose that messages are all periodic, and the objective is to schedule every period of a message before the next one arrives. Thus, if the period of message M i ; is slots to schedule every packet of that mes- sage. The fluid policy that assigns the constant rate of 1=p i to message M i accomplishes this task. The messages are generated randomly between input and output pairs with equal probability. The period of the message is selected uniformly in the range of f3; :::; 8g. The messages are selected such that in every experiment the utility of every node is in [0:90; 0:97]. In all experiments the average utility I utility l=0 l=1 l=2 late % is around 0.92. The critical nodes are inspected based on the constant rate, and the inspection horizon is set to N . We have done the simulation for different switch sizes, and for every switch size we generated 100 different message sets and investigate the performance of the algorithm. In our model there can be different messages from the same input to the same output, but the scheduler works with the aggregate rate of all these messages. This is important, because the complexity of the arbiter does not depend on the number of messages, which can be large. To schedule different messages from the same input to output an EDF scheduler is maintained in every input port. A packet is considered to be on time (zero latency) if it is scheduled within a period interval after its arrival. The latency of a packet, l is equal to k ? 0 if it is scheduled k time units after its deadline. The result of the simulations are given in the table below. As far as the scheduler is concerned, it is working with constant rate traffics. Therefore, the results of this simulation are also applicable in the case of fixed rate scheduler for the network switches The number of packets with different latencies are indicated in the table. The percentage of late packets is also shown in the last column. Every row correspond to a different switch size. The number of packets missing their deadlines are less than one percent, and the maximum latency is two time units in all cases. We believe that in many of the applications this is an acceptable performance. In many applications the excessive delay imposed by the arbiter is tolerable. In fact, in some of the applications we are allowed to miss some packets, so we can neglect the packets that miss their deadline. This in return aids us in on time scheduling of the other packets. The other important issue is the size of the switch. Many of the heuristic algorithms proposed degrades as the size of the switch increases [13]. In our case, we did not observe any such degradation. VI. AND CONCLUSION In this paper the notion of fluid policies and tracking policies are extended to the N \Theta N switches. These concepts are both useful in the high speed networks, where they aid us in providing guaranteed service to different applications, and in TDMA-SS with multi-periodic messages. The existence of tracking policy is proved for the special case of 2 \Theta 2 switches. For the general case of N \Theta N switches a heuristic algorithm is provided. The heuristic algorithm is mainly presented to confirm the validity of two important notions and measures in tracking policies, the port tracking and critical node concepts. The existence of tracking policy for the N \Theta N is still an open question. However, based on the results provided here for the special cases and the heuristic algorithm, we think that such a tracking policy always exist. However, we have shown that it is impossible to have a perfect tracking policy when the fluid policy is non-anticipative. This fact together with the complexity issue justify the need for better and less complicated heuristic tracking policies, with good but not perfect performances. --R A calculus for network delay. A calculus for network delay. Analysis and simulation of a fair queueing algorithm. Optimal multiplexing on single link: Delay and buffer requirements. Efficient network QoS provisioning based on per node traffic shaping. Scheduling multirate periodic traffic in a packet switch Shaping. A generalized processor sharing approach to flow control in integrated services networks: The single node case. A generalized processor sharing approach to flow control in integrated services networks: The multiple node case. Scheduling real-time messages in packet-switched networks --TR Data networks Analysis and simulation of a fair queueing algorithm Introduction to algorithms A generalized processor sharing approach to flow control in integrated services networks A generalized processor sharing approach to flow control in integrated services networks Efficient network QoS provisioning based on per node traffic shaping EDD Algorithm Performance Guarantee for Periodic Hard-Real-Time Scheduling in Distributed Systems Scheduling real-time messages in packet-switched networks --CTR Yong Lee , Jianyu Lou , Junzhou Luo , Xiaojun Shen, An efficient packet scheduling algorithm with deadline guarantees for input-queued switches, IEEE/ACM Transactions on Networking (TON), v.15 n.1, p.212-225, February 2007

QoS provisioning;input-queued switching;scheduling

504641

performance over end-to-end rate control and stochastic available capacity.

Motivated by TCP over end-to-end ABR, we study the performance of adaptive window congestion control, when it operates over an explicit feedback rate-control mechanism, in a situation in which the bandwidth available to the elastic traffic is stochastically time varying. It is assumed that the sender and receiver of the adaptive window protocol are colocated with the rate-control endpoints. The objective of the study is to understand if the interaction of the rate-control loop and the window-control loop is beneficial for end-to-end throughput, and how the parameters of the problem (propagation delay, bottleneck buffers, and rate of variation of the available bottleneck bandwidth) affect the performance.The available bottleneck bandwidth is modeled as a two-state Markov chain. We develop an analysis that explicitly models the bottleneck buffers, the delayed explicit rate feedback, and TCP's adaptive window mechanism. The analysis, however, applies only when the variations in the available bandwidth occur over periods larger than the round-trip delay. For fast variations of the bottleneck bandwidth, we provide results from a simulation on a TCP testbed that uses Linux TCP code, and a simulation/emulation of the network model inside the Linux kernel.We find that, over end-to-end ABR, the performance of TCP improves significantly if the network bottleneck bandwidth variations are slow as compared to the round-trip propagation delay. Further, we find that TCP over ABR is relatively insensitive to bottleneck buffer size. These results are for a short-term average link capacity feedback at the ABR level (INSTCAP). We use the testbed to study EFFCAP feedback, which is motivated by the notion of the effective capacity of the bottleneck link. We find that EFFCAP feedback is adaptive to the rate of bandwidth variations at the bottleneck link, and thus yields good performance (as compared to INSTCAP) over a wide range of the rate of bottleneck bandwidth variation. Finally, we study if TCP over ABR, with EFFCAP feedback, provides throughput fairness even if the connections have different round-trip propagation delays.

Introduction bottleneck link xmitter recvr ABR recvr end-to-end ABR source ABR Figure 1: The network under study has a large round trip delay and a single congested link in the path of the connection. The ABR connection originates at the host and ends at the destination node. We call this scenario "end-to-end" ABR. The Available Bit Rate (ABR) service in Asynchronous Transfer Mode (ATM) networks is primarily meant for transporting best-effort data traffic. Connections that use the ABR service (so called ABR sessions) share the network bandwidth left over after serving CBR and VBR traffic. This available bandwidth varies with the requirements of the ongoing CBR/VBR sessions, hence the switches carrying ABR sessions implement a rate-based feedback control for congestion avoidance. This control causes the ABR sources to reduce or increase their cell transmission rates depending on the availability of bandwidth in the network. As the ABR service does not guarantee end-to-end reliable transport of data to the applications above it, an additional protocol is needed between the application and the ATM layer to ensure reliable communication. In most deploy- IP host IP host ATM wide area network router/switch router/switch IP IP ABR IP IP ABR Figure 2: TCP/IP over edge-to-edge ATM/ABR, with the TCP connection split into one edge-to-edge TCP over ABR connection, and two end-to-edge TCP over IP connections. The edge switch/router regulates the flow of TCP acknowledgements back to the TCP senders. ments of ATM networks, the Internet's Transport Control Protocol (TCP) is used to ensure end-to-end reliability for data applications. TCP, however, has its own adaptive window based congestion control mechanism that serves to slow down sources during network congestion. Hence, it is very important to know whether the adaptive window control at the TCP level, and the rate-based control at the ABR level interact beneficially from the point of view of application level throughput. In this paper, we consider the situation in which the ATM network extends upto the TCP endpoints; i.e., end-to-end ABR (as opposed to edge-to-edge ABR), see Figure 1. Our results also apply to TCP over edge-to-edge ABR if the end-to-end TCP connection comprises a tandem of an edge-to-edge TCP connection, and two edge-to-end connections, with TCP spoofing being done at the edge-devices (see Figure 2). Consider the hypothetical situation in which the control loop has zero delay. In such a case, the ABR source of a session (i.e., the ATM network interface card (NIC) at the source node) will follow the variations in the bandwidth of the bottleneck link without delay. As a result, no loss will take place in the network. The TCP window will grow, and once the TCP window size exceeds the window required to fill the round trip pipe, the packets will be buffered in the source buffer. Hence, we can see that congestion is effectively pushed to the network edge. As the source buffer would be much larger than the maximum window size, the TCP window will remain fixed at the maximum window size and congestion control will become a purely rate based one. If ABR service was not used, however, TCP would increase its window, overshoot the required window size, and then due to packet loss, would again reduce the window size. Hence, it is clear that, for zero delay in the control loop, end-to-end ABR will definitely improve the throughput of TCP. When variations in the bottleneck bandwidth do occur, however, and there is delay in the ABR control loop, it is not clear whether there will be any improvement. In this paper, we study TCP over end-to-end ABR, and consider a TCP connection in a network with large round-trip delay; the connection has a single bottleneck link with time varying capacity. Current trends seem to indicate that at least in the near future, ATM will remain only a WAN transport technology. Hence ATM services will only extend to the network edge, whereas TCP will continue to be used as the end-to-end protocol, with interLAN IP packets being transported over ATM virtual circuits. When designing a wide-area intranet based on such IP over ATM technology one can effectively control congestion in the backbone by transporting TCP/IP traffic over ABR virtual circuits. A connection between hosts on two LANs would be split into a TCP over ABR wide-area edge-to-edge connection, and two end-to-edge connections over each of the LANs; see Figure 2. The edge devices would then control the TCP end-points on their respective LANs by regulating the flow of acknowledgements (ACKs) back to the senders. In this framework, our results would apply to the edge-to-edge TCP over ABR connection. Many simulation studies have been carried out to study the interaction between the TCP and ATM/ABR control loops. In [7], the authors study the effect of running large unidirectional file transfer applications on TCP over ABR. An important result from their study is that cell loss ratio (CLR) is not a good indicator of TCP perfor- mance. They also show that when maximum throughput is achieved, the TCP sources are rate limited by ABR rather than window limited by TCP. Reference [8] reports a study of the buffering requirements for zero cell loss for TCP over ABR. It is shown that the buffer capacity required at the switch is proportional to the maximum round trip time of all the VCs through the link, and is independent of the number of sources (or VCs). The proportionality factor depends on the switch algorithm. In further work, in [9], the authors introduce various patterns of VBR background traffic. The VBR background traffic introduces variations in the ABR capacity and the TCP traffic introduces variations in the ABR demand. In [3], the authors study the effect of ATM/ABR control on the throughput and fairness of running large unidirectional file transfer applications on TCP-Tahoe and TCP-Reno (see [15]) with a single bottleneck link with a static service rate. The authors in [12] study the performance of TCP over ATM with multiple connections, but with a static bottleneck link. The paper reports a simulation study of the relative performances of the ATM ABR and UBR service categories in transporting TCP/IP flows through an edge-to-edge ATM (i.e., the host nodes are not ATM endpoints) network. Their summary conclusion is that there does not seem to be strong evidence that for TCP/IP workloads the greater complexity of ABR pays off in better TCP throughputs. Their results are, however, for edge-to-edge ABR; they do not comment on end-to-end (i.e., the hosts have an ATM NIC) ATM which is what we study in this paper. All the studies above are primarily simulation studies, and analytical work on TCP over ABR does not seem to exist in the literature. In this paper, we make the following contributions: (i) We develop an analytical model for a TCP connection over explicit rate ABR when there is a single bottleneck link with time varying available capacity. In the analytical model we assume that the explicit rate feedback is based on the short term average available capacity; we think of this as instantaneous capacity feedback, and we call the approach INSTCAP feedback. (ii) We use a test-bed to validate the analytical results. This test-bed implements a hybrid simulation comprising actual Linux TCP code, and a network emula- tion/simulation implemented in the IP loopback code in the Linux kernel. (iii) We then develop an explicit rate feedback that is based on a longer term history of the bottleneck rate process. The computation is motivated from (though it is not the same as) the well known concept of effective capacity (derived from large deviations analysis of the bottleneck queue process). We call this EFFCAP feedback. EFFCAP is more effective in preventing loss at the bottleneck buffers. Since the resulting model is hard to analyse, the results for EFFCAP feedback are all obtained from the hybrid simulator mentioned above. Our results show that different types of bottleneck bandwidth feedbacks are needed for slowly varying bottleneck bandwidth, rapidly varying bottleneck bandwidth and the intermediate regime. EFFCAP feedback adapts itself to the rate of bandwidth variation. We then develop guidelines for choosing two parameters that arise in the on-line calculations of EFFCAP. (iv) Finally, we study the performance of two TCP connections that pass through the same bottleneck link, but have different round trip propagation delays. Our objective here is to determine whether TCP over ABR is fairer than TCP alone, and under what circumstances. In this study we only use EFFCAP feedback. The paper is organized as follows. In Section 2, we describe the network model under study. In Section 3 we develop the analysis of TCP over ABR with INSTCAP Segmentation Buffer Rate Feedback HOST COMPUTER ABR adaptive rate server bottleneck link Figure 3: The segmentation buffer of the system under study is in the host NIC card and extends into the host's main memory. The rate feedback from the bottleneck link is delayed by one round trip delay. feedback, and of TCP alone. In Section 4, we develop the EFFCAP algorithm; TCP over ABR with EFFCAP feedback is only amenable to simulation. In Section 5, we present analysis results for INSTCAP feedback, and simulation results for INSTCAP and EFFCAP. The performance of INSTCAP and EFFCAP feedbacks are compared. In Section 6, we study the choice of two parameters that arise in EFFCAP feedback. In Section 7 we provide simulation results for two TCP connections over ABR with EFFCAP feedback. Finally, in Section 8, we summarise the observations from our work. 2 The Network Model Consider a system consisting of a TCP connection between a source and destination node connected by a network with a large propagation delay as shown in Figure 1. The TCP congestion control does not implement fast-retransmit, and hence must time out for loss recovery. We assume that only one link (called the bottleneck link) causes significant queueing delays in this connection, the delays due to the other links being fixed (i.e., only fixed propagation delays are introduced due to the other links). A more detailed model of this is shown in Figure 3. The TCP packets are converted into ATM cells and are forwarded to the ABR segmentation buffer. This buffer is in the network interface card (NIC) and extends into the main memory of the computer. Hence, we can look upon this as an infinite buffer. The segmentation buffer server (also called the ABR source) gets rate feedback from the network. The ABR source service rate adapts to this rate feedback. The bottleneck link buffer represents either an ABR output buffer in an ATM switch (in case of TCP over ABR), or a router buffer (in case of TCP alone). The network carries other traffic (CBR/VBR) which causes the bottleneck link capacity (as seen by the connection of interest) to vary with time. The bottleneck link buffer is finite which can result in packet loss due to buffer overflow when rate mismatch between the source rate and the link service rate occurs. In our model, we will assume that a portion of the link capacity is reserved for best-effort traffic, and hence is always available to the TCP connection. In the ATM/ABR case such a reservation would be made by using the Minimum Cell Rate (MCR) feature of ABR, and would be implemented by an appropriate link scheduling mechanism. Thus when guaranteed service traffic is backlogged at this link, then the TCP connection gets only the bandwidth reserved for best-effort traffic, otherwise it gets the full bandwidth. Hence a two state model suffices for the available link rate. In the first part of our study, we assume that the ABR feedback is an instantaneous rate feedback scheme; i.e., the bottleneck link periodically feeds back its short term to the ABR source. This feedback reaches after one round trip propagation delay. The ABR source adapts to this value and transmits the cells at this rate. 3 TCP/ABR with Instantaneous Capacity Feedback At time t, the cells in the ATM segmentation buffer at the source are transmitted at a time dependent rate S \Gamma1 which depends on the ABR rate feedback (i.e., S t is the service time of a packet at time t). The bottleneck has a finite buffer B max and has time dependent service rate R \Gamma1 t packets=sec which is a function of an independent Markov chain. In our analysis, we assume that there is a 2 state Markov chain modulating the channel. In each state, the bottleneck link capacity is deterministic. If the buffer is full when a cell arrives to it, the cell is dropped. In addition, we assume that all cells corresponding to that TCP packet are dropped. This assumption allows us to work with full TCP packets only; it is akin to the Partial Packet Discard proposed in [13]. If the packet is not lost, it gets serviced at rate R \Gamma1 (assumed constant over the service time of the packet), and reaches the destination after some deterministic delay. The destination ATM layer reassembles the packet and delivers it to the TCP receiver. The TCP receiver responds with an ACK TCP/ABR Transmitter Bottleneck Link Propagation Delay Figure 4: Queueing model of TCP over end-to-end ABR (acknowledgement) which, after some delay (propagation processing delay) reaches the source. The TCP source responds by increasing the window size. The TCP window evolution can be modeled in several ways (see [11], [10]). In this study, we model the TCP window adjustments in the congestion avoidance phase (for the original TCP algorithm as proposed in [4] by Van Jacobson) probabilistically as follows: every time a non-duplicate ACK (an acknowledgement that requests for a packet that has not been acknowledged earlier) arrives at the source, the window size W t increases by one with probability 1 On the other hand, if a packet is lost at the bottleneck link buffer, the ACK packets for any subsequently received packets continue to carry the sequence number of the lost packet. Eventually, the source window becomes empty, timeout begins and at the expiry of the timeout, the threshold window W th t is set to half the maximum congestion window achieved after the loss, and the next slow start begins. 3.1 Queueing Network Model Figure 4 is a closed queueing network representation of the TCP over ABR session. We model the TCP connection during the data transfer phase; hence the data packets are assumed to be of fixed length. The buffer of the segmentation queue at the source host is assumed to be infinite in size. There are as many packets in this buffer as the number of untransmitted packets in the window. The service time at this buffer models the time taken to transmit an entire TCP packet worth of ATM cells. Owing to the feedback rate control, the service rate follows the rate of the bottleneck link. We assume that the rate does not change during the transmission of the cells from a single TCP packet. The service time (or equivalently, the service rate) follows the bottleneck link service rate with a delay of \Delta units of time, \Delta being the round trip (fixed) propagation delay. The bottleneck link is modeled as a finite buffer queue with deterministic packet service time with the service time (or rate) Markov modulated by an independent Markov chain on two states 0 and 1; the service rate is higher in state 0. The round trip propagation delay \Delta is modeled by an infinite server queue with service time \Delta. Notice that various propagation delays in the network (the source-bottleneck link delay, bottleneck link-destination delay and the destination-source return path delay) have been lumped into a single delay element (See Figure 4). This can be justified from the fact that even if the source adapts itself to the change in link capacity earlier than one round trip time, the effect of that change will be seen only after a round trip time at the bottleneck link. With "packets" being read as "full TCP packets", let A t be the number of packets in the segmentation buffer at the host at time t t be the number of packets in the bottleneck link buffer at time t D t be the number of packets in the propagation queue at time t R t be the service time of a packet at the bottleneck link; R t 2 fr g. We take . Thus, all times are normalized to the bottleneck link packet service time at the higher service rate. S t be the service time of a packet at the source link. S t follows R t with delay \Delta, the round trip propagation delay, i.e., S g. Since the instantaneous rate of the bottleneck link is fed back, we call this the instantaneous rate feedback scheme. (Note that, in practice, the instantaneous rate is really the average rate over a small window; that is how instantaneous rate feedback is modelled in our simulations to be discussed later; we will call this feedback INSTCAP.) 3.2 Analysis of the Queueing Model Consider the vector process Slow start phase no loss Window reaches w & ceases to grow Coarse timeout occurs log wLoss epoch round trip propagation delay Figure 5: The embedded process This process is hard to analyze directly. Instead, we study an embedded process, which with suitable approximations, turns out to be analytically tractable. consider the embedded process f ~ with ~ use the obvious notation ~ In the following analysis, we will make the following assumptions : (i) We assume that the rate modulating Markov chain is embedded at the epochs (ii) The source adapts immediately to the explicit rate feedback that it receives. This is true for the actual ABR source behaviour (as specified by the ATM Forum [1]) if the rate decreases. In the actual ABR source behaviour, an increase in the explicit rate results in an exponential growth of the source rate and not a sudden jump. We, however, assume that even an increase in the source rate takes place with a sudden jump. (iii) There is no loss in the slow start phase of TCP. In [11], the authors show that loss will occur in the slow start phase if Bmax \Delta 1even if no rate change occurs in the slow start phase. However, for the case of TCP over ABR, as the source and bottleneck link rates match, no loss will occur in this phase as long as rate changes do not occur during slow-start. Hence, this assumption is valid for the case of TCP alone only if Bmax \Delta Observe that packets in the propagation delay queue (see Figure 4) at t k will have departed from the queue by t k+1 . This follows as the service time is deterministic, equal to \Delta, and t \Delta. Further, any new packet arriving to the propagation delay queue during still be present in that queue at t k+1 . On the other hand, if loss occurs due to buffer overflow at the bottleneck link in (t k ; t k+1 ), we proceed as follows. Figure 5 shows a packet loss epoch in the interval This is the first loss since the last time that TCP went through a timeout and recovery. At this loss epoch, there are packets in the bottleneck buffer, and some ACKs "in flight" back to the transmitter. These ACKs and packets form an unbroken sequence, and hence will all contribute to the window increase algorithm at the transmitter (we assume that there is no ACK loss in the reverse path). The transmitter will continue transmitting until the window is exhausted and then will start a coarse timer. We assume that this timeout will occur in the interval (t k+2 ; t k+3 ) (see Figure 5), and that recovery starts at the embedded epoch t k+3 . Thus, when the first loss (after recovery) occurs in an interval then, in our model, it takes two more intervals to start recovery. s). Note that, since no loss has occurred (since last recovery) until t k , therefore, the TCP window at t k is a d. Now, given ~ assuming that (i) packet transmissions do not straddle the embedded epochs, and (ii) packets arrive back-to-back into the segmentation buffer during any interval (t (This leads to a conservative estimate of TCP throughput. See the discussion following Figure 8 below.) we can find the probability that a loss occurs during (t k ; t k+1 ), and the distribution of the TCP window at the time that timeout starts. Suppose this window is w, then the congestion avoidance threshold in the next recovery cycle will be m := d we. It will take approximately dlog 2 me round trip times (each of length \Delta) to reach the congestion avoidance threshold. Assuming that no loss occurs during the slow start phase (this is true if B max is not too small [11]), at k me, we can determine the distribution of ~ . With the above description in mind, define For k 1, loss occurs in (T loss occurs in (T the loss window is w and . For a particular realization of X k , we will write (a; b; d; loss occurs during (T k ; and for k 0 (7) Recalling the evolution of fT We now proceed to analyze the evolution of fX k ; k 0g. The bottleneck link modulating process, as mentioned earlier, is a two state Markov chain embedded at taking values in fr g. Let p 01 be the transition probabilities of the Markov chain. Notice that S also a Discrete time Markov chain (DTMC). Let Q be the transition probability matrix for (R k g. As explained above, given X is For particular can be determined using the probabilistic model of window evolution during the congestion avoidance phase. Consider the evolution of A k , the segmentation buffer queue process. If no loss occurs in (T k ; T k+1 ), s where N k is the increment in the TCP window in the interval, and is characterized as follows: During (T k ; T k+1 ), for each ACK arriving at the source (say, at time t), the window size increases by one with probability 1 . However, we further assume that the window size increases by one with probability 1 (where the probability does not change after every arrival but, instead, we use the window at Then, with this assumption, due to d arrivals to the source queue, the window size increases by the random amount N k . We see that for d ACKs, the maximum increase in window size is d. Let us define ~ N k such that ~ a+b+d ). Then , d)). We can similarly get recursive relations for B k+1 and Consider an example for explaining the evolution of X k to X k+1 . Let X (a; b; d; 2; 1), i.e., the source service rate is twice that of the bottleneck link server, and loss can take place. Further, d \Delta packets are in flight. These d packets arrive at the source queue, increase the window size by N k , and hence, min(a+d+N k ; \Delta) packets are transmitted into the bottleneck buffer (at most \Delta packets can be transmitted at rate 1 during an interval of length \Delta). If loss will occur. For a given b and d, we can compute the range of N k for which Equation 11 is satisfied. Suppose that loss occurs for N k x( 0). Then, ix Let us define Prfwindow achieved is w loss occurs in (T k We can compute this quantity in a manner similar to that outlined for the computation of p(x). When no loss occurs, U k is given by Equation 8. When loss occurs, given X the next cycle begins after the recovery from loss which includes the next slow start phase. Suppose that the window was 2m when loss occured. Then, the next congestion avoidance phase will begin when the TCP window size in the slow start phase after loss recovery reaches m. This will take dlog 2 me cycles. At the end of this period, the state of various queues is given by m). The channel state at the start of the next cycle can be described by the transition probability matrix of the modulating Markov chain. Hence, me with probability p(x):ff(x; 2m) (14) and From the above discussion, it is clear that given X k , the distribution of X k+1 can be computed without any knowledge of its past history. Hence, fX k ; k 0g is a Markov chain. Further, given T k and X k , the distribution of T k+1 can be computed without any knowledge of its past history. Hence, the process f(X k ; T k ); k 0g is a Markov Renewal Process (MRP) (See [17]). It is this MRP that is our model for TCP/ABR. 3.3 Computation of Throughput Given the Markov Renewal Process f(X we associate with the kth cycle that accounts for the successful transmission of packets. Let (x) denote the stationary probability distribution of the Markov chain fX k ; k 0g. Denote by fl TCP=ABR , the throughput of TCP over ABR. Then, by the Markov renewal-reward theorem ([17]), we have denotes the expectation w.r.t. the stationary distribution (x). The distribution (x) is obtained from the transition probabilities in Section 3.2. We have x is the expected reward in a cycle that begins with B(x) and D(x) the values of A, B and D in the state x. Then, in an interval (T k where no loss occurs, we take Thus for lossless intervals the reward is the number of acknowledgements returned to the source; note that this actually accounts for packets successfully received by the receiver in previous intervals. Loss occurs only if the ABR source is sending at the high rate and the link is transmitting at the low rate. When loss occurs in (T k \Delta), we need to account for the reward in the interval starting from T k until T k+1 when slow-start ends. Note that at T k the congestion window is A(x) D(x). The first component of the reward is D(x); all the B(x) buffered packets will result in ACKs, causing the left edge of the TCP window to advance. Since the link rate is half the source rate, loss will occur when 2(B packets enter the link buffer from the ABR source; these packets succeed and cause the left edge of the window to further advance. Further, we assume that the window grows by 1 in this process; hence, following the lost packet, at most A(x) packets can be sent. Thus we bound the reward before timeout occurs by D(x) loss and timeout, the ensuing slow-start phase successfully transfers some packets (as described earlier). Hence, an upper bound on the "reward" when loss occurs is A(x) the summation index w being over all window sizes. Actually, this is an optimistic reward as some of the packets will be transmitted again in the next cycle even though they have successfully reached receiver. We could also have a conservative accounting, where we assume that if loss occurs, all the packets transmitted in that cycle are retransmitted in future cycles. In the numerical results, we shall compare the throughputs with these two bounds. It follows that x Similarly we have x where U(x) is the mean cycle length when x at the beginning of the cycle. From the analysis in Section 3.2, it follows that Hence, x 3.4 TCP without ATM/ABR Without the ABR rate control, the source host would transmit at the full rate of its link; we assume that this link is much faster than the bottleneck link and model it as Constant Rate Arrival Process Figure server queue with time varying service capacity, being fed by a constant rate source. infinitely fast. The system model is then very similar to the previous case, the only difference being that we have eliminated the segmentation buffer. The assumptions we make in this analysis, however, lead to an optimistic estimate of the throughput. The analysis is analogous to that provided above. 4 TCP/ABR with Effective Capacity Feedback We now develop another kind of rate feedback. To motivate this approach, consider a finite buffer single server queue with a stationary ergodic service process (see Figure 6). Suppose that the ABR source sent packets at a constant rate. Then, we would like to find that rate which maximizes TCP throughput. Hence, let the input process to this queue be a constant rate deterministic arrival process. Given the buffer size B max and a desired Quality of Service (QoS) (say a cell loss probability ffl), we would like to know the maximum rate of the arrival process such that the QoS guarantee is met. We look at a discrete time approach to this problem (see [16]); in practice, the discrete time approach is adequate as the rate feedback is only updated at multiples of some basic measurement interval. Consider a slotted time queueing model where we can service C i packets in slot i and the buffer can hold B max packets. fC i g is a stationary and ergodic process; let EC be the mean of the process and C min be the minimum number of packets that can be served per slot. A constant number of packets (denoted by fl) arrive in each slot. We would like to find fl max such that the desired QoS (cell loss probability ffl) is achieved. In [16], the following asymptotic condition is considered. If X is a random variable that represents the stationary queue length, then, with lim log i.e., for large B max the loss probability is better then e \GammaffiB max . It is shown that this All logarithms are taken to the base e performance objective is met if lim log Ee \Gammaffi For the desired QoS we need . Let us denote the expression on the right hand side of Equation 25 as \Gamma eff . Then, \Gamma eff can be called the effective capacity of the server. which is what we intuitively expect. For all other values of ffl, \Gamma eff 2 (C min ; EC). Let us apply this effective capacity approach to our problem. Let the ABR source (see Figure adapt to the effective bandwidth of the bottleneck link server. In our anal- ysis, we have assumed a Markov modulated bottleneck link capacity, changes occurring at most once every \Delta units of time, \Delta being the round trip propagation delay. Hence, we have a discrete time model with fl being the number of packet arrivals to the bottle-neck link in \Delta units of time and C i being the number of packets served in that interval. We will compute the effective capacity of the bottleneck link server using Equation 25. However, before we can do this, we still need to determine the desired QOS, i.e, ffl or equivalently, ffi. To find ffi, we conduct the following experiment. We let the ABR source transmit at some constant rate, say ; For a given Markov modulating process, we find that which maximizes TCP throughput. We will assume that this is the effective capacity of the bottleneck link. Now, using Equation 25, we can find the smallest ffi that results in an effective capacity of this . If the value of ffi so obtained turns out to be consistent for a wide range of Markov modulating processes, then we will use this value of ffi as the QoS requirement for TCP over ABR. The above discrete time queueing model for TCP over ABR can be analyzed in a manner analogous to that in Section 3.2. We find from the analysis that for several sets of parameters, the value of ffi which maximizes TCP throughput is consistently very large (about 60-70). This is as expected since TCP performance is very sensitive to loss. 4.1 Algorithm for Effective Capacity Computation In practice, we do not know a priori the statistics of the modulating process. Hence, we need an on-line method of computing the effective bandwidth. In this section, we develop an algorithm for computing the effective capacity of a time varying bottleneck link carrying TCP traffic. The idea is based on Equation 25, and the observation at the end of the previous section that ffi is very large. Averages time Figure 7: Schematic of the windows used in the computation of the effective capacity bsed rate feedback. We take the measurement interval to be s time units; s is also the update interval of the rate feedback. We shall approximate the expression for effective bandwidth in Equation 25 by replacing n !1 by a large finite M . log Ee \Gammaffi What we now have is an effective capacity computation performed over Ms units of time. We will assume that the process is ergodic and stationary. Hence, we approximate the expectation by the average of N sets of samples, each set taken over Ms units of time. Note that since the process is stationary and ergodic, the N intervals need not be disjoint for the following argument to work. Then, denoting C ij as the ith link capacity value in the jth block of M intervals (j 2 f1; Ng), we have logN e \Gammaffi log 1 log e \Gammaffi As motivated above, we now take ffi to be large. This yields log e \Gammaffi(min We notice that this essentially means that we average capacities over N sliding blocks, each block representing Ms units of time, and feed back the minimum of these values (see Figure 7). The formula that has been obtained (Equation 31) has a particularly simple form. The above derivation should be viewed more as a motivation for this formula. The formula, however, has independent intuitive appeal; see below. In the derivation it was required that M and N should be large. We can, however, study the effect of the choice of M and N (large or small) on the performance of effective capacity feedback. This is done in Section 6, where we also provide guidelines for selecting values of M and N under various situations. The formula in Equation 31 is intuitively satisfying; we will call it EFFCAP feedback. Consider the case when the network changes are very slow. Then, all N values of the average capacity will be the same, and each one will be equal to the capacity of the bottleneck link. Hence, the rate that is fed back to the ABR source will be the instantaneous free capacity of the bottleneck link; i.e., in this situation EFFCAP is the same as INSTCAP. When the network variations are very fast, EFFCAP will be the mean capacity of the bottleneck link which is what should be done to get the best throughput. Hence, EFFCAP behaves like INSTCAP for slow network changes and adapts to the mean bottleneck link capacity for fast changes. For intermediate rates of changes, EFFCAP is (necessarily) conservative and feeds back the minimum link rate. There is another benefit we could get by using EFFCAP. As EFFCAP assumes a large value of ffi, this means the the cell loss probability ffl is very small. This implies that the TCP throughput will essentially be the ABR throughput. Thus EFFCAP when used along with the Minimum Cell Rate feature of the ABR service can guarantee a minimum throughput to TCP connections. Some of our simulation results to be presented demonstrate this. 5 Numerical and Simulation Results In this section, we first compare our analytical results for the throughput of TCP, without ABR and with ABR with INSTCAP feedback, with simulation results from a hybrid TCP simulator involving actual TCP code, and a model for the network implemented in the loopback driver of a Linux Pentium machine. We show that the performance of TCP improves when ABR is used for end-to-end data transport below TCP. We then study the performance of the EFFCAP scheme and compare it with the INSTCAP scheme. Efficiency Mean time per state (rtd) "Conservative Analysis, 10 packets" "Conservative Analysis, 12 packets" "Optimistic Analysis, 12 packets" "Testbed results, 10 packets" "Testbed results, 12 packets" Figure 8: Analysis and Simulation results: INSTCAP feedback. Throughput of TCP over ABR: The round trip propagation delay is 40 time units. The bottleneck link buffers are either 10 or 12 packets. Notice that, for / ? 80, the lowest two curves are test-bed results, the uppermost two curves are the optimistic analysis, and the middle two curves are the conservative analysis. 5.1 Instantaneous Rate Feedback Scheme We recall from the previous section that the bottleneck link is Markov modulated. In our analysis, we have assumed that the modulating chain has two states which we call the high state and the low state. In the low state, with some link capacity being used by higher priority traffic, the link capacity is some fraction of the link capacity in the high state (where the full link rate is available). In the set of results that we present in this section, we will assume that this fraction is 0.5. Further, we will also assume that the mean time in each state is the same, i.e., the Markov chain is symmetric. We denote the mean time in each state by , and denote the mean time in each state normalized to \Delta by /, i.e., / := . For example, if \Delta is 200msec, then means that the mean time per state is 400msec. Note that our analysis only applies to / ? 1. A large value of / means that the network changes are slow compared to the round trip propagation delay (rtd) , whereas / !! 1 means that the network transients occur several times per round trip time. In the Linux kernel implementation of our network simulator, the Markov chain can make transitions at most once every 30msec. Hence we take this also to be the measurement interval, and the explicit rate feedback interval (i.e., We denote one packet transmission time at the bottleneck link in the high rate state as one time unit. Thus, in all the results presented here, the packet transmission time in the low rate state is 2 time units. We plot the bottleneck link efficiency vs. mean time that it spends in each state (i.e. We define efficiency as the throughput as a fraction of the mean capacity of the bottleneck link. We include the TCP/IP headers in the throughput, but account for ATM headers as overhead. We use the words throughput and efficiency interchangeably. With the modulating Markov chain spending the same time in each state, the mean capacity of the link is 0.75. Figure 8 shows the throughput of TCP over ABR with the INSTCAP scheme 2 . Here, we compare an optimistic analysis, a conservative one (see Section 3.3), and the test-bed (i.e., simulation) results for different buffer sizes. In our analysis, the processes are embedded at multiples of one round trip propagation delay, and the feedback from the bottleneck link is sent once every rtd. This feedback reaches the ABR source after one round trip propagation delay. In the simulations, however, feedback is sent to the ABR source every 30msec. This reaches the ABR source after one round trip propagation delay. In Figure 8, we can see that, except for very small /, the analysis and the simulations match to within a few percent. Both the analyses are less than the observed throughputs by about 10-20% for small /. This can be explained if we note that in our model, we assume that packets arrive (leave) back to back to (from) the ABR source. When a rate change occurs at the bottleneck link, as the packets arrive back to back, and the source sends at twice the rate of the bottleneck link (in our example); for every two packets arriving to the bottleneck link, one gets queued. However, in reality, the packets need not arrive back to back and hence, the queue buildup is slower. This means that the probability that packet loss occurs at the bottleneck link buffer is actually lower than in our analytical model. This effect becomes more and more significant as the rate of bottleneck link variations increase. However, we observe from the simulations that this effect is not significant for most values of /. Figure 9 shows the throughput of TCP without ABR. We can see that the simulation results give a throughput of upto 20% less than the analytical ones. This occurs due to two reasons. Even if / !1, the throughput of TCP over ABR will not go to 1 because of ATM overheads. For every 53 bytes transmitted, there are 5 bytes of ATM headers. Hence, the asymptotic throughput is approximately 90%. Efficiency Mean time per state (rtd) Figure 9: Analysis and Simulation results; throughput of TCP without ABR : The round propagation delay is 40 time units. The bottleneck link buffers are either 10 or 12 packets. We observe that TCP is sensitive to bottleneck link buffer size changes. (i) We assumed in our analysis that no loss occurs in the slow-start phase. It has been shown in [11] that if the bottleneck link buffer is less than 1of the bandwidth-delay product (which corresponds to about 13 packets or 6500 byte buffer), loss will occur in the slow-start phase. (ii) We optimistically compute the throughput of TCP by using an upper bound on the "reward" in the loss cycle. We see from Figures 8 and 9 that ABR makes TCP throughput insensitive to buffer size variations. However, with TCP alone, there is a significant worsening of throughput with buffer reduction. This can be explained by the fact that once the ABR control loop has converged, the buffer size is immaterial as no loss takes place when source and bottleneck link rate are the same. However, without ABR, TCP loses packets even when no transients occur. It is useful to observe that since the times in the above curves are normalized to the packet transmission time (in the high rate state), results for several different ranges of parameters can be read off these curves. To give an example, if the link has a capacity of 155Mbps during its high rate state, and TCP packets have a size of 500 bytes each, then one time unit is 25:8sec. The round trip propagation delay (\Delta) is means that changes in link bandwidth occur Efficiency Mean time per state (rtd) Alone, 8 packets" Alone, 12 packets" Figure 10: Simulation results for TCP with and without ABR (INSTCAP feedback) for small values of /. The rtd is 40 time units. on an average, once every 103.2msec. Consider another example where the link capacity is 2Mbps during the high rate period. Let the packet size by 1000bytes. Then, the delay here corresponding to 40 time units is 160msec. here corresponds to changes occurring once every 16 seconds. These two examples illustrate the fact the the curves are normalized and can be used to read off numbers for many scenarios. 3 From Figures 8 and 9, we can see that the performance of TCP improves by about 20% when ABR is employed for data transport, INSTCAP feedback is used, and the changes in link rate are slow. This improvement in performance with ABR is due to the fact that ABR pushes the congestion to the network edge. After the TCP window grows beyond the point where the pipe gets full, i.e., W t r , the packets start queueing in the ABR segmentation buffer, which being on the end system, is very large. The window size increases till the maximum window size advertised by the receiver is reached. No loss occurs in the network as the ABR source sends at just the right rate to the bottleneck link. Hence, we end up with the TCP window size fixed at the maximum window size and the pipe is always full. The assumptions in our analysis render it inapplicable for very small /. Figure 10 compares the simulation results for TCP with ABR (INSTCAP feedback) and without ABR for various buffer sizes. These results are for / starting with going 3 Note however that \Delta is an absolute parameter in these curves, since it governs the round trip "pipe". Thus, although / is normalized to \Delta, the curves do not yield values for fixed / and varying \Delta. down to We note that even though the throughput improvement due to ABR is not great for small /, the performance of TCP does not significantly worsen due to ABR. In the next section, we will see how a better rate feedback in ABR can result in a distinct improvement of TCP/ABR throughput even for this range of /. We see from Figure 10 that when / becomes less than 1, the throughput of TCP increases. This can be explained by the fact that the rate mismatch occurs for an interval of time less than one round trip propagation delay. As a result, the buffer size required to handle the overload becomes less. As / becomes very small, each packet is sent at a different rate and hence, the ABR source effectively sends at the mean capacity. Then loss very rarely occurs as the buffers can handle almost all rate mismatches and hence, the throughput increases. 5.2 Comparison of EFFCAP and INSTCAP Performance: Simulation Efficiency Mean time per state (rtd) "Effective Capacity, 8 packets" "Effective Capacity, 10 packets" "Effective Capacity, 12 packets" "Instantaneous rate feedback, 8 packets" "Instantaneous rate feedback, 10 packets" "Instantaneous rate feedback, 12 packets" Figure 11: Simulation results; Comparison of the EFFCAP and INSTCAP feedback schemes for TCP over ABR for various bottleneck link buffers (8-12 packets). As before, the rtd is 40 time units. Here, Figure 7). In this figure, we compare their performances for relatively large /. In Figure 11, we use results from the test-bed to compare the relative performances of the EFFCAP and INSTCAP feedback schemes for ABR. Recall that the EFFCAP algorithm has two parameters, namely M , the number of samples used for each block average, and N , the number of blocks of M samples over which the minimum is taken. Efficiency Mean time per state (rtd) "Effective Capacity, 8 packets" "Effective Capacity, 10 packets" "Effective Capacity, 12 packets" "Instantaneous rate feedback, 8 packets" "Instantaneous rate feedback, 10 packets" "Instantaneous rate feedback, 12 packets" Figure 12: Simulation results; Comparison of the EFFCAP and INSTCAP feedback schemes for TCP over ABR for various bottleneck link buffers (8-12 packets). As before, the rtd is 40 time units. Here, Figure 7). In this figure, we compare their performances for small values of /. In this figure, the EFFCAP scheme uses i.e, we average over one round trip propagation delay 4 worth of samples. We also maintain a window of 8 rtd worth of averages, i.e, we maintain averages over which the bottleneck link returns the minimum to the ABR source. The source adapts to this rate. In the case of the INSTCAP scheme, in the simulation, the rate is fed back every 30msec. We can see from Figure 11 that for large /, the throughput with EFFCAP is worse than that with the INSTCAP scheme by about 3-4%. This is because of the conservative nature of the EFFCAP algorithm (it takes the minimum of the available capacity over several blocks of time in an interval). However, we can see from Figure 12 that for small /, the EFFCAP algorithm improves over the INSTCAP approach by 10-20%. This is a significant improvement and it seems worthwhile to lose a few percent efficiency for large / to gain a large improvement for small /. To summarize, in Figures 13 and 14, we have plotted the throughput of TCP over ABR using the two different feedback schemes. We have compared these results with the throughput of TCP without ABR. We can see that the throughput of TCP improves 4 A new sample is generated every 30msec. The rtd is 200msec in this example. Hence, 6.667 which we round up to 7. Efficiency Mean time per state (rtd) "Effective Capacity, 10 packets" "Instantaneous rate feedback, 10 packets" Figure 13: Simulation results; Comparison of throughput of TCP over ABR with effective capacity scheme, instantaneous rate feedback scheme and TCP without ABR for a buffer of 10 packets, the other parameters remaining the same as in other simulations.0.50.60.70.80.9 Efficiency Mean time per state (rtd) "Effective Capacity, 10 packets" "Instantaneous rate feedback, 10 packets" Figure 14: Simulation results; Comparison of throughput of TCP over ABR with effective capacity scheme, instantaneous rate feedback scheme and TCP without ABR for a buffer of 10 packets, the other parameters remaining the same as in other simulations. if ABR is employed for link level data transport. These plots clearly brings out the merits of the effective capacity scheme. We can see that for all values of /, the EFFCAP scheme performs considerably better than TCP alone. For large /, we have a throughput improvement of about 30% while for small /, the improvement is of the order of 10-15%. Further, while being adaptive to /, EFFCAP succeeds in keeping the TCP throughput better than the minimum link rate, which INSTCAP fails to do (for small /). Thus an MCR in the ABR connection may be used to guarantee a minimum TCP throughput. 6 Choice of M and N for EFFCAP 6.1 Significance of M and N We begin by recalling the results shown in Figures 11 and 12. From these figures, we can identify three broad regions of performance in relation to /. For / very large (/ ? 50), the rate mismatch occurs for a small fraction of . Also the rate mismatches are infrequent, implying infrequent losses, thereby increasing the throughput. Hence, it is sufficient to track the instantaneous available capacity by choosing small values of M and N . This is verified from Figure 11 which shows that the INSTCAP feedback performs better in this region. On the other hand, when is a small fraction of \Delta (/ ! 0:2) there are frequent rate mismatches but of very small durations as compared to \Delta. This reduces the buffer requirement, and hence losses occur rarely. Because of the rapid variations in the capacity, even a small M provides the mean capacity. Also all the N averages roughly equal the mean capacity. Thus, the source essentially transmits at the mean capacity in EFFCAP as well as INSTCAP feedback. Hence a high throughput for both feedbacks is seen from Figure 12. For the intermediate values of / (0:5 ! / ! 20), the throughput drops substantially for both types of feedback. For these values of /, is comparable to \Delta. Hence rate mismatch is frequent, and persists relatively longer causing the buffer to build up to a larger value. This leads to frequent losses. Because of frequent losses the throughput is adversely affected by TCP's blind adaptation window control. In this range, we expect to see severe throughput loss for sessions with large \Delta. Therefore, in this region, we need to choose M and N properly; essentially, to avoid rate mismatch and hence loss, the capacity estimate should yield the minimum capacity, implying the need for small M and large N . A small M helps to avoid averaging over many samples, and a large N helps to pick out the minimum. The selection of M and N cannot be based on the value of / alone however. \Delta is an absolute parameter in TCP window control and has a major effect on TCP throughput, and hence on the selection of M and N . This can be seen from the results in Section 6.2. 6.2 Simulation Results and Discussion Simulations were carried out on the hybrid simulator that was also used in Section 5. As before, the capacity variation process is a two state Markov chain. In the high state, the capacity value is 100KB/sec (KB= Kilo Bytes) while in the low state it is 50KB/sec. The mean capacity is thus 75KB/sec. In all the simulations, the measurement and feedback 30ms. Throughput is measured for data transfer of a 10MB file; the average throughput for 4 file transfers is reported. The TCP packet size is 500 bytes and the maximum TCP window is 32KB. M denotes the number of measurement samples in the averaging window. The effective capacity method uses N such overlapping windows (see Figure 7) to determine the minimum average. Thus N corresponds to the 'memory' of the algorithm. We introduce the following notation in the simulation results. means that each average is calculated over measurement intervals corresponding to the round trip propagation delay, that is, d \Delta means that e averages are compared (or the memory of the algorithm is k round trip times). For example, let 200ms and (i.e., minimum of 49 averages). 6.2.1 Study of N In this section, we study the effect of N on the throughput by carrying out two sets of simulations. Case 1: Fixed \Delta; varying Figure 15 shows the effect of N on the throughput of a TCP session with a given \Delta, when (or equivalently the rate of capacity variation) is varying. These results corroborate the discussion at the beginning of Section 6.1 for fixed \Delta; for large / (/ ? 60), a small value of N performs slightly better, whereas when / is very small (/ ! 0:3), where the throughput increases steeply, there is negligible throughput gain by increasing N . When 0:3 ! / ! 1, as expected, an improvement is seen for larger N . But for only a slight improvement is seen with varying N . Efficiency Mean Time per State (rtd) Efficiency Mean Time per State (rtd) Figure 15: Efficiency variation with / for increasing values of N . is varied from 32ms to 40s. The link buffer is 10 pkts or 5000KB. The advantage of choosing a larger value of N in the intermediate range of / (0:2 clearly seen. Case 2: Fixed ; varying \Delta0.30.40.50.6 Efficiency Mean Time per State (rtd) Figure Efficiency vs. is fixed to 1000ms. \Delta is varied (right to left) from 50ms to 500ms. M : \Delta; link buffer=10 Efficiency Mean Time per State (rtd) Figure 17: Efficiency vs. is fixed to 100ms. \Delta is varied (right to left) from 50ms to 500ms. M : \Delta, link buffer=10 pkts. Figures and 17 show the Efficiency variation with / for different values of N when is fixed and \Delta is varied. We note that, according to the notation, N is different for different \Deltas on a N : k\Delta curve. For example, N on the N : 4\Delta curve for (i.e., the memory of the algorithm is 4 rtds) 100ms is respectively 6 and 12. We notice that compared to Figure 15, Figures 16 and 17 show different Efficiency variations with /. This is because, in the former case is varied and \Delta kept constant, whereas in the latter case is fixed and \Delta varied. As indicated in Section 6.1, \Delta is an absolute parameter which affects the throughput Figure 16 corresponds to \Delta=500ms, and Figure 17 corresponds to 50ms). The considerable throughput difference demonstrates the dependence on the absolute value of \Delta. It can be observed that the throughput for larger values of \Delta is lower than that for small values of \Delta. This difference is because a TCP session with larger \Delta needs a larger window 5 to achieve the desired throughput. A single packet loss causes the TCP window to drop. After a packet loss, it takes a longer time for a session with larger \Delta to rebuild the window than a session with a smaller \Delta. In the intermediate range of /, as explained in Section 6.1, losses are frequent and high. Hence, the throughput of a session with large \Delta is severely affected. In Figure 17, a substantial improvement in the throughput is seen as the memory of the EFFCAP algorithm increases. A larger memory gives better throughput over a wider range of \Delta as compared to lower values. The reason for these observations is as follows. For a given \Delta, as N increases we are able to track the minimum capacity value better. The minimum capacity is 50KB/sec, which is 66% of the mean capacity 75KB/sec. Hence, as N increases we see Efficiency increasing above 0.6. that for small / , the average over \Delta yields the average rate, whereas for large / the average over \Delta yields the peak or minimum rate. Thus for large /, the minimum over just few \Deltas (4 to 6) is adequate to yield a high throughput, whereas for small / many more averages need to be minimized over to get the minimum rate. Figure shows that for / ! 8, larger values of N improve the throughput according to the argument given above. When we see that smaller N performs better, but the improvement is negligible. The conclusions that can be drawn from above results are as follows. The choice of N is based on / and \Delta. For very large / (/ ? 20) N should be small (along with M ); the limiting case being 1. For very small / (/ ! 0:2) N does not matter much. In the intermediate range, a large N is better. 6.2.2 Study of M In this section we study the performance of M , the averaging parameter. We have already seen from Figure 11 that for / ? 60, a small value of M should be selected. We now study the lower ranges of /. 5 Actually TCP window depends directly on the bandwidth delay product. In all the numerical results however the link speed is fixed, hence the window depends simply on \Delta. Efficiency Window Size (measurement intervals) RTT 50ms RTT 100ms Efficiency Window Size (measurement intervals) RTT 50ms RTT 100ms RTT 200ms Figure Efficiency variation with varying M . is set to 1000ms, and three \Delta values- 50ms, 100ms and 200ms are considered. In the left-hand graph, N is set according to and in the right-hand N : 12\Delta. M is varied from 1 to 10, i.e., the averaging interval is varied from 30ms to 300ms. The link buffer is 10 pkts. / ranges from 5 (for Efficiency Window Size (measurement intervals) RTT 50ms RTT 100ms RTT 200ms0.50.70.9 Efficiency Window Size (measurement intervals) RTT 50ms RTT 100ms RTT 200ms Figure 19: Efficiency variation with varying M . is set to 100ms, and three \Delta values- 50ms, 100ms and 200ms are considered. In the left-hand graph, N is set according to and in the right-hand N : 12\Delta. M is varied from 1 to 10. The link buffer is 10 pkts. / ranges from 0.5 (for To study the effect of M , we vary the window size for two N settings, so that the effect of N could be differentiated. M is varied from 1 to 10 measuring intervals. The results are shown in Figure Figure 19 100ms). The values of \Delta are 50ms, 100ms and 200ms. Thus the range of /(= under consideration is 0.5 to 20. We have already found that M needs to be small for / ? In the range under consideration, however, we observe that the throughput is not very sensitive to M . In Figure 18, a clear advantage of increasing N is seen for 200ms. For ms in this figure, that is, we see a slight trend in decreasing throughput with increasing M . This is because \Delta is small and with small M it is possible to track the instantaneous capacity better for larger . For larger \Delta values and 1000ms the above mentioned trend is greatly reduced making the throughput insensitive to M . However, a slight decrease in the throughput is seen in case of ms when M takes larger values. The reason is as follows. As discussed in Section 6.1, larger value of M makes the response of the algorithm sluggish. Hence to track the minimum in the intermediate range of /, a larger N is needed. In this case N is fixed, hence the decrease for larger M . A similar effect is seen in the Figure 19 for \Delta=50ms and 100ms. In Figure 19, / is in the range 0.5 to 2. For the throughput is insensitive to the variation in M . Insensitivity is observed in the case of N : small / but for larger /, that is, \Delta=50ms and 100ms, a 10-15% decrease in the throughput is seen. The reason, as explained above, is the inability to track the minimum because of the smaller value of N . We conclude that in the intermediate range of /, the throughput is not very sensitive to M . For small \Delta and larger / (e.g. performs better since it is possible to track the instantaneous rate. In general, a small value of M improves the throughput in the intermediate range. For larger values of M , N needs to be increased to enable tracking of the minimum. 7 TCP/ABR with EFFCAP Feedback: Multiple Sessions Sharing a Bottleneck Link Throughput fairness is a major issue for multiple sessions sharing a bottleneck link. It is seen that TCP alone is unfair towards sessions that have larger round trip times. It may be expected however, that TCP sessions over ABR will get a fair share of the available capacity. In [14], the fairness of the INSTCAP feedback was investigated and it was shown that for slow variations of the available capacity, TCP sessions over ABR employing the INSTCAP feedback achieve fairness. In this section we study the fairness of TCP sessions over ABR with the EFFCAP feedback scheme In the simulations, we use 240ms as the round-trip time for Session 1 and 360ms for Session 2. The link buffer size is bytes. Denote by \Delta 1 and \Delta 2 , the round-trip times for Session 1 and Session 2 respectively. Other notations are as described earlier (subscripts denote the session number). In the following graphs / is (mean time per state of the Markov chain) divided by larger 360ms. Simulations are carried out by calculating the EFFCAP by two different ways as explained below. Case 1: Effective Capacity with In this case, we calculate the EFFCAP for each session independently. This is done by selecting M i proportional to \Delta i , that is (with a 30ms update interval) we select for Session 1 and 2. We take 132 (see section 6). EFFCAP i is computed with M i and N i ; session i is fedback 1of Figure 20 shows the simulation results. We see that for very small values of / 0:3), the sessions receive equal throughput. However, for 0:3 unfairness is seen towards the session with larger propagation delay. This can be explained from the discussion in Section 6.1. In this range of /, due to frequent rate mismatches and hence losses, TCP behavior is dominant. A packet drop leads to greater throughput decrease for a session with larger \Delta than for a session with smaller \Delta.0.250.350.450.550 Efficiency Mean time per state (larger rtd) session 1:240ms session 2:360ms0.250.350.450.550 Efficiency Mean time per state (larger rtd) session 1:240ms session 2:360ms Figure 20: Efficiency variation with / (mean time per state normalized to larger in the case of two sessions sharing a link. \Delta 1 is 240ms and \Delta 2 is 360ms. Link buffer is 18pkts. Each session is fed back the fair share (half) of the EFFCAP calculated. 7.2 Case 2: Effective Capacity with simulation M corresponds to the average of \Delta 1 and \Delta 2 , i.e., 300ms or 10 measurement intervals. Correspondingly, with choosing Efficiency Mean time per state (larger rtd) session 1:240ms session 2:360ms0.250.350.450.550 Efficiency Mean time per state (larger rtd) session 1:240ms session 2:360ms Figure 21: Efficiency variation with / (mean time per state normalized to larger in the case of two sessions sharing a link. \Delta 1 is 240ms and \Delta 2 is 360ms. Link buffer is 18pkts. averages. Each session is fed back the fair share (half) of the EFFCAP calculated. M and N this way, we are making rate calculation independent of individual round-trip times. We observe from Figure 21 that the EFFCAP calculated in this way yields somewhat better fairness than the scheme used in Case 1. We also see that better fairness is obtained even in the intermediate range of /. However, there is a drop in the overall efficiency. This is because the throughput of the session with smaller \Delta is reduced. Figures 22 and 23 show the comparison of TCP alone with TCP over ABR with EFFCAP feedback for a longer range of /. These curves include results from Figures 20 and 21. We see that for / ? 20, EFFCAP gives fairness to the sessions whereas TCP is grossly unfair to the session with larger \Delta. There is a slight decrease in the overall efficiency with TCP over ABR; but note that with TCP over ABR the link actually carries 10% more bytes (the ATM overhead) than with TCP alone! We also see from Figure even when / ! 20 which is not observed in the case of INSTCAP in [14]. Conclusions In this paper, we first developed an analytical model for a wide-area TCP connection over end-to-end ABR with INSTCAP feedback, running over a single bottleneck link with time-varying capacity. We have compared our analytical results for the performance of TCP without ABR, and with ABR (INSTCAP rate feedback) with results from a hybrid Efficiency Mean time per state (larger rtd) session 1: EFFCAP session 2: EFFCAP session 1: TCP alone session 2: TCP alone Figure 22: Comparison between Efficiency of sessions with TCP alone and TCP over ABR employing EFFCAP feedback (Case 1). \Delta 1 is 240ms and \Delta 2 is 360ms. Both the cases use a link buffer of pkts. simulation. We have seen that the analysis and simulation results for TCP over ABR match quite well, whereas the analysis overestimates the performance of TCP without ABR. Our results show that the throughput improvement by running TCP over ABR depends on the relative rate of capacity variation with respect to the round trip delay in the connection. For slow variations of the link capacity the improvement is significant (25% to 30%), whereas if the rate variations are comparable to the round trip delay then the TCP throughput with ABR can be slightly worse than with TCP alone. We have also proposed EFFCAP, an effective capacity based algorithm for rate feedback. We have simulated TCP over ABR with EFFCAP feedback and shown that, unlike INSTCAP feedback, EFFCAP succeeds in keeping the TCP throughput higher than the minimum bandwidth of the bottleneck link; see Figure 12. The EFFCAP computation involves two parameters M and N . The throughput variation with /, of a TCP session over ABR employing EFFCAP feedback, can be broadly divided into three regions. The selection of parameters M and thus, depends on the range of / as well as on the round-trip propagation delay. When / is very large (? 60) M and N need to be small. Ideally Efficiency Mean time per state (larger rtd) session 1: EFFCAP session 2: EFFCAP session 1: TCP alone session 2: TCP alone Figure 23: Comparison between Efficiency of sessions with TCP alone and TCP over ABR employing EFFCAP feedback (Case 2). \Delta 1 is 240ms and \Delta 2 is 360ms. Both the cases use a link buffer of pkts. INSTCAP) performs the best in this region. When / is small (! 0:3), it is sufficient for the source to send at the mean bottleneck rate; for such /, because of rapid capacity variations, the mean capacity can be measured by selecting a large M . However, as / becomes very small, the choice of M and N does not matter much. When / is in the intermediate range (0:5 ! / ! 20), the throughput decreases due to frequent losses, hence M and N need to be selected carefully. In this region a large value of N improves the throughput whereas the performance is insensitive towards M . In general, a small value of M performs better for a given value of N . The throughput drop in this region can be compensated by choosing a large buffer, thereby, reducing the losses. In summary, as a broad guideline, for the buffer sizes that we studied, using provides good throughput performance for TCP over ABR over a wide range of / and \Delta values. In the case of multiple sessions, EFFCAP feedback provides fairness over a wider range of / then INSTCAP. EFFCAP feedback based on the average round-trip times of the sessions is seen to provide fairness even when the capacity variations are in the intermediate range. This is an advantage over INSTCAP which is fair only when the rate variations are slow compared to \Delta [14]. --R The ATM Forum Traffic Management Specification Version 4.0 "A Simulation of TCP Performance in ATM Networks" " Impact of ATM ABR Control on the Performance of TCP-Tahoe and TCP-Reno" "Congestion avoidance and control" "Modified TCP Congestion Avoidance Algorithm" " The ERICA Switch Algorithm for ABR Traffic Management in ATM Networks," " Performance of TCP/IP over ABR Service on ATM Networks" " Buffer Requirements for TCP/IP over ABR" " Performance of TCP over ABR on ATM Backbone and with Various VBR Traffic Patterns" "Comparative performance analysis of versions of TCP in a local network with a lossy link" "The Performance of TCP/IP for Networks with High Bandwidth Delay Products and Random Loss, " "TCP over ATM: ABR or UBR" "Dynamics of TCP Traffic over ATM Networks" "TCP over End-to-End ABR: A Study of TCP Performance with End-to-End Rate Control and Stochastic Available Capacity" Slow Start, Congestion Avoidance, Fast Retransmit, and Fast Recovery Algorithms" "Effective Bandwidths: Call Admission, Traffic Policing and Filtering for ATM Networks" Stochastic Modeling and the Theory of Queues Prentice Hall --TR The performance of TCP/IP for networks with high bandwidth-delay products and random loss Comparative performance analysis of versions of TCP in a local network with a lossy link Analysis of source policy and its effects on TCP in rate-controlled ATM networks TCP over wireless with link level error control The ERICA switch algorithm for ABR traffic management in ATM networks Modeling TCP Reno performance A new approach for asynchronous distributed rate control of elastic sessions in integrated packet networks --CTR Aditya Karnik , Anurag Kumar, Performance of TCP congestion control with explicit rate feedback, IEEE/ACM Transactions on Networking (TON), v.13 n.1, p.108-120, February 2005 Jung-Shian Li , Chuan-Gang Liu , Cheng-Yu Huang, Achieving multipoint-to-multipoint fairness with RCNWA, Journal of Systems Architecture: the EUROMICRO Journal, v.53 n.7, p.437-452, July, 2007 Ahmed E. Kamal, Discrete-time modeling of TCP Reno under background traffic interference with extension to RED-based routers, Performance Evaluation, v.58 n.2+3, p.109-142, November 2004

TCP over ABR;congestion control;TCP performance

504650

Jitter control in QoS networks.

We study jitter control in networks with guaranteed quality of service (QoS) from the competitive analysis point of view: we propose on-line algorithms that control jitter and compare their performance to the best possible (by an off-line algorithm) for any given arrival sequence. For delay jitter, where the goal is to minimize the difference between delay times of different packets, we show that a simple on-line algorithm using a buffer of B slots guarantees the same delay jitter as the best off-line algorithm using buffer space B/2. We prove that the guarantees made by our on-line algorithm hold, even for simple distributed implementations, where the total buffer space is distributed along the path of the connection, provided that the input stream satisfies a certain simple property. For rate jitter, where the goal is to minimize the difference between inter-arrival times, we develop an on-line algorithm using a buffer of size 2B + h for any h 1, and compare its jitter to the jitter of an optimal off-line algorithm using buffer size B. We prove that our algorithm guarantees that the difference is bounded by a term proportional to B/h.

Introduction The need for networks with guaranteed quality of service (QoS) is widely recognized today (see, e.g., [8, 11]). Unlike today's ``best effort'' networks such as the Internet, where the user has no guarantee on the performance it may expect from the network, QoS networks guarantee the end-user application a certain level of performance. For example, ATM networks support guaranteed QoS in various parameters, including end-to-end delay and delay jitter (called Cell Transfer Delay and Cell Delay Variation, respectively [5, 12]). Jitter measures the variability of delay of packets in the given stream, which is an important property for many applications (for example, streaming real-time applications). Ideally, packets should be delivered in a perfectly periodic fashion; however, even if the source generates an evenly spaced stream, unavoidable jitter is introduced by the network due to the variable queuing and propagation delays, and packets arrive at the destination with a wide range of inter-arrival times. The jitter increases at switches along the path of a connection due to many factors, such as conflicts with other packets wishing to use the same links, and non-deterministic propagation delay in the data-link layer. Jitter is quantified in two ways. One measure, called delay jitter, bounds the maximum difference in the total delay of different packets (assuming, without loss of generality, that the abstract source is perfectly periodic). This approach is useful in contexts such as interactive communication (e.g., voice and video tele-conferencing), where a guarantee on the delay jitter can be translated to the maximum buffer size needed at the destination. The second measure, called rate jitter, bounds the difference in packet delivery rates at various times. More precisely, rate jitter measures the difference between the minimal and maximal inter-arrival times (inter-arrival time between packets is the reciprocal of rate). Rate jitter is a useful measure for many real-time applications, such as a video broadcast over the net: a slight deviation of rate translates to only a small deterioration in the perceived quality. Another important reason for keeping the jitter under control comes from the network management itself, even if there are no applications requiring jitter guarantees. For example, it is well known that traffic bursts tend to build in the network [8, 15]. Jitter control provides a means for regulating the traffic inside the network so that the behavior of internal traffic is more easily manageable. A more subtle argument in favor of jitter control (given by [17]) proceeds as follows. When a QoS network admits a connection, a type of "contract" is agreed upon between the network and the user application: the user is committed to keeping its traffic within certain bounds (such as peak bandwidth, maximal burst size etc.), and the network is committed to providing certain service guarantees (such as maximal delay, loss rate etc. Since the network itself consists of a collection of links and switches, its guarantees must depend on the guarantees made by its components. The guarantees made by a link or a switch, in turn, are contingent on some bounds on the locally incoming traffic. As mentioned above, unless some action is taken by the network, the characteristics of the connection may in fact get worse for switches further down the path, and thus they can only commit to lower QoS. Jitter control can be useful in allowing the network to ensure that the traffic incoming into a switch is "nicer," and get better guarantees from the switch. Jitter control implementation is usually modeled as follows [17, 8]. Traffic incoming into the switch is input into a jitter-regulator, which re-shapes the traffic by holding packets in an internal buffer. When a packet is released from the jitter-regulator, it is passed to the link scheduler, which schedules packet transmission on the output link. In this work we focus on studying jitter-regulators. Nature of our results. Before we state concrete results, we would like to explain the insight we seek. Prior to our work, performance of jitter control algorithm was measured either by worst-case behavior, or under statistical assumptions. Thus the properties of the algorithms were either deterministic (given deterministic worst-case assumptions on the input stream), or probabilistic (given stochastic assumptions on the input stream). In this work, we prove relativistic guarantees: we compare the performance of the algorithm in question to the performance of the best possible algorithm, which we treat as an adversary we compete against. The adversary algorithm is not assumed to be constrained by the on-line nature of the problem: it is assumed to produce the best possible output for the given input, even if the best output may be computable only in hindsight (hence the adversary algorithm is sometimes called the off-line algorithm). Algorithms whose performance can be bounded with respect to the performance of an off-line adversary are called competitive [10, 7, 1]. We argue that proving that an algorithm is competitive is meaningful, and sometimes superior, to proving deterministic or stochastic guarantees: first, deterministic or stochastic guarantees say nothing about the case where the underlying assumptions do not hold for some reason (even worse, the underlying assumptions-in particular, tractable stochastic assumptions- are notoriously hard to justify). On the other hand, a competitive algorithm does not assume anything about the input, and therefore its guarantees are more robust in this sense. Secondly, worst-case guarantees usually do not say much about individual cases: for example, an algorithm may be called deterministically optimal even if it performs always as bad as the worst case; competitive algorithms, by contrast, are guaranteed to do relatively well on each and every instance. Thirdly, if we add an assumption about the input sequence, the relativistic guarantee would immediately translate to a specific deterministic guarantee. We remark that unlike conventional competitive analysis, in most cases we shall compare the performance of our on-line algorithms to the performance of an (optimal, off-line) adversary which is restricted to use less buffer space. For example, we prove statements such as "an algorithm Z using space B produces jitter which never more than the jitter produced by an optimal algorithm, for the given arrival sequence, using space B=2." One possible interpretation for this result is that algorithm Z always uses at least half of its buffer space optimally-as if it knew the future in advance. Our Results. We consider both delay- and rate-jitter. For delay-jitter, we give a very simple on-line algorithm, and prove that the delay-jitter in its output is no more than the delay-jitter produced by an optimal (off-line) algorithm using half the space. We give a lower bound on delay-jitter showing that doubling the space is necessary. We also consider a distributed implementation of our algorithm, where the total space of 2B is distributed along a path. We prove that the distributed algorithm guarantees the same delay-jitter of a centralized, off-line algorithm using space B, provided that an additional condition on the beginning of the sequence is met. To complete the picture, we also describe an efficient optimal off-line algorithm. For all our delay-jitter algorithms, we assume that the average inter-arrival time of the input stream (denoted X a ) is given ahead of time. One way to view the relativistic gurantee of our algorithm is the following. Assume that the specific arrival sequence is such that using a buffer of size B one can reduce the jitter comletely (i.e. zero jitter). In such a case, our online algorithm, using space 2B would also output a completely periodic sequence (i.e. zero jitter). For rate jitter, we assume that the on-line algorithm receives, in addition to X a , two parameters denoted I max and I min , which are a lower and an upper bound on the desired time between consecutive packets in the output stream. The on-line algorithm we present uses a buffer of size 2B + h where h 1 is a parameter, and B is such that an off-line algorithm using buffer space B can release the packets with inter-departure times in the interval [I min ; I max ] (but the optimal jitter may be much lower). The algorithm guarantees that the rate-jitter of the released sequence is at most the best off-line jitter plus an additive term of 2(B We also show how can the algorithm adapt to unknown X a . Finally, we prove that on-line algorithms using less than 2B buffer space are doomed to have trivial rate-jitter guarantees with respect to an off-line algorithm using space B. Related Work. QoS has been the subject of extensive research in the current decade, starting with the seminal work of Ferrari [2] (see [16] for a comprehensive survey). A number of algorithms has been proposed for jitter control. Partridge [9] proposed to time-stamp each message at the source, and fully reconstruct the stream at the destination based on a bound on the maximal end-to-end delay. Verma et al. [13] proposed the jitter-EDD algorithm, where a jitter controller at a switch computes for each packet its eligibility time, before which the packet is not submitted for to the link scheduler. The idea is to set the eligibility time to the difference between maximum delay for the previous link and the actual delay for the packet: this way the traffic is completely reconstructed at each jitter node. Note that jitter- EDD requires nodes to have synchronized clocks. The Leave-in-Time algorithm [3] replaces the synchronized clocks requirement of jitter-EDD with virtual clocks [19]. Golestani [4] proposed the Stop-and-Go algorithm, which can be described as follows. Time is divided to frames; all packets arriving in one frame are released in the following frame. This allows for high flexibility in re-shaping the traffic. Hierarchical Round-Robin (HRR), proposed in [6], guarantees that in each time frame, each connection has some predetermined slots in which it can send packets. A comparative study of rate-control algorithms can be found in [18]. A new jitter control algorithm was proposed in [14]. Paper Organization. In Section 2 we give the basic definitions and notations. In Section 3 we study delay jitter for a single switch. In Section 4 we extend the results of Section 3 to a distributed implementation. In section 5 we study rate jitter. Model jitter-control algorithm packet arrival sequence FIFO buffer packet release sequence Figure 1: Abstract node model. The jitter control algorithm controls packet release from the buffer, based on the arrival sequence. We consider the following abstract communication model for a node in the network (see Fig. 1). We are given a sequence of packets denoted 0; arrives at time arrival(k). Packets are assumed to have equal size. Each packet is stored in the buffer upon arrival, and is released some time (perhaps immediately) after its arrival. Packets are released in FIFO order. The time of packet release (also called packet departure or packet send) is governed by a jitter control algorithm. Given an algorithm A and an arrival time sequence, we denote by send A (k) the time in which packet k is released by A. We consider jitter control algorithms which use bounded-size buffer space. We shall assume that each buffer slot is capable of storing exactly one packet. All packets must be delivered, and hence the buffer size limitation can be formalized as follows. The release time sequence generated by algorithm A using a buffer of size B must satisfy the following condition for all 0 k n: where we define The lower bound expresses the fact that a packet cannot be sent before it arrives, and the upper bound states that when packet k +B arrives, packet k must be released due to the FIFOness and the limited size of the buffer. We call a sequence of departure times B-feasible for a given sequence of arrival times if it satisfies Eq. (1), i.e., it can be attained by an algorithm using buffer space B. An algorithm is called on-line if its action at time t is a function of the packet arrivals and releases which occur before or at t; an algorithm is called off-line if its action may depend on future events too. A times sequence is a non-decreasing sequence of real numbers. We now turn to define properties of times sequences, which are our main interest in this paper. Given a times sequence i=0 , we define its average, minimum, and maximum inter-arrival times as follows. ffl The average inter-arrival time of oe is X oe n . ffl The minimum inter-arrival time of oe is X oe ng. ffl The maximum inter-arrival time of oe is X oe ng. We shall omit the oe superscript when the context is clear. The average rate of oe is simply 1=X oe a . We shall talk about the jitter of oe. We distinguish between two different kinds of jitter. The delay jitter, intuitively, measures how far off is the difference of delivery times of different 1 Note that our definition allows for 0-length intervals where more than B packets are in the system. This formal difficulty can be overcome by assuming explicitly that each event (packet arrival or release) occurs in a different time point. For clarity of exposition, we prefer this simplified model, although our results hold in both models. packets from the ideal time difference in a perfectly periodic sequence, where packets are spaced exactly X a time units apart. Formally, given a times sequence i=0 , we define the delay jitter of oe to be 0i;kn We shall also be concerned with the rate jitter of oe, which can be described intuitively as the maximal difference between inter-arrival times, which is equivalent to the difference between rates at different times. Formally, we define the rate jitter of oe to be 0i;j!n The following simple property shows the relationship between delay and rate jitter. Lemma 2.1 Let oe be a times sequence. (1) The delay jitter of oe equals 0 if and only if the rate jitter of oe equals 0. (2) If the delay jitter of oe is J, then the rate jitter of oe is at most 2J. (3) For all ffl ? 0, and M , there exists a sequence oe ffl;M with rate jitter at most ffl and delay jitter at least M . Proof: Suppose that i=0 . 1. The delay jitter of oe is 0 iff for all 0 i n we have t a , which is true iff the rate jitter of oe is 0. 2. If the delay jitter of oe is J , then for all 0 and by the triangle inequality we have that the rate jitter of oe is at most 2J . 3. Let ffl Choose an even number n ? i=0 be defined inductively as follows. For Clearly, the resulting oe is a times sequence with average inter-arrival rate X a and rate jitter at most ffl 0 ffl. However, we have that hence the delay jitter is at least nffl 0? M by choice of n. Our means for analyzing the performance of jitter control algorithms is competitive analysis [1]. In our context, we shall measure the (delay or rate) jitter of the sequence produced by an on-line algorithm against the best jitter attainable for that sequence. As expected, finding the release times which minimize jitter may require knowledge of the complete arrival sequence in advance, i.e., it can be computed only by an off-line algorithm. Our results are expressed in terms of the performance of our on-line algorithms using buffer space B on as compared to the best jitter attainable by an off-line algorithm using space B off , where usually . We are interested in two parameters of the algorithms: the jitter (guaranteed by our on-line algorithms as a function of the best possible off-line guarantee) and the buffer size (used by the on-line algorithm, as a function of the buffer size used by an optimal off-line algorithm). 3 Delay-Jitter Control In this section we analyze the best achievable delay-jitter. We first present an efficient off-line algorithm which attains the best possible delay jitter using a given buffer with space B. We then proceed to the main result of this section, which is an on-line delay-jitter control algorithm which attains the best jitter guarantee that can be attained by any (off-line) algorithm which uses half the buffer space. Finally, we present a lower bound which shows that any on-line algorithm whose jitter guarantees are a function of the jitter guarantees of an off-line algorithm, must have at least twice the space used by the off-line algorithm. 3.1 Off-line Delay-Jitter Control We start with the off-line case. Suppose we are given the complete sequence farrival(k)g n of packet arrival times. We wish to find a sequence of release times fsend off (k)g n k=0 which minimizes the delay jitter, using no more than B buffer space. The off-line algorithm is defined as follows. Algorithm A: off-line delay-jitter control. 1. For each 0 k n, define the interval where we define 2. Find a minimal interval M which intersects all intervals E k . 3. For each packet k, let Theorem 3.1 The sequence fsend off (k)g n k=0 is a non-deceasing, B-feasible sequence with minimal delay jitter. Proof: It is straightforward to see from the definitions that send off [arrival(k); arrival(k +B)] and hence the resulting sequence is B-feasible. Proving FIFOness is done as follows. By definitions, it is sufficient to prove that P k P k+1 +X a . To see this, first note that by definition, We distinguish between two cases now. If min(M) Eq. (2) we have that P k+1 P k , and we are done. The second case is that min(M) ! In this case Eq. (2) implies that and the proof of correctness is complete. The optimality of the solution follows immediately from the minimality of M . 3.2 On-line Delay-Jitter Control Algorithm We now turn to our main result for delay-jitter control: an on-line algorithm using 2B buffer space, which guarantees delay-jitter bounded by the best jitter achievable by an off-line algorithm using B space. The algorithm is simple: first the buffer is loaded with B packets, and when the (B+1)-st packet arrives, the algorithm releases the first buffered packet. From this time on, the algorithm tries to release packet k after kX a time. Formally, the algorithm is defined as follows. Algorithm B: on-line delay-jitter control. Define send on a for all n. The release sequence is defined by send on send on (k); if arrival(k) send on on on Clearly, Algorithm B is an on-line algorithm. We prove its jitter-control property. Theorem 3.2 If for a given arrival sequence, an off-line algorithm using space B can attain delay jitter J , then the release sequence generated by Algorithm B has delay-jitter at most J using no more than 2B buffer space. Proof: Obviously, the buffer space used by Algorithm B is at most 2B. The bound on the delay-jitter follows from Lemma 3.4 and Lemma 3.6 proved below. time packet number Figure 2: An example for oriented jitter bounds. A point at coordinates (x; y) denotes that packet y is released at time x. The slope of the dashed lines is 1=X a . The following definition is useful in the analysis (see Fig. 2). Definition 3.1 Let k=0 be a time sequence. The oriented jitter bounds for packet k are 0in a g 0in a g Intuitvely, J oe (k) says by how much packet k is late compared to the earliest packet, and J oe (k) says by how much k is premature comapred to the latest packet. We have the following immediate properties for oriented jitter bounds. Lemma 3.3 Let k=0 be a time sequence with average inter-arrival times X a and delay jitter J. Then (1) For all k, J(k) 0 and J(k) 0. (2) For all k, (3) There exist k and k 0 such that Proof: 1. Follows by choosing 3.1. 2. Let i be such that j. Rearranging, we have that Assume w.l.o.g. that t i 0 . From the definition it follows that t i 0 g. Therefore, 0in and 0in Summing Eqs. (3,4), the result follows. 3. Follows by choosing Eqs. (3,4), respectively. The following lemma shows that the deviation of the actual release time generated by Algorithm B from the ideal 0-jitter sequence of fsend on (k)g k is bounded. Somewhat surpris- ingly, it is bounded by the oriented jitter bounds of two specific packets in any B-feasible sequence. Lemma 3.4 Let k=0 be any B-feasible sequence for a given arrival sequence. Then for all 0 k n, we have \GammaJ oe (0) send on Proof: We proceed by case analysis. If send on on (k) then we are done by Lemma 3.3 (1). If send on (k) ? send on (k), then by the specification of Algorithm B, we have that send on arrival(k). In this case the lemma is proved by the following inequality. send on send off (k) by definition of J oe (0) send on since send on (0) send off (0) on The last case to consider is send on on (k). In this case, by the specification of Algorithm B, we have that send on 2B). The lemma in this case is proved by the following inequality. send on send off (k +B) by B-feasibility of off-line by definition of J oe (B) send on on (0) send off (B) on The reader may note that since Lemma 3.3 (1,2) implies that J oe (0); J oe (0) J oe , Lemma 3.4 can be used to easily derive a bound of 2J oe on the delay-jitter attained by Algorithm B. Proving the promised bound of J oe requires a more refined analysis of the oriented jitter bounds. To facilitate it, we now introduce the following concept. Definition 3.2 Let k=0 be a times sequence. Let t t k . The times sequence oe perturbed at k to t is Intuitively, is the sequence obtained by assigning release time t to packet k, and changing the times of other packets to preserve the FIFO order (see Fig. 3 for an example): if packet k is to be released earlier than t k , then some packets before k may be moved as well; and if packet k is to be released later than t k , then some packets after k may be moved. The following properties for perturbed sequences are a direct consequence of the definition k=0 be a times sequence, let k be any packet, and let t be any time point. packet number time packet number time Figure 3: An example of perturbation. Left: A sequence oe. Right: oe(5 : t). Note that in were moved with respect to oe. Proof: The simplest way to verify these claims is geometrical: Consider Figure 3, which corresponds to the case of t Assertion (B1) says that if point k is not moved left to the left diagonal line, then all points remain between the two diagonal lines, and that there are points which lie on the diagonal lines. Assertion (B2) states that the horizontal distance between point k and left diagonal line strictly decreases, and Assertion states that for points below point k, the horizontal distance to the left diagonal line does not increase. The case of t analogous. To prove Theorem 3.2, we prove an interesting property of oriented jitter bounds in optimal sequences. Intuitively, the lemma below says the following. Fix an arrival sequence, and consider all optimal release sequences using B buffer space. Fix any two packets at most B apart. Then it cannot be the case that in all optimal release sequences both the first packet is too early and the second packet is too late. Formally, we have the following. Lemma 3.6 Let J be the minimal delay jitter for a given arrival sequence using space B, and let 0 i j n be packets such that j i +B. Then there exists a B-feasible sequence oe for the given arrival sequence with delay jitter J such that J oe (i) Note that Lemma 3.6 with combined with Lemma 3.4, completes the proof of Theorem 3.2. We shall use the general statement in Section 4. Proof: Let oe be an optimal release sequence attaining jitter J for the given arrival se- quence, in which J oe (i) + J oe (j) is minimal among all optimal sequences. First, note that if either J oe then we are done since by Lemma 3.3 (1,2) we have that J oe (i); J oe (j) J . So assume from now on that J oe (i) ? 0 and J oe (j) ? 0. We claim that in this case, t are released together (and hence all packets are released together). We prove this claim by contradiction: suppose that t Then it must be the case that either (i) t i ! arrival(j) or (ii) t j ? arrival(j), or both (i) and (ii) hold. If case (i) holds, let t and consider the perturbed sequence oe(i : t) in which packet i is released at time t. By choice of t, we have that (i). The perturbed sequence oe(i : t) has the following properties. is B-feasible, since it may differ from oe at most by packets These packets are held a little longer in oe(i : but they are released at time t ! arrival(j) arrival(i +B). The claim now follows for case (i), since Properties (1,2) imply that oe(i : t) is a sequence using B buffer space which attains jitter J , but Properties (3,4) contradict the assumed minimality of J oe (i) (j). A similar argument shows that if case (ii) holds, then for arrival(j)g, the perturbed sequence oe(j : t) contradicts the minimality of J oe (i) Thus we have proved that for an optimal sequence, either J oe which cases the lemma is proved), or else, for a sequence minimizing J oe (i) must be the case that t . We now proceed to bound J oe (i) using the fact that t First, note that since by definition there exists a packet k 1 such that t k 1 (j), and since by definition we get from the fact that Similarly, we have that Adding Equations (5,6), we get that we conclude that as required. 3.3 A Lower Bound for On-line Delay-Jitter Control Algorithms We close this section with a lower bound for on-line delay-jitter control algorithms. The following theorem says that any on-line algorithm using less than 2B buffer space pays heavily in terms of delay jitter when compared to an off-line algorithm using space B. Theorem 3.7 Let 1 ' ! B. There exist arrival sequences for which an off-line algorithm using space B gets jitter 0, and any on-line algorithm using 2B \Gamma ' buffer space gets delay-jitter at least 'X a . Moreover, there exist arrival sequences for which an off-line algorithm using space B gets 0-jitter, and no on-line algorithm using less than B buffer space can guarantee any finite delay jitter. Proof: Consider the following scenario. At time 0, packets arrive, and at time arrive. First, note that there is an off-line algorithm attaining 0 jitter by releasing each packet k at time k \Delta X a . Consider now any on-line algorithm Z. We first claim that Z cannot release packet 0 before packet B arrives: otherwise, packet B may arrive arbitrarily far in the future, making the delay jitter of the on-line algorithm arbitrarily large. Hence, at time B \Delta X a , when B+1 new packets arrive, algorithm Z still stores the first packets, and since it has buffer space 2B \Gamma ' by assumption, it is forced to release at least immediately. Since the delays of packets 0 and ' are equal, it follows from the definition of delay-jitter that the delay-jitter of the release sequence is at least 'X a . For the case of an on-line algorithm with less than B space, consider the scenario where a batch of B packet arrive together at time 0, and then a batch of B more packets arrive at time T for some very large T . Since the on-line algorithm has to release packet 0 at time 0, we have that its delay jitter is at least T=(B \Gamma 1), which can be arbitrarily large. 4 Distributed Delay-Jitter Control In Section 3 we have considered a single delay-jitter regulator. In this section we prove an interesting property of composing many delay-jitter regulators employing our Algorithm B. Specifically, we consider a path of m links connecting nodes v is the source and v m is the destination. We make the simplifying assumption that the propagation delay in each link is deterministic. We denote the event of the arrival of packet k at node j by arrival(k; j), and the release of packet k from node v j by send(k; j). The input stream, generated by the source, is fsend(k; 0)g k (or farrival(k; 1)g k ), and the output stream is fsend(k; m)g k . Each node has 2B=m buffer space, and for simplicity we assume that m divides B. The distributed algorithm is the following. Algorithm BD: distributed on-line delay-jitter control. For each 1 j m, node buffer space 2B=m. Specifically, node j sets send on (k; a , and it releases packet k as close as possible to send on (k; j) subject to 2B=m-feasibility (see Algorithm B). We prove that the jitter control capability of Algorithm BD is the same as the jitter control capability of a centralized jitter control algorithm with B total buffer space, under a certain condition for the beginning of the sequence (to be explained shortly). Put differently, one does not lose jitter control capability by dividing the buffer space along the path. The precise result is given in the theorem below. Theorem 4.1 Suppose that for a given arrival sequence k=0 , there exists a centralized off-line algorithm attaining jitter J using space B, with packet 0 released before time arrival(B=m). Then if oe is the release sequence of node v 0 , the release sequence fsend on (k; m)g k generated by Algorithm BD at node v m has delay jitter at most J. Intuitively, the additional condition is that there is a way to release the first packet relatively early by a centralized optimal algorithm. This condition suffices to compensate for the distributed nature of Algorithm BD. The condition is also necessary for the algorithm to are input into the system at the start of the algorithm, then an off-line algorithm can still wait arbitrarily long before starting to release packets, while Algorithm BD is bound to start releasing packets even if only 2B The proof is essentially adapting the proofs of Algorithm B in Section 3 to the distributed setting. We highlight the distinguishing points. Let the propagation delay over link (v j the total delay of links on the path. The first lemma below bounds the desired release times of all packets at one node in terms of the desired release times in upstream nodes. Lemma 4.2 For all nodes 1 j i m and all packets k, send on (k; on (k; i) send on Proof: Consider the lower bound first. By the algorithm, we have that for all ', send on (0; send on (0; '). Since for all ' ? 1, send on (0; ') send on (0; we obtain by induction on on (k; i) send on (k; proving the lower bound. We now prove the upper bound. First, we claim that for all 1 i m, send on (k; ') send on (k; ') for 0 k Eq. (7) follows from the fact that by the specification of Algorithm B, a node starts releasing packets only if all first B are in its buffer, and therefore none of the first B packets is released too late in any node. We now prove the upper bound by induction on j. The base case, trivial. For the inductive step, fix j and consider i have send on a by algorithm send on (0; i) m )X a by (7) send on (0; )X a by induction on (k; rearranging on For the case of underflow, we argue that if a packet is "late" in the output node v n , then it was late in all nodes on its way. Lemma 4.3 If send on (k; m) ? send on (k; m), then send on (k; m) Proof: First, we show that for any node v j , if send on (k; on (k; j), then send on (k; on 1). This is true since by the specification of Algorithm B, at time send on (k; j) the buffer at node j is empty, and hence node v j \Gamma1 has not sent packet k by time send on (k; on on (k; this implies that send on (k; send on 1). Therefore, for all nodes v j , we have that send on (k; and by summation we obtain that send on (k; m) = send on (k; For the case of overflow, we show the analogous property: if there is an overflow in the output node, then it is the result of a "chain reaction" of overflows in all nodes. Lemma 4.4 If send on (k; m) ! send on (k; m), then send on (k; m) Proof: We prove that if send on (k; i) ! send on (k; i) then send on send on by the bound on the buffer size send on (k; i) \Gamma d i by our assumption send on send In other words, if packet k is overflowing at node m, then packet k m is overflowing in node i, for each 1 i m. Hence for each i, we have send on . The lemma follows. The lemmas above are used in the proof of the following variant of Lemma 3.4. Lemma 4.5 Let k=0 be any B-feasible sequence for a given arrival sequence such that send off (0) arrival(B=m; 1). Then for all 0 k m, we have \GammaJ oe (0) send on (k; m) \Gamma send on (k; m) J oe (B=m). Proof: If send on (k; m) = send on (k; m) we are done by Lemma 3.3 (1). If send on (k; m) ? send on (k; m), then send on (k; m) = send on (k; 0) +D by Lemma 4.3 send off (k) +D since send on (k; send on (0; since send off (0) arrival( send on (0; m) on (k; m) If send on (k; m) ! send on (k; m), then send on (k; m) send off (k +B) +D since off-line has B space send off ( B arrival( send on (0; send on Theorem 4.1 follows from Lemma 4.5, when combined with Lemma 3.6 (which is independent of the on-line algorithm), with 5 Rate-Jitter Control In this section we consider the problem of minimizing the rate-jitter, i.e., how to keep the rate at which packets are released within the tightest possible bounds. We shall use the equivalent concept of minimizing the difference between inter-departure times. We present an on-line algorithm for rate-jitter control using space 2B compare it to an off-line algorithm using space B and guaranteeing jitter J . Our algorithm guarantees rate jitter at most J constant c. We also show how to obtain rate jitter which is a multiplicative factor from optimal, with a simple modification of the algorithm. The algorithm can work without knowledge of the exact average inter-arrival time: in this case, jitter guarantees will come into effect after an initial period in which packets may be released too slowly. We also show that without doubling the space, no guarantees in terms of the optimal rate-jitter can be made. As an aside, we remark that off-line rate-jitter control can be solved optimally using linear-programming technique. 5.1 On-line Rate-Jitter Control Algorithm We now turn to describe the main result for this section: an on-line algorithm for rate-jitter control. The algorithm is specified with the following parameters: ffl B, the buffer size of an off-line algorithm, i.e. B space parameter for the on-line algorithm, such that B on ffl I min ; I bounds on the minimum and maximum inter-departure time of an off-line algorithm. ffl X a , the average inter-departure time in the input (and also the output) sequence. The parameters I min and I max can be thought of as requirements: these should be the worst rate jitter bounds the application is willing to tolerate. The goal of a rate-jitter control algorithm is to minimize the rate jitter, subject to the assumption that space B is sufficient (for an off-line algorithm) to bound the inter-departure times in the range [I min ; I max ]. A trivial choice for I min and I max is X min and X max , which are the minimal and maximal inter arrival times in the input sequence. However, using tighter I min and I max , one may get a much stronger guarantee. The jitter guarantees will be expressed in terms of B; h; I and J , the best rate jitter for the given arrival sequence attainable by an off-line algorithm using space B. Note that for an on-line algorithm, even achieving rate jitter I max \GammaI min may be non-trivial. These are bounds on the performance of an off-line algorithm, whose precise specification may depend on events arbitrarily far in the future. The basic idea in our algorithm is that the next release time is a monotonically decreasing function of the current number of packets in the buffer. In other words, the more packets there are in the buffer, the lower the inter-departure time between the packets (and thus the higher the release rate). Algorithm C: on-line rate-jitter control. The algorithm uses B on space. With each possible number 0 j 2B + h of packets in the buffer, we associate an time denoted IDT(j), defined as follows. Let I I I Note that IDT(j) is a monotonically decreasing function in j. The algorithm starts with a buffer loading stage, in which packets are only accumulated (and not released) until the first time that the number j of packets in the buffer satisfies IDT(j) X a . Let a g, and let T denote the first time in which the number of packets in the buffer reaches S. At time T , the loading stage is over: the first packet is released and the following rule governs the remainder of the execution of the algorithm. A variable last departure is maintained, whose value is the time at which the last packet was sent. If at time t, we have t last departure+IDT(j), where j is the number of packets currently in the buffer, then we deliver a packet and update last departure. The rate-jitter bound of Algorithm C is given in the following theorem. Theorem 5.1 Let J be the best rate-jitter attainable (for an off-line algorithm) using buffer space B for a given arrival sequence. Then the maximal rate-jitter in the release sequence generated by Algorithm C is at most J h , and never more than I The idea in the proof of Theorem 5.1 is that the number of packets in the buffer is never more than slots away from the slots which correspond to rates generated by an optimal off-line algorithm. We now formally analyze Algorithm C. Fix an optimal execution of the off-line algorithm. Let us denote the maximum and minimum inter-departure times of the off-line execution by Y max and Y min , respectively. (Hence the jitter attained by the off-line algorithm is Y these quantities, we also define the following terms. I I Note that L S U . We shall also use the following shorthand notation. Let B on (t) and denote the number of packets stored in time t in the buffers of the Algorithm C and of the off-line algorithm, respectively, and let i.e., how many packets does the Algorithm C has more than the off-line algorithm at time t. We use extensively the following trivial property of the difference. Lemma 5.2 For all t, \GammaB diff(t) B on (t). Proof: Immediate from the fact that 0 B off (t) B. Let S on the number of packets sent by Algorithm C in the time interval analogously for the off-line algorithm. The following lemma states that the difference is modified according only to the difference in the packets released. Lemma 5.3 For any two time points Proof: Consider the events in the time interval arrival increases the number of stored packets for both the off-line and Algorithm C, and hence does not change their difference. It follows that diff(t 2 exactly the difference in the number of packets sent by the two algorithms in the given interval. The significance of Lemma 5.3 is in that it allows us to ignore packet arrivals when analyzing the space requirement of an algorithm: all we need is to consider the difference from the space requirement of the off-line algorithm. The following lemma, which bounds the minimal inter-departure time of Algorithm C, is an example for that. Lemma 5.4 For all times t, diff(t) U + 1. Proof: Let t be any point in time. If B on (t) U + 1, the lemma follows immediately. So assume that B on (t) ? U be a point such that B on (t 0 ) U and B on (t 0 ) U for all t Such a point exists since B on (T Consider the time interval t]: in this interval, at most Y min were released by the off-line algorithm, while Algorithm C has released at least I U Y min , and hence S on Since by Lemma 5.2 we have that diff(t 0 ) U , the result follows from Lemma 5.3. Similarly, we bound the difference from below. Lemma 5.5 For all times t ? T , diff(t) Proof: Let t ? T be a point in time. The case of B on (t) be a point such that B on (t 0 ) L and B on (t 0 ) L for all t]. The point t 0 must exist since B on (T L. For the time interval we have that S off , and S on I L B, the result follows from Lemma 5.3. We now prove Theorem 5.1. Proof of Theorem 5.1: By Lemma 5.4, at all times t, B on (t) hence the minimal inter-departure time of Algorithm C is smaller than Y min by less than (B+2) Imax \GammaI min h . By Lemma 5.5, for all times t ? T , the maximal inter-departure time of Algorithm C is larger than Y max by less than (B h . Since since no packets is released before time T , the theorem follows. It is worthwhile noting that doubling the space is mandatory for on-line rate-jitter control (as well as for delay-jitter control), as the following theorem implies. Theorem 5.6 Let 1 ' ! B. There exist arrival sequences for which an off-line algorithm using space B gets 0-jitter, and any on-line algorithm using 2B \Gamma ' buffer space gets rate-jitter at least X a The proof of Theorem 5.6 is similar to the proof of Theorem 3.7, and we therefore omit it. 5.2 Adapting to Unknown X a We can avoid the need of knowing X a in advance, if we are willing to tolerate slow rate in an initial segment of the online algorithm. This is done by changing the specification of the loading stage of Algorithm C to terminate when the buffer contains B packets (which corresponds to inter-arrival time of I max , as opposed to inter-arrival time of X a in the original specification). Thereafter, the algorithm starts releasing packets according to the specification of IDT. Call the resulting algorithm C \Gamma . Below, we bound the time which elapses in an execution of C \Gamma until the buffer size will reach the value of L. Clearly, from that point onward, all guarantees made in Theorem 5.1 hold true for Algorithm C \Gamma as well. Lemma 5.7 Consider an execution of Algorithm C \Gamma . Let T be the time where the initial loading ends, and let T + be the first time such that B on (T Proof: For to be the first time after T where B on (t i ) Consider a time interval its length by i . Denote the number of packets arriving in the interval by A i . Consider the off-line algorithm: For all 1 i have that Consider now the execution of Algorithm in the time interval the inter-departure time is at least Y therefore S on Ymax+iffi . Using Eq. (8) and since B on (t by definition, we have i.e., \GammaY Summing over noting that and that \GammaB 5.3 Multiplicative Rate Jitter For some applications, it may be useful to define jitter as the ratio between the maximal and minimal inter-arrival times. We call this measure the multiplicative rate jitter, or m-rate jitter for short. It is easy to adapt Algorithm C to the case where we are interested in the m-rate jitter. All that is needed is to define I I I for . In this case we obtain the following result, using the same proof technique as for Theorem 5.1. Theorem 5.8 Let J be the best m-rate-jitter attainable (for an off-line algorithm) using buffer space B for a given arrival sequence. Then the maximal m-rate-jitter in the release sequence generated by Algorithm C using function IDTm is at most J \Delta h . 6 Conclusion In this paper we have studied jitter control algorithms, measured in terms of guarantees relative to the best possible by an off-line algorithm. Our results for delay jitter show that the simple algorithm of filling half the buffer has a very strong relative property. For rate jitter, we proposed a simple algorithm where the release rate is proportional to the fill level of the buffer, and showed that its relative guarantees are quite strong as well. We have studied a very simple distributed model for jitter control. We leave for further work analyzing more realistic models of systems, including multiple streams and more interesting network topology. --R Online Computation and Competitive Analysis. Client requirements for real-time communication services ATM Networks: Concepts Rate control servers for very high-speed networks Competitive snoopy caching. An Engineering Approach to Computer Networking. Isochronous applications do not require jitter-controlled networks Amortized efficiency of list update and paging rules. Computer Networks. The ATM Forum Technical Committee. Guaranteeing delay jitter bounds in packet switching networks. Charcterizing traffic behavior and providing end-to-end service guarantees within ATM networks Service disciplines for guaranteed performance service in packet-switched networks Comparison of rate-based services disciplines A New Architecture for Packet Switched Network Protocols. --TR Amortized efficiency of list update and paging rules A stop-and-go queueing framework for congestion management Comparison of rate-based service disciplines On per-session end-to-end delay distributions and the call admission problem for real-time applications with QOS requirements ATM networks (2nd ed.) Leave-in-Time Computer networks (3rd ed.) An engineering approach to computer networking Online computation and competitive analysis Characterizing Traffic Behavior and Providing End-to-End Service Guarantees within ATM Networks --CTR Khuller, Problems column, ACM Transactions on Algorithms (TALG), v.2 n.1, p.130-134, January 2006 Pal , Mainak Chatterjee , Sajal K. Das, A two-level resource management scheme in wireless networks based on user-satisfaction, ACM SIGMOBILE Mobile Computing and Communications Review, v.9 n.4, October 2005 Yiping Gong , Bin Liu , Wenjie Li, On the performance of input-queued cell-based switches with two priority classes, Proceedings of the 15th international conference on Computer communication, p.507-514, August 12-14, 2002, Mumbai, Maharashtra, India

buffer overflow and underflow;competitive analysis;streaming connections;jitter control;quality of service networks

504917

Prefetching for improved bus wrapper performance in cores.

Reuse of cores can reduce design time for systems-on-a-chip. Such reuse is dependent on being able to easily interface a core to any bus. To enable such interfacing, many propose separating a core's interface from its internals by using a bus wrapper. However, this separation can lead to a performance penalty when reading a core's internal registers. In this paper, we introduce prefetching, which is analogous to caching, as a technique to reduce or eliminate this performance penalty, involving a tradeoff with power and size. We describe the prefetching technique, classify different types of registers, describe our initial prefetching architectures and heuristics for certain classes of registers, and highlight experiments demonstrating the performance improvements and size/power tradeoffs. We further introduce a technique for automatically designing a prefetch unit that satisfies user-imposed register-access constraints. The technique benefits from mapping the prefetching problem to the well-known real-time process scheduling problem. We then extend the technique to allow user-specified register interdependencies, using a Petri net model, resulting in even more efficient prefetch schedules.

Overview Separating a core's interface behavior and internal behavior can lead to performance penalties. For example, consider the core architectures shown in Figures 1(a), 1(b) and 1(c), showing a core with no bus wrapper, a core with a bus wrapper (BW) but without prefetching, and a core with a BW with prefetching, respectively. The latter two architectures are similar to that being proposed by the VSIA. The BW interfaces with the system bus, whose protocol may be arbitrarily complex, include a variety of features like arbitration. The BW also interfaces with the core internals, over a core internal bus; this bus is typically extremely simple, implementing a straightforward data transfer. It is this internal bus that the VSI On-Chip Bus group is standardizing. Without a BW, a read of a core's internal register from the on-chip bus may take as little as two cycles, as shown in Figure 2(a). With a BW, the read of a core's internal register may require four cycles, two from the internal module to the BW, and ACM Transactions on Design Automation of Electronic Systems, Vol. 7, No. 1, January 2002. Prefetching for Improved Bus Wrapper Performance 7 Fig. 3. PVCI's location in a system-on-a-chip. two from the BW to the bus. Thus, a read may require extra cycles compared with a core whose interface and internal behavior are combined. However, a core with its interface behavior separated into a bus wrapper is believed to be much easier to retarget to different buses than a core whose interface behavior is integrated with its internal behavior. By standardizing the interface between the core's internals and the bus wrapper, retargeting of a core may become easier. 3.2 PVCI After deciding that a single on-chip bus standard was unlikely, the VSIA developed the VCI [Virtual Socket Interface Association 1997b]. The VCI is a proposed standard interface between a core's internals and a core's bus wrapper, as illustrated in Figure 3. Retargeting a core using VCI will involve roughly the same changes to the bus wrapper, since the VCI ensures that the changes are limited to the wrapper and not the internals, and since a bus provider can even provide bus wrapper templates between the bus and the VCI. The VCI is a far simpler protocol than a typical bus protocol, since it is a point-to-point transfer protocol. In contrast, a bus protocol may involve more advanced features, such as arbitration, data multiplexing, pipelining, and so on. Thus, standardizing the VCI is far simpler than standardizing a bus protocol. The PVCI is a simpli?ed version of the VCI, speci?cally intended for periph- erals. PVCI cores would reside on a lower-speed peripheral bus as shown in Figure 3, and thus would not need some of the high-speed features of the VCI, e.g., packet chaining. The general structure of the PVCI is shown in Figure 4. It consists of two unidirectional buses. One bus leads from the wrapper to the internals. The wrapper sets the read line to indicate a read or a write, and sets the address lines with a valid address. For a write, it also sets the wdata lines. It asserts the val line to actually initiate the read or write. The wrapper must hold all these lines constant until the internals assert the ack line. For a write, this means that the internals have captured the write data. For a read, this means that the internals have put the read data on the rdata bus. The transaction ACM Transactions on Design Automation of Electronic Systems, Vol. 7, No. 1, January 2002. 8 Lysecky and Vahid Fig. 4. PVCI's general structure. is completed on the next rising clock edge. A fast internals module can keep ack asserted continuously to provide for fast transfers, similar in spirit to the synchronous wait protocol [Vahid and Givargis 1999]. 3.3 Experiments with Bus Wrappers We sought to evaluate the impact of a wrapper and of PVCI using a simple peripheral bus. We used a bus with a two-phase handshake protocol to ensure that the communication was as fast as possible for a given peripheral. As previously demonstrated, using a wrapper results in a two-cycle overhead per read as compared with an integrated core. Figure 2(a) illustrates the timing of a read cycle of this peripheral bus for an integrated core. The peripheral bus master (in our case, the bridge) places an address on addr and then strobes rd. The peripheral responds by placing data on data and strobing rdy as early as once cycle after receiving the rd strobe. Thus, the total read cycle could be as little as two clock cycles. Figure 2(b) illustrates the read cycle of the bus for a core using a bus wrapper. After the bus master places the address and strobes rd, the wrapper responds by translating this read request into a read request over the internal bus. This translation involves translating the address to one appropriate for the core and then placing that address on wrp addr, and then asserting wrp read. The core's internals respond by placing data on wrp data and then asserting wrp rdy. The wrapper receives the data, puts it on the peripheral bus, and strobes rdy. A write cycle need not incur any performance overhead in the wrapper ver- sions. When the bus master sets the addresses and strobes the appropriate ready line, the wrapper can respond immediately by capturing the data and strobing the ready line, just like an integrated core will do. The wrapper can then proceed to write the captured data to the core internals, while the bus master proceeds with other activities. The example we evaluated was a simple version of a digital camera system, illustrated in Figure 5. The camera system consists of a (simpli?ed) MIPS mi- croprocessor, BIOS, and memory, all on a system bus, with a bridge from the system bus to a peripheral bus (ISA) having a CCD (charge-coupled device) Prefetching for Improved Bus Wrapper Performance 9 Fig. 5. Digital camera example system. preprocessor and a simple CODEC (compressor/decompressor). The two-level bus structure is in accord with the hierarchical bus concept described in [Virtual Socket Interface Association 1997a]. The camera is written in register-transfer level synthesizable VHDL, and synthesizes to about 100,000 cells. We used the Synopsys Design Compiler as well as the Synopsys power analysis tools to evaluate different design metrics. Power and performance were measured for the processing of one frame. We made changes to the CCD preprocessor and CODEC cores since they represent the peripherals on the peripheral bus. These cores are used heavily while processing a frame. We created three versions of the camera system: (1) Integrated: The CCD preprocessor and CODEC cores were written with the interface behavior inlined into the internal behavior of the core. Thus, synthesis generates one entity for each core. (2) Non-PVCI wrapper: The CCD preprocessor and CODEC cores were written with the interface behavior separated into a wrapper. Thus, synthesis generates two connected entities for the core. The interface between these two wrapper and internal entities consisted of a single bidirectional bus, a strobe control line and a read/write control line, and however many address lines were necessary to distinguish among internal registers. (3) PVCI wrapper: Same as the previous version, except that the interface between the wrapper and internal entities was PVCI. The non-PVCI wrapper version was created for another purpose, well before the PVCI standard was developed and with no knowledge that the version would be used in these experiments. Thus, its structure was developed to be as simple as possible. Table I summarizes size, performance, and power results. Size is reported in equivalent NAND gates, time in nanoseconds, and power in milliwatts. The size overhead when using a bus-wrapper (non-PVCI) compared to the integrated version was roughly 1500 gates per core. This overhead comes from extra control ACM Transactions on Design Automation of Electronic Systems, Vol. 7, No. 1, January 2002. Table I. Comparison of Interface Versions Using a Custom Bus Version Ex. Size of Wrapper Size of Power for Total time I/O Time Integrated Wrapper CCD 1661 34556 8.11 79055 15520 CODEC 1674 1904 PVCI Wrapper CCD 1439 33978 7.98 79055 15520 CODEC 1434 1588 and registers. In the integrated version, the core's internals includes control to interface to the peripheral bus. In the wrapper version, this control is replaced by control for interfacing to the wrapper, so the size of the core's internals stays the same. However, the wrapper now must implement control for interfacing to the internals, and for interfacing to the peripheral bus, representing overhead. The wrapper must also include registers whose contents are copied to/from the internals, representing additional overhead. The reason that the non-PVCI wrapper version shows more size overhead than the PVCI wrapper version is because the non-PVCI version used a single bus for transfers both two and from the core internals, whereas PVCI speci?es two separate buses, resulting in less logic but more wires. Fifteen hundred gates of size overhead seems quite reasonable, given the continued increase of chips' gate capacities, and given that peripheral cores typically posses 20,000 gates or more [Mentor Graphics n.d. The system power overhead was only about 1%. The extra power comes from having to transfer items twice per access. On a write, an item must be transferred ?rst from the bus to the wrapper, then from the wrapper to the internals. On a read, an item must be transferred ?rst from the internals to the wrap- per, then from the wrapper to the bus. However, the power consumed by the memory, system bus, and processor dominate, so the extra power due to the wrappers is very small?even though the CCD and CODEC are heavily used when processing a frame. In Table I, we can see that there is a 100% increase in peripheral I/O access time when bus wrappers are employed. This overhead is due to the use of a wrapper, which would have occurred whether using PVCI or another wrapper. In our experiments, the CCD was accessed 256 times per image frame, while the CODEC was accessed a total of 128 times per frame. Because the MIPS processor executed approximately 5000 instructions per frame, the overall overhead of the bus wrappers amounts to approximately 5%. One difference between the non-PVCI and PVCI interface that does not appear in the results is the number of wires internal to the core. The non-PVCI version uses a multiplexed bus, and has fewer signals (some PVCI signals were not shown), and thus would have fewer internal wires. Noting that our CCD and CODEC cores are relatively small and have simple interfaces, it took us 6 designer hours, excluding synthesis and simulation time, to retarget a design from one wrapper to another, e.g., to convert the CCD's non-PVCI wrapper to a PVCI implementation. Synthesis time for the CCD and CODEC was approximately 1 hour. Simulation time for capturing ACM Transactions on Design Automation of Electronic Systems, Vol. 7, No. 1, January 2002. Prefetching for Improved Bus Wrapper Performance 11 one image frame was slightly over 10 hours and power analysis was an additional 5 hours. These times were obtained by synthesizing the models down to gates using Synopsys Design Compiler with medium mapping effort, using the lsi-10k library supplied by Synopsys, with no area or delay constraints speci?ed. We used a dual 200-MHz Ultra Sparc II machine to perform both our synthesis and simulation. Synthesis and simulation times were relatively the same between the integrated bus implementations and those using a bus wrapper. We note that peripheral devices capable of DMA or burst mode I/O with interrupts will require more time to integrate into a system. Although the use of bus wrappers improves the usefulness of a core by making it easier to retarget to varying systems, this reusability comes at a cost. Bus wrappers introduce both performance and power overhead, as we have demon- strated. In tightly constrained systems where peripheral access time is critical, this overhead is often infeasible. Ideally, the use of bus wrappers could allow for quick retargeting of a core while not degrading performance. In the next section, we present a technique called prefetching that effectively eliminates the performance overhead of bus wrappers. 4. BASIC PREFETCHING 4.1 Overview Our focus is to minimize this performance penalty in order to maximize the usefulness of the core. We seek to do so in a manner transparent to both the developers of the core internal behavior as well as developers of the on-chip bus. Because of the continued exponential growth in chip capacity, we seek to gain performance by making the tradeoff of increased size, since size constraints continue to ease. However, we note that our approach increases the switching activity of the core, and thus we must also evaluate the increased power consumption and seek to minimize this increase. We focus on peripheral cores, whose registers will be read by a microprocessor over an on-chip bus (perhaps via a bus bridge) with the idea being to minimize the read latency experienced by the microprocessor. The basic technique that we propose is called prefetching. Prefetching is the technique of copying a core's internal register data into a prefetch register in a core's BW, so that when a read request from the bus occurs, the core can immediately output prefetched data without spending extra cycles to ?rst get the data from the core's internal module. We use the terms hit and miss in a manner identical for caches; a hit means that the desired data is in a prefetch register, while a miss means that the data must ?rst be fetched into a prefetch register before being output to the on-chip bus. For example, Figure 2(c) shows that prefetching a core's internal register D into a BW register D0 results in a system read again requiring only two cycles, rather than four. 4.2 Classi?cation of Core Registers We immediately recognized the need to classify common types of registers found in peripheral cores, since different types would require different prefetching approaches. 12 Lysecky and Vahid After examining cores, primarily from the Inventra library [Mentor Graphics n.d.], focusing on bus peripherals, serial communication, encryption, and com- pression/decompression, we de?ned a register classi?cation scheme based on four attributes: update type, access type, noti?cation type, and structure type: (1) The update type of a register describes how the register's contents are mod- i?ed. Possible types include: (a) A static-update register is updated by the system only, where the system is the device (or devices) that communicate with the core over the on-chip bus. An example of a static register is a con?guration register. After the system updates the register, the register's content does not change until the system updates it again. (b) A volatile-update register is updated by a source other than the system (e.g., internally by the core or externally by the core's environment) at either a random or ?xed rate. An example is an analog-to-digital converter, which samples external data, converts the data to digital, and stores the result in a register, at a ?xed rate. (c) An induced-update register is updated as a direct result of another register within the core being updated. Thus, we associate this register with the inducing register. Typically, an induced register is one that provides status information. (2) The access type of a register describes whether the system reads and/or writes the register, with possible types including: (a) read-only access, (b) write-only access, and (c) read/write access. (3) The noti?cation type describes how the system is made aware that a register has been updated, with possible types including: (a) An interrupt noti?cation in which the core generates an interrupt when the register is updated. (b) A register-based ?ag noti?cation in which the core sets a ?ag bit (where that bit may be part of another register). (c) An output ?ag noti?cation in which the core has a speci?c output signal that is asserted when the register is updated. (d) No noti?cation in which the system is not informed of updates and simply uses the most recent register data. (4) The structure type of the register describes the actual storage capability of the register, with possible types including: (a) A singly structured register is accessed through some address and is internally implemented as one register. (b) A queue-structured register is a register that is accessed through some address but is internally implemented as a block of memory. A common example is a buffer register in a UART. (c) A block-structured register is a block of registers that can be accessed through consecutive addresses, such as a register ?le or a memory. Prefetching for Improved Bus Wrapper Performance 13 4.3 Commonly Occurring Register Types For our ?rst attempt at developing prefetching techniques for cores, we focused on the following three commonly occurring combinations of registers in cores: (1) Core1?con?guration registers: Many cores have con?gurable settings controlled by a set of con?guration registers. A typical con?guration register has the features of static update, read/write access, no noti?cation, and singly structured. We refer to this example as Core1. (2) Core 2?task registers: Many cores carry out a speci?c task from start to completion and have a combination of a data input register, a data output register, and a status register that indicates completion of the core's task. For example, a CODEC (compress/decompress) core typically has such a set of registers. We looked at how to prefetch the data output and status regis- ters. The data output register has the following features: volatile-update at a random rate, read-only access, register-based ?ag noti?cation with the ?ag stored in the status register, and singly structured. The status register has the following features: induced update by an update to the data output reg- ister, read-only access, no noti?cation, and singly structured. Although the data input register will not be prefetched, its features are: volatile-update at a random rate, write-only access, no noti?cation, and singly structured. We refer to this example as Core2. (3) Core3?input-buffer registers: Many cores have a combination of a queue data buffer that receives data and a status register that indicates the number of bytes in the buffer. A common example of such a core is a UART. Features of the data buffer include: volatile-update at a random rate, read-only access, register-based ?ag noti?cation stored in the status register, and queue-structured. The status register features include: induced-update by an update to the data register, read-only access, no noti?cation, and singly structured. We refer to this example as Core3. 4.4 Prefetching Architectures and Heuristics 4.4.1 Architecture. In order to implement the prefetching for each of the above listed combinations of registers, we developed architectures for bus wrappers for each. Figure 6 illustrates the architecture for each of the three combinations respectively. Each BW architecture has three regions: (1) Controller: The controller's main task is to interface with the on-chip bus. It thus handles reads and writes from and to the core's registers. For a write, the controller writes the data over the core internal bus to the core internal register. For a read, the controller outputs the appropriate prefetch register data onto the bus; for a hit, this outputting is done immediately, while for a miss, it is done only after forcing the prefetch unit to ?rst read the data from the core internals. (2) Prefetch registers: These registers are directly connected to the on-chip bus for fast output. Any output to the bus must pass through one of these registers. 14 Lysecky and Vahid Bus wrapper architecture and timing diagrams for (a) Core1, (b) Core2, and (c) Core3. Fig. 6. Prefetching for Improved Bus Wrapper Performance 15 (3) Prefetch unit: The PFU implements the prefetch heuristics, and is responsible for reading data from the core internals to the prefetch registers. Its goal is to maximize hits. The architecture for the Core1 situation is shown in Figure 6(a), showing one register D and its corresponding prefetch register D0. Since D is only updated by the on-chip bus, no prefetch unit is needed; instead, we can write to D0 whenever we write to D. Such a lack of a PFU is an exception to the normal situation. Figure 6(b) shows the architecture for the Core2 situation. The data output register DO and status register S both have prefetch registers in the BW, but the data input register DI does not since it is never read by the on-chip bus. The PFU carries out its prefetch heuristic (see next section), unless the controller asserts the ?writing? line, in which case the PFU suspends prefetching so that the controller may write to DI over the core internal bus. Figure 6(c) shows the architecture for the Core3 example, which has no write-access registers and hence does not include the bus between the controller and the core internal bus. 4.4.2 Heuristics. We applied the following prefetch heuristics within each core's bus wrapper: Upon a system write to the data register D, simultaneously write the data into the prefetched data register D0. This assumes that a write to the data register will occur prior to a read from the register. After the system writes to the data input register DI, we read the core's internal status register S into the prefetched status register S0.Ifthe status indicates completion, we read the core's internal data output register DO into the prefetched data-output register DO0. We repeat this process. We continuously read the core's internal status register S into the prefetched status register S0 until the status indicates the buffer is no longer empty. We then read the core's data register D into the prefetched data register D0. While waiting for the system to read the data, we continuously read the core's internal status register into the prefetched status register, thereby providing the most current status information. When the data is read by the system, depending on whether the buffer is empty, we either read the next data item from the core or repeat the process. Figure 6 shows timing diagrams for the three cores with a BW and prefetch- ing. In all three cores, the read latency for each core with a BW and prefetching was equal to the latency of that core without a BW, thus eliminating the performance penalty. Note that a BW's architecture and heuristic are dependent on the core internals. This is acceptable since the core developer builds the BW. The BW controller's bus interface is not, however, dependent on the core internals, as desired. 4.5 Experiments We implemented cores representing the three earlier common examples, in order to evaluate performance, power, and size tradeoffs achievable through ACM Transactions on Design Automation of Electronic Systems, Vol. 7, No. 1, January 2002. Table II. Impact of Prefetching on Several Cores Size w/o BW (gates) Size w/BW w/o PF (gates) Size w/BW w/PF (gates)26692638617211506Performance w/o BW (ns) Performance w/BW w/o PF (ns) Performance w/BW w/PF (ns)9835551555454305Power w/o BW (microwatts) Power w/BW w/o PF (microwatts) Power w/BW w/PF (microwatts)13994805601521Energy w/o BW (nJ) Energy w/BW w/o PF (nJ) Energy w/BW w/PF (nJ)13.762.653.116.55prefetching. Results are summarized in Table II. All three cores were written as soft cores in register-transfer-level behavioral VHDL. The three cores required 136, 220, and 226 lines of VHDL, respectively. We synthesized the cores using Synopsys Design Compiler. Performance, average power, and energy metrics were measured using Synopsys analysis tools, using a suite of core test vectors for each core. It is important to note that these cores have simple internal behavior and were used for experimentation purposes only. Although these examples are small, because the PFU unit is independent of the core internals our approach can be applied to larger examples as well. In all three cores, when prefetching was added to the BW's, any performance penalty was effectively eliminated. In Core2 and Core3, there was a trivial one-time 30-ns and 10-ns overhead associated with the initial time required to start and restart the prefetching process for the particular prefetch heuristics. The addition of a BW to cores adds size overhead to the design, but size constraints continue to relax as chip capacities continue their exponential growth. In the three cores described above, there was an average increase in the size of each core by 1352 gates. The large percentage increase in size for Core1 and was due to the fact that these cores were unusually small to begin with since they had only simple internal behavior, having only 1000 or 2000 gates; more typical cores would have closer to 10,000 or 20,000 gates, so the percentage increase caused by the few thousand extra gates would be much smaller. In order for prefetching to be a viable solution to our problem, power and energy consumption must also be acceptable. Power is a function of the amount of switching in the core, while energy is a function of both the switching and the total execution time. BWs without prefetching caused both an increase in power (due to additional internal transfers to the BW) and an increase in overall energy consumption (due to longer execution time) in all three cores. Compared to BWs without prefetching, BWs with prefetching may increase or decrease power depending on the prefetch heuristic and particular application. For ex- ample, in Core1 and Core3, there was an increase in power due to the constant activity of the prefetch unit, but in Core2, there was a decrease in power due to the periods of time during which the prefetch unit was idle. However, in all three cores, the use of prefetching in the BW decreased energy consumption over the ACM Transactions on Design Automation of Electronic Systems, Vol. 7, No. 1, January 2002. Prefetching for Improved Bus Wrapper Performance 17 Table III. Impact of Prefetching on Digital Camera Performance Reads Cycles w/o prefetching Cycles w/ prefetching CCD?Status CCD?Data CODEC?Status CODEC?Data25625710241028512514 Total for 2 cores 772 3088 1544 Digital Camera Peripheral I/O Access Digital Camera Processor Execution Digital Camera 48,616 47,072 Table IV. Impact of Prefetching on Digital Camera Power/Energy prefetching BW w/prefetching Power, mW 95.4 98.1 98.1 Energy, J 44.9 47.7 46.2 BW without prefetching because of reduced execution time. In addition, the increase in energy consumption relative to the core without a bus wrapper was fairly small. To further evaluate the usefulness of prefetching, we analyzed a digital camera as shown in Figure 5. We initially had implemented the CCD and CODEC cores using BWs without prefetching. We therefore modi?ed them to use prefetching, and compared the two versions of the digital camera system. Table III provides the number of cycles for reading status and data registers for the two cores to capture one picture frame. The number of cycles required for these cores with prefetching was half of the number of cycles required without prefetching. The improvement in performance for reads from the CCD and CODEC was 50%. The overall improvement in performance for the digital camera was over 1500 cycles just by adding prefetching to these two cores, out of a total of about 47,000 cycles to capture a picture frame. The prefetching performance increase of the digital camera was directly related to the ratio of I/O access to processor computation. Because the digital camera spends 78% of execution time performing computation and only 12% performing I/O access, prefetching did not have a large impact on overall performance. However, the increase in performance for peripheral I/O access was 25%. Therefore, for a design that is more I/O intensive, one would expect a greater percentage performance increase. Furthermore, if the processor was pipelined, the number of cycles required for program execution would decrease, and the percentage of time required for I/O access would increase. Thus, one would again expect a greater percentage performance increase from prefetching. Adding prefetching to other cores would of course result in even further reductions. The power and energy penalties are shown in Table IV. We see that, in this example, prefetching is able to eliminate any performance overhead associated with keeping interface and internals separated in a core. Prefetching enables elimination of the performance penalty while fully supporting the idea of a VSI standard for the internal bus between the BW and core internals. It can also be varied to tradeoff performance with size and power; ideally, a future tool would synthesize a BW satisfying the power, performance, and size constraints given by the user of a core. 5. ?REAL-TIME" PREFETCHING 5.1 Overview One of the drawbacks to the prefetching technique described above is that the prefetch unit was manually designed and created. We desired to also investigate an automatic solution to designing a prefetch unit. The bus wrapper in our automated approach has an identical architecture to our previous bus wrapper with prefetching. However, we now rede?ne the task of the prefetch unit (PFU). The prefetch unit is responsible for keeping the prefetch registers as up-to-date as possible, by prefetching the core's internal registers over the internal bus, when the internal bus is not being used for a write by the controller, i.e., during internal bus idle cycles. Only one register can be read from the core internals at a time. We assume we are given a list of the core's readable registers, which must be prefetched. We also assume that the bus wrapper can accommodate one copy of each such register. Each register in the list is annotated with two important read-access constraints: ?Register age constraint: This constraint represents the number of cycles old that data may be when read. In other words, it represents the period during which the prefetch register must be updated at least once. An age constraint of 0 means that the data must be the most recent data, which in turn means that the data must come directly from the core and hence prefetching is not allowed, since prefetched data is necessarily at least one cycle old. A constraint of 0 also means that the access-time constraint must be at least four cycles. ?Register access-time constraint: This constraint represents the maximum number of cycles that a read access may take. The minimum is two, in which case the register must be prefetched. An access-time constraint greater than 2 denotes that additional cycles may be tolerated. We wish to design a PFU that reads the core internal registers into the prefetch registers using a schedule that satis?es the age and access-time constraints on those registers. Note that certain registers may be prefetched more frequently than others if this is required to satisfy differing register access constraints. The tradeoff of prefetching is performance improvement at the expense of size and power. Our main goal is performance improvement, but we should ensure that size and power do not grow more than an acceptable amount. Future work may include optimizing a cost function of performance, size, and power. For example, Figure 7 shows a core with three registers, A, B, and C. We assume that registers A and B are independent registers that are read-only, and updated randomly by the core internals. Assume that A and B have register age constraints of four and six cycles, respectively. We might use a naive prefetching heuristic that prefetches on every idle cycle, reading A 60% and Prefetching for Improved Bus Wrapper Performance 19 Fig. 7. Bus wrapper with prefetching. Table V. Prefetch Scheduling for the Core in Figure 7 Idle cycle Schedule 1 Schedule 213579 A A A A A A A A A B 40% of the time, leading to Schedule 1 in Table V. However, we can create a more ef?cient schedule, as shown in Schedule 2. Although both schedules will meet the constraints, the ?rst schedule will likely consume more power. The naive scheduler also does not consider the effects of register writes, which will be taken into consideration using real-time scheduling techniques. During our investigation for heuristics to solve the prefetching problem, we noticed that the problem could be mapped to the widely studied problem of real-time process scheduling, for which a rich set of powerful heuristics and analysis techniques already exist. We now describe the mapping and then provide several prefetching heuristics (based on real-time scheduling heuristics) and analysis methods. 5.2 Mapping to Real-Time Scheduling A simple de?nition of the real-time scheduling problem is as follows. Given a set of N independent periodic processes, and a set of M processors, we must ACM Transactions on Design Automation of Electronic Systems, Vol. 7, No. 1, January 2002. 20 Lysecky and Vahid order the execution of the N processes onto the M processors. Each process has a period, Pi, a deadline, Di, and a computation time, Ci. The period of a process is the rate at which the process requests execution. The deadline is the length of time in which a process must complete execution after it requests to be executed. Finally, the computation time is the length of time a process takes to perform its computation. Therefore, real-time scheduling is the task of ordering the execution of the N processes among the M processors, to ensure that each process executes once every period Pi and within its deadline Di, where each process takes Ci time to complete. A mapping of the prefetching problem to the real-time process-scheduling problem is as follows. process: A register that must be scheduled for prefetching corresponds to a process that must be scheduled for execution. ?Internal bus ! processor: The internal bus can accommodate only one prefetch at a time. Likewise, a processor can accommodate only one process execution at a time. Thus, the internal bus corresponds to a processor. ?Prefetch ! process execution: A prefetch occurs over the internal bus, and thus corresponds to a process execution occurring on a processor. ?Register age constraint ! process period: The register age constraint de?nes the period during which the register must be prefetched, which corresponds to the period during which a process must be scheduled. ?Register access-time constraint ! process deadline: The access-time constraint de?nes the amount of time a read may take relative to the read request, which corresponds to the amount of time a process must complete its execution relative to the time it requested service. Process computation time: A prefetch corresponds to a process execution, so the time for a prefetch corresponds to the computation time for a process. In this paper, we assume a prefetch requires two cycles, although the heuristics and analysis would of course apply if we extended the register model to allow for (the rather rare) situation where different registers would require different amounts of time to read them from the core internals. Given this mapping, we can now use several known real-time scheduling and analysis techniques to solve the prefetching problem. 5.3 Heuristics 5.3.1 Cyclic Executive Approach. The cyclic executive approach [Burns and Wellings 1997] is a straightforward process scheduling method that can be used for a ?xed set of periodic processes. The approach constructs a ?xed repeating schedule called a major cycle, which consists of several minor cycles of ?xed duration. The minor cycle is the rate at which the process with the highest priority will be executed. The minor cycle is therefore equal to the smallest age of the registers to be prefetched. This approach is attractive due to its simplicity. However, it does not handle sporadic processes (in our case, sporadic writes), Prefetching for Improved Bus Wrapper Performance 21 Table VI. Prefetch Core Descriptions Core Register Max Age D Priority RM PF Time Response Time Util. Util. Bound A all process periods (register-age constraints) must be a multiple of the minor cycle time, and constructing the executive may be computationally infeasible for a large number of processes (registers). To serve as examples, we describe three cores with various requirements. Table VI contains data pertaining to all three of our cores. Table VI contains information regarding maximum register age constraint (Max Age), register access time constraint or deadline (D), rate monotonic priority assignment (Prior- ity RM), time required to prefetch register (PF Time), response time of register (Response Time), utilization for register set (Util.), and utilization bound for register set (Util. Bound). Core1 implements a single-channel DAC converter. Although the analog portion of the converter could not be modeled in VHDL, the technique for converting the analog input was implemented. The core has a single register, DATA, that is read-only and updated randomly externally from the system. Core2 calculates the Greatest Common Divisor (GCD) of three inputs while providing checksum information for the inputs and the result. The core contains three registers, GCD1, GCD2, and CS. The result from the GCD calculator is valid when GCD1 is equal to GCD2. Registers GCD1, GCD2, and CS are independent read-only registers that are updated externally from the system. Core3 has ?ve registers, STAT, BIAS, A, B, and RES. STAT is a status register that is read-only, and indicates the status of the core, i.e., busy or not busy. Registers A and B are read-only registers that are updated randomly from outside the system. RES is a read-only register containing the results of some computation on registers A, B, and BIAS, where BIAS is a write-only register that represents some programmable adjustment in the computation. We can use the cyclic executive approach to create a schedule for each of our three cores. For Core1, both the minor cycle and major cycles are three. For Core2, the minor cycle is 10 and the major cycle is 20. Finally, for Core3, we can construct a cyclic executive with a minor cycle of ?ve and a major cycle of 25. 5.3.2 Rate Monotonic Priority Assignment. A more general scheduling approach can be used for more complex examples, wherein we determine which process to schedule (register to prefetch) next based on a priority scheme. A rate monotonic priority assignment [Burns and Wellings 1997] assigns a priority to each register based upon its age. The register with the smallest age will have the highest priority. Likewise, the register with the largest age will have the lowest priority. For our examples we will use a priority of 1 to indicate the highest priority possible. Rate monotonic priority assignment is known to be ACM Transactions on Design Automation of Electronic Systems, Vol. 7, No. 1, January 2002. 22 Lysecky and Vahid optimal in the sense that if a process set can be scheduled with a ?xed-priority assignment scheme, then the set can also be scheduled with a rate monotonic assignment scheme. We again refer to Table VI for data pertaining to all three of our cores. For Core1, the register age constraint of the register DATA is three cycles. Given that DATA is the only register present, it is assigned the highest prior- ity. Core2's registers GCD1, GCD2, and CS have age constraints of 10, 10, and respectively. Therefore, the corresponding priorities from highest to lowest are GCD1, GCD2, and CS. However, because the register age constraint for GCD1 and GCD2 are equal, the priorities for Core2 could also be, from highest to lowest, GCD2, GCD1, and CS. It is important to note that the priorities of registers with the same age constraint can be assigned arbitrary relative priorities as long as the constraints are met. For Core3, the age constraints for the registers STAT, A, B, and RES are respectively 5, 25, 25, and 10. Therefore, the priority of the registers from highest to lowest would be STAT, RES, A, and B. 5.3.3 Utilization-Based Schedulability Test. The utilization-based schedulability test [8] is used to quickly indicate whether a set of processes can be scheduled, or in our case whether the registers can be prefetched. All N registers of a register set can be prefetched if Equation (1) is true, where Ci is the computation time for register i, Ai is the age constraint of register i, and N is the number of registers to be prefetched. The left-hand side of the equation represents the utilization bound for a register set with N registers, and the right-hand side represents the current utilization of the given register set: If the register set passes this test, all registers can be prefetched and no further schedulability analysis is needed. However, if the register set fails the test, a schedule for this register set that meets all constraints might still exist. In other words, the utilization-based schedulability test will indicate that a register set can be prefetched, but does not indicate that a register set cannot be prefetched. We can analyze our cores to determine whether we can schedule them. From Table VI, we can see that both Core1 and Core2 pass the utilization-based schedulability test with respective utilizations of 66.7% and 50.0%, where the corresponding utilization bounds were 100% and 78.0%. This indicates that we can create a schedule for both of these cores and we do not need to perform any further analysis. However, Core3 has a utilization of 86.0%, but the utilization bound for four registers is 75.7%. Therefore, we have failed the utilization-based schedulability test, though a schedule might still exist. 5.3.4 Response-Time Analysis. Response-time analysis [Burns and Wellings 1997] is another method for analyzing whether a process set (in our case, register set) can be scheduled. However, in addition to testing the schedulability of a set of registers, it also provides the worst-case response time for each register. We calculate the response of a register using Equation ACM Transactions on Design Automation of Electronic Systems, Vol. 7, No. 1, January 2002. Prefetching for Improved Bus Wrapper Performance 23 (2), where Ri is the response time for register i, Ci is the computation time of register i, and Ii is the maximum interference that register i can experience in any time interval [t, t C Ri). The interference of a register is the amount of time that a process must wait while other higher-priority processes execute. Ri D Ci C Ii (2) A register set is schedulable if all registers in the set have a response time less than or equal to their age constraint. From Table VI, we can see that the registers of all three cores will meet their register age constraints. Therefore, it is possible to create a prefetching schedule for all three cores. It is interesting to note that although the utilization-based schedulability test failed for Core3, response time analysis indicates that all of the registers can be prefetched. We refer the reader to Burns and Wellings [1997] for further details on response-time analysis. Writes. We now consider the impact of writes to core registers. Writes come at unknown intervals, and a write ties up the core's internal bus and thus delays prefetches until done. We can therefore view a register write as a high-priority sporadic process. We can attribute a maximum rate at which write commands will be sent to the core. We will also introduce a deadline for a write. The deadline of a write is similar to the access-time for a register being prefetched. This deadline indicates that when a write occurs, it must be completed within the speci?ed number of cycles. In order to analyze how a register write will impact this scheduling, we can create a dummy register, WR, in our register set. The age of the WR register will be the period that corresponds to the maximum rate at which a write will occur. WR's access-time will be equal to its deadline. We can now analyze the register set to determine if a prefetching schedule exists for it. This analysis will provide us with an analysis of the worst case scenario in which a write will occur once every period. 5.3.6 Deadline Monotonic Priority Assignment. Up to this point, we have been interested mainly in a static schedule of the register set. However, because writes are sporadic, we must provide some dynamic mechanism for handling them. Thus, a dynamic scheduling technique should be used because we cannot accurately predict these writes. Therefore, we can use a more advanced priority assignment scheme, deadline monotonic priority assignment [Burns and Wellings 1997]. Deadline monotonic priority assignment assigns a priority to each process (register) based upon its deadline (access-time), where a smaller access-time corresponds to a higher priority. We can still incorporate rate monotonic priority assignment in order to assign priorities to registers with equal access-times. Deadline monotonic priority assignment is known to be optimal in the sense that if a process set can be scheduled by a priority scheme, then it can be scheduled by deadline monotonic priority assignment. For example, in order to accommodate writes to the BIAS register in Core3, we can add the BIAS register to the prefetching algorithm. The deadline for the BIAS register will be such that we can ensure that writes will always have ACM Transactions on Design Automation of Electronic Systems, Vol. 7, No. 1, January 2002. Fig. 8. Performance in ns (top), size in gates (middle), and energy in nJ (bottom). the highest priority when we use the deadline monotonic priority assignment. Using this priority assignment mechanism, the priority of the registers from highest to lowest would be BIAS, STAT, RES, A, and B. 5.4 Experiments with ?Real-Time? Prefetching In addition to implementing the ADJUST core as described above, we implemented two additional examples in order to evaluate the impact on perfor- mance, size, and energy using our real-time pre-fetching. The CODEC core contains three registers DIN, DOUT, and STAT. This core behaves like a simple compressor/decompressor, whereby the input data is modi?ed via some arbitrary translation, after which the STAT register is updated to re?ect completion. The FIFO core contains two registers DATA and STAT. This core represents a simple FIFO that has data stored in DATA and the current number of items in the FIFO stored in STAT. We modeled the cores as synthesizable register-transfer VHDL models, requiring 215, 204, and 253 lines of code, respectively?note that we intentionally did not describe internal behavior of the cores, but rather just the register- access-related behavior, so we could see the impacts of prefetching most clearly. We used Synopsys Design Compiler for synthesis as well as Synopsys power analysis tools. Figure 8 summarizes the results for the three cores. For each core, we examined three possible bus wrapper con?gurations: no bus wrapper (No BW), a bus wrapper without prefetching (BW), and a bus wrapper with real-time prefetching (RTPF). The ?rst chart in Figure 8 summarizes performance results. Using our real-time prefetching heuristic, we can see a good performance improvement when compared to a bus wrapper without prefetching. However in FIFO, we ACM Transactions on Design Automation of Electronic Systems, Vol. 7, No. 1, January 2002. Prefetching for Improved Bus Wrapper Performance 25 only see a very small performance improvement using real-time prefetch- ing. This small improvement is due to fact that the DATA register in FIFO cannot be prefetched using this approach. If we were to prefetch using real-time prefetching, we would empty the FIFO and lose data. Fur- thermore, without any prefetching, we can see a signi?cant performance penalty. The second chart in Figure 8 summarizes size results. As expected, the size of the cores increased when a bus wrapper was added, and further increased when prefetching was added to the bus wrapper. The average increase in size caused by adding real-time prefetching to the bus wrapper was only 1.4K gates. This increase in design complexity was due to the need to keep track of current register ages. Furthermore, this size increase was relatively small when compared to a typical core size of 10K to 20K gates. The third chart in Figure 8 summarizes energy consumption for our test vectors. In all three cores, there was an overall increase in energy consumption when a bus wrapper was added to the core. However, the addition of prefetching to the bus wrappers did not always strictly increase or decrease energy consumption. In fact, real-time prefetching increased energy consumption in CODEC and FIFO, and decreased energy consumption in ADJUST. As expected, when compared to the core without a bus wrapper, prefetching resulted in an increase in energy consumption. 6. UPDATE-DEPENDENCY BASED PREFETCHING USING PETRI NETS 6.1 Overview In some cases, a core designer may be able to provide us more information regarding when the core's internal registers get updated?in particular, update dependencies among registers, e.g., if register A is updated externally, then register B will be updated one cycle later. Using this information, we can design a schedule that performs fewer prefetches to satisfy given constraints, and thus can yield advantages of being able to handle more complex problems, or of using less power. 6.2 General Register Attributes We need a method for capturing the information a designer provides regarding register updates. In Section 3.2, we provided a taxonomy of register attributes can be used to categorize how a register is used. We extend this by introducing update dependecies. Update dependencies provide further details on when a register gets updated as a result of other updates (inducements). There are two kinds of update dependencies: ?Internal dependencies: Dependencies between registers must be accurately described. Dependencies between registers affect both the operation of the core and the time at which registers are updated. Therefore, these dependencies are extremely important in providing an accurate model of a core's behavior. 26 Lysecky and Vahid ?External dependencies: Updates to registers via reads and writes over the OCB also need to be included in our model. This information is important because reads and writes can directly update registers or trigger updates to other registers, e.g., a write to a control register of a CODEC core will trigger an event that will update the output data register. Likewise, updates from external ports to internal core registers must also be present in our model. These events occur at random intervals and cannot be directly monitored by a bus wrapper and are therefore needed to provide a complete model of a core. We needed to create a model to capture the above information. After analyzing many possible models to describe both internal and external update depen- dencies, we concluded that a Petri net model would best ?t our requirements. 6.3 Petri Net Model Construction As in all Petri net models, we have places, arcs, and transitions. In our model, a place represents data storage, i.e., a register, or any bus that the bus wrapper can monitor. In this model, a bus place will generate tokens that will be out- puted over all outgoing arcs and consumed by data storage places whenever an appropriate transition is ?red. A transition represents an update dependency between either the bus and a register or between two registers. Transitions may be labeled with conditions that represent some requirement on the data coming into a transition. However, in many cases, a register may be updated from some external source, i.e., the register's update-type is volatile. Therefore, we need a mechanism to describe such updates. We will use a transition without incoming arcs and without an associated condition to represent this behavior. We will refer to such a transition as a random transition. Given random transi- tions, tokens can also be generated by external sources that cannot be directly monitored by the bus wrapper. Thus, our model provides a complete description of the core's internal register dependencies without providing all details of the core's internal behavior. We implemented three core examples to analyze our update dependency model and prefetching technique. In order to demonstrate the usefulness of our model we will describe one of the cores we implemented, which we will refer to as ADJUST, and elaborate on this example throughout the paper. ADJUST contains three registers GO, MD, and S. First, we annotate each register with the general register attributes described earlier. The GO register has the attributes of static-update, write access, no noti?cation, and singly structured. The MD register has the attributes of volatile-update, read/write access, no noti?cation, and singly structured. Finally, the S register has the attributes of volatile update, read-only access, no noti?cation, and singly structured. Next, we constructed the Petri net for ADJUST. Figure 9 shows the register update dependency model for ADJUST. From this model we can see how each register is updated. GO is updated whenever a write request for GO is initiated on the OCB. S is updated randomly by some external event that is unknown to the prefetch unit. MD is updated when GO is equal to 1, a write request for MD is initiated on the OCB, and some external ACM Transactions on Design Automation of Electronic Systems, Vol. 7, No. 1, January 2002. Prefetching for Improved Bus Wrapper Performance 27 Fig. 9. ADJUST register dependencies. event occurs. Therefore, we now have a complete model of the ADJUST core that can be used to create a prefetching algorithm. Using the current model of ADJUST, we need three prefetch registers in the bus wrapper, namely, GO0,MD0, and S0.GO0 would be updated whenever a write to GO was initiated over the OCB. For MD and S, we need some method of refreshing the prefetch registers to keep them as up-to-date as possible. We will later discuss the heuristics for updating registers with incoming random transitions. However, we further know that prefetching the MD register would not be required until a write to MD was made over the OCB and the GO register was equal to 1. This simple interpretation of the model will reduce the power consumed by the prefetch unit by not prefetching MD if it is not needed. 6.4 Model Re?nement for Dependencies Further re?nement of our register update dependency model can be made to eliminate some random transitions. Although the model of a particular core may have many random transitions, there may exist some relationships between the registers with random transitions. If two registers are both updated by the same external event, it is possible that a relationship may exist between the registers. For example, in a typical CODEC core, we would ?nd a data register and a status register. When the data register is updated, the status register is also updated to indicate the operation has completed. Although both registers are updated at random times, we know that if the status register indicates comple- tion, then the data register has been updated. We can thus eliminate one random transition by replacing the random transition with a transition having an incoming arc from the related register and assigning an appropriate condition to this transition. Thus, we have successfully re?ned our model to eliminate a random transition. The goal of this re?nement is to eliminate as many random transitions as possible, but it is important to note that it is not possible to eliminate all random transitions. Therefore, we still need a method for refreshing the contents of registers with incoming random transitions. Figure shows a re?ned register update dependency model for the ADJUST core. In this new model, we have eliminated one random transition by replacing ACM Transactions on Design Automation of Electronic Systems, Vol. 7, No. 1, January 2002. 28 Lysecky and Vahid Fig. 10. Re?ned ADJUST register dependencies. it with a transition that will ?re if S is equal to 1. Hence, we need to modify our prefetching algorithm to accompany this change. We now know that we only need to prefetch MD if S is equal to 1, GO is equal to 1, and a write to MD was initiated over the OCB. This re?nement further simpli?es our prefetching algorithm and will again reduce power consumption. 6.5 Prefetch Scheduling Given an update dependency model of a core, we need to construct a schedule to prefetch the core's registers into the bus wrapper's prefetch registers. Figure 11 describes our update dependency model prefetching heuristic using pseudo- code. Our heuristic uses the update dependency model in conjunction with our real-time prefetching to create a schedule for prefetching the core's registers. The following description will further elaborate on the heuristic. In order to implement our prefetching heuristic, we will need two data struc- tures. The ?rst data structure needed is a prefetch register heap, or priority queue, used to store the registers that need to be prefetched. Second, we need a list of update arcs that must be analyzed after a register is prefetched or a read or write request is detected on the OCB. Using these data structures, we will next describe how the prefetch unit will be designed. The ?rst step in our prefetching heuristic is to add all registers with incoming random transitions to the prefetch register heap. These registers will always remain in the heap because they will need to be repeatedly prefetched in order to satisfy their register age constraints. Next, our prefetch heuristic needs to respond to read and write requests on the OCB. In the event of a read request, the prefetch unit will add any outgoing arcs to the list of arcs needed to be analyzed. As described in our real-time prefetching work, a write is treated as another register with special age and access-time constraints, i.e., the register age constraint is 0 and the access-time constraint is initially set to in?nity. Because the core internal bus may be currently in use performing a prefetch, we use this mechanism to eliminate any contention. As described below, by setting the access-time constraint on the write register to 0, we will ensure that the write will be the next action performed. Therefore, a write request will be handled by ?rst copying the data ACM Transactions on Design Automation of Electronic Systems, Vol. 7, No. 1, January 2002. Prefetching for Improved Bus Wrapper Performance 29 Fig. 11. General register model prefetching heuristic used to implement PFU. into the corresponding prefetch register, setting the access-time constraint to 0, and adding the write register to the prefetch register heap. In addition, any outgoing arcs will be added to the list of update arcs. We will use our real-time prefetching to prefetch registers according to their priorities as assigned by the deadline monotonic priority assignment. When two registers have the same priority assigned by this mechanism, we will use the priority assigned by the rate-monotonic priority assignment to schedule the prefetching. According to this heuristic, registers with an access-time constraint of 0 will be prefetched ?rst. That means that all write requests and, as we will describe later, all registers that have been updated will be prefetched ?rst. Note that writes will still take highest priority because their register age constraint is 0. If no write requests or registers without incoming random ACM Transactions on Design Automation of Electronic Systems, Vol. 7, No. 1, January 2002. transitions need to be prefetched, our prefetching heuristic will next schedule registers with incoming random transitions according to their rate-monotonic priority assignment. Therefore, our prefetch register heap will be sorted ?rst by deadline-monotonic priority assignment and further by rate-monotonic priority assignment. After each register prefetch is made or a read or write request is detected on the OCB, we need to analyze all arcs in the update arc list. If any transition ?res, the outgoing arcs of this transition will be added to the list. If a token reaches another place, we set the corresponding register's access-time to 0 and add it to the heap, thus ensuring that this register is prefetched as soon as possible. In order to better understand this prefetching heuristic, we will look at the ADJUST core. In ADJUST, we have one random transition which is connected to the S register. We noticed that in our design, on average, we only needed to read the contents of S every six cycles. Therefore, we set the register age constraint to six cycles, and the register access-time constraint to two, indicating that the register S must be prefetched every six cycles. For MD, both the register age and access-time constraints are two cycles. GO, however, has neither an age constraint nor an access-time constraint because it is a write-only register. Note that even though GO is a write-only register, a copy must be maintained in the bus wrapper, as it is needed in order to analyze the update dependencies. Our prefetching algorithm will monitor the OCB. If a write to the GO register is made, the data will be copied into GO0, and the write register access-time will be set to 0. On a write to the MD register, the access-time of S will be set to Also if GO is equal to 1 and S is equal to 1, then set the access-time for MD to 0. Finally, we will use the scheduling above to prefetch the registers when needed and perform write operations. 6.6 Experiments with Update-Dependency Prefetching We implemented the update dependency prefetching on the same three cores as above, namely ADJUST, CODEC, and FIFO. Figure 12 summarizes the results for the three cores. We now have four possible bus wrapper con?gurations for each core: no bus wrapper (no BW), a bus wrapper without prefetching (BW), a bus wrapper with real-time prefetching (RTPF), and a bus wrapper with our update dependency prefetching model (UDPF). The ?rst chart in Figure 12 summarizes performance results. In all three cores, the use of our update dependency prefetching method almost entirely eliminated the performance penalty associated with the bus wrapper. There was still a slight overhead caused by starting the prefetch unit. Using our real-time prefetching heuristic, we can see that although there is a performance improvement when compared to a bus wrapper without prefetching, it did not perform as well as our update dependency model. The second chart in Figure 12 summarizes size results. The average increase in size caused by adding the update dependency prefetching technique to the bus wrapper was only 1.5K gates. In comparison, real-time prefetching resulted in an average increase of 1.4K gates. It is interesting to note why the two ACM Transactions on Design Automation of Electronic Systems, Vol. 7, No. 1, January 2002. Prefetching for Improved Bus Wrapper Performance 31 Fig. 12. Performance in ns (top), size in gates (middle), and energy in nJ (bottom). approaches, although quite different, resulted in approximately the same size increase. As stated earlier, using real-time prefetching, we increase the design complexity due to the need to keep track of current register ages. However, using our extended approach, complexity increases due to added logic needed to analyze update dependencies. The third chart in Figure 12 summarizes energy consumption for our test vectors. However, the addition of prefetching to the bus wrappers does not always strictly increase or decrease energy consumption. In fact, we can see that in ADJUST and FIFO, there is a decrease in energy consumption when our update dependency prefetching is added to the bus wrapper, but in CODEC, there is an increase. On the other hand, real-time prefetching increases energy consumption in CODEC and FIFO, and decreases energy consumption in ADJUST. More importantly, if we compare the results of our real-time prefetching to our update dependency prefetching, we notice that the update dependency prefetching results in signi?cantly less energy consumption. This is easily explained by the fact that this approach only prefetches registers when they have been updated whereas our real-time prefetching will prefetch registers more often to keep them as up-to-date as possible. Therefore, by eliminating the need to prefetch all registers within their register age constraints, we can reduce energy consumption. 7. CONCLUSIONS While keeping a core's interface and internal behavior separated is key to a core's marketability, we demonstrated that the use of such bus wrappers, both non-PVCI and PVCI, results in size, power, and performance overhead. Thus, the retargetability advantages of such a standard seem to come with some penalties. We introduced prefetching as a technique to overcome the performance over- head. We demonstrated that in some common cases of register combinations, prefetching eliminates the performance degradation at the expense of acceptable increases in size and power. By overcoming the performance degradation associated with bus wrappers, prefetching thus improves the usefulness of cores. We have further provided a powerful solution to this problem by mapping the problem to the real-time process-scheduling domain, and then applying heuristics and analysis techniques from that domain. We also provided a general register update dependency model that we used to construct a more ef?cient prefetching schedule, in conjunction with our real-time prefetching. We demonstrated the effectiveness of these solutions through several experiments, showing good performance improvements with acceptable size and energy increases. Furthermore, we demonstrated that using our update dependency model we were able to better prefetch registers when compared to our real-time prefetching methodology. The two approaches are thus complementary?the real-time approach can be used when only register constraints are provided, while the model-based approach of this paper can be used when register update information is also provided. 8. FUTURE WORK Although prefetching works well, there are many possibilities for improve- ments. In our current approach we assume that all registers of the core will be prefetched. However, for cores with large numbers of registers, this approach is not feasible. Thus, we are considering restricting the number of registers that can appear in a bus wrapper. This creates new cache-like issues such as mapping, replacement, and coherency issues that are not present in our current design. In addition, we can further evaluate the effects of prefetching on larger core examples. Another direction involves developing prefetching heuristics that optimize a given cost function of performance, power, and size. --R Interface co-synthesis techniques for embedded systems Fast prototyping: a system design Java driven codesign and prototyping of networked embedded systems. A new direction for computer architecture research. Description and simulation of hard- ware/software systems with Java Interface design for core-based systems Inventra core library. Computer architecture Introduction to rapid silicon prototyping: hardware-software co-design for embedded systems-on-a-chip ICs ASSOCIATION. <Year>1999</Year>. International technology roadmap for semiconductors: The case for a con? An object-oriented communication library for hardware-software co-design Experiences with system level design for consumer ICs. Constructing application-speci?c heterogeneous embedded architectures from custom HW/SW applications VIRTUAL SOCKET INTERFACE ASSOCIATION. VIRTUAL SOCKET INTERFACE ASSOCIATION. VIRTUAL SOCKET INTERFACE ASSOCIATION. accepted May --TR Computer architecture: a quantitative approach Real-time systems and their programming languages Interface co-synthesis techniques for embedded systems Constructing application-specific heterogeneous embedded architectures from custom HW/SW applications Interface-based design The case for a configure-and-execute paradigm Fast prototyping Java driven codesign and prototyping of networked embedded systems Scheduling Algorithms for Multiprogramming in a Hard-Real-Time Environment A New Direction for Computer Architecture Research Interface Design for Core-Based Systems An Object-Oriented Communication Library for Hardware-Software CoDesign Bus-Based Communication Synthesis on System-Level --CTR Ken Batcher , Robert Walker, Cluster miss prediction for instruction caches in embedded networking applications, Proceedings of the 14th ACM Great Lakes symposium on VLSI, April 26-28, 2004, Boston, MA, USA Minas Dasygenis , Erik Brockmeyer , Bart Durinck , Francky Catthoor , Dimitrios Soudris , Antonios Thanailakis, A combined DMA and application-specific prefetching approach for tackling the memory latency bottleneck, IEEE Transactions on Very Large Scale Integration (VLSI) Systems, v.14 n.3, p.279-291, March 2006

PVCI;VSIA;system-on-a-chip;interfacing;cores;bus wrapper;on-chip bus;design reuse;intellectual property

505243

An optimal minimum spanning tree algorithm.

We establish that the algorithmic complexity of the minimum spanning tree problem is equal to its decision-tree complexity. Specifically, we present a deterministic algorithm to find a minimum spanning tree of a graph with n vertices and m edges that runs in time O(T*(m,n)) where T* is the minimum number of edge-weight comparisons needed to determine the solution. The algorithm is quite simple and can be implemented on a pointer machine.Although our time bound is optimal, the exact function describing it is not known at present. The current best bounds known for T* are (m) and O(m (m,n)), where is a certain natural inverse of Ackermann's function.Even under the assumption that T* is superlinear, we show that if the input graph is selected from Gn,m, our algorithm runs in linear time with high probability, regardless of n, m, or the permutation of edge weights. The analysis uses a new martingale for Gn,m similar to the edge-exposure martingale for Gn,p.

Introduction The minimum spanning tree (MST) problem has been studied for much of this century and yet despite its apparent simplicity, the problem is still not fully understood. Graham and Hell [GH85] give an excellent survey of results from the earliest known algorithm of Boruvka [Bor26] to the invention of Fibonacci heaps, which were central to the algorithms in [FT87, GGST86]. Chazelle [Chaz97] presented an MST algorithm based on the Soft Heap [Chaz98] having complexity O(m(m;n) log (m; n)), where is a certain inverse of Ackermann's function. Recently Chazelle [Chaz00] modied the algorithm in [Chaz97] to bring down the running time to O(m (m; n)). Later, and in independent work, a similar algorithm of the same running time was presented in Pettie [Pet99], which gives an alternate exposition of the O(m (m; n)) result. This is the tightest time bound for the MST problem to date, though not known to be optimal. This is an updated version of UTCS Technical Report TR99-17 which includes performance analysis on random graphs and new references. Part of this work was supported by Texas Advanced Research Program Grant 003658- 0029-1999. Seth Pettie was also supported by an MCD Fellowship. All algorithms mentioned above work on a pointer machine [Tar79] under the restriction that edge weights may only be subjected to binary comparisons. If a more powerful model is assumed, the MST can be computed optimally. Fredman and Willard [FW90] showed that on a unit-cost RAM where the bit-representation of edge weights may be manipulated, the MST can be computed in linear time. Karger et al. [KKT95] presented a randomized MST algorithm that runs in linear time with high probability, even if edge weights are only subject to comparisons. It is still unknown whether these more powerful models are necessary to compute the MST in linear time. However, in this paper we give a deterministic, comparison-based MST algorithm that runs on a pointer machine in O(T (m; n)) time, where T (m; n) is the number of edge-weight comparisons needed to determine the MST on any graph with m edges and n vertices. Additionally, we show that our algorithm runs in linear time for the vast majority of graphs, regardless of density or the permutation of edge weights. Because of the nature of our algorithm, its exact running time is not known. This might seem paradoxical at rst. The source of our algorithm's optimality, and its mysterious running time, is the use of precomputed 'MST decision trees' whose exact depth is unknown but nonetheless provably optimal. A trivial lower bound on our algorithm is m); the best upper bound, O(m(m;n)), is due to Chazelle [Chaz00]. We should point out that precomputing optimal decision trees does not increase the constant factor hidden by big-Oh notation, nor does it result in a non-uniform algorithm. Our optimal MST algorithm should be contrasted with the complexity-theoretic result that any optimal verication algorithm for some problem can be used to construct an optimal algorithm for the same problem [Jo97]. Though asymptotically optimal, this construction hides astronomical constant factors and proves nothing about the relationship between algorithmic complexity and decision-tree complexity. See Section 8 for a discussion of these and other related issues. In the next section we review some well-known MST results that are used by our algorithm. In section 3 we prove a key lemma and give a procedure for partitioning the graph in an MST- respecting manner. Section 4 gives an overview of the optimal algorithm and discusses the structure and use of pre-computed decision-trees for the MST problem. Section 5 gives the algorithm and a proof of optimality. Section 6 shows how the algorithm may be modied to run on a pointer machine. In section 7 we show our algorithm runs in linear-time w.h.p. if the input graph is selected at random. Sections 8 & 9 discuss related problems and algorithms, open questions, and the actual complexity of MST. Preliminaries The input is an undirected graph E) where each edge is assigned a distinct real-valued weight. The minimum spanning forest (MSF) problem asks for a spanning acyclic subgraph of G having the least total weight. In this paper we assume for convenience that the input graph is connected, since otherwise we can nd its connected components in linear time and then solve the problem on each connected component. Thus the MSF problem is identical to the minimum spanning tree problem. It is well-known that one can identify edges provably in the MSF using the cut property, and edges provably not in the MSF using the cycle property. The cut property states that the lightest edge crossing any partition of the vertex set into two parts must belong to the MSF. The cycle property states that the heaviest edge in any cycle in the graph cannot be in the MSF. 2.1 Boruvka steps The earliest known MSF algorithm is due to Boruvka [Bor26]. The algorithm is quite simple: It proceeds in a sequence of stages, and in each stage it executes a Boruvka step on the graph G, which identies the set F consisting of the minimum-weight edge incident on each vertex in G, adds these edges to the MSF (since they must be in the MSF by the cut property), and then forms the graph as the input to the next stage, where GnF is the graph obtained by contracting each connected component formed by F . This computation can be performed in linear time. Since the number of vertices reduces by at least a factor of two, the running time of this algorithm is O(m log n), where m and n are the number of vertices and edges in the input graph. Our optimal algorithm uses a procedure called Boruvka2(G; F; G 0 ). This procedure executes two Boruvka steps on the input graph G and returns the contracted graph G 0 as well as the set of edges F identied for the MSF during these two steps. 2.2 Dijsktra-Jarnk-Prim Algorithm Another early MSF algorithm that runs in O(m log n) time is the one by Jarnk [Jar30], re-discovered by Dijkstra [Dij59] and Prim [Prim57]. We will refer to this algorithm as the DJP algorithm. Brie y, the DJP algorithm grows a tree T , which initially consists of an arbitrary vertex, one edge at a time, choosing the next edge by the following simple criterion: Augment T with the minimum weight edge (x; y) such that x 2 T and y 62 T . By the cut property, all edges in T are in the MSF. Lemma 2.1 Let T be the tree formed after the execution of some number of steps of the DJP algorithm. Let e and f be two arbitrary edges, each with exactly one endpoint in T , and let g be the maximum weight edge on the path from e to f in T . Then g cannot be heavier than both e and f . Proof: Let P be the path in T connecting e and f , and assume the contrary, that g is the heaviest edge in P [ fe; fg. Now consider the moment when g is selected by DJP and let P 0 be the portion of P present in the tree. There are exactly two edges in (P P which are eligible to be chosen by the DJP algorithm at this moment, one of which is the edge g. If the other edge is in P then by our choice of g it must be lighter than g. If the other edge is either e or f then by our assumption it must be lighter than g. In both cases g could not be chosen next by the DJP algorithm, a contradiction. 2 2.3 The Dense Case Algorithm The algorithms presented in [FT87, GGST86, Chaz97, Chaz00, Pet99] will nd the MSF of a graph in linear time if the graph is su-ciently dense, i.e., has a su-ciently large edge-to-vertex ratio. For our purposes, 'su-ciently dense' will mean an (3) n), where n is the number of vertices in the graph. All of the above algorithms run in linear time for that density. The procedure DenseCase(G; F ) takes as input an n-node graph G and returns the MSF F of G in linear time for graphs with density 263 (3) n). Our optimal algorithm will call DenseCase on a graph derived from an n-node, m-edge graph by contracting vertices so that the number of vertices is reduced by a factor of n). The number of edges in the contracted graph is no more than m. It is straightforward to see that DenseCase will run in O(m + n) time on such a graph. 2.4 Soft Heap The main data structure used by our algorithm is the Soft Heap [Chaz98]. The Soft Heap is a kind of priority queue that gives us an optimal tradeo between accuracy and speed. It supports the following operations: MakeHeap(): returns an empty soft heap. Insert(S; x): insert item x into heap S. Findmin(S): returns item with smallest key in heap S. delete x from heap S. create new heap containing the union of items stored in S 1 and S 2 , destroying S 1 and S 2 in the process. All operations take constant amortized time, except for Insert, which takes O(log( 1 To save time the Soft Heap allows items to be grouped together and treated as though they have a single key. An item adopts the largest key of any item in its group, corrupting the item if its new key diers from its original key. Thus the original key of an item returned by Findmin (i.e. any item in the group with minimum key) is no more than the keys of all uncorrupted items in the heap. The guarantee is that after n Insert operations, no more than n corrupted items are in the heap. The following result is shown in [Chaz98]. Lemma 2.2 Fix any parameter 0 < < 1=2, and beginning with no prior data, consider a mixed sequence of operations that includes n inserts. On a Soft Heap the amortized complexity of each operation is constant, except for insert, which takes O(log(1=)) time. At most n items are corrupted at any given time. 3 A Key Lemma and Procedure 3.1 A Robust Contraction Lemma It is well known that if T is a tree of MSF edges, we can contract T into a single vertex while maintaining the invariant that the MSF of the contracted graph plus T gives the MSF for the graph before contraction. In our algorithm we will nd a tree of MSF edges T in a corrupted graph, where some of the edge weights have been increased due to the use of a Soft Heap. In the lemma given below we show that useful information can be obtained by contracting certain corrupted trees, in particular those constructed using some number of steps from the Dijkstra-Jarnik-Prim (DJP) algorithm. Ideas similar to these are used in Chazelle's 1997 algorithm [Chaz97], and more explicitly in the recent algorithms of Pettie [Pet99] and Chazelle [Chaz00]. Before stating the lemma, we need some notation and preliminary concepts. Let V (G) and E(G) be the vertex and edge sets of G, and n and m be their cardinality, respectively. Let the G-weight of an edge be its weight in graph G (the G may be omitted if implied from context). For the following denitions, M and C are subgraphs of G. Denote by G * M a graph derived from G by raising the weight of each edge in M by arbitrary amounts (these edges are said to be corrupted). Let MC be the set of edges in M with exactly one endpoint in C. Let GnC denote the graph obtained by contracting all connected components induced by C, i.e. by replacing each connected component with a single vertex and reassigning edge endpoints appropriately. We dene a subgraph C of G to be DJP-contractible if after executing the DJP algorithm on G for some number of steps, with a suitable start vertex in C, the tree that results is a spanning tree for C. Lemma 3.1 Let M be a set of edges in a graph G. If C is a subgraph of G that is DJP-contractible w.r.t. G * M , then MSF (G) is a subset of MSF (C) [ MSF (GnC MC ) [ MC . Proof: Each edge in C that is not in MSF(C) is the heaviest edge on some cycle in C. Since that cycle exists in G as well, that edge is not in MSF(G). So we need only show that edges in GnC that are not in MSF(GnC MC are also not in MSF(G). hence we need to show that no edge in H MSF (H) is in MSF (G). Let e be the heaviest edge on some cycle in H (i.e. e 2 H MSF (H)). If does not involve the vertex derived by contracting C, then it exists in G as well and e 62 MSF (G). Otherwise, forms a path P in G whose end points, say x and y, are both in C. Let the end edges of P be (x; w) and included no corrupted edges with one end point in C, the G-weight of these edges is the same as their (G * M)-weight. Let T be the spanning tree of C * M derived by the DJP algorithm, Q be the path in T connecting x and y, and g be the heaviest edge in Q. Notice that P [ Q forms a cycle. By our choice of e, it must be heavier than both (x; y) and (w; z), and by Lemma 2.1, the heavier of (x; y) and (w; z) is heavier than the (G * M)-weight of g, which is an upper bound on the G-weights of all edges in Q. So w.r.t. G-weights, e is the heaviest edge on the cycle P [ Q and cannot be in MSF (G). 2 3.2 The Partition Procedure Our algorithm uses the Partition procedure which is given below. This procedure nds DJP- contractible subgraphs C in which edges are progressively being corrupted by the Soft Heap. Let MC i contain only those corrupted edges with one endpoint in C i at the time it is completed. Each subgraph C i will be DJP-contractible w.r.t a graph derived from G by several rounds of contractions and edge deletions. When C i is nished it is contracted and all incident corrupted edges are discarded. By applying Lemma 3.1 repeatedly we see that after C i is built, the MSF of G is a subset of MSF The Partition procedure is shown in Figure 1. The arguments appearing before the semicolon are inputs; the others are outputs. M is a set of edges and C=fC is a set of subgraphs of G. No edge will appear in more than one of M;C Initially, Partition sets every vertex to be live. The objective is to convert each vertex to dead, signifying that it is part of a component C i with maxsize vertices and part of a conglomerate of maxsize vertices, where a conglomerate is a connected component of the graph Intuitively a conglomerate is a collection of C i 's linked by common vertices. This scheme for growing components is similar to the one given in [FT87]. We grow the C i 's one at a time according to the DJP algorithm, except that we use a Soft Heap. A component is done growing if it reaches maxsize vertices or if it attaches itself to an existing component. Clearly if a component does not reach maxsize vertices, it has linked to a Partition(G; All vertices are initially "live" While there is a live vertex Increment i live vertex Create a Soft Heap consisting of v's edges (uses ) While all vertices in V i are live and jV i j < maxsize Repeat Find and delete min-weight edge (x; y) from Soft Heap Until y If y is live then insert each of y's edges into the Soft Heap all vertices in V i to be dead be the corrupted edges with one endpoint in V i Dismantle the Soft Heap Let C := fC z is the subgraph of G induced by V z Exit. Figure 1: The Partition Procedure. conglomerate of at least maxsize vertices. Hence all its vertices can be designated dead. Upon completion of a component C i , we discard the set of corrupted edges with one endpoint in C i . The running time of Partition is dominated by the heap operations, which depend on . Each edge is inserted into a Soft Heap no more than twice (once for each endpoint), and extracted no more than once. We can charge the cost of dismantling the heap to the insert operations which created it, hence the total running time is O(m log( 1 )). The number of discarded edges is bounded by the number of insertions scaled by , thus jM j 2m. Thus we have Lemma 3.2 Given a graph G, any 0 < < 1 2 , and a parameter maxsize, Partition nds edge-disjoint subgraphs a) For all b) For all i, jV (C i )j maxsize. c) For each conglomerate P 2 d) jE(M)j 2 jE(G)j 4 Overview of the Optimal Algorithm Here is an overview of our optimal MSF algorithm. In the rst stage we nd DJP-contractible subgraphs C with their associated set of edges consists of corrupted edges with one endpoint in C i . In the second stage we nd the MSF F i of each C i , and the MSF F 0 of the contracted graph Gn( 3.1, the MSF of the whole graph is contained within Note that at this point we have not identied any edges as being in the MSF of the original graph G. In the third stage we nd some MSF edges, via Boruvka steps, and recurse on the graph derived by contracting these edges. We execute the rst stage using the Partition procedure described in the previous section. We execute the second stage with optimal decision trees. Essentially, these are hardwired algorithms designed to compute the MSF of a graph using an optimal number of edge-weight comparisons. In general, decision trees are much larger than the size of the problem that they solve and nding optimal ones is very time consuming. We can aord the cost of building decision trees by guaranteeing that each one is extremely small. At the same time, we make each conglomerate formed by the C i to be su-ciently large so that the MSF F 0 of the contracted graph can be found in linear time using the DenseCase algorithm. Finally, in the third stage, we have a reduction in vertices due to the Boruvka steps, and a reduction in edges due to the application of Lemma 3.1. In our optimal algorithm both vertices and edges reduce by a constant factor, thus resulting in the recursive applications of the algorithm on graphs with geometrically decreasing sizes. 4.1 Decision Trees An MSF decision tree is a rooted tree having an edge-weight comparison associated with each internal node (e.g. weight(x; y) < weight(w; z)). Each internal node has exactly two children, one representing that the comparison is true, the other that it is false. The leaves of the tree list the edges in some spanning tree. An MSF decision tree is said to be correct if the edge-weight comparisons encountered on any path from the root to a leaf uniquely identify the spanning tree at that leaf as the MSF. A decision tree is said to be optimal if it is correct and there exists no correct decision tree with lesser depth. Let us bound the time needed to nd all optimal decision trees for graphs of r vertices by brute force search. There are fewer than 2 r 2 such graphs and for each graph we must check all possible decision trees bounded by a depth of r 2 . There are < r 4 possibilities for each internal node and < r 2 r 2 +O(1) decision trees to check. To determine if a decision tree is correct we generate all possible permutations of the edge weights and for each, solve the MSF problem on the given graph. Now we simultaneously check all permutations against a decision tree. First put all permutations at the root, then move them to the left or right child depending on the truth or falsity of the edge-weight comparison w.r.t to each permutation. Repeat this step until all permutations reach a leaf. If for each leaf, all permutations sharing that leaf agree on the MSF, then the decision tree is correct. This process takes no longer than (r for each decision tree. Setting allows us to precompute all optimal decision trees in o(n) time. Observe that in the high-level algorithm we gave in section 4, if the maximum size of each component C i is su-ciently small, the components can be organized into a relatively small number of groups of isomorphic components (ignoring edge weights). For each group we use a single precomputed optimal decision tree to determine the MSF of components in that group. In our optimal algorithm we will use a procedure DecisionTree(G; F), which takes as input a collection of graphs G, each with at most r vertices, and returns their minimum spanning forests in F using the precomputed decision trees. 5 The Algorithm As discussed above, the optimal MSF algorithm is as follows. First, precompute the optimal decision trees for all graphs with log (3) n vertices. Next, divide the input graph into subgraphs discarding the set of corrupted edges MC i as each C i is completed. Use the decision trees found earlier to compute the MSF F i of each C i , then contract each connected component spanned by F (i.e., each conglomerate) into a single vertex. The resulting graph has n= log (3) n vertices since each conglomerate has at least log (3) n vertices by Lemma 3.2. Hence we can use the DenseCase algorithm to compute its MSF F 0 in time linear in m. At this point, by Lemma 3.1 the MSF is now contained in the edge set F . On this graph we apply two Boruvka steps, reducing the number of vertices by a factor of four, and then compute recursively. The algorithm is given below. (this is used by the Soft Heap in the Partition procedure). Precompute optimal decision trees for all graphs with log (3) n 0 vertices, where n 0 is the number of vertices in the original input graph. If r := log (3) jV (G)j Partition(G; G a := Apart from recursive calls and using the decision trees, the computation performed by Opti- malMSF is clearly linear since Partition takes O(m log( 1 owing to the reduction in vertices, the call to DenseCase also takes linear time. For 8 , the number of edges passed to the nal recursive call is m=4 giving a geometric reduction in the number of edges. Since no MSF algorithm can do better than linear time, the bottleneck, if any, must lie in using the decision trees, which are optimal by construction. More concretely, let T (m; n) be the running time of OptimalMSF. Let T (m; n) be the optimal number of comparisons needed on any graph with n vertices and m edges and let T (G) be the optimal number of comparisons needed on a specic graph G. The recurrence relation for T is given below. For the base case note that the graphs in the recursive calls will be connected if the input graph is connected. Hence the base case graph has no edges and one vertex, and we have T (0; 1) equal to a constant. It is straightforward to see that if T (m; n) = O(m) then the above recurrence gives T (m; O(m). One can also show that T (m; n) = O(T (m; n)) for many natural functions for T (including n)). However, to show that this result holds no matter what the function describing T (m; n) is, we need to establish some results on the decision tree complexity of the MSF problem, which we do in the next section. 5.1 Some Results for MSF Decision Trees In this section we establish some results on MSF decision trees that allow us to establish our main result that OptimalMSF runs in O(T (m; n)) time. Proposition 5.1 T (m; n) m=2. Proposition 5.2 For xed m and Proposition 5.1 is obviously true since every edge should participate in a comparison to determine inclusion in or exclusion from the MSF. Proposition 5.2 holds since we can add isolated vertices to a graph, which obviously does not aect the MSF or the number of necessary comparisons. We now state a property that is used by Lemmas 5.4 and 5.5. Property 5.3 The structure of G dictates that are edge-disjoint subgraphs of G. are the components returned by Partition, it can be seen that the graph Denition 5.3 since every simple cycle in this graph must be contained in exactly one of the C i . To see this, consider any simple cycle and let i be the largest index such that C i contains an edge in the cycle. Since each C i shares no more than one vertex with contain an an edge from . The proof of the following lemma can be found in [PR99b]. Lemma 5.4 If Property 5.3 holds for G, then there exists an optimal MSF decision tree for G which makes no comparisons of the form e < f where e 2 C Proof: Consider a subset P of the permutations of all edge weights where for e 2 C holds that weight(e) < weight(f ). Permutations in P have two useful properties which can be readily veried. First, any number of inter-component comparisons shed no light on the relative weights of edges in the same component. Second, any spanning forest of a component is the MSF of that component for some permutation in P. Now consider any optimal decision tree T for G. Let T 0 be the subtree of T which contains only leaves that can be reached by some permutation in P. Each inter-component comparison node in must have only one child, and by the rst property, the MSF at each leaf was deduced using only intra-component comparisons. By the second property, T 0 must determine the MSF of each component correctly, and thus by Property 5.3 it must determine the MSF of the graph G correctly. Hence we can contract T 0 into a correct decision tree T 00 by replacing each one-child node with its only child. 2 Lemma 5.5 If Property 5.3 holds for G, then T Proof: Given optimal decision trees T i for the C i we can construct a decision tree for G by replacing each leaf of T 1 by T 2 , and in general replacing each leaf of T i by T i+1 and by labeling each leaf of the last tree by the union of the labels of the original trees along this path. Clearly the height of this tree is the sum of the heights of the T i , and hence T (G) need only prove that no optimal decision tree for G has height less than the sum of the heights of the T i . Let T be an optimal decision tree for G that has no inter-component comparisons (as guaranteed by Lemma 5.4). We show that T can be transformed into a 'canonical' decision tree T 0 for G of the same height as T , such that in T 0 , all comparisons for C i precede all comparisons for C i+1 , for each i, and further, for each i, the subgraph of T 0 containing the comparisons within C i consists of a collection of isomorphic trees. This establishes the desired result since T 0 must contain a path that is the concatenation of the longest path in an optimal decision tree for each of the C i . We rst prove this result for the case when there are only two components, C 1 and C 2 . Assume inductively that the subtrees rooted at all vertices at a certain depth d in T have been transformed to the desired structure of having the C 1 comparisons occur before he C 2 comparisons, and with all subtrees for C 2 within each of the subtrees rooted at depth d being isomorphic. (This is trivially the case when d is equal to the height of T .) Consider any node v at depth d 1. If the comparison at that node is a C 1 comparison, then all C 2 subtrees at descendent nodes must compute the same set of leaves for C 2 . Hence the subtree rooted at v can be converted to the desired format simply by replacing all C 2 subtrees by one having minimum depth (note that there are only two dierent C 2 subtrees { all C 2 subtrees descendent to the left (right) child of v must be isomorphic). If the comparison at v is a C 2 comparison, we know that the C 1 subtrees rooted at its left child x and its right child y must both compute the same set of leaves for C 1 . Hence we pick the C 1 subtree of smaller height (w.l.o.g. let its root be x) and replace v by x, together with the C 1 subtree rooted at x. We then copy the comparison at node v to each leaf position of this C 1 subtree. For each such copy, we place one of the isomorphic copies of the C 2 subtree that is a descendant of x as its left subtree, and the C 2 subtree that is a descendant of y as its right subtree. The subtree rooted at x, which is now at depth d 1 is now in the desired form, it computes the same result as in T , and there was no increase in the height of the tree. Hence by induction T can be converted into canonical decision tree of no greater height. Assume inductively that the result hold for up to k 1 2 components. The result easily extends to k components by noting that we can group the rst k 1 components as C 0 1 and let C k be C 0 . By the above method we can transform T to a canonical tree in which the C k comparisons appear as leaf subtrees. We now strip the C k subtrees from this canonical tree and then by the inductive assumption we can perform the transformation for remaining k 1 components. 2 Corollary 5.6 Let the C i be the components formed by the Partition routine applied to graph G, and let G have m edges and n vertices. Then, Corollary 5.7 For any m and n, We can now solve the recurrence relation for the running time of OptimalMSF given in the previous section. (Corollary 5.6) (Corollary 5.7 and Propositions 5.1, 5.2) c T (m; n) (for su-ciently large c; this completes the induction) This gives us the desired theorem. Theorem 5.8 Let T (m; n) be the decision-tree complexity of the MSF problem on graphs with m edges and n nodes. Algorithm OptimalMSF computes the MSF of a graph with m edges and n vertices deterministically in O(T (m; n)) time. 6 Avoiding Pointer Arithmetic We have not precisely specied what is required of the underlying machine model. Upon examina- tion, the algorithm does not seem to require the full power of a random access machine (RAM). No bit manipulation is used and arithmetic can be limited to just the increment operation. However, if procedure DecisionTree is implemented in the obvious manner it will require using a table lookup, and thus random access to memory. In this section we describe an alternate method of handling the decision trees which can run on a pointer machine [Tar79], a model which does not allow random access to memory. Our method is similar to that described in [B+98], but we ensure that the time overhead in performing the table lookups during a call to DecisionTree is linear in the size of the current input to DecisionTree. A pointer machine distinguishes pointers from all other data types. The only operations allowed on pointers are assignment, comparison for equality and dereferencing. Memory is organized into records, each of which holds some constant number of pointers and normal data words (integers, oats, etc. Given a pointer to a particular record, we can refer to any pointer or data word in that record in constant time. On non-pointer data, the usual array of logical, arithmetic, and binary comparison operations are allowed. We rst describe the representation of a decision tree. Each decision tree has associated with it a generic graph with no edge weights. This decision tree will determine the MST of each permutation of edge weights for this generic graph. At each internal node of the decision tree are four pointers, the rst two point to edges in the generic graph being compared and the second two point to the left and right child of the node. Each leaf lists the edges in some spanning tree of the generic graph. Since a decision tree is a pointer-based structure, we can construct each precomputed decision tree (by enumerating and checking all possibilities) without using table lookups. We now describe our representation of the generic graphs. The vertices of a generic graph are numbered in order by integers starting with 1, and the representation consists of a listing of the vertices in order, starting from 1, followed by the adjacency list for each vertex, starting with vertex 1. Each generic graph will have a pointer to the root of its decision tree. Recall that we precomputed decision trees for all generic graphs with at most log (3) n 0 vertices (where n 0 is the number of vertices in the input graph whose MSF we need to nd). The generic graphs will be generated and stored in lexicographically sorted order. Note that with our represen- tation, in the sorted order the generic graphs will appear in nondecreasing order of the number of vertices in the graph. Before using a decision tree on an actual graph (which must be isomorphic to the generic graph for that decision tree), we must associate each edge in the actual graph with its counterpart in the generic graph. Thus a comparison between edge weights in the generic graph can be substituted by the corresponding weights in the actual graph in constant time. On a random access machine, we can encode each possible graph in a single machine word (say, as an adjacency matrix), then index the generic graph in an array according to this representation. Thus given a graph we can nd the associated decision tree in constant time. On a pointer machine however, converting a bit vector or an integer to a pointer is specically disallowed. We now describe our method to identify the generic graph for each C i e-ciently. We assume that each C i is specied by the adjacency lists representation, and that each edge (x; y) has a pointer to the occurrence of (y; x) in y's adjacency list. Each edge also has a pointer to a record containing its weight. Let m and n be the number of edges and vertices in We rewrite each C i in the same form as the generic graphs, which we will call the numerical representation. Let C i have p vertices (note that p r). We assign the vertices numbers from 1 to p in the order in which they are listed in the adjacency lists representation, and we rewrite each edge as a pair of such numbers indicating its endpoints. Each edge will retain the pointer to its weight, but that is separate from its numerical representation. We then change the format for each graph as follows: Instead of a list of numbers, each in the range [1::r], we will represent the graph as a list of pointers. For this we initialize a linked list with r buckets, labeled 1 through r. If, in the numerical representation the number j appears, it will be replaced by a pointer to the j th bucket. We transform a graph into this pointer representation by traversing rst the list of vertices and then the list of edges in order, and traversing the list of buckets simultaneously, replacing each vertex entry, and the rst vertex entry for each edge by a pointer to the corresponding bucket. Thus edge (x; y), also appearing as (y; x), will now appear as (ptr(x); y) and (ptr(y); x). We then employ the twin pointers to replace the remaining y and x with their equivalent pointers. Clearly this transformation can be performed in O(m) time, where m is the sum of the sizes of all of the We will now perform a lexicographic sort [AHU74] on the sequence of C i 's in order to group together isomorphic components. With our representation we can replace each bucket indexing performed by traditional lexicographic sort by an access to the bucket pointer that we have placed for each element. Hence the running time for the pointer-based lexicographic sort is O( is the length of the i th vector and DecisionTree is called with graphs of size r = O(log (3) n), we have and the sum of the sizes of the graphs is O(m). Hence the radix sort can be performed in O(m Finally, we march through the sorted list of the C i 's and the sorted list of generic graphs, matching them up as appropriate. We will only need to traverse an initial sequence of the sorted generic graphs containing O(2 r 2 entries in order to match up the graphs. This takes time O(m 7 Performance on Random Graphs Even if we assume that MST has some super-linear complexity, we show below that our algorithm runs in linear time for nearly all graphs, regardless of edge weights. This improves upon the expected linear-time result of Karp and Tarjan [KT80], which depended on the edge weights being chosen randomly. Our result may also be contrasted with the randomized algorithm of Karger et al. [KKT95], which is shown to run in O(m) time w.h.p. by a proof that depends on the permutation of edge weights and random bits chosen, not the graph topology. In fact, none of the earlier published MST algorithms appear to have this property of running in linear time w.h.p. on random graphs for all edge-weights. Using the analysis of this section and suitably souped-up versions of earlier algorithms [FT87, GGST86, Chaz00], we may obtain the same high probability result. Our analysis hinges on the observation that for sparse random graphs, w.h.p. any subgraph constructed by the Partition routine has only a miniscule number in edges in excess of the number of spanning forst edges in that subgraph. The MST of such graphs can be computed in linear time, and hence the computation on optimal decision trees takes linear time on these graphs. Throughout this section will denote (m; n). Theorem 7.1 The MST of a graph can be found in linear time with probability e graph drawn from G n;m graph drawn from G n;p . Both (1) and (2) hold regardless of the permutation of edge weights. In the next section we describe the edge-addition martingale for the G n;m model. In section 7.2 we use this martingale and Azuma's inequality to prove part (1) of Theorem 7.1. Part (2) is shown to follow from part (1). 7.1 The Edge-Addition Martingale Consider the G n;m random graph model in which each graph with n labeled vertices and m edges is equally likely. For analytical purposes, we select a random graph by beginning with n vertices and adding one edge at a time [ER61]. Let X i be a random edge s.t. X i be the graph made up of the rst i edges, with G 0 being the graph on n vertices having no edges. A martingale is a sequence of random variables Y We now prove that if g is any graph-theoretic function and for is a martingale. Lemma 7.2 The sequence m, is a martingale, where g is any graph theoretic function, G 0 is the edge-free graph on n vertices, and G i is derived from G i 1 by adding a random edge not in G i 1 to G i 1 . g. Given that G i 1 has been xed, call the sequence proved to be a martingale in Lemma 7.2 the edge-addition martingale in contrast to the edge-exposure martingale for G n;p . We now recall the well-known Azuma's inequality (see, e.g., [AS92]). Theorem 7.3 (Azuma'a Inequality.) Let Y Ym be a martingale with jY m. Let > 0 be arbitrary. Then Pr[jY To facilitate the application of Azuma's inequality to our edge-addition martingale we establish the following lemma. Lemma 7.4 Consider the sequence proved to be a martingale in Lemma 7.2. Let g be any graph-theoretic function such that jg(G) g(G 0 )j 1 for any pair of graphs G and G 0 of the form are the average of range over their possible outcomes, given G i and G i 1 respectively. We identify each outcome of with equal-size disjoint sets of outcomes of X m which cover all outcomes of X m . Then may be regarded as an average of set averages. If, for each set corresponding to an outcome P of X m we establish that the set average diers from g(G i [ P ) by no more than 1, the Lemma follows. The correspondence is as follows. Let a. For each outcome x m , the corresponding set consists of outcomes x j ranges over all edges not appearing in G i 1 and x m . For each outcome i+1 and all Q in P 's associated set, since the graphs dier in at most one edge. Clearly holds as well, where the average is over outcomes Q in P 's associated set. 2 7.2 Analysis We dene the excess of a subgraph H to be jE(H)j jF (H)j, where F (H) is any spanning forest of H. Let f(G) be the maximum excess of the graph made up of intra-component edges, where the sets of components range over all possible sets returned by the Partition procedure. (Recall that the size of any component is no more than The key observation leading to our linear-time result is that each pass of our optimal algorithm denitely runs in linear time if f(G) m=(m;n). To see this, note that if this bound on f(G) holds, we can reduce the total number of intra-component edges to 2m= in linear time using log Boruvka steps, and then, clearly, the MST of the resulting graph can be determined in O(m) time. We show below that if a graph is randomly chosen from G n;m , f(G) m=(m;n) with high probability. We now show that Lemma 7.4 applies to the graph-theoretic function f , and then apply Azuma's inequality to obtain our desired result. Lemma 7.5 Let be two graphs on a set of labeled vertices which dier by no more than one edge. Then jf(G) f(G 0 )j 1. Proof: Suppose w.l.o.g. that f(G) f(G 0 ) > 1, then we could apply the optimal set of components of G to G 0 . Every intra-component edge of G remains an intra-component edge, except possibly e. This can reduce the excess by no more than one, a contradiction. The possibility that e 0 may become an intra-component edge can only help the argument. 2 Lemma Proof: Notice that if m simply impossible to have m= intra-component edges, so we assume m An upper bound on f E (G 0 ) is the expected number of indices i s.t. edge X i completed a cycle of length k in G i 1 , since all edges which caused f to increase must have satised this criterion. be the probability that X i completed a cycle of length k. By bounding the number of such cycles, and the probability they exist in the graph, we have Y <n nm (recall that i m) O n) In either case, f G be chosen from G n;m . Then Pr[f(G) > m=] < e Proof: By applying Azuma's inequality, we have that Pr[jf E (Gm Setting m gives the Lemma. Note that by Lemma 7.6 f insignicant. 2 We are now ready to prove Theorem 7.1. Proof: We examine only the rst log k passes of our optimal algorithm, since all remaining passes certainly take o(m) time. Lemma 7.7 assures us that the rst pass runs in linear time w.h.p. However, the topology of the graph examined in later passes does depend on the edge weights. Assuming the Boruvka steps contract all parts of the graph at a constant rate, which can easily be enforced, a partition of the graph in one pass of the algorithm corresponds to a partition of the original graph into components of size less than k c , for some xed c. Using k c in place of k does not aect Lemma 7.6, which gives the Theorem for G n;m , that is, part (1). For G n;p note that the probability that there are not (pn 2 ) edges is exponential in pn 2 ), hence the probability that the algorithm fails to run in linear time is dominated by the bound in part (1).For the sparse case where m < n=, Theorem 7.1 part (1) holds with probability 1, and for < 1=n, by a Cherno bound, part (2) holds with probability 1 e n=) . An intriguing aspect of our algorithm is that we do not know its precise deterministic running time although we can prove that it is within a constant factor of optimal. Results of this nature have been obtained in the past for sensitivity analysis of minimum spanning trees [DRT92] and convex matrix searching [Lar90]. Also, for the problem of triangulating a convex polygon, it was observed in [DRT92] that an alternate linear-time algorithm could be obtained using optimal decision trees on small subproblems. However, these earlier algorithms make use of decision trees in more straightforward ways than the algorithm presented here. As noted in Section 4.1, the construction of optimal decision trees takes sub-linear time. Thus, it is important to observe that our use of decision trees does not result in a large constant factor in the running time. Further, this construction of optimal decision trees is performed by a straightforward brute-force search, hence the resulting algorithm is uniform. It was mentioned in the introduction that an optimal algorithm can be constructed for any prob- lem, given an optimal verication algorithm for that problem [Jo97]. This construction produces an algorithm which enumerates programs (for some machine model) and executes them incrementally. Whenever one of the programs halts the verier checks its output for correctness. Using a linear-time MST verication algorithm such as [DRT92, K97, B+98], this construction yields an optimal MST algorithm, however it is unsatisfactory for several reasons. Aside from truly astronomical constant factors (roughly exponential in the size of the optimal program), the algorithm is optimal only with respect to a particular machine model (say a TM, a RAM, or a pointer machine). Our result, in contrast, is robust in that it ties the algorithmic complexity of MST to its decision-tree complexity, a limiting factor in any machine model. It is not always the case that algorithmic complexity and decision-tree complexity are asymptotically equivalent. In fact, one can easily concoct simple problems which are NP-hard but nevertheless have polynomial-depth decision-trees (e.g. nd the lightest edge on any Hamiltonian path). See [GKS93], [PR01, Section 8] for two sorting-type problems whose decision-tree complexity and algorithmic complexity provably diverge. 9 Conclusion We have presented a deterministic MSF algorithm that is provably optimal. The algorithm runs on a pointer machine, and on graphs with n vertices and m edges, its running time is O(T (m; n)), where T (m; n) is the decision-tree complexity of the MSF problem on n-node, m-edge graphs. Also, on random graphs our algorithm runs in linear time with high probability for all possible edge-weights. Although the exact running time of our algorithm is not known, we have shown that the time bound depends only on the number of edge-weight comparisons needed to determine the MSF, and not on any data structural issues. Determining the worst-case complexity of our algorithm is the main open question remaining in the MSF problem, however, there is a subtler open question. We have given an optimal uniform algorithm for the MSF problem. Is there an optimal uniform algorithm which does not use precomputed decision trees (or some similar technique)? More generally, are there problems where precomputation is necessary? One may wish to study this issue in a simpler setting, say the MSF verication problem on a pointer machine. Here there is still an (m; n) factor separating the best pointer machine algorithm which uses precomputed decision trees [B+98] and the one which does not [Tar79b]. One may also ask for the parallel complexity of the MSF problem. Here, resolved recently were the randomized work-time complexity [PR99] and the deterministic time complexity [CHL99] of the MSF problem on the EREW PRAM. An open question that remains here is to obtain a deterministic work-time optimal parallel MSF algorithm. Parallelizing our optimal algorithm is not at all straightforward. Although handling decision trees does not present any problems in the parallel context, we still need a method for identifying contractible components in parallel and a base case algorithm that performs linear work for graph-densities of log (3) n. Existing sequential algorithms which are suitable for the base case, such as the one in [FT87], are also not easily parallelizable. --R The Design and Analysis of Computer Algorithms. The Probabilistic Method. A faster deterministic algorithm for minimum spanning trees. A minimum spanning tree algorithm with inverse-Ackermann type complexity On the parallel time complexity of undirected connectivity and minimum spanning trees. A note on two problems in connexion with graphs. Fibonacci heaps and their uses in improved network optimization algorithms. On the history of the minimum spanning tree problem. Optimal randomized algorithms for local sorting and set- maxima Computability and Complexity: From a Programming Perspective. A randomized linear-time algorithm to nd minimum spanning trees Linear expected-time algorithms for connectivity problems A simpler minimum spanning tree veri An optimal algorithm with unknown time complexity for convex matrix searching. A randomized time-work optimal parallel algorithm for nding a minimum spanning forest <Proceedings>Proc An optimal minimum spanning tree algorithm. Computing undirected shortest paths with comparisons and additions. Finding minimum spanning trees in O(m Shortest connection networks and some generalizations. A class of algorithms which require nonlinear time to maintain disjoint sets. Applications of path compression on balanced trees. --TR Efficient algorithms for finding minimum spanning trees in undirected and directed graphs Fibonacci heaps and their uses in improved network optimization algorithms An optimal algorithm with unknown time complexity for convex matrix searching Verification and sensitivity analysis of minimum spanning trees in linear time Optimal randomized algorithms for local sorting and set-maxima Trans-dichotomous algorithms for minimum spanning trees and shortest paths A randomized linear-time algorithm to find minimum spanning trees Computability and complexity Linear-time pointer-machine algorithms for least common ancestors, MST verification, and dominators Applications of Path Compression on Balanced Trees The soft heap A minimum spanning tree algorithm with inverse-Ackermann type complexity Concurrent threads and optimal parallel minimum spanning trees algorithm Computing shortest paths with comparisons and additions Minimizing randomness in minimum spanning tree, parallel connectivity, and set maxima algorithms The Design and Analysis of Computer Algorithms A Randomized Time-Work Optimal Parallel Algorithm for Finding a Minimum Spanning Forest A Faster Deterministic Algorithm for Minimum Spanning Trees Finding Minimum Spanning Trees in O(m alpha(m,n)) Time --CTR Jess Cerquides , Ramon Lpez Mntaras, TAN Classifiers Based on Decomposable Distributions, Machine Learning, v.59 n.3, p.323-354, June 2005 Artur Czumaj , Christian Sohler, Estimating the weight of metric minimum spanning trees in sublinear-time, Proceedings of the thirty-sixth annual ACM symposium on Theory of computing, June 13-16, 2004, Chicago, IL, USA Tzu-Chiang Chiang , Chien-Hung Liu , Yueh-Min Huang, A near-optimal multicast scheme for mobile ad hoc networks using a hybrid genetic algorithm, Expert Systems with Applications: An International Journal, v.33 n.3, p.734-742, October, 2007 Seth Pettie, A new approach to all-pairs shortest paths on real-weighted graphs, Theoretical Computer Science, v.312 n.1, p.47-74, 26 January 2004 Ran Mendelson , Robert E. Tarjan , Mikkel Thorup , Uri Zwick, Melding priority queues, ACM Transactions on Algorithms (TALG), v.2 n.4, p.535-556, October 2006 Amos Korman , Shay Kutten, Distributed verification of minimum spanning trees, Proceedings of the twenty-fifth annual ACM symposium on Principles of distributed computing, July 23-26, 2006, Denver, Colorado, USA

optimal complexity;graph algorithms;minimum spanning tree

505422

Closing the smoothness and uniformity gap in area fill synthesis.

Control of variability in the back end of the line, and hence in interconnect performance as well, has become extremely difficult with the introduction of new materials such as copper and low-k dielectrics. Uniformity of chemical-mechanical planarization (CMP) requires the addition of area fill geometries into the layout, in order to smoothen the variation of feature densities across the die. Our work addresses the following smoothness gap in the recent literature on area fill synthesis. (1)The very first paper on the filling problem (Kahng et al., ISPD98 [7]) noted that there is potentially a large difference between the optimum window densities in fixed dissections vs. when all possible windows in the layout are considered. (2)Despite this observation, all filling methods since 1998 minimize and evaluate density variation only with respect to a fixed dissection. This paper gives the first evaluation of existing filling algorithms with respect to "gridless" ("floating-window") mode, according to both the effective and spatial density models. Our experiments indicate surprising advantages of Monte-Carlo and greedy strategies over "optimal" linear programming (LP) based methods. Second, we suggest new, more relevant methods of measuring a local uniformity based on Lipschitz conditions, and empirically demonstrate that Monte-Carlo methods are inherently better than LP with respect to the new criteria. Finally, we propose new LP-based filling methods that are directly driven by the new criteria, and show that these methods indeed help close the "smoothness gap".

INTRODUCTION Chemical-mechanical planarization (CMP) and other manufacturing steps in nanometer-scale VLSI processes have varying effects on device and interconnect features, depending on the local characteristics of the layout. To improve manufacturability and performance predictability, foundry rules require that a layout be made uniform with respect to prescribed density criteria, through insertion of area fill (dummy fill) geometries. All existing methods for synthesis of area fill are based on discretization: the layout is partitioned into tiles, and filling constraints or objectives (e.g., minimizing the maximum density variation) are enforced for square windows that each consists of r \Thetar tiles. Thus, to practically control layout density in arbitrary windows, density bounds are enforced in only a finite set of windows. More precisely, both foundry rules and EDA physical verification and layout tools attempt to enforce density bounds within r 2 overlapping fixed dissec- tions, where r determines the "phase shift" w=r by which the dissections are offset from each other. The resulting fixed r-dissection (see Figure 1) partitions the n \Theta n layout into tiles T ij , then covers the layout by w \Theta w-windows W ij , 1, such that each window W ij consists of Two main filling objectives are considered in the recent literature: ffl (Min-Var Objective) the variation in window density (i.e., maximum window density minus minimum window density) is minimized while the window density does not exceed the given upper bound U ; ffl (Min-Fill Objective) the number of inserted fill geometries is minimized while the density of any window remains in the given range (L; U ). Recent methods on area fill synthesis also focused exclusively on the fixed-dissection context, including: ffl Linear Programming (LP) methods based on rounding relaxation of the corresponding integer linear program formulations. The LP formulations for filling were first proposed by Kahng et al. in [6] and adapted to other objectives and CMP models in [12, 13]); ffl Greedy methods which iteratively find the best tile for the next filling geometry to be added into the layout. These methods were first used in [3] for ILD thickness control, and also used for shallow-trench isolation (STI) CMP model in [13]); ffl Monte-Carlo (MC) methods, which are similar to greedy methods but insert the next filling geometry randomly. Due to its efficiency and accuracy, these were used for both flat [3, 4] and hierarchical [2] layout density con- trol; and ffl Iterated Greedy (IGreedy) and Iterated Monte-Carlo methods, which improve the solution quality by iterating the insertions and deletions of dummy fill features with respect to the density variation ([3]). The motivation for our present work is a "smoothness gap" in the fill literature. All existing filling methods fail to consider the potentially large difference between extremal densities in fixed-dissection windows and extremal densities when all possible windows are considered. On the other hand, the very first paper in the fill synthesis literature (Kahng et al., ISPD98 [7]) already pointed out the gap between fixed-dissection and "gridless" analyses, for both tile density and window density. 1 The potential consequence of the smoothness gap is that the fill result will not satisfy either given upper bounds on post-fill window density, or given bounds on density variation between windows. As post-CMP variation for oxide ILD polishing is essentially monotone in window density variation [11], this smoothness gap can compromise manufacturability of the layout, particularly given the small values of r in recent design rules. We first address the discretization gap in existing analyses (i.e., evaluations) methods. Previous works compare density control methods only with respect to a given fixed grid, which underestimates the actual "gridless" density varia- tion, but has been justified on grounds that gridless analysis is impractical. In this paper, we show for the first time the viability of gridless or floating window analyses, originally developed for the spatial density model [6], and extend it for the more accurate effective density model [9]. Second, previous research in layout density control concentrated on the global uniformity achieved by minimizing the window density variation over the entire layout. However, the density variation between locations which are far from each other is actually not so critical, in that the pressure/speed of the polishing pad can be (self-)adjusted during CMP. Thus, we propose and analyze criteria for "local uniformity" as a measure of smoothness in filling solutions. We evaluate existing methods with respect to the new criteria, and we suggest LP-based methods that directly optimize filling solutions with respect to smoothness. The rest of the paper is organized as follows. In Section 2 we show how to apply floating window density analysis 1 Bounding the spatial density in a fixed set of w \Thetaw windows can incur substantial error, since other windows may still violate the density bounds [6]. methods (such as extremal-density window and multilevel density analyses) to spatial and effective density models. We then give the first "gridless" evaluation of existing filling algorithms, under the effective as well as spatial density models. Our experiments indicate surprising advantages of Monte-Carlo and greedy methods over "optimal" linear programming (LP) based methods. In Section 3 we introduce new Lipschitz-like measures for layout smoothness and describe new LP-based filling methods driven by such mea- sures. We also compare the results of existing and new filling approaches, with respect to the new smoothness criteria. Section 4 concludes with directions for future work. 2. LAYOUT DENSITY ANALYSES As noted above, for the sake of tractability, previous works have used fixed dissections to decide the amount and positions of dummy fill features[6]. A smoothness gap thus exists because a filling solution based on a fixed-dissection does not address the true post-filling density variation. Here we first summarize the two main density models used in the current literature. We then introduce two extremal-density analysis algorithms (for spatial and effective density, respectively) which we use to compute post-fill layout density. Finally, we evaluate existing filling methods according to (near-)gridless density variation. 2.1 Density Models for Oxide CMP We focus on layout density control for (oxide) interlevel dielectric CMP. 2 Several models have been proposed in [8], including the model of [10], where the interlevel dielectric thickness z at location (x; y) is calculated as: (1) The crucial element of this model is the determination of the effective initial pattern density, ae(x; y). The simplest model for ae(x; y) is the local areal feature density, i.e., the window density is simply equal to the sum: area(Tkl ) (2) where area(Tkl ) denotes the original layout area of the tile Tkl . This spatial density model is due to [6], which solved the resulting filling problem using linear programming. A more accurate model considers the deformation of the polishing pad during the CMP process [5]: effective local density ae(x; y) is calculated as the sum of weighted spatial pattern densities within the window, relative to an elliptical weighting function: with experimentally determined constants c0 , c1 , and c2 [12]. The discretized effective local pattern density ae for a window ij in the fixed-dissection regime (henceforth referred to as effective density) is: recent works, particularly by Wong et al., have studied alternative arenas for dummy fill, including shallow- trench isolation and dual-damascene copper. For such are- nas, density calculations and physical polish mechanisms are different from those in the oxide context. Consideration of these alternate models is orthogonal to our contribution; certainly, the concept of a "smoothness gap" applies to all filling contexts. arbitrary window W window tile shrunk fixed dissection window fixed dissection bloated fixed dissection window Figure 2: An arbitrary floating w \Theta w-window W always contains a shrunk (r \Gamma 1) \Theta (r \Gamma 1)-window of a fixed r-dissection, and is always covered by a bloated (r+1)\Theta(r+1)-window of the fixed r-dissection. A standard r \Theta r fixed-dissection window is shown with thick border. A floating window is shown in light gray. The white window is the bloated fixed- dissection window, and the dark gray window is the shrunk fixed-dissection window. area(Tkl where the arguments of the elliptical weighing function f are the x- and y-distances of the tile Tkl from the center of the window W ij . 2.2 Window Density Analyses The authors of [6] proposed optimal extremal-density (i.e., minimum or maximum window density in the layout) analysis algorithms. Their ALG1, with complexity O(k is the number of rectangles in layout), is proposed as a means of checking the gridless post-filling density variation. However, with a large number of original and dummy fill features, this algorithm may be infeasible in practice. Another method of [6] overcomes the intractability of optimal extremal-density analysis, based on the following fact (see Fig. 2). Lemma 1. Given a fixed r-dissection, any arbitrary w \Theta w window will contain some shrunk w(1 \Gamma 1=r) \Theta w(1 \Gamma 1=r) window of the fixed r-dissection, and will be contained in some bloated w(1 window of the fixed r-dissection. The authors of [6] implemented the above Lemma within a multi-level density analysis algorithm (see Fig. 3). Here used to denote the required user-defined accuracy in finding the maximum window density. The lists TILES and WINDOWS are byproducts of the analysis. Since any floating w \Theta w-window W is contained in some bloated window, the filled area in W ranges between Max (maximum w \Theta w- window filled area found so far) and BloatMax (maximum bloated window filled area found so far). The algorithm terminates when the relative gap between Max and BloatMax is at most 2 \Delta ffl, and then outputs the middle of the range (Max,BloatMax). We use this algorithm (with accuracy = 1.5%) throughout this paper to achieve an accurate, efficient post-filling density analysis. 3 To handle the effective density model, the 3 For the test cases used in this paper, the runtimes of the Multi-Level Density Analysis Algorithm Input: n \Theta n layout and accuracy ffl ? 0 Output: maximum density of w \Theta w window with accuracy ffl (1) Make a list ActiveTiles of all w=r \Theta w=r-tiles (3) While Accuracy do (a) Find all rectangles in w=r \Theta w=r-tiles from ActiveTiles (b) Find area of each window consisting of tiles from ActiveTiles, add such window to the list WINDOWS (c) maximum area of standard window with tiles from ActiveTiles maximum area of bloated window with tiles from ActiveTiles (e) For each tile T from ActiveTiles which do not belong to any bloated window of area more than Max do remove T from ActiveTiles (f) Replace in ActiveTiles each tile with four of its subtiles (4) Move all tiles from ActiveTiles to TILES Figure 3: Multi-level density analysis algorithm. multi-level density analysis based on bloated and shrunk windows must be refined somewhat. To obtain more accurate results, the multi-level density analysis algorithm divides the r-dissection into smaller grids, so that more windows will be considered. With the effective density model, the discretized formulation (effective) shows that effective local pattern density is dependent on the window size w and the r-dissection. That is, we have to consider the effect on the formulation of the further division of layout during post-filling density analysis. We assume here that the effective local pattern density is still calculated with the value of r-dissection used in the filling process. The only difference is that the windows phase-shift will be smaller. For example, in Figure 4(a) we calculate the effective density of the window shown in light gray by considering 5 \Theta 5 tiles (also called "cells") during the filling process. In Figure 4(b) the layout is further partitioned by a factor of 4. The effective density of the light gray window will be still calculated with the 5 \Theta 5 "cells". Here each "cell" has the same dimension as a tile in the filling process and consists of 2 \Theta 2 smaller tiles. More windows (e.g., the window with thick border) with smaller phase-shifts will be considered in the more gridded layout. 2.3 Accurate Analysis of Existing Methods Here we compare the performance of existing fill synthesis methods, using the accurate multilevel floating-window density analysis. All experiments are performed using part of a metal layer extracted from an industry standard-cell layout 4 (Table 2.3). Benchmark L1 is the M2 layer from an 8,131-cell design and benchmark L2 is the M3 layer from a 20,577-cell layout. multi-level analysis with accuracy = 1:5% appear reason- able. Our (unoptimized) implementation has the following runtimes for Min-Var LP solutions and the spatial density model: L1/32 (45 sec), L1/16 (183 sec), L2/28 (99 sec), L2/14 (390 sec), For the effective density model, the run-times are: L1/32 (49 sec), L1/16 (194 sec), L2/28 (109 sec), 4 Our experimental testbed integrates GDSII Stream input, conversion to CIF format, and internally-developed geometric processing engines, coded in C++ under Solaris. We use CPLEX version 7.0 as the linear programming solver. All runtimes are CPU seconds on a 300 MHz Sun Ultra-10 with 1GB of RAM. (a) (b) window tile cell Figure 4: Post-filling density analysis for the effective density model. (a): a fixed-dissection, where each window consists of 5 \Theta 5 cells (the same size as tiles); (b): a fixed-dissection for post-filling density analysis, where each window consists of smaller tiles and each cell consists of 2 \Theta 2 tiles. Table 2 shows that underestimation of the window density variation as well as violation of the maximum window density in fixed-dissection filling can be severe: e.g., for the LP method applied to the case L2/28/4 for the spatial (resp. effective) density model, the density variation is underestimated by 210% (resp. 264%) and the maximum density is violated by 21% (resp. 15%). Even for the finest grid (L2/28/16), the LP method may still yield considerable er- ror: 11% (resp. 23%) in density variation and 1.2% (resp. 3.2%) in maximum density violation. Note that the LP method cannot easily handle a finer grid since the runtime is proportional to r 6 . Our comparisons show that the winning method is IMC and the runner-up is IGreedy. IMC and IGreedy can be run for much finer grids since its runtime is proportional to r 2 log r). Although for L2/28/16 errors in density variation and the maximum density violation are similar, the iterative methods become considerably more accurate. test case L1 L2 layout size n 125,000 112,000 rectangles k 49,506 76,423 Table 1: Parameters of four industry test cases. Here 40 units are equivalent to 1 micron. 3. LOCAL DENSITY VARIATION The main objective of layout filling is to improve CMP and increase yield. Traditionally, layout uniformity has been measured by global spatial or effective density variation over all windows. Such a measure does not take in account that the polishing pad during CMP can change (adjust) the pressure and rotation speed according to the pattern distribution (see [11]). Boning et al. [1] further point out that while the effective density model is excellent for local CMP effect pre- diction, it fails to take into account global step heights. The influence of density variation between far-apart regions can be reduced by a mechanism of pressure adjustment, which leads to the contact wear model proposed in [1]. Within each local region, the area fill can be used to improve the CMP performance. Therefore, density variation between two windows in opposite corners of the layout will not cause problems because of the polishing dynamics. According to the integrated contact wear and effective density model, only a significant density variation between neighboring windows will complicate polishing pad control and may cause either dishing or underpolishing. Thus, it is more important to measure density variation between neighboring windows. 3.1 Lipschitz Measures of Smoothness Depending on the CMP process and the polishing pad movement relative to the wafer, we may consider different window "neighborhoods". Below we propose three relevant Lipschitz-like definitions of local density variation which differ only in what windows are considered to be neighbors. ffl Type I: The maximum density variation of every r neighboring windows in each row of the fixed-dissection. The intuition here is that the polishing pad is moving along window rows and therefore only overlapping windows in the same row define a neighborhood. ffl Type II: The maximum density variation of every cluster of windows which cover one tile. The idea here is that the polishing pad can touch all overlapping windows almost simultaneously. ffl Type III: The maximum density variation of every cluster of windows which cover one square consisting of r=2 \Theta r=2 tiles. The difference between this and the previous definition is the assumption that the polishing pad is moving slowly; if windows overlap but are still too far from each other, then we can disregard their mutual influence. We compared the behaviors of existing filling methods with respect to these Lipschitz characteristics. The results in Table 3 show that there is a gap between the traditional Min-Var objective and the new "smoothness" objectives: the solution with the best Min-Var objective value does not always have the best value in terms of "smoothness" objec- tives. For the spatial (resp. effective) density model, though LP yields the best result for the case L2/28/4 with Min-Var objective with fixed-dissection model, it can not obtain the best result with respect to Lipschitz type-I variation. Thus, "smoothness" objectives may be considered separately for the filling process. We also notice that Monte-Carlo methods can achieve better solutions than LP with respect to the "smoothness" objectives (note that although LP is "op- timal", it suffers from rounding and discreteness issues when converting the LP solution to an actual filling solution). 3.2 Smoothness Objectives for Filling Obviously, all Lipschitz conditions are linear and can be implemented as linear programming formulations. We describe four linear programming formulations for the "smooth- ness" objectives with respect to the spatial density model. (The linear programming formulations for the effective density model are similar.) The first Linear Programming formulation for the Min- Lip-I objective is: Minimize: L Subject to: st nr area(T st st LP Greedy MC IGreedy IMC Test case OrgDen FD Multi-Level FD Multi-Level FD Multi-Level FD Multi-Level FD Multi-Level T/W/r MaxD MinD DenV MaxD DenV DenV MaxD DenV DenV MaxD DenV DenV MaxD DenV DenV MaxD DenV Spatial Density Model Effective Density Model Table 2: Multi-level density analysis on results from existing fixed-dissection filling methods. Notation: T/W/r: Layout / window size / r-dissection; LP: linear programming method; Greedy: Greedy method; MC: Monte-Carlo method; IGreedy: iterated Greedy method; IMC: iterated Monte-Carlo method; OrgDen: density of original layout; FD: fixed-dissection density analysis; Multi-Level: multi-level density analysis; MaxD: maximum window density; MinD: minimum window density; DenV: density variation. Test case LP Greedy MC IGreedy IMC T/W/r LipI LipII LipIII LipI LipII LipIII LipI LipII LipIII LipI LipII LipIII LipI LipII LipIII Spatial Density Model Effective Density Model L1/16/4 4.048 4.333 3.864 5.332 5.619 5.190 3.631 4.166 3.448 3.994 4.254 3.132 4.245 4.481 3.315 L2/28/4 2.882 5.782 4.855 2.694 6.587 6.565 1.498 5.579 5.092 2.702 6.317 5.678 2.532 5.640 4.981 Table 3: Different behaviors of existing filling methods on "smoothness" objectives. Note: All data for effective density model have been timed by 10 3 . Notation: LipI: Lipschitz condition I; LipII: Lipschitz condition II; LipIII: Lipschitz condition III. Here, U is the given upper bound of the effective tile densi- ties. The constraints (5) imply that features can be added but not deleted from any tile. The slack constraints (6) are computed for each tile. The pattern-dependent coefficient pattern denotes the maximum pattern area which can embedded in an empty unit square. If a tile T ij is originally overfilled, then we set slack(T ij In the LP solution, the values of p ij indicate the fill amount to be inserted in each tile T ij . The constraint (7) says that no window can have density greater than than U (unless it was initially over- filled). The constraints (8) imply that the auxiliary variable L is an upper bound on all variation between (2r in the same row. The second Linear Programming formulation for the Min- Lip-II objective replaces the constraints (8) with the following constraints Here, the auxiliary variables minDen(i; j) and maxDen(i; are the minimum and maximum tile effective densities in centered at T i;j . The constraints above ensure that the density variations among all windows which cover T i;j is less than the auxiliary variable L. The Min-Lip-III objective strives to minimize the maximum density variation of every cluster of windows which cover one square consisting of k tiles. The constraints are changed to the following: r The constraints (10) ensure that the density variation between any two windows which cover r \Theta r tiles is less than the auxiliary variable L. Finally, in order to consider the "smoothness" objectives together with the Min-Var objective, we propose another LP formulation with the combined objective which is the linear summation of Min-Var, Lip-I, and Lip-II objectives with specific coefficients. Minimize: C0 M Lip-I Constraints (8), Lip-II (9) and Min-Var constraints are added for the combined objective: Here, the auxiliary variables LI and LII are the maximum Lipschitz condition type-I and type-II, and the auxiliary variable M is a lower bound on all tile densities. 3.3 Computational Experience We tested the smoothness of filling solutions generated using the same test cases, with smoothness evaluated using finest-r density analysis with 64. Runtimes of the new methods are substantially longer than for the original Min-Var LP formulation, because many more constraints are added for each layout window due to the Lipschitz condition objectives. For example, for L2/28/8 and the spatial den- Test case Min-Var LP LipI LP LipII LP LipIII LP Comb LP Spatial Density Model Effective Density Model Table 4: Comparison among the LP methods on Min-Var and Lipschitz condition objectives. Notation: Min- Var LP: LP with Min-Var objective; LipI LP: LP with Min-Lip-I objective; LipII LP: LP with Min-Lip-II objective; LipIII LP: LP with Min-Lip-III objective; Com LP: LP with combined objective. sity model, the runtime of Min-Var LP is 6.9 seconds, while the Lip-I LP runtime is 3.41 seconds, the Lip-II LP runtime is 994 seconds and the Lip-III LP runtime is 71.4 seconds. For L2/28/8 and the effective density model, the runtime of Min-Var LP is 2.3 seconds, while the Lip-I LP runtime is 708 seconds, the Lip-II LP runtime is 5084 seconds and the Lip-III LP runtime is 2495 seconds. Since fill generation is a post-processing step (currently performed in PV tools), we do not believe that these runtimes are prohibitive. Our major win is that LP is tractable with Lipschitz objectives (as opposed to intractable with large values of r). Of course, finding smoothness objectives that result in smaller LPs is a direction for future work. The performances of the new LP formulations with "smooth- ness" objectives are studied in Table 4. We use the coefficients (0.4/0.4/0.2) in the combined objective; these values were derived from the greedy testing of all coefficient combi- nations. Because of LP's rounding error, 5 some new LPs do not achieve the best value on certain test cases. From the comparison between the new LPs and Min-Var LP, it appears that neither Min-Var LP nor the Lipschitz condition- derived LPs are dominant. At the same time, when compared against existing filling methods in Table (3), the new LP with combined objective normally achieves the best comprehensive solutions in terms of trading off among the Min- Den, Lipschitz conditions I and II. Another interesting observation is that the LP with combined objective can achieve even smaller density variations than the Min-Var LP. This shows that the solution qualities of LP methods can be significantly damaged by rounding effects, and that a better non-LP method may be possible. 4. CONCLUSIONS&FUTURERESEARCH To improve manufacturability and performance predictabil- ity, it is necessary to "smoothen" a layout by the insertion of "filling (dummy) geometries". In this paper, we pointed out the potentially large difference between fixed-dissection filling results and the actual maximum or minimum window density in optimal density analyses. We compared existing filling algorithms in gridless mode using the effective as well as the spatial density models. We also suggested new methods of measuring local uniformity of the layout based on Lipschitz conditions and proposed new filling methods based on these properties. Our experimental results highlight the advantages of Monte-Carlo and greedy -based methods over previous linear program based approaches. 5 The desired fill area specified for each tile in the LP solution must be rounded to an area that corresponds to an an integer number of dummy fill features. Ongoing work addresses extensions of multi-level density analyses to measuring local uniformity ("smoothness") with respect to other CMP physical models. We also seek improved methods for optimizing fill synthesis with respect to our new (and possibly other alternative) local uniformity objectives. 5. --R "Models for Pattern Dependencies: Capturing Effects in Oxide, STI, and Copper CMP" "Hierarchical Dummy Fill for Process Uniformity" "Practical Iterated Fill Synthesis for CMP Uniformity" "New Monte-Carlo Algorithms for Layout Density Control" "Effect of Fine-line Density and Pitch on Interconnect ILD Thickness Variation in Oxide CMP Process" "Filling Algorithms and Analyses for Layout Density Control" "Filling and Slotting: Analysis and Algorithms" "Modeling of Chemical-Mechanical Polishing: A Review" "An Integrated Characterization and Modeling Methodology for CMP Dielectric Planarization" "A Closed-Form Analytical Model for ILD Thickness Variation in CMP Processes" "Rapid Characterization and Modeling of Pattern Dependent Variation in Chemical Mechanical Polishing" "Model-Based Dummy Feature Placement for Oxide Chemical Mechanical Polishing Manufacturability" "Dummy feature placement for chemical-mechanical polishing uniformity in a shallow trench isolation process " --TR Filling and slotting Model-based dummy feature placement for oxide chemical-mechanical polishing manufacturability Practical iterated fill synthesis for CMP uniformity Monte-Carlo algorithms for layout density control Dummy feature placement for chemical-mechanical polishing uniformity in a shallow trench isolation process Hierarchical dummy fill for process uniformity --CTR Hua Xiang , Liang Deng , Ruchir Puri , Kai-Yuan Chao , Martin D.F. Wong, Dummy fill density analysis with coupling constraints, Proceedings of the 2007 international symposium on Physical design, March 18-21, 2007, Austin, Texas, USA Performance-impact limited area fill synthesis, Proceedings of the 40th conference on Design automation, June 02-06, 2003, Anaheim, CA, USA Hua Xiang , Kai-Yuan Chao , Ruchir Puri , Martin D.F. Wong, Is your layout density verification exact?: a fast exact algorithm for density calculation, Proceedings of the 2007 international symposium on Physical design, March 18-21, 2007, Austin, Texas, USA Andrew B. Kahng, Research directions for coevolution of rules and routers, Proceedings of the international symposium on Physical design, April 06-09, 2003, Monterey, CA, USA

monte-carlo;dummy fill problem;VLSI manufacturability;density analysis;chemical-mechanical polishing

505522

High-speed architectures for Reed-Solomon decoders.

New high-speed VLSI architectures for decoding Reed-Solomon codes with the Berlekamp-Massey algorithm are presented in this paper. The speed bottleneck in the Berlekamp-Massey algorighm is in the iterative computation of discrepencies followed by the updating of the error-locator polynomial. This bottleneck is eliminated via a series of algorithmic transformations that result in a fully systolic architecture in which a single array of processors computes both the error-locator and the error-evaluator polynomials. In contrast to conventional Berlekamp-Massey architectures in which the critical path passes through two multipliers and 1+[log2(t +1)] adders, the critical path in the proposed architecture passes through only one multiplier and one adder, which is comparable to the critical path in architectures based on the extended Euclidean algorithm. More interestingly, the proposed architecture requires approximately 25% fewer multipliers and a simpler control structure than the architectures based on the popular extended Euclidean algorithm. For block-interleaved Reed-Solomon codes, embedding the interleaver memory into the decoder results in a further reduction of the critical path delay to just one XOR gate and one multiplexer, leading to speed ups of as much as an order of magnitude over conventional architectures.

Introduction Reed-Solomon codes [1], [3] are employed in numerous communications systems such as those for deep space, digital subscriber loops, and wireless systems as well as in memory and data storage systems. Continual demand for ever higher data rates makes it necessary to devise very high-speed implementations of decoders for Reed-Solomon codes. Recently reported decoder implementations [5], [19] have quoted data rates of ranging from 144 Mb/s to 1.28 Gb/s. These high throughputs have been achieved by architectural innovations such as pipelining and parallel processing. A majority of the implementations [2], [8], [15], [19] employ an architecture based on the extended Euclidean (eE) algorithm for computing the greatest common divisor of two polynomials [3]. A key advantage of architectures based upon the eE algorithm is regularity. In addition, the critical path delay in these architectures is at best Tmult Tmult , T add , and Tmux are the delays of the nite-eld multiplier, adder, and 2 1 multiplexer respectively, and this is su-ciently small for most applications. In contrast, relatively few decoder implementations have employed architectures based on the Berlekamp-Massey (BM) algorithm DRAFT June 19, 2000 [1], [3], [10], presumably because the architectures were found to be irregular and to have a longer critical path delay that was also dependent on the error-correcting capability of the code [5]. In this paper, we show that, in fact, it is possible to reformulate the BM algorithm to achieve extremely regular decoder architectures. Surprisingly, these new architectures can not only operate at data rates comparable to architectures based on the eE algorithm, but they also have lower gate complexity and simpler control structures. This paper begins with a brief tutorial overview of the encoding and decoding of Reed-Solomon codes in Section II. Conventional architectures for decoders based on the BM algorithm are described in Section III. In Section IV, we show that it is possible to algorithmically transform the BM algorithm so that a homogenous systolic array architecture for the decoder can be developed. Finally, in Section V, we describe a pipelined architecture for block-interleaved Reed-Solomon codes that achieves an order of magnitude reduction in the critical path delay over the architectures presented in Sections III and IV. II. Reed-Solomon Codes We provide a brief overview of the encoding and decoding of Reed-Solomon codes. A. Encoding of Reed-Solomon Codes data symbols (bytes) that are to be transmitted over a communication channel (or stored in memory.) These bytes are regarded as elements of the nite eld (also called Galois eld) GF(2 m ), 1 and encoded into a codeword (c of n > k bytes. These codeword symbols are transmitted over the communication channel (or stored in memory.) For Reed-Solomon codes over GF(2 m is odd, and the code can correct (n k)=2 byte errors. The encoding process is best described in terms of the data polynomial being transformed into a codeword polynomial polynomials C(z) are polynomial Addition (and subtraction) in GF(2 m ) is the bit-by-bit XOR of the bytes. The 2 m 1 nonzero elements of can also be regarded as the powers, of a primitive element (where so that the product of eld elements June 19, 2000 DRAFT multiples of G(z), the generator polynomial of the code, which is dened as Y typically 0 or 1. However, other choices sometimes simplify the decoding process slightly. Since 2t consecutive powers of are roots of G(z), and C(z) is a multiple of G(z), it follows that for all codeword polynomials C(z). In fact, an arbitrary polynomial of degree less than n is a codeword polynomial if and only if it satises (2). A systematic encoding produces codewords that are comprised of data symbols followed by parity-check symbols, and is obtained as follows. Let Q(z) and P (z) denote the quotient and remainder respectively when the polynomial z n k D(z) of degree n 1 is divided by G(z) of degree is a multiple of G(z). Furthermore, since the lowest degree term in z n k D(z) is d 0 z n k while P (z) is of degree at most n k 1, it follows that the codeword is given by and consists of the data symbols followed by the parity-check symbols. B. Decoding of Reed-Solomon Codes Let C(z) denote the transmitted codeword polynomial and let R(z) denote the received word polynomial. The input to the decoder is R(z), and it assumes that where, if e > 0 errors have occurred during transmission, the error polynomial E(z) can be written as It is conventional to say that the error values have occurred at the error locations . Note that the decoder does not know E(z); in fact, it does DRAFT June 19, 2000 not even know the value of e. The decoder's task is to determine E(z) from its input R(z), and thus correct the errors by subtracting o E(z) from R(z). If e t, then such a calculation is always possible, that is, t or fewer errors can always be corrected. The decoder begins its task of error correction by computing the syndrome values If all 2t syndrome values are zero, then R(z) is a codeword, and it is assumed that that is, no errors have occurred. Otherwise, the decoder knows that e > 0 and uses the syndrome polynomial S(z), which is dened to be to calculate the error values and error locations. Dene the error locator polynomial (z) of degree e and the error evaluator polynomial z) of degree at most e 1 to be e Y e e Y These polynomials are related to S(z) through the key equation [1], [3]: Solving the key equation to determine both (z) and z) from S(z) is the hardest part of the decoding process. The BM algorithm (to be described in Section III) and the eE algorithm can be used to solve (6). If e t, these algorithms nd (z) and z), but if e > t, then the algorithms almost always fail to nd (z) and z). Fortunately, such failures are usually easily detected. Once (z) and z) have been found, the decoder can nd the error locations by checking whether for each j, 0 j n 1. Usually, the decoder computes the value of just before the j-th received symbol r j leaves the decoder circuit. This process is called a Chien search [1], [3]. If is one of the error locations (say X i ). In other words, r j is in error, and needs to be corrected before it leaves the decoder. The decoder can June 19, 2000 DRAFT 6 SUBMITTED TO IEEE TRANS. VLSI SYSTEMS calculate the error value Y i to be subtracted from r j via Forney's error value formula [3]: z 0 (z) z= j denotes the formal derivative of (z). Note that the formal derivative simplies to 0 since we are considering codes over GF(2 m ). Thus, z which is just the terms of odd degree in (z). Hence, the value of z 0 (z) at can be found during the evaluation of (z) at z = j and does not require a separate computation. Note also that (7) can be simplied by choosing C. Reed-Solomon Decoder Structure In summary, a Reed-Solomon decoder consists of three blocks: the syndrome computation (SC) block, the key-equation solver (KES) block, and the Chien search and error evaluator (CSEE) block. These blocks usually operate in pipelined mode in which the three blocks are separately and simultaneously working on three successive received words. The SC block computes the syndromes via (3) usually as the received word is entering the decoder. The syndromes are passed to the KES block which solves (6) to determine the error locator and error evaluator polynomials. These polynomials are then passed to the CSEE block which calculates the error locations and error values via (7) and corrects the errors as the received word is being read out of the decoder. The throughput bottleneck in Reed-Solomon decoders is in the KES block which solves (6): in contrast, the SC and CSEE blocks are relatively straightforward to implement. Hence, in this paper we focus on developing high-speed architectures for the KES block. As mentioned earlier, the key equation (6) can be solved via the eE algorithm (see [19] and [17] for imple- mentations), or via the BM algorithm (see [5] for implementations). In this paper, we develop high-speed architectures for a reformulated version of the BM algorithm because we believe that this reformulated algorithm can be used to achieve much higher speeds than can be achieved by other implementations of the BM and eE algorithms. Furthermore, as we shall show in Section IV.B.4, these new architectures also have lower gate complexity and a simpler control structure than architectures based on the eE algorithm. DRAFT June 19, 2000 III. Existing Berlekamp-Massey (BM) Architectures In this section, we give a brief description of dierent versions of the Berlekamp-Massey (BM) algorithm and then discuss a generic architecture, similar to that in the paper by Reed et al. [13], for implementation of the algorithm. A. The Berlekamp-Massey Algorithm The BM algorithm is an iterative procedure for solving (6). In the form originally proposed by Berlekamp [1], the algorithm begins with polynomials (0; iteratively determines polynomials (r; z) and satisfying the polynomial congruence, for thus obtains a solution (2t; z) and d t; z) to the key equation (6). Two \scratch" polynomials B(r; z) and H(r; z) with initial values B(0; are used in the algorithm. For each successive value of r, the algorithm determines (r; z) and B(r; z) from (r 1; z) and B(r 1; z). Similarly, the algorithm determines z) and H(r; z) from r 1; z) and H(r 1; z). Since S(z) has degree 2t 1, and the other polynomials can have degrees as large as t, the algorithm needs to store roughly 6t eld elements. If each iteration is completed in one clock cycle, then 2t clock cycles are needed to nd the error-locator and error-evaluator polynomials. In recent years, most researchers have used the formulation of the BM algorithm given by Blahut [3] in which only (r; z) and B(r; z) are computed iteratively. Following the completion of the 2t iterations, the error-evaluator polynomial y t; z) is computed as the terms of degree t 1 or less in the polynomial product (2t; z)S(z). An implementation of this version thus needs to store only 4t eld elements, but the computation of t; z) requires an additional t clock cycles. Although this version of the BM algorithm trades space against time, it also suers from the same problem as the Berlekamp version, viz. during some of the iterations, it is necessary to divide each coe-cient of (r; z) by a quantity - r . These divisions are most e-ciently handled by rst computing - 1 r , the inverse of - r , and then multiplying each coe-cient of (r; z) by - 1 r . Unfortunately, regardless of whether this method is used or whether one constructs separate divider circuits for each coe-cient of (r; z), these divisions, which occur inside an iterative loop, are more time-consuming than multiplications. Obviously, if these divisions could be replaced by multiplications, the resulting circuit implementation would have a smaller critical June 19, 2000 DRAFT 8 SUBMITTED TO IEEE TRANS. VLSI SYSTEMS path delay and higher clock speeds would be usable. 2 A less well-known version of the BM algorithm [4], [13], has precisely this property, and has been recently employed in practice [13], [5]. We focus on this version of the BM algorithm in this paper. The inversionless BM (iBM) algorithm is described by the pseudocode shown below. The iBM algorithm actually nds scalar multiples (z) and z) instead of the (z) and dened in (4) and (5). However, it is obvious that the Chien search will nd the same error locations, and it follows from (7) that the same error values are obtained. Hence, we continue to refer to the polynomials computed by the iBM algorithm as (z) and z). As a minor implementation detail, in (4) and thus requires no latches for storage, but the iBM algorithm must store Note also that b 1 (r) which occurs in Steps iBM.2 and iBM.3 is a constant: it has value 0 for all r. The iBM Algorithm Initialization: t. for do begin Step iBM.1 Step iBM.2 Step iBM.3 if -(r) 6= 0 and k(r) 0 then begin else 2 The astute reader will have noticed that the Forney error value formula (7) also involves a division. Fortunately, these divisions can be pipelined because they are feed-forward computations. Similarly, the polynomial evaluations needed in the CSEE block (as well as those in the SC block) are feed-forward computations that can be pipelined. Unfortunately, the divisions in the KES block occur inside an iterative loop, and hence pipelining the computation becomes di-cult. Thus, as was noted in Section II, the throughput bottleneck is in the KES block. DRAFT June 19, 2000 begin t. For r < t, Step iBM.1 includes terms s 1 r+1 (r); s 2 r+2 involving unknown quantities s Fortunately, it is known [3] that deg (r; z) r so that therefore the unknown s i do not aect the value of -(r). Notice also the similarity between Steps iBM.1 and iBM.4. These facts have been used to simplify the architecture that we describe next. B. Architectures Based on the iBM algorithm Due to the similarity of Steps iBM.1 and iBM.4, architectures based on the iBM algorithm need only two major computational structures as shown in Fig. 1: The discrepancy computation (DC) block for implementing Step iBM.1, and The error locator update (ELU) block which implements Steps iBM.2 and iBM.3 in parallel. The DC block contains latches for storing the syndromes s i , the GF(2 m ) arithmetic units for computing the discrepancy -(r) and the control unit for the entire architecture. It is connected to the ELU block which contains latches for storing for (r; z) and B(r; z) as well as GF(2 m ) arithmetic units for updating these polynomials, as shown in Fig. 1. During a clock cycle, the DC block computes the discrepancy -(r) and passes this value together with (r) and a control signal MC(r) to the ELU block which updates the polynomials during the same clock cycle. operations are completed in one clock cycle, we assume that m-bit parallel arithmetic units are being employed. Architectures for such Galois eld arithmetic units can be found in numerous references including [7] and will not be discussed here. June 19, 2000 DRAFT block (in cycles 2t+1 to 3t) block Syndromes from the To the block To the l t (r) l t-1 (r) 0011 d Fig. 1. The iBM architecture. B.1 DC Block Architecture The DC block architecture shown in Fig. 2 has 2t latches constituting the DS shift register that are initialized such that the latches DS contain the syndromes In each of the rst 2t clock cycles, the t +1 multipliers compute the products in Step iBM.1. These are added in a binary adder tree of depth dlog 1)e to produce the discrepancy -(r). Thus, the delay in computing -(r) is T A typical control unit such as the one illustrated in Fig. 2 has counters for the variables r and k(r), and storage for (r). Following the computation of -(r), the control unit computes the OR of the m bits in -(r) to determine whether -(r) is nonzero. This requires m 1 two-input gates arranged in a binary tree of depth dlog 2 me. If the counter for k(r) is implemented in twos-complement representation, then k(r) 0 if and only if the most signicant bit in the counter is 0. The delay in generating signal MC(r) is thus me T or + T and . Finally, once the MC(r) signal is available, the counter for k(r) can be updated. Notice that a twos-complement arithmetic addition is needed if k(r 1. On the other hand, negation in twos-complement representation complements all the bits and then adds 1, and hence the update k(r only the complementation of all the bits in the k(r) counter. We note that it is possible to use ring counters for r and k(r), in which case k(r) is DRAFT June 19, 2000 SC block Syndromes from the CONTROL AND msb CNTR r Fig. 2. The discrepancy computation (DC) block. updated just Tmux seconds after the MC(r) signal has been computed. Following the 2t clock cycles for the BM algorithm, the DC block computes the error-locator polynomial z) in the next t clock cycles. To achieve this, the DS t ; DS latches are reset to zero during the 2t-th clock cycle, so that, at the beginning of the (2t cycle, the contents of the DS register (see Fig. 2) are s Also, the outputs of the ELU block are frozen so that these do not change during the computation of z). From Step iBM.4, it follows that the \discrepancies" computed during the next t clock June 19, 2000 DRAFT cycles are just the coe-cients ! 0 z). The architecture in Fig. 2 is an enhanced version of the one described in [13]. The latter uses a slightly dierent structure and dierent initialization of the DS register in the DC block, which requires more storage and makes it less adaptable to the subsequent computation of the error-locator polynomial. Note that the total hardware requirements of the DC block are 2t m-bit latches, t tipliers, t adders, and miscellaneous other circuitry (counters, arithmetic adder or ring counter, gates, inverters and latches), in the control unit. From Fig. 2, the critical path delay of the DC block is me T or B.2 ELU Block Architecture Following the computation of the discrepancy -(r) and the MC(r) signal in the DC block, the polynomial coe-cient updates of Steps iBM.2 and iBM.3 are performed simultaneously in the ELU block. The processor element PE0 (hereinafter the PE0 processor) that updates one coe-cient of (z) and B(z) is illustrated in Fig. 3(a). The complete ELU architecture is shown in Fig. 3(b), where we see that signals -(r), (r) and MC(r) are broadcast to all the PE0 processors. In addition, the latches in all the PE0 processors are initialized to zero except for which has its latches initialized to the element 1 latches and multipliers, and t adders and multiplexers are needed. The critical path delay of the ELU block is given by B.3 iBM Architecture Ignoring the hardware used in the control section, the total hardware needed to implement the iBM algorithm is 4t multiplexers. The total time required to solve the key equation for one codeword is 3t clock cycles. Alternatively, if t; z) is computed iteratively, the computations require only 2t clock cycles. However, since the computations required to update are the same as that of (r; z), a near-duplicate of the ELU block is needed. 3 This increases the hardware requirements to 6t t; z) < t, the array has only t PE0 processors. DRAFT June 19, 2000 d (r) d (r) l d (r) d (r) l (a) d (r)010101010101010101 PE0001100110011l l l l 0t-1 (b) Fig. 3. The ELU block diagram:(a) the PE0 processor, and (b) the ELU architecture. The latches in are initialized to 1 2 GF those in other PE0s are initialized to 0. In either case, the critical path delay of the iBM architecture can be obtained from Figs. 1, 2, and 3 as which is the delay of the direct path that begins in the DC block starting from the DS i latches, through a multiplier, an adder tree of height dlog (generating the signal -(r)), feeding into the ELU block multiplier and adder before being latched. We have assumed that the indirect path taken by -(r) through the control unit (generating signal MC(r)) feeding into the ELU block multiplexer is faster than the direct path, i.e., Tmult > dlog 2 me T or + T and . This is a reasonable assumption in most technologies. Note that more than half of T iBM is due to the delay in the DC block, and that this contribution increases logarithmically with the error correction capability. Thus, reducing the delay in the DC block is the key to achieving higher speeds. In the next section, we describe algorithmic reformulations of the iBM algorithm that lead to June 19, 2000 DRAFT 14 SUBMITTED TO IEEE TRANS. VLSI SYSTEMS a systolic architecture for the DC block and reduce its critical path delay to TELU . IV. Proposed Reed-Solomon Decoder Architectures The critical path in iBM architectures of the type described in Section III passes through two multipliers as well as the adder tree structure in the DC block. The multiplier units contribute signicantly to the critical path delay and hence reduce the throughput achievable with the iBM architecture. In this section, we propose new decoder architectures that have a smaller critical path delay. These architectures are derived via algorithmic reformulation of the iBM algorithm. This reformulated iBM (riBM) algorithm computes the next discrepancy -(r + 1) at the same time that it is computing the current polynomial coe-cient updates, that is, the and the b i (r 1)'s. This is possible because the reformulated discrepancy computation does not use the explicitly. Furthermore, the discrepancy is computed in a block which has the same structure as the ELU block, so that both blocks have the same critical path delay A. Reformulation of the iBM Algorithm A.1 Simultaneous Computation of Discrepancies and Updates Viewing Steps iBM.2 and iBM.3 in terms of polynomials, we see that Step iBM.2 computes while Step iBM.3 sets B(r+1; z) either to (r; z) or to zB(r; z). Next, note that the discrepancy -(r) computed in Step iBM.1 is actually - r (r), the coe-cient of z r in the polynomial product Much faster implementations are possible if the decoder computes all the coe-cients of (r; z) (and of (r; even though only - r (r) is needed to compute (r and to decide whether B(r is to be set to (r; z) or to z B(r; z). Suppose that at the beginning of a clock cycle, the decoder has available to it all the coe-cients of (r; z) and (r; z) (and, of course, of (r; z) and B(r; z) as well.) Thus, available at the beginning of the clock cycle, and the decoder can compute (r DRAFT June 19, 2000 Furthermore, it follows from (10) and (11) that set to either (r; or to z (r; z B(r; z) S(z). In short, (r are computed in exactly the same manner as are (r Furthermore, all four polynomial updates can be computed simultaneously, and all the polynomial coe-cients as well as - r+1 (r + 1) are thus available at the beginning of the next clock cycle. A.2 A New Error-Evaluator Polynomial The riBM algorithm simultaneously updates four polynomials (r; z), B(r; z), (r; z), and S(z). The 2t iterations thus produce the error-locator polynomial (2t; z) and also the polynomial (2t; z). Note that since c t; z) (2t; z) S(z) mod z 2t it follows from (11) that the low-order coe-cients of (2t; z) are just s t; z), that is, the 2t iterations compute both the error-locator polynomial (2t; z) and the error-evaluator polynomial y t; z) { the additional t iterations of Step iBM.4 are not needed. The high-order coe-cients of (2t; z) can also be used for error evaluation. Let (2t; z) where (h) (z) of degree at most e 1 contains the high-order terms. Since X 1 i is a root of (2t; z), it follows from (11) that (2t; X 1 0: Thus, (7) can be re-written as z 0 (z) z=X 1 We next show that this variation of the error evaluation formula has certain architectural advan- tages. Note that the choice m preferable if (12) is to be used. A.3 Further Reformulation Since the updating of all four polynomials is identical, the discrepancies can be calculated using an ELU block like the one described in Section III. Unfortunately, for the discrepancy - r (r) is computed in processor PE0 r . Thus, multiplexers are needed to route the June 19, 2000 DRAFT appropriate latch contents to the control unit and to the ELU block that computes (r Additional reformulation of the iBM algorithm, as described next, eliminates these multiplexers. We use the fact that for any i < aect the value of any later discrepancy - r+j (r j). Consequently, we need not store - i (r) and i (r) for i < r. Thus, for and the polynomials with initial values ^ It follows that these polynomial coe-cients are updated as set either to - i+1+r or to i+r (r). Note that the discrepancy - r is always in a xed (zero-th) position with this form of update. As a nal comment, note this form of update ultimately produces ^ thus (12) can be used for error evaluation in the CSEE block. The riBM algorithm is described by the following pseudocode. Note that b 1 for all values of r, and these quantities do not need to be stored or updated. The riBM Algorithm Initialization: t. for do begin Step riBM.1 Step riBM.2 if ^ then begin DRAFT June 19, 2000 else begin Next, we consider architectures that implement the riBM algorithm. B. High-speed Reed-Solomon Decoder Architectures As in the iBM architecture described in Section III, the riBM architecture consists of a reformulated discrepancy computation (rDC) block connected to an ELU block. B.1 The rDC Architecture The rDC block uses the processor PE1 shown in Fig. 4(a) and the rDC architecture shown in Fig. 4(b). Notice that processor PE1 is very similar to processor PE0 of Fig. 3(a). How- ever, the contents of the upper latch \ ow through" PE1 while the contents of the lower latch \recirculate". In contrast, the lower latch contents \ ow through" in processor PE0 while the contents of the upper latch \recirculate". Obviously, the hardware complexity and the critical path delays of processors PE0 and PE1 are identical. Thus, assuming as before that Tmult > dlog 2 me T or +T and , we get that T Note that the delay is independent of the error-correction capability t of the code. The hardware requirements of the proposed architecture in Fig. 4 are 2t PE1 processors, that is, 4t latches, 4t multipliers, 2t adders, and 2t multiplexers, in addition to the control unit which is the same as that in Fig. 2. June 19, 2000 DRAFT d (r) d (r) d (r) d (r) d (r) d (r) d d (a) s d (r) sw sw s (b) Fig. 4. The rDC block diagram:(a) the PE1 processor, and (b) the rDC architecture. B.2 The riBM Architecture The overall riBM architecture is shown in Fig. 5. It uses the rDC block of Fig. 4 and the ELU block in Fig. 3. Note that the outputs of the ELU block do not feed back into the rDC block. Both blocks have the same critical path delay of T add and since they operate in parallel, our proposed riBM architecture achieves the same critical path delay: which is less than half the delay T add of the enhanced iBM architecture. As noted in the previous subsection, at the end of the 2t-th iteration, the PE1 i s, contain the coe-cients of (2t; z) which can be used for error evaluation. Thus, 2t clock cycles are used to determine both (z) and (h) (z) as needed in (12). Ignoring the control unit, the hardware requirement of this architecture is 3t DRAFT June 19, 2000 d (r) l l l l 0t-1 ELU- block CONTROL rDC-block Fig. 5. The systolic riBM architecture. multiplexers. This compares very favorably with the multiplexers needed to implement the enhanced iBM architecture of Section III in which both the error-locator and the error-evaluator polynomial are computed in 2t clock cycles. Using only t 1 additional multipliers and t additional multiplexers, we have reduced the critical path delay by more than 50%. Furthermore, the riBM architecture consists of two systolic arrays and is thus very regular. B.3 The RiBM Architecture We now show that it is possible to eliminate the ELU block entirely, and to implement the BM algorithm in an enhanced rDC block in which the array of 2t PE1 processors has been lengthened into an array of 3t processors as shown in Fig. 6. In this completely systolic architecture, a single array computes both (z) and (z). Since the t processors eliminated from the ELU block re-appear as the t additional PE1 processors, the RiBM architecture has the same hardware complexity and critical path delay as the riBM architecture. June 19, 2000 DRAFT However, its extremely regular structure is esthetically pleasing, and also oers some advantage in VLSI circuit layouts.h t-1t d s CONTROL l (r) s Fig. 6. The homogenous systolic RiBM architecture. An array of PE0 processors in the riBM architecture (see Fig. 5) carries out the same polynomial computation as an array of PE1 processors in the RiBM architecture (see Fig 6), but in the latter array, the polynomial coe-cients shift left with each clock pulse. Thus, in the RiBM architecture, suppose that the initial loading of PE1 0 , PE1 1 , . , PE1 2t 1 is as in Fig. 4, while are loaded with zeroes, and the latches in PE1 3t are loaded as the iterations proceed, the polynomials ^ are updated in the processors in the left-hand end of the array (eectively, (r; z) and (r; z) get updated and shifted leftwards). After 2t clock cycles, the coe-cients of (h) (z) are in processors Next, note that PE1 3t contains (0; z) and B(0; z), and as the iterations proceed, (r; z) and B(r; z) shift leftwards through the processors in the right-hand end of the array, with i (r) and b i (r) being stored in processor PE1 3t r+i . After 2t clock cycles, processor PE1 t+i contains i (2t) and b i (2t) for t. Thus, the same array is carrying out two separate computations. These computations do not interfere with one another. Polynomials DRAFT June 19, 2000 are stored in processors numbered 3t r or higher. On the other hand, since deg (r; is known to be an upper bound on deg (r; z). It is known [3] that l(r) is a nondecreasing function of r, and that it has maximum value errors have occurred. Hence, 2t 1 r thus, as (r; z) and B(r; z) shift leftwards, they do not over-write the coe-cients of ^ We denote the contents of the array in the RiBM architecture as polynomials ~ ~ z) with initial values ~ z 3t . Then, the RiBM architecture implements the following pseudocode. Note that ~ - 3t+1 values of r, and this quantity does not need to be stored or updated. The RiBM Algorithm Initialization: ~ for do begin then begin ~ else begin ~ June 19, 2000 DRAFT 22 SUBMITTED TO IEEE TRANS. VLSI SYSTEMS B.4 Comparison of Architectures Table I summarizes the complexity of the various architectures described so far. It can be seen that, in comparison to the conventional iBM architecture (Berlekamp's version), the proposed riBM and RiBM systolic architectures require t 1 more multipliers and t more multiplexers. All three architectures require the same numbers of latches and adders, and all three architectures require 2t cycles to solve the key equation for a t-error-correcting code. The riBM and RiBM architectures require considerably more gates than the conventional iBM architecture (Blahut's version), but also require only 2t clock cycles as compared to the 3t clock cycles required by the latter. Furthermore, since the critical path delay in the riBM and RiBM architectures is less than half the critical path delay in either of the iBM architectures, we conclude that the new architectures signicantly reduce the total time required to solve the key equation (and thus achieve higher throughput) with only a modest increase in gate count. More important, the regularity and scalability of the riBM and RiBM architectures creates the potential for automatically generating regular layouts (via a core generator) with predictable delays for various values of t and m. Comparison of the riBM and RiBM architectures with eE architectures is complicated by the fact that most recent implementations use folded architectures in which each processor element in the systolic array has only a few arithmetic units, and these units carry out all the needed computations via time-division-multiplexing. For example, the hypersystolic eE architecture in [2] has 2t elements each containing only one multiplier and adder. Since each iteration of the Euclidean algorithm requires 4 multiplications, the processors of [2] need several multiplexers to route the various operands to the arithmetic units, and additional latches to store one addend until the other addend has been computed by the multiplier, etc. As a result, the architecture described in [2] requires not only many more latches and multiplexers, but also many more clock cycles than the riBM and RiBM architectures. Furthermore, the critical path delay is slightly larger because of the multiplexers in the various paths. On the other hand, nite-eld multipliers themselves consist of large numbers of gates (possibly as many as 2m 2 , but fewer if logic minimization techniques are used), and thus a complete comparison of gate counts for the two architectures requires very specic details about the multipliers. Nonetheless, a rough DRAFT June 19, 2000 I Comparison of Hardware Complexity and Path Delays Architecture Adders Multipliers Latches Muxes Clock Critical cycles path delay iBM iBM Euclidean Euclidean [2](folded) 2t comparison is that the riBM and RiBM architectures require three times as many gates as the hypersystolic eE architecture, but solve the key equation in one-sixth the time. It is, of course, possible to implement the eE algorithm with more complex processor elements, as described by Shao et al. [14]. Here, the 4 multiplications in each processor are computed using 4 separate multipliers. The architecture described in [14] uses only 2t+1 processors as compared to the 3t processors needed in the riBM and RiBM architectures, but each processor in [14] has 4 multipliers, 4 multiplexers, and 2 adders. As a result, the riBM and RiBM architectures compare very favorably to the eE architecture of [14] { the new architectures achieve the same (actually slightly higher) throughput with much smaller complexity. One nal point to be made with respect to the comparison between the riBM and RiBM architectures and the eE architectures is that the controllers for the systolic arrays in the former are actually much simpler. In the eE architecture of [14], each processor also has a \control section" that uses an arithmetic adder, comparator, and two multiplexers. 2dlog 2 te bits of arithmetic data are passed from processor to processor in the array, and these are used to generate multiplexer control signals in each processor. Similarly, the eE architecture of [2] has a separate control circuit for each processor. The delays in these control circuits are not accounted for in the critical path delays for the eE architectures that we have listed in Table I. In contrast, all the multiplexers in the riBM and RiBM architectures receive the same signal and the computations in these architectures is purely systolic in the sense that all processors carry out exactly the same computation in each cycle, with all the multiplexers set the same way in all the June 19, 2000 DRAFT processors { there are no cell-specic control signals. Preliminary Layout Results Preliminary layout results from a core generator are shown in Fig. 7 for the KES block for a 4-error-correcting Reed-Solomon code over GF(2 8 ). The processing element PE1 is shown in Fig. 7(a) where the upper 8 latches store the element ~ - r while the lower 8 latches store the element ~ r . A complete RiBM architecture is shown in Fig. 7(b) where the 13 PE1 processing elements are arrayed diagonally and the error locator and error evaluator polynomials output latches can be seen to be arrayed vertically. The critical path delay of the RiBM architecture as reported by the synthesis tool in SYNOPSYS was 2:13 ns in TSMC's 0:25m 3:3V CMOS technology. (a) (b) Fig. 7. The RiBM architecture synthesized in a 3:3V , 0:25m CMOS technology:(a) the PE1 processing element, and (b) the RiBM architecture. In the next section, we develop a pipelined architecture that further reduces the critical path delay by as much as an order of magnitude by using a block-interleaved code. DRAFT June 19, 2000 V. Pipelined Reed-Solomon Decoders The iterations in the original BM algorithm were pipelined using the look-ahead transformation [12] by Liu et al. [9], and the same method can be applied to the riBM and RiBM algorithms. However, such pipelining requires complex overhead and control hardware. On the other hand, pipeline interleaving (also described in [12]) of a decoder for a block-interleaved Reed-Solomon code is a simple and e-cient technique that can reduce the critical path delay in the decoder by an order of magnitude. We describe our results for only the RiBM architecture of Section IV but the same techniques can also be applied to the riBM architecture as well as to the decoder architectures described in Section III. A. Block-Interleaved Reed-Solomon Codes A.1 Block Interleaving Error-correcting codes for use on channels in which errors occur in bursts are often interleaved so that symbols from the same codeword are not transmitted consecutively. A burst of errors thus causes single errors in multiple codewords rather than multiple errors in a single codeword. The latter occurrence is undesirable since it can easily overwhelm the error-correcting capabilities of the code and cause a decoder failure or decoder error. Two types of interleavers, block interleavers and convolutional interleavers, are commonly used (see, e.g. [16], [18].) We restrict our attention to block-interleaved codes. Block-interleaving an (n; k) code to depth M results in an (nM; kM) interleaved code whose codewords have the property that (c (n 1)M+i ; c (n is a codeword in the (n; Equivalently, a codeword of the (nM; kM) code is a multichannel data stream in which each of the M channels carries a codeword of the (n; code. A.2 Interleaving via Memory Arrays The usual description (see, e.g., [16], [18]) of an encoder for the block-interleaved (nM; kM) code involves partitioning kM data symbols (d blocks of k consecutive symbols, and encoding each block into a codeword of the (n; codewords are stored row-wise into an M n memory array. The memory is then read out column-wise to form the block-interleaved codeword. Notice that the block-interleaved code- June 19, 2000 DRAFT 26 SUBMITTED TO IEEE TRANS. VLSI SYSTEMS word is systematic in the sense that the parity-check symbols follow the data symbols, but the Reed-Solomon encoding process described in Section II-A results in a block-interleaved codeword in which the data symbols are not transmitted over the channel in the order in which they entered the encoder. 4 At the receiver, the interleaving process is reversed by storing the nM received symbols column-wise into an M n memory array. The memory is then read out row-wise to received words of length n that can be decoded by a decoder for the (n; code. The information symbols appear in the correct order in the de-interleaved stream, and the decoder output is passed on to the destination. A.3 Embedded Interleavers An alternative form of block interleaving embeds the interleaver into the encoder, thereby transforming it into an encoder for the (nM; kM) code. For interleaved Reed-Solomon codes, the mathematical description of the encoding process is that the generator polynomial of the interleaved code is G(z M ), where G(z) denotes the generator polynomial of the (n; dened in (1), and the codeword is formed as described in Section II-A { i.e., with D(z) now denoting the data polynomial d kM 1 z of degree kM 1, the polynomial z (n k)M D(z) is divided by G(z M ) to obtain the remainder P (z) of degree (n k)M 1. The transmitted codeword is z (n k)M D(z) P (z). In essence, the data stream treated as if it were a multichannel data stream and the stream in each channel is encoded with the (n; code. The output of the encoder is a codeword in the block-interleaved Reed-Solomon code (no separate interleaver is needed) and it has the property that the data symbols are transmitted over the channel in the order in which they entered the encoder. The astute reader will have observed already that the encoder for the (nM; kM) code is just a delay-scaled encoder for the (n; code. The delay-scaling transformation of an architecture replaces every delay (latch) in the architecture with M delays, and re-times the architecture to account for the additional delays. The encoder treats its input as a multichannel data stream and produces a multichannel output data stream, that is, a block-interleaved Reed-Solomon In fact, the data symbol ordering is that which is produced by interleaving the data stream in blocks of k symbols to depth M . DRAFT June 19, 2000 codeword. Note also that while the interleaver array has been eliminated, the delay-scaled encoder uses M times as much memory as the conventional encoder. Block-interleaved Reed-Solomon codewords produced by delay-scaled encoders contain the data symbols in the correct order. Thus, a delay-scaled decoder can be used to decode the received word of nM symbols, and the output of the decoder also will have the data symbols in the correct order. Note that a separate de-interleaver array is not needed at the receiver. However, the delay-scaled decoder uses M times as much memory as the conventional decoder. For example, delay-scaling the PE1 processors in the RiBM architecture of Fig. 6 results in the delay-scaled processor DPE1 shown in Fig. 8. Note that for 0 i 2t 1, the top and bottom sets of M latches in DPE1 i are initialized with the syndrome set S 0;M 1 where s i;j is the i th syndrome of the j th codeword. For 2t i 3t 1, the latches in DPE1 i are initialized to 0 while the latches in DPE1 3t are initialized to 1 2 GF(2 m ). After 2tM clock cycles, processors DPE1 0 { DPE1 t 1 contain the interleaved error-evaluator polynomials while processors DPE1 t { DPE1 2t contain the interleaved error-locator polynomials. MD MD Fig. 8. Delay-scaled DPE1 processor. Initial conditions in the latches are indicated in ovals. The delay-scaled RiBM architecture is obtained by replacing the PE1 processors in Fig. 6 with DPE1 processor and delay-scaling the control unit as well. We remark that delay-scaled decoders can also be used to decode block-interleaved Reed-Solomon codewords produced by memory array interleavers. However, the data symbols at the June 19, 2000 DRAFT 28 SUBMITTED TO IEEE TRANS. VLSI SYSTEMS output of the decoder will still be interleaved and an M k memory array is needed for de-interleaving the data symbols into their correct order. This array is smaller than the M n array needed to de-interleave the M codewords prior to decoding with a conventional decoder, but the conventional decoder also uses less memory than the delay-scaled decoder. Delay-scaling the encoder and decoder eliminates separate interleavers and de-interleaver and is thus a natural choice for generating and decoding block-interleaved Reed-Solomon codewords. However, a delay-scaled decoder has the same critical path delay as the original decoder, and hence cannot achieve higher throughput than the original decoder. On the other hand, the extra delays can be used to pipeline the computations in the critical path, and this leads to signicant increases in the achievable throughput. We discuss this concept next. B. Pipelined Delay-Scaled Decoders The critical path delay in the RiBM architecture is mostly due to the nite-eld multipliers in the processors. For the delay-scaled processors DPE1 shown in Fig. 8, these multipliers can be pipelined and the critical path delay reduced signicantly. We assume that M m and describe a pipelined nite-eld multiplier with m stages. B.1 A Pipelined Multiplier Architecture While pipelining a multiplier, especially if it is a feedforward structure, is trivial, it is not so in this case. This is because for RS decoders the pipelining should be done in such a manner that the initial conditions in the pipelining latches are consistent with the syndrome values generated by the SC block. The design of nite-eld multipliers depends on the choice of basis for the representation. Here, we consider only the standard polynomial basis in which the m-bit byte represents the Galois eld element The pipelined multiplier architecture is based on writing the product of two GF X and Y as Let pp i denote the sum of the rst i terms in the sum above. The multiplier processing element shown in Fig. 9(a) computes pp i+1 by adding either X i (if y DRAFT June 19, 2000 Simultaneously, MPE i multiplies X i by . Since is a constant, this multiplication requires only XOR gates, and can be computed with a delay of only T On the other hand, the delay in computing pp i+1 is T . Thus, the critical path delay is an order of magnitude smaller than T tremendous speed gains can be achieved if the pipelined multiplier architecture is used in decoding a block-interleaved Reed-Solomon code. Practical considerations such as the delays due to pipelining latches, clock skew and jitter will prevent the fullest realization of the speed gains due to pipelining. Nevertheless, the pipelined multiplier structure in combination with the systolic architecture will provide signicant gains over existing approaches. DD pp pp pp pp a (m-1)D (m-2)D D (b) Fig. 9. The pipelined multiplier block diagram:(a) the multiplier processing element (MPE), and (b) the multiplier architecture. Initial conditions of the latches at the y input are indicated in ovals. The pipelined multiplier thus consists of m MPE processors connected as shown in Fig. 9(b) with inputs pp and the y i 's. The initial conditions of the latches at the y input are zero, and therefore the initial conditions of the lower latches in the MPEs do not aect the circuit operation. The product XY appears in the upper latch of MPEm 1 after m clock cycles and each succeeding clock cycle thereafter computes a new product. Notice also that during the rst June 19, 2000 DRAFT clock cycles, the initial contents of the upper latches of the MPEs appear in succession at the output of MPEm 1 . This property is crucial to the proper operation of our proposed pipelined decoder. B.2 The Pipelined Control Unit If the pipelined multiplier architecture described above (and shown in Fig. is used in the DPE1 processors of Fig. 8, the critical path delay of DPE1 is reduced from Tmult + T add to just . Thus, the control unit delay in computing MC(r), which is inconsequential in the RiBM architecture (as well as in the iBM and riBM architectures, and the delay-scaled versions of all these), determines the largest delay in a pipelined RiBM architecture. Fortunately, the computation of MC(r) can also be pipelined in (say) stages. This can be done by noting that d delays from DPE1 0 in the M-delay scaled RiBM architecture (see Fig. 8) can be retimed to the outputs of the control unit and then subsequently employed to pipeline it. Note, however, that the d latches in DPE1 0 that are being retimed are initialized to S 0;d 1 i at the begininng of every decoding cycle. Hence, the retimed latches in the control unit will need to be initialized to values that are a function of syndromes S 0;d 1 i . This is not a problem because these syndromes will be produced by the SC block in the beginning of each decoding cycle. B.3 Pipelined Processors If pipelined multiplier units as described above are used in a delay-scaled DPE1 processor, and the control unit is pipelined as described above, then we get the pipelined PPE1 i processor shown in Fig. 10 (and the pipelined RiBM (pRiBM) architecture also described in Fig. 10). The initial values stored in the latches are the same as were described earlier for the DPE1 processors. Note that some of the latches that store the coe-cients of ~ are part of the latches in the pipelined multiplier. However, the initial values in the latches in the lower multiplier in Fig. 10 are 0. Thus, during the rst m clock cycles ow through into the leftmost latches without any change. From the above description, it should be obvious that the pRiBM architecture based on the PPE1 processor of Fig. 10 has a critical path delay of DRAFT June 19, 2000 mD mD d (Mr+j)(M-d)D d (M-m)D Fig. 10. Pipelined PPE1 processor. Initial conditions in the latches are indicated in ovals. The pipelined RiBM architecture is obtained by replacing the PE1 processors in Fig. 6 with PPE1 processor and employing the pipelined delay-scaled controller. Thus, the pRiBM architecture can be clocked at speeds that can be as much as an order of magnitude higher than those achievable with the unpipelined architectures presented in Sections III and IV. C. Decoders for Non-interleaved Codes The pRiBM architecture can decode a block-interleaved code at signicantly faster rates than the RiBM architecture can decode a non-interleaved code. In fact, the dierence is large enough that a designer who is asked to devise a decoder for non-interleaved codes should give serious thought to the following design strategy. Read in M successive received words into an block-interleaver memory array. Read out a block-interleaved received word into a decoder with the pRiBM architecture. Decode the block-interleaved word and read out the the data symbols into a block-deinterleaver memory array. Read out the de-interleaved data symbols from the deinterleaver array. Obviously, similar decoder design strategies can be used in other situations as well. For example, to decode a convolutionally interleaved code, one can rst de-interleave the received words, and then re-interleave them into block-interleaved format for decoding. Similarly, if a block- interleaved code has very large interleaving depth M , the pRiBM architecture may be too large to implement on a single chip. In such a case, one can de-interleave rst and then re-interleave to a suitable depth. In fact, the \de-interleave and re-interleave" strategy can be used to construct June 19, 2000 DRAFT a universal decoder around a single decoder chip with xed interleaving depth. VI. Concluding Remarks We have shown that the application of algorithmic transformations to the Berlekamp-Massey algorithm result in the riBM and RiBM architectures whose critical path delay is less than half that of conventional architectures such as the iBM architecture. The riBM and RiBM architectures use systolic arrays of identical processor elements. For block-interleaved codes, the de-interleaver can be embedded in the decoder architecture via delay-scaling. Furthermore, pipelining the multiplications in the delay-scaled architecture result in an order of magnitude reduction in the critical path delay. In fact, the high speeds at which the pRiBM architecture can operate makes it feasible to use it to decode non-interleaved codes by the simple stratagem of internally interleaving the received words, decoding the resulting interleaved word using the pRiBM architecture, and then de-interleaving the output. Future work is being directed towards integrated circuit implementations of the proposed architectures and their incorporation into broadband communications systems such as those for very high-speed digital subscriber loops and wireless systems. VII. Acknowledgments The authors would like to thank the reviewers for their constructive criticisms which has resulted in signicant improvements in the manuscript. --R Algebraic Coding Theory Theory and Practice of Error-Control Codes Applied Coding and Information Theory for Engineers Control Systems for Digital Communication and Storage --TR systems for digital communication and storage Applied coding and information theory for engineers --CTR Kazunori Shimizu , Nozomu Togawa , Takeshi Ikenaga , Satoshi Goto, Reconfigurable adaptive FEC system with interleaving, Proceedings of the 2005 conference on Asia South Pacific design automation, January 18-21, 2005, Shanghai, China Tong Zhang , Keshab K. Parhi, On the high-speed VLSI implementation of errors-and-erasures correcting reed-solomon decoders, Proceedings of the 12th ACM Great Lakes symposium on VLSI, April 18-19, 2002, New York, New York, USA Y. W. Chang , T. K. Truong , J. H. Jeng, VLSI architecture of modified Euclidean algorithm for Reed-Solomon code, Information Sciences: an International Journal, v.155 n.1-2, p.139-150, 1 October Zhiyuan Yan , Dilip V. Sarwate, Universal Reed-Solomon decoders based on the Berlekamp-Massey algorithm, Proceedings of the 14th ACM Great Lakes symposium on VLSI, April 26-28, 2004, Boston, MA, USA Jung H. Lee , Jaesung Lee , Myung H. Sunwoo, Design of application-specific instructions and hardware accelerator for reed-solomon codecs, EURASIP Journal on Applied Signal Processing, v.2003 n.1, p.1346-1354, January Zhiyuan Yan , Dilip V. Sarwate, New Systolic Architectures for Inversion and Division in GF(2^m), IEEE Transactions on Computers, v.52 n.11, p.1514-1519, November

systolic architectures;interleaved codes;berlekamp-massey algorithm;pipelined decoders;reed-solomon codes

505524

Delay fault testing of IP-based designs via symbolic path modeling.

Predesigned blocks called intellectual property (IP) cores are increasingly used for complex system-on-a-chip (SOC) designs. The implementation details of IP cores are often unknown or unavailable, so delay testing of such designs is difficult. We propose a method that can test paths traversing both IP cores and user-defined blocks, an increasingly important but little-studied problem. It models representative paths in IP circuits using an efficient form of binary decision diagram (BDD) and generates test vectors from the BDD model. We also present a partitioning technique, which reduces the BDD size by orders of magnitude and makes the proposed method practical for large designs. Experimental results are presented which show that it robustly tests selected paths without using extra logic and, at the same time, protects the intellectual contents of IP cores.

Introduction While reusable predesigned circuits called intellectual property (IP) circuits or cores are becoming increasingly popular for VLSI system-on-a-chip (SOC) designs [1, 3, 8, 11, 12, 14, 21], they present difficult testing problems that existing methodologies cannot adequately handle. Path delay verification of IP-based designs is among the most challenging problems because the implementation details of the IP circuits are hidden. This is particularly the case when the paths traverse both IP circuits and user-defined circuits. Conventional delay fault testing methods using standard scan [6, 22], boundary scan [3, 8, 21], or enhanced scan methods [7, 9] cannot test such paths effectively. We previously proposed a method called STSTEST [12] which can test complete paths between IP and UD circuits, but requires extra scan logic. Nikolos et al. [16] suggest calculating the delays of complete paths by measuring the delays of partial paths. This method appears suited to delay evaluation of a prototype circuit but is impractical for production testing of high-speed circuits due to the difficulty of accurately measuring analog delay values. To address these problems, we propose a delay testing method dubbed symbolic path modeling- based testing (SPMTEST) which can directly test selected complete paths between IP and UD circuits without using extra logic. It employs an IP modeling scheme that abstracts the information of the IP cir- cuit's paths using a special style of binary decision diagram (BDD), and protects the intellectual content of the IP circuits. We also present an associated ATPG algorithm that generates robust delay tests for the paths using the symbolic IP path models. Figure 1 shows an example design where a UD circuit UDB1 and an IP circuit IPB1 form a single combinational block. Like many delay fault testing methods [7, 9, 13, 17, 18], we assume that to sensitize the target paths, two test patterns are applied via an enhanced scan register R1 which uses two flip-flops in each scan cell to hold a pair of test patterns. Complete single-cycle paths exist from register R1 to that traverse both UDB1 and IPB1, such as the one marked by the thick solid line. Neither the IP providers nor Extra scan register (boundary scan or STS) Intellectual property (IP) circuit User-defined (UD) circuits Fig. testing of a circuit containing IP and UD blocks with boundary scan which can test only partial paths, or selectively transparent scan which can test complete paths. Enhanced scan Enhanced scan the system designers can generate tests for these complete paths using conventional ATPG methods for path-delay faults. This is because UDB1's implementation details are unknown to the IP providers, while IPB1's implementation is hidden from the system designers. For stuck-at fault testing, extra logic such as boundary scan registers [3, 8, 21] or multiplexers are often inserted between UDB1 and IPB1. However, precomputed tests applied to the IP circuit via such extra logic cannot detect a delay fault involving a complete path from R1 to R2. For example, precomputed tests applied via boundary scan in Fig. 1 can sensitize only a partial path such as the one indicated by the thick dashed line. To allow testing of the complete paths linking UD and IP circuits, the STSTEST method [12] we proposed previously employs a new type of scan register called a selectively transparent scan (STS) regis- ter. With the STS register in Fig. 1 replacing the boundary scan, any complete path like the highlighted one can be tested. In the test mode, part of the STS register on the path is made transparent, while other parts of the STS register hold values pre-selected to satisfy the conditions required for the path sensitization. An IP modeling technique for STSTEST is defined in [12] that can test complete paths of a specified delay range and protect the implementation details of the IP circuits. The overhead of the STS registers can limit their use in high-performance or area-critical circuits. This overhead tends to be more significant in designs like Fig. 2(a) where complete paths traverse more than one IP and UD block, and STS registers need to be inserted between every two blocks. The SPMTEST method proposed here and illustrated in Fig. 2(b) can test complete paths without needing extra scan registers. As in STSTEST, we require the IP providers to supply IP models that allow system designers to generate test vectors for complete paths. Unlike STSTEST which specifies a test cube (a) (b) Fig. Testing complete paths that traverse multiple IP and UD blocks using (a) STSTEST which requires extra STS registers between every two blocks, and (b) SPMTEST which requires no extra logic. Test IP models based IP models BDD-based (Control) (ALU) (ALU) (Control) register register register Enhanced scan registers for each selected path in its IP models, SPMTEST abstracts all the conditions required to compute tests for the selected paths by means of an efficient form of BDD [2]. This symbolic IP modeling technique eliminates the need for STS registers. To handle large IP circuits, we propose a circuit partitioning technique that decomposes the BDDs and leads to IP models of practical size. We also present an ATPG algorithm that acts directly on the decomposed BDDs and thus can protect the IP circuit's implementation details. Given symbolic IP models, SPMTEST finds 2-pattern robust tests for complete paths of a specified delay range, if the tests exist. Finally, we present a CAD tool implementing SPMTEST and experimental results which show that SPMTEST is a cost-efficient solution for the delay testing problem of IP-based designs. The remainder of the paper is organized as follows. Section 2 introduces the BDD-based IP modeling procedure, while Sec. 3 describes the circuit partitioning technique for large IP designs. Section 4 presents the ATPG procedure that computes the final test vectors using the IP models. Section 5 describes experimental results obtained with the ISCAS benchmark circuits. Modeling In order to allow the system designers to generate test vectors, an IP model should specify the sensitization conditions for selected paths in the IP circuit. First we show how we construct such a model using BDDs for the selected paths. Then we describe a path selection scheme that yields all complete paths whose delays exceed some specified threshold. Symbolic path modeling: The basic idea of our symbolic path modeling approach is inspired by (1) the conditional delay model proposed by Yalcin and Hayes [23] which employs BDDs to find a critical path, and (2) the BDD-based path-delay fault testing method in [2]. The conditional delay model demonstrates that a hierarchical representation of path sensitization conditions can efficiently identify many false paths. SPMTEST also exploits hierarchical structures consisting of IP and UD blocks to identify untestable paths and to generate test vectors. Bhattacharya et al. [2] show that BDDs can be successfully used for delay-fault ATPG, and report promising results for many benchmark circuits. They represent each path's sensitization conditions by a BDD from which a test vector can be derived. To avoid the difficulty of representing the rising and falling transition values by BDDs (which can represent only 1 and 0 values explicitly), they assume that all the off-path primary inputs have stable 0 or 1 values. This assumption allows a BDD to represent the conditions required to sensitize the path and avoid static hazards on the path. This assumption cannot be made for IP modeling, however, since any primary input of the IP circuits can receive transition or hazard signals from other blocks that drive the IP circuit's inputs. Therefore, we employ an encoding technique that represents 4each signal by 2-bit values so that any signal transitions and hazard conditions can be represented by BDDs. For example, Fig. 3 shows the ISCAS-85 benchmark circuit c17 regarded as an IP circuit. Suppose we want to model the highlighted paths P IP1 and P IP2 using the robust test conditions proposed in [13]. To test a path robustly, the following conditions on the side input values of the gates along the path that need to be satisfied. When the on-path input has a controlling final value, the side inputs must have non-controlling stable values; when the on-path input has a non-controlling final value, the side inputs must have non-controlling final values with arbitrary initial values. Here we use a 7-valued logic to represent the signal values as in [5, 11, 13]; this logic is defined in Fig. 4. . Robust test condition for . Robust test condition for Figure 4 also shows how the seven values are encoded for BDD representation. A similar encoding technique was employed earlier in a delay fault testing method based on a CNF formulation of the satisfiability problem [5]. Here v f represents a signal's final value, while v s represents the stability of the signal, that is, v the signal is stable. Let R(P IPi ) denote the robust test condition for path P IPi . The robust test conditions for P IP1 and P IP2 are encoded as follows: F R Fig. 3 : The ISCAS benchmark circuits c17 viewed as a small IP circuit. R F F Logic values BDD encoding (final value v f , stability v s ) Value Interpretation F Falling transition (0, R Rising transition (1, f Unknown-to-0 transition (0, X) r Unknown-to-1 transition (1, X) Unknown (X, X) Fig. 4 : The 7-valued logic for robust tests and the corresponding 2-bit encoding used for BDD representation. . R(P IP1 . R(P IP2 Note that R(P IPi ) is constructed by ANDing all v f 's and v s 's of non-X values. In order to construct BDDs representing R(P IPi ), the primitive logic operations AND, OR, and NOT are modified to apply to signal values of the form (v f , v s ); see Fig. 5. The same encoding scheme is found [5]. Each encoded output value (z f , z s ) in the tables of Fig. 5 is obtained by applying the indicated logic operations to x f , x s , y f , and y s . For example, in the AND case, z We apply the modified logic operations to every gate in the IP circuit recursively, starting from the primary inputs until all BDDs representing each encoded signal are obtained. Then, for each selected path path model is constructed by ANDing the BDDs representing each component of R(P IPi ). Figure 6 shows symbolic path models constructed for Fig. 3 in this way. The variables of these BDDs listed at the left are the primary inputs in the form of encoded value pairs (I if , I is ). It is not possible to reverse-engineer the symbolic path models to recover the circuit's gate-level structure, so this modeling method protects the intellectual content of the IP circuits. Symbolic path modeling also can easily identify untestable IP paths a priori and exclude them from the IP model. This follows from the fact that if P IPi is untestable, the BDD for R(P IPi ) must denote the zero function. The foregoing technique can be easily extended to handle other delay fault test conditions by using different encoding schemes. Our Fig. Encoded logic operations for BDD construction z f z s defined by is defined by (x f , x s ) is defined by CAD tool implementing SPMTEST can also handle the hazard-free robust test conditions [10] using a 3- bit signal encoding scheme. We focus only on robust testing with 2-bit encoding in this paper. The ATPG procedure which we discuss later computes tests by justifying the robust test conditions given by the symbolic path models of an IP block B via other IP or UD blocks that drive B. Consequently, the IP block's output functions are needed for test generation, so the IP models also contain BDDs representing functions of O jf and O js for all outputs O j 's of the IP block. The output functions of IP blocks often must be provided to the system designers for simulation and verification of the entire system, and are not intellectual content for many circuits such as arithmetic circuits whose functions are well known. Finally, for each selected IP path, we include in the IP model the following path information: (1) the input and output terminals of the path, (2) the transition direction R (rising) or F (falling) at the path terminals, and (3) the delay of the path. Figure 7 shows an IP model constructed for the example of Fig. 3. It consists of Fig. representing the robust test conditions for (a) P IP1 and (b) P IP2 in Fig. 3. R(P IP1 )(a) (b) Branch to the high child Branch to the low child Branch to the low child with complement Fig. 7 : IP model for c17 when two IP paths are selected. )Selected IP path information Path ID I/O terminal I/O transition Delay (ns) BDDs for the four output functions and the two selected paths, and the associated path information. We next describe the path selection method for constructing IP models. STSTEST introduced a path selection method for IPB's that derives all complete paths of a certain delay range in (UDB, IPB) block pairs. The same method is used by SPMTEST. However, SPMTEST can be also applied to any combination of IPBs and UDBs with only minor modifications to the path selection scheme. We first describe path selection for the (UDB, IPB) case, and then generalize to other cases. Due to the enormous number of paths in large circuits, we only test paths whose delays exceed some specified threshold, an approach commonly employed by delay fault testing methods [13, 22]. To test such complete paths in (UDB, IPB), therefore, we consider all IP paths that can potentially yield complete paths exceeding the threshold when combined with certain UD paths. Figure 8 shows an example (UDB, IPB) pair consisting of the smallest ISCAS-89/85 benchmark circuits cs27 and c17. I k denotes the k-th input port of IPB. We compute the path delays using the Synopsys cell library [20], and treat each path as two separate paths with rising and the falling transitions, as in [13, 17, 18]. If the IP models just include paths that meet a certain path-length threshold, they may not yield all required complete paths. For example, suppose only the critical IP path P marked by the dashed line exceeds the threshold delay and so is included in the IP model. Then, the critical complete path of (UDB, IPB) indicated by thick solid line cannot be derived from path P. To avoid such problems, we select IP paths by assuming that all UD paths have their maximum allowable delay (slack limit), which is the delay margin for each IPB input I k left after subtracting from the clock period the longest delay of the IP paths starting from I k . Figure 9(a) shows the maximum allowable delays (the length of the thick arrows) for one clock period, which is formed by positioning all the IP paths to align the longest IP paths starting from every I k with the right end of the critical IP path. From Fig. 9(a) we select IP paths that extend beyond the IP path threshold denoted by the dashed line; in this example P 1:6 is selected. It follows from Fig. 9(a) that all complete paths exceeding the complete-path threshold of Fig. 9(b) can be derived from the six selected I 2 I 4 I 1 I 0 Critical path P of IPB F R R F Fig. consisting of the ISCAS benchmark circuits cs27 and c17. The critical complete path R F R F F R R F . . g10 R IP paths. For example, Fig. 9(b) shows six such complete paths, which are guaranteed to be tested. This approach yields all complete paths longer than the threshold delay determined by the clock period and the IP path threshold delay. For convenience, we represent the IP path threshold by T IP - critical IP-path delay, where T IP denotes a threshold factor, 0 - T IP - 1. For example, if the IP provider chooses T for the IPB of Fig. 8, a total of paths will be included in the IP model in the form shown in Fig. 7. Next we discuss path selection for a few other IPB and UDB combinations. The above selection scheme for (UDB, IPB) can be modified for the (IPB, UDB) pair by reversing the IP paths in Fig. 9. Since the IPB drives the UDB in (IPB, UDB), we position all the IP paths to align the longest IP paths ending at every output port of the IPB with the left end of the critical IP path. Then we select IP paths whose left ends extend beyond the specified threshold. The path selection scheme for (UDB, IPB) can also be easily extended to the case of (UDB1, IPB, UDB2), where an IPB is surrounded by two UDBs. In this case, we assume that all paths within the UDBs have their maximum allowable delays. We position all the IP paths such that the longest IP paths having the same I/O terminals are aligned with the right or left end of the critical IP path. Then we select IP paths that exceed the specified threshold. Fig. 9 : (a) IP path selection using the method in STSTEST [12]; (b) all the complete paths corresponding to the IP paths in (a), which also exceed the complete path threshold delay. (a) I 1 I 2 I 4 n1-n10-n22 n3-n11-n16-n22 Slack limits (b) Complete path threshold delay Clock period for (UDB, IPB) for UD paths I 3 n6-n11-n16-n22 Critical IP path delay IP path threshold delay n3-n11-n16-n22 I 1 I 2 Clock period for (UDB, IPB) Complete path threshold delay I 3 n6-n11-n16-n22 UD paths IP paths n3-n11-n16-n22 n6-n11-n16-n22 I 3 Circuit Partitioning The fact that BDD size can explode when handling large circuits limits the applications of many BDD-based methods [2] to control circuits or relatively small datapath circuits. In order to enable SPMTEST to handle a broad range of IP circuits, we use circuit partitioning to reduce the BDDs to a manageable size. Functional decomposition techniques that reduce BDD size have been previously proposed for formal design verification [15]. Here we use a structural BDD decomposition technique that partitions an IP circuit into a set of blocks and constructs BDDs for the partitioned blocks. This approach has the advantage of reducing the number of paths that must be included in the IP models, since a few partitioned paths often cover a large number of paths. We can also easily adapt existing structural ATPG algorithms to deal with partitioned BDDs. Symbolic models for the partitioned IP paths may not identify all untestable IP paths a priori. To alleviate this drawback, we propose an algorithm dubbed SPM-PART that maximizes the chance of untestable paths being identified by exploiting a property of untestable paths. A untestable path P contains a fanout point F i and a reconvergence point R j that conflict with the robust test conditions for P. Symbolic path modeling is guaranteed to identify P as untestable, if for every on P, all paths linking F i and R j are in the same partition. SPM-PART partitions an IP circuit in a way that maximizes the number of and the paths linking F i and R j contained in the same partition, while limiting the partition to a specified size. We describe SPM-PART below. Let N i,j be the number of fanin lines to R j that have a path from F i , and let D i,j be the distance between F i and R j in terms of the number of levels. SPM-PART first computes for every pair defined as N i,j / D i,j . Combining an of large G(F i , R j ) can lead to a small partition (due to a small D ij ) that contains a large number of paths linking F i and R j . SPM-PART creates each partition B k by adding such one at a time. It first selects an F i from the primary inputs of IPB and inserts F i into the current partition B k . SPM-PART then selects an R j that maximizes the sum of the G(F i , R j )'s for all the F i 's in B k that have a path to R j . It inserts into B k all non-partitioned gates in the transitive fanin region of R j to maximize the number of paths in B k linking the F i 's and R j . If the current partition exceeds a specified size, B k+1 is set to the current partition. In this way, SPM-PART continues to insert the next R j 's into B k , until no R j remains. The complexity of SPM-PART including the gain factor computation is O(N 2 ). Figure 10(a) shows a 2-bit adder viewed as an IPB and partitioned into two blocks by SPM-PART. Figure 10(b) shows a graph whose nodes represent the F i 's and R j 's in IPB and whose edges represent the Figure 10(c) lists the gain factors computed for every limits of each partition in terms of I/O line numbers is set to 3/2. With a 0 as the first F i inserted in B1, we select p 0 as R i , since p 0 yields the largest gain factor G(a 0 , partition B1 indicated by A is formed by including p 0 and its transitive fanin node b 0 . Next s 0 is selected which yields the largest gain factor G(a 0 , s 0 )+G(b 0 , and by including s 0 and c in , B1 now becomes the partition indicated by B in Fig. 10(b). After including c 1 , B1 exceeds the size limit, so the next nodes are added to B2. Figure 10(a) indicates by dashed lines the final two partitions created in this way. Observe that in Fig. 10(a), most like the ones marked by X and the paths linking F i and R j are contained in the same partition. In this example, the symbolic path modeling can identify all the untestable paths such as the ones highlighted. Figure 11(a) shows the same circuit but partitioned arbitrarily without using SPM-PART. In this case, symbolic path modeling cannot identify any untestable a 0 c in a 1 c out A Fig. partition produced by the proposed algorithm; (b) fanout-reconvergence graph and partition steps; (c) gain values computed for every a 0 c in a 1 c out (a) (b) (c) a 0 c in a 0 . s 0 c out (a) (b) Fig. Arbitrary partition of the 2-bit adder obtained without using our algorithm; (b) comparison of the partitions. Partition type No. of paths in IP model No. of untestable paths identified BDD size of IP model partitioning Partition of Fig. 10(a) Partition of Fig. 11(a) paths, because for all the that make the paths untestable, F i and R j are in different partitions. Figure 11(b) compares the unpartitioned IPB with the partitions of Fig. 10(a) and 11(a). Note that the partition of Fig. 10(a) allows all untestable paths to be identified, so the test generation procedure needs to be run for only 74 testable paths. On the other hand, the partition of Fig. 11(a) requires the test generation procedure to be run for all 90 paths in the IPB. Although the BDD size reduction looks minor in this small example, very significant reductions are obtained for larger circuits, as our experiments show. Test Generation Assuming that the IP models for all IPBs are constructed by the method described above, system designers can generate test vectors using the ATPG procedure SPMTEST (Fig. 12) which is an extension of the PODEM algorithm to handle BDDs in a block framework. SPMTEST takes symbolic models of IPBs as inputs, and creates symbolic models for UDBs. For example, Fig. 13 shows the (UDB, IPB) pair of Fig. 8 where each block is treated as a black box specified by its symbolic model. Let P Bi denote a partial path derived from (partitioned) block B i . SPMTEST selects a complete path P B1 -P B2 -P Bn that exceeds the complete path threshold delay derived by the method describe in Sec. 2. For example, Fig. 13 shows one such complete path P UD1 -P IP1 . To speed up the test generation, SPMTEST simplifies the BDDs of the symbolic models by setting v primary input, and v (Note that enhanced scan can assign stable values to the primary inputs.) For example, Fig. 13 shows such values assigned for the case of P UD1 -P IP1 . Figures 14(a) and (b) show IPB's BDDs Symbolic IP models Create symbolic models Select a complete path P B1 -P B2 -P Bn Simplify BDDs by assigning initial input values BDD-based PODEM algorithm Select objective: one input variable of P Bi Backtrace to a target primary input and assign a value Evaluate BDDs using cofactor-based implication Backtrack, if any R(P Bi Repeat until - i =1:N R(P Bi Repeat until all target complete paths are tested 2-pattern test cubes Fig. test generation algorithm UDBs from IP providers substantially simplified with these assigned values; compare with the original BDDs in Figs. 6(a) and 7, respectively. Next the BDD-based PODEM algorithm attempts to satisfy the condition - i=1:N R(P Bi which is represented by conjunction of the BDDs for all P Bi 's symbolic path models. For each R(P Bi ), it first selects as an objective support variable s i from R(P Bi )'s maximal cube. The backtrace step then finds a primary input as follows: for the output function f i corresponding to s i , select f i 's support variable s i from f i 's maximal cube; repeat this step until s i is a primary input. In the example of Fig. 13, first we select n6 as an objective from R(P IP1 )'s maximal cube n3 f n3 s n6 f next we select g0 cube stop backtracing, since g0 f is a primary input. The next phase of SPMTEST is a ternary implication step that evaluates all the BDDs with their variables assigned the values 0, 1, and X. (Note that the initial values of all the BDD variables are X.) We implement ternary implication by computing the cofactors of a BDD with respect to its non-X input values. If a resulting cofactor is constant 1 (0), the BDD evaluates to 1 (0); otherwise, the result is X. For example, given input values n3 f n3 s n6 f 1X0X, the cofactor of R(P IP1 ) with respect to n3 f n6 shown in Fig. 14(c). Since this cofactor is not constant, R(P IP1 ) is found to be X. These steps are repeated until - i=1:N R(P Bi test cube is obtained. For the example of Fig. 13, the test cube Fig. 13 : The (UDB, IPB) pair of Fig. 8 represented by black boxes. F F Simplified BDDs for (a) the output functions of IPB and (b) the symbolic path model of P IP1 in Fig. 13; (c) the cofactor of (b) with respect to n3 f n6 Cofactor of (c) (b) (a) For each complete path, SPMTEST either computes a robust test or concludes that the path is robustly untestable. Since it acts on a structure consisting of multiple (partitioned) blocks in symbolic form, SPMTEST can handle any combination of IPBs and UDBs without needing extra scan registers. 5 Experimental Results We have implemented the SPMTEST method in a CAD tool composed of 17,000 lines of C++ code and an existing BDD package CUBDD [19]. We have applied it to a number of benchmark circuits, including ISCAS-85, ISCAS-89 (combinational versions), and datapath circuits, that have been artificially paired as UD and IP blocks. Figure 15 compares the symbolic IP models constructed by SPMTEST with and without circuit partitioning. The first column lists the benchmark circuits regarded as IPBs, while the next three columns give the circuit partitioning results. For the specified limits on the number of I/O lines of each partition, the number of resulting partitioned blocks and the CPU time spent for partitioning are listed. Next, the results of symbolic IP modeling using the partitioned IP circuits are listed for the specified IP path delay threshold factor T IP . The untestable path identification ratio UPI is given for circuits that have a large number of untestable paths. UPI is defined as the number of untestable IP paths identified by the IP models divided by the number of all untestable IP paths. Then the BDD size in terms of the number of Fig. modeling with and without circuit partitioning. Bench- mark circuit Circuit partitioning for symbolic IP modeling Symbolic IP modeling With partitioning Without partitioning I/O limits No. of partitions CPU time Untestable paths identification (UPI) BDD size CPU cs1423 20/15 14 1.43 0.8 70.8% 10425 4.8 541812 ~2 hours hours Exploded >12 hours c2670 25/25 17 2.57 0.8 93.7% 44425 29.69 Exploded >12 hours c3540 25/25 26 14.67 0.9 72.9% 68502 62.86 Exploded >12 hours Exploded >12 hours c7552 25/25 38 20.61 0.8 56.1% 164055 171.2 Exploded >12 hours nodes, and the CPU time spent for IP modeling are listed. The last two columns list the results of symbolic IP modeling without partitioning, in which case UPI is 100%. In all cases, IP modeling with partitioning finishes within reasonable CPU time with relatively small BDDs. For example, modeling the largest ISCAS-85 circuit c7552 is completed in 171 seconds with BDDs containing a total of 164K nodes. On the other hand, IP modeling for most large circuits without partitioning either takes several hours or cannot finish due to the excessive BDD size. It is well known that BDDs representing larger ISCAS-85 benchmark circuits such as c2670 and c7552 tend to explode, so BDD-based methods like that of [2] have not been applied to these circuits. Furthermore, IP modeling with partitioning can identify a large number of the untestable paths in most circuits. For example, cs9234's IP model identifies 99.9% of untestable paths with 20 partitioned blocks, which indicates that the proposed partition algorithm is highly efficient. Some low UPI ratios for circuits like cs1196 can be explained by their structural property that the separation of most fanout-reconvergence pairs is very large, so it is difficult to contain such pairs within the same partition. Figure gives the results of applying SPMTEST to a number of (UDB, IPB) pairs whose IP models appear in Fig. 15. Although we limit our attention to (UDB, IPB) pairs, other combinations of IPBs and UDBs show similar results. The first two columns list the (UDB, IPB) pairs tested. The next two columns Circuit pair UDB partition results UD/IP pair test results UDB IPB I/O limits of each partition No. of partitions Complete path threshold T C No. of complete paths Tried for test generation Robustly tested 28 19.22 cs1238 shift32 20/15 23 0.801 343 150 91.92 shif32 cs1238 30/30 8 0.860* 347 shift32 shift16 c5315 25/25 2 0.839 4118 558 1491.3 shift32 Fig. test generation results for benchmark circuits configured as (UDB, IPB) pairs. show the symbolic modeling results for the UDBs. The next column lists the complete path threshold delay factor T C for each (IPB, UDB) pair. Given T IP and the clock period T clock , T C is determined by T D IP (1- T IP ) / T clock , where D IP is IPB's critical path delay; see Fig. 9. All testable complete paths exceeding the threshold delay T C - T clock are guaranteed to be tested. In the cases indicated by * in column T C , we have chosen values of T C smaller than the values calculated in the above way, because either the (IPB, UDB) pairs have too many complete paths, or T IP = 0. The next column lists the number of complete paths tried for test generation and the number of complete paths robustly tested. In most cases, the tried complete paths are much fewer in number than all complete paths meeting T C , because many untestable paths are eliminated a priori in the IP modeling step, which speeds up test generation. The fact that only a few complete paths are robustly tested in many cases is not surprising, because the artificial (functionally meaning- less) connections between UDBs and IPBs tend to make a large number of complete paths untestable. The time listed in the last column of Fig. 16 is reasonable for most cases except the circuit pair containing c2670. This large ISCAS-85 benchmark circuit is well known to have very few robustly testable paths due to its large amount of reconvergent fanout, and so path delay testing for it is inherently very difficult. In STSTEST [12], some untestable complete paths are robustly testable due to the STS registers, whereas in SPMTEST, only robustly testable complete paths are considered as testable and counted in Fig. 16. Therefore, the results of SPMTEST cannot be directly compared with those of STSTEST. Comparison with other methods is also difficult, since most delay testing methods are not aimed at IP-based designs; in the case of [16], no experimental results are provided. 6 Conclusions We have presented the SPMTEST method for path delay testing of designs containing IP cores, a difficult problem not addressed by existing methods. SPMTEST can test complete paths linking IP and user-defined blocks via a symbolic modeling technique that abstracts an IP block's paths in a compact form. Hence it does not require extra scan logic, an advantage over STSTEST. The ATPG algorithm in SPMTEST generates tests for the complete paths using only symbolic models, and hence protects the implementation details of the IP blocks. Our experimental results show that for all the benchmark circuits chosen, SPMTEST constructs compact symbolic IP models, and robustly tests all testable complete paths of a specified delay range. Therefore SPMTEST appears to an ideal approach to path delay testing of IP-based designs. SPMTEST has a limitation that some complex circuits such as multipliers can require IP models of excessive size. To address this problem, we are investigating alternative symbolic modeling approaches. --R "Scan Chain Design for Test Time Reduction in Core-Based ICs," "Test Generation for Path Delay Faults Using Binary Decision Diagrams," "Hierarchical Test Access Architecture for Embedded Cores in an Integrated Circuit," "On Variable Clock Methods for Path Delay Testing of Sequential Circuits," "A Satisfiability-Based Test Generation for Path Delay Faults in Combinational Circuits," "Robust Delay-Fault Test Generation and Synthesis for Testability Under a Standard Scan Design Methodology," "A Partial Enhanced-Scan Approach to Robust Delay-Fault Test Generation for Sequential Circuits," "Test Methodology for Embedded Cores which Protects Intellectual Property," "Design for Testability: Using Scanpath Techniques for Path-Delay Test and Measurement," "Synthesis of Robust Delay-Fault-Testable Circuits: Theory," "High-Coverage ATPG for Datapath Circuits with Unimplemented Blocks," "Delay Fault Testing of Designs with Embedded IP Cores," "On Delay Fault Testing in Logic Circuits," "Testing ICs: Getting to the Core of the Problem," "Partitioned ROBDDs-A Compact, Canonical and Efficiently Manipulable Representation for Boolean Functions," "Path Delay Fault Testing of ICs with Embedded Intellectual Property Blocks," "NEST: A Nonenumerative Test Generation Method for Path Delay Faults in Combinational Circuits," "Advanced Automatic Test Pattern Generation Techniques for Path Delay Faults," CUDD: CU Decision Diagram Package Synopsys Inc. "Testing Embedded Cores Using Partial Isolation Rings," A Path-Delay Test Generator for Standard Scan Designs," "Hierarchical Timing Analysis Using Conditional Delays," --TR Robust delay-fault test generation and synthesis for testability under a standard scan design methodology Hierarchical timing analysis using conditional delays A satisfiability-based test generator for path delay faults in combinational circuits Partitioned ROBDDsMYAMPERSANDmdash;a compact, canonical and efficiently manipulable representation for Boolean functions Path delay fault testing of ICs with embedded intellectual property blocks Logic Synthesis and Verification Algorithms Testing ICs Test Generation for Path Delay Faults Using Binary Decision Diagrams Fastpath A Partial Enhanced-Scan Approach to Robust Delay-Fault Test Generation for Sequential Circuits Design for Testability Scan chain design for test time reduction in core-based ICs Testing embedded-core based system chips High-coverage ATPG for datapath circuits with unimplemented blocks Testing Embedded Cores Using Partial Isolation Rings 1.1 Test methodology for embedded cores which protects intellectual property Delay Fault Testing of Designs with Embedded IP Cores

automatic test pattern generation ATPG;binary decision diagram BDD decomposition;system-on-a-chip SOC;delay fault testing;intellectual property IP core testing

505588

An algebraic approach to IP traceback.

We present a new solution to the problem of determining the path a packet traversed over the Internet (called the traceback problem) during a denial-of-service attack. This article reframes the traceback problem as a polynomial reconstruction problem and uses algebraic techniques from coding theory and learning theory to provide robust methods of transmission and reconstruction.

Introduction A denial of service attack is designed to prevent legitimate access to a resource. In the context of the Internet, an attacker can "flood" a victim's connection with random packets to prevent legitimate packets from getting through. These Internet denial of service attacks have become more prevalent recently due to their near untraceability and relative ease of execution [9]. Also, the availability of tools such as Stacheldraht [11] and TFN [12] greatly simplify the task of coordinating hundreds or even thousands of compromised hosts to attack a single target. These attacks are so difficult to trace because the only hint a victim has as to the source of a given packet is the source address, which can be easily forged. Although ingress filtering can help by preventing a packet from leaving a border network without a source address from the border network [14], attackers have countered by choosing legitimate border network addresses at random. The traceback problem is also difficult because many attacks are launched from compromised systems, so finding the source of the attacker's packets may not lead to the attacker. Disregarding the problem of finding the person responsible for the attack, if a victim was able to determine the path of the attacking packets in near real-time, it would be much easier to quickly stop the attack. Even finding out partial path information would be useful because attacks could be throttled at far routers. This paper presents a new scheme for providing this traceback data by having routers embed information randomly into packets. This is similar to the technique used by Savage, et al [24], with the major difference being that our schemes are based on algebraic techniques. This has the advantage of providing a scheme that offers more flexibility in design and more powerful techniques that can be used to filter out attacker generated noise and separate multiple paths. Our schemes share similar backwards compatibility and incremental deployment properties to the previous work. More specifically, our scheme encodes path information as points on polynomials. We then use algebraic methods from coding theory to reconstruct these polynomials at the victim. This appears to be a powerful new approach to the IP traceback problem. We note that although the study of traceback mechanisms was motivated by denial of service attacks, there are other applications as well. These methods might be useful for the analysis of legitimate traffic in a network. For example, congestion control, robust routing algorithms, or dynamic network reconfiguration might benefit from real-time traceback mechanisms. The rest of the paper is organized as follows: Section 2 discusses related work, Section 3 contains an overview of the problem and our assumptions, Section 4 presents our approach for algebraically coding paths, Section 5 gives detailed specifications for some of our schemes, Section 6 provides a mathematical analysis of the victim's re-construction task, Section 7 discusses the issue of encoding marking data in IP packets, and 8 gives conclusions and future work. Related Work The idea of randomly encoding traceback data in IP packets was first presented by Savage, et al [24]. They proposed a scheme in which adjacent routers would randomly insert adjacent edge information into the ID field of packets. Their key insight was that traceback data could be spread across multiple packets because a large number of packets was expected. They also include a distance field which allows a victim to determine the distance that a particular edge is from the host. This prevents spoofing of edges from closer than the nearest attacker. The biggest disadvantage of this scheme is the combinatorial explosion during the edge identification step and the few feasible parameterizations. The work of Song and Perrig provides a more in depth analysis of this scheme [25]. There have been many other notable proposals for IP traceback since the original proposal. Bellovin has proposed having routers create additional ICMP packets with traceback information at random and a public key infrastructure to verify the source of these packets [4]. This scheme can also be used in a non-authenticated mode, although the attackers can easily forge parts of routes that are farther from the victim than the closest of the attackers. Song and Perrig have an improved packet marking scheme that copes with multiple attackers [25]. Unfortunately, this scheme requires that all victims have a current map of all upstream routers to all attackers (although Song and Perrig describe how such maps can be maintained). Additionally, it is not incrementally deployable as it requires all routers on the attack path to participate (although Song and Perrig note that it also suffices for the upstream map to indicate which routers are participating). Doeppner, Klein, and Koyfman proposed adding traceback information to an IP option [13]. Besides the large space overhead, this solution would cause serious problems with current routers, as they are unable to process IP packets with options in hardware. R 4 R 5 R 6 R 7 A 3 A 4 Figure 1: Our example network. It also causes others issues, for example, adding the option may require the packet to be fragmented. Burch and Cheswick have a scheme that uses UDP packets and does not require the participation of intermediate ISPs [8]. This scheme, however, assumes that the denial of service attack is coming from a single source network. This differs from us as we aim to distinguish multiple attacking hosts. Lee and Park have analyzed packet marking schemes in general [19]. Their paper contains general tradeoffs between marking probability, recovered path length, and packets received, that can be applied to any of the probabilistic marking schemes, including the one in this paper. We refer the reader to Savage's paper for a discussion of other methods to detect and prevent IP spoofing and denial of service attacks. The algebraic techniques we apply were originally developed for the fields of coding theory [15] and machine learning [2]. For an overview of algebraic coding theory, we refer the reader to the survey by Sudan [27] or the book by Berlekamp [6]. Overview This paper addresses what Savage, et al call the approximate traceback problem. That is, we would like to recover all paths from attacker to victim, but we will allow for paths to have invalid prefixes. For example, for the network shown in Figure 1, the true path from the attacker A 1 to the victim V is R 4 R 2 R 1 . We will allow our technique to also produce paths of the form R 2 R 6 R 4 R 2 R 1 because the true path is a suffix of the recovered path. Our family of algebraic schemes was motivated by the same assumptions as used in previous work: 1. Attackers are able to send any packet 2. Multiple attackers can act together 3. Attackers are aware of the traceback scheme 4. Attackers must send at least thousands of packets 5. Routes between hosts are in general stable, but packets can be reordered or lost 6. Routers can not do much per-packet computation 7. Routers are not compromised, but not all routers have to participate Algebraic Coding of Paths We will now present a series of schemes that use an algebraic approach for encoding traceback information. All of these schemes are based on the principal of reconstructing a polynomial in a prime field. The basic idea is that for any polynomial f (x) of degree d in the prime field GF(p), we can recover f (x) given f (x) evaluated at (d +1) unique points. Let A 1 ; A be the 32-bit IP addresses of the routers on path P. We associate a packet id x j with the jth packet. We then somehow evaluate f P as the packet travels along the path, accumulating the result of the computation in a running total along the way. When enough packets from the same path reach the destination, then f P can be reconstructed by interpolation. The interpolation calculation might be a simple set of linear equa- tions, if all of the packets received at the destination traveled the same path. Otherwise, we will need to employ more sophisticated interpolation strategies that succeed even in the presence of incorrect data or data from multiple paths [5, 28, 15, 7]. These methods were developed originally for use in coding theory and learning theory. A naive way to evaluate f P (w) would be to have the jth router add A j w n j into an accumulator that kept the running total. Unfortunately, this would require that each router know its position in the path and the total length of the path. We could eliminate the need for each router to know the total length of the path (while still requiring each router to know its position in the path) by reordering the coefficients of f However, we can do even better by sticking with our original ordering, and using an alternative means of computing the polynomial. Specifically, to compute f P (w), each router R j multiplies the amount in the accumulator by w, adds returns the result to the accumulator, and passes the packet on to the next router in the path (Horner's rule [18]). For example, ((((0 w)+R 1 )w+R 2 )w+R 3 )w+R Notice that the router doesn't need to know the total length of the path or its position in the path for this computation of f P . 4.1 Deterministic Path Encoding The simplest scheme that uses this algebraic technique encodes an entire path. At the beginning of a path, let FullPath 0; Each router i on the path calculates FullPath is a random value passed in each packet, R i is the router's IP address and p is the smallest prime larger than 2 32 1. The value FullPath passed in the packet, along with x j , to the next router. At the packet's destination FullPath will equal (R n x which can be reconstructed by solving the following matrix equation over GF(p):B B B @ FullPath n;1 FullPath n;2 FullPath n;3C C C A As long as all of the x i 's are distinct, the matrix is a Vandermonde matrix (and thus has full rank) and is solvable in O(n 2 ) field operations [22]. Assuming that we get a unique x j in each packet, we can recover a path of length d with only d packets. The downside, however, is that this scheme would require log bits per packet (the first term is the encoding of the running FullPath and the second term is the encoding of the x j and y j values). Even for modest maximum path lengths of 16, the space required (68 bits, counting 4 bits for recording the number of routers in the path, and 32 bits each for the x coordinate and y coordinate of the point on the polynomial) far exceeds the number of bits available to us in an IP header. We could split a router's IP address into c chunks and add dlog 2 (c)e bits to indicate which chunk was represented in a given packet. Another approach would be to have each router add all of its chunks into each packet. That is, each router would update FullPath c times, substituting each chunk of their IP address in order. The destination could then trivially reconstruct the IP addresses by interpolating to recover - R 1;1 +R 1;2 are the successive chunks of R j . This would increase the degree of f by a factor of c, which would impact the performance of the reconstruction algorithm. 4.2 Randomized Path Encoding In the above schemes, we require FullPath 0; This implies that there is some way for a router to know that it is the "first" participating router on a particular path. In the current Internet architecture there is no reliable way for a router to have this information. We must therefore extend our scheme to mitigate this problem. In our revised scheme a router first flips a weighted coin. If it came up tails the router would assume it was not the first router and simply follow the FullPath algorithm presented above, adding its IP address (or IP address chunk) data. On the other hand, if the coin came up heads, the router would assume it was the first router and randomly choose a x j to use for the path. We will refer to this state as "marking mode." This overall approach - which might be called the "reset paradigm" - was also used by Savage et al. for their traceback solutions. At the destination, we would receive a number of different polynomials, all representing suffixes of the full path. In our example network, packets from A 1 could contain R We could change our marking strategy slightly. Whenever a router receives a packet, it still flips a weighted coin. But now, instead of simply going into marking mode for one packet when the coin comes up heads, the router could stay in marking mode for the next t packets it receives. More generally, the reset behavior could follow any Markov Process. One problem is that attackers can cause more false paths than true paths to be received at the victim. This is due to the fact the our choice of a small p creates large number of packets in which no router on the packet's path is in marking mode. The attacker can thus insert any path information he wishes into such packets. Because the attacker can generally find out the path to his victim (using traceroute, for example) he can compute FullPath 0; This choice will cause the victim to receive FullPath When trying to reconstruct paths, the victim will have no indication as to which paths are real and which paths are faked. Two solutions to this problem are to increase p or to store a hop count (distance field) in the packet that each participating router would increment. Increasing the probability makes it even harder to receive long paths. Adding a hop count would prevent an attacker from forging paths (or suffixes of paths) that are closer than its actual distance from the victim but would require dlog 2 (d)e more bits in the packet. Our schemes could also make use of the HMAC techniques discussed by Song and Perrig to ensure that edges are not faked, but this would require us to either use additional space in the packets to store the hash or lose our incremental deployment properties [25]. If we decided to make one of these tradeoffs, our scheme would be comparably secure against multiple attackers. 4.3 Edge Encoding We could add another parameter, ', that represents the maximum length of an encoded path. The value of ' is set by the marking router and decremented by each participating router who adds in their IP information. When the value reaches 0, no more routers add in their information. For example, in the full path encoding scheme encoding of edges between routers. When we call this an "algebraic edge encoding" scheme. The benefit of this change would be to decrease the maximum degree d of the polynomials in order to reduce the number of packets needed out of a given set or packets to recover a route. The cost of this change is that it would add dlog bits to the packets. Of course, if ' is less than the true path length, then reconstruction finds arbitrary subsequences of the path (not just suffixes as in Full Path encoding). The victim still has some work to do to combine these subsequences properly (as described in Savage et al. Thus reconstruction in this scheme has an algebraic step followed by a combinatorial step. 5 Pseudocode for Sample Algebraic Schemes In this section, we present pseudocode for some sample algebraic marking schemes that are based on the principles described in the previous section. Recall that each router has a unique 32-bit id. 5.1 Algebraic Edge Encoding Here is the router's pseudocode for Edge1, an algebraic edge encoding scheme. Each packet is marked with is the number of x values. The degree of the polynomial is one. Marking procedure at router R: for each packet w with probability p w.xval := random; w.yval := R; w.flag := otherwise if w.flag w.yval := w.yval * w.xval w.flag := 0 Here is Edge2, algebraic edge encoding with c "chunks" per "hop". Each packet is marked with d32=ce+ dlogne+1 bits. The degree of the polynomial is 2c 1. Marking procedure at router R: for each packet w with probability p w.xval := random; w.yval := R[c] w.xval-{c-1} w.flag := otherwise if w.flag w.yval := w.yval * w.xval-c w.flag := 0 Here is Edge3, which is identical to Edge2 except that each packet also has a distance field ("hop count"). Following Savage et al., we reserve five bits for the distance field. Each packet is marked with d32=ce+ dlogne+6 bits. The degree of the polynomial is 2c 1. Marking procedure at router R: for each packet w with probability p w.xval := random; w.yval := R[c] w.xval-{c-1} w.flag := w.dist := 0; otherwise if w.flag w.yval := w.yval * w.xval-c w.flag := 0; w.dist Here is Edge4, which is identical to Edge3 except that the second router only contributes half of the bits of its router id. This lowers the degree of the polynomial, and introduces a little uncertainty into the reconstruction process (if two routers at the same distance from the victim had router id's that agreed on all of the contributed bits). Each packet is marked with d32=ce+ dlogne +6 bits. The degree of the polynomial is 1:5c 1. Marking procedure at router R: for each packet w with probability p w.xval := random; w.yval := R[c] w.xval-{c-1} w.flag := w.dist := 0; otherwise if w.flag w.yval := w.yval * w.xval-{c/2} w.flag := 0; w.dist 5.2 Algebraic Full Path Encoding Here is the router's pseudocode for Full1, the full path encoding scheme. Each packet is marked with 32+dlogne bits, where n is the number of possible x values. The degree of the path polynomial is at most L, the length of the path. Full1 Marking procedure at router R: for each packet w with probability p w.xval := random; w.yval := 0; w.yval := w.yval * w.xval Here is the router's pseudocode for Full2, the full path encoding scheme with a distance field ("hop count"). Following Savage, we reserve five bits for the distance field, so each packet is marked with 37+ dlogne bits. Full2 Marking procedure at router R: for each packet w with probability p w.xval := random; w.yval := 0; w.dist := 0; w.yval := w.yval * w.xval w.dist 6 Path Reconstruction by the Victim In this section, we look more closely at the problem of path reconstruction by the victim. Let k denote the number of attack paths. Let L denote the expected length of an attack path. For simplicity, we will assume that all attack paths are very close to L in length. For the main scheme of Savage et al. (which uses a total of 16 bits), the complexity of path reconstruction by the victim is O(Lk 8 ). The exponent of eight reflects a combinatorial task that the victim must try by brute force. Of course, if they had more room to work with in their marking scheme, then the reconstruction complexity would go down. For example, if they used 23 bits for their marking scheme (and divided the "padded" router id into four 16-bit chunks), then the victim's reconstruction task reduces to O(Lk 4 ). Our goal is to design algebraic schemes that improve on the reconstruction complexity of Savage et al. There are two main algebraic reconstruction approaches that we consider: Reed-Solomon List Decoding: Given polynomials of degree at most d that pass through at least m of these points. Guruswami-Sudan give an algorithm to solve this problem in time O(N 3 ) when N < m 2 =d. An improvement by Olshevsky and Shokrollahi reduces the time to O(N 2:5 ). More precisely, the reconstruction algorithm due to Guruswami and Sudan [15] can be implemented in a number of ways. The most straightforward implementation would take time O(n 3d ) to recover all edges for which we received at least dn out of packets. However, this drops to O(n 3 requiring only slightly more packets: dn(1+d) out of n, for any d 1. By scaling d appropriately, this allows us to trade off computation time (and memory) for accuracy. A recent algorithmic breakthrough by Olshevsky and Shokrollahi would reduce our reconstruction time even further, to O(n 2:5 ) [21]. Moreover, this new algorithm is highly parallelizable (to up to O(n) processors), which suggests that distributing the reconstruction task might speed things up even more. Noisy Polynomial Interploation: Given at most m, find all polynomials f of degree at most d such that f Bleichenbacher-Nguyen give an algorithm to solve this problem whenever m < n=d, with running time identical to the Reed-Solomon List Decoding problem. They give other algorithms that work even when the bound m < n=d is not met. Types of Packets: Let us assume that each packet that the victim receives is one of three possible types. A "true packet" contains a point on a polynomial that corresponds to a real attack path. A "bogus packet" contains a point created by an attacker outside the periphery, and never reset by any honest router along an attack path. A "stray packet" contains a point on a polynomial that corresponds to normal non-attack traffic. When a denial of service attack is underway, we assume that the fraction of stray packets is very small compared to true and bogus packets. False Positives: A "false positive" is a polynomial that is recovered by the reconstruction algorithm, but does not correspond to part of an actual attack path. For Reed-Solomon list decoding, the expected number of false positives in a random sample is about (N!=(m!(N m)!)) (1=q) m d 1 . For noisy polynomial interpolation, the expected number of false positives in a random sample is about m n =q n d 1 . For the main scheme of Savage et al., the expected number of false positives is about m 8 =2 When the marking scheme has no distance field, then we must also be concerned with "bogus edges" or "bogus paths" that the attacker can cause to appear in our sam- ple. We will consider this separately from the issue of false positives that arise at random. A moderate number of false positives is not a serious problem. Consider our marking scheme Edge3. The victim reconstructs a set of candidate edges for each distance. Each set of candidate edges includes true edges and "false positive" (but no "bogus edges" from the attacker assuming that no attacker is within this distance of the vic- tim). Now the victim attempts to assemble paths by connecting edges from distance ' with edges from distance ' + 1. There is certainly no problem unless the first endpoint of a false positive edge from some distance ' matches the second endpoint of a false positive or true edge from distance ' + 1. Let f be the expected number of false positives at each distance, and let k be the number of true edges at each distance. Then there are f expected false positives at distance expected false positives and true edges at distance ' + 1. Let M be the number of distinct router id's or partial router id's that are possible (e.g., 2 for Edge3). The probability of an accidental match is less than 1 ((M f )=M) f +k which is very close to 1 e f ( f +k)=M . (When f is close to this probability is unacceptably high.) The probability of an accidental match at any distance is less than is the length of the longest path. We now analyze the effectiveness of these approaches to path reconstruction. For each approach, the best known algorithms impose constraints on the design parameters for our marking schemes. It will be convenient for us to consider separately marking schemes that have a distance field and marking schemes that do not. 6.1 Reconstruction With a Distance Field When the marking scheme has a distance field, the task of the victim is simplified. The victim can select a sample of packets for which w.dist = ', for any given '. As long as no attacker is within distance ' of the victim, this sample will contain only true packets with points on polynomials that were last reset by routers at distance '. 6.1.1 Guruswami-Sudan Reconstruction With a Distance Field The path reconstruction problem faced by the victim can be viewed as a Reed-Solomon list decoding problem. The distinct points are chosen from a random sample of the distinct points in packets that reach the victim. The victim can filter out packets that were last reset at distance ', for every '. This simplifies the Reed-Solomon list decoding problem, by creating a smaller problem instance for each distance. We need nk packets from distance ' to have n distinct points from each of k polynomials. The victim collects the largest possible sample of distinct points from packets with w.dist = ' for every '. We need N < n 2 =d to reconstruct the polynomials using the Guruswami-Sudan algorithm. Lastly, we need for the efficiency of reconstruction to improve on Savage et al. This has at least a few solutions, but the improvements are not so compelling. For example, using Edge3 with three 11-bit chunks can be competitive with Savage et al. for certain values of k. 6.1.2 Bleichenbacher-Nguyen Reconstruction With a Distance Field The problem faced by the victim can be viewed as a noisy polynomial interpolation problem. The values x are all of the possible x values. Each set S i contains all of the distinct y values such that occurs in some packet within a random sample of all received packets. The polynomial f could be any of the polynomials that corresponds to a true attack path or a stray path. The victim could proceed as follows. He looks at a sample of N packets (for suitably large N), and for each x i he chooses a set S i of size m from all of the in the sample. If the number of distinct y values for which occurs in the sample is greater than m, then the victim chooses which m values to include in S i at random. The victim can filter out packets that were last reset at distance ', for every '. This creates a smaller problem instance for each distance. For each problem instance, the number of S i sets is equal to n, the number of possible x values. The size of each S i is k, the number of attack paths. The degree of each polynomial is at most d, which depends on the particular algebraic encoding method we are using. False Positives: There are k n ways of taking one x value from each set. Each of these will actually be a polynomial of degree d or less with probability at most 1=q n d 1 . Here q is the size of the finite field, which is essentially the number of distinct y values. We need the expected number of false positives k n =q n d 1 to be reasonably small. For the basic reconstruction algorithm of Bleichenbacher-Nguyen, we need k < n=d. Three other algorithms by Bleichenbacher-Nguyen work for many k;n;d even if they do not satisfy k < n=d. A "meet-in-the-middle" algorithm has running time (n d)m n=2 with precomputation that uses memory which is O(m n=4 logq). Note that this is independent of k. A "Grobner basis reduction" algorithm computes a Grobner basis reduction on a system of k polynomial equations in d +1 unknowns. The best known Grobner basis algorithms are super-exponential in d, but reasonably efficient for small d (e.g., d < 20). A "lattice basis reduction" algorithm performs a lattice basis reduction on an 1)-dimensional lattice in Z nk over a finite field of size about n. This method will be ineffective for our application because the size of the finite field is too small. For efficiency over Savage et al., we need (nk) 2:5 < k 8 . Here are some interesting instantiations of our schemes with respect to this method of reconstruction: Example 1: Edge3 encoding with 12 distinct x values (represented in a 4-bit xval field) and 8-bit yval field. Then the noisy polynomial problem has 12 S i sets, where the size of each S i is k, the number of attack paths. The degree of each polynomial is at most 7. The size of the finite field is 256. The meet-in-the-middle algorithm takes time 8k 6 , which compares favorably to the k 8 required by Savage et al. The Grobner basis reduction algorithm should also be reasonably efficient here. The total size of this marking scheme is bits. However, the number of false positives is unacceptably high here: k 12 =2 Example 2: Edge4 encoding with 12 distinct x values (in a 4-bit xval field) and 8- bit yval field. Then the degree of each polynomial is at most 5. The running time for the meet-in-the-middle algorithm is about 8k 6 . The running time for the Grobner basis reduction algorithm is faster than in the previous example. The expected number of false positives is lower than in the previous example: k 12 =2 48 . If we expect about one false positive at each distance. (Of course, the risk from false positives is slightly greater than in the previous case, because the number of possible partial router id's is only 2 22 . Thus there will be slightly more accidental matches of endpoints involving a bogus edge, but it is not significantly worse.) The total size of this marking scheme is bits. Example 3: Edge4 encoding with Then the degree of each polynomial is at most 4 (using 22 bits of second router id). Running time for meet-in-the-middle is about 6k 5 , versus k 8 for Savage et al. Number of false positives is about Savage. For example, this is expected positives for false positives for are quite manageable. The total size of this marking scheme is 21 bits. Example 4: Edge4 encoding with of each polynomial at most 4. Running time for meet-in-the-middle is about 4k 4 . Number of false positives is about k 8 =2 33 (e.g., 1=2 when The total size of this marking scheme is 20 bits. Example 5: Edge4 encoding with of each polynomial at most 4. Running time for meet-in-the-middle is about 8k 6 . Number of false positives is about k 12 =2 77 (e.g., 2 5 when both of which are quite manageable). The total size of this marking scheme is 21 bits. 6.2 Reconstruction Without a Distance Field For reconstruction when the marking scheme does not have a distance field, we do not achieve schemes that are competitive with Savage et al. Our analysis will begin with some facts and simplifying assumptions about the distribution of received packets by the victim. 6.2.1 Distribution of Received Packets be the fraction of packets arriving on the ith attack path that reach the victim as bogus packets. Let T i be the fraction of packets on the ith attack path that reach the victim as true packets. Let F i be the fraction of packets on the ith attack path that reach the victim as true packets that were only reset by the furthest router on that path. By Assumption For all of the encoding schemes (unless "marking mode" is used), we have F Viewed as a function of p over fraction takes on its maximum value at implies that F For all of the encoding schemes, we have When this implies that B . The fact that there can be such a large fraction of bogus packets arriving on each path has serious consequences for our marking schemes without a distance field. Let B;T;F be the fractions of bogus packets, true packets, and furthest packets for all paths to the victim. If we assume that the arrival rate of packets on all attack paths is approximately the same, then When "marking mode" is used, the probability that a router is not in reset mode is Coupon Collector's Bound: A sample of lC logC elements, drawn with replacement from according to the uniform distribution, is very likely to contain all C possible values, for some small constant l. 6.2.2 Guruswami-Sudan Reconstruction Without a Distance Field The victim can choose a random sample of distinct points in the packets that reach him. Without a distance field, he cannot partition the packets into smaller samples by last Assume that the routers are using an edge encoding scheme, and assume that we succeed if we can reconstruct all of the furthest edge polynomials. Let us also assume that we will search for polynomials that pass through n distinct points, where n is the number of distinct x values. There are actually three distinct levels of reconstruction success that can be con- sidered: (a) The sample of N points contains at least n points on every furthest edge polynomial with overwhelming probability; (b) The sample of N points contains at least n points on some furthest edge polynomial with overwhelming probability; (c) The sample of N points contains at least n points on some furthest edge polynomial with non-negligible probability q. For case (a), the Guruswami-Sudan algorithm needs to be applied only once to a random sample of N points. For case (b), the algorithm needs to be applied lklogk times to independent random samples of N points (coupon collector's problem on the set of k furthest edge polynomials). For case (c), the algorithm needs to be applied lklogk=q times to independent random samples of N points. For case (a), it suffices to have N (lnklog(nk))=(p(1 p) L 1 ). That is because we are very likely to get a complete set of all n possible x values for all k edge poly- nomials. This implies that lnklog(nk) "samples" are sufficient. By the analysis of the preceding subsection, a "sample" from a furthest edge polynomial is expected in fraction of all of the packets. When combined with the Guruswami-Sudan bound, we get n 2 =d > (lnklog(nk))=(p(1 p) L 1 ). Assuming that we have For case (b), it suffices to have N M n;k =(p(1 p) L 1 , where M n;k is the answer to the following "occupancy problem": Throw M n;k balls into k bins, and expect to find lnlogn balls in the bin with the most balls. By the Pigeonhole Principle, it is certainly true that M n;k < lnklogn. In fact, the actual value for M n;k is quite close to this. Combined with the Guruswami-Sudan bound, we get n 2 =d > lnklogn=(p(1 Assuming that ln=logn > dk(L 1)e: For case (c), we can reduce the value of M n;k a little, but it doesn't appear to be significant for our purposes. Of course, for any of (a), (b), or (c), we can reduce N by eliminating from the sample any duplicate points. Since as many as 1=e of all packets in the sample could be bogus packets from the attacker, removing duplicate points will have limited benefit. We can find no solution that yields a marking scheme that is more efficient than Savage et al. Moreover, for any plausible instantiation, the number of false positives and bogus edges (or bogus paths) is unacceptably high. 6.2.3 Bleichenbacher-Nguyen Reconstruction Without a Distance Field The victim can proceed as described at the start of Section 6.1.2, although without a distance field the packets cannot be partitioned by last reset distance. Version H. Length Type of Service (8-bit) Total Length Fragment ID (16-bit) Flags Fragment Offset Time to Live Protocol Header Checksum Source IP Address Destination IP Address Figure 2: The IP Header. Darkened areas represent underutilized bits. Suppose that N is chosen to be large enough that all of the points on some furthest polynomial are included with high probability. The probability that a given S i includes a point from f is at least m=n. The probability that reconstruction succeeds is at least (m=n) n . The basic algorithm of Bleichenbacher and Nguyen solves the noisy polynomial interpolation whenever m < n=k. This approach does not seem too promising when k > 1. In this case, (m=n) n < 2 n . Thus reconstruction is unlikely to succeed. the victim can choose positive integer c. The probability that reconstruction succeeds is at least (1 c=n) n which is about e c . Un- fortunately, either the number of false positives is unacceptably large, or the success probability is unacceptably small. Another approach would be to have the victim bias his sample with respect to how frequently different points occurred in the packets that reached him. Unfortunately, this does not appear to work well either. Since the victim will not be able to recognize the true packets that contain points from the furthest polynomials. We conclude that when the marking scheme does not have a distance field, we do not see how to use the Bleichenbacher-Nguyen method of polynomial reconstruction, at least using their simplest algorithm. It is possible that their other algorithms, e.g., based on Grobner basis reduction, might be more effective. 7 Encoding Path Data We now need a way to store our traceback data in IP packets. We will try to maximize the number of bits available to us while preserving (for the most part) backwards compatibility. 7.1 IP options An IP option seems like the most reasonable alternative for storing our path informa- tion. Unfortunately, most current routers are unable to handle packets with options in hardware [3]. Even if future routers had this ability, there are a number of problems associated with this approach as presented by Savage, et al [24]. For all of these reasons we have concluded that storing data in an IP option is not feasible. 7.2 Additional Packets Instead of trying to add our path data to the existing IP packets, we could instead send the data out of band using a new protocol that would encapsulate our data. While this may have limited uses for special cases (such as dealing with IP fragments), a general solution based on inserting additional packets requires a means of authenticating these packets. This is because, presumably, the number of inserted packets is many orders of magnitude less than the number of packets inserted by the attacker. Thus, because we assume that an attacker can insert any packet into the network, the victim can be deluged with fake traceback packets, preventing any information to be gained from the legitimate packets. 7.3 The IP Header Our last source of bits is the IP header. There are several fields in the header that may be exploited for bits, with varying tradeoffs. As shown in Figure 2, we have found 25 bits that might possibly be used.xs 7.3.1 The TOS Field The type of service field is an 8 bit field in the IP header that is currently used to allow hosts a way to give hints to routers as to what kind of route is important for particular packets (maximized throughput or minimized delay, for example) [1]. This field has been little used in the past, and, in some limited experiments, we have found that setting this field arbitrarily makes no measurable difference in packet delivery. There is a proposed Internet standard [20] that would change the TOS field to a "differentiated services field." Even the proposed DS field has unused bits, however, there are already other proposed uses for these bits (e.g. [23]). 7.3.2 The ID Field The ID field is a field used by IP to permit reconstruction of fragments. Naive tampering with this field breaks fragment reassembly. Since less than 0:25% of all Internet traffic is fragments [26], we think that overloading this field is appropriate. A more in-depth discussion of the issues related to its overloading can be found in Savage's work [24]. 7.3.3 The Unused Fragment Flag There is an unused bit in the fragment flags field that current Internet standards require to be zero. We have found that setting this bit to one has no effect on current imple- mentations, with the exception that when receiving the packet, some systems will think it is a fragment. The packet is still successfully delivered however, because it looks to those systems as though it is fragment 1 of 1. Our Selection We could choose to use up to 25 bits out of the ID, flag, and TOS fields. This would suffice for all of the examples given in Section 6.1.2. The implications of using multiple fields in the IP header simultaneously are modest, since the lost functionality appears to be the union of what would break due to overwriting each field separately. The impact on header checksum calculation is modest, as this can be done in hardware using the standard algorithm. Of course, the algebraic marking scheme is independent of the choice of bits. The decision of where to put the marking data must be seen as conditional, subject to change as new standards arise. 7.4 IPsec The interoperability of a traceback scheme with IPsec should be considered. The Encapsulated Payload (ESP) [17], which encrypts a datagram for confidentiality, provides no problem for a traceback scheme, as it does not assume anything about the surrounding datagram's headers. The Authentication Header (AH) [16], does present an issue for IPv4. Using the AH, the contents of the surrounding datagram's headers are hashed. Certain header fields are considered mutable (e.g., Fragment Offset), and not included in the hash computation. Unfortunately, the mutable fields in the IPv4 header are unusable for traceback: they either are necessary for basic IP functionality, or reusing them breaks backward compatibility with current IP implementations. 7.5 IPv6 Since IPv6 does not have nearly as many backwards compatibility issues as IPv4, the logical place to put traceback information is a hop-by-hop option in the IPv6 header [10]. However, schemes such as those presented here are still valuable because they use a fixed number of bits per packet thereby avoiding the generation of fragments. Unlike the case in IPv4, we can set the appropriate bit in the Option Type field to indicate that the data in the option is mutable, and should be treated as zero for the purposes of the Authentication Header. We have not worked out the best way to accommodate IPv6's 128-bit addresses, but note that due to alignment issues, one is likely to select an option length of 8n+6 bytes, n 0. It would likely be the case that 0 8 Conclusion and Future Work We have presented a new algebraic approach for providing traceback information in IP packets. Our approach is based on mathematical techniques that were first developed for problems related to error correcting codes and machine learning. Though we have proposed it in the context of a probabilistic packet marking scheme, our algebraic approach could also be applied to an out-of-packet scheme. The resulting scheme would have the desirable property of allowing multiple routers to act on the extra packet while it remains at a small constant size. Our marking schemes have applications for other network management scenarios besides defense against denial of service. One important open problem is to find better instantiation of the specific methods we have proposed. In particular, a successful approach based on full path tracing would be attractive. More generally, it would be interesting to explore resource and security tradeoffs for the many parametrizations of our schemata. Lower bounds on the size of any marking scheme would be most helpful. It would also be interesting to explore the use of algebraic geometric codes in marking schemes. Acknowledgments We would like to thank David Goldberg and Dan Boneh for valuable discussions. We would also like to thank Dawn Song, Adrian Perrig, Ramarathnam Venketesan, Glenn Durfee, and the anonymous referees for helpful comments on earlier versions of this paper. --R Type of service in the internet protocol suite. Reconstructing algebraic functions from mixed data. Personal Communications ICMP traceback messages. correction of algebraic block codes. Algebraic Coding Theory. Bleichenbacher and Nguyen. Tracing anonymous packets to their approximate source. CERT coordination center denial of service attacks. Internet protocol "stacheldraht" "Tribe Flood Network" Using router stamping to identify the source of IP packets. Network ingress filtering: Defeating denial of service attacks which employ IP source address spoofing. Improved decoding of Reed-Solomon and algebraic-geometric codes IP authentication header IP encapsulating security payload (ESP) The Art of Computer Programming On the effectiveness of probabilistic packet marking for ip traceback under denial of service attack. Definition of the Differentiated Services field (DS field) in the IPv4 and IPv6 headers. A displacement approach to efficient decoding of algebraic-geometric codes Numerical Recipes in FORTRAN: The Art of Scientific Computing. A proposal to add Explicit Congestion Notification (ECN) to IP. Practical network support for IP traceback. Advanced and authenticated marking schemes for IP traceback. Providing guaranteed services without per flow management. Algorithmic issues in coding theory. Decoding of Reed Solomon codes beyond the error-correction bound --TR Numerical recipes in FORTRAN (2nd ed.) The art of computer programming, volume 2 (3rd ed.) Decoding of Reed Solomon codes beyond the error-correction bound A displacement approach to efficient decoding of algebraic-geometric codes Providing guaranteed services without per flow management Practical network support for IP traceback Using router stamping to identify the source of IP packets Algorithmic Issues in Coding Theory --CTR Karthik Lakshminarayanan , Daniel Adkins , Adrian Perrig , Ion Stoica, Taming IP packet flooding attacks, ACM SIGCOMM Computer Communication Review, v.34 n.1, January 2004 Hikmat Farhat, Protecting TCP services from denial of service attacks, Proceedings of the 2006 SIGCOMM workshop on Large-scale attack defense, p.155-160, September 11-15, 2006, Pisa, Italy Florian P. Buchholz , Clay Shields, Providing process origin information to aid in computer forensic investigations, Journal of Computer Security, v.12 n.5, p.753-776, September 2004 David G. Andersen, Mayday: distributed filtering for internet services, Proceedings of the 4th conference on USENIX Symposium on Internet Technologies and Systems, p.3-3, March 26-28, 2003, Seattle, WA Katerina Argyraki , David R. Cheriton, Loose source routing as a mechanism for traffic policies, Proceedings of the ACM SIGCOMM workshop on Future directions in network architecture, August 30-30, 2004, Portland, Oregon, USA Haining Wang , Danlu Zhang , Kang G. Shin, Change-Point Monitoring for the Detection of DoS Attacks, IEEE Transactions on Dependable and Secure Computing, v.1 n.4, p.193-208, October 2004 Andrey Belenky , Nirwan Ansari, On deterministic packet marking, Computer Networks: The International Journal of Computer and Telecommunications Networking, v.51 n.10, p.2677-2700, July, 2007 Sherif Khattab , Rami Melhem , Daniel Moss , Taieb Znati, Honeypot back-propagation for mitigating spoofing distributed Denial-of-Service attacks, Journal of Parallel and Distributed Computing, v.66 n.9, p.1152-1164, September 2006 Brent Waters , Ari Juels , J. Alex Halderman , Edward W. Felten, New client puzzle outsourcing techniques for DoS resistance, Proceedings of the 11th ACM conference on Computer and communications security, October 25-29, 2004, Washington DC, USA Hassan Aljifri, IP Traceback: A New Denial-of-Service Deterrent?, IEEE Security and Privacy, v.1 n.3, p.24-31, May Zhiqiang Gao , Nirwan Ansari, A practical and robust inter-domain marking scheme for IP traceback, Computer Networks: The International Journal of Computer and Telecommunications Networking, v.51 n.3, p.732-750, February, 2007 Christos Siaterlis , Vasilis Maglaris, One step ahead to multisensor data fusion for DDoS detection, Journal of Computer Security, v.13 n.5, p.779-806, October 2005 Christos Douligeris , Aikaterini Mitrokotsa, DDoS attacks and defense mechanisms: classification and state-of-the-art, Computer Networks: The International Journal of Computer and Telecommunications Networking, v.44 n.5, p.643-666, 5 April 2004

traceback;internet protocol

505628

Extraction and Optimization of B-Spline PBD Templates for Recognition of Connected Handwritten Digit Strings.

Recognition of connected handwritten digit strings is a challenging task due mainly to two problems: poor character segmentation and unreliable isolated character recognition. In this paper, we first present a rational B-spline representation of digit templates based on Pixel-to-Boundary Distance (PBD) maps. We then present a neural network approach to extract B-spline PBD templates and an evolutionary algorithm to optimize these templates. In total, 1,000 templates (100 templates for each of 10 classes) were extracted from and optimized on 10,426 training samples from the NIST Special Database 3. By using these templates, a nearest neighbor classifier can successfully reject 90.7 percent of nondigit patterns while achieving a 96.4 percent correct classification of isolated test digits. When our classifier is applied to the recognition of 4,958 connected handwritten digit strings (4,555 2-digit, 355 3-digit, and 48 4-digit strings) from the NIST Special Database 3 with a dynamic programming approach, it has a correct classification rate of 82.4 percent with a rejection rate of as low as 0.85 percent. Our classifier compares favorably in terms of correct classification rate and robustness with other classifiers that are tested.

Introduction In the field of researching on automated handwritten document processing system, hand-written digit recognition still poses a very challenge to scientists and practitioners. The difficulties arise not only from the different ways in which the single digit can be written, but also from varying requirements imposed by the specific applications, such as recognition of connected digit strings in ZIP codes reading and cheque reading. The techniques for recognition of connected digit strings can be categorized into two classes: Segmentation-based algorithms [1], [2], [3], [4], [5], [6], and Segmentation-free algorithm [7], [8], [9], [10]. Recently, there are also some techniques reported to combined the two techniques together to achieve a more reliable performance [11], [12]. In the first class of algorithms, the segmentation procedures were applied before the recognition, and in the later class algorithms, they combined the segmentation and recognition together. The segmentation-based techniques were first applied in the recognition of connected char- acters, while in recent years, more attention have been focused on the segmentation-free techniques. The reason is that segmentation is ambiguous and prone to failure. Anther reason that researchers paying more attention to segmentation-free techniques, especially September 16, 1998 Fig. 1. Some samples selected randomly from NIST Special Database 3: (a), single digits; (b), connected digit strings: from left to right, two connected digit strings, three connected digit strings, four connected digit strings. in handwritten word recognition, is that they believe the lexical information being used in post-processing is powerful enough to achieve a reliable performance, and only need a not very robust but fast classifier [13]. However, in the recognition of connected digit strings, little lexical information can be applied in post-processing to help in choosing the correct result from a set of recognition candidates, as in the word recognition. And from experience gained from our previous research in segmentation-based handwritten digit string recognition [14], [5], [6], we believe that, for both techniques, a reliable and robust classifier that can distinguish legible digits from unsure patterns is utmost important to the system performance, and this is the motivation that we developed the new classifier, which will be introduced in this paper. Fig. 1 are some samples of single and connected digits. We can see that degradation caused by connection, overlap increase the difficult of the recognition. Many new techniques have been presented in recently years for the recognition of isolated digits, they are differ in the feature extraction and classification methods employed. Some comprehensive reviews were given by Govindan [15] and Impedovo [16]; S. Lee et al gave a classifier based on Radial Basis Network and Spatio-Temporal Features [17]; Z. Chi et al proposed a classifier with Combined ID3-Derived Fuzzy Rules and Markov Chains [18]; another classifier based on Multilayer Cluster Neural Network and Wavelet features was presented by S. W. Lee et al [19]; and D. H. Cheng et al presented a classifier based on morphology [20]. Besides these, Trier presented a survey of the feature extraction techniques [21]. Another technique applied in classifier designing is template matching, which is one of the earliest pattern recognition techniques applied in digit recognition. In these years, it re-attracted a lot of attention from many researchers because it is intuitive and maybe have the ability to achieve a practical recognition performance. A number of studies have been reported in the literature which have applied template-based techniques to digit recognition. H. Yan proposed an Optimized Nearest-Neighbor Classifier for the recognition of handprint digits [22], the templates are 8 \Theta 8 gray-scale images, rescaled from the original 160 \Theta 160 normalized binary image. Wakahara uses iterated local affine transformation (LAT) operations to deform binary images to match templates [23], the templates here are 32 \Theta 32 binary images. Nishida presented a structure-based template matching algorithm [24]. The templates are composite with straight lines, arcs and corners, and the matching is based on the geometrical deformation of the templates. Cheung et al proposed a kind of templates based on splines [25], he model digits with a spline, and assume the spline parameters have a multivariate Gaussian distributions. Revow et al gives another digit recognition algorithm based on spline templates [26], the templates are elastic spline curves with ink-generating Gaussian "beads" strung along their length. K. Jain et al presented a deformable templates matching algorithm based on object contours [27], [28], the recognition procedure is to minimize the objective function by iteratively updating the transformation parameters to alter the shape of the templates so that the best match between the templates and unknown digits. In this paper, we present a new template presentation scheme, a neural network approach 5for extracting templates and a evolutionary algorithm based templates optimization approach based on this new presentation. Each template presented by a distance- to-boundary distribution map is approximated by a rational B-spline surface with a set of control knots. In training, first a cost function of the amplitude and gradient of the distribution map is minimized by a neural network; then, the templates were optimized by a optimization algorithm based on evolutionary algorithm. While in the matching step, a similarity measure that takes into account of both the amplitude and gradient of the distribution map is adopted to match an input pattern to templates. The rest of this paper is organized as follows. The presentation of new templates is given in section II; section III introduced the neural network for templates extraction; the templates optimization part based on evolutionary algorithm is proposed in section IV; section V is the experimental results and discussions; and the conclusion is given in section VI. II. Presentation of Templates In our algorithm, the templates are presented by a rational B-spline surface with a set of control knots. Curve and surface presentation and manipulation using the B-spline form (non-rational or rational) are widely used in geometric design. In 1975, Versprille proposed a rational B-splines to geometric design of curves and surfaces [29]. The work outlined the important properties of rational B-splines, such as continuity, local control property, etc. In recent years, B-spline curve (snake) have been used in the description of objects [30], [31], as well as digits [25], [26]. With the B-spline estimation, information concerning the shape of desired objects can be incorporated into the control parameters of the curve or surface based templates. However, there is a shortcoming in these approaches: the estimation of B-spline curve exist a lot of uncertainties and prune to failure, because the corresponding parameters of templates and input data should reflect the information in the same aspect of image. So, in our algorithm, we use B-spline surface to estimate the distribution map of the digit image, The number of the control points of surface, and the position of them are pre-specified, so for each control point, the control area is certain, and the shortcoming can be avoided. O fg O bg O b Fig. 2. Definition of the distance-to-boundary distribution. Input Binary Data Distance Map Distribution Map Fig. 3. Converting binary data into a distribution map (one-dimensional example). A. Presentation of Binary Digit Images A binary digit image can be converted into a distribution map according to the distance from pixel O (x;y) to the boundary between the foreground and background, as shown in Fig. II-A. For a foreground pixel, we have d(x; y) 0, for a background pixel, we assume it has a negative value, that is d(x; \GammajD bg (1) September 16, 1998 Fig. 4. Control points of a template. Let I(x; y) denote the value of pixel O (x;y) in the distribution map, it can be defined via the following transformation: where d max is the maximum value in distribution map, and is a constant to make sure that a point on the boundary, O b with I(O b shown in Fig. 3. Obviously, the value in distribution map reflect the posterior possibility of pixels in input binary image belonging to foreground, and ridge of the map form the skeleton of digit. B. Template Presented by rational B-spline Surface Assume that the rational B-spline surfaces for the templates are fS(P j where are control points of template j, as shown in Figure II-B. The rational B-spline surface can be defined as: Normally, r is the order of the B-spline basis, and B r (x) is a vector of base function given by BN \Gamma1;r (x)C C C C C C C C A For each basis, we have and is a knot vector of length N u , It is assumed that The value of u i determine the position of control knots and should be pre- defined. Here, to simplify to the B-spline surface, we define a knot vector as f r z - r z - In most applications, the number of input data N i AE (N \Theta N ), the number of control points in a templates. III. Template Extraction The templates extraction algorithm is adapted from Yan's optimized nearest neighbor classifier [32], which can be presented as a multi-layer neural network, as shown in Figure 5. The input data is a distribution map I. The hidden layer contains the patterns to be selected. Each hidden node corresponds to a pattern. is the output of the network, corresponding to N classes. Input image Output layer Hidden Layer Input Layer Fig. 5. The template extraction model. The output of the hidden node is given by: OE j is a distance function that measure the similarity between the input feature I and pattern S(P j ). The connection weight from hidden node j to output node m is w jm , so the output of the node is: here, w jm can be a constance or linear function, or even non-linear function, but must be pre-defined. Assume that the desired and actual activities of the output node m are z m0 and z m respectively for input I. Then for each j, we need to find p kl so that the following energy function is minimized: According to generalized delta rule, a locally minimized p jk j can be found by successively applying an increment given by September 16, 1998 \Deltap kl @p kl @zm @zm @p kl @p kl where ff is a learning factor. In this algorithm, not only the value of the distribution map, but also the gradient directions are considered: 1. Assuming the input distribution map is continuous, we have where C s is a smooth factor, and and fi j (x; y) is the angle between the gradient of prototype S j (x; y) and the gradient of the distribution map I(x; y) at point O (x;y) . Obviously, 0 OE 1;j 1 and 0 OE 2;j 1. If When an input image is given, its distribution map I can be obtained, so are @I @x @x and I 2 dy, therefore we have sum Z 1Z 1i @S j @x I 0 @y I 0 x I 0 x Considering the differentiation property of the B-spline [33], @x r (x) @x similarly, @y @y r (x)P y where ur \Gammau 0 ur \Gammau 0 ur \Gammau 0 ur \Gammau 0 ur \Gammau 0 ur \Gammau 0 The gradients of templates are also rational B-spline surfaces. Therefore, we have September 16, 1998 kl kl kl kl I 2 sum I e kl (B i;r (x)B j;r (y)) dx dy (24) kl I x dx dy (25) where e kl is the effective control region of p kl and nh x I 0 y (S y io io In order to simplify the extraction algorithm, clustering algorithm was adapted: At iteration t of the templates extraction procedure, as to an input distribution map, instead of calculate all templates, only the templates with maximum OE j are upgraded. So, w jm can be defined as: where Sm is the set in which each index corresponds to a hidden node that represents a prototype that belongs to the class represented by output node m, and m is the class that the input distribution map belongs to. So, \Deltap j kl can be given by: kl kl kl IV. Template Optimization by Evolutionary Algorithm Evolutionary algorithms for optimization have been studied for more than three decades, since L. J. Fogel et al published the first works on evolutionary simulation. Many years of research and application have demonstrated that Evolutionary Algorithms, which simulate the search process of natural evolution, are powerful for the global optimization. Current Evolutionary Algorithms (EAs) include Evolutionary Programming (GP), Evolutionary Strategies (ES), Genetic Algorithm (GA) and Genetic Programming (GP) [34]. T. Back et al gave us a comparative study of the first three approaches [35]. And an introduction of evolutionary techniques on optimization is given by D. B. Fogel [36]. Several applications of evolutionary templates optimization have been reported in [37], [38], [39], [40]. However, in these algorithms, only one best template is extraction from one set of training samples. Obviously, for most of object recognition problems, only one template for a class of object is not enough for a reliable recognition system, more templates should be extracted to achieve a better performance. Sarkar [41] presented a fitness based clustering algorithm, which can be adapted for templates optimization. In the algorithm, one opponent (parent or offspring) is a set of clustering centers with different number. In its selection procedure, the fitness values of parents and offsprings are compared. In comparison, a number of opponents are selected randomly as the comparison reference. The opponent with better performance to each reference opponent can receive a win. Based on the wins, some opponents are selected as the parents of next generation. The shortcoming of the algorithm is that the computation requirement is in a exponential relationship with the cluster number. If the number of cluster number is very large, the algorithm ceases to be feasible. Unlike Sarkar [41] use a set of cluster centers as a component in evolution and select only one set as the winner, we use templates as the components directly in our algorithm and the selected survivors are a group of templates. The evolutionary processing is to September 16, 1998 simulate the evolution of a nature social system, such as ants, bees and human society, or even the nature system. The study of social system, as a part of Sociobiology, began at the middle of 19th cen- tury, after Charles Darwin published his most famous book, On the Origin of Species by Means of Natural Selection in 1859. Sociobiology seeks to extend the concept of natural selection to social systems and social behavior of animals, including humans. The sociobi- olists regard a social system, in simple, as a system consists in a plurality of individuals, heterogenous and homogeneous, interacting with each other in a situation which has at least a physical or environmental aspect, individuals who are motivated in terms of a tendency to the "optimization of gratification" and who relation to their situations, including each other, is defined and mediated in terms of a system of culturally structured and shared symbols [42]. The social behavior of the individuals can be categorized into two classes: competition and cooperation, which prompt the development of the system and keep the system. In our algorithm, we define a small, simple and homogeneous social system. Each template is just like a individual in it, and the society can only tolerate a limited number of individuals, but the number of the individuals can generate a larger number of offsprings in each generation, so only some of the offsprings with better abilities, based on the "selection rule", can be kept and others must be discarded, that is, "survival of the fittest". Moreover, the relationship between the individuals is not only competition: only winners can survive, but also cooperation: the properties of individuals are reflected by their performance in all. After the evolution by generations and generations, a group of templates are selected and the whole of they can achieve a good performance of the recognition system. Our templates optimization algorithm works as fellow: 1. Initially, generate a population N P of templates P (call them parents). In our al- gorithm, they are just the templates extracted by neural networks, as introduced in section. III. 2. Then, create a set of offsprings O in a number of NO based on the parents set P by evolutionary operations, which will describe in section IV-A. 3. Compare all offsprings and select a subset of templates in a number of N P as the September 16, 1998 parents of next generation, as present in section IV-B. 4. if the number of generation is less than a pre-specified constant, go to step2. In our algorithm, the templates in each digit class are optimized separately, that is, in each computation, only one class of templates and training samples are considered. The advantage of the algorithm can be easily adapted to parallel computer. A. Generation of Offsprings For better explanation of the selection procedure, we define: S tr the training sample; tr the ith training sample; P the parent set; k the kth component in set P ; O the offspring set; O k the kth component in set O; the selected subset of O as the next generation; k the control parameter in position (x; y) of O k ; k the control parameter in position (x; y) of P k . The generation of offsprings includes three ways: replication, mutation and recombination Replication is just to copy the value from parents to generate new offsprings. O The number of replicated offsprings is N r is the number of parents. Mutation is to add some perturbation on the parents. In our algorithm, the perturbation G are Gaussian noises. O The number of mutated offsprings is N m is a positive integer. Recombination mechanisms are used in EAs either in their usual form, producing one new individual from two randomly selected parents, or in their global form, allowing of taking components for one new individual from potentially all individuals available in the parent population [35]. In our algorithm, the values of recombined offsprings are selected randomly from the two randomly selected parents, some Gaussian noise are added as perturbations. The positions for adding perturbation are also controlled by a random value. R is a continues random value of 0 and 1, R p is a discrete random value of 0 or 1 to control perturbation position, j1 and j2 are index of two randomly selected parents. The number of recombined templates are N c O , which is pre-specified. The number of offsprings O +N c O . B. Selection Procedure The selection procedure is to select a subset of templates P 0 in a number of N P from the the offspring set O. We set the output from equation 8 between template j and training sample i is OE i j , and fitness of P 0 is: The selected subset should satisfy that the fitness(P 0 ) is bigger than the fitness of any other subsets of the offspring set O. Considering the number of possible subsets of O is NO , it is a extremely large number if we set N In order to save the computation resource to make this algorithm practical, a fast searching technology is developed, which is directly based on the value OE i . The whole selection procedure can be divided into following steps: 1. For each training sample S i tr , a number of N top templates with top function OE outputs are recorded and sorted in the decreasing order. The others are discarded. We let fT i;j be the set of the recorded output values and fT i;j be the index from T i;j to the September 16, 1998 templates, so the relation between them can be presented as: index and for training sample S i tr l (34) for 8 l 2 O and l = index g. Moreover, a flag is also set to each training sample: T i 2. For each template O k , add the values of remaining outputs together: index =k 3. Find the highest Sum k and move the template O k 0 (max(Sum k )) to the parent set P 0 for next generation: . The training samples, which index are given a flag 4. If there still exist some flags of training samples that T i for each template O k left in set O, get the new value of Sum k index =k where, f lag 0 and f lag f lag 0 and f lag September 16, 1998 if the flags of all training samples satisfy T i f lag 0, replace equation 37 with equation 35. 5. Similar to Step 3, find the template O k 0 with highest Sum k in O and move it to P 0 , re-set the flag of training samples which f lag 0, replace it with lag 6. Repeat step 4, 5 till the number of templates in set P 0 reach N P . Finally, the selected subset of templates can be considered as the desired output, and the computation resource required by the selection procedure is available to most of current scientific calculating devices. V. Experimental Results and Discussion In total, 10,426 numeral samples from NIST Special Database 3 were extracted as the training set. Each numeral was normalized into a 48 \Theta 48 binary image with 8-pixel- wide background borders for the distribution transform. So the actual image size is 64 \Theta 64. Another independent 10,426 numeral samples were also extracted from NIST Special Database 3 as the test set. First, 1000 templates were first extracted from the training set by the neural network approach discussed in section III with each numeral class having 100 templates. Every template contain 11 \Theta 11 control points. Because the amplitude of a pixel on four borders is close to 0, we set the control points on borders to 0, only the inner control points that is, 9 \Theta 9 = 81 control points were upgraded in each iteration. In this step, we test samples with different fl values. Table I is the results of the comparison, we use three sets of templates extracted with different fl values: 0:5 the algorithm achieves a reasonable good performance. Then, the 1000 templates were used as the input of our optimization algorithm to get another 1000 new templates with Because each numeral class has 100 templates, in each optimization computation N and In simulation, we also set N top = 5. Figure 6 shows some of the optimized templates, which are presented in grey-scale images. For a comparison, we have also applied a few other algorithms to recognize the same September 16, 1998 Fig. 6. Extracted templates in grey-scale images. Fig. 7. Some examples of unsure patterns. Experimental results of test samples with different fl values. ff Without rejection With rejection II A performance comparison of our approach with other numeral recognition algorithms by using test set. Techniques With Rejection Without Rejection Rejection Optimized Templates 26.7 98.9 0.0 96.4 Extracted Templates 27.9 98.7 0.0 95.8 MLP 34.4 99.9 0.0 96.9 numeral samples. These algorithms include a three-layer MLP (Multi-Layer Perceptron), the ONNC (Optimized Nearest-Neighbor Classifier) algorithm proposed by Yan [22], and initial templates extracted from an MLP. Sixty-four intensities on the 8 \Theta 8 image, rescaled from the original 160 \Theta 160 normalized binary image, were used as the inputs of the MLP based approach and ONNC. The size of the MLP is 64-75-10, As to the ONNC algorithm, it returns both the assigned class j and the distance D j of the image from the closest class prototype. The quotient of the distance D j and M i , termed as the "recognition were used as the estimation of the reliability of a classification [43], where, M i is the mean value of the recognition measures for the correctly classified numerals in the training samples that belong to class i. Table II shows the test results of out algorithm together with MLP, ONNC, and the approach with the initial templates. As it is expected, the performance with new templates from the evolutionary algorithm September 16, 1998 Experimental results of unsure patterns with different classifiers (the threshold used here is same as the threshold we used in Table II "With Rejection"). Techniques Rejection Optimized Templates 90.7 Previous Templates 87.5 MLP 44.5 ONNC 69.3 is better than that with the templates extracted from an MLP. However, compared to the MLP and ONNC classifiers, the performance of our approach is slightly lower in dealing with such isolated numerals. To further comparing these algorithms, we also used 10,426 unsure patterns (see the examples shown in Figure 7) to verify the reliability of the classifiers in rejecting illegible patterns. This ability is very important for connected character recognition. The unsure patterns used in the experiment were generated by merging two parts of isolated numerals, the left part of the right numeral and the right part of the left numeral. The choosing of two numerals, the width and relative position of each parts and the overlap degree are set randomly. Table III shows the experimental results of our algorithm together with the other two techniques. The threshold used for each classifier is the same as the setting for obtaining the results shown in Column 1 of Table II on isolated numerals. We can see that the rejection rate of optimized templates is the highest on the unsure patterns, that is, our approach can achieve a more reliable performance in rejecting illegible numerals, which is a desirable property for connected handwriting numeral recognition. We can also see from the experimental results that the optimized templates by using the evolutionary algorithm can achieve a better performance than the initial templates extracted by a neural network. It seems that the new approach is one step closer in achieving global optimization. As any other application with evolutionary algorithms, our approach also took long to find the solution. Two PentiumII-333 and one PentiumII- 300 with Linux operation system were used for the optimization computation. With 200 iterations, the computation took about 14 days for the ten classes of numerals. However, the training is required once only, the approach is still feasible. For further verification of the performance of the templates, we also applied them to recognize some 4-connected digit strings with a dynamic-programming approach. The main idea of the dynamic-programming approach is to apply the classifier move through digit string image with rectangle windows in different width, and match the containing in windows with optimized templates. In the experiment, digit string images were pre-processed in advance by the method presented in [44]. We use Ww and W h to present the width and height of window, and W s as the step length. Ww and W s are determined by the W h . We use nine windows in different width fW i w g with values: hIn the matching between the window and templates, for the consideration of possible include parts of neighboring digits in window, we do not simply use the equation 8, some techniques are applied to decrease their effect. A normalization method is first applied: find the maximal and minimal y value of the foreground, y max and y min , between the region and at the same time, find y j and y min j at region in template that fit S j (x; y) 0:5; then, scale and translate the window so that y . Except normalization, we also make a variation to equation 8, instead of executing the calculation in whole image, it happens only at is a threshold between 0 and 1. That is, equation 14 and equation 15 are replace by: H\Omega dx dy H\Omega I 2 (x; y) dx dy In experiment, we set The whole approach can be simply descript by several steps: (a) (b) Fig. 8. Examples of correctly recognized digit string, (a) original coonected digit strings; (b) separated digits. 1. from left of the image, apply rectangle windows with different widths on the image, match the containing in windows with templates; 2. if the output from a window is good enough (?0.5, for example), keep the result and move windows to the right, the begin place is at some distance (three steps in our experiment) left to the right edge of the previous window and of course, the windows should be in various width; 3. if the outputs of all the windows are not good enough, move the begin place of windows a step to right, and go to step 2; 4. repeat step 2,3, till the window reach the right end of the digit string image. It is highly possible that more than one sequence of digits will be generated from this approach. A score value is assigned to each sequence by: are the results in this sequence, and L is the length. We choose the sequence with maximal score value as the output of the approach. Figure 8 is an example of correctly recognized digit string. The first row is the original September 16, 1998 Fig. 9. Examples of recognition-failed digit string. connected digit string image, and the second row is the separated and recognized digits. And figure 9 are some recognition-failed examples. The reasons are that: 1) the writing style of the sample is not included in trainging sets (left sample); 2) degradation (right sample). Recognizing this samples cost about 2-7 hours in our PentiumII-333 with linux operation system. Obviously, this does not make for a practical approach for recognizing connected digit strings. One way to reduce this computational burden would to bind a simple classifier with it. Windows are first classified by the simple classifier, those can not be satisfactorily assigned to a class are passed to the optimized templates classifier. Another way is to re-design the dynamic-programming scheme, obviously it is not a efficient method. VI. Conclusion In this paper we first propose a new presentation of handwritten numeral templates with the rational B-spline surfaces of distance distribution maps in order to develop a classifier that can reliablly rejects illegible patterns while achiving high recognition rate. A two-step templates extraction algorithm is then presented. First, a multi-layer neural network approach is adapted to extract templates from a set of training samples, then we use an evolutionary algorithm to optimize the extracted templates. In the evolutionary algorithm, instead of making use of a fitness measure, we directly used the function that measures the difference between a template and a training sample for offspring selection directly. Experimental results on NIST Special Database 3 show that the optimized templates can achieve better performance than the intitial templates. When compared to the MLP and ONNC classifiers, the recognition rates from our new template presentation are slightly lower in dealing with isolated numerals. However, the classifiers with the new template September 16, 1998 presentation can achieve a more reliable performance in rejecting illegible patterns, which is highly desirable in connected handwritten character recognition. --R Vertex directed segmentation of handwritten numerals. Segmentation and recognition of connected handwritten numeral strings. Separation of single- and double-touching numeral strings Recognition of handwritten script: A hidden markov model based approach. Character recognition without segmentation. Handprinted word recognition on a nist data set. Modeling and recognition of cursive words with hidden Markov models. Handwritten word recognition using segmentation-free hidden Markov modeling and segmentation-based dynamic programming technique A lexion driven approach to handwritten word recognition for real-time appli- cations Handwritten word recognition with character and inter-character neural networks Character recognition - a review Optical character recognition - a survey Unconstrained handwritten numeral recognition based on radial basis competitive and cooperative networks with spatio-temporal features representation Handwritten digit recognition using combined id3-derived fuzzy rules and Markov chains Multiresolution recognition of unstrained handwritten numerals with wavelet transform and multilayer cluster neural network. Recognition of handwritten digits based on contour information. Feature extraction methods for character recognition - a survey Design and implementation of optimized nearest neighbor classifiers for handwritten digit recognition. Shape matching using lat and its application to handwritten numeral recognition. A structural model of shape deformation. A unified framework for handwritten character recognition using deformable models. Using generative models for handwritten digit recognition. Object matching using deformable templates. Representation and recognition of handwritten digits using deformable templates. On modelling An affine-invariant active contour model (ai-snake) for model-based segmenta- tion Handwritten digit recognition using optimized prototypes. Multiobjective optimization and multiple constrain handling with evolutionary algorithms part i: A unified formulation. An overview of evolutionary algorithms for parameter optimization. An introduction to simulated evolutionary optimization. Learned deformable templates for object recognition. Genetic algorithm and deformable geometric models for anatomical object September Human face location in image sequences using genetic templates. A clustering algorithm using an evolutionary programming-based approach The Social System. Separation of single- and double-touching handwritten numeral strings Length estimation of digits strings using a neural network with struture based features. --TR Context-directed segmentation algorithm for handwritten numeral strings Character recognitionMYAMPERSANDmdash;a review An overview of evolutionary algorithms for parameter optimization Using Generative Models for Handwritten Digit Recognition A Lexicon Driven Approach to Handwritten Word Recognition for Real-Time Applications Representation and Recognition of Handwritten Digits Using Deformable Templates A clustering algorithm using an evolutionary programming-based approach Postprocessing of Recognized Strings Using Nonstationary Markovian Models Shape Matching Using LAT and its Application to Handwritten Numeral Recognition Character Recognition Without Segmentation Computer-aided design applications of the rational b-spline approximation form. --CTR Chenn-Jung Huang, Clustered defect detection of high quality chips using self-supervised multilayer perceptron, Expert Systems with Applications: An International Journal, v.33 n.4, p.996-1003, November, 2007

template optimization;evolutionary algorithm;nearest neighbor classifier;digit templates;pixel-to-boundary distance map;connected handwritten digit recognition;multilayer perceptron classifier;b-spline fitting

505929

Local search characteristics of incomplete SAT procedures.

Effective local search methods for finding satisfying assignments of CNF formulae exhibit several systematic characteristics in their search. We identify a series of measurable characteristics of local search behavior that are predictive of problem solving efficiency. These measures are shown to be useful for diagnosing inefficiencies in given search procedures, tuning parameters, and predicting the value of innovations to existing strategies. We then introduce a new local search method, SDF ("smoothed descent and flood"), that builds upon the intuitions gained by our study. SDF works by greedily descending in an informative objective (that considers how strongly clauses are satisfied, in addition to counting the number of unsatisfied clauses) and, once trapped in a local minima, "floods" this minima by re-weighting unsatisfied clauses to create a new descent direction. The resulting procedure exhibits superior local search characteristics under our measures. We show that this method can compete with the state of the art techniques, and significantly reduces the number of search steps relative to many recent methods. Copyright 2001 Elsevier Science B.V.

Introduction Since the introduction of GSAT (Selman, Levesque, & Mitchell 1992) there has been considerable research on local search methods for finding satisfying assignments for CNF formulae. These methods are surprisingly effective; they can often find satisfying assignments for large CNF formulae that are far beyond the capability of current systematic search methods (however see (Bayardo & Schrag 1997; competitive systematic search results). Of course, local search is incomplete and cannot prove that a formula has no satisfying assignment when none exists. However, despite this limitation, incomplete methods for solving large satisfiability problems are proving their worth in applications ranging from planning to circuit design and diagnosis (Selman, Kautz, & McAllester 1997; Kautz & Selman 1996; Larrabee 1992). Significant progress has been made on improving the speed of these methods since the development of GSAT. In fact, a series of innovations have led to current search methods that are now an order of magnitude faster. Copyright c 2000, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. Perhaps the most significant early improvement was to incorporate a "random walk" component where variables were flipped from within random falsified clauses (Selman & Kautz 1993). This greatly accelerated search and led to the development of the very successful WSAT procedure (Selman, Kautz, & Cohen 1994). A contemporary idea was to keep a tabu list (Mazure, Sa - is, & Gr-egoire 1997) or break ties in favor of least recently flipped variables (Gent &Walsh 1993; 1995) to prevent GSAT from repeating earlier moves. The resulting TSAT and HSAT procedures were also improvements over GSAT, but to a lesser extent. The culmination of these ideas was the development of the Novelty and R Novelty procedures which combined a preference for least recently flipped variables in a WSAT-type random walk (McAllester, Selman, & Kautz 1997), yielding methods that are currently among the fastest known. A different line of research has considered adding clause- weights to the basic GSAT objective (which merely counts the number of unsatisfied clauses) in an attempt to guide the search from local basins of attraction to other parts of the search space (Frank 1997; 1996; Cha & Iwama 1996; 1995; Morris 1993; Selman & Kautz 1993). These methods have proved harder to control than the above techniques, and it has only been recent that clause re-weighting has been developed to a state of the art method. The series of "discrete Lagrange multiplier" (DLM) systems developed in (Wu & Wah 1999; Shang & Wah 1998) have demonstrated competitive results on benchmark challenge problems in the DIMACS and SATLIB repositories. Although these developments are impressive, a systematic understanding of local search methods for satisfiability problems remains elusive. Research in this area has been largely empirical and it is still often hard to predict the effects of a minor change in a procedure, even when this results in dramatic differences in search times. In this paper we identify three simple, intuitive measures of local search effectiveness: depth, mobility, and coverage. We show that effective local search methods for finding satisfying assignments exhibit all three characteristics. These, however, are conflicting demands and successful methods are primarily characterized by their ability to effectively manage the tradeoff between these factors (whereas ineffective methods tend to fail on at least one measure). Our goal is to be able to distinguish between effective and ineffective search strategies in a given problem (or diagnose problems with a given method, or tune parameters) without having to run exhaustive search experiments to their completion. To further justify our endeavor, we introduce a new local search procedure, SDF ("smoothed descent and flood") that arose from our investigation of the characteristics of effective local search procedures. We show that SDF exhibits uniformly good depth, mobility, and coverage values, and consequently achieves good performance on a large collection of benchmark SAT problems. Local search procedures In this paper we investigate several dominant local search procedures from the literature. Although many of these strategies appear to be only superficial variants of one an- other, they demonstrate dramatically different problem solving performance and (as we will see) they exhibit distinct local search characteristics as well. The local search procedures we consider start with a random variable assignment make local moves by flipping one variable x 0 a time, until they either find a satisfying assignment or time out. For any variable assignment there are a total of n possible variables to consider, and the various strategies differ in how they make this choice. Current methods uniformly adopt the original GSAT objective of minimizing the number of unsatisfied clauses, perhaps with some minor variant such as introducing clause weights or considering how many new clauses become unsatisfied by a flip (break count) or how many new clauses become satisfied (make count). The specific flip selection strategies we investigate (along with are as follows. GSAT() Flip the variable x i that results in the fewest total number of clauses being unsatisfied. Break ties randomly. (Selman, Levesque, & Mitchell 1992) HSAT() Same as GSAT, but break ties in favor of the least recently flipped variable. (Gent & Walsh 1993) WSAT-G(p) Pick a random unsatisfied clause c. With probability p flip a random x i in c. Otherwise flip the variable in c that results in the smallest total number of unsatisfied clauses. (McAllester, Selman, & Kautz 1997) WSAT-B(p) Like WSAT-G except, in the latter case, flip the variable that would cause the smallest number of new clauses to become unsatisfied. (McAllester, Selman, & Kautz 1997) WSAT(p) Like WSAT-B except first check whether some variable x i would not falsify any new clauses if flipped, and always take such a move if available. (Selman, Kautz, Pick a random clause c. Flip the variable x i in c that would result in the smallest total number of unsatisfied clauses, unless x i is the most recently flipped variable in c. In the latter case, flip x i with probability 1 p and otherwise flip the variable x j in c that results in the second smallest total number of unsatisfied clauses. (McAllester, that after the clause c is selected, flip a random x i in c with probability h, otherwise continue with Novelty. (Hoos 1999) Note that, conventionally, these local search procedures have an outer loop that places an upper bound, F , on the maximum number of flips allowed before re-starting with a new random assignment. However, we will not focus on random restarts in our experiments below because any search strategy can be improved (or at the very least, not dam- aged) by choosing an appropriate cutoff value F (Gomes, In fact, it is straightforward and well known how to do this optimally (in principle): For a given search strategy and problem, let the random variable f denote the number of flips needed to reach a solution in a single run, and let f F denote the number of flips needed when using a random restart after every F flips. Then we have the straightforward equality (Parkes & Walser 1996) Note that this always offers a potential improvement since for any cutoff F > 0. In particular, one could choose the optimal cutoff value F We report this optimal achievable performance quantity for every procedure below, using the empirical distribution of f over several runs to estimate EfF . Thus we will focus on investigating the single run characteristics of the various variable selection policies, but be sure to report estimates of what the optimum achievable performance would be using random restarts. Measuring local search performance In order to tune the parameters of a search strategy, determine whether a strategic innovation is helpful, or even debug an implementation, it would be useful to be able to measure how well a search is progressing without having to run it to completion on large, difficult problems. To begin, we consider a simple and obvious measure of local search performance that has no doubt been used to tune and debug many search strategies in the past. Depth measures how many clauses remain unsatisfied as the search proceeds. Intuitively, this indicates how deep in the objective the search is remaining. To get an overall summary, we take a depth average over all search steps. Note that it is desirable to obtain a small value of depth. Although simple minded, and certainly not the complete story, it is clear that effective search strategies do tend to descend rapidly in the objective function and remain at good objective values as the search proceeds. By contrast, strategies that fail to persistently stay at good objective values usually have very little chance of finding a satisfying assignment in a reasonable number of flips (McAllester, Selman, To demonstrate this rather obvious effect, consider the problem of tuning the noise parameter p for the WSAT procedure on a given problem. Here we use the uf100-0953 100 runs on Avg. Avg. Est. opt. uf100-0953 depth flips w/cutoff Figure 1: Depth results Flips Depth rank rank best 1 2 3 worst 4 best 1 .82 .09 .05 .04 worst 4 .03 .16 .25 .57 Mobility rank best worst 4 Figure 2: Large scale experiments (2700 uf problems) problem from the SATLIB repository to demonstrate our point. 1 Figure 1 shows that higher noise levels cause WSAT to stay higher in the objective function and significantly increase the numbers of flips needed to reach a solution. This result holds both for the raw average number of flips but also for the optimal expected number of flips using a maximum flips cutoff with random restarts, ^ EfF . The explanation is obvious: by repeatedly flipping a random variable in an unsatisfied clause, WSAT is frequently "kicked out" to higher objective values-to the extent that it begins to spend significant time simply re-descending to a lower objective value, only to be prematurely kicked out again. Although depth is a simplistic measure, it proves to be very useful for tuning noise and temperature parameters in local search procedures. By measuring depth, one can determine if the search is spending too much time recovering from large steps up in the objective and not enough time exploring near the bottom of the objective. More importantly, maintaining depth appears to be necessary for achieving reasonable search times. Figure 2 shows the results of a large experiment conducted over the entire collection of 2700 uf problems from SATLIB. This comparison ranked four comparable methods-SDF (introduced below), Novelty, Nov- elty+, and WSAT-in terms of their search depth and average flips. For each problem, the methods were ranked in terms of their average number of flips and average depth. Each (flips rank, depth rank) pair was then recorded in a ta- ble. The relative frequencies of these pairs is summarized in Figure 2. This figure shows that the highest ranked method in terms of search efficiency was always ranked near the best (and almost never in the bottom rank) in terms of search depth. Although useful, depth alone is clearly not a sufficient criterion for ensuring good search performance. A local search 1 The uf series of problems are randomly generated 3-CNF formulae that are generated at the phase transition ratio of 4:3 clauses to variables. Such formulae have roughly a 50% chance of being satisfiable, but uf contains only verified satisfiable instances. Avg. Distance Time Lag Avg. Distance Time Lag Avg. Distance Time Lag Novelty Avg. Distance Time Lag Novelty WSAT Avg. Distance Time Lag Novelty WSAT WSAT-G Avg. Distance Time Lag Novelty WSAT WSAT-G GSAT 100 runs on Avg. Avg. Avg. Est. opt. uf100-0953 mobility depth flips w/cutoff Figure 3: Mobility results could easily become stuck at a good objective value, and yet fail to explore widely. To account for this possibility we introduce another measure of local search effectiveness. Mobility measures how rapidly a local search moves in the search space (while it tries to simultaneously stay deep in the objective). We measure mobility by calculating the Hamming distance between variable assignments that are k steps apart in the search sequence, and average this quantity over the entire sequence to obtain average distances at time lags etc. It is desirable to obtain a large value of mobility since this indicates that the search is moving rapidly through the space. Mobility is obviously very important in a local search. In fact, most of the significant innovations in local search methods over the last decade appear to have the effect of substantially improving mobility without damaging depth. This is demonstrated clearly in Figure 3, again for the uf100-0953 problem. It appears that the dramatic improvements of these methods could have been predicted from their improved mobility scores (while maintaining comparable depth scores). Figure 3 covers several highlights in the development of local search methods for satisfiability. For example, one of the first useful innovations over GSAT was to add a preference for least recently flipped variables, resulting in the superior HSAT procedure. Figure 3 shows that one benefit of this change is to increase mobility without damaging search depth, which clearly corresponds to improved solution times. Another early innovation was to incorporate "random walk" in GSAT. Figure 3 shows that WSAT-G also delivers a noticeable increase in mobility-again resulting in a dramatic reduction in solution times. It is interesting to note that the apparently subtle distinction between WSAT- G and WSAT in terms of their definition is no longer sub- 100 runs on Avg. Avg. Avg. Avg. Est. uf100-0953 cover. rate mob. dep. flips opt. 26 4.1 1,355 1,355 Figure 4: Coverage results tle here: WSAT offers a dramatic improvement in mobility, along with an accompanying improvement in efficiency. Fi- nally, the culmination of novelty and random walk in the Novelty procedure achieves even a further improvement in mobility, and, therefore it seems, solution time. We have observed this effect consistently over the entire range of problems we have investigated. Thus it appears that, in addition to depth, mobility also is a necessary characteristic of an effective local search in SAT problems. To establish this further, Figure 2 shows the results of a large experiment on the entire collection of 2700 uf problems in SATLIB. The same four procedures were tested (SDF, Nov- elty, Novelty+, WSAT) and ranked in terms of their search mobility and solution time. The results show that the highest ranked in terms of mobility is almost always ranked near the top in problem solving efficiency, and that low mobility tends to correlate with inferior search efficiency. A final characteristic of local search behavior that we consider is easily demonstrated by a simple observation: Hoos presents a simple satisfiable CNF formula with five variables and six clauses that causes Novelty to (sometimes) get stuck in a local basin of attraction that prevents it from solving an otherwise trivial problem. The significance of this example is that Novelty exhibits good depth and mobility in this case, and yet fails to solve what is otherwise an easy problem. This concern led Hoos to develop the slightly modified procedure Novelty+ in (Hoos 1999). The characteristic that Novelty is missing in this case is coverage. Coverage measures how systematically the search explores the entire space. We compute a rate of coverage by first estimating the size of the largest "gap" in the search space (given by the maximum Hamming distance between any unexplored assignment and the nearest evaluated assign- ment) and measuring how rapidly the largest gap size is being reduced. In particular, we define the coverage rate to be (n max gap)=search steps. Note that it is desirable to have a high rate of coverage as this indicates that the search is systematically exploring new regions of the space as it proceeds. Figure 4 shows that Hoos's modified Novelty+ procedure improves the coverage rate of Novelty on the uf100-0953 problem. Space limitations do not allow a full description, but Novelty+ demonstrates uniformly better coverage than Novelty while maintaining similar values on other measures, and thus achieves better performance on nearly every problem in the benchmark collections. our results lead us to hypothesize that local search procedures work effectively because they descend quickly in the objective, persistently explore variable assignments with good objective values, and do so while moving rapidly through the search space and visiting very different variable assignments without returning to previously explored regions. That is, we surmise that good local search methods do not possess any special ability to predict whether a local basin of attraction contains a solution or not-rather they simply descend to promising regions and explore near the bottom of the objective as rapidly, broadly, and systematically as possible, until they stumble across a solution. Although this is a rather simplistic view, it seems supported by our data and moreover it has led to the development of a new local search technique. Our new procedure achieves good characteristics under these measures and, more impor- tantly, exhibits good search performance in comparison to existing methods. A new local search strategy: SDF Although the previous measures provide useful diagnostic information about local search performance, the main contribution of this paper is a new local search procedure, which we call SDF for "smoothed descent and flood." Our procedure has two main components that distinguish it from previous approaches. First, we perform steepest descent in a more informative objective function than earlier methods. Second, we use multiplicative clause re-weighting to rapidly move out of local minima and efficiently travel to promising new regions of the search space. Recall that the standard GSAT objective simply counts the number of unsatisfied clauses for a given variable as- signment. We instead consider an objective that takes into account how many variables satisfy each clause. Here it will be more convenient to think of a reversed objective where we seek to maximize the number of satisfied clauses instead of minimize the number of unsatisfied clauses. Our enriched objective works by always favoring a variable assignment that satisfies more clauses, but all things being equal, favoring assignments that satisfy more clauses twice (subject to satisfying the same number of clauses once), and so on. In effect, we introduce a tie-breaking criterion that decides, when two assignments satisfy the same number of clauses, that we should prefer the assignment which satisfies more clauses on two distinct variables, and if the assignments are still tied, that we should prefer the assignment that satisfies more clauses on three distinct variables, etc. This tie-breaking scheme can be expressed in a scalar objective function that gives a large increment to the first satisfying vari- able, and then gives exponentially diminishing increments for subsequent satisfying variables for a given clause. For k-CNF formulas with m clauses, such a scoring function is c 's that satisfy c) Our intuition is that performing steepest ascent in this objective should help build robustness in the current variable assignment which the search can later exploit to satisfy new clauses. In fact, we observe this phenomenon in our exper- iments. Figure 5 shows that following a steepest ascent in Avg. Depth Time Steps Avg. Depth Time Steps GSAT Figure 5: Descent results f ABE descends deeper in the original GSAT objective than the GSAT procedure itself (before either procedure reaches a local extrema or plateau). This happens because plateaus in the GSAT objective are not plateaus in f ABE ; in fact, such plateaus are usually opportunities to build up robustness in satisfied clauses which can be later exploited to satisfy new clauses. This effect is systematic and we have observed it in every problem we have examined. This gives our first evidence that the SDF procedure, by descending deeper in the GSAT objective, has the potential to improve the performance of existing local search methods. The main outstanding issue is to cope with local maxima in the new objective. That is, although f ABE does not contain many plateaus, the local search procedure now has to deal with legitimate (and numerous) local maxima in the search space. While this means that plateau walking is no longer a significant issue, it creates the difficulty of having to escape from true traps in the objective function. Our strategy for coping with local maxima involves the second main idea behind the SDF procedure: multiplicative clause re-weighting. Note that when a search is trapped at a local maxima the current variable assignment must leave some subset of the clauses unsatisfied. Many authors have observed that such local extrema can be "filled in" by increasing the weight of the unsatisfied clauses to create a new search direction that allows the procedure to escape (Wu & Wah 1999; Frank 1996; Morris 1993; Selman & Kautz 1993). How- ever, previous published re-weighting schemes all use additive updates to increment the clause weights. Unfortunately, additive updates do not work very well on difficult search problems because the clauses develop large weight differences over time, and this causes the update mechanism to lose its ability to rapidly adapt the weight profile to new regions of the search space. Multiplicative updating has the advantage of maintaining the ability to swiftly change the weight profile whenever necessary. One final issue we faced was that persistently satisfied clauses would often lose their weight to the extent that they would become frequently falsified, and consequently the depth of search (as measured in the GSAT objective) would deteriorate. To cope with this effect, we flattened the weight profile of the satisfied clauses at each re-weighting by shrinking them towards their common mean. This increased the weights of clauses without requiring them to be explicitly falsified and had the overall effect of restoring search depth and improving performance. The final SDF procedure we tested is summarized as follows. the variable x i that leads to the greatest increase in the weighted objective c w(c) score(# x i 's that satisfy c) If the current variable assignment is a local maximum and not a solution, then re-weight the clauses to create a new ascent direction and continue. Multiplicatively re-weight the unsatisfied clauses and re-normalize the clause weights so that the resulting largest difference in the f WABE objective (when flipping any one variable) is -. (That is, create a minimal greedy search direction.) Then flatten the weight profile of the satisfied clauses by shrinking them 1 of the distance towards their common mean (to prevent the weights from becoming too small and causing clauses to be falsified gratuitously). One interesting aspect of this procedure is that it is almost completely deterministic (given that ties are rare in the ob- jective, without re-starts) and yet seems to perform very well in comparison to the best current methods, all of which are highly stochastic. We claim that much of the reason for this success is that SDF maintains good depth, mobility, and coverage in its search. This is clearly demonstrated in Figures 1-3 which show that SDF obtains superior measurements in every criteria. Evaluation We have conducted a preliminary evaluation of SDF on several thousand benchmark SAT problems from the SATLIB and DIMACS repositories. The early results appear to be very promising. Comparing SDF to the very effective Nov- elty+ and WSAT procedures, we find that SDF typically reduces the number of flips needed to find a solution over the best of Novelty+ and WSAT by a factor of two to four on random satisfiable CNF formulae (from the uf, flat, and aim collections), and by a factor of five to ten on non-random CNF formulae (from the SATLIB blocks-world and ais problem sets). These results are consistent across the vast majority of the problems we have investigated, and hold up even when considering the mean flips without restart, median flips, and optimal expected flips using restarts estimated from (1). However, our current implementation of SDF is not optimized and does not yet outperform current methods in terms of CPU time. Details are reported below. In all experiments, each problem was executed 100 times and results averaged over these runs. All problems were tried with SDF( :2 WSAT(:5). Furthermore, smaller problems were also tried with HSAT, GSAT, and simulated annealing. 2 There are We have as yet been unable to replicate the reported results for the DLM procedures from the published descriptions (Wu & Wah 1999; Shang & Wah 1998), and so did not include them in our study. This remains as future work. Avg. Est. opt. Avg. Est. opt. Est. opt. uf125 1906 1563 5160 2712 5876 Failed Avg. flips Est. opt. Est. opt. SDF / Nov+ huge 2561 2560 11104 0 / 0 logistics.c SDF WSAT % Failed Avg. flips Est. opt. Est. opt. SDF / WSAT Figure Search results three sets of experiments reported in Figure 6. The first set covers a wide array of random SAT problems from both the SATLIB and DIMACS repositories. The results shown are averaged over all problems in the respective problem set and are shown for the runs with SDF, Novelty+ (which was 2nd best), and WSAT. The second set covers large planning problems. The results are shown for SDF and Novelty+ (2nd best), and the failure rates of each are compared. The third set covers the ais (All-Interval Series) problem set and shows results for SDF and WSAT (2nd best). In all experiments, the mean flips without restart and optimal expected flips are reported for SDF, and the optimal expected flips is reported for the other algorithms (when significantly smaller than the mean without restarts). The results for the non-random blocks-world and ais problems are particularly striking. These problems challenge state of the art local search methods (verified in Figure and yet SDF appears to solve them relatively quickly. This suggests that, although SDF shares many similarities to other local search methods currently in use, if might offer a qualitatively different approach that could yield benefits in real world problems. The current implementation of SDF is unfortunately not without its limitations. We are presently using a non-optimized floating-point implementation, which means that even though SDF executes significantly fewer search steps (flips) to solve most problems, each search step is more expensive to compute. The overhead of our current implementation is about factor of six greater than that of Novelty or WSAT per flip, which means that in terms of CPU time, SDF is only competitive with the current best methods in some cases (e.g. bw large.b). However, the inherent algorithmic complexity of each flip computation in SDF is no greater than that of GSAT, and we therefore expect that an optimized implementation in integer arithmetic will speed SDF up considerably-possibly to the extent that it strongly outperforms current methods in terms of CPU time as well. --R Using CSP look-back techniques to solve real-world SAT instances Performance test of local search algorithms using new types of random CNF formulas. Adding new clauses for faster local search. Weighting for Godot: Learning heuristics for GSAT. Learning short-tem weights for GSAT Towards an understanding of hill-climbing procedures for SAT Unsatisfied variables in local search. In Hallam Boosting combinatorial search through randomization. On the run-time behavior of stochastic local search algorithms for SAT Pushing the envelope: Plan- ning Test pattern generation using boolean satisfia- bility Evidence for invariants in local search. The breakout method for escaping from local minima. Tuning local search for satisfiability testing. Noise strategies for improving local search. Ten challenges in propositional reasoning and search. A new method for solving hard satisfiability problems. A discrete Lagrangian based global search method for solving satisfiability problems. Trap escaping strategies in discrete Lagrangian methods for solving hard satisfiability and maximum satisfiability problems. --TR Noise strategies for improving local search Boosting combinatorial search through randomization On the run-time behaviour of stochastic local search algorithms for SAT Trap escaping strategies in discrete Lagrangian methods for solving hard satisfiability and maximum satisfiability problems Heavy-Tailed Phenomena in Satisfiability and Constraint Satisfaction Problems A Discrete Lagrangian-Based Global-Search Method for Solving Satisfiability Problems An Efficient Global-Search Strategy in Discrete Lagrangian Methods for Solving Hard Satisfiability Problems Generating Satisfiable Problem Instances Local Search Characteristics of Incomplete SAT Procedures --CTR Monte Lunacek , Darrell Whitley , James N. Knight, Measuring mobility and the performance of global search algorithms, Proceedings of the 2005 conference on Genetic and evolutionary computation, June 25-29, 2005, Washington DC, USA Alex Fukunaga, Automated discovery of composite SAT variable-selection heuristics, Eighteenth national conference on Artificial intelligence, p.641-648, July 28-August 01, 2002, Edmonton, Alberta, Canada Holger H. Hoos, A mixture-model for the behaviour of SLS algorithms for SAT, Eighteenth national conference on Artificial intelligence, p.661-667, July 28-August 01, 2002, Edmonton, Alberta, Canada Benjamin W. Wah , Yixin Chen, Constraint partitioning in penalty formulations for solving temporal planning problems, Artificial Intelligence, v.170 n.3, p.187-231, March 2006

constraint satisfication;experimental analysis;satisfiability;local search

505996

Improving Latency Tolerance of Multithreading through Decoupling.

AbstractThe increasing hardware complexity of dynamically scheduled superscalar processors may compromise the scalability of this organization to make an efficient use of future increases in transistor budget. SMT processors, designed over a superscalar core, are therefore directly concerned by this problem. This work presents and evaluates a novel processor microarchitecture which combines two paradigms: simultaneous multithreading and access/execute decoupling. Since its decoupled units issue instructions in-order, this architecture is significantly less complex, in terms of critical path delays, than a centralized out-of-order design, and it is more effective for future growth in issue-width and clock speed. We investigate how both techniques complement each other. Since decoupling features an excellent memory latency hiding efficiency, the large amount of parallelism exploited by multithreading may be used to hide the latency of functional units and keep them fully utilized. Our study shows that, by adding decoupling to a multithreaded architecture, fewer threads are needed to achieve maximum throughput. Therefore, in addition to the obvious hardware complexity reduction, it places lower demands on the memory system. Since one of the problems of multithreading is the degradation of the memory system performance, both in terms of miss latency and bandwidth requirements, this improvement becomes critical for high miss latencies, where bandwidth might become a bottleneck. Finally, although it may seem rather surprising, our study reveals that multithreading by itself exhibits little memory latency tolerance. Our results suggest that most of the latency hiding effectiveness of SMT architectures comes from the dynamic scheduling. On the other hand, decoupling is very effective at hiding memory latency. An increase in the cache miss penalty from 1 to cycles reduces the performance of a 4-context multithreaded decoupled processor by less than 2 percent. For the nondecoupled multithreaded processor, the loss of performance is about 23 percent.

Introduction The gap between the speeds of processors and memories has kept increasing in the past decade and it is expected to sustain the same trend in the near future. This divergence implies, in terms of clock cycles, an increasing latency of those memory operations that cross the chip boundaries. In addition, processors keep growing their capabilities to exploit parallelism by means of greater issue widths and deeper pipelines, which makes even higher the negative impact of memory latencies on the performance. To alleviate this problem, most current processors devote a high fraction of their transistors to on-chip caches, in order to reduce the average memory access time. Several prefetching techniques have also been developed, both hardware and software [3]. Some processors, commonly known as out-of-order processors [40, 20, 18, 8, 9], include dynamic scheduling techniques, most of them based on Tomasulo's algorithm [34] or variations of it, that allow them to tolerate both memory and functional unit latency, by overlapping it with useful computations of other independent instructions. To implement it, the processor is capable of filling issue slots with independent instructions by looking forward in the instruction stream, into a limited instruction window. This is a general mechanism that aggressively extracts the instruction parallelism available in the instruction window. As memory latencies continue to grow in the future, out-of-order processors will need larger instruction windows to find independent instructions to fill the increasing number of empty issue slots, and this number will grow even faster with greater issue widths. The increase in the instruction window size will have an obvious influence on the chip area, but its major negative impact will strike at the processor clock cycle time. As reported recently [21], the networks involved in the issue wake-up and bypass mechanisms, and also - although to a less extent - those of the renaming stage, are in the critical path that determines the clock cycle time. In their analysis, the authors of that study state that the delay function of these networks has a component that increases quadratically with the window length. And, although linearly, it also depends strongly on the issue width. Moreover, higher density technologies only accelerate the increase in these latencies. Their analysis suggest that out-of-order architectures could find in the future a serious boundary on their clock speeds. Different kinds of architectures have been proposed recently, either in-order or out-of-order, which address the clock cycle problem by partitioning critical components of the architecture and/or providing less complex scheduling mechanisms [30, 6, 16, 21, 41]. They follow different partitioning strategies. One of them is the access/execute paradigm, which was first proposed for early scalar architectures to provide them with dual issue and a limited form of dynamic scheduling that is especially oriented to tolerate memory latency. We believe that decoupled access/execute architectures can regain progressively interest as far as issue widths and memory latencies keep growing and demanding larger instruction windows, because these trends will make it worth trading issue complexity for clock speed. Typically, a decoupled access/execute architecture [26, 27, 7, 39, 38, 23, 2, 12] splits, either statically or dynamically, the instruction stream into two. The access stream is composed of all those instructions involved in the fetch of data from memory, and it runs asynchronously with respect to the execute stream, which is formed by the instructions that process these data. Both streams are executed on independent processing units (called AP and EP respectively, in this paper). The AP is expected to execute in advance of the EP and to prefetch data from memory into the appropriate buffering structures, so that the EP can consume them without any delay. This anticipation or slippage, may involve multiple conditional branches. However, the amount of slippage between the AP and the EP highly depends on the program ILP, because data and control dependences can force both units to synchronize - the so called Loss of Decoupling events [2, 35] - producing a serious performance degradation. The decoupling model presented in this paper performs dynamic code partitioning, as in [27, 12], by following a simple scheme which is based on the instruction data types, i.e. integer or fp. Although this rather simplistic scheme mostly benefits to numerical programs, it still provides a basis for our study which is mainly focused on the latency hiding potential of decoupling and its synergy with multithreading. Recent studies [22, 24] have proposed other alternative compiler-assisted partitioning schemes that address the partitioning of integer codes. Since one of the main arguments for the decoupled approach is the reduced issue logic complexity, it has been chosen to issue instructions in-order within each processing unit. Such a decoupled architecture adapts to higher memory latencies by scaling much simpler structures than an out-of-order, i.e. scaling at a lower hardware cost, or conversely scaling at a higher degree with similar cost. It may be argued that in-order processors have a limited potential to exploit ILP. However, current compiling techniques can extract much ILP and thus, the compiler can pass this information to the hardware instead of using run-time schemes. This is the approach that emerging EPIC (Explicitly Parallel Instruction Computing) architectures take [10]. We propose a new decoupled architecture which provides both the AP and the EP with a powerful dynamic scheduling mechanism: simultaneous multithreading [37, 36]. Each processing unit has several contexts, each issuing instructions in the above mentioned decoupled mode, which are active simultaneously and compete for the issue slots, so that instructions from different contexts can be issued in the same cycle. We show in this study that the combination of decoupling and mulithreading takes advantage of their best features: while decoupling is a simple but effective technique for hiding high memory latencies with a reduced issue complexity, multithreading provides enough parallelism to hide functional unit latencies and keep functional units busy. In addition, multithreading also helps to hide memory latency when a program decouples badly. However, as far as decoupling succeeds in hiding memory latency, few threads are needed to keep the functional units busy and achieve a near-peak issue rate. This is an important result, since having few threads reduces the memory pressure, which has been reported to be the major bottleneck in multithreading architectures, and reduces the hardware cost and complexity. The rest of this paper is organized as follows. Section 2 describes the base decoupled architecture. It is then analyzed in Section 3, providing justification for multithreading. Section 4 describes and evaluates the proposed multithreaded decoupled architecture. Finally, we summarize our conclusions in Section 5. 2. The basic decoupled architecture model The baseline decoupled architecture considered in this paper (Figure 1) consists of two superscalar decoupled processing units: the Address Processing unit (AP) and the Execute Processing unit (EP). The decoupled processor executes a single instruction stream, based on the DEC-alpha ISA [5], by splitting it dynamically and dispatching the instructions to either the AP or the EP. There are two separate physical register files, one in the AP with 64 integer registers, the other in the EP with 96 FP registers. Both units share a common fetch and dispatch stage, while they have separate issue, execute and write-back stage pipelines. Next, there is a brief description of each stage: Memory Subsystem Store Addres s Figure 1: Scheme of the base decoupled processor Fetch Decode & Rename Instruction Reg. File Reg. File Map Table Register The fetch stage reads up to 4 consecutive instructions per cycle (but less than 4 if there is a taken branch among them) from an infinite I-cache. Notice that I-cache miss ratios for SPEC FP95 are usually very low, so this approximation introduces a small perturbation. It is also provided with a conditional branch prediction scheme based on a 2K entry Branch History Table, with a 2- bit saturating counter per entry [25]. The dispatch stage decodes and renames up to 4 instructions per cycle and sends them to either the AP or to the instruction queue IQ (48 entries) of the EP, depending on whether they are integer or floating point instructions. All memory instructions are dispatched to the AP. The IQ allows the AP to execute ahead of the EP, providing the necessary slippage between them to hide the memory latency. Exceptions are kept precise by means of a reorder buffer, a graduation mechanism, and the register renaming map table [13, 28]. Other decoupled architectures [27] had chosen to steer memory instructions to both units to allow copying data from the load queue to registers. Since preliminary studies showed that such code expansion would significantly reduce the performance, we implemented dynamic register renaming, which avoids any duplication. That is, data fetched from memory is written into a physical register rather than into a data queue, eliminating the need for copying. It is also a convenient way to manage the disordered completion of loads when a lockup-free cache is present. Duplication of conditional branch instructions, also used in [27], may be avoided by incorporating similar speculation and recovery mechanisms as it uses the MIPS R10000 to identify the instructions to squash in case of a misprediction. Both the AP and the EP are provided with 2 general purpose, fully pipelined functional units whose latencies are 1 cycle (AP) and 4 cycles (EP), respectively. Each processing unit can read and issue up to 2 instructions per cycle. To better exploit the parallelism between the AP and the EP, the instructions can issue and execute speculatively beyond up to four unresolved branches (as the MIPS R10000 [40] or the PowerPC 620 [20]). This feature may become sometimes a key factor to enable the AP to slip ahead of the EP. Store addresses are held in the SAQ queue (32 entries) until the stores graduate. Loads are issued to the cache after being disambiguated against all the addresses held in the SAQ. Whenever a dependence is encountered, the data from the pending store is immediately bypassed to the register if it is available. Otherwise, the load is put aside until this data is forwarded to it. The primary data cache is on-chip, 2-ported [31], direct-mapped, 64 KB sized, with a block length, and it implements a write-back policy to minimize off-chip bus traffic. It is a lockup-free cache [17], modelled similarly to the MAF of the Alpha 21164 [5]. It can hold up to outstanding (primary) misses to different lines, each capable to merge up to 4 (secondary) per pending line. We assume that L1 cache misses always hit in an infinite multibanked off-chip L2 cache, and they have a 16 cycle latency plus any penalty due to bus contention. The L1-L2 interface consists of a fast 128-bit wide data bus, capable to deliver 16 bytes per cycle, like that of the R10000 (the bus is busy during 2 cycles for each line that is fetched or copied back). 3. Quantitative evaluation of a decoupled processor In this section it is first characterized the major sources of wasted cycles in a typical single-threaded decoupled processor. Next, the latency hiding effectiveness of this architecture is evaluated, identifying the main factors that influence the latency tolerance of the architecture. Other studies on decoupled machines have been carried out before [1, 26, 7, 29, 27, 39, 38, 19, 14], but they did not incorporate techniques like store-load forwarding, control speculation or lockup-free caches. This section also provides the motivation for the multithreaded decoupled architecture that is analyzed in Section 4. 3.1. Experimental framework The experiments were carried out with a trace driven simulator. The binary code was obtained by compiling the SPEC FP95 benchmark suite [33], for a DEC AlphaStation 600 5/266, with the DEC compiler applying full optimizations. The trace was generated by running this code previously instrumented with the ATOM tool [32]. The simulator modelled, cycle-by-cycle, the architecture described in the previous section, and run the SPEC FP95 benchmarks, fed with their largest available input data sets. Since it is very slow, due to the detail of the simulations, we run only a portion of 100M instructions of each benchmark., after skipping an initial start-up phase. To determine the appropriate initial discarded offset we compared the instruction-type frequencies of such a fragment starting at different points, with the full run frequencies. We found that this phase has not the same length for all the benchmarks: about 5000 M instructions for 101.tomcatv and 1000 M for 104.hydro2d and 146.wave5; and just 100 M for the rest of the benchmarks. 3.2. Sources of wasted cycles Figure 2 shows the throughput of the issue stage in terms of the percentage of committed instructions over the total issue slot count (i.e. percent of issue slots where it is really doing useful work) for the AP and the EP. The wasted throughput is also characterized, by identifying the cause for each empty issue slot. Four different configurations have been evaluated, which differ in whether lockup-free cache is included and whether the store-load forwarding mechanism is enabled. To stress the memory system, in this section we assume an 8 KB L1 data cache. As shown in Figure 2, when a lockup-free cache is not present (first and second bars), the AP is stalled by load misses and the EP is starved for most of the time. Miss latency increases the AP cycle count far above the EP cycle count. The AP execution time becomes the bounding limit of the global performance, and decoupling can hardly hide memory latencies. The nature of these Figure 2: Issue slot breakdown for several decoupled architectures, that show the effects of a lockup-free cache and a store-load forwarding mechanism (8 KB L1 cache size). none forwd l-free l-free +forwd none forwd l-free l-free +forwd Configuration1030507090%of Issue slots wrong-path instr. or idle wait operand from FU wait operand from memory blocking miss st/ld hazard other useful work stalls is a structural hazard. When a lockup-free cache is used, this kind of stalls are almost eliminated (third and fourth bars). Of course, this uncovers other overlapped causes, but the overall improvement in performance achieves an impressive 2.3 speed-up (from 0.98 to 2.32 IPC). A memory data hazard can occur between a store and a FP load, and it is detected during memory disambiguation. When store-load forwarding is not enabled (first and third bars), a memory hazard produces a stall on the AP until the store is issued to the cache. In addition, it causes a slippage reduction between the two units - we call this event a loss of decoupling, or LOD [2, 35] - that may expose the EP to be penalized by the memory latency in case of a subsequent load miss. The amount of slippage reduction between the AP and the EP caused by a memory hazard depends on how close the load is scheduled after the matching store. The results depicted in Figure 2 show that the AP stalls (labelled st/ld hazards) are almost completely removed when the store-load forwarding is enabled. However, the average improvement on the EP performance is almost negligible (overall IPC increases just by 1.8%). This latter fact suggests that either the stores are scheduled enough in advance of the matching loads, or there is little probability to get a subsequent miss. Finally, for a full featured configuration (fourth bar in the graph), it can be observed that the major source of wasted slots in the EP are true data dependences between register operands (labelled wait operand from FU), and that these stalls are less than those caused by misses (labelled wait operand from memory). Notice that although there are many more loads to the EP registers than loads to the AP registers, the stalls caused by misses are similar on both processor units because each integer load miss produces a higher penalty, as this will be more clearly illustrated in the next section. 3.3. Latency hiding effectiveness The interest of a decoupled architecture is closely related to its ability to hide high memory latencies without resorting to other more complex issue mechanisms. The latency hiding potential of a decoupled processor depends strongly on the decoupling behaviour of the programs being tested. For some programs, the scheduling ability of the compiler to remove LOD events, which force the AP and the EP to synchronize, is also a key factor. However, the compiler we have used (Digital f77) is not especially tailored to a decoupled processor. Therefore, since the latency hiding effectiveness of decoupling provides the basis for our proposed multithreaded decoupled architecture, in order to validate our conclusions, we are interested in having an assessment of it in our base architecture, without any specific compiler support. For this purpose, we have run the 10 benchmarks with the external L2 cache latency varying from 1 to 256 cycles. The simulations assume all the architectural parameters described in Section 2 except that all the architectural51525 L2 Latency (cycles)51525 Average Perceived FP-Load Miss Latency (cycles) tomcatv su2cor hydro2d mgrid applu turb3d apsi tomcat swim su2cor hydro mgrid applu turb3d apsi fpppp wave5 Benchmark2060100 Miss latency stores loads20601001401 L2 Latency (cycles)2060100140Average Perceived I-Load Miss Latency (cycles) tomcatv su2cor hydro2d mgrid applu turb3d apsi L2 Latency (cycles) loss su2cor hydro2d mgrid applu turb3d apsi Figure 3-a: Perceived miss latency of FP loads. Figure 3-b: Perceived miss latency of Integer loads. Figure 3-c: Miss Ratios of Loads and Stores, when L2 latency is 256 cycles. Figure 3-d: Impact of latency on performance (loss relative to the 1-cycle L2 latency case). queues and physical register files are scaled up proportionally to the L2 latency. In addition to the performance, we have also measured separately the average "perceived" latency of integer and FP load misses. Since we are interested in the particular benefit of decoupling, independently of the cache miss ratio, this average does not include load hits. The perceived latency of FP load misses measures the EP stalls caused by misses, and reveals the "decoupled behavior" of a program, i.e. the amount of slippage of the AP with respect to the EP. As shown in Figure 3-a, except for fpppp, more than 96% of the FP load miss latency is always hidden. The perceived latency of integer load misses measures the AP stalls caused by misses, and it depends on the ability of the compiler to schedule integer loads ahead of other dependent instructions. As shown in Figure 3-b, fpppp, su2cor, turb3d and wave5 are the programs that experience the largest integer load miss stalls. Regarding the impact of the L2 latency on performance (see Figure 3-d), although programs like fpppp or turb3d have quite high perceived load miss latencies, they are hardly performance degraded due to their extremely low miss ratios (depicted in Figure 3-c). The most performance degraded programs are those with both high perceived miss latencies and significant miss ratios: hydro2d, wave5 and su2cor. To summarize, performance is little affected by the L2 latency when either it can be hidden efficiently (tomcatv, swim, mgrid, applu and apsi), or when the miss ratio is low (fpppp and turb3d), but it is seriously degraded for programs that lack both features (su2cor, wave5 and hydro2d). The hidden miss latency of FP loads depends on the good decoupling behavior of the programs, while that of integer loads relies exclusively on the static instruction scheduling. 4. A multithreaded decoupled architecture As shown in the previous section, most of the stalls of a decoupled processor may be removed, except those caused by true data dependences between register operands in the EP (Figure 2 right, labelled wait operand from FU), because of the restricted ability of the in-order issue model to exploit ILP. If both the AP and the EP were provided with some dynamic scheduling capability, most of these stalls could also be removed. Simultaneous multithreading (SMT) is a dynamic scheduling technique that increases processor throughput by exploiting thread level parallelism. Multiple contexts simultaneously active compete for issue slots and functional units. Previous studies of SMT focused on several dynamic instruction scheduling mechanisms [4, 11, 37, 36, among others] other than decoupling. In this paper, we analyze its potential when implemented on a decoupled processor. We still refer to it as simultaneous although there are obvious substantial differences from the original SMT, because it retains the key concept of issuing from different threads during a single cycle. Since decoupling provides excellent memory latency tolerance, and multithreading supplies enough amounts of parallelism to remove the remaining stalls, we expect important synergistic effects in a new microarchitecture which combines these two techniques. In this section we present and evaluate the performance and memory latency tolerance of the multithreaded decoupled access/execute architecture, and we analyze the mutual benefits of both techniques, especially when the miss latency is large. 4.1. Architecture overview Our proposal is a multithreaded decoupled architecture (Figure 4). That is, each thread executes in a decoupled mode, sharing the functional units and the data cache with other threads. The base Memory Subsystem Store Addres s Figure 4: Scheme of the multithreaded decoupled processor Instruction Reg. Files Reg. Files Map Tables Register Fetch Dispatch & Rename multithreaded decoupled architecture is based on the decoupled design of the previous section with some extensions: it can run up to 6 threads and issue up to 8 instructions per cycle (4 at the AP and 4 at the EP) to 8 functional units. The L1 lockup-free data cache is augmented to 4 ports. The fetch and dispatch stages - including branch prediction and register map tables - and the register files and queues are replicated for each context. The issue logic, functional units and the data cache are shared by all the threads. In our model, all the threads are allowed to compete for each of the 8 issue slots each cycle, and priorities among them are determined in pure round-robin order (similar to the full simultaneous issue scheme reported in [37]). Each cycle, only two threads have access to the I- cache, and each of them can fetch up to 8 consecutive instructions (up to the first taken branch). The chosen threads are those with less instructions pending to be dispatched (similar to the RR-2.8 with I-COUNT schemes, reported in [36]). 4.2. Experimental evaluation The multithreaded decoupled simulator is fed with t different traces, corresponding to t independent threads. The trace of every thread is built by concatenating the first 10 million instructions of the 10 traces used in the previous section - each thread using a different permutation - thus totalling 100 million instructions per thread. In this way, all threads have different traces but balanced workloads, similar miss-ratios, etc. Figure 5 shows the wasted issue slots when varying the number of threads from 1 to 6. Since different threads may be candidates for the same slot, and each can lose it because of a different cause, in order to characterize the loss of performance, we have classified the wasted issue slots proportionally to the causes that prevent individual threads from issuing. 4.3. Wasted issue slots in the multithreaded decoupled architecture The first column in Figure 5 represents the case with a single thread, and it reveals, as expected, that the major bottleneck is caused by the EP functional units latency (caused by the lack of parallelism of the in-order issue policy, as discussed in Section 3). When two more contexts are added, the multithreading mechanism reduces drastically these stalls in both units, and produces a 2.31 speed-up (from 2.68 IPC to 6.19 IPC). Since with 3 threads the AP functional units are nearly saturated (90.7%), negligible additional speed-ups are obtained by adding more contexts (6.65 IPC is achieved with 4 threads). Notice that although the AP almost achieves its maximum throughput, the EP functional units do not saturate due to the load imbalance between the AP and the EP. Therefore, the effective peak performance is reduced by 17%, from 8 to 6.65 IPC. This problem could be addressed with a different choice of the number of functional units in each processor unit, but this is beyond the scope of this study. Another important remark is that when the number of threads is increased, the combined working set is larger, and the miss ratios increase progressively, putting greater demands on the external bus bandwidth. On average, there are more pending misses, thus increasing the effective load miss latency, and increasing the EP stalls caused by waiting operands from memory (see rightmost graph of Figure 5). On the other hand, the AP stalls due to integer load misses (see operands from memory in the leftmost graph of Figure 5) are almost eliminated by multithreading since these loads do not benefit from decoupling. Number of threads1030507090%of Issue Cycles idle wait operand from FU wait operand from memory other useful work Number of threads1030507090%of Issue Cycles empy i-queue wait operand from FU wait operand from memory other useful work Figure 5: AP (left) and EP (right) issue slots breakdown for the multithreaded decoupled architecture. 4.4. Latency hiding effectiveness Multithreading and decoupling are two different approaches to tolerate high memory latencies. We have run some experiments, similar to those of Section 3.3, for a multithreaded decoupled processor having from 1 to 4 contexts to quantify its latency tolerance. In addition, some other experiments are also carried out to reveal the contribution of each mechanism to the latency hiding effect. They consist of a set of identical runs on a degenerated version of our multithreaded architecture where the instruction queues are disabled (i.e. a non-decoupled multithreaded architecture). Figure 6-a shows the average perceived load miss latency from the point of view of each individual thread, for the 8 configurations mentioned above, by varying L2 latency from 1 to 256 cycles. This metric expresses the average number of times an instruction of a scheduled thread cannot issue because its operand depends on a pending load miss. Figure 6-b shows the corresponding relative performance loss (with respect to the 1-cycle L2 latency) of each of the 8 configurations. Notice that this metric compares the tolerance of these architectures to memory latency, rather than their absolute performance. Several conclusions can be drawn from these graphs. First, we can observe in Figure 6-a that the average load miss latency perceived by an individual thread is quite low when decoupling is enabled (less than 6 cycles, for a L2 latency of 256 cycles) but it is much higher when decoupling is disabled. Second, the load miss latency perceived by an individual thread is slightly longer when more threads are running. Although having more threads effectively reduces the number of stall cycles of each thread, it also increases the miss ratio (due to the larger combined working set) and produces longer bus contention delays, which becomes the - slightly - dominant effect. Third, it is shown in Figure 6-b that when the L2 memory latency is increased from 1 cycle to cycles, the decoupled multithreaded architecture experiences performance drops of less than 3.6% (less than 1.5%, with 4 threads), while the performance degradation observed in all non- -decoupled configurations is greater than 23%. Even for a huge memory latency of 256 cycles, the performance loss of all the decoupled configurations is lower than 39% while it is greater than 79% for the non-decoupled configurations. Fourth, multithreading provides some additional latency tolerance improvements, especially in the non-decoupled configurations, but it is much lower than the latency tolerance provided by decoupling. Some other conclusions can be drawn from Figure 6-c. While multithreading raises the performance curves, decoupling makes them flatter. In other words, while the main effect of L2 Latency (cycles)1030507090110 Perceived Load Miss L2 Latency (cycles) loss (relative tocycle latency) Figure 6-a: Average perceived load miss latency of individual threads. Figure 6-b: Latency tolerance: performance loss is relative to the 1-cycle L2 latency case Figure 6-c: Contribution of decoupling and multithreading to performance.13571 L2 Latency (cycles)1357IPC 4 T, decoupled 3 T, decoupled T, decoupled 4 T, non-decoupled 3 T, non-decoupled T, non-decoupled non-decoupled multithreading is to provide more throughput by exploiting thread level parallelism, the major contribution to memory latency tolerance, which is related to the slope of the curves, comes from decoupling, and this is precisely the specific role that decoupling plays in this hybrid architecture. 4.5. Hardware context reduction and the external bus bandwidth bottleneck Multithreading is a powerful mechanism that highly improves the processor throughput, but it has a cost: it needs a considerable amount of hardware resources. We have run some experiments that illustrate how decoupling reduces the hardware context requirements. We have measured the performance of several configurations having from 1 to 8 contexts, both with a decoupled multithreaded architecture and a non-decoupled multithreaded architecture (see Figure 7-a). While the decoupled configuration achieves the maximum performance with just 3 or 4 threads, the non- decoupled configuration needs 6 threads to achieve similar IPC ratios. One of the traditional claims of the multithreading approach is its ability to sustain a high processor throughput even in systems with a high memory latency. Since hiding a longer latency may require a higher number of contexts and, as it is well known, this has a strong negative impact on the memory performance, the reduction in hardware context requirements obtained by decoupling may become a key factor when L2 memory latency is high. To illustrate this, we have run the previous experiment having a L2 memory latency of 64 cycles. As shown in Figure 7-b,13571 2 3 4 5 6 7 8 Number of Threads1357IPC decoupled non-decoupled Number of Threads1357IPC decoupled non-decoupled Figure 7-a: Decoupling reduces the number of hardware contexts Figure 7-b: Maximum performance without decoupling cannot be reached due to external bus saturation. while the decoupled architecture achieves the maximum performance with just 4 or 5 threads, the non-decoupled architecture cannot reach similar performance with any number of threads, because it would need so many that would they saturate the external L2 bus: the average bus utilization is 89% with 12 threads, and 98% for threads. Moreover, notice that the decoupled architecture requires just 3 threads to achieve about the same performance as the non-decoupled architecture with 12 threads. Thus, decoupling significantly reduces the amount of parallelism required to reach a certain level of performance. The previous result suggests that the external L2 bus bandwidth is a potential bottleneck in this kind of architectures. To further describe its impact, we have measured the performance and bus utilization of several configurations having from 1 to 6 hardware contexts, for three different external bus bandwidths of 8, 16 and 32 bytes/cycle. Results are shown in Figure 8-a and Figure 8-b. For an 8 bytes/cycle bandwidth, the bus becomes saturated when more than 3 threads are running, and performance is degraded beyond this point. To summarize, decoupling and multithreading complement each other to hide memory latency and increase ILP with reduced amounts of thread-level parallelism and low issue logic complexity. Figure 8-a: IPC, for several bus bandwidths Figure 8-b: External L2 bus utilization, for several bus bandwidths2060100 Number of Threads2060100 External Bus Utilization cycles) 8 bytes/cycle Number of Threads1357IPC 8 bytes/cycle 5. Summary and conclusions In this paper we have analized the synergy of multithreading and access/execute decoupling. A multithreaded decoupled architecture aims at taking advantage of the latency hiding effectiveness of decoupling, and the potential of multithreading to exploit ILP. We have analyzed the most important factors that determine its performance and the synergistic effect of both paradigms. A multithreaded decoupled architecture hides efficiently the memory latency: the average load miss latency perceived by an individual thread is less than 6 cycles in the worst case (with 4 threads and a L2 latency of 256 cycles). We have also found that, for L2 latencies lower than cycles, their impact on the performance is quite low: less than 3.5% IPC loss, relative to the 1- cycle latency scenario, and it is quite independent of the number of threads. However, this impact is greater than a 23% IPC loss if decoupling is disabled. This latter fact shows that the main contribution to the memory latency tolerance corresponds to the decoupling mechanism. The architecture reaches maximum performance with very few threads, significantly less than in a non-decoupled architecture. The number of simultaneously active threads supported by the architecture has a significant impact on the hardware chip area (e.g. number of registers, instruction queues) and complexity (e.g. the instruction fetch and issue mechanisms) and consequently in clock cycle. Reducing the number of threads also reduces the cache conflicts and the required memory bandwidth, which is usually one of the potential bottlenecks of a multithreaded architecture. We have shown how the external L2 bus bandwidth becomes a bottleneck when the miss latency is 64 cycles, if decoupling is disabled, preventing it from achieving the maximum performance with any number of threads. In summary, we can conclude that decoupling and multithreading techniques complement each other to exploit instruction level parallelism and to hide memory latency. This particular combination obtains its maximum performance with few threads, has a reduced issue logic complexity, and it is hardly performance degraded by a wide range of L2 latencies. All of these features make it a promising alternative for future increases in clock speed and issue width. 6. --R A Decoupled Access/Execute Architecture for Efficient Access of Structured Data. The Effectiveness of Decoupling. A performance study of software and hardware data prefetching schemes. The Concurrent Execution of Multiple Execution Streams on Super-scalar Processors Alpha 21164 Microprocessor Hardware Reference Manual The Multicluster Architecture: Reducing Cycle Time Through Partitioning. PIPE: A VLSI Decoupled Architecture. 's P6 Uses Decoupled Superscalar Design. Digital 21264 Sets New Standard. HP Make EPIC Disclosure. An Elementary Processor Architecture with Simultaneous Instruction Issuing from Multiple Threads. Designing the TFP Microprocessor. Superscalar Microprocessor Design. A Limitation Study into Access Decoupling. Improving Direct-Mapped Cache Performance by the Addition of a Small Fully-Associative Cache and Prefetch Buffers PEWs: A Decentralized Dynamic Scheduler for ILP Processing. Memory Latency Effects in Decoupled Architectures. The PowerPC 620 Decoupling Integer Execution in Superscalar Processors. Structured Memory Access Architecture. Exploiting Idle Floating-Point Resources For Integer Execution A Study of Branch Prediction Strategies. Decoupled Access/Execute Computer Architectures. Implementation of Precise Interrupts in Pipelined Processors. A Simulation Study of Decoupled Architecture Computers. Multiscalar Processors. ATOM: A System for Building Customized Program Analysis Tools. Standard Performance Evaluation Corporation. An Efficient Algorithm for Exploiting Multiple Arithmetic Units. Compiling and Optimizing for Decoupled Architectures. Exploiting Choice: Instruction Fetch and Issue on an Implementable Simultaneous Multithreading Processor. Simultaneous Multithreading: Maximizing On-Chip Par- allelism MISC: A Multiple Instruction Stream Computer. An Evaluation of the WM Architecture. The Mips R10000 Superscalar Microprocessor. --TR A simulation study of decoupled architecture computers The ZS-1 central processor High-bandwidth data memory systems for superscalar processors An elementary processor architecture with simultaneous instruction issuing from multiple threads Evaluation of the WM architecture MISC The effectiveness of decoupling ATOM Designing the TFP Microprocessor Compiling and optimizing for decoupled architectures Simultaneous multithreading Multiscalar processors Decoupling integer execution in superscalar processors Exploiting choice Complexity-effective superscalar processors Trace processors The multicluster architecture Exploiting idle floating-point resources for integer execution Performance modeling and code partitioning for the DS architecture Improving direct-mapped cache performance by the addition of a small fully-associative cache and prefetch buffers Implementation of precise interrupts in pipelined processors Decoupled access/execute computer architectures The MIPS R10000 Superscalar Microprocessor Memory Latency Effects in Decoupled Architectures A Limitation Study into Access Decoupling The PowerPC 620 microprocessor Lockup-free instruction fetch/prefetch cache organization A study of branch prediction strategies A Cost-Effective Clustered Architecture The Latency Hiding Effectiveness of Decoupled Access/Execute Processors

simultaneous multithreading;instruction-level parallelism;hardware complexity;latency hiding;access/execute decoupling

506024

Fault Detection for Byzantine Quorum Systems.

AbstractIn this paper, we explore techniques to detect Byzantine server failures in asynchronous replicated data services. Our goal is to detect arbitrary failures of data servers in a system where each client accesses the replicated data at only a subset (quorum) of servers in each operation. In such a system, some correct servers can be out-of-date after a write and can therefore, return values other than the most up-to-date value in response to a client's read request, thus complicating the task of determining the number of faulty servers in the system at any point in time. We initiate the study of detecting server failures in this context, and propose two statistical approaches for estimating the risk posed by faulty servers based on responses to read requests.

Introduction Data replication is a well-known means of protecting against data unavailability or corruption in the face of data server failures. When servers can suffer Byzantine (i.e., arbitrary) failures, the foremost approach for protecting data is via state machine replication [Sch90], in which every correct server receives and processes every request in the same order, thereby producing the same output for each re- quest. If the client then accepts a value returned by at least t to t arbitrary server failures can be masked. Numerous systems have been built to support this approach (e.g., [PG89, SESTT92, Rei94, KMM98]). To improve the efficiency and availability of data access while still protecting the integrity of replicated data, the use of quorum systems has been proposed. Quorum preprint of paper to appear in Seventh IFIP International Working Conference on Dependable Computing for Critical Applications (DCCA-7) January, 1999, San Jose, California y Department of Computer Science, University of Texas, Austin, Texas; lorenzo@cs.utexas.edu. This work was funded in part by a NSF CAREER award (CCR- 9734185), a DARPA/SPAWAR grant number N66001-98-8911 and a NSF CISE grant (CDA- z AT&T Labs-Research, Florham Park, New Jersey; dalia@research.att.com x Department of Computer Science, University of Texas, Austin, Texas; tumlin@cs.utexas.edu - Bell Laboratories, Lucent Technologies, Murray Hill, New Jersey; reiter@research.bell- labs.com Alvisi, Malkhi, Pierce, Reiter systems are a family of protocols that allow reads and updates of replicated data to be performed at only a subset (quorum) of the servers. In a t-masking quorum system, the quorums of servers are defined such that any two quorums intersect in at least 2t In a system with a maximum of t faulty servers, if each read and write operation is performed at a quorum, then the quorum used in a read operation will intersect the quorum used in the last preceding write operation in at least t+1 correct servers. With appropriate read and write protocols, this intersection condition ensures that the client is able to identify the correct, up- to-date data [MR97a]. A difficulty of using quorum systems for Byzantine fault tolerance is that detecting responsive but faulty servers is hard. In state machine replication, any server response that disagrees with the response of the majority immediately exposes the failure of the disagreeing server to the client. This property is lost, however, with quorum systems: because some servers remain out of date after any given write, a contrary response from a server in a read operation does not necessarily suggest the server's failure. Therefore, we must design specific mechanisms to monitor the existence of faults in a quorum-replicated system, e.g., to detect whether the number of failures is approaching t. In this paper, we initiate the study of Byzantine fault detection methods for quorum systems by proposing two statistical techniques for estimating the number of server failures in a service replicated using a t-masking quorum system. Both of our methods estimate the total number of faulty servers from responses to a client's read requests executed at a quorum of servers, and are most readily applicable to the threshold quorum construction of [MR97a], in which a quorum is defined as any set of size d n+2t+1 e. The first method has the advantage of requiring essentially no change to the read and write protocols proposed in [MR97a]. The second method does require an alteration of the read and write protocols, but has the advantages of improved accuracy and specific identification of a subset of the faulty servers. Furthermore, the fault identification protocol of the second method is applicable without alteration to all types of t-masking quorum systems, and indeed to other types of Byzantine quorum systems as proposed in [MR97a]. Both methods set an alarm line t a ! t and issue a warning whenever the number of server failures exceeds t a . We show how the system can use information from each read operation to statistically test the hypothesis that the actual number of faults f in the system is at most t a . As we will show, if t a is correctly selected and read operations are frequent, both methods can be expected to issue warnings in a timely fashion, i.e., while it is still the case that f ! t. The service can then be repaired (or at least disabled) before the integrity of the data set is compromised. As an initial investigation into the statistical monitoring of replicated data, this paper adopts a number of simplifying assumptions. First, we perform our statistical analysis in the context of read operations that are concurrent with no write oper- ations, as observing partially completed writes during a read substantially complicates the task of inferring server failures. Second, we assume that clients are correct; distinguishing a faulty server from a server into which a faulty client has Fault Detection for Byzantine Quorum Systems written incorrect data raises issues that we do not consider here. Third, we restrict our attention to techniques that modify the read and write protocols only minimally or not at all and that exploit data gathered from a single read only, without aggregating data across multiple reads. (As we will show in this paper, a surprising amount of information can be obtained without such aggregation.) Each of these assumptions represents an area for possible future research. The goal of our work is substantially different from that of various recent works that have adapted failure detectors [CT96] to solve consensus in distributed systems that can suffer Byzantine failures [MR97b, DS97, KMM97]. These works focus on the specification of abstract failure detectors that enable consensus to be solved. Our goal here is to develop techniques for detecting Byzantine failures specifically in the context of data replicated using quorum systems, without regard to abstract failure detector specifications or the consensus problem. Lin et al. [LRM98] analyze the process of gradual infection of a system by malicious entities. Their analysis attempts to project when failures exceed certain thresholds by extrapolating from observed failures onto the future, on the basis of certain a priori assumptions about the communication patterns of processes and the infection rate of the system. Our methods do not depend on these assumptions, as they do not address the propagation of failures in the system; rather, they attempt to measure the current number of failures at any point in time. To summarize, the contributions of this paper are twofold: we initiate the direction of fault monitoring and detection in the context of Byzantine quorum sys- tems; and we propose two statistical techniques for performing this detection for t-masking quorum systems under the conditions described above. The rest of this paper is organized as follows. In Section 2 we describe our system model and necessary background. In Sections 3-4 we present and analyze our two statistical methods using exact formulae for alarm line placement in relatively small systems. In Section 5 we present an asymptotic analysis for estimating appropriate alarm line placement in larger systems for both methods. We conclude in Section 6. Preliminaries 2.1 System model Our system model is based on a universe U of n data servers. A correct server is one that behaves according to its specification, whereas a faulty server deviates from its specification arbitrarily (Byzantine failure). We denote the maximum allowable number of server failures for the system by t, and the actual number of faulty servers in the system at a particular moment by f . Because our goal in this paper is to detect faulty servers, we stipulate that a faulty server does in fact deviate from its I/O specification, i.e., it returns something other than what its specification would dictate (or it returns nothing, though unresponsive servers are ignored in this paper and are not the target of our detection methods). It is hardly fruitful to attempt to detect "faulty" servers whose visible behavior is consistent with correct execution. Alvisi, Malkhi, Pierce, Reiter Our system model also includes some number of clients, which we assume to be correct. Clients communicate with servers over point-to-point channels. Channels are reliable, in the sense that a message sent between a client and a correct server is eventually received by its destination. In addition, a client can authenticate the channel to a correct server; i.e., if the client receives a message from a correct server, then that server actually sent it. 2.2 Masking quorum systems We assume that each server holds a copy of some replicated variable Z , on which clients can execute write and read operations to change or observe its value, respectively. The protocols for writing and reading Z employ a t-masking quorum system [MR97a, MRW97], i.e., a set of subsets of servers Q ' 2 U such that Intuitively, if each read and write is performed at a quorum of servers, then the use of a t-masking quorum system ensures that a read quorum Q 2 intersects the last write quorum Q 1 in at least t which suffices to enable the reader to determine the last written value. Specifically, we base our methods on threshold masking quorum systems [MR97a], defined by i.e., the quorums are all sets of servers of size d n+2t+1 2 e. These systems are easily seen to have the t-masking property above. We consider the following protocols for accessing the replicated variable Z , which were shown in [MR97a] to give Z the semantics of a safe variable [Lam86]. Each server u maintains a timestamp T u with its copy Z u of the variable Z . A client writes the timestamp when it writes the variable. These protocols require that different clients choose different timestamps, and thus each client c chooses its timestamps from some set T c that does not intersect T c 0 for any other client c 0 . Client operations proceed as follows. Write: For a client c to write the value v to Z , it queries each server in some quorum Q to obtain a set of value/timestamp pairs a timestamp T 2 T c greater than the highest timestamp value in A and greater than any timestamp it has chosen in the past, and updates Z u and T u at each server u in some quorum Q 0 to v and T , respectively. Read: For a client to read a variable Z , it queries each server in some quorum Q to obtain a set of value/timestamp pairs From among all pairs returned by at least t servers in Q, the client chooses the pair !v; T? with the highest timestamp T , and then returns v as the result of the read operation. If there is no pair returned by at least t servers, the result of the read operation is ? (a null value). In a write operation, each server u updates Z u and T u to the received values !v; T? only if T is greater than the present value of T u ; this convention guarantees the seri- Fault Detection for Byzantine Quorum Systems alizability of concurrent writes. As mentioned in Section 1 we consider only reads that are not concurrent with writes. In this case, the read operation will never return ? (provided that the assumed maximum number of failures t is not exceeded). 2.3 Statistical building blocks The primary goal of this paper is to draw conclusions about the number f of faulty servers in the system, specifically whether f exceeds a selected alarm threshold t a , where t, using the responses obtained in the read protocol of the previous subsection. To do this, we make extensive use of a statistical technique called hypothesis testing. To use this technique, we establish two hypotheses about our universe of servers. The first of these is an experimental hypothesis HE that represents a condition to be tested for, e.g., that f exceeds the alarm threshold t a , and the second is a null hypothesis H 0 complementing it. The idea behind hypothesis testing is to examine experimental results (in our case, read operations) for conditions that suggest the truth of the experimental hypothesis, i.e., conditions that would be "highly unlikely" if the null hypothesis were true. We define "highly unlikely" by choosing a rejection level identifying a corresponding region of rejection for H 0 , where the region of rejection is the maximal set of possible results that suggest the truth of HE (and thus the falsity of H 0 ) and whose total probability given H 0 is at most ff. For the purposes of our work, HE will be . (Note that although these hypotheses are not strictly complementary, the region of rejection for H 0 encompasses that of every a , where therefore the rejection level of the truly complementary hypothesis f - t a is bounded by that of H 0 . This treatment of the null hypothesis is a standard statistical procedure.) In this paper we will typically choose t a to be strictly less than the maximum assumed number t of failures in the system, for the reason that it is of little use to detect a dangerous condition after the integrity of the data has been compromised. The "safest" value for t a is 0, but a higher value may be desirable if small numbers of faults are common and countermeasures are expensive. In order for our statistical calculations to be valid, we must be able to treat individual quorums and the intersection between any two quorums as random samples of the universe of servers. Given our focus on a quorum system consisting of all sets of size d n+2t+1 2 e, this can be accomplished by choosing quorums in such a way that each quorum (not containing unresponsive servers) is approximately equally likely to be queried for any given operation. As in any statistical method, there is some possibility of false positives (i.e., alarms sent when the fault level remains below t a ) and false negatives (failure to detect a dangerous fault level before the threshold is exceeded). As we will show, however, the former risk can be kept to a reasonable minimum, while the latter can be made essentially negligible. 1 1 Except in catastrophically unreliable systems. Neither our method nor any other of which we Alvisi, Malkhi, Pierce, Reiter 3 Diagnosis using justifying sets Our first method of fault detection for threshold quorum systems uses the read and write protocols described in Section 2.2. As the random variable for our statistical analysis, we use the size of the justifying set for a read operation, which is the set of servers that return the value/timestamp pair !v; T? chosen by the client in the read operation. The size of the justifying set is at least 2t there are no faulty servers, but can be as small as t. The justifying set may be as large as d n+2t+1 e in the case where the read quorum is the same as the quorum used in the last completed write operation. Suppose that a read operation is performed on the system, and that the size of the justifying set for that read operation is x. We would like to discover whether this evidence supports the hypothesis that the number of faults f in the system exceeds some value t a , where t a ! t. We do so using a formula for the probability distribution for justifying set sizes; this formula is derived as follows. Suppose we have a system of n servers, with a quorum size of q. Given f faulty servers in the system, the probability of exactly j failures in the read quorum can be expressed by a hypergeometric distribution as follows: q\Gammaj Given that the number of failures in the read quorum is j, the probability that there are exactly x correct servers in the intersection between the read quorum and the previous write quorum is formulated as follows: the number of ways of choosing x correct servers from the read quorum is x , and the number of possible previous write quorums that intersect the read quorum in exactly those correct servers (and some number of incorrect ones) is . The probability that the previous write quorum intersects the read quorum in exactly this way is therefore: x To get the overall probability that there are exactly x correct servers in the intersection between the read and most recent write quorums, i.e., that the justifying set size (size) is x, we multiply the conditional probability given j failures in the read quorum by the probability of exactly j failures in the read quorum, and sum the result for f x q\Gammaj (1) are aware will protect against sudden near-simultaneous Byzantine failures in a sufficiently large number (e.g., greater than t) of servers. Fault Detection for Byzantine Quorum Systems This formula expresses the probability that a particular read operation in a t-masking quorum system will have a justifying set size of x given the presence of f faults. For a given rejection level ff, then, the region of rejection for the null hypothesis a is defined as x - highreject, where highreject is the maximum value such highreject X t a x ji t a ji n\Gammat a q\Gammaj The left-hand expression above represents the significance level of the test, i.e., the probability of a false positive (false alarm). If there are in fact failures in the system, the probability of detecting this condition on a single read is: highreject X x q\Gammaj If we denote this value by fl, then the probability that k consecutive reads fail to detect the condition is As shown in the following examples, k need not be very large for this probability to become negligible. Example 1: Consider a system of fault tolerance threshold In order to test whether there are any faults in the system, we set t a = 0, so that the null hypothesis H 0 is and the experimental hypothesis HE is f ? 0. Plugging these numbers into (1) over the full range of x yields the results in Table 1. For all other values of x not shown in Table 1, the probability of a justifying set of size x given 0:071, the region of rejection for defined as x - 53; if a read operation has a justifying set of size 53 or less, the client rejects the null hypothesis and concludes that there are faults in the system. This test has a significance level of that is, there is a probability of 0.019 that the client will detect faults when there are none. (If this level of risk is unacceptable for a particular system, ff can be set to a lower value, thus creating a smaller region of rejection.) Suppose that there are actually f failures in the system. The probability that this experiment will detect the presence of failures during any given read is:X x=26 f x f 101\Gammaf Table shows these values for 1 - f - 20. Although the probability of detecting faults during a given read in this system is relatively low for very small values of f , it would appear that this test is reasonably Alvisi, Malkhi, Pierce, Reiter x 54 .051857 67 9:03 \Theta 10 \Gamma07 Table 1: Probability distribution on justifying set sizes for Example 1 9 .739333 19 .997720 Table 2: Probability of detecting f ? 0 in Example 1 Fault Detection for Byzantine Quorum Systems 9 .130284 Table 3: Probability of detecting f ? 5 in Example 2 powerful. Even for fault levels as low as 4 or 5, a client can reasonably expect to detect the presence of failures within a very few reads; e.g., if then the probability of detecting that f ? t a in only 6 reads is already :921. As the fault levels rise, the probability of such detection within a single read approaches near-certainty. Example 2: Consider a much smaller system consisting of a quorum size q = 46 and a fault tolerance threshold Furthermore, suppose that the administrator of this system has decided that no action is called for if only a few failures occur, so that t a is set at 5 rather than 0. Given 0:05, the region of rejection for the null hypothesis H a is x - 27. The probabilities of detecting this condition for actual values of f between 8 and 12 inclusive are shown in Table 3. As one might expect, error conditions are more difficult to detect when they are more narrowly defined, as the contrast between examples 1 and 2 shows. Even in the latter experiment, however, a client can reasonably expect to detect a serious but non-fatal error condition within a small number of reads. For the probability that the alarm is triggered within six read operations is approximately 96.5 percent. The probability that it is triggered within ten reads is over 99.6 percent. We can therefore reasonably consider this technique to be a useful diagnostic in systems where read operations are significantly more frequent than server failures, particularly if the systems are relatively large. While the ability to detect faulty servers in threshold quorum systems is a step forward, this method leaves something to be desired. It gives little indication of the specific number of faults that have occurred and provides little information toward identifying which servers are faulty. In the next section we present another diagnostic method that addresses both these needs. 4 Diagnosis using quorum markers The diagnostic method presented in this section has two distinct functions. First, it uses a technique similar to that of the previous section to estimate the fault distribution over the whole system, but with greater precision. Second, it pinpoints Alvisi, Malkhi, Pierce, Reiter specific servers that exhibit detectably faulty behavior during a given read. The diagnostic operates on an enhanced version of the read/write protocol for masking quorum systems: the write marker protocol, described below. 4.1 The write marker protocol The write marker protocol uses a simple enhancement to the read/write protocol of Section 2.2: we introduce a write quorum marker field to all variables. That is, for a replicated variable Z , each server u maintains, in addition to Z u and T u , a third value W u , which is the name of the quorum (e.g., an n-bit vector indicating the servers in the quorum) used to complete the write operation in which Z u and T u were last written. The write protocol proceeds as in Section 2.2, except that in the last step, in addition to updating Z u and T u to v and T at each server u in a quorum , the client also updates W u with (the name of) Q 0 . Specifically, to update Z u , T u , and W u at all (correct) servers in Q 0 , the client sends a message containing to each u 2 Q 0 . Because communication is reliable (see Section 2), the writer knows that Z u , T u and W u will be updated at all correct servers in Q 0 . As before, each server u updates Z u , T u , and W u to the received values !v; T; only if T is greater than the present value of T u . The read protocol proceeds essentially as before, except that each server returns the triple !Z in response to a read request. From among all triples returned from at least t servers, the client chooses the triple with the highest timestamp. Below we describe two ways of detecting faults by making use of the set of triples returned by the servers. 4.2 Statistical fault detection Our revised statistical technique uses the quorum markers to determine the set S of servers whose returned values would match the accepted triple in the absence of faults, and the set S 0 of servers whose returned values actually do match that triple. Because of the size-based construction of threshold quorum systems and the random selection of the servers that make up the quorum for a given operation, the set S can be considered a random sample of the servers, of which jS nS 0 j are known to be faulty. Taking a random variable y to be the number of faulty servers in the sample, we can use similar calculations to those in Section 3 to analyze with greater precision the probability that f exceeds t a . As shown in Section 3, the probability of finding y faults in a sample of size s given a universe of size n containing f faults is expressed by the hypergeometric y s\Gammay s Fault Detection for Byzantine Quorum Systems 9 .999660 19 .999999 Table 4: Probability of detecting f ? 0 in Example 3 For a rejection level ff, the region of rejection for the hypothesis a is therefore defined by the lowest value lowreject such that: s y=lowreject t a y n\Gammat a s\Gammay s Again, the left-hand expression represents the parameterized probability of a false alarm. For this method, experiments in which t a = 0 are a degenerate case. The presence of any faults in the intersection set is visible and invalidates the null hypoth- esis; the probability of a false positive in such cases is zero, as the formula above confirms. Likewise, as the number of faults increases, the probability of detecting faults within one or two reads rapidly approaches certainty. Example 3: Consider again the system of servers, with a fault tolerance threshold of quorum size of q = 76, and t a = 0, and suppose that a given read quorum overlaps the previous write quorum in servers (the most likely overlap, with a probability of about 0.21). The probability of alarm on a single read operation for various values of f ! t, is shown in Table 4. A comparison of this table with Table 2 illustrates the dramatically higher precision of the write-marker method over the justifying set method. This precision has additional advantages when t a is set to a value greater than 0. Example 4: Consider again the system of servers, with a fault tolerance threshold of quorum size of q = 46, and t a = 5, and suppose that a given read quorum overlaps the previous write quorum in the most common intersection size servers. The region of rejection for the null hypothesis calculated using the formula above, is y - 5. The probability of alarm on a single read operation for various values of f , t a t, is shown in Table 5. Alvisi, Malkhi, Pierce, Reiter 9 Table 5: Probability of detecting f ? 5 in Example 4 Again, the increased strength of the write-marker method is evident (see Table 3). Like the method presented in Section 3, the write-marker technique also has the advantage of flexibility. If we wish to minimize the risk of premature alarms (i.e., alarms that are sent without the alarm threshold being exceeded) we may choose a smaller ff at the risk of somewhat delayed alarms. In fact, the greater precision of this method decreases the risks associated with such a course: even delayed alarms can be expected to be timely. 4.3 Fault identification The write marker protocol has an even stronger potential as a tool for fault de- tection: it allows the client to identify specific servers that are behaving incorrectly. By keeping a record of this list, the client can thereafter select quorums that do not contain these servers. This allows the system to behave somewhat more efficiently than it would otherwise, as well as gathering the information needed to isolate faulty servers for repair so that the system's integrity is maintained. The fault identification algorithm accepts as input the triples f!Z that the client obtained from servers in the read protocol, as well as the triple that the client chose as the result of the read operation. It then computes the set S n S 0 where is the set of servers that returned in the read operation. The servers in S n S 0 are identified as faulty. Note that the fault identification protocol does not depend in any way on the specific characteristics of threshold quorum systems, and is easily seen to be applicable to masking quorum systems in general. 5 Choosing alarm lines for large systems The analysis of the previous two sections is precise but computationally cumbersome for very large systems. A useful alternative is to estimate the performance of possible alarm lines by means of bound analysis. In this section we present an asymptotic analysis of the techniques of Sections 3 and 4 that shows how to choose an alarm line value for arbitrarily large systems. Fault Detection for Byzantine Quorum Systems Let us denote the read quorum Q, the write quorum Q 0 , the set of faulty servers by B, and the hypothesized size of B (i.e., the alarm line) by t a . We define a random variable which is the justifying set size. We can compute the expectation of X directly. For each server u 62 B define an indicator random variable I u such that I I otherwise. For such u we have P (I since Q and Q 0 are chosen independently. By linearity of expectation, Intuitively, the distribution on X is centered around its expectation and decreases exponentially as X moves farther away from that expectation. Thus, we should be able to show that X grows smaller than its expectation with exponentially decreasing probability. A tempting approach to analyzing this would be to use Chernoff bounds, but these do not directly apply because the selection of individual servers in Q (similarly, Q 0 ) is not independent. In the analysis below, we thus use a more powerful tool, martingales, to derive the anticipated Chernoff-like bound. We bound the probability using the method of bounded differences, by defining a suitable Doob martingale sequence and applying Azuma's inequality (see [MR95, Ch. 4.4] for a good exposition of this technique; Appendix A provides a brief introduction). Here, a Doob martingale sequence of conditional random variables is defined by setting q, to be the expected value of X after i selections are made in each of Q and Q 0 . Then, and it is not difficult to see that jX This yields the following bound (see Appendix A). We use this formula and our desired rejection level ff to determine a ffi such that probability value is our probability of a false alarm and can be diminished by decreasing ff and recalculating ffi. The value defines our region of rejection (see Section 2.3). In order to analyze the probability that our alarm is triggered when the number of faults in the system is t 0 ? t a , we define a second random variable X 0 identical to X except for the revised failure hypothesis. This gives us: An analysis similar to the above provides the following bound: To summarize, these bounds can now be used as follows. For any given alarm line t a , and any desired confidence level ff, we can compute the minimum ffi to satisfy Alvisi, Malkhi, Pierce, Reiter We thus derive the following test: An alarm is triggered whenever the justifying set size is less than (n \Gamma t a ffi. The analysis above guarantees that this alarm will be triggered with false positive probability at most our computed bound ff. If, in fact, f faults occur and f is sufficiently larger than t a , then there exists by the analysis above, the probability of triggering the alarm is greater than 8q . In the case of the write marker protocol, we can tighten the analysis by using the (known) intersection size between Q and Q 0 as follows. Bj. Y has a hypergeometric distribution on s, )=n. The appropriate Doob martingale sequence in this case defines Y i , s, to be the expected value of Y after i selections are made in S. Then, jY and so to set the region of rejection we can use 6 Conclusion In this paper, we have presented two methods for probabilistic fault diagnosis for services replicated using t-masking quorum systems. Our methods mine server responses to read operations for evidence of server failures, and if necessary trigger an alarm to initiate appropriate recovery actions. Both of our methods were demonstrated in the context of the threshold construction of [MR97a], i.e., in which the quorums are all sets of size d n+2t+1e, but our techniques of Section 4 can be generalized to other masking quorum systems, as well. Our first method has the advantage of requiring no modifications to the read and write protocols proposed in [MR97a]. The second method requires minor modifications to these protocols, but also offers better diagnosis capabilities and a precise identification of faulty servers. Our methods are very effective in detecting faulty servers, since faulty servers risk detection in every read operation to which they return incorrect answers. Future work will focus on generalizations of these techniques, as well as uses of these techniques in a larger systems context. In particular, we are presently exploring approaches to react to server failures once they are detected. --R Unreliable failure detectors for reliable distributed sys- tems Muteness detectors for consensus with Byzantine pro- cesses Solving consensus in a Byzantine environment using an unreliable failure detector. The SecureRing protocols for securing group communication. On interprocess communication (part II: algorithms). On the resilience of multicasting strategies in a failure-propagating environment Randomized algorithms. Byzantine quorum systems. Unreliable intrusion detection in distributed computation. The load and availability of Byzantine quorum systems. Secure agreement protocols: Reliable and atomic group multicast in Rampart. Reliable scheduling in a TMR database system. Implementing fault-tolerant services using the state machine ap- proach: A tutorial Principal features of the VOLTAN family of reliable node architectures for distributed systems. --TR Randomized algorithms A <inline-equation> <f> <rad><rcd>N</rcd></rad></f> </inline-equation> algorithm for mutual exclusion in decentralized systems Unreliable failure detectors for reliable distributed systems Synchronous Byzantine quorum systems Probabilistic Byzantine quorum systems The Load and Availability of Byzantine Quorum Systems Intrusion Detection An Architecture for Survivable Coordination in Large Distributed Systems Fault Detection for Byzantine Quorum Systems Unreliable Intrusion Detection in Distributed Computations A comparison connection assignment for diagnosis of multiprocessor systems --CTR Andreas Haeberlen , Petr Kouznetsov , Peter Druschel, The case for Byzantine fault detection, Proceedings of the 2nd conference on Hot Topics in System Dependability, p.5-5, November 08, 2006, Seattle, WA Meng Yu , Peng Liu , Wanyu Zang, Specifying and using intrusion masking models to process distributed operations, Journal of Computer Security, v.13 n.4, p.623-658, July 2005

byzantine fault tolerance;replicated data;fault detection;quorum systems

506154

Task assignment with unknown duration.

We consider a distributed server system and ask which policy should be used for assigning jobs (tasks) to hosts. In our server, jobs are not preemptible. Also, the job's service demand is not known a priori. We are particularly concerned with the case where the workload is heavy-tailed, as is characteristic of many empirically measured computer workloads. We analyze several natural task assignment policies and propose a new one TAGS (Task Assignment based on Guessing Size). The TAGS algorithm is counterintuitive in many respects, including load unbalancing, non-work-conserving, and fairness. We find that under heavy-tailed workloads, TAGS can outperform all task assignment policies known to us by several orders of magnitude with respect to both mean response time and mean slowdown, provided the system load is not too high. We also introduce a new practical performance metric for distributed servers called server expansion. Under the server expansion metric, TAGS significantly outperforms all other task assignment policies, regardless of system load.

Introduction In recent years, distributed servers have become commonplace because they allow for increased computing power while being cost-effective and easily scalable. In a distributed server system, requests for service (tasks) arrive and must be assigned to one of the host machines for processing. The rule for assigning tasks to host machines is known as the task assignment policy. The choice of the task assignment policy has a significant effect on the performance perceived by users. Designing a distributed server system often comes down to choosing the "best" task assignment policy for the given model and user requirements. The question of which task assignment policy is "best" is an age-old question which still remains open for many models. In this paper we consider the particular model of a distributed server system in which tasks are not preemptible - i.e. we are concerned with applications where context switches are too costly. For example, one such application is batch computing environments where the hosts themselves are parallel processors and the tasks are parallel. Context switching between tasks involves reloading all the processors and memory to return them to the state before the context switch. Because context switching is so expensive in this environment, tasks are always simply run to completion. Note, the fact that context switches are too expensive does not preclude the possibility of killing a job and restarting it from scratch. We assume furthermore that no a priori information is known about the task at the time when the task arrives. In particular, the service demand of the task is not known. We assume all hosts are identical and there is no cost (time required) for assigning tasks to hosts. Figure 1 is one illustration of a distributed server. In this illustration, arriving tasks are immediately dispatched by the central dispatcher to one of the hosts and queue up at the host waiting for service, where they are served in first-come-first-served (FCFS) order. Observe however that our model in general does not preclude the possibility of having a central queue at the dispatcher where tasks might wait before being dispatched. It also does not preclude the possibility of an alternative scheduling discipline at the hosts, so long as that scheduling discipline does not require preempting tasks and does not rely on a priori knowledge about tasks. Our main performance goal, in choosing a task assignment policy, is to minimize mean waiting time and more importantly mean slowdown. A task's slow-down is its waiting time divided by its service demand. All means are per-task averages. We consider mean slowdown to be more important than mean waiting time because it is desirable that a task's delay be proportional to its size. That is, in a system in which task sizes are highly variable, users are likely to anticipate short delays for short tasks, and are likely to tolerate long delays for longer tasks. Later in the paper we introduce a new performance metric, called server expansion which is related to mean slowdown. A secondary performance goal is fairness. We adopt the standard definition of fairness that says all tasks, large or small, should experience the same expected slowdown. In particular, large tasks shouldn't be penalized - slowed down by a greater factor than are small tasks. 1 Consider some task assignment policies commonly proposed for distributed server systems: In the Random task assignment policy, an incoming task is sent to Host i with probability 1=h, where h is the number of hosts. This policy equalizes the expected number of tasks at each host. In Round-Robin task as- signment, tasks are assigned to hosts in a cyclical fashion with the ith task being assigned to Host i mod h. This policy also equalizes the expected number of tasks at each host, and has slightly less variability in interarrival times than does Random. In Shortest-Queue task assignment, an incoming task is immediately dispatched to the host with the fewest number of tasks. This policy has the benefit of trying to equalize the instantaneous number of tasks at each host, rather than just the expected number of tasks. All the above policies have the property that the tasks arriving at each host are serviced in FCFS order. The literature tells us that Shortest-Queue is in fact the best task assignment policy in a model where the following conditions are met: (1) there is no a priori knowledge about tasks, (2) tasks are not preemptible, (3) each host services tasks in a FCFS order, (4) incoming tasks are immediately dispatched to a host, and (5) the task size distribution is Exponential (see Section 2). If one removes restriction (4), it is possible to do even better. What we'd really like to do is send a task to the host which has the least total outstanding work (work is the sum of the task sizes at the host) because that host would afford the task the smallest waiting time. However, we don't know a priori which host currently has the least work, since we don't know task sizes. It turns out this is actually easy to get around: we simply hold all tasks at the dispatcher in a FCFS queue, and only when a host is free does it request the next task. It is easy to prove that this holding method is exactly equivalent to immediately dispatching arriving tasks to the host with least outstanding work (see [6] for a proof and Figure 2 for an illustration). We will refer to this policy as Least-Work-Remaining since it has the effect of sending each task to the host with the currently least remaining work. Observe that Least-Work-Remaining comes closest to obtaining instantaneous load balance. It may seem that Least-Work-Remaining is the best possible task assignment policy. Previous literature shows that Least-Work-Remaining outperforms all of the above previously-discussed policies under very general conditions (see Section 2). Previous literature also suggests that Least-Work-Remaining 1 For example, Processor-Sharing (which requires infinitely-many preemptions) is ultimately fair in that every task experiences the same expected slowdown. DISPATCHER OUTSIDE ARRIVALS FCFS FCFS FCFS FCFS Figure 1: Illustration of a distributed server. DISPATCHER OUTSIDE ARRIVALS FCFS FCFS FCFS FCFS Send to host with least work FCFS (a) (b) Figure 2: Two equivalent ways of implementing the Least-Work-Remaining task assignment policy. (a) Shows incoming tasks immediately being dispatched to the host with the least remaining work, but this requires knowing a priori the sizes of the tasks at the hosts. (b) Shows incoming tasks pooled at a FCFS queue at the dispatcher. There are no queues at the individual hosts. Only when a host is free does it request the next task. This implementation does not require a priori knowledge of the task sizes, yet achieves the same effect as (a). may be the optimal (best possible) task assignment policy in the case where the task size distribution is Exponential (see Section 2 for a detailed statement of the previous literature). But what if task size distribution is not Exponential? We are motivated in this respect by the increasing evidence for high variability in task size distri- butions, as seen in many measurements of computer workloads. In particular, measurements of many computer workloads have been shown to fit heavy-tailed distributions with very high variance, as described in Section 3 - much higher variance than that of an Exponential distribution. Is there a better task assignment policy than Least-Work-Remaining when the task size variability is characteristic of empirical workloads? In evaluating various task assignment policies, we will be interested in understanding the influence of task size variability on the decision of which task assignment policy is best. For analytical tractability, we will assume that the arrival process is Poisson - our simulations indicate that the variability in the arrival process is much less critical to choosing a task assignment policy than is the variability in the task size distribution. In this paper we propose a new algorithm called TAGS - Task Assignment by Guessing Size which is specifically designed for high variability workloads. We will prove analytically that when task sizes show the degree of variability characteristic of empirical (measured) workloads, the TAGS algorithm can out-perform all the above mentioned algorithms by several orders of magnitude. In fact, we will show that the more heavy-tailed the task size distribution, the greater the improvement of TAGS over the other task assignment algorithms. The above improvements are contingent on the system load not being too high. 2 In the case where the system load is high, we show that all the task assignment policies have such poor performance that they become impractical, and TAGS is especially negatively affected. In practice, if the system load is too high to achieve reasonable performance, one adds new hosts to the server (without increasing the outside arrival rate), thus dropping the system load, until the system behaves as desired. We refer to the "number of new hosts which must be added" above as the server expansion requirement. We will show that TAGS outperforms all the previously-mentioned task assignment policies with respect to the server expansion metric (i.e., given any initial load, TAGS requires far fewer additional hosts to perform well). We will describe three flavors of TAGS. The first, called TAGS-opt-meanslowdown is designed to minimize mean slowdown. The second, called TAGS-opt-meanwaitingtime 2 For a distributed server, system load is defined as follows: System load = Outside arrival rate \Delta Mean task size = Number of hosts For example, a system with 2 hosts and system load .5 has same outside arrival rate as a system with 4 hosts and system load .25. Observe that a 4 host system with system load ae has twice the outside arrival rate of a 2 host system with system load ae. is designed to minimize mean waiting time. Although very effective, these algorithms are not fair in their treatment of tasks. The third flavor, called TAGS-opt-fairness, is designed to optimize fairness. While managing to be fair, TAGS-opt-fairness still achieves mean slowdown and mean waiting time close to the other flavors of TAGS. Section 2 elaborates in more detail on previous work in this area. Section 3 provides the necessary background on measured task size distributions and heavy-tails. Section 4 describes the TAGS algorithm and all its flavors. Section 5 shows results of analysis for the case of 2 hosts and Section 6 shows results of analysis for the multiple-host case. Section 7 explores the effect of less-variable job size distributions. Lastly, we conclude in Section 8. Details on the analysis of TAGS are described in the Appendix. Previous Work on Task Assignment 2.1 Task assignment with no preemption The problem of task assignment in a model like ours (no preemption and no a priori knowledge) has been extensively studied, but many basic questions remain open. One subproblem which has been solved is that of task assignment under the restriction that all tasks be immediately dispatched to a host upon arrival and each host services its tasks in FCFS order. Under this restricted model, it has been shown that when the task size distribution is exponential and the arrival process is Poisson, then the Shortest-Queue task assignment policy is optimal, Winston [19]. In this result, optimality is defined as maximizing the discounted number of tasks which complete by some fixed time t. Ephremides, Varaiya, and Walrand [5] showed that the Shortest-Queue task assignment policy also minimizes the expected total time for the completion of all tasks arriving by some fixed time t, under an exponential task size distribution and arbitrary arrival process. The actual performance of the Shortest-Queue policy is not known exactly, but the mean response time is approximated by Nelson and Phillips [11], [12]. Whitt has shown that as the variability of the task size distribution grows, the Shortest-Queue policy is no longer optimal [18]. Whitt does not suggest which policy is optimal. The scenario has also been considered, under the same restricted model described in the above paragraph, but where the ages (time in service) of the tasks currently serving are known, so that it is possible to compute an arriving task's expected delay at each queue. In this scenario, Weber [17] considers the Shortest-Expected-Delay rule which sends each task to the host with the least expected work (note the similarity to the Least-Work-Remaining policy). Weber shows that this rule is optimal for task size distributions with increasing failure rate (including Exponential). Whitt [18] shows that there exist task size distributions for which this rule is not optimal. Wolff, [20] has proven that Least-Work-Remaining is the best possible task assignment policy out of all policies which do not make use of task size. This result holds for any distribution of task sizes and for any arrival process. Another model which has been considered is the case of no preemption but where the size of each task is known at the time of arrival of the task. Within this model, the SITA-E algorithm (see [7]) has been shown to outperform the Random, Round-Robin, Shortest-Queue, and Least-Work-Remaining algorithms by several orders of magnitude when the task size distribution is heavy- tailed. In contrast to SITA-E, the TAGS algorithm does not require knowledge of task size. Nevertheless, for not-too-high system loads (! :5), TAGS improves upon the performance of SITA-E by several orders of magnitude for heavy-tailed workloads. 2.2 When preemption is allowed and other generalizations Throughout this paper we maintain the assumption that tasks are not pre- emptible. That is, once a task starts running, it can not be stopped and re- continued where it left off. By contrast there exists considerable work on the very different problem where tasks are preemptible (see [8] for many citations). Other generalizations of the task assignment problem include the scenario where the hosts are heterogeneous or there are multiple resources under contention The idea of purposely unbalancing load has been suggested previously in [3] and in [1], under different contexts from our paper. In both these papers, it is assumed that task sizes are known a priori. In [3] a distributed system with preemptible tasks is considered. It is shown that in the preemptible model, mean waiting time is minimized by balancing load, however mean slowdown is minimized by unbalancing load. In [1], real-time scheduling is considered where tasks have firm deadlines. In this context, the authors propose "load profiling," which "distributes load in such a way that the probability of satisfying the utilization requirements of incoming tasks is maximized." 3 Heavy Tails As described in Section 1, we are concerned with how the distribution of task sizes affects the decision of which task assignment policy to use. Many application environments show a mixture of task sizes spanning many orders of magnitude. In such environments there are typically many small tasks, and fewer large tasks. Much previous work has used the exponential distribution to capture this variability, as described in Section 2. However, recent measurements indicate that for many applications the exponential distribution is a poor model and that a heavy-tailed distribution is more accurate. In general a heavy-tailed distribution is one for which 2. The simplest heavy-tailed distribution is the Pareto distribu- tion, with probability mass function and cumulative distribution function A set of task sizes following a heavy-tailed distribution has the following properties 1. Decreasing failure rate: In particular, the longer a task has run, the longer it is expected to continue running. 2. Infinite variance (and if ff - 1, infinite mean). 3. The property that a very small fraction (! 1%) of the very largest tasks make up a large fraction (half) of the load. We will refer to this important property throughout the paper as the heavy-tailed property. The lower the parameter ff, the more variable the distribution, and the more pronounced is the heavy-tailed property, i.e. the smaller the fraction of large tasks that comprise half the load. As a concrete example, Figure 3 depicts graphically on a log-log plot the measured distribution of CPU requirements of over a million UNIX processes, taken from paper [8]. This distribution closely fits the curve PrfProcess Lifetime ? In [8] it is shown that this distribution is present in a variety of computing en- vironments, including instructional, research, and administrative environments. In fact, heavy-tailed distributions appear to fit many recent measurements of computing systems. These include, for example: ffl Unix process CPU requirements measured at Bellcore: 1 - ff - 1:25 [10]. ffl Unix process CPU requirements, measured at UC Berkeley: ff - 1 [8]. ffl Sizes of files transferred through the Web: 1:1 - ff - 1:3 [2, 4]. ffl Sizes of files stored in Unix filesystems: [9]. ffl I/O times: [14]. ffl Sizes of FTP transfers in the Internet: :9 - ff - 1:1 [13]. In most of these cases where estimates of ff were made, ff tends to be close to 1, which represents very high variability in task service requirements. In practice, there is some upper bound on the maximum size of a task, because files only have finite lengths. Throughout this paper, we therefore model task sizes as being generated i.i.d. from a distribution that follows a power law, but has an upper bound - a very high one. We refer to this distribution as a Bounded Pareto. It is characterized by three parameters: ff, the exponent of the power law; k, the smallest possible observation; and p, the largest possible observation. The probability mass function for the Bounded Pareto B(k; p; ff) is defined as: In this paper, we will vary the ff-parameter over the range 0 to 2 in order to observe the effect of changing variability of the distribution. To focus on the effect of changing variance, we keep the distributional mean fixed (at 3000) and the maximum value fixed (at which correspond to typical values taken from [2]. In order to keep the mean constant, we adjust k slightly as ff Note that the Bounded Pareto distribution has all its moments finite. Thus, it is not a heavy-tailed distribution in the sense we have defined above. How- ever, this distribution will still show very high variability if k - p. For example Figure 4 (right) shows the second moment E \Psi of this distribution as a function of ff for is chosen to keep E fXg constant at 3000, 1500). The figure shows that the second moment explodes exponentially as ff declines. Furthermore, the Bounded Pareto distribution also still exhibits the heavy-tailed property and (to some extent) the decreasing failure rate property of the unbounded Pareto distribution. We mention these properties because they are important in determining our choice of the best task assignment policy. Distribution of process lifetimes (log plot) (fraction of processes with duration > T) Duration (T secs.)1/2 Figure 3: Measured distribution of UNIX process CPU lifetimes, taken from [HD97]. Data indicates fraction of jobs whose CPU service demands exceed T seconds, as a function of T . power law w/ exponent Second Moment of Bounded Pareto Distribution alpha Figure 4: Parameters of the Bounded Pareto Distribution (left); Second Moment of as a function of ff, when E OUTSIDE ARRIVALS Figure 5: Illustration of the flow of tasks in the TAGS algorithm. 4 The TAGS algorithm This section describes the TAGS algorithm. Let h be the number of hosts in the distributed server. Think of the hosts as being numbered: h. The ith host has a number s i associated with it, where s TAGS works as shown in Figure 5: All incoming tasks are immediately dispatched to Host 1. There they are serviced in FCFS order. If they complete before using up s 1 amount of CPU, they simply leave the system. However, if a task has used s 1 amount of CPU at Host 1 and still hasn't completed, then it is killed (remember tasks cannot be preempted because that is too expensive in our model). The task is then put at the end of the queue at Host 2, where it is restarted from scratch 3 . Each host services the tasks in its queue in FCFS order. If a task at host i uses up s i amount of CPU and still hasn't completed it is killed and put at the end of the queue for Host i + 1. In this way, the TAGS algorithm "guesses the size" of each task, hence the name. The TAGS algorithm may sound counterintuitive for a few reasons: First of all, there's a sense that the higher-numbered hosts will be underutilized and the 3 Note, although the task is restarted, it is still the same task, of course. We are therefore careful in our analysis not to assign it a new service requirement. first host overcrowded since all incoming tasks are sent to Host 1. An even more vital concern is that the TAGS algorithm wastes a large amount of resources by killing tasks and then restarting them from scratch. 4 There's also the sense that the big tasks are especially penalized since they're the ones being restarted. TAGS comes in 3 flavors; these only differ in how the s i 's are chosen. In TAGS-opt-meanslowdown, the s i 's are chosen so as to optimize mean slowdown. In TAGS-opt-meanwaitingtime, the s i 's are chosen so as to optimize mean waiting time. As we'll see, TAGS-opt-meanslowdownand TAGS-opt-meanwaitingtime are not necessarily fair. In TAGS-opt-fairness the s i 's are chosen so as to optimize fairness. Specifically, the tasks whose final destination is Host i experience the same expected slowdown under TAGS-opt-fairness as do the tasks whose final destination is Host j, for all i and j. TAGS may seem reminiscent of multi-level feedback queueing, but they are not related. In multi-level feedback queueing there is only a single host with many virtual queues. The host is time-shared and tasks are preemptible. When a task uses some amount of service time it is transferred (not killed and restarted) to a lower priority queue. Also, in multi-level feedback queueing, the tasks in that lower priority queue are only allowed to run when there are no tasks in any of the higher priority queues. 5 Analysis and Results and For the Case of 2 Hosts This section contains the results of our analysis of the TAGS task assignment policy and other task assignment policies. In order to clearly explain the effect of the TAGS algorithm, we limit the discussion in this section to the case of 2 hosts. In this case we refer to the tasks whose final destination is Host 1 as the small tasks and the tasks whose final destination is Host 2 as the big tasks. Until Section 5.3, we will always assume the system load is 0:5 and there are 2 hosts. In Section 5.3, we will consider other system loads, but still stick to the case of 2 hosts. Finally, in Section 6 we will consider distributed servers with multiple hosts. We evaluate several task assignment policies, all as a function of ff, where ff is the variance-parameter for the Bounded Pareto task size distribution, and ff ranges between 0 and 2. Recall from Section 3 that the lower ff is, the higher the variance in the task size distribution. Recall also that empirical measurements of task size distributions often show ff - 1. 4 My dad, Micha Harchol, would add that there's also the psychological concern of what the angry user might do when he's told his task's been killed to help the general good. We will evaluate the Random, Least-Work-Remaining, and TAGS policies. The Round-Robin policy (see Section 1) will not be evaluated directly because we showed in a previous paper [7] that Random and Round-Robin have almost identical performance. As we'll explain in Section 5.1, our analysis of Least-Work-Remaining is only an approximation, however we have confidence in this approximation because our extensive simulation in paper [7] showed it to be quite accurate in this setting. As we'll discuss in Section 5.1, our analysis of TAGS is also an approximation, though to a lesser degree. Figure 6(a) below shows mean slowdown under TAGS-opt-slowdown as compared with the other task assignment policies. The y-axis is shown on a log scale. Observe that for very high ff, the performance of all the task assignment policies is comparable and very good, however as ff decreases, the performance of all the policies degrades. The Least-Work-Remaining policy consistently outperforms the Random policy by about an order of magnitude, however the TAGS-opt-slowdown policy offers several orders of magnitude further improvement: At 1:5, the TAGS-opt-slowdown policy outperforms the Least-Work-Remaining policy by 2 orders of magnitude; at ff - 1, the TAGS-opt-slowdown policy outperforms the Least-Work-Remaining policy by over 4 orders of magnitude; at :4 the the TAGS-opt-slowdown policy out-performs the Least-Work-Remaining policy by over 9 orders of magnitude, and this increases to 15 orders of magnitude for Figures 6(b) and (c) show mean slowdown of TAGS-opt-waitingtime and TAGS-opt-fairness, respectively, as compared with the other task assignment policies. Since TAGS-opt-waitingtime is optimized for mean waiting time, rather than mean slowdown, it is understandable that its performance improvements with respect to mean slowdown are not as dramatic as those of TAGS-opt-slowdown. However, what's interesting is that the performance of TAGS-opt-fairness is very close to that of TAGS-opt-slowdownand yet TAGS-opt-fairness has the additional benefit of fairness. Figure 7 is identical to Figure 6 except that in this case the performance metric is mean waiting time, rather than mean slowdown. Again the TAGS al- gorithm, especially TAGS-opt-waitingtime, shows several orders of magnitude improvement over the other task assignment policies. Why does the TAGS algorithm work so well? Intuitively, it seems that Least-Work-Remaining should be the best performer, since Least-Work-Remaining sends each task to where it will individually experience the lowest waiting time. The reason why TAGS works so well is 2-fold: The first part is variance reduction (Section 5.1) and the second part is load unbalancing (Section 5.2). using TAGS-opt-slowdown: 2 hosts, load .5 alpha . Least-Work-Left Approx. (b) Results: MEAN SLOWDOWN using TAGS-opt-waitingtime: 2 hosts, load .5 alpha . Least-Work-Left Approx. (c) using TAGS-opt-fairness: 2 hosts, load .5 alpha . Least-Work-Left Approx. TAGS-opt-fairness Figure Mean slowdown for distributed server with 2 hosts and system using TAGS-opt-slowdown: 2 hosts, load .5 alpha . Least-Work-Left Approx. TAGS-opt-fairness (b) Results: MEAN WAITING TIME using TAGS-opt-waitingtime: 2 hosts, load .5 alpha . Least-Work-Left Approx. TAGS-opt-fairness (c) using TAGS-opt-fairness: 2 hosts, load .5 alpha . Least-Work-Left Approx. TAGS-opt-fairness Figure 7: Mean waiting time for distributed server with 2 hosts and system 5.1 Variance Reduction Variance reduction refers to reducing the variance of task sizes that share the same queue. Intuitively, variance reduction is important for improving performance because it reduces the chance of a small task getting stuck behind a big task in the same queue. This is stated more formally in Theorem 1 below, which is derived from the Pollaczek-Kinchin formula. Theorem 1 Given an M/G/1 FCFS queue, where the arrival process has rate -, X denotes the service time distribution, and ae denotes the utilization fXg). Let W be a task's waiting time in queue, S be its slowdown, and Q be the queue length on its arrival. Then, Proof: The slowdown formulas follow from the fact that W and X are independent for a FCFS queue, and the queue size follows from Little's formula. Observe that every metric for the simple FCFS queue is dependent on \Psi , the second moment of the service time. Recall that if the workload is heavy-tailed, the second moment of the service time explodes, as shown in Figure 4. We now discuss the effect of high variability in task sizes on a distributed server system under the various task assignment policies. Random Task Assignment The Random policy simply performs Bernoulli splitting on the input stream, with the result that each host becomes an independent queue. The load at the ith host, ae i , is equal to the system load, ae. The arrival rate at the ith host is 1=h-fraction of the total outside arrival rate. Theorem 1 applies directly, and all performance metrics are proportional to the second moment of B(k; p; ff). Performance is generally poor because the second moment of the B(k; p; ff) is high. Round Robin The Round Robin policy splits the incoming stream so each host sees an E h =B(k; p; ff)=1 queue, with utilization ae This system has performance close to the Random policy since it still sees high variability in service times, which dominates performance. Least-Work-Remaining The Least-Work-Remaining policy is equivalent to an M/G/h queue, for which there exist known approximations, [16],[21]: QM=G=h QM=M=h where X denotes the service time distribution, and Q denotes queue length. What's important to observe here is that the mean queue length, and therefore the mean waiting time and mean slowdown, are all proportional to the second moment of the service time distribution, as was the case for the Random and Round-Robin task assignment policies. In fact, the performance metrics are all proportional to the squared coefficient of variation (C ) of the service time distribution. TAGS The TAGS policy is the only one which reduces the variance of task sizes at the individual hosts. Let p i be the fraction of tasks whose final destination is Host i. Consider the tasks which queue at Host i: First there are those tasks which are destined for Host i. Their task size distribution is B(s because the original task size distribution is a Bounded Pareto. Then there are the tasks which are destined for hosts numbered greater than i. These tasks are all capped at size s i . Thus the second moment of the task size distribution at Host i is lower than the second moment of the original B(k; p; ff) distribution (for all hosts except the highest-numbered host, it turns out). The full analysis of the TAGS policy is presented in the Appendix and is relatively straightforward except for one point which we have to fudge and which we explain now: For analytic convenience, we need to be able to assume that the tasks arriving at each host form a Poisson Process. This is of course true for Host 1. However the arrivals at Host i are those departures from Host exceed size s They form a less bursty process than a Poisson Process since they are spaced apart by at least s i\Gamma1 . Throughout our analysis of TAGS, we make the assumption that the arrival process into Host i is a Poisson Process. 5.2 Load Unbalancing The second reason why TAGS performs so well has to do with "load unbalancing." Observe that all the other task assignment policies we described specifically try to balance load at the hosts. Random and Round-Robin balance the expected load at the hosts, while Least-Work-Remaining goes even further in trying to balance the instantaneous load at the hosts. In TAGS we do the opposite. Figure 8 shows the load at Host 1 and the load at Host 2 for TAGS-opt-slowdown, TAGS-opt-waitingtime, and TAGS-opt-fairness as a function of ff. Observe that all 3 flavors of TAGS (purposely) severely underload Host 1 when ff is low but for higher ff actually overload Host 1 somewhat. In the middle range, ff - 1, the load is balanced in the two hosts. We first explain why load unbalancing is desirable when optimizing overall mean slowdown of the system. We will later explain what happens when optimizing fairness. To understand why it is desirable to operate at unbalanced loads, we need to go back to the heavy-tailed property. The heavy-tailed property says that when a distribution is very heavy-tailed (very low ff), only a miniscule fraction of all tasks - the very largest ones - are needed to make up more than half the total load. As an example, for the case turns out that less than 10 \Gamma6 fraction of all tasks are needed to make up half the load. In fact not many more tasks, still less than 10 \Gamma4 fraction of all tasks, are needed to make up :99999 fraction of the load. This suggests a load game that can be played: We choose the cutoff point (s 1 ) such that most tasks fraction) have Host 1 as their final destination, and only a very few tasks (the largest fraction of all tasks) have Host 2 as their final destination. Because of the heavy-tailed property, the load at Host 2 will be extremely high (.99999) while the load at Host 1 will be very low (.00001). Since most tasks get to run at such reduced load, the overall mean slowdown is very low. When the distribution is a little less heavy-tailed, e.g., ff - 1, we can't play this load unbalancing game as well. Again, we would like to severely underload Host 1 and send .999999 fraction of the load to go to Host 2. Before we were able to do this by making only a very small fraction of all tasks (! 10 \Gamma4 fraction) go to Host 2. However now that the distribution is not as heavy-tailed, a larger fraction of tasks must have Host 2 as its final destination to create very high load at Host 2. But this in turn means that tasks with destination Host 2 count more in determining the overall mean slowdown of the system, which is bad since tasks with destination Host 2 experience larger slowdowns. Thus we can only afford to go so far in overloading Host 2 before it turns against us. When get to ff ? 1, it turns out that it actually pays to overload Host 1 a little. This seems counter-intuitive, since Host 1 counts more in determining the overall mean slowdown of the system because the fraction of tasks with destination Host 1 is greater. However, the point is that now it is impossible to create the wonderful state where almost all tasks are on Host 1 and yet Host 1 is underloaded. The tail is just not heavy enough. No matter how we choose the cutoff, a significant portion of the tasks will have Host 2 as their destination. Thus Host 2 will inevitably figure into the overall mean slowdown and so we need to keep the performance on Host 2 in check. To do this, it turns out we need to slightly underload Host 2, to make up for the fact that the task size variability is so much greater on Host 2 than on Host 1. The above has been an explanation for why load unbalancing is important with respect to optimizing the system mean slowdown. However it is not at all clear why load unbalancing also optimizes fairness. Under TAGS-opt-fairness, 20.10.30.50.70.9Loads at hosts under TAGS-opt-slowdown: 2 hosts, load .5 alpha host 1 - Load host 2 (b) 20.10.30.50.70.9Loads at hosts under TAGS-opt-waitingtime: 2 hosts, load .5 alpha host 1 - Load host 2 (c) 20.10.30.50.70.9Loads at hosts under TAGS-opt-fairness: 2 hosts, load .5 alpha host 1 - Load host 2 Figure 8: Load at Host 1 as compared with Host 2 in a distributed server with 2 hosts and system load .5 under (a) TAGS-opt-slowdown, (b) TAGS-opt-waitingtime, and (c) TAGS-opt-fairness. Observe that for very low ff, Host 1 is run at load close to zero, and Host 2 is run at load close to 1, whereas for high ff, Host 1 is somewhat overloaded. (a) System load 0:3 using TAGS-opt-slowdown: 2 hosts, load .3 alpha . Least-Work-Left Approx. (b) System load 0:5 Results: MEAN SLOWDOWN using TAGS-opt-slowdown: 2 hosts, load .5 alpha . Least-Work-Left Approx. (c) System load 0:7 using TAGS-opt-slowdown: 2 hosts, load .7 alpha . Least-Work-Left Approx. Figure 9: Mean slowdown under TAGS-opt-slowdown in a distributed server with 2 hosts with system load (a) 0:3, (b) 0:5, and (c) 0:7. In each figure the mean slowdown under TAGS-opt-slowdown is compared with the performance of Random and Least-Work-Remaining. Observe that in all the figures TAGS outperforms the other task assignment policies under all ff. However TAGS is most effective at lower system loads. the mean slowdown experienced by the small tasks is equal to the mean slowdown experienced by the big tasks. However it seems in fact that we're treating the big tasks unfairly on 3 counts: 1. The small tasks run on Host 1 which has very low load (for low ff). 2. The small tasks run on Host 1 which has very low E 3. The small tasks don't have to be restarted from scratch and wait on a second line. So how can it possibly be fair to help the small tasks so much? The answer is simply that the small tasks are small. Thus they need low waiting times to keep their slowdown low. Big tasks on the other hand can afford a lot more waiting time. They are better able to amortize the punishment over their long lifetimes. It is important to mention, though, that this would not be the case for all distributions. It is because our task size distribution for low ff is so heavy-tailed that the big tasks are truly elephants (way bigger than the smalls) and thus can afford to suffer more. 5 5.3 Different Loads Until now we have studied only the model of a distributed server with two hosts and system load :5. In this section we consider the effect of system load on the performance of TAGS. We continue to assume a 2 host model. Figure 9 shows the performance of TAGS-opt-slowdown on a distributed server with 2 hosts run at system load (a) 0:3, (b) 0:5, and (c) 0:7. In all three figures TAGS-opt-slowdown improves upon the performance of Least-Work-Remaining and Random under the full range of ff, however the improvement of TAGS-opt-slowdown is much better when the system is more lightly loaded. In fact, all the task assignment policies improve as the system load is dropped, however the improvement in TAGS is the most dramatic. In the case where the system load is 0:3, TAGS-opt-slowdown improves upon Least-Work-Remaining by over 4 orders of magnitude at by 6 or 7 orders of magnitude when and by almost 20 orders of magnitude when When the system load is 0:7 on the other 5 It may interest the reader to understand the degree of unfairness exhibited by TAGS-opt-slowdown and TAGS-opt-waitingtime. For TAGS-opt-slowdown, our analysis shows that the expected slowdown of the big tasks always exceeds that of the small tasks and the ratio increases exponentially as ff drops, so that at EfSlowdown(smalls)g, and at In contrast, for TAGS-opt-waitingtime, the expected slowdown of the big tasks is approximately equal to that of the small tasks until ff drops below 1, at which point the expected slowdown of the big tasks drops way below that of the small tasks, the ratio of bigs to smalls decreasing superexponentially as ff drops. hand, TAGS-opt-slowdown behaves comparably to Least-Work-Remaining for most ff and only improves upon Least-Work-Remaining in the narrower range of however that at ff - 1, the improvement of TAGS-opt-slowdown is still about 4 orders of magnitude. Why is the performance of TAGS so correlated with load? There are 2 reasons, both of which can be understood by looking at Figure 10 which shows the loads at the 2 hosts under TAGS-opt-slowdown in the case where the system load is (a) 0:3, (b) 0:5, and (c) 0:7. The first reason for the ineffectiveness of TAGS under high loads is that the higher the load, the less able TAGS is to play the load-unbalancing game described in Section 5.2. For lower ff, TAGS reaps much of its benefit at the lower ff by moving all the load onto Host 2. When the system load is only 0:5, TAGS is easily able to pile all the load on Host 2 without exceeding load 1 at Host 2. However when the system load is 0:7, the restriction that the load at Host 2 must not exceed 1 becomes a bottleneck for TAGS since it means that Host 1 can not be as underloaded as TAGS would like. This is seen by comparing Figure 10(b) and Figure 10(c) where in (c) the load on Host 1 is much higher for the lower ff than it is in (b). The second reason for the ineffectiveness of TAGS under high loads has to do with what we call excess. Excess is the extra work created in TAGS by tasks being killed and restarted. In the 2-host case, the excess is simply equal to is the outside arrival rate, p 2 is the fraction of tasks whose final destination is Host 2, and s 1 is the cutoff differentiating small tasks from big tasks. An equivalent definition of excess is the difference between the actual sum of the loads on the hosts and h times the system load, where h is the number of hosts. Notice that the dotted line in Figure 10(a)(b)(c) shows the sum of the loads on the hosts. Until now we've only considered the distributed servers with 2 hosts and system load 0:5. For this scenario, excess has not been a problem. The reason is that for low ff, where we need to do the severe load unbalancing, excess is basically non-existent for loads 0:5 and under, since p 2 is so small (due to the heavy-tailed property) and since s 1 could be forced down. For high ff, excess is present. However all the task assignment policies already do well in the high ff region because of the low task size variability, so the excess is not much of a handicap. When we look at the case of system load 0:7, however, excess is much more of a problem, as is evidenced by the dotted line in Figure 10(c). One reason that the excess is worse is simply that overall excess increases with load because excess is proportional to - which is in turn proportional to load. The other reason that the excess is worse at higher loads has to do with s 1 . In the low ff range, although p 2 is still low (due to the heavy-tailed property), s 1 cannot be forced low because the load at Host 2 is capped at 1. Thus the excess for low ff is very high. To make matters worse, some of this excess must be heaped on Host 1. In the high ff range, excess again is high because p 2 is high. Fortunately, observe that for higher loads excess is at its lowest point at ff - 1. In fact, it is barely existent in this region. Observe also that the region is the region where balancing load is the optimal thing to do (with respect to minimizing mean slowdown), regardless of the system load. This "sweet spot" is fortunate because ff - 1 is characteristic of many empirically measured computer workloads, see Section 3. 6 Analytic Results for Case of Multiple Hosts Until now we have only considered distributed servers with 2 hosts. For the case of 2 hosts, we saw that the performance of TAGS-opt-slowdown was amazingly good if the system load was 0:5 or less, but not nearly as good for system load ? 0:5. In this section we consider the case of more than 2 hosts. The phrase "adding more hosts" can be ambiguous because it is not clear whether the arrival rate is increased as well. For example, given a system with hosts and system load 0:7, we could increase the number of hosts to 4 hosts without changing the arrival rate, and the system load would drop to 0:35. On the other hand, we could increase the number of hosts to 4 hosts and increase the arrival rate appropriately (double it) so as to maintain a system load of 0:7. In our discussions below we will attempt to be clear as to which view we have in mind. One claim that can be made straight off is that an h host system (h ? with system load ae can always be configured to produce performance which is at least as good as that of a 2 host system with system load ae. To see why, observe that we can use the h host system (assuming h is even) to simulate a 2 host system as illustrated in Figure 11: Rename Hosts 1 and 2 as Subsystem 1. Rename Hosts 3 and 4 as Subsystem 2. Rename Hosts 5 and 6 as Subsystem 3, etc. Now split the traffic entering the h host system so that 2=hth of the tasks go to each of the h=2 Subsystems. Now apply your favorite task assignment policy to each Subsystem independently - in our case we choose TAGS. Each Subsystem will behave like a 2 host system with load ae running TAGS. Since each Subsystem will have identical performance, the performance of the whole host system will be equal to the performance of any one subsystem. (Observe that the above cute argument works for any task assignment policy). Figure 12 shows the mean slowdown under TAGS-opt-slowdown for the case of a 4 host distributed server with system load 0:3. Comparing these results to those for the 2 host system with system load 0:3 (Figure 9(a)), we see that: (a) System load 0:3 Loads at hosts under TAGS-opt-slowdown: 2 hosts, load .3 alpha host 1 - Load host 2 . Sum (b) System load 0:5 Loads at hosts under TAGS-opt-slowdown: 2 hosts, load .5 alpha host 1 - Load host 2 . Sum (c) System load 0:7 20.611.41.8Loads at hosts under TAGS-opt-slowdown: 2 hosts, load .7 alpha host 1 - Load host 2 . Sum Figure 10: Load at Host 1 and Host 2 under TAGS-opt-slowdown shown for a distributed server with 2 hosts and system load (a) 0:3 (b) 0:5 (c) 0:7. The dotted line shows the sum of the loads at the 2 hosts. If there were no excess, the dotted line would be at (a) 0:6 (b) 1:0 and (c) 1:4 in each of the graphs respectively. In figures (a) and (b) we see excess only at the higher ff range. In figure (c) we see excess in both the low ff and high ff range, but not around ff - 1. DISPATCHER OUTSIDE ARRIVALS TAGS SUBSYSTEM TAGS SUBSYSTEM TAGS Figure 11: Illustration of the claim that an h host system (h ? 2) with system load ae can always be configured to produce performance at least as good as a 2 host system with system load ae (although the h host system has much higher arrival rate). Results: MEAN SLOWDOWN using TAGS-opt-slowdown: 4 hosts, load .3 alpha . Least-Work-Left Approx. Figure 12: Mean slowdown under TAGS-opt-slowdown compared with other task assignment policies in the case of a distributed server with 4 hosts and system load 0:3. The cutoffs for TAGS-opt-slowdown were optimized by hand. In many cases it is possible to improve upon the results shown here by adjusting the cutoffs further, so the slight bend in the graph may not be meaningful. Observe that the mean slowdown of TAGS almost never exceeds 1. 1. The performance of Random stayed the same, as it should. 2. The performance of Least-Work-Remaining improved by a couple orders of magnitude in the higher ff region, but less in the lower ff region. The Least-Work-Remaining task assignment policy is helped by increasing the number of hosts, although the system load stayed the same, because having more hosts increases the chances of one of them being free. 3. The performance of TAGS-opt-slowdown improved a lot. So much so, that the mean slowdown under TAGS-opt-slowdown is never over 6 and almost always under 1. At ff - 1, TAGS-opt-slowdown improves upon Least-Work-Remainingby 4-5 orders of magnitude. At improves upon Least-Work-Remaining by 8-9 orders of magnitude. At improves upon Least-Work-Remaining by over 25 orders of magnitude! The enhanced performance of TAGS on more hosts may come from the fact that more hosts allow for greater flexibility in choosing the cutoffs. However it is hard to say for sure because it is difficult to compute results for the case of more than 2 hosts. The cutoffs in the case of 2 hosts were all optimized by Mathematica, while in the case of 4 hosts it was necessary to perform the optimizations by hand (and for all we know, it may be possible to do even better). For the case of system load 0:7 with 4 hosts we ran into the same type of problems as we did for the 2 host case with system load 0:7. 6.1 The Server Expansion Performance Metric There is one thing that seems very artificial about our current comparison of task assignment policies. No one would ever be willing to run a system whose expected mean slowdown In practice, if a system was operating with mean slowdown of the number of hosts would be increased, without increasing the arrival rate, (thus dropping the system load) until the system's performance improved to a reasonable mean slowdown, like 3 or less. Consider the following example: Suppose we have a 2-host system running at system load .7 and with variability parameter :6. For this system the mean slowdown under TAGS-opt-slowdown is on the order of 10 9 , and no other task assignment policy that we know of does better. Suppose however we desire a system with mean slowdown of 3 or less. So we double the number of hosts (without increasing the outside arrival rate). At 4 hosts, with system load 0:35, TAGS-opt-slowdown now has mean slowdown of around 1, whereas Least-Work-Remaining's slow-down has improved to around 10 8 . It turns out we would have to increase number of hosts to 13 for the performance of Least-Work-Remaining to improve to the point of mean slowdown of under 3. And for Random to reach that level it would require an additional 10 9 hosts! The above example suggests a new practical performance metric for distributed servers, which we call the server expansion metric. The server expansion metric asks how many additional hosts must be added to the existing server (without increasing outside arrival rate) to bring mean slowdown down to a reasonable level (where we'll arbitrarily define ``reasonable'' as 3 or less). Figure 13 compares the performance of our task assignment policies according to the server expansion metric, given that we start with a 2 host system with system load of 0:7. For TAGS-opt-slowdown, the server expansion is only 3 for no more than 2 for all the other ff. For Least-Work-Remaining, on the other hand, the number of hosts we need to add ranges from 1 to 27, as ff decreases. Still Least-Work-Remaining is not so bad because at least its performance improves somewhat quickly as hosts are added and load is decreased, the reason being that both these effects increase the probability of a task finding an idle host. By contrast Random, shown in Figure 13(b), is exponentially worse than the others, requiring as many as 10 5 additional hosts when ff - 1. Although Random does benefit from increasing the number of hosts, the effect (a) Non-log scale alpha -o- TAGS-opt-slowdown .o. Least-Work-Remaining (b) Log scale Server expansion requirement -o- Random .o. Least-Work-Remaining -o- TAGS-opt-slowdown alpha Figure 13: Server expansion requirement for each of the task assignment policies, given that we start with a 2 host system with system load of 0:7. (a) Shows just Least-Work-Remaining and TAGS-opt-slowdown on a non-log scale (b) Shows Least-Work-Remaining, TAGS-opt-slowdown, and Random on a log scale. Second Moment of Bounded Pareto Distribution B(k,p,alpha) where alpha Figure 14: Second moment of B(k; p; ff) distribution, where now the upper bound, p, is set at The mean is held fixed at 3000 as ff is varied. Observe that the coefficient of variation now ranges from 2, when isn't nearly as strong as it is for TAGS and Least-Work-Remaining. 7 The effect of the range of task sizes The purpose of this section is to investigate what happens when the range of task sizes (difference between the biggest and smallest possible task sizes) is smaller than we have heretofore assumed, resulting in a smaller coefficient of variation in the task size distribution. Until now we have always assumed that the task sizes are distributed according to a Bounded Pareto distribution with upper bound fixed mean 3000. This means, for example, that when ff - 1 (as agrees with empirical data), we need to set the lower bound on task sizes to 167. However this implies that the range of task sizes spans 8 orders of magnitude! It is not clear that most applications have task sizes ranging 8 orders in mag- nitude. In this section we rederive the performance of all the task assignment policies when the upper bound p is set to still holding the mean of the task size distribution at 3000. This means, for example, that when ff - 1 using TAGS-opt-slowdown: 2 hosts, load .5,p =10 7 alpha Figure 15: Mean slowdown under TAGS-opt-slowdown in a distributed server with 2 hosts with system load 0:5, as compared with the performance of Random and Least-Work-Remaining. In this set of results the task size distribution is (as agrees with empirical data), we need to set the lower bound on task sizes to which implies the range of task sizes spans just 5 orders of magnitude. Figure 14 shows the second moment of the Bounded Pareto task size distribution as a function of ff when . Comparing this figure to Figure 4, we see that the task size variability is far lower when therfore so is the coefficient of variation. Lower variance in the task size distribution suggests that the improvement of TAGS over the other task assignment policies will not be as dramatic as in the higher variability setting (when This is in fact the case. What is interesting, however, is that even in this lower variability setting the improvement of TAGS over the other task assignment policies is still impressive, as shown in Figure 15. Figure 15 shows the mean slowdown of TAGS-opt-slowdown as compared with Random and Least-Work-Left for the case of two hosts with system load 0:5. Observe that for ff - 1, TAGS improves upon the other task assignment policies by over 2 orders of magnitude. As ff drops, the improvement increases. This figure should be contrasted with Figure 9(b), which shows the same scenario where 8 Conclusion and Future Work This paper is interesting not only because it proposes a powerful new task assignment policy, but more so because it challenges some natural intuitions which we have come to adopt over time as common knowledge. Traditionally, the area of task assignment, load balancing and load sharing has consisted of heuristics which seek to balance the load among the multiple hosts. TAGS, on the other hand, specifically seeks to unbalance the load, and sometimes severely unbalance the load. Traditionally, the idea of killing a task and restarting from scratch on a different machine is viewed with skepticism, but possibly tolerable if the new host is idle. TAGS, on the other hand, kills tasks and then restarts them at a target host which is typically operating at extremely high load, much higher load than the original source host. Furthermore, TAGS proposes restarting the same task multiple times. It is interesting to consider further implications of these results, outside the scope of task assignment. Consider for example the question of scheduling CPU-bound tasks on a single CPU, where tasks are not preemptible and no a priori knowledge is given about the tasks. At first it seems that FCFS scheduling is the only option. However in the fact of high task size variability, FCFS may not be wise. This paper suggests that killing and restarting tasks may be worth investigating as an alternative, if the load on the CPU is low enough to tolerate the extra work created. Task assignment also has applications outside of the context of a distributed server system described in this paper. A very interesting recent paper by Shaikh, Rexford, and Shin [15] discusses routing of IP flows (which also have heavy-tailed size distributions) and recommends routing long flows differently from short flows. --R Load profiling: A methodology for scheduling real-time tasks in a distributed system Task assignment in a distributed system: Improving performance by unbalancing load. A simple dynamic routing problem. Task assignment in a distributed server. Task assignment in a distributed server. Exploiting process lifetime distributions for dynamic load balancing. Unix file size survey - <Year>1993</Year> An approximation to the response time for shortest queue routing. An approximation for the mean response time for shortest queue routing with general interarrival and service times. Fractal patterns in DASD I/O traffic. Approximations in finite capacity multiserver queues with poisson arrivals. On the optimal assignment of customers to parallel servers. Deciding which queue to join: Some counterexamples. Optimality of the shortest line discipline. An upper bound for multi-channel queues Stochastic Modeling and the Theory of Queues. --TR Deciding which queue to join: Some counterexamples An approximation to the response time for shortest queue routing An approximation for the mean response time for shortest queue routing with general interarrival and service times area traffic Exploiting process lifetime distributions for dynamic load balancing Self-similarity in World Wide Web traffic Task assignment in a distributed system (extended abstract) Heavy-tailed probability distributions in the World Wide Web Load-sensitive routing of long-lived IP flows Load-balancing heuristics and process behavior On choosing a task assignment policy for a distributed server system Implementing Multiprocessor Scheduling Disciplines Theory and Practice in Parallel Job Scheduling Improved Utilization and Responsiveness with Gang Scheduling Valuation of Ultra-scale Computing Systems A parallel workload model and its implications for processor allocation Evaluation of Task Assignment Policies for Supercomputing Servers Load profiling --CTR Jianbin Wei , Cheng-Zhong Xu, Design and implementation of a feedback controller for slowdown differentiation on internet servers, Special interest tracks and posters of the 14th international conference on World Wide Web, May 10-14, 2005, Chiba, Japan Konstantinos Psounis , Pablo Molinero-Fernndez , Balaji Prabhakar , Fragkiskos Papadopoulos, Systems with multiple servers under heavy-tailed workloads, Performance Evaluation, v.62 n.1-4, p.456-474, October 2005 Victoria Ungureanu , Benjamin Melamed , Phillip G. Bradford , Michael Katehakis, Class-Dependent Assignment in cluster-based servers, Proceedings of the 2004 ACM symposium on Applied computing, March 14-17, 2004, Nicosia, Cyprus Victoria Ungureanu , Benjamin Melamed , Michael Katehakis , Phillip G. Bradford, Deferred Assignment Scheduling in Cluster-Based Servers, Cluster Computing, v.9 n.1, p.57-65, January 2006 Mor Harchol-Balter , Cuihong Li , Takayuki Osogami , Alan Scheller-Wolf , Mark S. Squillante, Cycle stealing under immediate dispatch task assignment, Proceedings of the fifteenth annual ACM symposium on Parallel algorithms and architectures, June 07-09, 2003, San Diego, California, USA Jianbin Wei , Xiaobo Zhou , Cheng-Zhong Xu, Robust Processing Rate Allocation for Proportional Slowdown Differentiation on Internet Servers, IEEE Transactions on Computers, v.54 n.8, p.964-977, August 2005 James Broberg , Zahir Tari , Panlop Zeephongsekul, Task assignment with work-conserving migration, Parallel Computing, v.32 n.11-12, p.808-830, December, 2006 T. Madhusudan , Young-Jun Son, A simulation-based approach for dynamic process management at web service platforms, Computers and Industrial Engineering, v.49 n.2, p.287-317, September 2005

load sharing;fairness;job scheduling;supercomputing;distributed servers;contrary behavior;high variance;load balancing;task assignment;heavy-tailed workloads;clusters

506161

Mesh Partitioning for Efficient Use of Distributed Systems.

Mesh partitioning for homogeneous systems has been studied extensively; however, mesh partitioning for distributed systems is a relatively new area of research. To ensure efficient execution on a distributed system, the heterogeneities in the processor and network performance must be taken into consideration in the partitioning process; equal size subdomains and small cut set size, which results from conventional mesh partitioning, are no longer the primary goals. In this paper, we address various issues related to mesh partitioning for distributed systems. These issues include the metric used to compare different partitions, efficiency of the application executing on a distributed system, and the advantage of exploiting heterogeneity in network performance. We present a tool called PART, for automatic mesh partitioning for distributed systems. The novel feature of PART is that it considers heterogeneities in the application and the distributed system. Simulated annealing is used in PART to perform the backtracking search for desired partitions. While it is well-known that simulated annealing is computationally intensive, we describe the parallel version of simulated annealing that is used with PART. The results of the parallelization exhibit superlinear speedup in most cases and nearly perfect speedup for the remaining cases. Experimental results are also presented for partitioning regular and irregular finite element meshes for an explicit, nonlinear finite element application, called WHAMS2D, executing on a distributed system consisting of two IBMSPs with different processors. The results from the regular problems indicate a 33 to 46 percent increase in efficiency when processor performance is considered as compared to the conventional even partitioning. The results indicate a 5 to 15 percent increase in efficiency when network performance is considered as compared to considering only processor performance; this is significant given that the optimal improvement is 15 percent for this application. The results from the irregular problem indicate up to 36percent increase in efficiency when processor and network performance are considered as compared to even partitioning.

Introduction Distributed computing has been regarded as the future of high performance computing. Nation-wide high speed networks such as vBNS [25] are becoming widely available to interconnect high-speed computers, virtual environments, scientific instruments and large data sets. Projects such as Globus [15] and Legion [20] are developing software infrastructure that integrates distributed computational and informational resources. In this paper, we present a mesh partitioning tool for distributed systems. This tool, called PART, takes into consideration the heterogeneity in processors and networks found in distributed systems as well as heterogeneities found in the applications. Mesh partitioning is required for efficient parallel execution of finite element and finite difference applications, which are widely used in many disciplines such as biomedical engineering, structural mechanics, and fluid dynamics. These applications are distinguished by the use of a meshing procedure to discretize the problem domain. Execution of a mesh-based application on a parallel or distributed system involves partitioning the mesh into subdomains that are assigned to individual processors in the parallel or distributed system. Mesh partitioning for homogeneous systems has been studied extensively [2, 4, 14, 31, 36, 37, 41]; however, mesh partitioning for distributed systems is a relatively new area of research brought about by the recent availability of such systems. To ensure efficient execution on a distributed system, the heterogeneities in the processor and network performance must be taken into consideration in the partitioning process; equal size subdomains and small cut set size, which results from conventional mesh partitioning, are no longer desirable. PART takes advantage of the following heterogeneous system features: (1) processor speed; (2) number of processors; (3) local network performance; and wide area network performance. Further, different finite element applications under consideration may have different computational complexity, different communication patterns, and different element types, which also must be taken into consideration when partitioning. In this paper, we discuss the major issues in mesh partitioning for distributed systems. In particular, we identify a good metric to be used to compare different partitioning results, present a measure of efficiency for a distributed system, and discuss optimal number of cut sets for remote communication. The metric used with PART to identify good efficiency is estimated execution time. We also present a parallel version of PART that significantly improves performance of the partitioning process. Simulated annealing is used in PART to perform the backtracking search for desired partitions. However, it is well known that simulated annealing is computationally intensive. In the parallel PART, we use the asynchronous multiple Markov chain approach of parallel simulated annealing [21]. PART is used to partition six irregular meshes into 8, 16, and 100 subdomains using up to 64 client processors on an IBM SP2 machine. The results show superlinear speedup in most cases and nearly perfect speedup for the rest. The results also indicate that the parallel version of PART produces partitions consistent with the sequential version of PART. Using partitions from PART, we ran an explicit, 2-D finite element code on two geographically distributed IBM SP machines. We used Globus software for communication between the two SPs. We compared the partitions from PART with that generated using the widely-used partitioning tool, METIS [26], which considers only processor performance. The results from the regular problems indicate a increase in efficiency when processor performance is considered as compared to the conventional even partitioning; the results indicate 5 \Gamma 15% increase in efficiency when network performance is considered as compared to considering only processor performance; this is significant given that the optimal is 15% for this application. The result from the irregular problem indicate up to 36% increase in efficiency when processor and network performance are considered as compared to even partitioning. The remainder of the paper is organized as follows: Section 2 provides background. Section 3 discusses issues. Section 4 describes PART in detail. Section 5 is experimental results. Section 6 gives previous work and finally conclusion. Background 2.1 Mesh-based Applications Finite element method has been the fundamental numerical analysis technique to solve partial differential equations in the engineering community for the past three decades [24, 3]. There are three basic procedures in the finite element method. The problem is first formulated in variational or weighted residual form. In the second step, the problem domain is discretized into complex shapes called elements. The last major step is to solve the resulting system of equations. The procedure of discretizing the problem domain is called meshing. Applications that involve a meshing procedure are referred to as mesh-based applications. Mesh-based applications are naturally suitable for parallel or distributed systems. Implementing the finite element method in parallel involves partitioning the global domain of elements into connected subdomains that are distributed among P processors; each processor executes the numerical technique on its assigned subdomain. The communication among processors is dictated by the types of integration method and solver method. Explicit integration finite element problems do not require the use of a solver since a lumped matrix (which is a diagonal matrix) is used. Therefore, communication only occurs among neighboring processors that have common data and is relatively simple. For implicit integration finite element problems, however, communication is determined by the type of solver used in the application. The application used in this paper is an explicit, nonlinear finite code, called WHAMS2D [6], which is used to analyze elastic plastic materials. While we focus on the WHAMS2D code, the concept can be generalized to implicit as well as other mesh-based applications. 2.2 Distributed System Distributed computing consists of a platform with a network of resources. These resources may be clusters of workstations, cluster of personal computers, or parallel machines. Further, the resources maybe located at one site or distributed among different sites. Figure 1 shows an example of a distributed system. Distributed systems provide an economical alternative to costly massively parallel computers. Researchers are no longer limited by the computing resources at individual sites. The distributed computing environment also provides researchers opportunities to collaborate and share ideas through the use of collaboration technologies. In a distributed system, we define "group" as a set of processors that share one interconnection network and have the same performance. A group can be an SMP, a parallel computer, or a cluster of workstations or personal computers. Communication occurs both within a group and between groups. We refer to communication within a group as local communication; and those between processors in different groups as remote communication. The number of groups in the distributed system is represented by the term S. Supercomputer SMPs NOW Supercomputer Figure 1: A distributed system. 2.3 Problem Formulation Mesh partitioning for homogeneous systems can be viewed as a graph partitioning problem. The goal of the graph partitioning problem is to find a small vertex separator and equal sized subsets. Mesh partitioning for distributed system, however, is a variation of the graph partitioning problem; the goal differs from regular graph partitioning problem in that equal sized subsets may not be desirable. The distributed system partitioning problem can be stated as follows: Given a graph E) of jV and the maximum of a cost function f over all V i is minimized: Min In this paper, the cost function f is the estimate of execution time of a given application on a distributed system. This function is discussed further in Section 4. Graph partitioning has been proven to be NP-complete. The mesh partitioning problem for distributed system is also NP-complete as proven in Appendix 1. Therefore, we focus on heuristics to solve this problem. 3 Major Issues In this section, we discuss the following major issues related to the mesh partitioning problem for distributed systems: comparison metric, efficiency, and number of cuts between groups. 3.1 Comparison Metric The de facto metric for comparing the quality of different partitions for homogeneous parallel systems has been equal subdomains and minimum interface (or cut set) size. Although there have been objections and counter examples [14], this metric has been used extensively in comparing the quality of different partitions. It is obvious that equal subdomain size and minimum interface is not valid for comparing partitions for distributed systems. One may consider an obvious metric for a distributed system to be unequal subdomains (pro- portional to processor performance) and small cut set size. The problem with this metric is that heterogeneity in network performance is not considered. Given the local and wide area networks are used in distributed system, it is the case that there will be a big difference between local and remote communication, especially in terms of latency. We argue that the use of an estimate of execution time of the application on the target heterogeneous system will always lead to a valid comparison of different partitions. The estimate is used for relative comparison of different partition methods. Hence a coarse approximation of the execution is appropriate for the comparison metric. It is important to make the estimate representative of the application and the system. The estimate should include parameters that correspond to system heterogeneities such as processor performance, local and remote communication. It should also reflect the application computational complexity. 3.2 Efficiency The efficiency for the distributed system is equal to the ratio of the relative speedup to the effective number of processors, V . This ratio is given below: (1) where E(1) is the sequential execution time on one processor and E is the execution time on the distributed system. The term V is equal to the summation of each processor's performance relative to the performance of the processor used for sequential execution. This term is as follows: (2) where k is the processor used for sequential execution. For example, with two processors having processor performance F 2, the efficiency would be if processor 1 is used for sequential execution; the efficiency is if processor 2 is instead used for sequential execution. 3.3 Network Heterogeneity It is well-known that heterogeneity of processor performance must be considered with distributed systems. In this section, we identify conditions for which heterogeneity in network performance must be considered. For a distributed system, recall that we define a group to be a collection of processors that have the same performance and share a local interconnection network. Remote communication corresponds to communication between two groups. Given that some processors require remote and local communication, while others only require local communication, there will be a disparity between the execution times of these processors corresponding to the difference in remote and local communications (assuming equal computational loads). 3.3.1 Ideal Reduction in Execution Time A retrofit step is used with the PART tool to reduce the computational load of processor with local and remote communication to equalize the execution time among the processors in a group. This step is described in detail in Section 6.2. The reduction in execution time that occurs with this retrofit is demonstrated by considering a simple case, stripe partitioning, for which communication occurs with at most two neighboring processors. Assume there exists two groups having the same processor and local network performance; the groups are located at geographically distributed sites requiring a WAN for interconnection. Figure 2 illustrates one such case. G processors local communication local communication remote communication Figure 2: Communication Pattern for Stripe Partitioning. Processor i (as well as processor local and remote communication. The difference between the two communication times is: where x is the percentage of the difference of CR and CL in the total execution time E. Assume that E represents the execution time taking into consideration only processor performance. Since it is assumed that all processors have the same performance, this entails an even partition of the mesh. This time can be written as: Now consider the case of partitioning to take into consideration the heterogeneity in network performance. This is achieved by decreasing the load assigned to processor i and increasing the loads of the processors in group 1. The same applies to processor j in group 2. The amount of the load to be redistributed is CR \Gamma CL or x%E and this amount is distributed to G processors. This is illustrated in Figure 6, which is discussed with the retrofit step of PART. The execution time is now: G The difference between E and E 0 is: x G G Therefore, by taking the network performance into consideration when partitioning, the percentage reduction in execution time is approximately x%E(denoted as \Delta(1; G)) which includes the following: (1) the percentage of communication in the application and (2) the difference in the remote and local communication. Both factors are determined by the application and the partitioning. If the maximum number of processors among the groups that have remote communication is -, then the reduction in execution time is as follows: G \Delta(-; For example, for the WHAMS2D application in our experiments, we calculated the ideal reduction to be 15% for the regular meshes with 8. For those partitions, only one processor in each group has local and remote communication, therefore, it is relatively easy to calculate the ideal performance improvement. 3.3.2 Number of Processors in a Group with Local and Remote Communication The major issue to be addressed with the reduction is how to partition the domain assigned to a group to maximize the reduction. In particular, this issue entails a tradeoff between the following two scenarios: 1. Many processors in a group having local and remote communication, resulting in small message sizes for which the execution time without the retrofit step is smaller than that for case 2. However, given that many processors in a group have remote and local communication, there are fewer processors that are available for redistribution of the additional load. This is illustrated in Figure 3 where a mesh of size n \Theta 2n is partitioned into 2P blocks. Each block oe -? oe -? Figure 3: A size n \Theta 2n mesh partitioned into 2P blocks. is n \Theta n assuming all processors have equal performance. The mesh is partitioned into two groups, each group having P processors. Processors on the group boundary incur remote communication as well as local communication. Part of the computational load of these processors need to be moved to processors with only local communication to compensate for the longer communication times. Assume there is no overlap in communication messages and message aggregation is used for the communication of one node to the diagonal processor, the communication time for a processor on the group boundary is approximately: local+remote For a processor with only local communication, the communication time is approximately (again, message aggregation and no overlapping is assumed): local Therefore, the communication time difference between a processor with local and remote communication and a processor with only local communication is approximately: local There are a total of P number of processors with local and remote communication. There- fore, using Equation 1, the ideal reduction in execution time in Group 1 (and Group 2) -: . n Figure 4: A size n \Theta 2n mesh partitioned into 2P stripes. is: blocks 2. Only one processor in a group has local and remote communication, resulting in large message sizes which result in the execution time without the retrofit step larger than that for case 1. However, there are more processors that are available for redistribution of additional load. This is illustrated in Figure 4 where the same mesh is partitioned into stripes; there is only one processor in each group that have local and remote communication. Following a similar analysis as in Figure 3, the communication time difference between a processor with both local and remote communication and a processor with only local communication is approximately: local There is only one processor with remote communication in each group. Hence, using Equation 1, the ideal reduction in execution time is: stripes comm (stripes) (15) Therefore, the total execution time for stripe and block partitioning are: stripes reduction blocks reduction The difference in total execution time between block and stripe partition is: \DeltaT blocks\Gammastripes ff L A Therefore, the difference in total execution time between block and stripe partitioning is determined by . The term A and C are positive since P ? 1; while the term B is negative 4, the block partition has a higher execution time, i.e., the stripe partitioning is advantageous. If P ? 4, however, block partitioning will still have a higher execution time unless n is so large that the absolute value of term B is larger than the sum of the absolute values of A and C. Note that ff L and ff R are one to two orders of magnitude larger than fi L . In our experiments, we calculated that block partitioning has a lower execution time only if n ? 127KB. In the meshes that we used, however, the largest n is only about 10KB. 4 Description of PART PART considers heterogeneities in both the application and the system. In particular, PART takes into consideration that different mesh based applications may have different computational complexities and the mesh may consist of different element types. For distributed systems, PART takes into consideration heterogeneities in processor and network performance. Figure 5 shows a flow diagram of PART. PART consists of an interface program and a simulated annealing program. A finite element mesh is fed into the interface program and produces the proposed communication graph, which is then fed into a simulated annealing program where the final partitioning is computed. This partitioned graph is translated to the required input file format for the application. This section describes the initial interface program and the steps required to partition the graph. Problem domain # of groups # of processors per group Computational models Communication models Partitioned data Finite element mesh Interface program graph Simulated Annealing Partitioned graph Interface program Input to processors Figure 5: PART flowchart. 4.1 Mesh Representation We use a weighted communication graph to represent a finite element mesh. This is a natural extension of the communication graph. As in the communication graph, vertices represent elements in the original mesh. A weight is added to each vertex to represent the number of nodes within the element. Same as in communication graph, edges represent the connectivity of the elements in the weighted communication graph. A weight is also added to each edge to represent the number of nodes of which information need to be exchanged between the two neighboring elements. 4.2 Partition Method PART entails three steps to partition a mesh for distributed systems. These steps are: 1. Partition the mesh into S subdomains for the S groups, taking into consideration heterogeneity in processor performance and element types. 2. Partition each subdomain into G parts for the G processors in a group, taking into consideration heterogeneity in network performance and element types. 3. If necessary, globally retrofit the partitions among the groups, taking into consideration heterogeneity in the local networks among the different groups. Each of the above steps is described in detail in the following subsections. Each subsection includes a description of the objective function used with simulated annealing. The key to a good partitioning by simulated annealing is the cost function. The cost function used by PART is the estimate of execution time. For one particular supercomputer, let E i be the execution time for the i-th processor (1 - i - p). The goal here is to minimized the variance of the execution time for all processors. While running the simulated annealing program, we found that the best cost function is: instead of the sum of the 2 . So (20) is the actual cost function used in the simulated annealing program. In this cost function, Ecomm includes the communication cost for the partitions that have elements that need to communicate with elements on a remote processor. Therefore, the execution time will be balanced. - is the parameter that needs to be tuned according to the application and problem size. Partition The first step generates a coarse partitioning for the distributed systems. Each group gets a subdomain that is proportional to its number of processors, the performance of the processors, and the computational complexity of the application. Hence computational cost is balanced across all the groups. The cost function is given by: where S is the number of groups in the system. 4.2.2 Step 2: Retrofit In the second step, the subdomain that is assigned to each group from Step 1 is partitioned among its processors. Within each group, simulated annealing is used to balance the execution time. In this step, variance in network performance is considered. Processors that entails inter group communication will have reduced computational load to compensate for the longer communication time. The step is illustrated in Figure 6 for two supercomputers, SC1 and SC2. In SC1, four processors are used; and two processors are used in SC2. Computational load is reduced for P3 since it communicates with a remote processor. The amount of reduced computational load is represented as ffi. This amount is equally distributed to the other three processors. Assuming the cut size remains unchanged, the communication time will not change, hence the execution time will be balanced after this shifting of computational load. comm. comp. comm. p3 p3 comp. retrofit d d/4 Figure An illustration of the retrofit step for two supercomputers assuming only two nearest neighbor communication. This step entails generating imbalanced partitions in group i that take into consideration that some processors communicate locally and remotely and other processors communicate only locally. The imbalance is represented with the term \Delta i . This term is added to processors that require local and remote communication. Adding this term results in a decrease in Ecomm i as compared to processors requiring only local communication. The cost function is given by the following equation: where p is the number of processors in a given group; \Delta i is the difference in the estimation of local and remote communication time. For processors that only communicate locally, 4.2.3 Step 3: Global Retrofit The third step addresses the global optimization, taking into consideration differences in the local interconnect performance of various groups. Again, the goal is to minimize the variance of the execution time across all processors. In this step, elements on the boundaries of partitions are moved according to the execution time variance between neighboring processors. This step is only executed if there is a large difference in the performance of the different local interconnects. For the case when a significant number of elements are moved between the groups in Step 3, the second step is executed again to equalize the execution time in a group given the new computational load. After Step 2, processors in each group will have a balanced execution time. However, execution time of the different groups may not be balanced. This may occur when there is a large difference in the communication time of the different groups. To balance the execution among all the groups, we take the weighted average of execution times from all the groups. The weight for each group equals to the computing power of that group versus the total computing power. The computing power for a particular group is the multiplication of the ratio of the processor performance with respect to the slowest one among all the groups and the number of processors used from that group. We denote this weighted average as - E. Under the assumption that communication time will not change much (i.e., the separators from Step 1 will not incur a large change in size), E is the optimal execution time that can be achieved. To balance the execution time so that each group will have an execution time of - E, we first compute the difference of E i with - E: This \Gamma i is then added to each E comp i in the cost function. The communication cost Ecomm i is now again the remote communication cost for group i. The cost function is therefore given by: where S is the number of groups in the system. For groups whose the domain will increase; for groups whose the domain will decrease. If Step 3 is necessary, then Step 2 is performed again to partition within each group. 5 Parallel Simulated Annealing PART uses simulated annealing to partition the mesh. Figure 7 shows the serial version of the simulated annealing algorithm. This algorithm uses the Metropolis criteria (line 8 to 13 in Figure 7) to accept or reject moves. The moves that reduce the cost function are accepted; the moves that increase the cost function may be accepted with probability e \Gamma \DeltaE avoiding being trapped in local minima. This probability decreases when the temperature is lowered. Simulated Annealing is computationally intensive, therefore, a parallel version of simulated annealing is used in the parallel version of PART. There are three major classes of parallel simulated annealing [19]: serial- like [32, 39], parallel moves [1], and multiple Markov chains [5, 21, 34]. Serial like algorithms essentially break up each move into subtasks and parallelize the subtasks (parallelizing line 6 and 7 in Figure 7). For the parallel moves algorithms, each processor generates and evaluates moves cost function calculation may be inaccurate since processors are not aware of moves by other processors. Periodic updates are normally used to address the effect of cost function error. Parallel moves algorithms essentially parallelize the for loop in Figure 7 (line 5 to 14). For the multiple Markov chains algorithm, multiple simulated annealing processes are started on various processors with different random seeds. Processors periodically exchange solutions and the best is selected and given to all the processors to continue their annealing processes. In [5], the multiple Markov chain approach was shown to be most effective for VLSI cell placement. For this reason, the parallel version of PART uses the multiple Markov chain approach. Given P processors, a straightforward implementation of the multiple Markov chain approach would be initiating simulated annealing on each of the P processors with a different seed. Each processor performs moves independently and then finally the best solution from those computed by 1. Get an initial solution S 2. Get an initial temperature T ? 0 3. While stopping criteria not met f 4. number of moves per temperature 5. for 6. Generate a random move 7. Evaluate changes in cost function: \DeltaE 8. if (\DeltaE ! 9. accept this move, and update solution S 10. g else f 11. accept with probability 12. update solution S if accepted 13. g 14. g /*end for loop*/ 15. 16.g /*end while loop */ Figure 7: Simulated annealing. all processors is selected. In this approach, however, simulated annealing is essentially performed P times which may result a better solution but not speedup. To achieve speedup, P processors perform an independent simulated annealing with a different seed, but each processor performs only M=P moves (M is the number of moves performed by the simulated annealing at each temperature). Processors exchange solutions at the end of each temperature. The exchange of data occurs synchronously or asynchronously. In the synchronous multiple Markov chain approach, the processors periodically exchange solutions with each other. In the asynchronous approach, the client processors exchange solutions with a server processor. It has been reported that the synchronous approach is more easily trapped in a local optima than the asynchronous [21], therefore the parallel version of PART uses the asynchronous approach. During solution exchange, if the client solution is better, the server processor is updated with the better solution; if the server solution is better, the client gets updated with the better solution and continues from there. Each processor exchanges solution with the server processor at the end of each temperature. To ensure that each subdomain is connected, we check for disconnected components at the end of PART. If any subdomain has disconnected components, the parallel simulated annealing is repeated with a different random seed. This process continues until there are no disconnected subdomains or the number of trials exceed three times. A warning message is given in the output if there are disconnected subdomains. 6 Experiments In this section, we present the results from two different experiments. The first experiment focuses on the speedup of the parallel version of PART. The second experiment focuses on the quality of the partitions generated with PART. 6.1 Speedup Results Table 1: Parallel PART execution time (seconds): 8 partitions. # of proc. barth4 barth5 inviscid labarre spiral viscous 4 44.4 54.7 46.6 32.4 41.2 53.7 2.1 2.2 2.4 2.5 1.5 2.9 PART is used to partition six 2D irregular meshes with triangular elements: barth4 (11451 labarre (14971 elem.), spiral (1992 elem.), and viscous (18369 elem. The running time of partitioning the six irregular meshes into 8, and 100 Parallel PART Speedup: 8 partitions2060100140 barth4 barth5 inviscid labarre spiral viscous Figure 8: Parallel PART speedup for 8 partitions. subdomains are given in Tables 1 and 2, respectively. It is assumed that the subdomains will be executed on a distributed system consisting of two IBM SPs, with equal number of processors but different processor performance. Further, the machines are interconnected via vBNS for which the performance of the network is given in Table 4 (discussed in Section 6.2). In each table, column 1 is the number of client processors used by PART, and columns 2 to 6 are the running time of PART in seconds for the different meshes. The solution quality of using two or more client processors is within 5% of that of using one client processor. In this case, the solution quality is the estimate of the execution time of WHAMS2D. Figures 8 and 9 are graphical representations of the speedup of the parallel version of PART relative to one client processor. The figures show that when the meshes are partitioned into 8 subdomains, superlinear speedup occurs in all cases. When the meshes are partitioned into 100 subdomains, superlinear speedup occurs only in the cases of two smallest meshes, spiral and inviscid. Other cases show slightly less than perfect speedup. This superlinear speedup is attributed to the use of multiple client processors conducting a search, for which all the processors benefit from the results. Once a good solution is found by any one of the clients, this information is given to other clients quickly, thereby reducing the effort of continuing to search for a solution. The superlinear Table 2: Parallel PART execution time (seconds): 100 partitions. # of proc. barth4 barth5 inviscid labarre spiral viscous 4 3982.2 4666.4 3082.3 4273.5 1304.4 3974.0 288.3 426.3 192.8 291.2 62.7 391.7 speedup results are consistent with that reported in [33]. 6.2 Quality of Partition 6.2.1 Regular Meshes PART was applied to an explicit, nonlinear finite code, called WHAMS2D [6], that is used to analyze elastic plastic materials. The code uses MPI built on top of Nexus for interprocessor communication within a supercomputer and between supercomputers. Nexus is a runtime system that allows for multiple protocols within an application. The computational complexity is linear with the size of the problem. The code was executed on the IBM SP machines located at Argonne National Laboratory and the Cornell Theory Center. These two machines were connected by the Internet. Macro benchmarks were used to determine the network and processor performance. The results of the network performance analysis are given in Table 3. Further, experiments were conducted to determine that the Cornell nodes were 1.6 times faster than the Argonne nodes. The problem mesh consists of 3 regular meshes. The execution time is given for 100 time steps corresponding to 0.005 seconds of application time. Generally, the application may execute for 10; 000 to 100; 000 time steps. The recorded execution time represents over 100 runs, taking the data from the runs with standard deviation less than 3%. The regular problems were executed on Parallel PART Speedup: 100 partitions1030507090barth4 barth5 inviscid labarre spiral viscous Figure 9: Parallel PART speedup for 100 partitions. Table 3: Values of ff and fi for the different networks. Argonne SP Vulcan Switch ff a machine configuration of 8 processors (4 at ANL IBM SP and 4 at CTC IBM SP). Table 4 presents the results for the regular problems. Column 1 is the mesh configuration. Column 2 is the execution time resulting from the conventional equal partitioning. In particular, we used Chaco's spectral bisection. Column 3 is the result from the partitioning taken from the end of the first step for which the variance in processor performance and computational complexity are considered. Column 4 is the execution time resulting from the partitioning taken from the end of the second step for which the variance in network performance is considered. The results in Table 4 shows that approximately increase in efficiency can be achieved by balancing the computational cost; another 5 \Gamma 16% efficiency increase can be achieved by considering the variance in network performance. The small increase in efficiency by considering the network performance Table 4: Execution time using the Internet 8 processors: 4 at ANL, 4 at CTC Case Chaco Proc. Perf. Local Retrofit 9 \Theta 1152 mesh 102.99 s 78.02 s 68.81 s efficiency 0.46 0.61 0.71 efficiency 0.47 0.61 0.68 36 \Theta 288 mesh 103.88 s 73.21 s 70.22 s efficiency 0.46 0.67 0.70 is due to communication being a small component of the WHAMS2D application. However, recall that the optimal increase in performance is 15% for the regular problem as described earlier. The global optimization step, which is the last step of PART that balances execution time across all supercomputers, did not give significant increase in efficiency (it is not included in Table 4). This is expected since the two supercomputers we used, the Argonne IBM SP and the Cornell IBM SP, both have interconnection networks that have very similar performance as indicated in Table 3. The results indicate the performance gains achievable with each step in comparison to conventional methods that evenly partition the mesh. Given that it is obvious that considering processor performance results in significant gains, the following section on irregular meshes only considers performance gains resulting from considering network performance. 6.2.2 Irregular Meshes The experiments on irregular meshes were performed on the GUSTO testbed, which is not available when we experimented on the regular meshes. This testbed includes two IBM SP machines, one located at Argonne National Laboratory (ANL) and the other located at the San Diego Supercomputing Center (SDSC). These two machines are connected by vBNS (very high speed Backbone Network Service). We used Globus [15, 16] software to allow multimodal communication within the application. Macro benchmarks were used to determine the network and processor performance. The results of the network performance analysis are given in Table 5. Further, experiments were conducted to determine that the SDSC SP processors nodes were 1.6 times as fast as the ANL ones. Table 5: Values of ff and fi for the different networks. ANL SP Vulcan Switch ff SDSC SP Vulcan Switch ff PART is used to partition five 2D irregular meshes with triangular elements: barth4 (11451 labarre (14971 elem.), viscous (18369 elem.), and inviscid (6928 (called PART without restriction). A sightly modified version of PART (called PART with restriction) is used to partition the meshes so that only one processor has remote communication in each group. METIS 3.0 [26] is used to generate partitions that take into consideration processor performance (each processor's compute power is used as one of the inputs). These three partitioners are used to identify the performance impact of considering heterogeneity of networks in addition to that with processors. Further, the three partitioners highlight the difference when forcing remote communication to occur on one processor in a group versus having multiple processors with remote communication in a group. We consider 6 configurations of the two machines: 4 at ANL and 4 at SDSC, 8 at ANL and 8 at SDSC, and 20 at ANL and 20 at SDSC. The two groups correspond to the two IBM SPs at ANL and SDSC. We used up to 20 processors from each SP due to limitations in co-scheduling computing resources. The execution time is given for 100 time steps. The recorded execution time represents an average of 10 runs, and the standard deviation is less than 3%. Tables 6 to Table 8 show the experimental results from the 3 configurations. Column one identifies the irregular meshes and the number of elements in each mesh (included in parenthesis). Column two is the execution time resulting from the partitions from PART with the restriction that only one processor per group entails remote communication. For Columns 2 to 4, the number - indicates the number of processors that has remote communication in a group. Column three is similar to Column two except that the partition does not have the restriction that remote communication be on one processor. Column four is the execution time resulting from METIS which takes computing power into consideration (each processor's compute power is used as one of Table Execution time using the vBNS on 8 processors: 4 at ANL, 4 at SDSC. Mesh PART w/ restriction PART w/o restriction Proc. Perf. (METIS) efficiency viscous (18369 elem.) 150.0s (-=1) 169.0s (-=3) 170.0s (-=3) efficiency 0.86 0.75 0.75 labarre (14971 elem.) 133.0s (-=1) 142.0s (-=2) 146.0s (-=3) efficiency 0.79 0.73 0.71 efficiency 0.79 0.68 0.68 inviscid (6928 elem.) 73.2s (-=1) 85.5s (-=3) 88.5s(-=3) efficiency 0.66 0.56 0.55 the inputs to the METIS program). The results show that by using PART without restrictions, a slight decrease (1-3%) in execution time is achieved as compared to METIS. By forcing all the remote communication on one processor, the retrofit step can achieve more significant reduction in execution time. The results in Tables 6 to Table 8 show that efficiency is increased by up to 36% as compared to METIS, and the execution time is reduced by up to 30% as compared to METIS; This reduction comes from the fact that even on a high speed network such as the vBNS, the difference of message start up cost on remote and local communication is very large. From Table 5, we see this difference is two orders of magnitude for message start up as compared to approximately one order of magnitude for bandwidth. Restricting remote communication on one processor allows PART to redistribute the load among more processors thereby achieving close to the ideal reduction in execution time. 7 Previous Work The problem of domain partitioning for finite element meshes is equivalent to partitioning the graph associated with the finite element mesh. Graph partitioning has been proven to be an Table 7: Execution time using the vBNS on processors: 8 at ANL, 8 at SDSC. Mesh PART w/ restriction PART w/o restriction Proc. Perf. (METIS) efficiency 0.72 0.62 0.59 viscous (18369 elem.) 82.9s(-=1) 100.8s(-=4) 106.0s(-=5) efficiency 0.77 0.64 0.61 labarre (14971 elem.) 75.8s(-=1) 83.7s(-=3) 88.6s(-=3) efficiency 0.69 0.62 0.59 efficiency 0.74 0.50 0.48 inviscid (6928 elem.) 42.2s(-=1) 62.8s(-=3) 67.2s(-=4) efficiency 0.57 0.39 0.36 NP-complete problem [17]. Many good heuristic static partitioning methods have been proposed. Kernighan-Lin [31] proposed a locally optimized partitioning method. Farhat [13, 14] proposed an automatic domain decomposer based on Greedy algorithm. Berger and Bokhari [4] proposed Recursive Coordinate Bisection (RCB) which utilizes spatial nodal coordinate information. Nour-Omid et al. [35, 40] proposed Recursive Inertial Bisection (RIB). Simon [37] proposed Recursive Spectral Bisection (RSB) which computes the Fiedler vector for the graph using the Lanczos algorithm and then sorts vertices according to the size of the entries in Fiedler vector. Recursive Graph Bisection (RGB) is proposed by George and Liu [18], which uses SPARSPAK RCM algorithm to compute a level structure and then sort vertices according to the RCM level structure. Barnard et al. in [2] proposed a multilevel version of RSB which is faster. Hendrickson and Leland [23, 22] also reported a similar multilevel partitioning method. Karypis and Kumar [27, 28, 30] proposed a new coarsening heuristic to improve the multilevel method. Most of the aforementioned decomposition methods are available in one of three automated tools: Chaco [22], METIS [26, 29] and TOP/DOMDEC [38]. Chaco, the most versatile, implements inertial, spectral, Kernighan-Lin, and multilevel algorithms. These algorithms are used to Table 8: Execution time using the vBNS on 40 processors: 20 at ANL, 20 at SDSC. Mesh PART w/ restriction PART w/o restriction Proc. Perf. (METIS) efficiency 0.69 0.54 0.45 viscous (18369 elem.) 38.7s(-=1) 58.6s(-=5) 64.9s(-=7) efficiency 0.67 0.44 0.40 labarre (14971 elem.) 33.8s(-=1) 51.2s(-=3) 53.5s(-=6) efficiency 0.62 0.41 0.40 efficiency 0.39 0.34 0.32 inviscid (6928 elem.) 33.5(-=1) 34.7s(-=4) 46.8s(-=5) efficiency recursively bisect the problem into equal sized subproblems. METIS uses the method for fast partitioning of the sparse matrices, using a coarsening heuristic to provide the speed. TOP/DOMDEC is an interactive mesh partitioning tool. All these tools produce equal size partitions. These tools are applicable to systems with the same processors and one interconnection network. Some tools such as METIS, can produce partitions with unequal weights. However, none of these tools can take network performance into consideration in the partitioning process. For this reason, these tools are not applicable to distributed systems. Crandall and Quinn [7, 8, 9, 10, 11, 12] developed a partitioning advisory system for network of workstations. The advisory system has three built-in partitioning methods (contiguous row, contiguous point, and block). Given information about the problem space, the machine speed, and the network, the advisory system provides ranking of the three partitioning methods. The advisory system takes into consideration of variance in processor performance among the workstations. The problem, however, is that linear computational complexity is assumed for the application. This is not the case with implicit finite element problems, which are widely used. Further, variance in network performance is not considered. 8 Conclusion In this paper, we addressed issues in mesh partitioning problem for distributed systems. These issues include comparison metric, efficiency, and cut sets. We present a tool, PART, for automatic mesh partitioning for distributed systems. The novel feature of PART is that it considers heterogeneities in both the application and the distributed system. The heterogeneities in the distributed system include processor and network performance; the heterogeneities in the application include computational complexity. We also demonstrate the use of a parallel version of PART for distributed systems. The novel part of the parallel PART is that it uses the asynchronous multiple Markov chain approach of parallel simulated annealing for mesh partitioning. The parallel PART is used to partition 6 irregular meshes into 8, 16, and 100 subdomains using up to 64 client processors on an IBM SP2 machine. Results show superlinear speedup in most cases and nearly perfect speedup for the rest. We used Globus software to run an explicit, 2-D finite element code using mesh partitions from the parallel PART. Our testbed includes two geographically distributed IBM SP machines. Experimental results are presented for 3 regular meshes and 4 irregular finite element meshes for the WHAMS2D application executing on a distributed system consisting of two IBM SPs. The results from the regular problems indicate a increase in efficiency when processor performance is considered as compared to even partitioning; the results also indicate an additional 5 \Gamma 16% increase in efficiency when network performance is considered. The result from the irregular problem indicate a 38% increase in efficiency when processor and network performance are considered as compared to even partitioning. Experimental results from the irregular problems also indicate up to 36% increase in efficiency compared with using partitions that only take processor performance into consideration. This improvement comes from the fact that even on a high speed network such as the vBNS, the message start up cost on remote and local communication still has a large difference. Appendix 1: Proof of NP-complete of the Mesh Partitioning Problem for Distributed Systems partitioning problem for distributed systems is NP-complete. Proof 1 We transform a proven NP-Complete problem, MINIMUM SUM OF SQUARES [17], to the Partition problem for distributed systems. Let set A , with for each a 2 A be an arbitrary instance of MINIMUM SUM OF SQUARES. We shall construct a graph the desired partition exists for G if and only if A has a sum of squares. The basic units of MINIMUM SUM OF SQUARES instance are a n. The local replacement substitute for each a i 2 A is the collection E i of 3 edges shown in Figure 10. Therefore, E) is defined as the following: a a [2] a [1] Figure 10: Local replacement for a i 2 A for transforming MINIMUM SUM OF SQUARES to the Partition problem for distributed systems. It is easy to see this instance of Partition problem for distributed systems can be constructed in polynomial time from the MINIMUM SUM OF SQUARES instance. are the disjoint k partitions of A such that the sum of squares is minimized, then the corresponding k disjoint partitions of V is given by taking fa i [1]; a i [2]; a i g for each a i in every subset of A. We also restrict the cost function f i to be the same as is in MINIMUM SUM OF SQUARES: and a 2 A i . This ensures that the partition sum of squares of the cost function. Conversely, if is a disjoint k partition of G with minimum sum of squares of the cost function, the corresponding disjoint k partition of set A is given by choosing those vertices a i such that fa i [1]; a i [2]; a i Hence the minimum sum of squares for the cost function over k disjoint partitions ensures that the sum of squares of s(a) on k disjoint set of A is also minimized. We conclude that the Partition problem for distributed systems is NP-Complete. Appendix 2: Nomenclature Estimated execution time on processor i. Estimated computational time on processor i. Estimated communication time on processor i. Performance of processor i as measured by a computation kernel. ff L - Per message cost of local communication. message cost of remote communication. cost of local communication. cost of remote communication. - Size of message. CL - Local communication time for a processor. CR - Remote communication time for a processor. - The difference between CR and CL for one processor. The difference between CR and CL for processor i. - The maximum number of processors in a group that have both local and remote communication. Coefficient of computational complexity. Parameter used to equalize the contribution of the computation and communication to execution time. Number of elements in partition i. - Number of processors in the system. Number of processors in group i. G - Number of processors in a particular group (same as P i ). - Number of groups in the system. The i-th group in the system. The ratio of the speed of processors in S i relative to the slowest processor in the system. --R Parallel simulated annealing algorithms for cell placement on hypercube multiprocessors. A fast multilevel implementation of recursive spectral bisection for partitioning unstructured problems. Finite Element Procedures in Engineering Analysis. A partitioning strategy for non-uniform problems on multiproces- sors An evaluation of parallel simulated annealing strategies with application to standard cell placement. WHAMS3D project progress report PR-2 Data partitioning for networked parallel processing. Problem decomposition in parallel networks. Block data partitioning for partial-homogeneous parallel networks Evaluating decomposition techniques for high-speed cluster computing A partitioning advisory system for networked data-parallel processing A simple and efficient automatic fem domain decomposer. Automatic partitioning of unstructured meshes for the parallel solution of problems in computational mechanics. Managing multiple communication methods in high-performance networked computing systems Software infrastructure for the i-way meta-computing experiment Computers and Intractability: A Guide to the Theory of NP- Completeness Computer Solution of Large Sparse Positive Definite Systems. Parallel simulated annealing techniques. the Legion team. Simulated annealing based parallel state assignment of finite state machines. The chaco user's guide. A multilevel algorithm for partitioning graphs. The Finite Element Method. The internet fast lane for research and education. A fast and high quality multilevel scheme for partitioning irregular graphs. A fast and high quality multilevel scheme for partitioning irregular graphs. Multilevel k-way partitioning scheme for irregular graphs Multilevel k-way partitioning scheme for irregular graphs Parallel multilevel k-way partitioning scheme for irregular graphs An efficient heuristic procedure for partitioning graphs. Placement by simulated annealing on a multiprocessor. Introduction to Parallel Computing: Design and Analysis of Algorithms. Asynchronous communication of multiple markov chain in parallel simulated annealing. Solving finite element equations on concurrent computers. Partitioning sparse matrices with eigenvectors of graphs. Partitioning of unstructured problems for parallel processing. Top/domdec: a software tool for mesh partitioning and parallel processing. Parallel n-ary speculative computation of simulated annealing A study of the factorization fill-in for a parallel implementation of the finite element method A retrofit based methodology for the fast generation and optimization of large-scale mesh partitions: Beyond the minimum interface size criterion --TR A partitioning strategy for nonuniform problems on multiprocessors Partitioning sparse matrices with eigenvectors of graphs Parallel simulated annealing techniques Introduction to parallel computing Three-dimensional grid partitioning for network parallel processing The Legion vision of a worldwide virtual computer Managing multiple communication methods in high-performance networked computing systems Simulated annealing based parallel state assignment of finite state machines Computer Solution of Large Sparse Positive Definite Computers and Intractability Parallel Simulated Annealing Algorithms for Cell Placement on Hypercube Multiprocessors Parallel N-ary Speculative Computation of Simulated Annealing Mesh Partitioning for Distributed Systems Problem Decomposition in Parallel Networks --CTR Kyungmin Lee , Dongman Lee, A scalable dynamic load distribution scheme for multi-server distributed virtual environment systems with highly-skewed user distribution, Proceedings of the ACM symposium on Virtual reality software and technology, October 01-03, 2003, Osaka, Japan Zhiling Lan , Valerie E. Taylor , Greg Bryan, Dynamic load balancing of SAMR applications on distributed systems, Scientific Programming, v.10 n.4, p.319-328, December 2002 Zhiling Lan , Valerie E. Taylor , Greg Bryan, Dynamic load balancing of SAMR applications on distributed systems, Proceedings of the 2001 ACM/IEEE conference on Supercomputing (CDROM), p.36-36, November 10-16, 2001, Denver, Colorado

simulated annealing;distributed systems;mesh partitioning

506198

Speculative Versioning Cache.

Dependences among loads and stores whose addresses are unknown hinder the extraction of instruction level parallelism during the execution of a sequential program. Such ambiguous memory dependences can be overcome by memory dependence speculation which enables a load or store to be speculatively executed before the addresses of all preceding loads and stores are known. Furthermore, multiple speculative stores to a memory location create multiple speculative versions of the location. Program order among the speculative versions must be tracked to maintain sequential semantics. A previously proposed approach, the Address Resolution Buffer (ARB) uses a centralized buffer to support speculative versions. Our proposal, called the Speculative Versioning Cache (SVC), uses distributed caches to eliminate the latency and bandwidth problems of the ARB. The SVC conceptually unifies cache coherence and speculative versioning by using an organization similar to snooping bus-based coherent caches. Our evaluation for the Multiscalar architecture shows that hit latency is an important factor affecting performance and private cache solutions trade-off hit rate for hit latency.

Introduction Modern microprocessors extract instruction level parallelism (ILP) from sequential programs by issuing instructions from an active instruction window. Data dependences among instructions, and not the original program order, determine when an instruction may be issued from the win- dow. Dependences involving register data are detected easily because register designators are completely specified within instructions. However, dependences involving memory data (e.g. between a load and a store or two stores) are ambiguous until the memory addresses are computed. A straightforward solution to the problem of ambiguous memory dependences is to issue loads and stores only after their addresses are determined. Furthermore, a store is not allowed to complete and commit its result to memory until all preceding instructions are known to be free of ex- ceptions. Each such store to a memory location creates a speculative version of that location. These speculative versions are held in buffers until they can be committed. Multiple speculative stores to the same location create multiple versions of the location. To improve performance, loads are allowed to bypass buffered stores, as long as they are to different addresses. If a load is to the same address as a buffered store, it can use data bypassed from the store when the data becomes available. An important constraint of this approach is that a load instruction cannot be issued until the addresses of all the preceding stores are determined. This approach may diminish ILP unnecessarily, especially in the common case where the load is not dependent on preceding stores. More aggressive uniprocessor implementations issue load instructions as soon as their addresses are known, even if the addresses of all previous stores may not be known. These implementations employ memory dependence speculation [8] and predict that a load does not depend on previous stores. Furthermore, one can also envision issuing and computing store addresses out of order. Such memory dependence speculation enables higher levels of ILP, but more advanced mechanisms are needed to support this specula- tion. These aggressive uniprocessors dispatch instructions from a single instruction stream, and issue load and store instructions from a common set of hardware buffers (e.g. reservation stations). Using a common set of buffers allows the hardware to maintain program order of the loads and stores via simple queue mechanisms, coupled with address comparison logic. The presence of such queues provides support for a simple form of speculative versioning. However, proposed next generation processor designs use replicated processing units that dispatch and/or issue instructions in a distributed manner. These future approaches partition the instruction stream into sub streams called tasks [11] or traces [10]. Higher level instruction control units distribute the tasks to the processors for execution and, the processors execute the instructions within each task leading to a hierarchical execution model. Proposed next generation multiprocessors [9, 12] that provide hardware support for dependence speculation also use such execution models. A hierarchical execution model naturally leads to memory address streams with a similar hierarchical structure. In particular, each individual task generates its own address stream, which can be properly ordered (dis- ambiguated) within the processor that generates it, and at the higher level, the multiple address streams produced by the processors must also be properly ordered. It is more challenging to support speculative versioning for this execution model than a superscalar execution model because a processor executes loads and stores without knowing those executed by other processors. The Address Resolution Buffer [3] (ARB) provides speculative versioning support for such hierarchical execution models. Each entry in the ARB buffers all versions of the same memory location. However, there are two significant performance limitations of the ARB: 1. The ARB is a single shared buffer connected to the multiple processors and hence, every load and store incurs the latency of the interconnection network. Also, the ARB has to provide sufficient bandwidth for all the processors. 2. When a task completes all its instructions, the ARB commits its speculative state into the architected storage (or copies all the versions created by this task to the data cache). Such write backs generate bursty traffic and can increase the time to commit a task, which delays the issue of new task to that processor and lowers the overall performance. We propose a new solution for speculative versioning called the Speculative Versioning Cache [2, 5] (SVC), for hierarchical execution models. The SVC comprises a private cache for each processor, and the system is organized similar to a snooping bus-based cache coherent Symmetric Multiprocessor (SMP). Memory references that hit in the private cache do not use the bus as in an SMP. Task commits do not write back speculative versions en masse. Each cache line is individually handled when it is accessed the next time. Section 2 introduces the hierarchical execution model briefly and identifies the issues in providing support for speculative versioning for such execution models. Section 3 presents the SVC as a progression of designs to ease under- standing. Section 4 gives a preliminary performance evaluation of the SVC to highlight the importance of a private cache solution for speculative versioning. We derive conclusions in section 5. 2. Speculative versioning First, we discuss the issues involved in providing support for speculative versioning for current generation processors. Second, we describe the hierarchical execution model used by the proposed next generation processors. Third, we discuss the issues in providing support for speculative versioning for this execution model and use examples to illustrate them. Finally, we present similarities between multiprocessor cache coherence and speculative versioning for the hierarchical execution model and use this unification to motivate our new design, the speculative versioning cache. Speculative versioning involves tracking the program order among the multiple buffered versions of a location to guarantee the following sequential program semantics: ffl A load must eventually read the value created by the most recent store to the same location. This requires that (i) the load must be squashed and re-executed if it executes before the store and incorrectly reads the previous version and, (ii) all stores (to the same location) that follow the load in program order must be buffered until the load is executed. ffl A memory location must eventually have the correct version independent of the order of the creation of the ver- sions. Consequently, the speculative versions of a location must be committed to the architected storage in program order. 2.1. Hierarchical execution model In this execution model, the dynamic instruction stream of a program is partitioned into fragments called tasks. These tasks form a sequence corresponding to their order in the dynamic instruction stream. A higher level control unit predicts the next task in the sequence and assigns it for execution to a free processor. Each processor executes the instructions in the task assigned to it and buffers the speculative state created by the task. The Wisconsin Multiscalar [11] is an example architecture that uses the hierarchical execution model. When a task misprediction is detected, the speculative state of all the tasks in the sequence including and after the incorrectly predicted task are invalidated 1 and the corresponding processors are freed. This is called a task squash. The correct tasks in the sequence are then assigned for exe- cution. When a task prediction has been validated, it commits by copying the speculative buffered state to the architected storage. Tasks commit one by one in the order of the sequence. Once a task commits, its processor is free to execute a new task. Since the tasks commit in program order, tasks are assigned to the processors in program order. An alternative model for recovery invalidates only the dependent (a) (b) Figure 1: Task commits and squashes: example. Figure commits and task squashes. Ini- tially, tasks 0, 1, 99 and 3 are predicted and speculatively executed in parallel by the four processors as shown in Figure 1(a). When the misprediction of task 99 is detected, tasks 99 and 3 are squashed and their buffered states are invalidated. New tasks 2 and 3 are then executed by the processors as show in Figure 1(b). Tasks that are currently executing are said to be active. When task 0 completes exe- cution, the corresponding processor is freed and task 4 is assigned for execution as shown in Figure 1(c). The program order, represented by the sequence among the tasks, enforces an implicit total order among the processors; the arrows show this order. When the tasks are speculatively executed in parallel, the multiple speculative load/store streams from the processors are merged in arbitrary order. Providing support for speculative versioning for such execution models requires mechanisms that establish program order among these streams. The following subsections outline how the order is established using the sequence among the tasks. 2.1.1. Loads A task executes a load as soon as its address is available, speculating that stores from previous tasks in the sequence do not write to the same location. The closest previous version of the location is supplied to the load; this version could have been created either by the same task or by a previous task. A load that is supplied a version from a previous task is recorded to indicate a use before a potential definition. If such a definition (a store to the same location from a previous task) occurs, the load was supplied with an incorrect version and memory dependence was violated. 2.1.2. Stores When a task executes a store to a memory location, it is communicated to all later active tasks in the sequence 2 . When a task receives a new version of a location from a previous task, it squashes if a use before definition is recorded for that location - a memory dependence violation is detected. All tasks after the squashed task are also squashed as on a task misprediction (simple squash model). 2.1.3. Task commits and squashes The oldest active task is non-speculative and can commit its speculative memory state (versions created by stores from this task) to ar- chains of instructions by maintaining information at a finer granularity. This paper assumes the simpler model. In reality, the store has to be communicated only until the task that has created the next version, if any, of the location. chitected storage. Committing a version involves logically copying the versions from the speculative buffers to the architected storage (data cache). As we assume the simple task squash model, the speculative state associated with a task is invalidated when it is squashed. 2.2. Examples for speculative versioning Figure 2 illustrates the issues involved in speculative versioning using an example program and a sample execution of the program on a four processor hierarchical system. We use the same example in the later sections to explain the SVC design. Figure 2(a) shows the loads and stores in the example program and the task partitioning. Other instructions are not of direct relevance here. Figure 2(b) shows two snapshots of the memory system during a sample execution of the program. Each snapshot contains four boxes, one for each active task and shows the load or store that has been executed by the corresponding task. The program order among the instructions translates to a sequence among the tasks which imposes a total order among the processors executing them; solid arrowheads show the program order and hollow arrowheads show the execution time order in all the examples. Example Program (a) (b) Sample Execution st 3, A ld r, A st 0, A23 st 0, A st 1, A st 0, A st 1, A st 3, A st 5, A ld r, A ld r, A Program Order31st 1, A Dependence Violation23 Figure 2: Speculative versioning example. The first snapshot is taken just before task 1 executes a store to address A. Tasks 0 and 3 have already stored and 3 to A and task 2 has executed a load to A. The load is supplied the version created and buffered by task But, according to the original program, this load must be supplied the value 1 created by the store from task 1, i.e., the store to load dependence has been violated. This violation is detected when task 1 stores to address A and all the tasks including and after task 2 are squashed and re- executed. The second snapshot is taken after the tasks have been squashed and re-started. 2.3. Coherence and speculative versioning The actions performed on memory accesses and task commits and squashes are summarized in Table 1. The functionality in this table requires the hardware to track the active tasks or processors that executed a load/store to a location and the order among the different copies/versions of this location. Cache coherent Symmetric MultiProcessors use similar functionality to track the caches that have a copy of every memory location. SMPs, however, need not track the order among these copies since all the copies are of a single version. Event Actions Load Record use before definition by the task; supply the closest previous version. Store Communicate store to later tasks; later tasks look for memory dependence violations. Commit Write back buffered versions created by the task to main memory. Squash Invalidate buffered versions created by the task. Table 1: Versioning: events and actions. SMPs typically use snooping [4] to implement a Multiple Reader/Single Writer protocol, which uses a coherence directory that is a collection of sets, each of which tracks the sharers of a line. In a snooping bus based SMP, the directory is typically implemented in a distributed fashion comprising state bits associated with each cache line. On the other hand, the Speculative Versioning Cache (SVC) implements a Multiple Reader/Multiple Writer protocol that tracks copies of multiple speculative versions of each memory location. This protocol uses a version directory that maintains ordered sets for each line, each of which tracks the program order among the multiple speculative versions of a line. This ordered set or list, called the Version Ordering List (VOL), can be implemented in several different ways - the SVC, proposed in this paper, uses explicit pointers in each line to implement it as a linked list (like in SCI [1]). The following sections elaborate on a design that uses pointers in each cache line to maintain the VOL. The private cache organization of the SVC makes it a feasible memory system for proposed next generation single chip multiprocessors that execute sequential programs on tightly coupled processors using automatic parallelization [9, 12]. Previously, ambiguous memory dependences limited the range of programs chosen for automatic parallelization. The SVC provides hardware support to overcome ambiguous memory dependences and enables more aggressive automatic parallelization of sequential programs 3. SVC design In this section, we present the Speculative Versioning Cache (SVC) as a progression of designs to ease under- standing. Each design improves the performance over the previous one by tracking more information. We begin with a brief review of snooping bus-based cache coherence and then present a base SVC design which provides support for speculative versioning with minimal modifications to the cache coherence scheme. We then highlight the performance bottlenecks in the base design and introduce optimizations one by one in the rest of the designs. 3.1. Snooping bus based cache coherence Figure 3 shows a 4-processor SMP with private L1 caches that uses a snooping bus to keep the caches consis- tent. Each cache line comprises an address tag that identifies the data that is cached, the data that is cached, and two bits (valid and store) representing the state of the line. The valid (V ) bit is set if the line is valid. The store (S) or dirty bit is set when a processor stores to the line. Bus Arbiter V: Valid S: Store or dirty Tag V S Data Next level memory Snooping Bus Figure 3: SMP coherent cache. A cache line is in one of three states: Invalid, Clean and Dirty. A request (load or store) from a processor to its L1 cache hits if a valid line with the requested tag is in an appropriate state; otherwise, it misses. Cache misses issue bus requests while cache hits do not. More specifically, a load from a clean or dirty line and a store to a dirty line result in cache hits. Otherwise, the load(store) misses and the cache issues a BusRead(BusWrite) request. The L1 caches and the next level memory snoop the bus on every request. If a cache has a valid line with the requested tag, it issues an appropriate response according to a coherence protocol. A store to a clean line misses and the cache issues a BusWrite request. An invalidation-based coherence protocol invalidates copies of this line in all other caches, if any. This protocol allows a dirty line to be present in only one of the caches. However, a clean line can be present in multiple caches simultaneously. The cache with the dirty line supplies the data on a BusRead request. A cache issues a BusWback request to cast out a dirty line on a replacement. This simple protocol can be extended by adding an exclusive bit to the state of each line to cut down traffic on the shared bus. If a cache line has the exclusive bit set, then it has the only valid copy of the line and can perform a store to that line locally. The SVC designs we discuss in the following sections also use an invalidation-based protocol. Z Y Z Y Z Y Z Y BusWback Flush Replace BusWrite Invalidate ld r, A BusRead Flush st 1, A 0State Data W.Z: Caches Figure 4: Cache coherence example. Figure 4 shows snapshots of the cache lines with tag or address A in an SMP with four processors, W , X , Y , and Z. The state of the cache line is shown in a box corresponding to that cache. An empty box corresponding to a cache represents that the line is not present in that cache. The first snapshot is taken before processor Z issues a load from A and misses in its private cache. The cache issues a BusRead request and cache X supplies the data on the bus. The second snapshot shows the final state of the lines; they are clean. Later, processor Y issues a BusWrite request to perform a store to A. The clean copies in caches X and Z are invalidated and the third snapshot shows the final state. chooses to replace this line, it casts out the line to memory by issuing a BusWback request; the final state is shown in the fourth snapshot; only the next level memory contains a valid copy of the line. 3.2. Base SVC design The organization of the private L1 caches in the SVC design is shown in Figure 5; all the SVC designs use the same organization. The base design minimally modifies the memory system of the snooping bus-based cache coherent SMP to support speculative versioning for processors based on the hierarchical execution model. We assume that memory dependences among loads and stores executed by an individual processor are ensured by a conventional load-store queue; our design guarantees program order among loads and stores from different processors. The base design also assumes that the cache line size is one word; a later design relaxes this assumption. First, we introduce the modifications to the SMP coherent cache, and then discuss how the individual operations listed in Table 1 are performed. 1. Each cache line maintains an extra state bit called the load (L) bit, as shown in Figure 6. The L bit is set when a task loads from a line before storing to the line - a potential Bus Arbiter Version Control Logic Next level memory Snooping Bus Version Control Logic Task assignment information VCL responses to each cache States of snooped lines from each cache Figure 5: Speculative versioning cache. V: Valid S: Store L: Load Tag Pointer Data Figure Base SVC design: structure of a line. violation of memory dependence in case a previous task stores to the same line. 2. Each cache line maintains a pointer that identifies the processor (or L1 cache) that has the next copy/version, if any, in the Version Ordering List (VOL) for that line. Thus, the VOL for a line is stored in a distributed fashion among the private L1 cache lines. It is important to note that the pointer identifies a processor rather than a task. Storing the VOL explicitly in the cache lines using pointers may not be necessary for the base design. However, it is necessary to explicitly store the VOL for the advanced designs and we introduce it in the base design to ease the transition to the advanced designs. 3. The SVC uses combinational logic called the Version Control Logic (VCL) that provides support for speculative versioning using the VOL. A processor request that hits in the private L1 cache does not need to consult the VOL and hence does not issue a bus request; the VCL is also not used. Cache misses issue a bus request that is snooped by the L1 caches and the next level memory. The states of the requested line in each L1 cache and the VOL are supplied to the VCL. The VCL uses the bus re- quest, the program order among the tasks, and the VOL to compute appropriate responses for each cache. Each cache line is updated based on its initial state, the bus request and the VCL response. A block diagram of the Version Control Logic is shown in Figure 5. For the base design, the VCL responses are similar to that of the disambiguation logic in the ARB [3]. The disambiguation logic searches for previous or succeeding stages in a line to execute a load or store, respectively. 3.2.1. Loads Loads are handled in the same way as in an SMP except that the L bit is set if the line was initially in- valid. On a BusRead request, the VCL locates the closest previous version by searching the VOL in the reverse order beginning from the requestor; this version, if any, is supplied to the requestor. If a previous version is not buffered in any of the L1 caches, the next level memory supplies the data. Task assignment information is used to determine the position of the requestor in the VOL. The VCL can search the VOL in reverse order because it has the entire list available and the list is short. 0.3: Tasks Data Pointer State W.Z: Caches Z ld r, A Execution time order Program order Figure 7: Base SVC design: example load. We illustrate the load executed by task 2 to address A in the example program. Figure 7 shows two snapshots: one before the load executes and one after the load completes. Each box shows the line with tag or address A in an L1 cache (the valid bit is not explicitly shown). The number adjacent to a box gives a processor/cache identifier and a task identifier. The processor identifiers are used by the explicit pointers in each line to represent the VOL, whereas, the task identifiers serve only to ease the explanation of the examples. Task 2 executes a load that misses in cache Z and results in a bus request. The VCL locates cache Z in the VOL for address A using program order and then searches the VOL in the reverse order to find the correct version to supply, which is the version in cache Y (the version created by task 1). 3.2.2. Stores The SVC performs more operations on a store miss as compared to a cache coherent SMP. When a BusWrite request is issued on a store miss, the VCL sends invalidation responses to the caches beginning from the re- questor's immediate successor (in task assignment order) to the cache that has the next version (including it, if it has the L bit set). This invalidation response allows for multiple versions of the same line to exist and also serves to detect memory dependence violations. A cache sends a task squash signal to its processor when it receives an invalidation response from the VCL and the L bit is set in the line. Z Z Z st 3, A st 1, A W/- Figure 8: Base SVC design: example stores. We illustrate the stores executed by tasks 1 and 3 in the example program. Figure 8 shows four snapshots of the cache lines with address A. The first snapshot is taken before task 3 executes a store that results in a BusWrite re- quest. Since task 3 is the most recent in program order, the store by task 3 does not result in any invalidations. Note that a store to a line does not invalidate all other cache lines (unlike an SMP) to allow for multiple versions of the same line. The second snapshot is taken after the store from task 3 completes and before task 1 executes its store. Based on task assignment information, the VCL sends an invalidation response to each cache from the one after cache Y until the one before cache W , which has the next version of the line (cache W is not included since it does not have the L bit set) sends an invalidation response to cache Z. But, the load executed by task 2, which follows the store by task 1 in program order, has already executed. Cache Z detects a memory dependence violation since the L bit is set when it receives an invalidation response from the VCL. Tasks 2 and 3 are squashed as shown in the third snapshot by shaded boxes. The final snapshot is taken after the store by task 1 has completed. 3.2.3. Task commits and squashes The base SVC design handles task commits and squashes in a naive man- ner. When a processor commits a task, all dirty lines in its L1 cache are immediately written back to the next level memory and all other lines are invalidated. To write back all the dirty lines immediately, a list of the stores executed by the task is maintained by the processor. When a task is squashed, all lines in the corresponding cache are invalidated 3.3. Base design performance drawbacks The base design just described has two significant performance limitations that make it less desirable: (i) write backs performed when a processor commits a task lead to bursty bus traffic that may increase the time to commit the task and delay issuing a new task to that processor, (ii) clean lines are also invalidated when a task commits or squashes because the buffered versions could be stale for the new task allocated on the same processor; the correct version may be present in other caches. Consequently, every task begins execution with a cold L1 cache, increasing the bandwidth demand. The following advanced designs eliminate these problems by tracking additional information. 1. The first advanced design, the ECS design (section 3.5), makes task commits and squashes more efficient. To ease the understanding of this design, we first present an intermediate design, the EC design (section 3.4), that makes task commits efficient by distributing the write backs of dirty lines over time. Also, it retains read-only data in the L1 caches across task commits by careful book-keeping. However, it assumes that mispredictions do not occur. Then, we present the ECS design that extends the EC design to allow task squashes. Task squashes are as simple as in the base design, but are more efficient as they retain non-speculative data in the caches across task squashes. 2. The second advanced design (section 3.6) boosts the hit-rate of the ECS design by allowing requests to snarf [6] the bus to account for reference spreading. Snarfing involves copying the data supplied on a bus request issued by another processor in an attempt to combine bus requests indirectly. 3. The final design (section 3.7) is realistic and allows the size of a cache line to be more than one word. 3.4. Implementing efficient task commits (EC) The EC design avoids expensive cache flushes on task commits by maintaining an extra state bit, called the commit bit, in each cache line. Task commits do not stall until all lines with speculative versions are written back. The EC design eliminates write back bursts on the bus during task commits. Also, no extra hardware is necessary to maintain a list of stores performed by each task. Further, the EC design improves cache utilization by keeping the L1 caches warm across tasks. Tag S L V Data V: Valid L: Load C: Commit S: Store T: sTale Figure 9: EC design: structure of a line. The structure of a cache line in the EC design is shown in Figure 9. When a processor commits a task, the C bit is set in all its cache lines. This operation is entirely local to the L1 cache and does not issue a bus request. A dirty committed line is written back, if necessary, when it is accessed the next time either on a processor request or on a bus request. Therefore, committed versions could remain in the caches until much later in time since the task that created the version committed. The order among committed and uncommitted versions is still maintained by the explicit pointers in the line. This order among the versions is necessary to write back the correct committed version and to supply the correct version on a bus request. The EC design uses an additional state bit, the sTale (T ) bit, to retain read-only data across tasks. First, we discuss how loads and stores are handled when caches have both committed and uncommitted versions and then discuss the stale bit. 3.4.1. Loads and stores Loads to committed lines are handled like cache misses and issue a bus request. The VCL searches the VOL in the reverse order beginning from the requestor for the closest previous uncommitted version; this version, if any, is supplied to the requestor. If no such version is found, the VCL supplies the most recent committed version, if any. This version is the first committed version that is encountered on the reverse search. All other committed versions need not be written back and are invalidated. On a store miss, committed versions are purged in a similar fashion. ld r, A Figure 10: EC design: example load. We illustrate the load executed by task 2 in the example program. Figure 10 shows two snapshots: one before the load executes and one after the load completes. Versions 0 and 1 have been committed (the C bit is set in the lines in caches X and Y ). Task 2 executes a load that misses in cache Z and results in a bus request. The VCL knows that task 2 is the head task and determines that cache X has the most recent committed version. Cache X supplies the data which is also written back to the next level memory. Other committed versions (version are invalidated and are never written back to memory. The VCL also inserts the new copy of version 1 into the VOL by modifying the pointers in the lines accordingly - the second snapshot shows the modified VOL. Figure 11 illustrates the actions performed on a store miss. The first snapshot is taken before a store is executed by task 5. Versions 0 and 1 have been committed. Task 5 executes a store that misses in cache Y and results in a BusWrite request even though the line has a committed ver- sion. The VCL purges all committed versions of this line - it determines that version 1 has to be written back to the next level memory and the other versions (version be invalidated. Purging the committed versions also makes space for the new version (version 5). The modified VOL shown in the second snapshot contains only the two uncommitted versions. Y st 5, A Figure 11: EC design: example store. 3.4.2. Stale copies The EC design makes task commits efficient by delaying to commit each cache line until a later time. Therefore, a cache line could have a stale copy because versions more recent than the version buffered by the committed task could be present in other caches. The base SVC design does not introduce stale copies because it invalidates all non-dirty lines whenever a task commits. The EC design uses the stale (T ) bit to distinguish stale copies from correct copies and avoids issuing a bus request on accesses to correct copies. This additional information allows the EC design to retain read-only data (correct copies) across task commits. First, we illustrate that stale and correct copies are indistinguishable without the T bit and then show how the T bit is used. ld r, A Z CS Y CL ld r, A CL W/3 Y/5 CS CS CSS W/7Z Z Figure 12: EC design: correct and stale copies. Figure 12 shows two execution time lines - one that leaves a correct copy of address A (shown using solid lines) in cache Z and another that leaves a stale copy of address A in the same cache (shown using dashed lines). The first time line shows a sample execution of a modified version of our example program - task 3 in Figure 2 does not execute the store. The second time line shows an execution of our original program. The first snapshot is the same for both time lines. The second snapshot in the second time line is taken after tasks 0 and 1 have committed. The C bit is set in their caches and new tasks 4 and 5 have been allocated. The final snapshot in both time lines are taken when tasks 4 to 7 are active and before task 6 executes a load. In the first time line, the data in cache Z is a correct copy, since no versions were created after version 1; the load can be supplied data by just resetting the C bit and without issuing a bus request. In the second time line, the copy in cache Z is stale since the creation of version 3 and hence the load misses resulting a bus request. However, cache Z cannot distinguish between these two scenarios and has to issue a request in both cases to consult the VOL and obtain a copy of the correct version. The EC design uses the stale (T ) bit to distinguish between these two scenarios and avoids the bus request whenever a copy is not stale. The design maintains the invariant: the most recent version of an address and its copies have the T bit reset and the other copies and versions have the T bit set. This invariant is easily guaranteed by resetting the T bit in the most recent version, or a copy thereof, when it is created and setting the T bit in the previous versions, if any. The T bits are updated on the BusWrite request issued to create a version or a BusRead request issued to copy a version and hence do not generate additional bus traffic. Since stores in different tasks can be executed out of program order, an active task could execute a store to a copy that has the T bit set (the copy is not stale for this task, but is stale for the next task allocated to the same processor). Figure 13 shows the two time lines in our example with the status of the T bit. Cache Z can distinguish between the correct copy (T bit is not set) and the stale copy (T bit is set). The load hits if a correct copy is present and no bus request is issued. ld r, A CSTCL ld r, A Z CST CST CLT Y Figure 13: EC design: Using the stale bit. The EC design eliminates the serial bottleneck in flushing the L1 cache on task commits by using the commit (C) bit. Also, this design retains non-dirty lines after task commits as long as they are not stale. More generally, read-only data used by a program is fetched only once into the L1 caches and never invalidated unless chosen to be replaced on a cache miss. Further a task commits by just setting the C bit in all lines in its L1 cache. 3.5. Implementing efficient task squashes (ECS) The ECS design extends the EC design to allow task squashes for the EC design. Also, the ECS design makes the task squashes more efficient than in the base design by retaining non-speculative data in the caches across squashes using another state bit, the Architectural (A) bit. The structure of a line in the ECS design is shown in Figure 14. V: Valid S: Store L: Load C: Commit T: sTale A: Architectural Tag S L Figure 14: ECS design: structure of a line. When a task squashes, all uncommitted lines (lines with the C bit reset) are invalidated by resetting the valid (V ) bit. The invalidation makes the pointers in these lines and their VOLs inexact. The VOL has a (dangling) pointer in the last valid (or unsquashed) copy or version of the line and the status of the T bit in the lines are incorrect. The ECS design repairs the VOL of such a line when the line is accessed later either on a processor request or on a bus request. Updating the T bits is not necessary because it is only a hint to avoid a bus request and a squash would not incorrectly reset a stale version to be correct. However, the ECS design updates the T bit on this bus request by consulting the repaired VOL. W/- CST ST ST W WZ Figure 15: ECS design: VOL repair. Figure repair with an example time line with three snapshots. The first snapshot is taken just before the task squash occurs. Tasks 3 and 4 are squashed; only version 3 is invalidated. The VOL with incorrect T bits and the dangling pointer are shown in the second snapshot. executes a load that misses in cache W and results in a bus request. The VCL resets the dangling pointer and the T bit in cache Y . The VCL then determines the version to supply the load. Also, the most recent committed version (version back to the next level memory. The third snapshot is taken after the load has completed. 3.5.1. Squash invalidations The base design invalidates non-dirty lines in the L1 cache on task squashes. This includes both speculative data from previous tasks and architectural data from the next level memory (or the committed tasks). The base design invalidates these lines because it does not track the creator of a speculative versions for each line and hence cannot determine whether the version in a line has been committed or squashed. Squashing non-speculative data leads to higher miss rates for tasks that are squashed and restarted multiple times. To distinguish between copies of speculative and architectural versions, we add the architectural (A) bit to each cache line as shown in Figure 14. The A bit is set in a copy if either the next level memory or a committed version supplies data when a bus request issued to obtain the copy; else the A bit is reset. One of the VCL responses on a bus request specifies whether the A bit should be set or reset. Copies of architectural versions are not invalidated on task squashes, i.e., the ECS design only invalidates lines that have both the A and C bits reset. Further, a copy of a speculative version used by a task becomes an architectural copy when the task commits. However, the A bit is not set until the line is accessed by a later task, when the C bit is reset and the A bit is set. 3.6. Hit rate optimizations The base and ECS designs incur severe performance penalties due to reference spreading. When a uniprocessor program is executed on multiple processors with private L1 caches, successive accesses to the same line that hit after missing once in a shared L1 cache could result in a series of misses. This phenomenon is also observed for parallel programs where the miss rate for read-only shared data with private caches is higher than that with a shared cache. We use snarfing [6] to mitigate this problem. Our SVC implementation snarfs data on the bus if the corresponding cache set has a free line available. However, an active task's cache can only snarf the version that the task can use unlike an coherent cache. The VCL determines whether a task can copy a particular version or not and informs caches of an opportunity to snarf data on a bus request. 3.7. Realistic line size The base and ECS designs assume that the line size of the L1 caches is one word. The final SVC design however allows lines to be longer than a word. Similar to an SMP coherent cache, we observe effects due to false sharing. In addition to causing higher bus traffic, false sharing leads to more squashes when a store to a cache line from a task is executed out-of-order with a load from a different byte or word in the same line from a later task. We mitigate the effects of false sharing by using a technique similar to the sector cache [7]. Each line is divided into sub-blocks and the L and S bits are maintained for each sub-block. The size of a sub-block or versioning block is less than that of the address block (storage unit for which an address tag is maintained). Also, when a store miss results in a BusWrite request, mask bits that indicate the versioning blocks modified by the store are also made available on the bus. 4. Performance evaluation We report preliminary performance results for the SVC using the SPEC95 benchmarks. The goal of our implementation and evaluation is to prove the SVC design not just to analyze its performance. We underline the importance of a private cache solution by first showing how performance degrades rapidly as the hit latency for a shared cache solution is increased; the Address Resolution Buffer (ARB) is the shared cache solution we use for this evaluation. We mitigate the commit time bottlenecks in the ARB (by using an extra stage that contains architectural data) to isolate the effects of pure hit latency from other performance bottlenecks 4.1. Methodology and configuration All the results in this paper were collected on a simulator that faithfully models a Multiscalar processor. The simulator dynamically switches between a functional and a detailed cycle-by-cycle model to provide accurate and fast simulation of a program. The memory system model includes a fair amount of detail including an off chip cache, DRAM banks and interconnects between the different levels of memory hierarchy. The Multiscalar processor used in the experiments has 4 processors each of which can issue 2 instructions out-of-order. Each processor has 2 simple integer ALUs, 1 complex integer unit, 1 floating point unit, 1 branch unit and 1 address calculation unit, all of which are assumed to be completely pipelined. Inter-processor register communication latency is 1 cycle and each processor can send as many as two registers to its neighbor in every cycle. Loads and stores from each processor are executed in program order by using a load/store queue of 16 entries The ARB is a fully-associative set of 32-byte lines with a total of 8KB storage per stage and five stages; the shared data cache that backs up the ARB is 2-way set associative and 64KB in size. The off chip cache is 4MB in size with a total peak bandwidth of 16 bytes per processor clock to the L1 data, instruction and task caches. Main memory access time for the first word is 24 processor clocks and has a RAMBUS-like interface that operates at half the speed of the processors to provide a peak bandwidth of 8 bytes every bus clock. All the caches and memory are 4-way in- terleaved. Both the ARB and the L1 data cache have MSHRs/writebuffers each; each buffer can combine up to 8 accesses to the same line. Disambiguation is performed at the byte-level. The base ARB hit time is varied from 1 to 3 cycles in the experiments. Both the tags and data RAMs are single ported in all the caches. The private caches that comprise the SVC are connected together and with the off chip cache by an 8-word split-transaction snooping bus where a typical transaction requires 3 processor cycles 3 . Each processor has its own private L1 cache with 16KB of 4-way set-associative storage in lines. Both loads and stores are non-blocking with 8 MSHRs/writebuffers per cache. Each buffer can combine up to 4 accesses to the same line. Disambiguation is performed at the byte-level. L1 cache hit time is fixed at 1 cy- cle. The tag RAM is dual ported to support snooping while the data RAM is single ported. 4.2. Benchmarks We used the following programs from the SPEC95 benchmark suite with train inputs except in the cases listed: compress, gcc (ref/jump.i), vortex, perl, ijpeg (test/specmun.ppm), mgrid (test/mgrid.in), apsi, fpppp, and turb3d. All programs were stopped after executing 1 billion instructions. From past experience, we know that for these programs performance change is not significant beyond 1 billion instructions. 4.3. Experiments Figure presents the instructions per cycle (IPC) for a Multiscalar processor with either the ARB or the SVC. The configurations keep total data storage of the SVC and ARB/cache storage roughly the same. The percentage miss rates for the ARB and the SVC are shown on top of the IPC bar clusters (in that order). For the SVC, an access is counted as a miss if data is supplied by the next level mem- data transfers between the L1 caches are not counted as misses. From these preliminary experiments, we make three ob- servations: (i) the hit latency of data memory significantly affects ARB performance, (ii) the SVC trades-off hit rate for hit latency and the ARB trades-off hit latency for hit rate to achieve performance, and (iii) for the same total data storage, the SVC performs better than the ARB having a hit latency of 2 or more cycles as shown in Figure 16. The graphs in these figures show that performance improves in the range of 5% to 20% when decreasing the hit latency of the ARB from 3 cycles to 1 cycle. This improvement indicates that techniques that use private caches to improve hit latency are an important factor in increasing overall per- formance, even for latency tolerant processors like a Multiscalar processor. 3 Bus arbitration occurs only once for cache to cache data transfers. An extra cycle is used to flush a committed version to the next level memory. compress gcc vortex perl ijpeg mgrid apsi turb3d IPC 3.1/7.5 2.1/3.6 2.6/2.4 8.1/9.3 2.3/3.4 1.1/2.2 6.9/8.1 ARB (3 cycle) Figure The distribution of storage for the SVC produces higher miss rates than for the ARB. We attribute the increase in miss rates for the SVC to two factors. First, distributing the available storage results in reference spreading [6] and replication of data reduces available storage. Second, a latest version of a line that caches fine-grain shared data between Multiscalar tasks constantly moves from one L1 cache to another (migratory data). Such fine-grain communication may increase the number of total misses as well. 5. Conclusion Speculative versioning is important to overcome limits on Instruction Level Parallelism (ILP) due to ambiguous memory dependences in a sequential program. Our pro- posal, called the Speculative Versioning Cache(SVC), uses distributed caches to eliminate the latency and bandwidth problems of a previous solution, the Address Resolution Buffer, which uses a centralized buffer. The SVC conceptually unifies cache coherence and speculative versioning by using an organization similar to snooping bus-based coherent caches. A preliminary evaluation for the Multiscalar architecture shows that hit latency is an important factor affecting performance, and private cache solutions trade-off hit rate for hit latency. The SVC provides hardware support to break ambiguous memory dependences allowing proposed next generation multiprocessors to use aggressive parallelizing software for sequential programs. Acknowledgements We thank Scott Breach, Andreas Moshovos, Subbarao Palacharla and the anonymous referees for their comments and valuable suggestions on earlier drafts of the paper. This work was supported in part by NSF Grants CCR- 9303030 and MIP-9505853, ONR Grant N00014-93-1- 0465, and by U.S. Army Intelligence Center and Fort Huachuca under Contract DABT63-95-C-0127 and ARPA order no. D346 and a donation from Intel Corp. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the U. S. Army Intelligence Center and Fort Huachuca, or the U.S. Government. --R IEEE Standard for Scalable Coherent Interface (SCI) 1596- 1992 Data memory alternatives for multiscalar pro- cessors ARB: A hardware mechanism for dynamic reordering of memory references. Using cache memory to reduce processor-memory traffic Speculative Versioning Cache. Memory reference behavior and cache performance in a shared memory multi- processor Structural aspects of the system/360 model 85 part II: The cache. Dynamic speculation and synchronization of data de- pendences The case for a single-chip multiprocessor Trace processors: Moving to fourth-generation microarchitectures Multiscalar processors. The potential for thread-level data speculation in tightly-coupled multiprocessors --TR The Wisconsin multicube: a new large-scale cache-coherent multiprocessor The expandable split window paradigm for exploiting fine-grain parallelsim Boosting the performance of hybrid snooping cache protocols Multiscalar processors The case for a single-chip multiprocessor Improving superscalar instruction dispatch and issue by exploiting dynamic code sequences Dynamic speculation and synchronization of data dependences Complexity-effective superscalar processors Trace processors Data speculation support for a chip multiprocessor A scalable approach to thread-level speculation Architectural support for scalable speculative parallelization in shared-memory multiprocessors IEEE Standard for Scalable Coherent Interface, Science Using cache memory to reduce processor-memory traffic The Potential for Using Thread-Level Data Speculation to Facilitate Automatic Parallelization Speculative Versioning Cache Hardware for Speculative Parallelization of Partially-Parallel Loops in DSM Multiprocessors --CTR Arun Kejariwal , Xinmin Tian , Wei Li , Milind Girkar , Sergey Kozhukhov , Hideki Saito , Utpal Banerjee , Alexandru Nicolau , Alexander V. Veidenbaum , Constantine D. Polychronopoulos, On the performance potential of different types of speculative thread-level parallelism: The DL version of this paper includes corrections that were not made available in the printed proceedings, Proceedings of the 20th annual international conference on Supercomputing, June 28-July 01, 2006, Cairns, Queensland, Australia

speculative versioning;memory disambiguation;snooping cache coherence protocols;speculative memory

506342

Strong normalizability of the non-deterministic catch/throw calculi.

The catch/throw mechanism in Common Lisp provides a simple control mechanism for non-local exits. We study typed calculi by Nakano and Sato which formalize the catch/throw mechanism. These calculi correspond to classical logic through the Curry-Howard isomorphism, and one of their characteristic points is that they have non-deterministic reduction rules. These calculi can represent various computational meaning of classical proofs. This paper is mainly concerned with the strong normalizability of these calculi. Namely, we prove the strong normalizability of these calculi, which was an open problem. We first formulate a non-deterministic variant of Parigot's -calculus, and show it is strongly normalizing. We then translate the catch/throw calculi to this variant. Since the translation preserves typing and reduction, we obtain the strong normalization of the catch/throw calculi. We also briefly consider second-order extension of the catch/throw calculi. Copyright 2002 Elsevier Science B.V

Introduction The catch and throw mechanism provides a means to implement non-local exits. The following simple example written in Common Lisp [19] shows how to use the catch and throw mechanism: (defun multiply (x) (catch 'zero (multiply2 x))) (defun multiply2 (x) (if (null x) 1 (if (= (car x) (* (car x) (multiply2 (cdr x)))))) The rst function multiply sets up the catch-point with the tag zero, and immediately calls the second function. The second one multiply2 performs the actual computation by recursion. Given a list of integers, it calculates the multiplication of the members in the list. If 0 is found in the list, then the result must be 0 without computing any further, so it returns 0 by the throw-expression. The catch/throw mechanism is useful if one wants to escape from nested function calls at a time, especially in run-time errors. Nakano [11{14] proposed calculi with inference rules which give logical interpretations of the catch/throw constructs in Lisp. His calculi dier from the actual catch/throw- constructs in Common Lisp in the following two ways. (1) He changed the scope rule of the catch-construct from a dynamic one to a lexical one. In the above example, the expression (throw 'zero 0) is not lexically in the scope of the corresponding catch-expression, which indicates that the catch-expression has dynamic scope in Common Lisp. 1 In Nakano's calculi, tags are variables rather than constants, and the correspondence between throw and catch is represented as the ordinary variable binding mechanism, in which the scope of binders is lexical. (2) He introduced the tag-abstraction and tag-application mechanisms which do not exist in Common Lisp. 2 The motivation of this was to recover the expressivity which was lost by changing the scope rule of the catch-construct. Let us see how the above example can be written in Nakano's style: (defun multiply (x) (catch 'zero (multiply2 x 'zero))) (defun multiply2 (x u) (if (null x) 1 (if (= (car x) (throw u (* (car x) (multiply2 (cdr x) u))))) In this modied program, the catch-construct has lexical scope so that the scope of the tag zero is (multiply2 x 'zero) only. To throw an object from another function multiply2, the function is abstracted by the tag variable u. When using the function multiply2 we must provide the tag zero as the second parameter. Nakano also introduced a new type constructor (called \otherwise") for the tag abstraction mechanism; if a is a term of type A, and u is a tag-variable of type B, then the abstraction of a by u has type A B. The characteristic points in Nakano's formulation were (1) L c=t has restriction (side- condition) in the implication-introduction rule, and it excludes terms which corresponds to classical proofs. Actually L c=t corresponds to an intuitionistic calculus through the Curry-Howard isomorphism. (2) L c=t allows as many reductions as possible, hence it is non-deterministic (not con uent). These two features may look strange, since classical logic is said to be essentially non-con uent, while intuitionistic logic is con uent. 3 We consider that the classical version of L c=t , which is obtained by removing the restriction, is a more natural calculus, and is suitable for extracting algorithmic meaning from classical proofs. We call L K c=t as the classical version of L c=t . 1 Similarly the exception mechanism in the Standard ML has dynamic scope. 2 The exception mechanism in the Standard ML has abstraction/application. 3 We refer to Girard [6] and Parigot [15] for the discussion on the con uence and the classical logic. A few years later than Nakano, the second author (Sato) proposed another formulation for the catch/throw mechanism [17]. His motivation was to eliminate the type of the tag abstraction (\otherwise") in L c=t , since it is equivalent to disjunction. By unifying the throw-expression and the tag-abstraction mechanism, he obtained a simpler calculus NJ c=t . He also showed that L c=t can be interpreted in NJ c=t . NJ c=t has essentially the same restriction in the implication-introduction rule, hence it corresponds to intuitionistic logic. He also dened NK c=t by throwing away the restriction, and showed that it corresponds to classical logic. In summary, there are proposed four calculi for the catch/throw mechanism: Author Intuitionistic Logic Classical Logic Nakano L c=t L K c=t Sato NJ c=t NK c=t In this paper, we investigate the strong normalizability (SN) of the above four calculi, in particular, L K c=t and NK c=t . The SN of L c=t was proved by Nakano [14], but his proof was based on complex model-theoretic arguments. In our previous works, we proved the SN of NJ c=t in [8], and the SN of a large fragment of L K c=t in [9], but the SN of the full fragments of classical calculi L K c=t and NK c=t was an open problem. This paper solves this problem in an armative way. We rst formulate a non-deterministic variant of Parigot's -calculus by adding several reduction rules, and prove its strong normalizability using the reducibility method. We then translate the catch/throw calculi to this variant. Since this translation preserves typing as well as reduction, we obtain a proof of the strong normalizability of all the four calculi. We nally brie y discuss second-order extension of them. 2. The Catch/Throw Calculi 2.1. Nakano's Formulation Nakano proposed several calculi for the catch/throw mechanism. Among them, L c=t given in [14] is the strongest one. In this paper we also study L K c=t , an extension of L c=t . Although Nakano himself did not present L K c=t in published papers, the latter can be obtained from L c=t by simply throwing away the restriction in the implication introduction rule, therefore we regard L K c=t as one of Nakano's calculi. In the following, we shall dene L K c=t and mention the dierence of L K c=t and L c=t . We assume that there are nitely many atomic types (we use K as a metavariable for atomic including ? (falsity). Denition 2.1 (Type) In this denition, !, ^, _ are the types for the function space, product, and sum. By the Curry-Howard isomorphism, we may identify them with logical connectives implication, conjunction, and disjunction. The connective was introduced to give a type to tag abstraction. As usual, we abbreviate A !? as :A. We assume that, for each type A, there are innitely many individual variables x A of type A and innitely many tag variables u A of type A. We use x A ; y A ; z A for individual variables and u A ; v A ; w A for tag variables. We regard u A and u B as dierent tag variables if A 6 B. This implies that we may sometimes use the same variable-name for dierent entities (dierent types). Preterms of L c=t and L K c=t are dened as follows. Denition 2.2 (Preterm) Among the preterms above, the constructs catch, throw, , and tapp were introduced by Nakano to represent the catch and throw mechanism. We refer to the following table for the correspondence to similar constructs in Common Lisp and Standard ML. c=t Common Lisp Standard ML As noted in the introduction, tags in Common Lisp (exception names in Standard ML) are represented as tag-variables rather than constants. The preterm u:t is the tag-abstraction mechanism like the -abstraction x:t, and the preterm tapp(t; u) is the tag-application mechanism 4 like the functional application apply(t; u). We sometimes omit the types in variables. We also write apply(a; b) as ab. An individual variable is bound by the -construct and the case-construct, and a tag variable is bound by the catch-construct and the -construct. We identify two terms which are equivalent under renaming of bound individual/tag variables. FV(t) and FTV(t) denote the set of free individual variables and the set of free tag variables in t, respectively. The type inference rules are given in the natural deduction style, and listed in Table 1. The inference rules are used to derive a judgment of the form ' a : A ; where is a nite set in the form fx g, and is a nite set in the form g. In both sets we understand each variable appears only once. is a context of individual variables, and is a context of tag variables. In L c=t , the implication-introduction rule (marked (*)) has a restriction on free tag variables in b. L K c=t has no restriction. In the intuitionistic calculus L c=t , a preterm x A :b is well-typed only when x A does not essentially occur in the scope of any throw-construct in b. One of Nakano's main results was that, this restriction neatly corresponds to intuitionistic propositional calculus through the Curry-Howard isomorphism. The actual Actually, Nakano did not use the word tapp. Rather, he simply wrote tu for tapp(t; u). In this paper, we use dierent function symbols for dierent term-construction to clarify the syntax. Table Type Inference Rules of L c=t and L K c=t ' a :? ; restriction is complex due to the existence of the case-construct. In this paper we do not give the precise denition of \essential occurrence". We refer to [11] and [14] for details. Among the inference rules, the rst ten are standard. The rules for throw and catch re ect their intended semantics, namely, aborts the current context so that this term can be any type regardless of the type of b, and the type of catch(u a) is the same as a and also the same as the type of possibly thrown terms. The term u B :a is tag-abstraction, and it is assigned a new type A B. Conversely, if a is of type A B, then applying a tag variable u B to it generates a term of type A. An example of the type inference is as follows (which corresponds to the double negation Ag. The above one is a type inference gure in L K c=t , but not in L c=t . This is because, in the formation of x A :throw(u A ; x A ), the abstracted variable x A occurs free in throw(u does not t into Nakano's restriction. Let a; b; c; be metavariables for terms. If ' a : A ; is derived by the inference rules, we say a is a term of type A under contexts and . One-step reduction rules of L c=t and L K c=t are given by Table 2. Table One-Step Reduction Rules of L c=t and L K c=t a 6 x and x 2 FV (a)) In this denition, C[ ] represents a context with a hole [ dened as usual. Also substitution a[b=x] and a[v=u] are dened as usual. As an instance, we have the following reductions: tapp((v:(throw(v; Instead of having a one-step reduction like catch(u; a[throw(u; b)=x]) ; 1 b, the catch/throw mechanism splits into two steps as follows: Since we did not restrict any evaluation strategy, the reduction in L K c=t is non-deterministic, moreover it is not con uent. For instance, we have the following reduction sequences where we put t catch(u We dene a ; b (zero or more step reduction), and a ;+ b (one or more step reduction) as usual. Theorem 2.1 (Nakano) The subject reduction property holds for L c=t and L K c=t . 2.2. Sato's Formulation In [17], Sato proposed another formulation of the catch/throw mechanism. His primary motivation was to get rid of the logical connective from L K c=t , yet to obtain a system which is as powerful as L K c=t . From the logical point of view, is redundant, since it is equivalent to disjunction. Sato successfully eliminated from the calculus by unifying the two binders of tag variables, catch and . We shall give the denition of NK c=t in the following. NJ c=t is obtained from NK c=t by restricting the !-introduction rule in the same way as L c=t from L K c=t . Types are those of L c=t with deleted. Preterms are dened as follows. Denition 2.3 (Preterm) Individual variables are bound by the - and the case-constructs, and tag variables are bound by the ?-constructs. The ?-construct replaces catch and in L c=t , the !-construct replaces throw in L c=t , and the tapply-construct replaces tapp in L c=t . The type inference rules for the new constructs are given by Table 3. Table Type Inference Rules of NJ c=t and NK c=t The inference rule for the !-construct is the same as that of throw in L c=t . The term ?u may be constructed even if the type of a diers from B. The meaning of ?u B :a is that, if the computation of a ends normally and returns a 0 , then it returns inj 1 (a 0 ), and if a term b is thrown during the computation of a, then it returns inj 2 (b). Hence ?u B :a has type A _ B if a is of type A. The tapply-construct may be dicult to understand, but it is an inverse operation of tag abstraction. So tapply(?u B reduces to a[v B =u B ]. Type inference rules of other constructs are the same as before. The calculus with the restriction in the implication introduction rule is called NJ c=t , and the one without the restriction is NK c=t . The former corresponds to intuitionistic logic and the latter to classical logic. One-step reduction rules for the new constructs are given as follows: (a) (if u 62 FTV(a)) The last reduction may look strange, but it is useful in writing concise proofs [17], and necessary to simulate the reduction tapp(v:a; u) ; 1 a[u=v] in L c=t /L K c=t . Theorem 2.2 (Sato) The subject reduction property holds for NJ c=t and NK c=t . 2.3. Non-determinism and Classical Logic All the four calculi for the catch/throw mechanism have non-deterministic reduction rules, and are not con uent. We do not think that this is defect because: (1) as far as the strong normalizability is concerned, it is good to have as many reduction rules as possible. As a corollary of the strong normalizability of the strongest calculus, we obtain the strong normalizability of any subcalculus, and (2) classical logic is said to be inherently non-deterministic. In order to express all possible computations in classical proofs, our calculus should be non-deterministic. Later we can choose one answer by xing an evaluation strategy. Murthy gave examples which show classical proofs may contain multiple computational meanings [10]. The second author showed in [18] that Murthy's example can be expressed in in the NK c=t -style calculus. 3. A Non-deterministic Variant of Parigot's In this section, we give a non-deterministic variant of Parigot's as a target of translation from the catch/throw calculi. Parigot's -calculus [16] is a second-order propositional calculus for classical logic. It is a natural-deduction system whose sequents have multiple consequents. The -calculus is a quite nice formulation of classical logic, and at the same time, it is computationally interesting, since various control structures can be represented by the -construct whose typing is given as follows: The most important reduction rule for the -construct (called structural reduction) is: where af[]c := [](cb)g is the term obtained from a by substituting [](cb) for every subterm in the form []c where this is free in a. We refer to [16] for the denition of the -calculus. We can simulate a simplied version of the catch/throw mechanism in L K c=t by the - construct as follows: catch(u; a) as u:[u]a throw(u; a) as v:[u]a (where v does not appear in [u]a) However, the catch/throw calculi we consider are not con uent. Moreover, one term reduces to dierent variables x A and y A as we saw in the previous section. Since the calculus is a con uent calculus, direct simulation of the catch/throw calculi by is not possible. A possible solution is to add more reductions to , for instance, the call-by-value version of the structural reduction (the symmetric structural reduction). However, it is not known that a system which has both the structural reduction and the symmetric structural reduction is strongly normalizing or not 5 . Instead of naively adding reduction rules, we slightly modify the -calculus, then add non-deterministic reductions. Namely, we classify uses of into three cases: (1) u:[u]a (2) u:[v]a with u 62 FTV([v]a) (3) u:[v]a with u 6 v and u 2 FTV(a) We need (1) and (2) to simulate the catch-construct and the throw-construct, respec- tively. We only need to extend the reduction rules for (2), and the reduction rules for (1) remain the same. We do not need (3) to simulate the catch/throw calculi, so such a term construction will be excluded. Another modication to the -calculus is that we no longer have distinction of individual variables and tag variables. The named term [u]a will be represented by ordinary application ua. By this modication, we can directly -abstract over variables which correspond to names such as []. This is the key to simulate the tag-abstraction/tag- application mechanism in L K c=t . This representation is essentially due to de Groote [3], who formalized the exception mechanism for ML. Fujita [4] recently studied a similar calculus for the exception mechanism. As notational convenience, we write u:a for the term u:[u]a, and abort(va) for the term u:[v]a. We also extend reduction rules for the abort-construct to have non-deterministic features. We call the resulting system ND . 3.1. A Non-deterministic Calculus ND The types of ND are dened as follows: Denition 3.1 (Type) Recently, Fujita[5] indicated that such a system is shown to be strongly normalizing by translating it to Barbanera and Berardi's symmetric -calculus if we restrict the system to rst-order. However, we need the second-order version in this paper. Since ND is second-order, the type ? is redundant from the logical point of view. We, however, include ? as a primitive type, since we want to interpret ? dierently from 8X:X. The type variable X is bound by the type abstraction 8X, and we identify two types which are identical modulo renaming of bound type variables. We abbreviate A !? as :A. The preterms are as follows. Note that we adopt the Curry-style (implicit typing) for ND as in the -calculus. 6 Hence we do not attach types to variables, and consider reduction rules. Denition 3.2 (Preterm) Contrary to the original , we have only one sort of variables. A variable x may be used for an ordinary variable and also for a name (a tag-variable in our sense). Also we have no distinction between ordinary terms and named terms. Variables are bound by and constructs, and we again identify two terms which diers only in the bound variables. The preterm abort(t) is new to ND as we explained above. A judgment in ND is in the form ' a : A where is a nite set in the form g. The type inference rules which derive judgments are shown in Table Table Type Inference Rules of ND ' a :? In the 8-introduction rule (marked (*)), X may not occur freely in . If ' a : A is derived using the above rules, we say a is a (typable) term of type A (sometimes written as a : A). The reduction rules are derived from the -calculus, but we added several rules for abort which makes ND non-deterministic. Since we shall use the substitution in the 6 In [16], Parigot also denes a Church-style system. Table One-Step Reduction Rules of ND form [x:u(xb)=u] many times, it is abbreviated as [b= u]. When using this notation, we always assume that x is a fresh variable. We also abbreviate composition of substitutions We often write b for the sequence hb 1 ; ; b n i. Hence successive application ( (ab 1 ) b n ) is abbreviated as a b and successive substitution In the last case we assume that b 1 ; ; b n do not contain u free. We also use simultaneous substitution [b 1 =x um are mutually distinct and c do not contain u As before, we use the notation a ; b and a ;+ b. The following lemma can be proved easily. Lemma 3.1 Let be the substitution [b 1 =x um ]. If a ; b, then a ; b. 3.2. Strong Normalizability of ND In this subsection, we prove the strong normalizability (SN) of ND . The proof is a slight modication of Parigot's original proof of the SN of . Nevertheless, we give the proof here for completeness. Let T be the set of preterms in ND , and SN be the set of strongly normalizing preterms in ND . Note that we do not restrict T and ND be subsets of typable terms, following [16]. (a) is the maximum length of reduction sequences starting from a if a 2 SN , and is undened if a 62 SN . For F T , let F <! be the set of nite sequences of elements of F . Namely, In particular, F <! contains the empty sequence hi. Let F and G be subsets of T , and S be a subset of T <! . Then we introduce the following notations: As a special case, we have G. Denition 3.3 (Reducibility Candidate) A reducibility candidate is a subset of T , and is inductively dened as follows: 1. SN is a reducibility candidate. 2. If F and G are reducibility candidates, so is F ! G. 3. If fF i g i2I is a family of reducibility candidates for a non-empty set I, then is a reducibility candidate. (Note that the index set I may be innite.) The set of the reducibility candidates is denoted as RC. Lemma 3.2 For any F 2 RC, the following four clauses hold. 1. F SN . 2. All variables are contained in F . 3. If a 2 SN , then abort(a) 2 F . 4. There exists a set S such that S SN <! and The clause 3 was added from Parigot's original proof. It means that abort(a) with a strongly normalizing term a is contained in any reducibility candidate. The main dierence of our proof and Parigot's is that, in our case, a term in the form C[abort(a)] may reduce to abort(a), so we should always consider abort(a) as a reduct. However such a term is contained in any reducibility candidate if a is strongly normalizing by this lemma, and therefore we can always handle this term easily. Proof. We prove all the four clauses simultaneously by induction on \F 2 RC". (Case: F is SN ) Clause 4 is proved by taking fhig as S, and other clauses are trivial. (Case: F is G induction hypothesis (abbreviated as IH), we have x proves Clause 1. By IH, G SN , and there exists a set S 0 SN <! such that by taking G; we have G which proves Clause 4. Let x be a variable, a 2 SN , b since all its reducts are in the form xb 0 n or abort(d)bk+1 b 0 proving Clause 2. We also have abort(a)b proving Clause 3. (Case: F is T are easily proved from IH. Also by IH, for each I there is an S i SN <! such that G then we have which proves Clause 4. ut From Clause 4 in the above lemma, we put F ? as the largest such S. Namely, for any A preterm a is neutral if it is either a variable or in the form bc. Lemma 3.3 For any F 2 RC, the following two clauses hold. 1. For any a 2 F , if a ; 1 a 0 then a 0 2 F . 2. If a is neutral, and a 0 2 F for any a 0 such that a ; 1 a 0 , then a 2 F . Proof. This lemma is proved by induction on \F 2 RC". The key case is F G ! H. We shall prove Clause 2 only. Suppose a is neutral, and a 0 2 G ! H for any a 0 such that a ; 1 a 0 . Take an arbitrary preterm b 2 G. We shall prove ab 2 H by induction on (a) (b) (since a and b are SN). The preterm ab reduces in one step to either one of a 0 b (with a ; 1 a 0 ), ab (with a a 00 [abort(c)=x]), or abort(d) (with b b 00 [abort(d)=x]). We can easily prove that all the four terms belong to H, and by IH, we have ab 2 H. Consequently a 2 G ! H. ut Denition 3.4 (Interpretation of Types) An interpretation is a map from type variables to reducibility candidates. Note that there exists an interpretation which maps all the type variables to SN . An interpretation is naturally extended to any types in the following way: where an interpretation [F=X] is dened as for Y 6= X. Lemma 3.4 Let A; B be types, and be an interpretation. Then This lemma can be proved by induction on the structure of A. Lemma 3.5 Let F 2 RC, x; u be variables, a; b be terms, and c be a sequence of terms. 1. If a[b=x] 2 F and b is SN, then (x:a)b 2 F . 2. If (u: (a[d= u])d)c 2 SN and c 3. If c 4. If a 2 SN , then u: a 2 SN . Proof. 1. We can prove this clause by induction on (a[b=x]) + (b) using Lemma 3.1 and Lemma 3.3. We must take care that the reducts of (x:a)b may be of the form abort(c), but this case can be treated using Clause 3 of Lemma 3.2. 2. We can prove this clause by induction on ((u: (a[d= u])d)c) 3. From Clause 1 above, all we have to prove is u(bc) 2 SN for any b 2 F . Since 4. This can be proved by analyzing reduction rules. ut Theorem 3.1 Assume is derived in ND . Assume also that is an interpretation, b then we have At rst look, the statement of the theorem looks ambiguous. For instance, given a proof of x C ; y :D ' a : A, we may split the lefthand side of ' in two ways, each of which results in the following conclusion: holds for any b 1 2 (C) and b 2 2 (:D). holds for any b 2 (C) and c 2 (D) ? . Actually the theorem implies that both hold, so no ambiguity arises. We now state the proof of the theorem. Proof. The theorem is proved by induction on the type inference of a. Let be a substitution [b 1 =x as . (Case: Assumption-Rule) In this case a x. We have to prove x 2 (A). There are two subcases. (i) x x i for some i. Then A C i and x b i . By the assumption b i 2 (C i ), so x 2 (A). (ii) x u i for some i. Then A :D i and x z:u i (zc). From Lemma 3.5, we have x 2 (D (Case: !-introduction) In this case a x: c. We have A B ! C and c is a term of type C. By a suitable renaming, we have a x: c. Take any d 2 (B). By IH, Hence by Lemma 3.5, we have (x: c)d 2 (C), hence C). (Case: !-elimination) In this case, we have a bc, Hence a (b)(c) 2 (A). (Case: 8-introduction) In this case, A 8X: B and a : B is derivable. Let F 2 RC and 0 [F=X]. Since the type variable X does not occur freely in , b m. Hence, by IH, we have a 2 0 (B). Finally a 2 (8X: B). (Case: 8-elimination) In this case, a : 8X: B and A B[C=X]. We have a 2 (8X: B) from IH, hence a 2 [(C)=X](B). By Lemma 3.4, a 2 (B[C=X]). (Case: -introduction) In this case, a u:b, and b : A. By a suitable renaming, we have a u: b. We have By IH, we have Hence (b 0 )c 2 SN , therefore by Clause 4 of Lemma 3.5, we have u: (b 0 )c 2 SN . By Clause 2 of Lemma 3.5, (u: b)c 2 SN . Consequently, (Case: ?-elimination) In this case, a abort(b). By IH, b 2 3.2, we have a abort(b) 2 (A). ut By choosing x m) in the theorem above, we obtain a 2 (A) for any term a of type A and an interpretation . Since there exists an interpretation , and (A) SN , we have the following theorem: Corollary 3.1 ND is strongly normalizing. 4. Translation from the Catch/Throw Calculi to ND This section gives translations from the catch/throw calculi to ND . In the following we give only translations from the classical catch/throw calculi L K c=t and NK c=t , but the translations work also for L c=t and NJ c=t , since they are subcalculi. 4.1. Translation of Nakano's Calculus We shall translate L K c=t to ND . The translation is the same as the standard encoding of propositional logic in second-order logic except the catch/throw-constructs. First, we translate types. (other than ?) in L K c=t , and variables in ND . The point here is that the type AB is translated to :B ! A. This translation re ects our intention that the -abstraction is translated to the -abstraction. We then translate preterms in L K c=t to ND . We assume that, for each individual variable x A in L K c=t , x is a variable in ND , and for each tag variable u A in L K c=t , u is a variable in ND . We also assume that this mapping on variables are injective. Preterms are translated as follows: x A x abort(a) abort(a) x A :a x:a ab ab (a) a(xy:x) (a) xy:xa tapp(a; The translation is extended to contexts for variables in the following way. Let be a context for individual variables fx A 1: A and be a context for tag variables in L K c=t . Then we dene: Note that the types of tag variables are negated through the translation. The translation preserves typing and reduction as we shall see. Lemma 4.1 (Preservation of Type Assignment) If ' a : A ; is derived in L K c=t , then ; ' a : A is derived in ND . Proof. Since the translation for propositional connectives are standard, we verify the other cases only. From IH, we have ; ' a : A. Since fu : Ag is fu : :Ag, and a) is u:a, we can derive ; fu A : Ag ' catch(u A ; a) : A by the -introduction rule. (throw) From IH, we have ; ' a : A. Since [ fu A : Ag is a) is abort(ua), we can derive by the !-elimination rule and the ?-elimination rule. From IH, we have ; ' a : A. Since and A B is :B ! A, we can derive ; by the !-introduction rule. (tapp) From IH, we have by the !-elimination rule, we can derive Lemma 4.2 The translation is compatible with substitution. Namely, a[b=x A ] a[b=x], and a[v B a[v=u]. This lemma is proved by straightforward induction on the construction of a, and is omitted. Lemma 4.3 (Preservation of Reduction) If a and b are typable terms and a ; 1 b in L K c=t , then a ;+ b in ND . Proof. This lemma is proved by induction on the structure of the term a. We prove the key cases only. 1. a 6 x and x 2 FV(a) By Lemma 4.2, we have a[throw(u A ; b)=x] a[abort(ub)=x]. By induction on the term a, we have a[abort(c)=x] ;+ abort(c) for any c, so we have a[abort(ub)=x] ;+ abort(ub). 2. catch(u A ; a) ; 1 a with u A 62 FTV(a) Since u 62 FV (a), we have catch(u A ; a) u:a ; 1 a. 3. catch(u A ; throw(u A ; We have catch(u A ; throw(u A ; a)) u:abort(ua), and u 62 FV(a). 4. tapp(u:a; We have tapp(u:a; v) (u:a)v, and it reduces to a[v=u]. By Lemma 4.2, a[v=u] a[v=u], hence we are done. ut From the above lemmas, we have the following theorem. Theorem 4.1 The system L K c=t is strongly normalizing. Hence L c=t is strongly normalizing. Remark. The translation from L K c=t to ND does not really need the second order quantier. Namely, if we eliminate 8, and add ^ and _ to ND , then we can translate c=t to this modied calculus. Since we can prove the SN of this modied calculus by an elementary method as in [16], we can also prove the SN of L K c=t by an elementary method. However, we need the second order quantier 8 in translating NK c=t as we shall see in the next section. We therefore proved the SN of L K c=t based on the reducibility method. 4.2. Translation of Sato's Calculus In this subsection we translate NK c=t to ND . Before dening the translation, we try to give a naive translation from NK c=t to L K c=t , and explains why it fails. A natural candidate of the translation is: A A for any type A tapply(a; where the type C will be supplied from the type inference of each term. At this moment let us ignore how to obtain this C. By the above translation we can interpret all but one reduction rules in NK c=t . The only exception is the following one: The lefthand side is interpreted as tapply(?u:a; v) case(catch(u; inj 1 which does not necessarily reduce to a[v=u]. Hence the naive translation from NK c=t to c=t fails. Moreover, it seems very dicult to nd a suitable extension of L K c=t which is strongly normalizing and which can reduce the above term to a[v=u]. However, the situation changes if we consider the second-order calculus where the disjunctive type is no more primitive, but is dened as A _ B As we shall see later, the term tapply(?u:a; v) reduces to a[v=u] through this encoding. Now let us dene the translation from NK c=t to ND . The translation of types are the same as the translation from L K c=t to ND . The translation of preterms except are the same. The translation of the new constructs is dened as follows: tapply(a; where we assume x; y are not used in the term a. This translation may look complex, but it is the result of the second-order encoding of the above naive translation from NK c=t to c=t . The translation is extended to a context for individual variables in the same way as before. For a context for tag variables , we need to change the translation, since a tag variable of type B should be translated to a variable of type C _ B where C is the type of the body of enclosing ?-expression. In other words, we cannot determine the type C until we reach the enclosing ?-expression. To solve this problem, we introduce a mapping from a set of tag variables (in NK c=t ) to a set of types in ND , and we make the translation of contexts for tag variables dependent on . Let be a context for tag variables fu B1: g, and be a mapping from fu B1 n g to types in ND . We dene In this denition _ 0 is an abbreviation dened as C _ 0 D 8X: the same as the result of the translation of _). Lemma 4.4 (Preservation of Type Assignment) If ' a : A ; is derivable in NK c=t , then, for any mapping whose domain contains all the tag variables in , we have ; ' a : A is derivable in ND . Proof. This is proved by induction on the derivation of ' a : A ; . We only verify the key cases. We have to derive ; any . Fix a mapping . Suppose (otherwise the proof is shorter), and 0 Bg. Let 0 be a mapping such that 0 (v) (v) for v 6 u and 0 (u) A. From IH, we can derive ; 0 ' a : A. We have 0 is B)g. Also we have From these facts, we can derive the desired type inference by the -introduction rule. We have to derive for any . Fix any . From IH, we can derive ; ' a : B. We have is abort(u(xy:ya)). We can derive the desired type inference by the ?-elimination rule. We have to derive for any . From IH, we can derive ; ' a : A _ B. We have B)g, and tapply(a; u B ) is a(x:x)(y:abort(u(zw:wy))). By calculating the type of this term, we can derive the desired type inference. ut The next lemma is used in proving the preservation of reduction through the translation. Lemma 4.5 Let a be a typable term in NK c=t . Let be a substitution [x:v(xt 1 t 2 )=v] where v is the result of the translation of a tag variable v B in NK c=t , t 1 is x:x, and t 2 is y:abort(u(zw:wy)). Then we have a ; a[u=v] in ND . Proof. We prove this lemma by induction on the structure of the term a. We state here the key cases only. (Case a !v We have the following reduction sequence: (Case a tapply(b; v B We have the following reduction sequence: Lemma 4.6 (Preservation of Reduction) If a and b are typable terms and a ; 1 b in ND , then a ;+ b in ND . Proof. We only check the key cases (the tapply-expression). In the following we put t 1 x:x, and t 2 y:abort(u(zw:wy)). 1. tapply(inj 1 We have tapply(inj 1 2. tapply(inj 2 We have tapply(inj 2 (a); abort(u(zw:wa)). The last term is !u B :a. 3. tapply(?v:a; u) We have: tapply(?v:a; u) (v:zw:za)t 1 t 2 a[u=v] Hence we have the desired property. ut From the lemmas we have the following theorem. Theorem 4.2 The system NK c=t (and hence NJ c=t ) is strongly normalizing. Remark. In our proof, the use of the second order quantier 8 is indispensable to give a translation of NK c=t . Since NK c=t is a rst-order system, one may think that our proof used too strong a method, and that the SN of NK c=t could be proved by a more elementary method. At present, we do not have an answer to this question. Our trial to apply an elementary method to NK c=t was not successful. 5. Extension of the Catch/Throw Calculi Having given the translation to ND , it is easy to introduce the second-order quantier to the four catch/throw calculi without loss of the nice properties such as the strong normalization. Since the catch/throw calculi are formulated in the Church-style (variables are explicitly labeled with their types), we should introduce term-constructs for type abstraction/application. As usual we let X:a denote the former and aX for the latter. The typing rules are given as follows: In the 8-introduction rule (marked (*)), X may not occur freely in nor . Also the reduction rule (X:a)B ; 1 a[B=X] is added. By adding these rules to L K c=t , we obtain a second-order catch/throw calculus L K2 c=t . Similarly we can obtain NK 2 c=t . The calculi L K2 c=t and NK 2 c=t enjoy nice properties such as the subject reduction and the strong normalization. Here we brie y mention the expressivity of these calculi. structures such as the integers and the binary trees can be encoded by the second-order quantier [6], we can dene functions with the catch/throw mechanism over various data types in the extended calculi. For instance the function multiply mentioned before is typed as follows: It Int U:uft:tUuf It IntList W:wft:tWwf Here Times(a; b) is the multiplication of two integers a and b and dened as usual. It Int and It IntList are iterators on the type Int and IntList. The term It Int (a; z:b; x) with x 62 FV (b) is the if-then-else expression, namely it reduces to a if x is 0, and b otherwise. It is easily seen that the above function multiply does the same computation as one given in the introduction. Since the above representation of free structures is not so good in computation [6], we may consider another direction of extension. Namely, we may add inductive data types. In [8], the rst author already proposed to add inductive data types to NJ c=t without loss of the SN of the calculus, and showed that higher-order functions which use the catch/throw mechanism can be represented in the extended calculus. However, we have not fully studied this direction for the classical catch/throw calculi, so it is left for future work. 6. Concluding Remarks We have investigated the four catch/throw calculi by Nakano and Sato, in particular, the calculi which correspond to classical logic through the Curry-Howard isomorphism. We dene a non-deterministic variant of Parigot's -calculus, and proved the strong normalizability of this variant. We gave faithful translations from the catch/throw calculi to this variant, and as a corollary, we obtained the strong normalizability of the four calculi. We also discussed their extension brie y. Recently, Fujita [4] studied exc , a call-by-name variant of de Groote's formulation of the exception mechanism in Standard ML. His calculus is a subcalculus of the rst-order version of -calculus. Since the catch/throw mechanism and the exception mechanism are essentially the same, his motivation and ours are similar. The main dierences of his calculus and our ND is that (1) his calculus is con uent, while ours are non-deterministic, so we have more computations, (2) he uses the rst-order version (actually, the implicational while we use the second-order version, and (3) his calculus has two sorts of variables (reminiscent of individual variables and tag variables), while we use one sort of variables, thus we can directly abstract over tags. Crolard [2] also studied a con uent calculus for the catch/throw mechanism. Since his calculus can be translated to Parigot's -calculus, it is similar to Fujita's formulation, thus diers from the calculi studied in this paper. Extracting algorithmic contents from classical proofs is now a quite active research area. Many researchers in this area aim at obtaining con uent calculi for classical logic. However, classical logic is said to be inherently non-deterministic, namely, classical proofs may contain multiple computational meanings. Therefore, if we want to represent as many computational meanings as possible, it is natural to begin with non-deterministic calculi. Our approach is to design and study non-deterministic calculi rst, then to study con uent subcalculi. We believe that the catch/throw calculi presented in this paper can be good basis for this approach. Barbanera and Berardi's calculus [1] is another non-deterministic calculus for classical proofs, so their calculus could be also a good basis. Further studies on extracting computational meaning from classical proofs are left for future work. Acknowledgement We would like to express our heartful thanks to Hiroshi Nakano, Makoto Tatsuta, and Izumi Takeuti for helpful comments on earlier works. We also thank to Ken-etsu Fujita for pointing out references and errors, and to anonymous referees for valuable comments for improvement. The author is supported in part by Grant-in-Aid for Scientic Research from the Ministry of Education, Science and Culture of Japan, No. 09780266 and No. 10143105. --R Proofs and Types (Cambridge University Press "A Classical Catch/Throw Calculus with Tag Abstractions and its Strong Normalizability" --TR Common LISP: the language Proofs and types A formulae-as-type notion of control Intuitionistic and classical natural deduction systems with the catch and the throw rules Lambda-My-Calculus Extracting Constructive Content from Classical Logic via Control-like Reductions A Simple Calculus of Exception Handling Classical Brouwer-Heyting-Kolmogorov Interpretation --CTR Emmanuel Beffara , Vincent Danos, Disjunctive normal forms and local exceptions, ACM SIGPLAN Notices, v.38 n.9, p.203-211, September

catch and throw;classical logic;type system;strong normalizability

506345

Least and greatest fixed points in intuitionistic natural deduction.

This paper is a comparative study of a number of (intensional-semantically distinct) least and greatest fixed point operators that natural-deduction proof systems for intuitionistic logics can be extended with in a proof-theoretically defendable way. Eight pairs of such operators are analysed. The exposition is centred around a cube-shaped classification where each node stands for an axiomatization of one pair of operators as logical constants by intended proof and reduction rules and each arc for a proof- and reduction-preserving encoding of one pair in terms of another. The three dimensions of the cube reflect three orthogonal binary options: conventional-style vs. Mendler-style, basic ("[co]iterative") vs. enhanced ("primitive-[co]recursive"), simple vs. course-of-value [co]induction. Some of the axiomatizations and encodings are well known; others, however, are novel; the classification into a cube is also new. The differences between the least fixed point operators considered are illustrated on the example of the corresponding natural number types.

Introduction This paper is a comparative study of a number of least and greatest xed point operators, or inductive and coinductive denition operators, that natural- deduction (n.d.) proof systems for intuitionistic logics (typed lambda calculi with product and sum types) can be extended with as logical constants (type-language constants), either by an axiomatization by intended proof and reduction rules (\implicit denition") or by a proof- and reduction-preserving encoding in terms of some logical constants already present (\explicit deni- tion"). One of the reasons why such logical or type-language constants are interesting lies in their useful programming interpretation: inductive types behave as data types, their introductions as data constructors and eliminations as recursors; coinductive types may be viewed as codata types, their introductions as corecursors and eliminations as codata destructors. In the literature, a fairly large number of axiomatizations and encodings of both particular [co]inductively dened types and general [co]inductive denition operators can be found, see e.g., [1,14,19,20,24,25,15,7]. The paper grew out of a wish to better understand their individual properties and their relations to each other. The contribution of the paper consists in a coordinated analysis of eight intensional-semantically distinct pairs of [co]inductive denition operators, arranged into a cube-shaped taxonomy, which resulted from an attempt to t the various known axiomatizations and encodings into a single picture and to nd llers for the holes. Each node of the cube stands for an axiomatization by proof and reduction rules of one pair of logical constants and each arc for a proof- and reduction-preserving encoding of one pair in terms of another. Some axiomatizations and encodings rely on the presence in the system of certain other logical constants (the standard propositional connectives, 2nd-order quantiers, or a \retractive" recursive denition operator ). The three dimensions of the cube re ect three orthogonal binary choices: conventional-style vs. Mendler-style, basic (\[co]iterative") vs. enhanced (\primitive-[co]recursive"), simple vs. course-of-value [co]induction. The cube looks as follows: q and (with optional superscripts) are conventional-style inductive and coinductive denition operators; m and n (with optional superscripts) are Mendler- style operators. The superscript ' q ' marks the \enhanced" feature, the superscript indicates the \course-of-value" feature. The distinctions between basic and enhanced, simple and course-of-value [co]in- duction are distinctions between essentially dierent forms of [co]induction, with dierent associating schemes of (total) [co]recursion. Basic [co]induction gives [co]iteration, enhanced [co]induction gives (full) primitive [co]recursion. All axiomatizations and encodings we have found in the literature deal with simple forms of [co]induction. The axiomatizations and encodings for course- of-value [co]induction in this paper are ours, we think. The dierence between conventional- and Mendler-style [co]induction (named after Mendler [19,20]) is more technical and harder to spell out informally, but not shallow. A conventional-style [co]inductive denition operator applies to a proposition-function only if it is positive; the associating reduction rule refers then to a proof of its monotonicity (all positive proposition-functions are monotonic wrt the preorder of inclusion). Mendler-style operators apply to any proposition-functions. The axiomatizations of enhanced and course-of- value conventional-style operators rely on the presence in the system of other logical constants, those of Mendler-style operators do not. Thus, in more than one sense, Mendler-style operators are more uniform than conventional-style operators; resorting to programming jargon, one might for instance want to say that the Mendler-style operators are generic, whereas the conventional- style ones are only polytypic. These uniformity features have a price the proof rules of the Mendler-style operators involve implicit (\external") 2nd-order quantication at the level of premisses. Throughout the paper, the semantics that we keep in mind is intensional, so we only consider -reduction, not -conversion. Some remarks are in order regarding the technical machinery that we use. By natural deduction, we mean a proof system style where instead of axioms involving implications and universal quantications we systematically prefer to have proof rules involving hypothetical and schematic judgements (\exter- nalized" implications and universal quantications), in sharp contrast to the Hilbert style of proof systems. For us therefore, natural deduction is really the \extended" natural deduction of Schroeder-Heister [30,31]: we allow proof rules to be of order higher than two: not only may conclusions have premisses and these have premisses in their turn, but even the latter may be hypo- thetical. This choice makes axiomatizations of dierent logical constants very compact, but on the expense of certain added complexity in their encodings in terms of other logical constants. In order to compactify the notation and to get around the technicalities related to -conversion and substitution, we use a simple meta-syntax, a higher-oder abstract syntax derived from logical frameworks such as de Bruijn's AUT-PI and AUT-QE [5], Martin-Lof's system of arities [22, chapter 3], and Harper, Honsell, and Plotkin's LF [12]. denotes the schematization of s wrt. x denotes the instantiation of s with Schematization and instantiation are stipulated to satisfy the following rules: are not free in s, then We have made an eort to make the paper self-contained; for the omitted details, we refer to Uustalu [35]. A preliminary report of the present work appeared as [37]. We also refer to Matthes [17], an in-depth study of extensions of system F with constructors of basic and enhanced conventional- and Mendler-style inductive types, which in regard to the clarication of the relationship between the conventional- and Mendler-style induction builds partly upon our work. The paper is organized as follows. In section 2, we lay down our starting point: it is given by systems that we denote NI and NI 2 , the n.d. proof systems for 1st- and 2nd-order intuitionistic propositional logics, optionally extended with a \retractive" recursive denition operator . Then, in section 3, we rst present the basic [co]induction operators, both in conventional and Mendler- style and then continue with their encodings in terms of the 2nd-order quan- tiers and each other. In sections 4 and 5, we describe enhanced [co]induction and course-of-value [co]induction operators respectively and their encodings via the operators of the basic kind. In section 6, we give a survey of related work on inductive and coinductive types. Finally, in section 7, we conclude and mention some directions for future work. Preliminaries In principle, the [co]inductive denition operators described below can be added to the n.d. proof system of any intuitionistic propositional logic. (They also admit a straightforward generalization for predicate logics.) The most natural base system for such extensions however is NI, the standard n.d. proof system for (full) 1st-order intuitionistic propositional logic. The logical constants of NI are ^ (conjuction), _ (disjunction), > (verum), ? (falsum), and ! (implication). These propositional connectives are axiomatized by the proof and reduction rules listed in Figure 1. (To save space, the reduction rules are given not for proofs, but for (untyped) term codes of proofs; the reduction rules for proofs are easy to recover. The reduction relation on terms satises subject reduction.) ^I _I L e (c) e c(e) Figure 1: Proof and reduction rules for standard propositional connectives. Another important base system is NI 2 , the n.d. proof system for 2nd-order intuitionistic propositional logic. This system extends NI with 8 2 and 9 2 , the standard 2nd-order quantiers. The proof rules for 8 2 and 9 2 are presented in Figure 2. In the encodings of enhanced [co]induction in terms of basic [co]induction, we shall need a logical constant , a \retractive" recursive denition oper- ator. This is a proposition-valued operator on proposition-functions that are positive. The proof and reduction rules for appear in Figure 3. The introduction and elimination rules for behave as an embedding-retraction pair. The c c e(c) e(c) Figure 2: Proof and reduction rules for 8 2 , 9 2 . extensions of NI and NI 2 with will be denoted by NI() and NI 2 (). Of importance for us is the fact that NI 2 () is strongly normalizing (i.e., every proof of NI 2 () is strongly normalizing); consult Mendler [19,20] and Urzyczyn [34]. Figure 3: Proof and reduction rules for . The syntactic concepts of positivity and negativity of proposition-functions are system-dependent. For any particular system, these concepts are dened by mutual structural induction on proposition-functions denable in this sys- tem. In NI and its extensions considered in this paper, a proposition-function (X)F is dened to be positive [negative] if every occurrence of X in F appears within an even [odd] number of antecedents of implications. Also for any particular system and by a similar induction, explicit denitions can be given for the derivable proof rules M and M + establishing that positive [neg- ative] proposition-functions are monotonic [antimonotonic] wrt. the preorder of proposition inclusion. These proof rules appear in Figure 4. F F positive F negative Figure 4: Derivable proof rules M and M + . As an example, we shall consider the proposition-function N dened by setting N is obviously positive. The corresponding monotonicity witness map N is dened as follows: 3 Basic [co]induction The logical constants from the two lower front nodes of the cube provide the most fundamental forms of [co]inductive denition of propositions, viz. the basic (in other words, \[co]iterative") forms of conventional- and Mendler-style [co]inductive denition. and are operators of conventional-style induction and coinduction and apply to positive proposition-functions are their Mendler-style counterparts applicable without restrictions to any proposition-functions. Their proof and reduction rules are given in Figures 5 and 6. The proof rules for m and n are more complex than those for and , but their reduction rules, in compensation, are simpler and more uniform: their right-hand sides do not refer to the M + proof rule. e( cata F cata F (wrap F (c); e) e(map F ()cata F (; e))) e( open open F (ana F (c; e)) map F Figure 5: Proof and reduction rules for and . e( e( Figure Proof and reduction rules for m and n. From the algebraic semantics point of view, F is a least prexed point of F wrt. the inclusion preorder of propositions: it is both itself a prexed point of F (by the I-rule) and a lower bound of the set of all prexed points of F (by the E-rule). (Recall that R is said to be a prexed point of F , if F (R) is less than R.) F , dually, is a greatest postxed point of F . 3 Since a least prexed [postxed] point of a monotonic function is also its least xed point, F and F are also least and greatest xed points of F . In a similar fashion, mF can be thought of as a least robustly prexed point of F : it is both itself a robustly prexed point of F and a lower bound of all robustly prexed points of F . Here, R is considered to be a robustly prexed point of F , if not only is F (R) less than R, but F (Y ) is less than R for all Y 's less than R. But mF is also a least (ordinary) prexed point of a function sending any R to a supremum of the set of all F (Y )'s such that Y is less than R. F e (which is always positive) appears to be a least monotonic majorant of F wrt. the pointwise \lifting" of the inclusion preorder of propositions to a preorder of proposition-functions. If F is monotonic, then F and F e are equivalent (pointwise). The dualization is obvious: nF is a greatest robustly postxed point of F and a greatest (ordinary) postxed point of a function F a [F a (R) 8 2 ((Y )(R!Y )!F (Y ))] sending any R to an inmum of the set of all F (Y )'s such that Y is greater than R. Under the programming interpretation, F is a data type, with wrap F a data constructor and cata F an iterator, and F is a codata type, with ana F a coiterator and open F a codata destructor, in the most standard sense. mF , with mapwrap and iter, and nF , with coit and mapopen, are Mendler-style versions of these things. This is best explained on an example. The type of standard natural numbers Nat, with zero and succ the constant zero and the successor function and natcata the iterator, is normally axiomatized as follows: These typing and reduction rules are essentially nothing else than those for 3 Note here that, in a preorder (also in a Heyting algebra), it may turn out that all monotonic functions have least [greatest] prexed [postxed] points; hence allowing and to apply to any positive F should not lead to inconsistencies (the encodability of , in terms of 8 2 , 9 2 demonstrates that this is the case indeed). conventional basic induction with N as the underlying proposition-function. Indeed, making the following denitions ensures the required typing and reduction properties: Nat (N) zero wrap N (inl(hi)) succ(c) wrap N (inr(c)) This suggests a similar specialization of Mendler-style basic induction for N by the following denitions: Nat m(N) mapzero(d) mapwrap(inl(hi); d) d) mapwrap(inr(c); d) The type Nat of Mendler-style natural numbers, with mapzero, mapsucc and natiter the Mendler-style constant zero, successor function, and iterator, obeys the following typing and reduction rules. Here, it may be helpful to think of Q as some chosen type of representations for naturals and d as a method for converting representations of this type to naturals. A natural, hence, is constructed from nothing or a representation (for its predecessor), together with a method for converting representations to naturals. Using Nat as Q, the standard constructors of naturals are denable as follows: zero mapzero(()) succ(c) mapsucc(c; ()) natcata and natiter are iterators. Iteration is a very simple form of total re- cursion: the result of an iteration on a given natural is only dependent of the result on the predecessor. If the \straightforward" denition of a function follows some more complex form of recursion, then denitions by iteration can get clumsy. The factorial of a given natural, for instance, depends not only on the factorial of its predecessor, but also on the predecessor itself. An iterative denition of the factorial has to dene both the factorial and the identity function \in parallel" and then project the factorial component out. zero zero Exactly the same trick of \tupling" is also needed to program the Fibonacci function: the Fibonacci of a given natural number depends not only on the Fibonacci of its predecessor, but also on the Fibonacci of its pre-predecessor. An iterative denition of Fibonacci has to dene both Fibonacci and the \one- step-behind Fibonacci" \in parallel". bo(c) fst(natcata(c; zero; case@ snd( bo(c) fst(natiter(c; (-) zero; case@ snd(-( These examples show how other forms of recursion can be captured by iteration using \tupling". Such modelling is not without drawbacks, however. First, it is more transparent to dene a function using its \native" form of re- cursion. Second, the intensional behavior of iterative denitions is not always satisfactory. It is well known, for instance, that the predecessor function can be programmed using iteration, but the programs take linear time to compute (and only work as desirable on numerals, i.e., closed natural number terms). pred(c) cata N (c; ( pred(c) iter(c; ( The more complex forms of induction considered in the following sections remedy these problems by oering more advanced forms of recursion. Basic [co]induction vs. 2nd-order quantiers Both , and m, n can be encoded in terms of 8 2 , 9 2 in a proof- and reduction- preserving manner. Proposition 1 The following is a proof- and reduction-preserving encoding of , in terms of 8 F (c) (() map cata F ana F open F (c) map F (fst(c) snd(c); ()hfst(c); i) This encoding is a proof theory recapitulation of the Knaster{Tarski xed point theorem [33] stating that an inmum [supremum] of the set of all prexed [postxed] points of a monotonic function is its least [greatest] prexed [postxed] point. In its general form, the encoding seems to be a piece of folk- lore. For the special case of \polynomial" proposition-functions (such as N), essentially the same encoding was rst given by Bohm and Berarducci [1] and Leivant [14]. For naturals, our encoding specializes to the following: Nat zero natcata (In Bohm and Berarducci's encoding, Nat zero c e z e s .) Proposition 2 The following is a proof- and reduction-preserving encoding of m, n in terms of 8 iter coit mapopen This encoding builds on the following robust analog of the Knaster-Tarski xed point theorem: an inmum [supremum] of the set of all robustly prexed [postxed] points of any function (monotonic or not) is its least [greatest] robustly prexed [postxed] point. Corollary 3 NI 2 () (and also its any fragment, including NI) extended with operators ; or m; n is strongly normalizing and con uent. Mendler-style vs. conventional [co]induction It is also possible to encode , in terms of m, n and vice versa. For the encoding in the latter direction, 8 2 , 9 2 have to be available. Proposition 4 The following is a proof- and reduction-preserving encoding of , in terms of m, n: F (c) mapwrap(c; ()) cata ana F (e( open F (c) mapopen(c; ()) Proposition 5 The following is a proof- and reduction-preserving encoding of m, n in terms of , in the presence of 9 2 , iter F a (R) 8 2 ((Y )(R!Y )!F (Y coit mapopen d) open F a(c) (d) The encoding of m, n in terms of , is a proof-theoretic version of the observation that a least [greatest] prexed [postxed] point of F e [F a ] is a least [greatest] robustly prexed [postxed] point of F . Enhanced [co]induction The logical constants from the two upper front nodes of the cube capture the enhanced (in other words, \primitive-[co]recursive") forms of conventional- and Mendler-style [co]inductive denition. q and q are operators of enhanced induction and coinduction; m q and n q are their Mendler-style counterparts. Their proof and reduction rules are given in Figures 7 and 8. Adding q , q to a proof system presupposes the presence of ^, _; there is no corresponding restriction governing the addition of m q , n q . q I e( para F (wrap q para F (fst(); e); e( (R _ q open q open q Figure 7: Proof and reduction rules for q and q . From the algebraic semantics point-of-view, q F is a least \recursive" prexed point of a given (necessarily monotonic) F , i.e., a least element of the set of all R's such that F (R ^ q F ) is less than R (note the recurrent occurrence of q F here!). q F is a greatest \recursive" postxed point of F . is a least \recursive" robustly prexed point of a given F , ie., a least element of the set of all R's such that F (Y ) is less than R for all Y 's less than not only R but also m q F (note again the circularity!). n q F , dually, is a greatest \recursive" robustly postxed point of F . For programming, q F is a \recursive" data type, with wrap q F a \recursive" data constructor and para F a primitive recursor, and q F is a \recursive" codata type, with apo F a primitive corecursor and open q F a \recursive" codata e( rec(mapwrap q (c; d; i); e) e(c; ()rec(d(); e); ()i()) e( mapopen q (c; d; mapopen q (cor(c; e); d; i) e(c; ()d(cor(; e)); ()i()) Figure 8: Proof and reduction rules for m q and n q . destructor. m q F , with mapwrap q and rec, and n q F , with cor and mapopen q , are their Mendler-style equivalents. Returning to our running example of naturals, specializing enhanced induction for N yields the type Nat q of \recursive" natural numbers, with zero q , succ q and natpara the \recursive" constant zero, \recursive" successor function and primitive recursor. Nat q q (N) zero q wrap q succ q (c) wrap q The typing and reduction rules for Nat q are the following: Note that a non-zero \recursive" natural is constructed from a pair of naturals. In the reduction rule, the rst of them is used as the argument for the recurrent applications of the function being dened, while the second one is used directly. In principle, the two naturals can be unrelated, but the normal usage of the construction is that the second natural is equal to the rst (the predecessor), so the standard successor function is recovered by duplicating its argument. zero zero q succ(c) succ q (hc; ci) The type Nat q of \recursive" Mendler-style naturals is dened as follows: Nat mapzero q (d; i) mapwrap q (inl(hi); d; i) Nat q obeys the following typing and reduction rules: mapzero q (d; e z natrec(mapzero q (d; i); e z A non-zero \recursive" Mendler-style natural is constructed from a representation (for the predecessor), a method for converting representations to naturals and another function from representations to naturals. In the normal usage of the construction, the second method is also a conversion method. Choosing Nat q as the type of representations, the standard constructors are obtained as follows: zero mapzero q ((); ()) succ(c) mapsucc q (c; (); ()) On \recursive" naturals constructed using the standard constructors, natpara and natrec capture standard primitive recursion. The factorial function, for instance, can be programmed as follows: A degenerate application of primitive recursion, which only uses the \direct- access" predecessors of non-zero naturals, gives a fast (constant time) program for the predecessor function: pred(c) natpara(c; inl(hi); ( pred(c) natrec(c; (-; )inl(hi); ( Enhanced vs. basic [co]induction Both , and m, n can be encoded in terms of q , q and m q , n q . The converse is also true, but only if the retractive recursive denition operator is available. Proposition 6 The following is a proof- and reduction-preserving encoding of , in terms of q , F (c) wrap q F F cata F F ana F F open F (c) map F (open q Proposition 7 The following is a proof- and reduction-preserving encoding of m, n in terms of m q , d) mapwrap q (c; d; d) iter coit mapopen d) mapopen q (c; d; d) Proposition 8 The following is a proof- and reduction-preserving encoding of q , q in terms of , in the presence of : q F q (R) F (R ^ q F (c) i(wrap F q F q para F cata F q q F q (R) F (R _ q apo F open q F q (open F q (o(c)); ()i()) Proposition 9 The following is a proof- and reduction-preserving encoding of m q , n q in terms of m, n in the presence of : F q (R) (R!m q mapwrap q rec F q (R) (n q cor In the last two encodings, we would really like to dene q cannot (because of the circularity). Resorting to is a way to overcome this obstacle. From the result in [32], it follows that using is a necessity, one cannot possibly do without it. The rst of these encodings is implicit in [25] and [15]. It also appears in [7]. The second seems to be new. (and also its any fragment, including NI) extended with operators q ; q or m q ; n q is strongly normalizing and con uent. 5 Course-of-value [co]induction The logical constants from the two lower rear nodes of the cube capture the course-of-value forms of conventional- and Mendler-style [co]inductive deni- tion. ? and ? are operators of course-of-value induction and coinduction; m ? and n ? are their Mendler-style counterparts. Their proof and reduction rules are given in Figures 9 and 10. Adding ? , ? to a proof system presupposes the presence of ^, , _, ; there is no corresponding restriction governing the addition of m ? (R 4 F )(P ) F e( cvcata F cvcata F (wrap ? cvcata F (fst(open 4F ( snd(open 4F ( (R e( open ? open ? F (cvana F (c; e)) F ()cata 5F (; ( )wrap 5F (case@ Figure 9: Proof and reduction rules for ? and ? . e( e( mapopen ? (c; d; mapopen ? (cvcoit(c; e); d; Figure 10: Proof and reduction rules for m ? and n ? . From the algebraic semantics point-of-view, ? F is a least course-of-value prexed point of a given (necessarily monotonic) F , i.e., a least element of the set of all R's such that F ((Z)R ^ F (Z)) is less than R. ? F is a greatest course-of-value postxed point of F . is a least course-of-value robustly prexed point of a given F , i.e., a least element of the set of all R's such that F (Y ) is less than R for all Y 's less than not only R but also F (Y dually, is a greatest course-of-value robustly postxed point of F . For programming, ? F is a course-of-value data type, with wrap ? F a course- of-value data constructor and cvcata F a course-of-value iterator, and ? F is a course-of-value codata type, with cvana F a course-of-value iterator and open ? F a course-of-value codata destructor. m ? F , with mapwrap ? and cviter, and n ? F , with cvcoit and mapopen ? , are their Mendler-style equivalents. Specializing course-of-value induction for N yields the type Nat ? of \course-of- value" natural numbers, with zero ? , succ ? and natcviter the \course-of-value" versions of constant zero, successor function and iterator respectively. Nat ? ? (N) zero ? wrap ? The specialized typing and reduction rules for these constants are the following Similarly to the \recursive" case, non-zero \course-of-value" naturals are not constructed from a single preceding natural. The argument of the \course-of- value" successor function is a colist-like structure of naturals. The coiteration in the reduction rule applies the function being dened recurrently to every element of the colist. In principle, again, the naturals in the colist can be unrelated. The normal usage, however, is that the tail of the colist is the ancestral of its head (the predecessor of the natural being constructed). (By the ancestral of a natural, we mean the colist of all lesser naturals in the descending order.) The standard successor function for naturals is therefore easily recovered from the \course-of-value" successor function by rst coiteratively applying the predecessor function to its argument. zero zero ? )h pred( The predecessor function, however, does not admit a very straightforward definition (this is a problem that vanishes in the case of course-of-value primitive recursion). But it is denable in terms of the ancestral function, which itself is denable by course-of-value iteration in the same way as the predecessor function is denable by simple iteration. pred(c) (pred ? (c); ()fst(open())) pred ? (c) cvcata N (c; ( The specialization of course-of-value Mendler-style induction for N yields the Nat ? of \course-of-value" Mendler-style naturals. Nat mapzero ? (d; The derived typing and reduction rules for the above-dened constants are the following: e z A non-zero \course-of-value" Mendler-style natural is constructed from three components. The rst two are the same as in the case of simple Mendler- style naturals: a representation for a natural (the predecessor) and a method to convert representations to naturals. The additional third component gives a method for converting a representation (for some natural) into nothing or another representation (normally for the predecessor of this natural). So, using Nat ? as the type of representations, we obtain the standard constructors of naturals as follows: zero mapzero ? ((); pred) succ(c) mapsucc ? (c; (); pred) To dene the predecessor function, we again need also the ancestral function. pred(c) (pred ? (c); ()fst(open())) pred ? (c) cviter(c; ( On \course-of-value" naturals constructed using the standard constructors, natcvcata and natcviter capture standard course-of-value iteration. The Fibonacci function, for instance, can be programmed using natcvcata as follows. bo(c) natcvcata(c; zero; ( )case@ snd(open( Using natcviter, the denition of the Fibonacci function becomes even more straightforward, as, instead of having to manipulate an intermediate colist of values that Fibonacci returns, we can \roll back" on inputs to it. bo(c) natcviter(c; (-; )zero; ( Course-of-value vs. basic [co]induction Encoding , and m, n in terms of ? , ? and m ? , n ? is very similar to encoding these constants in terms of q , q and m q , n q . Also encoding in the opposite direction is analogous and, in fact, even simpler (as in not needed). Proposition 11 The following is a proof- and reduction-preserving encoding of , in terms of ? , F (c) wrap ? F cata )case@ ()fst(open 4F i ana F F open F (c) map Proposition 12 The following is a proof- and reduction-preserving encoding of m, n in terms of m ? iter coit mapopen Proposition 13 The following is a proof- and reduction-preserving encoding of ? , ? in terms of , : F wrap F ?(c) cvcata F (c; e) cata F ?(c; e) cvana F F open F ?(c) Proposition 14 The following is a proof- and reduction-preserving encoding of m ? , n ? in terms of m, n: mapwrap ?] (c; d; cviter cvcoit mapopen ?] (c; d; Corollary 15 NI 2 () (and also its any fragment, including NI) extended with operators strongly normalizing and con uent. 6 Related work The rst author to extend an intuitionistic n.d. system with (basic conventional- style) inductively dened predicates uniformly by axiomatization was Martin- Lof, with his \theory of iterated inductive denitions" [21]. Bohm and Berarducci [1] and Leivant [14] were the rst authors to describe how to encode \polynomial" (basic conventional-style) inductive types in 2nd- order simply typed lambda calculus (Girard and Reynold's system F; the n.d. proof system for the !,8 2 -fragment of 2nd-order intuitionistic propositional logic). This method is often referred to as the impredicative encoding of inductive types (keeping in mind only basic conventional-style induction). Mendler [19] described the extension by axiomatization of 2nd-order simply typed lambda calculus with enhanced inductive and coinductive types of his style. Mendler [20] discussed a similar system with basic Mendler-style inductive and coinductive types. Extensions of the n.d. proof systems for 2nd-order intuitionistic predicate logic with constructors of (basic) conventional- and Mendler-style inductive predicates were described in Leivant's [15], a paper on extracting programs in (extensions of) 2st-order simply typed lambda calculus from proofs in (extensions of) the n.d. proof system for 2nd-order intuitionistic predicate logic. Parigot's work [24,25] on realizability-based \programming with proofs" bears connection to both Leivant's and Mendler's works. Greiner [10] and Howard [13, chapter 3] considered programming in an extension of 1st-order simply typed lambda calculus with axiomatized constructors of conventional-style (co)inductive types with (co)iteration and data destruction (codata construction). Both had their motivation in Hagino's category-theoretic work cited below and studied thus not barely -reduction, but even -conversion, driven by denite semantic considerations. Howard implemented his system in a programming language Lemon. Geuvers [7] carried out a comparative study of basic vs. enhanced, conventional- vs. Mendler- style inductive and coinductive types in extensions of 2nd-order simply typed lambda calculus. In the spirit of Leivant, Paulin-Mohring [26] extracted programs in Girard's F ! from proofs in Coquand and Huet's CC (calculus of constructions). The milestone papers on inductive type families in extensions of CC and Luo's ECC (extended calculus of constructions, a combination of CC and Martin-Lof's type theory) are Pfenning and Paulin-Mohring [28], Coquand and Paulin-Mohring [4] and Ore [23]. Paulin-Mohring [27] formulated the calculus of inductive constructions, which extends CC with inductive type families with primitive recursion by axiomatization. The Coq proof development system developed at INRIA-Rocqencourt and ENS-Lyon is an implementation of this last system. In category theory, (basic conventional-style) inductive and coinductive types are modelled by initial algebras and terminal coalgebras for covariant functors. Hagino [11] designed a typed functional language CPL based on distributive categories with initial algebras and terminal coalgebras for strong covariant functors. The implemented Charity language by Cockett et al. [3] is a similar programming language. The \program calculation" community is rooted in the Bird-Meertens formalism or Squiggol [2], which, originally, was an equational theory of programming with the parametric data type of lists. Malcolm [16] made the community aware of Hagino's work, and studied program calculation based on bi-Cartesian closed categories with initial algebras and terminal coalgebras for !-cocontinuous resp. !-continuous covariant functors. Meertens [18] was the rst author to give a treatment of primitive-recursion in this setting. Some classic references in the area are Fokkinga's [6] and Sheard and Fegaras' [29]. 7 Conclusion and Future Work In this paper, we studied least and greatest xed point operators that intuitionistic n.d. systems can be extended with. We described eight pairs of such operators whose eliminations and introductions behave as recursors and corecursors of meaningful kinds. We intend to continue this research with a study of the perspectives of the utility of intuitionistic n.d. systems with least and greatest xed point operators in program construction from specications; this concerns both specication methodology and computer assistance in synthesis. We have also started to study the relating categorical deduction systems (typed combinatory logics a la Curien), their utility in \program calculation" and the relevant categorical theory [38,36,39,40]. We also intend to nd out the details of the apparent close relationship of enhanced course-of-value Mendler-style (co)recursion to Gimenez' new formulation of guarded (co)recursion [9] (for systems with sub- and supertyping and quantication with upper and lower bounds; radically dierent from the older, very syntactical formulation of [8]). Acknowledgements We are thankful to our anonymous referees for a number of helpful comments and suggestions, especially in regards to matters of presentation. The proof gures and diagrams appearing in the paper were typeset using the proof.sty by Makoto Tatsuta and the XYpic generic by Kristoer C. Rose, respectively. --R An introduction to the theory of lists Yellow Series Report 92/480/18 Inductively de A survey of the project AUTOMATH Law and order in algorithmics Inductive and coinductive types with iteration and recursion A categorical programming language A framework for de Fixed points and extensionality in typed functional programming languages Reasoning about functional programs and complexity classes associated with type disciplines Contracting proofs to programs Data structures and program transformation Extensions of system F by iteration and primitive recursion on monotone inductive types Recursive types and type constraints in second-order lambda- calculus Inductive types and type constraints in the second-order lambda- calculus The extended calculus of constructions (ECC) with inductive types a second order type theory Recursive programming with proofs -Mohring, Extracting F! -Mohring, Inductive de -Mohring, Inductively de A fold for all seasons A natural extension of natural deduction Generalized rules for quanti A lattice-theoretical xpoint theorem and its applications Positive recursive type assignment Natural deduction for intuitionistic least and greatest A cube of proof systems for the intuitionistic predicate Primitive (co)recursion and course-of-value (co)iteration Coding recursion --TR An introduction to the theory of lists Extracting MYAMPERSANDohgr;''s programs from proofs in the calculus of constructions Inductively defined types Programming in Martin-LoMYAMPERSANDuml;f''s type theory: an introduction Data structures and program transformation Inductively defined types in the calculus of constructions Recursive programming with proofs A framework for defining logics The extended calculus of constructions (ECC) with inductive types A fold for all seasons Fixed points and extensionality in typed functional programming languages Type fixpoints Programming with Proofs Positive Recursive Type Assignment Inductive Definitions in the system Coq - Rules and Properties Structural Recursive Definitions in Type Theory Codifying Guarded Definitions with Recursive Schemes Mendler-style inductive types, categorically Programming with Inductive and Co-Inductive Types A categorical programming language --CTR Gilles Barthe , Tarmo Uustalu, CPS translating inductive and coinductive types, ACM SIGPLAN Notices, v.37 n.3, p.131-142, March 2002 G. Barthe , M. J. Frade , E. Gimnez , L. Pinto , T. Uustalu, Type-based termination of recursive definitions, Mathematical Structures in Computer Science, v.14 n.1, p.97-141, February 2004

coding styles;least and greatest fixed points;coinductive types;schemes of total corecursion;typed lambda calculi;natural deduction

506696

Skepticism and floating conclusions.

The purpose of this paper is to question some commonly accepted patterns of reasoning involving nonmonotonic logics that generate multiple extensions. In particular, I argue that the phenomenon of floating conclusions indicates a problem with the view that the skeptical consequences of such theories should be identified with the statements that are supported by each of their various extensions.

Introduction One of the most striking ways in which nonmonotonic logics can differ from classical logic, and even from standard philosophical logics, is in allowing for multiple sanctioned conclusion sets, known as extensions. The term is due to Reiter [12], who thought of default rules as providing a means for extending the strictly logical conclusions of a knowledge base with plausible information. Multiple extensions arise when a knowledge base contains conflicting default rules, suggesting different, often inconsistent ways of supplementing its strictly logical conclusions. The purpose of this paper is to question some commonly accepted patterns of reasoning involving theories that generate multiple extensions. In particular, I argue that the phenomenon of floating conclusions indicates a problem with the view that the skeptical consequences of such theories should be identified with the statements that are supported by each of their various extensions. Multiple extensions The canonical example of a knowledge base with multiple extensions is the Nixon Diamond, depicted in Figure 1. Here, the statements Qn, Rn, and Pn represent the propositions that Nixon is a Quaker, a Republican, and a pacifist; statements of the form A ordinary logical implications and "default" implications respectively, with A abbreviating A ) :B and A ! :B; and the special statement ? represents truth. What the knowledge base tells us, of course, is this: Nixon is both a Quaker and a Republican, the fact that he is a Quaker provides a good reason for concluding that he is a pacifist, and the fact that he is a Republican provides a good reason for concluding that he is not a pacifist. This example can be coded into default logic as the theory representing the basic facts of the situation and representing the two defaults. The theory yields two extensions: f:Png). The first results when the basic facts of the situation are extended by an application of the default concerning Quakers; the second results when the facts are extended by an application of the default concerning Republicans. In light of these two extensions, what are we to conclude from the initial information: is Nixon a pacifist or not? More generally, when a default theory leads to more than one extension, what should we actually infer from that theory-how should we define its set of consequences, or conclusions? Several proposals have been discussed in the literature. One option is to suppose that we should arbitrarily select a particular one of the theory's several extensions and endorse the conclusions contained in it; a second option is to suppose that we should be willing to endorse a conclusion just in case it is contained in some extension of the theory. These first two options are sometimes said to reflect a credulous reasoning policy. A third option, now generally described as skeptical, is to suppose that we should endorse a conclusion just in case it is contained in every extension of the theory. 1 Of these three options, the first-pick an arbitrary extension-really does seem to embody a sensible policy, or at least one that is frequently employed. Given conflicting defeasible information, we often simply adopt some internally coherent point of view in which the conflicts are resolved in some particular way, regardless of the fact that there are other coherent points of view in which the conflicts are resolved in different ways. Still, although this reasoning policy may be sensible, it 1 The use of the credulous/skeptical terminology in this context was first introduced by Touretzky et al. [15], but the distinction itself is older than this; it was already implicit in Reiter's paper on default logic, and was described explicitly by McDermott [10] as the distinction between brave and cautious reasoning. Makinson [8] refers to the first of the two credulous options described here as the choice option. s s s s \Gamma' @ @ @ @ @ @I Rn Pn Qn Figure 1: The Nixon Diamond is hard to see how it could be codified in a formal consequence relation. If the choice of extension really is arbitrary, different reasoners could easily select different extensions, or the same reasoner might select different extensions at different times. Which extension, then, would represent the real conclusion set of the original theory? The second of our three options-endorse a conclusion whenever it is contained in some extension- could indeed be codified as a consequence relation, but it would be a peculiar one. According to this policy, the conclusion set associated with a default theory need not be closed under standard logical consequence, and might easily be inconsistent, even in cases in which the underlying default theory itself seems to be consistent. The conclusion set of the theory representing the Nixon Diamond, for example, would contain both Pn and :Pn, since each of these formulas belongs to some extension of the default theory, but it would not contain Pn - :Pn, since this formula is not contained in any extension. One way of avoiding these peculiar features of the second option is to think of the conclusions generated by a default theory as being shielded by a kind of modal operator. Where A is a statement, let B(A) mean there is good reason to believe that A; and suppose a theory provides us with good reason to believe a statement whenever that statement is included in some extension of the theory, some internally coherent point of view. Then we could define the initial conclusions of a default theory as the set that extends W with a formula B(A) whenever A belongs to some extension of \Delta, and we could go on to define the theory's conclusion set as the logical closure of its initial conclusions. This variant of the second option has some interest. It results in a conclusion set that is both closed under logical consequence and consistent as long as W itself is consistent. And Reiter's original paper on default logic [12, Section 4] provides a proof procedure, sound and complete under certain conditions, that could be used in determining whether B(A) belongs to the conclusion set as defined here. Unfortunately, however, this variant of the second option also manages to sidestep our original question. We wanted to know what conclusions we should actually draw from the information provided by a default theory-whether or not, given the information from the Nixon Diamond, we should conclude that Nixon is a pacifist, for example. But according to this variant, we are told only what there is good reason to believe-that both B(Pn) and B(:Pn) are consequences of the theory, so that there is good reason to believe that Nixon is a pacifist, but also good reason to believe that he is not. This may be useful information, but it is still some distance from telling us whether or not to conclude that Nixon is a pacifist. 2 Of our three options for defining a notion of consequence in the presence of multiple exten- sions, only the third, skeptical proposal-endorse a conclusion whenever it is contained in every extension-seems to hold any real promise. This option leads to a single conclusion set, which is both closed under logical consequence and consistent as long as the initial information is. And it provides an answer that is at least initially attractive to our original question concerning proper conclusions. In the case of the Nixon Diamond, for example, since neither Pn nor :Pn belongs to every extension, this third option tell us that we should not conclude that Nixon is a pacifist, but that we should not conclude that Nixon is not a pacifist either. Since there is a good reason for each of these conflicting conclusion, we should remain skeptical. Floating conclusions logic defines a direct, unmediated relation between a particular default theory and the statement sets that form its extensions. Another class of formalisms-known as argument systems-takes a more roundabout approach, analyzing nonmonotonic reasoning through the study of interactions among competing defeasible arguments. 3 Although the arguments themselves that are studied in these argument systems are often com- plex, we can restrict our attention here entirely to linear arguments, analogous to the reasoning paths studied in theories of defeasible inheritance. 4 These linear arguments are formed by starting with a true statement and then simply stringing together strict and defeasible implications; each such argument can be said to support the final statement it contains as its conclusion. As an abstract example, the structure ? can be taken to represent an argument of the form "A is true, which defeasibly implies B, which strictly implies :C," supporting the conclusion :C. As a less abstract example, we can see that the Nixon Diamond provides the materials for constructing the two arguments ? supporting the conflicting conclusions Pn and :Pn. Where ff is an argument, we will let ff represent the particular conclusion supported by ff. Where \Phi is a set of arguments, we will let \Phi represent the set of conclusions supported by the arguments in \Phi-that is, the set containing the statement ff for each argument ff belonging to \Phi. The primary technical challenge involved in the development of an argument system is the specification of the coherent sets of arguments that an ideal reasoner might be willing to accept on the basis of a given body of initial information. We will refer to these coherent sets of arguments as argument extensions, to distinguish them from the statement extensions defined by theories such as default logic. Again, the actual definition of argument extensions is often complicated in ways that need not concern us here. Without going into detail, however, we can simply note that, just as theories like default logic allow multiple statement extensions, argument systems often associate multiple argument extensions with a single body of initial information. In the case of the Nixon Diamond, for example, an argument system patterned after multiple-extension theories of defeasible 2 Note that this objection is directed only against the use of modal operators to capture the epistemic interpretation of default logic. Other interpretations, involving other modal operators, are possible; it is shown in [4], for example, that a deontic interpretation, with default conclusions wrapped inside of deontic operators, generates a logic for normative reasoning corresponding to that originally suggested by van Fraassen [16]. 3 A recent survey of a variety of argument systems can be found in Prakken and Vreeswijk [11]. 4 The development of the path-based approach to inheritance reasoning was initiated by Touretzky [14]; a survey can be found in [5]. inheritance would generate the two extensions The first results from supplementing the initial information with the argument that Nixon is a pacifist because he is a Quaker, the second from supplementing this information with the argument that Nixon is not a pacifist because he is a Republican. When a knowledge base leads to multiple argument extensions, there are, as before, several options for characterizing the appropriate set of conclusions to draw on the basis of the initial infor- mation. Again, we might adopt a credulous reasoning policy, either endorsing the set of conclusions supported by an arbitrary one of the several argument extensions, or perhaps endorsing a conclusion as believable whenever it is supported by some extension or another. In the case of the Nixon Diamond, this policy would lead us to endorse either \Phi 1 Png or \Phi 2 as the conclusion set of the original knowledge base, or perhaps simply to endorse the statements belonging to the union of these two sets as believable. As before, however, we might also adopt a kind of skeptical policy in the presence of these multiple argument extensions, defining the appropriate conclusion set through their intersection. In this case, though, since these new extensions contain arguments rather than statements, there are now two alternatives for implementing such a policy. First, we might decide to endorse an argument just in case it is contained in each argument extension associated with an initial knowledge base, and then to endorse a conclusion just in case that conclusion is supported by an endorsed argument. Formally, this alternative leads to the suggestion that the appropriate conclusions of an initial knowledge base \Gamma should be the statements belonging to the set is an extension of \Gammag): Or second, we might decide to endorse a conclusion just in case that conclusion is itself supported by each argument extension of the initial knowledge base \Gamma, leading to the formal suggestion that the appropriate conclusions of the knowledge base should be the statements belonging to the set f \Phi : \Phi is an extension of \Gammag; where the order of and T is reversed. Of course, these two alternatives for implementing the skeptical policy come to the same thing in the case of the Nixon Diamond: both lead to fQn; Rng as the appropriate conclusion set. But there are other situations in which the two alternatives yield different results. A well-known example, due to Ginsberg, appears in Figure 2, where Qn and Rn are interpreted as before, Dn and Hn are interpreted to mean that Nixon is a dove or a hawk respectively, and En as meaning that Nixon is politically extreme (regarding the appropriate use of military force). What this diagram tells us is that Nixon is both a Quaker and a Republican, that there is good reason to suppose that Nixon is a dove if he is a Quaker, a hawk if he is a Republican, and that he is politically extreme if he is either a dove or a hawk. Again, a system patterned after multiple-extension inheritance theories would associate two argument extensions with this knowledge base, as follows: @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ s s Rn Hn Dn Qn En Figure 2: Is Nixon politically extreme? Since no arguments except for the trivial ? ) Qn and ? ) Rn are contained in both of these ex- tensions, the first of our two alternatives for implementing the skeptical policy, which involves intersecting the argument extensions themselves, would lead to fQn; Rng as the appropriate conclusion set, telling us nothing more than the initial information that Nixon is a Quaker and a Republican. On the other hand, each of these two argument extensions supports the statement En-one through the argument ? the other through the argument ? En. The second of our two alternatives for implementing the skeptical policy, which involves intersecting supported statements rather than the arguments that support them, would therefore lead to the conclusion set fQn; Rn; Eng, telling us also that Nixon is politically extreme. Statements like En, which are supported in each extension associated with a knowledge base, but only by different arguments, are known as floating conclusions. This phrase, coined by Makinson and Schlechta [9], nicely captures the picture of these conclusions as floating above the different and conflicting arguments that might be taken to support them. The phenomenon of floating conclusions was first investigated in the context of defeasible inheritance reasoning, particularly in connection with the theory developed by Thomason, Touretzky, and myself in [6]. In contrast to the multiple-extension accounts considered so far, that theory first defined a single argument extension that was thought of as containing the "skeptically accept- able" arguments based on a given inheritance network. The skeptical conclusions were then defined simply as the statements supported by those skeptically acceptable arguments. Ginsberg's political extremist example was meant to show that no approach of this sort, relying on a single argument extension, could correctly represent skeptical reasoning. A single argument extension could not consistently contain both the arguments ? since the strict information in the knowledge base shows that each of these arguments conflicts with an initial segment of the other. The single argument extension could not contain either of these arguments without the other, since that would involve the kind of arbitrary decision appropriate only for credulous reasoning. And if the single argument extension were to contain neither of these two arguments, it would not support the conclusion En, which Ginsberg considers to be an intuitive consequence of the initial information: "given that both hawks and doves are politically [extreme], Nixon certainly should be as well" [3, p. 221]. 5 Both Makinson and Schlechta [9] and Stein [13] also consider floating conclusions in the context of defeasible inheritance reasoning. Makinson and Schlechta share Ginsberg's view that the appropriate conclusions to derive from a knowledge base are those that are supported by each of its argument extensions: It is an oversimplification to take a proposition A as acceptable . iff it is supported by some [argument] path ff in the intersection of all extensions. Instead A must be taken as acceptable iff it is in the intersection of all outputs of extensions, where the output of an extension is the set of all propositions supported by some path within it [9, pp. 203-204]. From this they likewise argue, not only that the particular theory developed in [6] is incorrect, but more generally, that any theory attempting to define the skeptically acceptable conclusions by reference to a single set of acceptable arguments will be mistaken. And Stein reaches a similar judgment, for similar reasons: The difficulty lies in the fact that some conclusions may be true in every credulous extension, but supported by different [argument] paths in each. Any path-based theory must either accept one of these paths-and be unsound, since such a path is not in every extension-or reject all such paths-and with them the ideally skeptical conclusion- and be incomplete [13, p. 284]. What lies behind these various criticisms, of course, is the widely-held assumption that the second, rather than the first, of our two skeptical alternatives is correct-that floating conclusions should be accepted, and that a system that fails to classify them among the consequences of a defeasible knowledge base is therefore in error. The purpose of this paper is to question that assumption. 4 An example Why not accept floating conclusions? Their precarious status can be illustrated through any number of examples, but we might as well choose a dramatic one. Suppose, then, that my parents have a net worth of one million dollars, but that they have divided their assets in order to avoid the United States inheritance tax, so that each parent currently possess half a million dollars apiece. And suppose that, because of their simultaneous exposure to a fatal disease, it is now settled that both of my parents will die within a month. This is a fact: medical science is certain. Imagine also, however, that there is some expensive item-a yacht, say-whose purchase I believe would help to soften the blow of my impending loss. Although the yacht I want is currently 5 Although, as far as I know, this example was first published in the textbook cited here, it had previously been part of the oral tradition for many years-I first heard it during the question session after the AAAI-87 presentation of [6], when Ginsberg raised it as an objection to that theory. available, the price is good enough that it is sure to be sold by the end of the month. I can now reserve the yacht for myself by putting down a large deposit, with the balance due in six weeks. But there is no way I can afford to pay the balance unless I happen to inherit at least half a million dollars from my parents within that period, and if I fail the pay the balance on time, I will lose my large deposit. Setting aside any doubts concerning the real depth of my grief, let us suppose that my utilities determine the following conditional preferences: if I believe I will inherit half a million dollars from my parents within six weeks, it is very much in my benefit to place a deposit on the yacht; if I do not believe this, it is very much in my benefit not to place a deposit. Now suppose I have a brother and a sister, both of whom are extraordinarily reliable as sources of information. Neither has ever been known to be mistaken, to deceive, or even to misspeak- although of course, like nearly any source of information, they must be regarded as defeasible. My brother and sister have both talked with our parents about their wills, and feel that they understand the situation. I have written to each of them describing my delicate predicament regarding the yacht, and receive letters back. My brother writes: "Father is going to leave his money to me, but Mother will leave her money to you, so you're in good shape.'' My sister writes: "Mother is going to leave her money to me, but Father will leave his money to you, so you're in good shape." No further information is now available: the wills are sealed, my brother and sister are trekking together through the Andes, and our parents, sadly, have slipped into a coma. Based on my current information, what should I conclude? Should I form the belief that I will inherit half a million dollars-and therefore place a large deposit on the yacht-or not? The situation is depicted in Figure 3, where the statement letters are interpreted as follows: F represents the proposition that I will inherit half a million dollars from my father, M represents the proposition that I will inherit half a million dollars from my mother, BA(:F - M ) represents the proposition that my brother asserts that I will inherit my mother's money but not my father's, and represents the proposition that my sister asserts that I will inherit my father's money but not my mother's. The defeasible links BA(:F - M reflect the fact that any assertion by my brother or sister provides good reason for concluding that the content of that assertion is true. The strict links in the diagram record various implications and inconsistencies. Notice that, although the contents of my brother's and sister's assertions-the statements :F -M and F -:M-are jointly inconsistent, the truth of either entails the disjunctive which is, of course, all I really care about. As long as I can conclude that I will inherit half a million dollars from either my father or my mother, I should go ahead and place a deposit on the yacht. A multiple-extension approach would associate the following two argument extensions with this knowledge base: @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ s s Figure 3: Should I buy the yacht? Again, the first of our two alternatives for implementing the skeptical reasoning policy, which involves intersecting arguments, would lead to fBA(:F - M ); SA(F - :M )g as the appropriate conclusion set, telling me only that my brother and sister asserted what they did. But since each of the two extensions contains some argument supporting the statement F -M , the second alternative, which involves intersecting supported statements, leads to the conclusion set fBA(:F-M ); SA(F- telling me also-as a floating conclusion-that I will inherit half a million dollars from either my father or my mother. In this situation, then, there is a vivid practical difference between the two skeptical alternatives. If I were to reason according to the first, I would not be justified in concluding that I am about to inherit half a million dollars, and so it would be foolish for me to place a deposit on the yacht. If I were to reason according to the second, I would be justified in drawing this conclusion, and so it would be foolish for me not to place a deposit. Which alternative is correct? I have not done a formal survey, but most of the people to whom I have presented this example are suspicious of the floating conclusion, and so favor the first alternative. Most do not feel that the initial information from Figure 3 would provide sufficient justification for me to conclude, as the basis for an important decision, that I will inherit half a million dollars. Certainly, this is my own opinion-that the example shows, contrary to the widely-held assumption, that it is at least coherent for a skeptical reasoner to withhold judgment from floating conclusions. Although both my brother and sister are reliable, and each supports the conclusion that I will inherit half a million dollars, the support provided by each of these reliable sources is undermined by the other; there is no unopposed reason supporting the conclusion. Since either my brother or sister must be wrong, it is therefore easy to imagine that they might both be wrong, and wrong in this way: perhaps my father will leave his money to my brother and my mother will leave her money to my sister, so that I will inherit nothing. 5 Comments on the example First, in case this example does not yet seem convincing, it might help to modify things a bit. Suppose, then, that I had written only to my brother, and received his response-that my father had named him as sole beneficiary, but that my mother would leave her money to me. That is, suppose my starting point is the information depicted in the left-hand side of Figure 3. In this new situation, should I conclude that I will inherit half a million dollars, and therefore place a deposit on the yacht? Some might say no-that even in this simplified situation I should not make such a important decision on the basis of my brother's word alone. But this objection misses the point. Most of what we know, we know through sources of information that are, in fact, defeasible. By hypothesis, we can suppose that my brother is arbitrarily reliable, as reliable as any defeasible source of information could possibly be-as reliable as perception, for instance, or the bank officer's word that the money has actually been deposited in my account. If we were to reject information like this, it is hard to see how we could get by in the world at all. When a source of defeasible information that is, by hypothesis, arbitrarily reliable tells me that I will inherit half a million dollars, and there is no conflicting evidence in sight, it is reasonable for me to accept this statement, and to act on it. Note that both of the two skeptical alternatives yield this outcome in our simplified situation, since the initial information, represented by the left-hand side of Figure 3, generates only a single argument extension, in which the conclusion that I will inherit half a million dollars is supported by a single argument. Now suppose that, at this point, I hear from my equally reliable sister with her conflicting information-that she is my mother's beneficiary, but that my father will leave his money to me. As a result, I am again in the situation depicted in the full Figure 3, with two argument extensions, and in which the statement that I will inherit half a million dollars is supported only as a floating conclusion. Ask yourself: should my confidence in the statement that I will inherit half a million dollars be diminished in this new situation, now that I have heard from my sister as well as my brother? If it seems that my confidence can legitimately be diminished-that this new information casts any additional doubt on the outcome-then it follows that floating conclusions are somewhat less secure than conclusions that are uniformly supported by a common argument. And that is all we need. The point is not that floating conclusions might be wrong; any conclusion drawn through defeasible reasoning might be wrong. The point is that a statement supported only as a floating conclusion seems to be less secure than the same statement when it is uniformly supported by a common argument. As long as there is this difference in principle, it is coherent to imagine a skeptical reasoner whose standards are calibrated so as to accept statements that receive uniform support, but to reject floating conclusions. As a second comment, notice that, if floating conclusions pose a problem, it is not just a problem for argument systems, but also for traditional nonmonotonic formalisms, such as default or model- preference logics. Indeed, the problem is even more serious for these traditional formalisms. With argument systems, where the extensions generated are argument extensions, it is at least possible to avoid floating conclusions by adopting the first of our two skeptical alternatives-endorsing only those arguments belonging to each extension, and then endorsing only the conclusions of the endorsed arguments. Since arguments are represented explicitly in these systems, they can be used to filter out floating conclusions. In most traditional nonmonotonic logics, arguments are suppressed, and so the materials for carrying out this kind of filtering policy are not even available. To illustrate, a natural representation of the information from our yacht example in default logic is the theory describes what my brother and sister said and reflects the defaults that whatever my brother and sister say should be taken as true. This theory has two extensions: The extensions of default logic are statement extensions, and so the only possible policy for skeptical reasoning appears to be: intersect the extensions. Since the statement F - M belongs to both extensions, skeptical reasoning in default logic tells me, immediately and without ambiguity, that I will inherit half a million dollars. Of course, default logic is essentially a proof-theoretic formalism, and it is easy to see how it could be modified so that the extensions defined would contain proofs rather than statements; such a modification would then allow for floating conclusions to be filtered out by a treatment along the lines of our first alternative. 6 It is harder to see how floating conclusions could be avoided in model-preference logics. In a circumscriptive theory, for instance, the yacht example could naturally be expressed by supplementing the facts BA(:F - M ) and SA(F - :M ) with the statements (BA(:F - M ) - :Ab b preferring those models in which as few as possible of the propositions Ab b and Ab s -the abnormalities associated with the rare situations in which my brother or sister is mistaken-are true. Of course, there can be no models in which neither Ab b not Ab s is true. The most preferred models will therefore be those in which only one of these abnormalities holds. The statement F - M is true in all of these models, and would therefore follow as a circumscriptive consequence. 7 6 Objections to the example I have heard two objections worth noting to the yacht example as an argument against floating conclusions. The first focuses on the underlying methodology of logical formalization. Even though what my brother literally said is "Father is going to leave his money to me, but Mother will leave her money to you," one might argue that the real content of his statement-what he really meant-is better conveyed through the two separate sentences "Father is going to leave his money to me" and "Mother will leave her money to you." In that case, rather than formalizing my brother's assertion through the single conjunction :F - M , it would be more natural to represent its content through the separate statements :F and M ; and the content of my sister's assertion could likewise be formalized through the separate statements F and :M . Considered from the standpoint of default logic, the situation could then be represented through the new default theory 6 One suggested modification of default logic that is particularly relevant, because it bears directly on examples of the sort considered here, can be found in Brewka and Gottlob [2]. 7 This form of circumscription, which involved minimizing the truth of statements rather than the extensions of predicates, is a special case of the more usual form; see Lifschitz [7, pp. 302-303] for a discussion. describing what now appear to be the four independent assertions made by my brother and sister, and with carrying the defaults that any assertion by my brother or sister should be taken as true, if possible. This new default theory would then have four extensions: And since not all of these extensions contain the statement F - M , the policy of defining skeptical conclusions simply by intersecting the statements supported by each extension no longer leads, in this case, to the conclusion that I will inherit half a million dollars. The idea behind this objection is that the problems presented by floating conclusions might be avoided if we were to adopt a different strategy for formalizing the statements taken as inputs by the logical system, which would involve, among other things, articulating conjunctive inputs into their conjuncts. This idea is interesting, has some collateral benefits, and bears certain affinities to proposals that have been suggested in other contexts. 8 Nevertheless, in the present setting, the strategy of factoring conjunctive statements into their conjuncts in order to avoid undesirable floating conclusions suggests a procedure that might be described as "wishful formalization"- carefully tailoring the inputs to a logical system so that the system then yields the desired outputs. Ideally, a logic should take as its inputs formulas conforming as closely as possible to the natural language premises provided by a situation, and then the logic itself should tell us what conclusions follow from those premises. Any time we are forced to adopt a less straightforward representation of the input premises in order to avoid inappropriate conclusions-replacing conjunctions with their conjuncts, for example-we are backing away from that ideal. By tailoring the inputs in order to assure certain outputs, we are doing some work for the logic that, in the ideal case, the logic should be doing for us. The second objection to the yacht example as an argument against floating conclusions concerns the method for evaluating supported statements. Part of what makes this example convincing as a reason for rejecting the floating conclusion that I will inherit half a million dollars is the fact that it is developed within the context of an important practical decision, where an error carries significant consequences: I will lose my large deposit. But what if the consequences were less significant? Suppose the deposit were trivial: one dollar, say. In that case, many people would then argue that the support provided for the proposition that I will inherit half a million dollars-even as a floating 8 Imagine, for example, that my brother asserts a statement of the form P - Q, where it turns out that P is a logical contradiction-perhaps a false mathematical statement-but Q expresses a perfectly sensible proposition that just happens to be conjoined with P for reasons of conversational economy. Here, the representation of the situation through the default theory hW;Di with would prevent us from drawing either P or Q as a conclusion, since the justification for the default could not be satisfied. But if the situation were represented through the articulated theory hW;Di with and we could at least draw the conclusion Q. This idea of articulating premises into simpler components, in order to draw the maximum amount of information out of a set of input statements without actually reaching contradictory conclusions, has also been studied in the context of relevance logic; a carefully formulated proposal can be found in Section 82.4 of Anderson et al. [1]. conclusion- would be sufficient, when balanced against the possibility for gain, to justify the risk of losing my small deposit. The general idea behind this objection is that the proper notion of consequence in defeasible reasoning is sensitive to the risk of being wrong. The evaluation of a logic for defeasible reasoning cannot, therefore, be made outside of some particular decision-theoretic setting, with particular costs assigned to errors; and there are certain settings in which one might want to act even on the basis of propositions supported only as floating conclusions. This is an intriguing objection. I will point out only that, if accepted, it suggests a major revision in our attitude toward nonmonotonic logics. Traditionally, a logic-unlike a system for probabilistic or evidential reasoning-is thought to classify statements into only two categories: those that follow from some set of premises, and those that do not. The force of this objection is that nonmonotonic logics should be viewed, instead, as placing statements into several categories, depending on the degree to which they are supported by a set of premises, with floating conclusions then classified, not necessarily as unsupported, but perhaps only as less firmly supported than statements that are justified by the same argument in every extension. 7 Other examples Once the structure of the yacht example is understood, it is easy to construct other examples along similar lines: just imagine a situation in which two sources of information, or reasons, support a common conclusion, but also undermine each other, and therefore undermine the support that each provides for the common conclusion. Suppose you are a military commander pursuing an enemy that currently holds a strong defensive position. It is suicide to attack while the enemy occupies this position in force, but you have orders to press ahead as quickly as possible, and so you send out your reliable spies. After a week, one spy reports back that there can now be only a skeleton force remaining in the defensive position; he has seen the main enemy column retreating through the mountains, although he also noticed that they sent out a diversionary group to make it appear as if they were retreating along the river. The other spy agrees that only a skeleton force remains in the defensive position; he has seen the main enemy column retreating along the river, although he notes that they also sent out a diversionary group to make it appear is if they were retreating through the mountains. Based on this information, should you assume at least that the main enemy force has retreated from the defensive position-a floating conclusion that is supported by both spies-and therefore commit your troops to an attack? Not necessarily. Although they support a common conclusion, each spy undermines the support provided by the other. Perhaps the enemy sent out two diversionary groups, one through the mountains and one along the river, and managed to fool both your spies into believing that a retreat was in progress. Perhaps the main force still occupies the strong defensive position, awaiting your attack. Or suppose you attend a macroeconomics conference during a period of economic health, with low inflation and strong growth, and find that the community of macroeconomic forecasters is now split right down the middle. One group, working with a model that has been reliable in the past, predicts that the current strong growth rate will lead to higher inflation, triggering an economic downturn. By tweaking a few parameters in the same model, the other group arrives at a prediction according to which the current low inflation rate will actually continue to decline, leading to a dangerous period of deflation and triggering an economic downturn. Both groups predict an economic downturn, but for different and conflicting reasons-higher inflation versus deflation-and so the prediction is supported only as a floating conclusion. Based on this information, should you accept the prediction, adjusting your investment portfolio accordingly? Not necessarily. Perhaps the extreme predictions are best seen as undermining each other and the truth lies somewhere in between: the inflationary and deflationary forces will cancel each other out, the inflation rate will remain pretty much as it is, and the period of economic health will continue. There is no need to labor the point by fabricating further examples in which floating conclusions are suspect. But what about the similar cases, exemplifying the same pattern, that have actually been advanced as supporting floating conclusions, such as Ginsberg's political extremist example from Figure I have always been surprised that this particular example has seemed so persuasive to so many people. The example relies on our understanding that individuals adopt a wide spectrum of attitudes regarding the appropriate use of military force, but that Quakers and Republicans tend to be doves and hawks, respectively-where doves and hawks take the extreme positions that the use of military force is either never appropriate, or that it is appropriate in response to any provocation, even the most insignificant. Of course, Nixon's own position on the matter is well known. But if I were told of some other individual that he is both a Quaker and a Republican, I would not be sure what to conclude. It is possible that this individual would adopt an extreme position, as either a dove or a hawk. But it seems equally reasonable to imagine that such an individual, rather than being pulled to one extreme of the other, would combine elements of both views into a more balanced, measured position falling toward the center of the political spectrum-perhaps believing that the use of military force is sometimes appropriate, but only as a response to serious provocation. Given this real possibility, it might be appropriate to take a skeptical attitude, not only toward the questions of whether this individual would be a dove or a hawk, but also toward the question whether he would adopt a politically extreme position at all. Another example appears in Reiter's original paper on default logic, where he suggests [12, pp. 86-87] defaults representing the facts that people tend to live in the same cities as their spouses, but also in the cities in which they work, and then asks us to consider the case of Mary, whose spouse lives in Toronto but who works in Vancouver. Coded into default logic, this information leads to a theory with two extensions, in one of which Mary lives in Toronto and in one of which she lives in Vancouver. Reiter seems to favor the credulous policy of embracing a particular one of these extensions, either concluding that Mary lives in Toronto or concluding that Mary lives in But then, in a footnote, he also mentions what amounts to the skeptical possibility of forming only the belief that Mary lives in either Toronto or Vancouver-where this proposition is supported, of course, as a floating conclusion. Given the information from this example, I would, in fact, be likely to conclude that Mary lives either in Toronto or Vancouver. But I am not sure this conclusion should follow as a matter of logic, even default logic. In this case, the inference seems to rely on a good deal of knowledge about the particular domain involved, including the vast distance between Toronto and Vancouver, which effectively rules out any sort of intermediate solution to Mary's two-body problem. By contrast, consider the happier situation of Carol, who works in College Park, Maryland, but whose spouse works in Alexandria, Virginia; and assume the same two defaults-that people tend to live in the same cities as their spouses, but also tend to live in the cities in which they work. Represented in default logic, this information would again lead to a theory with multiple extensions, in each of which, however, Carol would live either in College Park or in Alexandria. Nevertheless, I would be reluctant to accept the floating conclusion that Carol lives either in College Park or in Alexandria. Just thinking about the situation, I would consider it equally likely that Carol and her spouse live together in Washington, DC, within easy commuting distance of both their jobs. Why is it so widely thought that floating conclusions should be accepted by a skeptical reasoner, so that a system that fails to generate these conclusions is therefore incorrect? It is hard to be sure, since this point of view is generally taken as an assumption, rather than argued for, but we can speculate. Suppose an agent believes that either the statement B or the statement C holds, that B implies A, and that C also implies A. Classical logic then allows the agent to draw A as a conclusion; this is a valid principle of inference, sometimes known as the principle of constructive dilemma. The inference to a floating conclusion is in some ways similar. Suppose a default theory has two , that the extension contains the statement A, and that the extension also contains the statement A. The standard view is that a skeptical reasoner should then draw A as a conclusion, even if it is not supported by a common argument in the two extensions. Notice the difference between these two cases, though. In the first case, the classical reasoning agent believes both that B and C individually imply A, and also that either B or C holds. In the second case, we might as well suppose that the skeptical reasoner knows that A belongs to both the extensions , so that both E 1 individually imply A. The reasoner is therefore justified in drawing A as a conclusion by something like the principle of constructive dilemma-as long as it is reasonable to suppose, in addition, that either E 1 or is correct. This is the crucial assumption, which underlies the standard view of skeptical reasoning and the acceptance of floating conclusions. But is this assumption required? Is it necessary for a skeptical reasoner to assume, when a theory leads to multiple extensions, that one of those extensions must be correct? Suppose that each of the theory's multiple extensions is endorsed by some credulous reasoner. Then the assumption that one of the theory's extensions must be correct is equivalent to the assumption that one of these credulous reasoners is right. But why should a skeptical reasoner assume that some credulous reasoner, following an entirely different reasoning policy, must be right? Of course, there may be situations in which it is appropriate for a skeptical reasoner to adopt this standard view-that one of the various credulous reasoners must be right, but that it is simply unclear which one. That might be the extent of the skepticism involved. But there also seem to be situations in which a deeper form of skepticism is appropriate-where each of the multiple extensions is undermined by another to such an extent that it seems like a real possibility that all of the credulous reasoners could be wrong. The yacht, spy, and economist examples illustrate situations that might call for this deeper form of skepticism. As a policy for reasoning with conflicting defaults, the notion of skepticism was originally introduced into the field of nonmonotonic logic to characterize the particular system presented in [6], which did not involve the assumption that one of a theory's multiple extensions must be correct, and did not support floating conclusions. By now, however, the term is used almost uniformly to describe approaches that do rely on this assumption, so that the "skeptical conclusions" of a theory are generally identified as the statements supported by each of its multiple extensions, including the floating conclusions. Of course, there is nothing wrong with this usage of the term, as a technical description of the statements supported by each extension-except that it might tend to cut off avenues for research, suggesting that we now know exactly how to characterize the skeptical conclusions of a theory, so that the only issues remaining are matters concerning the efficient derivation of these conclusions. On the contrary, if we think of skepticism as the general policy of withholding judgment in the face of conflicting defaults, rather than arbitrarily favoring one default or another, there is a complex space of reasoning policies that could legitimately be described as skeptical, many of which involve focusing on the arguments that support particular conclusions, not just the conclusions themselves. Acknowledgments This paper got its start in a series of conversations with Tamara Horowitz. I am grateful for valuable comments to Aldo Antonelli, David Makinson, and Richmond Thomason, and to a number of participants in the Fifth International Symposium on Logical Formalizations of Commonsense Reasoning, particularly Leora Morgenstern, Rohit Parikh, Ray Reiter, and Mary Anne Williams. --R Entailment: The Logic of Relevance and Necessity Essentials of Artificial Intelligence. Moral dilemmas and nonmonotonic logic. Some direct theories of nonmonotonic inheritance. A skeptical theory of inheritance in nonmonotonic semantic networks. General patterns in nonmonotonic reasoning. "directly skeptical" Logics for defeasible argumentation. A logic for default reasoning. Resolving ambiguity in nonmonotonic inheritance hierarchies. The Mathematics of Inheritance Systems. A clash of intuitions: the current state of nonmonotonic multiple inheritance systems. Values and the heart's command. --TR The mathematics of inheritance systems A skeptical theory of inheritance in nonmonotonic semantic networks Floating conclusions and zombie paths Resolving ambiguity in nonmonotonic inheritance hierarchies Essentials of artificial intelligence General patterns in nonmonotonic reasoning Some direct theories of nonmonotonic inheritance Circumscription Well-founded semantics for default logic --CTR Shingo Hagiwara , Satoshi Tojo, Stable legal knowledge with regard to contradictory arguments, Proceedings of the 24th IASTED international conference on Artificial intelligence and applications, p.323-328, February 13-16, 2006, Innsbruck, Austria Pietro Baroni , Massimiliano Giacomin , Giovanni Guida, SCC-recursiveness: a general schema for argumentation semantics, Artificial Intelligence, v.168 n.1, p.162-210, October 2005 Yoshitaka Suzuki, Additive Consolidation with Maximal Change, Electronic Notes in Theoretical Computer Science (ENTCS), 165, p.177-187, November, 2006 Pietro Baroni , Massimiliano Giacomin , Giovanni Guida, Self-stabilizing defeat status computation: dealing with conflict management in multi-agent systems, Artificial Intelligence, v.165 n.2, p.187-259, July 2005

default logic;nonmonotonic logic;skeptical reasoning

506794

Anomaly Detection in Embedded Systems.

By employing fault tolerance, embedded systems can withstand both intentional and unintentional faults. Many fault-tolerance mechanisms are invoked only after a fault has been detected by whatever fault-detection mechanism is used, hence, the process of fault detection must itself be dependable if the system is expected to be fault tolerant. Many faults are detectable only indirectly as a result of performance disorders that manifest as anomalies in monitored system or sensor data. Anomaly detection, therefore, is often the primary means of providing early indications of faults. As with any other kind of detector, one seeks full coverage of the detection space with the anomaly detector being used. Even if coverage of a particular anomaly detector falls short of 100 percent, detectors can be composed to effect broader coverage, once their respective sweet spots and blind regions are known. This paper provides a framework and a fault-injection methodology for mapping an anomaly detector's effective operating space and shows that two detectors, each designed to detect the same phenomenon, may not perform similarly, even when the event to be detected is unequivocally anomalous and should be detected by either detector. Both synthetic and real-world data are used.

Introduction As computer systems become more miniaturized and more pervasive, they will be embedded in everyday devices with increasing frequency, even to the point at which domestic and industrial consumers may not be aware of their presence. Some truck tires, for example, will soon have a processor and a pressure sensor/transponder embedded in the rubber, because this is cheaper than fitting in-hub pressure sensors in the wheels of old trailers on big rigs. Some laptop-computer batteries contain an embedded computer to track the charge remaining, thereby ensuring that battery's memory travels with the battery even when it is moved to another laptop computer. Disk drives may contain one or two embedded computers (one controller, one DSP chip). Even operating systems like Unix are being embedded in television set-top boxes (enabling pausing a live television broadcast), vending machines, Internet appliances, and the International Space Station (for controlling vibration damping) [1]. Many of these devices with embedded computers will be intrinsically safety-critical or mission-critical, and therefore will require a higher level of dependability than usual - automobile and aircraft engine controllers are one example. It is presumed that fault tolerance, which is one way of achieving high dependability, will be employed in such devices. Several methods of fault tolerance require that a fault be detected before bringing fault-tolerance measures to bear on it. One salient example is recovery blocks [2] [3]. Fault detection, therefore, is an essential first step in achieving dependability. If the detector is not reliable, the fault-tolerating mechanisms will not be effective, because they will not be activated. Faults can be detected either explicitly or implicitly. When a fault is detected explicitly it is typically through pattern recognition, wherein a signature is detected that is directly linked to a particular fault. When a fault is detected implicitly, it is usually due to having detected some indirect indi- cator, such as an anomaly, that may have been caused by the fault. System- performance anomalies are often the only indicators of problems, because some faults have no stable, explicit signature; they're indicated only through unusual behaviors. One such example is the fault condition known as an Ethernet broadcast storm, which is indicated indirectly by anomalously-high packet traffic on a network [4]. Another example is an increasing error rate reported by software sensors, and observed in system event logs; disk surface failures can be indicated this way [5] [6]. In the Ethernet example, measures of packet traffic served as a sensor. In system event logs, many different measures are available [7]. As noted in [4] there can be many sensors measuring the state of a network, system or process. These sensors can be hardware or software, although recent trends have been mainly toward software sensors [8]. The data produced by such sensors are referred to as sensor data or a sensor-data stream. The data in the sensor-data stream can be numeric or categorical. Numeric data are usually continuous, are on a ratio scale, have a unique zero point, and have mathematical ordering properties (e.g., taking differences or ratios of these measures makes sense). Categorical data, sometimes referred to as nominal data, are discrete, usually consist of a series of unique labels as categories, and have no mathematical ordering properties (e.g., an apple is not twice an orange) [9]. It seems likely that as computing power increases, more of the sensor data will be in the form of categorical data [8] [10], hence anomaly detectors will be required to operate primarily on categorical data, presenting a real challenge to developers and users of such sensors, because categorical data are much more difficult to handle statistically than numeric data are. This paper focuses on detecting anomalies in categorical data. An anomaly occurring in such sensor data is often the indirect or implicit manifestation of a fault or condition somewhere in the monitored system or process. Detecting such anomalies, therefore, can be an important aspect of maintaining the integrity, reliability, safety, security and general dependability of a system or process. Since anomaly detection is on the front line of many fault-tolerance and dependability mechanisms, it is essential that it is, itself, reliable. One way to gauge its reliability is by its coverage. Coverage is a figure of merit that gauges the effectiveness of a detection or testing process. Historically, a system's coverage has been said to be the proportion of faults from which a system recovers automatically; the faults in this class are said to be covered by the recovery strategy [11]. Coverage can also be viewed as the probability that a particular class of conditions or events is detected before a system suffers consequences from a missed or false detection. Another definition of coverage, and the one used in this paper, is: a specification or enumeration of the types of conditions against which a particular detection scheme guards [12]. More succinctly, the coverage of an anomaly detector is the extent to which it detects, correctly, the events of a particular anomaly class. The motivation for the concern with coverage is that one needs to know if and when one's anomaly detection system will experience a Type I error (a true null hypothesis is incorrectly rejected) or a Type II error (a false null hypothesis fails to be rejected), so that one can take precautionary measures to compensate for such errors. If an anomaly detector does not achieve complete coverage, its suitability for use should be scrutinized carefully. Anomaly classes will be discussed in Section 6. This paper addresses the issues of how to assess the coverage of an anomaly detector, how to acquire ground-truth test data to aid in that as- sessment, how to inject anomalous events into the test data, and how to map the coverage of the detector in terms of sweet spots (regions of adequate detection) and blind regions (regions of inadequate detection). Once a de- tector's coverage map is ascertained, it can be used to judge the suitability of the detector for various situations. For example, if the environment in which the detector is deployed will never experience a condition in the detector's blind region, then the detector can be used without adverse consequences. Problem and objective A critical problem is that there is little clarity in the literature regarding the conditions under which anomaly detection works well or works poorly. To gain that clarity it is necessary to understand the details of precisely how an anomaly detector works, i.e., what the detector sees, and what phenomena affect its performance as the stream of sensor data passes through the detec- tor's range of perception. Similarly, one would want to know how a sorting algorithm views the data it is sorting, as well as which characteristics of the data, e.g., presortedness, impinge on the algorithm's efficacy. The objectives of the present work are (1) to understand the details of how an anomaly detector works, as a stream of sensor data passes through the detector's purview; and (2) to use that understanding to guide fault-injection experiments in which anomalies are injected into normal background data, the outcome of which is a map illustrating the detector's regions of sensitivity and/or blindness to anomalies. The results will address such questions as: ffl What is the coverage of an anomaly detector? ffl How does one assess that coverage? ffl Do all anomaly detectors have the same coverage, for a given set of anomalies embedded in background data? ffl Can anomaly detectors be composed to attain greater coverage than that achieved by a single anomaly detector used alone? These issues are addressed using fault injection, a well-known technique for evaluating detection systems [13] [14]. Synthetic data are used to address the usual problem of determining ground truth. To facilitate simplicity and clarity, the most basic type of anomaly detection is used, namely that of a sliding-window detector moving over a univariate stream of categorical data. Anomalous sequences are considered to be contiguous. Temporal anomalies, not addressed here, can be treated similarly if they are aggregated using an appropriate feature-extraction mechanism. 3 What is an anomaly? According to Webster's dictionary, an anomaly is something different, abnormal or peculiar; a deviation from the common rule; a pattern or trait taken to be atypical of the behavior of the phenomenon under scrutiny. This definition, fitting as it may be for everyday purposes, is too vague to be scientifically useful. A more apt definition is this: an anomaly is an event (or object) that differs from some standard or reference event, in excess of some threshold, in accordance with some similarity or distance metric on the event. The reference event is what characterizes normal behavior. The similarity metric measures the distance between normal and abnormal. The threshold establishes the minimum distance that encompasses the variation of normalcy; any event exceeding that distance is considered anomalous. The specifics of establishing the reference event, the metric and the threshold are often situation-specific and beyond the scope of this paper, although they will be addressed peripherally in Section 7.1. Determining what constitutes an anomalous element in a stream of numeric data is intuitive; a data element that exceeds, say, the mean plus three standard deviations may be considered anomalous. Determining what constitutes an anomaly in categorical data, which is the specific problem addressed here, is less intuitive, since it makes no sense to compute the mean and standard deviation (or any other numerically-based measure) of categorical values such as cat or blue, even if these categories are translated into numbers. In categorical data, anomalous events are typically defined by the probabilities of encountering particular juxtapositions of symbols or subsequences in the data stream; i.e., symbols and subsequences in an anomaly are juxtaposed in unexpected ways. Categorical data sets are comprised of sequences of symbols. The collection of unique symbols in a data set is called the alphabet. Typically, a data set will be characterized in terms of what constitutes normal behavior for the environment from which the data were drawn. The data set so characterized is called the training data. Training data may be obtained from some appli- cation, e.g., a process-control application that is providing monitored data for consumption by various analysis programs. Training data are obtained from the process over a period of time during which the process is judged to be running normally. Within these normal data, the juxtapositions of symbols and subsequences would be considered normal, provided that no faults or unusual conditions prevailed during the collection period. Once the training data are characterized (termed the training phase), characterizations of new data, monitored while the process is in an unknown state (either normal or anomalous), are compared to expectations generated by the training data. Any sufficiently unexpected juxtaposition in the new data would be judged anomalous, the possible manifestation of a fault. Anomaly causes and manifestations What causes an anomaly, and what does an anomaly look like? An example from a semiconductor fabrication process illustrates. If one attaches a sensor to an environment, such as the plasma chamber of a reactive ion etcher, the sensor data will comprise a series of normal categorical values (given a normally operating etcher). When a fault occurs in the fabrication process, the fault will be manifested as an event (a series of sensor values) embedded in otherwise normal sensor data. That event will contain one or more data values that are related to the normal data values in one of two ways: (1) the embedded event could contain symbols commonly found and commonly juxtaposed in normal data; (2) the embedded event could contain symbols and symbol juxtapositions that are anomalous with respect to those found in normal data. Thus the fault could manifest itself as an event injected into a normal stream of data, and that event could be regarded as normal or it could be regarded as anomalous. There are three phenomena that could make an event anomalous: Foreign symbols. A foreign symbol is a symbol not included in the training-set alphabet. For example, any symbol, such as a Q, not in the training-set alphabet comprising A B C D E F would be considered a foreign symbol. Detection of events containing foreign symbols, called foreign- symbol-sequence anomalies, is straightforward. Foreign n-grams/sequences. An n-gram (a set of n ordered elements) not found in the training dataset (and also not containing a foreign symbol) is considered a foreign n-gram or foreign sequence, because it is foreign to the training dataset. A foreign n-gram event contains n-grams not present in the training data. For example, given an alphabet of A B C D E F, the set of all bigrams would contain AA AB AC . FF, for a total of 6 (in general, for an alphabet of ff symbols, the total possible If the training data contained all bigrams except CC, then CC would be a foreign n-gram. Note that if a foreign symbol (essentially a foreign unigram) appears in an n-gram, that would be a foreign-symbol event, not a foreign n-gram event. In real-world, computer-based data it is quite common that not all possible n-grams are contained in the training data, partly due to the relatively high regularity with which computers operate, and partly due to the large alphabets in, for example, kernel-call streams. Rare n-grams/sequences. A rare n-gram event, also called a rare sequence, contains n-grams that are infrequent in the training data. In the example above, if the bigram AA constituted 96% of the bigrams in the sequence, and the bigrams BB and CC constituted 2% each, then BB and CC would be rare bigrams. An n-gram whose exact duplicate is found only rarely in the training dataset is called a rare n-gram. The concept of rare is determined by a user-specified threshold. A typical threshold might be .05, which means that a rare n-gram would have a frequency of occurrence in the training data of not more than 5%. The selection of this threshold is arbitrary, but should be low enough for "rare" to be meaningful. 5 The ken of an anomaly detector An anomaly detector determines the similarity, or distance, between some standard event and the possibly-anomalous events in its purview; it can't make decisions about things it can't see. The purview of a sliding-window detector is the length of the window. Since not all anomalies are the same size as the detector window, and such size differentials can affect what the A A A A A A d d d d d d d d d d Background Whole Internal Encompassing Boundary Condition Boundary Condition Figure 1: The ken of an anomaly detector: different views of an anomaly (depicted by AAAAAA) embedded in a sensor-data stream (depicted by ddddd) from the perspective of a sliding-window anomaly detector. Arrows indicate direction of data flow. detector detects, it is useful to pursue the idea of a detector's ken, or range. The word ken means the extent or range of one's recognition, comprehen- sion, perception or understanding; one's horizon or purview. Thus it seems appropriate to ask, what is the ken of an anomaly detector? The univariate case is shown in Figure 1 which depicts a stream of sensor data (ddddd) into which an anomalous event (AAAAAA) has been injected. The right- directed arrows indicate that the data are moving to the right with respect to the detector window, as time and events pass. The width of the window through which the detector apprehends the anomaly can take on any value, typically based on the constraints of the environment or situation in which the detector is being used. The extent to which the detector window overlaps the anomaly can be thought of as the detector's view of the anomaly. It is natural to focus on the case in which the window is the same size as the anomaly and the entire anomaly is captured exactly within the window. This is called the whole view. There are, however, a number of other cases, illustrated in the figure. When the size of the detector window is less than the length of the anomaly, the detector has what is called an internal view. For the case in which the detector window is larger than the anomaly, both anomalous and normal background data are seen - this is the encompassing view. Irrespective of the width of the window, as time passes and an anomalous event moves through the window, the event presents itself to the detector in different perspectives. Of particular interest are situations termed external boundary conditions, used interchangeably here with the term boundary conditions. These arise at both ends of an injected sequence embedded in normal data, when the leading or trailing element of the anomaly abuts the normal data. Boundary conditions occur independently of the relative sizes of the detector window and anomaly (except in the degenerate case of size one). In a boundary condition, the detector sees part of the anomaly and part of the background data. The background view sees only background data, and no anomalies. It will be shown later that the detector views and conditions just discussed will be important in determining precisely what an anomaly detector is capable of detecting, as well as what may cause an anomaly detector to raise an alarm, even when it should not. Note that these conditions depend on the size of the injected event relative to the size of the detector window. Table 1 summarizes the conditions. Conditions Internal x Boundary x x x Encompassing x Table 1: Conditions of interest that ensue with respect to detector-window size (DW) and anomaly size (AS). 6 Anomaly space It is important to note that anomalies can be composed of subsequences of various types, three of which were identified in Section IV: foreign symbols, foreign n-grams and rare n-grams. A fourth type is a common n-gram, an n-gram that appears commonly (not rarely) in the normal data. Henceforth the terms n-gram and sequence will be used interchangeably, i.e., foreign n-gram and foreign sequence refer to the same thing. That an anomalous sequence can be composed of several different kinds of subsequences, along with the concept of internal sequences and boundary sequences, gives rise to the idea of creating a map of the anomaly space for sliding-window detectors. Given such a map, it should be possible to determine the extent to which that map is covered by a particular anomaly detector. It is not the goal of this paper to do that, but rather to show that two detectors can have unexpectedly different coverages, even when encountering the same events embedded in the same data; that is, different detection capabilities will arise from the use of different metrics and different detectors. An anomaly-space map is shown in Figure 2. The map is described in the figure caption and in the paragraph following it. The window size of the detector, relative to the size of the anomaly, is shown in the three columns of the figure: detector window size less than anomaly size, detector window size equal to the anomaly size, and detector window size greater than the anomaly size. For each of these conditions the figure addresses three kinds of anomalies: foreign-symbol-sequence anomalies (sequences comprising only foreign symbols); foreign-sequence anomalies (sequences comprising only foreign sequences); and rare-sequence anomalies (sequences comprising only rare sequences). The following material expands the description of a cell by selecting as an example the anomaly type FF AI AB, depicted at the upper left of the figure. This is a sequence of foreign symbols (FF) composed of alien internal sequences (AI) and having alien external boundaries (AB). The term alien is an umbrella term used to refer to sequences that do not exist in the normal (training) data, irrespective of the characteristics that make them foreign, unlike the more closely-defined terms foreign symbol and foreign sequence. FF is a foreign-symbol-sequence anomaly composed only of foreign sym- bols. In this specific case, when the anomalous sequence FF AI AB slides past a detector window whose size is less than the size of the anomaly, the detector will first encounter the leading edge of the anomaly. That leading edge will be alien, i.e., the sequence containing the first element of the anomaly and the normal element immediately preceding it is not a sequence that exists in the normal (training) data, and therefore will be anomalous. As the anomaly moves through the detector window, each internal, detector- window-sized subsequence of the anomaly will be alien. As the anomaly Foreign-Symbol-Sequence Anomalies FF AI AB FF - AB FF AE AB Foreign-Sequence Anomalies FS AI AB FS AE AB FS RI AB FS - AB FS RE AB FS CI AB FS CE AB FS AI RB FS AE RB FS RI RB FS - RB FS RE RB FS CI RB FS CE RB FS AI CB FS AE CB FS RI CB FS - CB FS RE CB FS CI CB FS CE CB Rare-Sequence Anomalies RS AI AB RS AE AB RS RI AB RS - AB RS RE AB RS CI AB RS CE AB RS AI RB RS AE RB RS RI RB RS - RB RS RE RB RS CI RB RS CE RB RS AI CB RS AE CB RS RI CB RS - CB RS RE CB RS CI CB RS CE CB Figure 2: Anomaly space. The first two letters in each cell identify the type of anomalous sequence (FS: foreign-sequence anomaly; RS: rare-sequence foreign-symbol-sequence anomaly). The next two letters identify the type of condition (internal (alien, rare or common) or encompassing (alien, rare or common)); the last two letters refer to the boundary conditions (alien, rare or common). DW ! AS indicates detector window smaller than anomaly analogously indicates window larger; when AS there are no internal or encompassing conditions, indicated by dashes replacing the middle two letters. Impossible conditions are struck out. elements comprising anomalous sequence. length of anomalous sequence A A A A A A Sliding a size 6 detectortecto Size of detector window: 6 marks the elements of the 26556 -3 data that have been incorporated over an injected anomaly. into the sequences comprising the external boundary conditions. A: marks the elements of thee 26310 anomalous sequence. Size of foreign-symbol sequence injected: 6 elements comprising external boundary conditions. Figure 3: Foreign-symbol-sequence anomaly injected into background data. External boundary conditions are shown for detector window size of 6 and anomaly size of passes out of the window, its trailing edge will form another alien boundary. Figure 3 illustrates a sliding-window detector moving over an anomaly injected (synthetically) into a data stream. The detector window and the anomaly size are the same: 6. The shaded boxes depict the injected FF anomaly, which raises an alarm as the detector window is positioned exactly over it. Ten sequences result from the interaction between the background and the injected anomaly. These ten sequences comprise the boundary conditions that may affect the response of the detector as its window slides over the injected anomaly, depending on whether or not additional anomalies are caused by the anomaly-background boundary interac- tions. The composition of an injected anomaly, as well as the position of the injection in the background data, must be carefully controlled to avoid the creation of additional anomalies at the boundaries of the injection; Section 7.4 provides details. Note that the anomalies depicted in Figure 2 reflect the restricted needs of an experimental regime, and do not express all of the conditions that might be encountered in a real-world environment. In the FSRIRB anomaly, for example, all of the subsequences comprising the internal condition are rare, and all the subsequences that make up the external condition are rare. In the real world, the subsequences comprising these conditions might be a mixture of rare, foreign and common. The anomaly space is constructed as it is in order to effect experimental control and to reduce confounding in which a detector's response cannot be attributed to any single phenomenon, but rather is due to the interaction of several phenomena. 7 Mapping the detection regions Different detectors may cover different parts of the anomaly space depicted in Figure 2. This section describes an experiment showing how well a selected portion of the anomaly space is covered by two different detectors, Markov and Stide, whose detection mechanisms will be explained below. The selected portion of the space is foreign-sequence anomalies as shown in the fifth row of the foreign-sequence section of the figure: FS RI RB, FS - RB, and FS RE RB. Because the last of these is not possible, focus is limited to the first two. These cells were chosen because they contain events that would unequivocally be termed anomalous by both of the detectors used in the experiment: both anomalies are foreign sequences with rare boundary conditions. For the case in which the detector window size is less than the anomaly size, the subsequences that make up the internal conditions are all rare, hence FS RI RB. For the case in which the detector window size is equal to the anomaly size, no internal conditions will be extant, hence FS - RB. The sequence FS RE RB is not possible, because the sequences that make up the encompassing condition will contain the foreign sequence FS. Sequences that contain foreign subsequences will themselves be foreign sequences; consequently it is not possible to have a rare sequence that contains a foreign subsequence. The following subsections describe the detectors used in the coverage- mapping experiment, the methods for generating the data used to test detector coverage (background data, anomaly data and anomaly-injected testing data), and the regime for running the detectors in the experiment. 7.1 Detectors To illustrate that different detectors may cover different parts of the anomaly space, two detectors were tested: Markov and Stide. Each of these is described below. 7.1.1 Markov detector Most engineered processes, including ones used by or being driven by com- puters, consist of a series of events or states. While the process is running, the state of the process will change from time to time, typically in an orderly fashion that is dictated by some aspect of the process itself. Certain kinds of orderly behavior are plausible approximations to problems in real-world anomaly detection and, moreover, they facilitate rigorous statistical treatment and conclusions. The anomaly-detection work in this paper focuses on the kind of orderly behavior that corresponds to Markov models. The Markov anomaly detector determines whether the states (events) in a sequential data stream, taken from a monitored process, are normal or anomalous. It calculates the probabilities of transitions between events in a training set, and uses these probabilities to assess the transitions between events in a testing set. These states and probabilities can be described by a Markov model. The key aspect of a Markov model is that the future state of the modeled process depends only on the current state, and not on any previous states [15, 16]. A Markov model consists of a collection of all possible states and a set of probabilities associated with transitioning from one state to another. A graphical depiction of a Markov model with four states is shown in Figure 4 in which the states are labeled with the letters A, B, C and D. Although the arcs are not explicitly labeled in the figure, they can be thought of as being labeled with the probabilities of transitioning from one state to another, e.g., from state A to state B. The transition probabilities can be written in a transition matrix, as shown in Figure 5, in which the letters indicate states and the numbers indicate transition probabilities. The probability of transitioning from D to A, for example, is 1; from D to any other state is 0. The transition probabilities are based on a key property, called the Markov assumption, and can be written formally as follows. If X t is the state of a system at time t, then: Figure 4: Four-state Markov model alphabetcomprisedoffoursymbols. Letters indicate states; arrows indicate transition probabilities. A transition can be made from any state to any other state, with a given probability. Hence the probability of being in state X t+1 = y at time t+1 depends only on the immediately preceding state X t = x at time t, and not on any previous state leading to the state at time t. Therefore the transition probability, P xy , denoting the progression of the system from state x to state y, can be defined as: Readers interested in further details are encouraged to consult the large literature on Markov models, e.g., [15]. Weather prediction provides a nice illustration of a Markov process. In general, one can usually predict tomorrow's weather based on how the weather is today. Over short time periods, tomorrow's noontime temperature depends only on today's noontime temperature. The previous day's temperature is correlated, but provides no additional information beyond that contained in today's measurement. So, there is some reasonably high probability that tomorrow will be like today. If tomorrow's weather is not like today's, then one's expectations would be violated, and one would feel surprised. The degree of surprise can be used in anomaly detection: the more surprised one is to observe a certain event or outcome, the more anomalous is the event, and the more it draws one's attention. A 0.00 1.00 0.00 0.00 Transition sequence: ABCDABCD. Figure 5: Transition matrix for four-state Markov model (alphabet comprised of four symbols: A, B, C and D). Letters indicate states; numbers indicate probabilities of transitioning from one state to another. Example: probability of transitioning from D to A is 1; probability of transitioning from D to any other state is 0. Basing an anomaly detector on a discrete Markov process requires three steps. First, a state transition matrix is constructed, using the training data; the training data represent the conditions that are considered to be normal. For example, if sensor data are collected from an aircraft turbine that is running under normal operating conditions, these data would be used as training data. From these data would be constructed the state transition matrix that represents normal turbine behavior. The second step is to establish a metric for surprise. This is generally a distance (or similarity) measure that determines how dissimilar from normal a process can be, while remaining within the bounds of what is considered to be normal operating behavior. If the threshold of dissimilarity is exceeded, then the observed behavior, as reflected in the sensor data, is judged to be abnormal, or anomalous. In the case of the Markov-model approach, if a transition is judged to be highly probable (e.g., has a probability of 0.9), then its surprise factor is 0:1. If, in this example, the threshold were set at 0:9, then a surprise factor of 0:1 would not be anomalous. If the surprise factor had been 0:98, for example, then the transition would have been considered anomalous. The more the surprise factor exceeds the surprise threshold, the more anomalous the event will seem. Given a threshold of 0:9, a surprise factor of 0:91 would be deemed anomalous, but a surprise factor of 0:99 would be regarded as fractionally more anomalous. The third step is to examine the test data to see if they fall within the expectations established by the training data. As each state transition in the test data is observed, its probability of occurring in normal data is retrieved from the transition matrix derived from the training data. If the transition under scrutiny has a surprise factor that exceeds the surprise threshold, then the event in the testing data is considered anomalous. The Markov-based anomaly detector that is used in this paper is based on the ideas presented in this section. The states in the model do not necessarily need to correspond to single events or unigrams; a state can be composed of a series of events, too. In a case where multiple events comprise a state, the collection of states in the Markov model spans the combinations of unigrams of a specified length as present in the training data. Consider, for example, the sequence A B C D E F. A 3-element window is moved through the sequence one event at a time: the first window position would contain A B C; the second window position would contain B C D, and so forth. In using a Markov model to assess the surprise factor of the transition between the first and second windows, the states would comprise a series of three events or unigrams, i.e., equivalent to the window size. Notice that in the transition from the first window to the second window, the event A is eliminated, and the event D is added. Since the events B C are common to both windows, it is the addition of event D which drives the surprise factor. Therefore, the resulting surprise factor reflects on the event D following the series of events A B C. A formal description of the training and testing processes used in conjunction with such a Markov model is given next. Markov training stage Primitives similar to those described in [17] are defined to facilitate the description of the training procedure. Let \Sigma denote the set of unique elements (i.e., the alphabet of symbols, or the set of states) in a sequential stream of data. A state in a Markov model is denoted by s and is associated with a sequence (window) of length N over the set \Sigma. A transition is a pair of states, (s; s 0 ) that denotes a transition from state s to state s 0 . The primitive operation shift(oe; z) shifts a sequence oe left by one, and appends the element z, where z 2 \Sigma, to the end. For instance, if the sequence which is equal to the new sequence bcz. The primitive operation next(oe) returns the first symbol of the sequence oe, then left shifts oe by one to the next symbol. This function is analogous to popping the first element from the top of a stack, where the top of the stack is the beginning of the sequence. For example, given a sequence abcde, next(abcde) returns a and updates the sequence to bcde. The construction of the Markov model for normal behavior based on training data can be described as follows: Initialize: ffl current state = first N elements of training data and, training-data stream minus first N elements. Until all the sequences of size N have been scanned from the training data: 1. Let next(oe). 2. Set next state to shift(current state; c). 3. Increment counter for the state current state and for the transition (current state; next state). 4. Set current state to be next state. After the entire stream of training data has been processed, the probability of the transition is computed as (s; s F are the counters associated with the transition (s; s 0 ) and s respectively. Markov testing stage Let 0.00 indicate normal, and let 1.00 indicate anomalous. The surprise factor (sometimes called an anomaly signal) can be calculated from test data as follows: Initialize: ffl current state = first N elements of training data and, training-data stream minus first N elements. Until all the sequences of size N have been scanned from the test data: 1. Let next(oe). 2. Surprise minus the transition probability of (current state; next state). 3. Set current state to be next state. 7.1.2 Stide detector Stide is a sequence, time-delay, embedding anomaly detector inspired by natural immune systems that distinguish self (normal) from nonself (anomalous) [18] [19]. The reference to "time" recognizes the time-series nature of the categorical data on which the detector is typically deployed. Stide has been applied to streams of system kernel-call data in which the manifestations of maliciously altered code are regarded as anomalies [20]. Stide mimics natural immune systems by constructing templates of "self" and then matching them against instances of "nonself." It achieves this in several stages. Stide training stage A database consisting of templates of "self" is constructed from a stream of data considered to be normal (self); these are the training data. The stream is broken into contiguous, n-element, overlapping subsequences, or n-grams. The value of n is typically determined empirically [21]. Duplicate n-grams are removed from the collection, leaving only the unique ones. These unique n-grams are stored for future fast access. This completes the training stage. Stide testing stage Stide compares n-grams from an unknown dataset (testing data) to each of its unique "self" n-grams. Any unknown n-gram that does not match a "self" n-gram is termed a mismatch. Finally, a score is calculated on the basis of the number of n-gram comparisons made within a temporally localized region (termed "locality frame") [21]. Each comparison in step two (testing) receives a score of either zero or one. If the comparison is an exact match, the score is zero; if the comparison is not a match, the score is one. These scores are summed within a local region to obtain an anomaly signal. An example illustrates. Within a local region of 20 comparisons made between "self" and unknown, if all 20 are mismatches, the score will be 20 for that particular region; if only 8 are mismatches, the score will be 8 for that region. There are many overlapping regions of 20 in any given data stream. Stide calculates which of these 20- element regions has the highest score, and concludes that that region is the locus of the anomaly. The Stide algorithm can be described formally as follows. Let N be the length of a sequence. The similarity between the sequence and the sequence defined by the function: The expression above states that the function Sim(X;Y ) returns 0 if two sequences of the same length are element-by-element identical; otherwise the function returns 1. Each sequence of size N in the test data is compared to every sequence of size N in the normal database. Let Norm be the number of sequences of size N in the normal database. Given the set of sequences in the normal database, for the ordered set of sequences fX in the test data, where X where Z is the number of elements in the data sample, the final similarity measure assigned the sequence X s is The expression above states that when a sequence, X s , from the test data is compared against all sequences in the normal database, the function Sim f (X s ) returns 1 if no identical sequence can be found (i.e., a mismatch); otherwise the function returns 0 to indicate the presence of an identical sequence (a match) in the normal database. The locality frame count (LFC) for each size N sequence in the test data is described as follows. Let L be the size of the locality frame and let Z be the number of elements in a data sample. For the ordered set of sequences the LFC can be described by: 7.2 Constructing the synthetic training data The training data serve as the "normal" data into which anomalous events are injected. The requirements for the training data are that a large proportion of the data be comprised of common sequences, that they contain a small proportion of rare sequences, and that there is a relatively high predictability from one symbol to another. The common sequences are required to facilitate the creation of background test data that will contain no noise in the form of naturally-occurring rare or foreign sequences. This is necessary so that a detector's response to the injected anomaly can be observed without confounding by such phenomena. The rare sequences in the training data are needed so that anomalous events composed of rare sequences can be drawn from the normal training data, and then injected into the test data; see Section 7.3 below for details. Finally, a modicum of predictability is convenient for emulating certain classes of real-world data (e.g., system kernel calls) for which detectors like Stide are said to be well suited [22]. The alphabet has eight symbols: A B C D E F G and H. A larger alphabet could have been used, but it would not have demonstrated anything that an 8-symbol alphabet could not demonstrate; increasing the alphabet size would not change the outcome. Moreover, substantially more computation time is required as the alphabet size goes up. It is noted that alphabet sizes in real-world data are typically much larger than 8; for example, the number of unique kernel-call commands in BSM [23] audit data exceeds 200. However, the current goal is to evaluate a detector in terms of its ability to detect anomalies as higher-level abstract concepts, and while alphabet size does influence the size of the set of foreign sequences and the set of possible sequences that populate the normal dataset, foreign sequences and rare sequences retain their character irrespective of alphabet size. Maintaining a relatively small alphabet size facilitates a more manageable experiment, yet permits direct study of detector response. To accommodate the requirements for predictability and data content, the training data were generated from an eight-by-eight state transition matrix with probability 0.9672 in one cell of each row, and 0.004686 in every other cell, resulting in a sequence of conditional entropy 0.1 (see [24] for details). One million data elements (symbols) were generated so that there would be a sufficient variety of rare sequences in the sample to use them in the construction of anomalous sequences for the test data. Ninety-eight percent of the training data consisted of repetitions of the sequence A B C D E F G H, seeding the data set with common sequences. This is the data set used to train the two detectors used in this study, i.e., to establish a model of normalcy against which unknown data can be compared. 7.3 Constructing the synthetic test data Test data, containing injected anomalies, are used to determine how well the detector can capture anomalous events and correctly reject events that are not anomalous. The test data consist of two components: a background, into which anomalies are injected, and the anomalies themselves. Each is generated separately, after which the anomalies are injected into the background under strict experimental control. The background consisted of repeated sequences of A B C D E F G H, the most common sequence in the training data. This was done so that the test data would not conflict with the training data, i.e., would not contain spurious rare or foreign sequences. 7.4 Constructing the anomalous injections Once the background data are available, anomalous events must be injected into them to finalize the test data. The anomalies must be chosen carefully so that when they are injected into the test data they do not introduce unintended anomalous perturbations, such as external boundary conditions. If this were to happen, then a detector could react to those conditions, confounding the outcomes of interest. Hence, scrupulous control is necessary. The goal is to map the detection capability of both the Stide and the Markov anomaly detectors, and to show that their detection capabilities may vary with respect to identical and unequivocally anomalous phenomena. Given this objective, a single anomaly type that both detectors must be able to detect is selected from the anomaly space in Figure 2 for the experiments. The anomaly type selected is a foreign sequence of length AS for which all subsequences of length less than AS that make up the internal sequences and the boundary sequences are rare. Rare is defined to be any sequence of detector-window length that occurs in the training data less than one percent of the time. It is within the scope of this study to map out only one region or type in the anomaly space in order to illustrate what can be learned and gained by the effort. Once that anomaly type is determined, e.g., FS RI RB, as described in Section 7, the next step is to inject a foreign sequence composed of rare sequences into the test data. A catalog of rare n-grams is obtained from the training data. Rare n-grams are drawn from the catalog, and composed to form a foreign sequence of the appropriate size. For example, the bigrams BA, AF, FH, HE, EC, CC and CF each occurred less than 0.06% of the time in the training data; consequently, these are rare bigrams. Combining these seven bigrams produces one octagram (BAFHECCF) whose internal sequences are made up of rare sequences of size two. This octagram was injected into the background data. Once the composed foreign sequence is injected into the background data, its boundary conditions must be checked to ensure that they are all rare (be- cause rare boundary conditions are consistent with the anomaly class being examined). If the boundary conditions are satisfied, the procedure is finished; if not, an attempt is made to handcraft the boundary conditions with the help of semiautomated tools. If the handcrafting fails, then a different set of rare sequences is selected from the catalog, and a new foreign sequence is com- posed. The new foreign sequence is injected into the background data, and its boundary conditions are checked. This entire procedure is repeated until a successful injection is obtained. Note that when the size of the detector window is greater than the size of the injection, an encompassing condition ensues, not an internal condition; however, care is still required to ensure that the external boundary conditions remain pertinent, even though the focus has been moved from internal conditions to encompassing conditions. Eight injection sizes and fourteen detector-window sizes were tested. The procedure outlined above for creating the anomalous events, and for injecting them, is repeated for each combination of injection size and window size, resulting in 112 total data sets. 7.5 Scoring detector performance Anomaly detectors are capable of only two kinds of decisions: yes, an anomaly exists; or, no, an anomaly does not exist. Detectors are usually judged in terms of hits, misses and false alarms. A hit occurs when the detector determines that an anomaly is present, and an anomaly actually is present. A miss occurs when the detector determines that no anomaly is present when actually there is one present. A false alarm occurs when the detector determines that there is an anomaly present when in fact there is no anomaly present. A perfect detector would have a 100% hit rate, no misses and no false alarms. To effect proper scoring, ground truth (a statement of undisputed fact regarding the test data) must be known. That is, it must be determined exactly where anomalies have been injected into the test data, so that when the detector issues a decision, the correctness of that decision can always be assured. The injector creates a key that indicates the exact location of every injection. Using this key, one can determine whether or not a detector's decisions are correct. The usual procedure for this is for the detector to create an output file containing its decisions: 0 for no, and 1 for yes. This file can be compared against the key, which contains a similar set of zeroes and ones. If the two files match perfectly, the detector's performance is a perfect 100%; otherwise, the percent of hits, misses and false alarms can be calculated easily. Using the injection key file, however, is not as straightforward as it might first appear. Due to the interaction of the detector-window size and the particular composition of the injected event, the detector's responses may not always be aligned perfectly with the ones and zeroes in the key file. For example, if the key file contains a one at the leading edge of an injected event (which seems sensible), the detector's response might not match. In fact, the detector might make a variety of responses, some of them multiple, to an injected event, and the key file must be constructed to facilitate correct scoring for any set of responses the detector might make. For example, the detector might decide, incorrectly, that an anomaly exists at the leading-edge boundary of an injected event; in fact, at that boundary there is no anomaly, so the detector's decision would be wrong. Another example is that the detector, depending on its window size relative to the size of the injected event, may respond only to subsequences internal to the injected event, but not to the event in toto. There is a danger that a detector will respond several times to a single injected event (because it may encounter several different views of that event), and thereby be judged mistakenly to have false-alarmed - to have decided yes in error. The problems with matching detector decisions against ground-truth keys can be addressed in a variety of ways. In the present work the primary concern is to determine whether or not a detector is completely blind to an injected event, so the most worrisome response would be no response at all - a miss. If the detector does not decide yes to any part of the injected event, whether internal, whole, encompassing or boundary, then it is blind to the event. Alternatively, if the detector does respond, there needs to be a way to mitigate the problem of key mismatch, as described above. This problem is addressed by requiring that the detector respond positively at least once within the span of the injected event and the elements comprising its boundary conditions (collectively called the incident span) in order to have judged the event to be a hit. 7.6 Procedure Each of the two detectors, Markov and Stide, was provided with the same set of training data. From the training data, the detectors learned their models of normal behavior. Then each detector was tested, using each of the 112 test data sets described in Section 7.4. The size of the detector window was varied from two to fifteen, and the size of the injected events was varied from two to nine. The restrictions on these dimensions were due to resource limitations, since computation time and memory increase with an increase in either dimension. Moreover, nothing new would be learned from raising either parameter. Each detector, for each testing session, produced a decision file which was compared to the key files for each test data set. Comparisons that revealed detector blindness (no detection of injected anomalous events) were charted. Note that Stide's locality-frame-count feature was ignored, because it operates only as an amplifier for detected anomalies (which Stide represents as mismatches). If a foreign-sequence anomaly is not detected by a mismatch, then applying the locality frame count will not make the anomaly visible. Since the task at hand is to determine whether or not the injected anomalies were detected, amplification, which can be viewed as a post-detection process, is not relevant for mismatch/anomaly detection under present experimental conditions. As a consequence of this, Stide's maximum anomalous response is 1. Detection and blind regions for the Markov and Stide detectors are depicted in Figures 6 and 7 respectively. These decision maps illustrate the detection capability of both the Markov and the Stide detector with respect to an injected foreign sequence composed of rare sequences. The x-axis of each map marks the increasing size of the foreign sequence injected into the test data; the y-axis marks the size of the detector window required to detect a foreign sequence of a specified size. Each star indicates successful detection of the foreign-sequence anomaly whose size is indicated on the x-axis, using a detector window whose size is marked on the x-axis; detection specifically means that at least one positive response occurred in the incident span, where "positive" connotes the most-anomalous value possible in the detector's range of responses. In both detectors, the most anomalous response is 1. The areas that are bereft of stars indicate either detection blindness or an undefined region. Detection blindness means that the detector was unable to detect the injected foreign sequence whose corresponding size is marked on the x-axis, i.e., the maximum anomalous response recorded along the entire incident span was 0 - to the detector, the anomaly appears as being completely normal. Note that no false alarms occurred, because background data were constructed from common sequences which do not raise alarms. Size of foreign-sequence anomaly Size of detector window3579111315Undefined region Detection region Figure Markov sweet and blind regions. The undefined region is an artifact of the Markov detector and anomaly type. Since the Markov detector is based on the Markov assumption, i.e., the next state is dependent only upon the current state, the smallest possible window size is 2, or a bigram. This means that the next expected, single, categorical element is dependent only on the current, single, categorical el- ement. As a result, the y-axis marking the detector window sizes in Figure 6 begins at 2. Although it is possible to run Stide on a window size of 1, doing so would produce results that do not include sequential ordering of Size of foreign-sequence anomaly Size of detector window3579111315Blind region Undefined region Detection region Figure 7: Stide sweet and blind regions. events, a property that comes into play in all window sizes larger than 1. This, together with the fact that there is no equivalent window size of 1 on the side of the Markov detector, argued against running Stide with a window of 1. The y-axis also begins at 2, because the type of anomalous event on which the detectors are being evaluated requires that a foreign sequence be composed of rare sequences. A foreign sequence of size one is an event that contains a single element that must be both foreign and rare at the same time; this is not possible. As a result, both Figures 6 and 7 show an undefined region corresponding to a detector window of size two and an anomaly of size one. By charting the performance spaces of Stide and the Markov-based detector with respect to a foreign sequence composed of rare sequences, one is able to observe the nature of the gain achieved by employing the conditional probabilities of the Markov detector (that are absent in Stide). The significant gain in detection capability endowed by the use of conditional probabilities is illustrated by the blind region depicted in Figure 7. It is interesting to note that, for the exact same datasets, using a detector window of length 9, Stide's detection coverage is just 56% of Markov's. As the detector window size decreases to 2, Stide's coverage decreases to only 12.5% of that of the Markov detector. At this window size, however, the Markov detector still has 100% coverage of the space, which is a tremendous difference. The results show that although the Markov and Stide detectors each use the concept of a sliding window, and are both expected to be able to detect foreign sequences, their differing similarity metrics significantly impact their detection capabilities. In the case of Stide, even though there is a foreign sequence present in the data stream, it is visible only if the size of the detector window is at least as large as the foreign sequence composed of rare subsequences - a requirement that the Markov detector does not have. Therefore, even if a fault does manifest as a foreign sequence in the data, it doesn't necessarily mean that Stide, which claims to be able to detect "unusual" sequences, will detect such a manifestation. It should be noted, therefore, that the selection of a similarity metric can have critical effects on the performance of a detector, and these choices should be made with care and with understanding of the metric. 9 Real-world data The results shown in previous sections were based on synthetic data that were generated specifically to test the different anomaly detectors described. This section provides a link to real-world data, and shows that the manifestations of live anomalies in system kernel-call data are consistent with the anomaly- space map of Figure 2. The live experiment consisted of a cyberattack on a RedHat 6.2 Linux system. The attack exploited a vulnerability in glibc (standard C library) through the su program (a program that switches a user from one account to another, given that the user provides appropriate credentials). The library allows a user to write an arbitrary format string. The exploiting program writes a carefully crafted string which interrupts the execution of su, allowing the user to run a script or program with root privileges. The exploit permits the user to switch accounts without providing credentials. Running su with and without the exploit should produce kernel-call data with and without anomalies due to the exploit itself. Kernel-call data on the victim machine were logged using the IMMSEC kernel patch, provided by the Computer Immune Systems Research Group at the University of New Mexico. The attack was scripted so that it could be repeated reliably and auto- matically. The following procedure was run three times, using standard su (with normal user interaction, providing credentials to switch from user to root) to obtain normal (training) data, and run three more times using the su exploit to obtain the anomalous (test) data: 1. Start a session as a regular user. 2. Turn on syscall logging for the su program. 3. Run the exploit as the regular user; verify that it was successful in giving the user a shell with root privileges. 4. Turn off syscall logging for the su program; move the log file to a permanent location. 5. Clean up the environment and log out. The monitored kernel-call data were examined by both an experienced system programmer and a set of automated tools to find all the minimal foreign sequences that appeared in the attack data, but not in the normal data. A minimal foreign sequence is a foreign sequence in which no shorter foreign sequence is embedded. The programmer and tool results were compared and found to be mutually consistent. The system programmer confirmed, through systematic analysis, that all of the foreign sequences were direct manifestations of the attacks. Seventeen foreign-sequence anomalies were discovered in the su exploit; no foreign symbols were found, and there was no variability in the kernel-call data for the three attacks. The foreign-sequence anomalies ranged in length from 2 to 5, with one anomaly of length 5, five anomalies of length 3, and eleven anomalies of length 2. Some anomalies were unique; others were repeated in the data. Details are shown in Figure 8. When using a detector window of size 2, which is the smallest possible size that covers an anomaly of length two, the real-world foreign-sequence anomalies in Figure 8 had compositional characteristics like the ones shown in the anomaly space in Figure 2. All of the anomaly descriptions are the same as the ones described in the anomaly space, except for the second one, which is eight characters instead of six or four, like the rest. The anomaly (FS RI RB FB) shown in the figure is from live data, and its composition reflects Anomaly Size Anomaly Contents Anomaly Description Figure 8: Foreign-sequence anomalies, discovered in real-world information- warfare attack data, showing the size of each of the 17 anomalies, the events comprising each anomaly, and the anomaly description in accordance with the anomaly space of Figure 2. Anomaly contents are numerical encodings of kernel calls. the broader set of conditions pertaining to uncontrolled, real-world data, as opposed to the more compact formulations in the anomaly space which are for well-behaved anomalies like the ones found in the synthetic data. The FS, as usual, indicates the base type of the anomaly: foreign sequence. The RI indicates that all of the internal conditions are rare. The RB indicates that all of the sequences comprising the left boundary are rare, and the AB indicates that all the sequences comprising the right-boundary are alien. Discussion and conclusion This paper has addressed fundamental issues in anomaly detection. The results are applicable in any domain in which anomaly detection in categorical data is conducted. The paper has shown how to assess the coverage of an anomaly detec- tor, and has also illustrated many subtleties involved in doing so. There are myriad factors to be considered carefully; the process is not straightfor- ward. Meticulous attention needs to be paid to the interactions between an anomalous event and the normal environment in which it is embedded, i.e., external boundary conditions, internal conditions, encompassing conditions, and common, rare and foreign sequences that compose an anomalous event. Unless all of these factors are accounted for, error may be the biggest enemy of a correct mapping. The coverage maps for two different anomaly detectors were shown to be strikingly different. This might come as a surprise to someone who believes that applying any anomaly detector to a stream of sensor data would be satisfactory, or that either of two detectors would detect the same events. One detector, Stide, which was specifically designed to detect foreign sequences, was shown to be blind to over half of a region it purports to cover. When used in its original role as a detector for information-warfare intrusions, Stide has been operated in a region of the detection space that is about six by six in terms of window size vs. anomaly size. It is interesting that in that region Stide is blind to 36% of the space, whereas the Markov detector covers 100% of that same region. It is not necessarily bad for an anomaly detector to have less than perfect coverage, as long as the user knows the limitations of the detector. If a detector has suboptimal coverage, it may be possible to deploy the detector in situations where it doesn't need to operate in the part of the space in which it is blind. It will never be possible to assure this, however, if the detector's coverage is not mapped. Can multiple anomaly detectors be composed to attain greater coverage than that achieved by a single anomaly detector used alone? It seems clear from the two maps produced here that one detector can be deployed to compensate for the deficiencies of another. In the present case it may appear that the Markov detector should simply replace Stide altogether, but because each detector has a different operational overhead, it may not be straightforward to determine the best mix for a compositional detection sys- tem. Also, one should be reminded that in the present work only one cell of the anomaly space depicted in Figure 2 has been examined; determination of overall coverage awaits examination of the rest of the cells as well. When deploying anomaly detectors in embedded or mission-critical systems, it is essential to understand the precise capabilities of the detectors, as well as the characteristics of the spaces in which they will operate. Although the real-world experiment with live systems and data was limited in scope, it still illustrates two important things. First, the anomaly types depicted in Figure 2 were demonstrated to exist in real-world data; they are not mere artifacts of a contrived environment. Second, the response of a detector to a specified type of anomaly will not change, whether the anomaly is found in synthetic data or in real-world data; consequently, the results obtained from having evaluated an anomaly detector on synthetic data will be preserved faithfully when applied to real data; that is, predictions made with synthetic data will be sustained when transferred to real-world environments. Some important lessons have been learned. A blind region in an anomaly- space map will always grow as the foreign sequence grows. This means that longer foreign sequences may constitute vulnerabilities for the detection algorithms considered here. Nevertheless, it is undoubtedly better to know the performance boundaries of a detector so that compensations can be made for whatever its weaknesses may be. Synthetic data have been effective in mapping anomaly spaces. Synthetic data may be the only avenue for creating such maps, because they permit running experiments in which all confounding conditions can be controlled, allowing absolute calibration of ground truth. Although real-world data is appealing for testing detection systems, real-world ground truth will always be difficult to obtain, and not all of the desired conditions for testing will occur in real data in a timely way. Finally, and most importantly, the anomaly-space framework provides a mechanism that bridges the gap between the synthetic and real worlds, allowing evaluation results to transfer to any domain through the anomaly- space abstraction. Acknowledgements The work herein was supported by the U.S. Defense Advanced Research Projects Agency (DARPA) under contracts F30602-99-2-0537 and F30602- 00-2-0528. Many other people contributed in various ways; the authors are grateful to Kevin Killourhy, Pat Loring, Bob Olszewski, Sami Saydjari and Tahlia Townsend for their help. This paper draws on Kymie Tan's forthcoming dissertation [25]. --R "Little Linuxes," Principles and Prac- tice "System structure for software fault tolerance," "A case study of ethernet anomalies in a distributed computing environment," Trend Analysis and Fault Prediction "Symptom based diagnosis," Computer Event Monitoring and Analysis "A survey of intrusion-detection techniques," "On the theory of scales of measurement," "On-line monitoring: A tutorial," "The concept of coverage and its effect on the reliability model of a repairable system," Reliable Computer Systems "Fault injection techniques and tools," "Fault injection - a method for validating computer-system dependability," The Theory of Stochastic Processes Time Series Analysis "Markov chains, classifiers, and intrusion detection," "Computer immunology," "A sense of self for unix processes," "Intrusion detection using sequences of system calls," "De- tecting intrusions using system calls: Alternative data models," "Self-nonself discrimination in a computer," "Sunshield basic security module guide," "Benchmarking anomaly-based detection systems," Defining the operational limits of anomaly-based intrusion detectors --TR Reliable computer systems (2nd ed.) Computer event monitoring and analysis A survey of intrusion detection techniques Computer immunology Little Linuxes Fault Tolerance Fault Injection On-Line Monitoring Fault Injection Techniques and Tools Benchmarking Anomaly-Based Detection Systems Markov Chains, Classifiers, and Intrusion Detection Self-Nonself Discrimination in a Computer A Sense of Self for Unix Processes --CTR Rodrigo M. Santos , Jorge Santos , Javier D. Orozco, A least upper bound on the fault tolerance of real-time systems, Journal of Systems and Software, v.78 n.1, p.47-55, October 2005 Tom Goldring, Scatter (and other) plots for visualizing user profiling data and network traffic, Proceedings of the 2004 ACM workshop on Visualization and data mining for computer security, October 29-29, 2004, Washington DC, USA Weng-Keen Wong , Andrew Moore , Gregory Cooper , Michael Wagner, What's Strange About Recent Events (WSARE): An Algorithm for the Early Detection of Disease Outbreaks, The Journal of Machine Learning Research, 6, p.1961-1998, 12/1/2005 Daniel P. Siewiorek , Ram Chillarege , Zbigniew T. Kalbarczyk, Reflections on Industry Trends and Experimental Research in Dependability, IEEE Transactions on Dependable and Secure Computing, v.1 n.2, p.109-127, April 2004

coverage;anomaly;dependability;anomaly detection

506833

Optimal partition of QoS requirements on unicast paths and multicast trees.

We investigate the problem of optimal resource allocation for end-to-end QoS requirements on unicast paths and multicast trees. Specifically, we consider a framework in which resource allocation is based on local QoS requirements at each network link, and associated with each link is a cost function that increases with the severity of the QoS requirement. Accordingly, the problem that we address is how to partition an end-to-end QoS requirement into local requirements, such that the overall cost is minimized. We establish efficient (polynomial) solutions for both unicast and multicast connections. These results provide the required foundations for the corresponding QoS routing schemes, which identify either paths or trees that lead to minimal overall cost. In addition, we show that our framework provides better tools for coping with other fundamental multicast problems, such as dynamic tree maintenance.

INTRODUCTION Broadband integrated services networks are expected to support multiple and diverse applications, with various quality of service (QoS) requirements. Accordingly, a key issue in the design of broadband architectures is how to provide the resources in order to meet the requirements of each connection. Supporting QoS connections requires the existence of several network mechanisms. One is a QoS routing mechanism, which sets the connection's topology, i.e., a unicast path or multicast tree. A second mechanism is one that provides QoS guarantees given the connection requirements and its topology. Providing these guarantees involves allocating resources, e.g., bandwidth and buffers, on the various network elements. Such a consumption of resources has an obvious cost in terms of network per- formance. The cost at each network element inherently depends on the local availability of resources. For instance, consuming all the available bandwidth of a link, considerably increases the blocking probability of future connections. Clearly, the cost of establishing a connection (and allocating the necessary re- sources) should be a major consideration of the connection (call) admission process. Hence, an important network optimization problem is how to establish QoS connections in a way that minimizes their implied costs. Addressing this problem impacts both the routing process and the allocation of resources on the selected topology. The latter translates into an end-to-end QoS requirement partition problem, namely local allocation of QoS requirements along the topology. The support of QoS connections has been the subject of extensive research in the past few years. Several studies and proposals considered the issue of QoS routing, e.g., [2], [8], [9], [21], [24] and references therein. Mechanisms for providing various QoS guarantees have been also widely investigated, e.g. [7], [22]. email:fdeanh@tx,ariel@eeg.technion.ac.il Although there are proposals for resource reservation, most notably RSVP [3], they address only the signaling mechanisms and do not provide the allocation policy. Indeed, the issue of optimal resource allocation, from a network perspective, has been scarcely addressed. Some studies, e.g. [13], consider the spe- cific, simple, case of constant link costs, which are independent of the QoS (delay) supported by the link. Pricing, as a network optimization mechanism, has been the subject of recent studies, however they either considered a basic best effort service envi- ronment, e.g. [14], [18], or simple, single link [17] and parallel link [20] topologies. In this paper, we investigate the problem of optimal resource allocation for end-to-end QoS requirements on given unicast paths and multicast trees. Specifically, we consider a frame-work in which resource allocation is based on the partition of the end-to-end QoS requirement into local QoS requirements at each network element (link). We associate with each link a cost function that increases with the severity of the local QoS requirement. As will be demonstrated in the next section, this framework is consistent with the proposals for QoS support on broadband networks. Accordingly, the problem that we address is how to partition an end-to-end QoS requirement into local requirements, such that the overall cost is minimized. This is shown to be intractable even in the (simpler) case of unicast connections. Yet, we are able to establish efficient (polynomial) solutions for both unicast and multicast connections, by imposing some (weak) assumptions on the costs. These results provide the required foundations for the corresponding QoS routing schemes, which identify either paths or trees that lead to minimal overall cost. Moreover, we indicate how the above frame-work provides better tools for coping with fundamental multi-cast problems, such as the dynamic maintenance of multicast trees. A similar framework was investigated in [4], [19]. There too, it was proposed that end-to-end QoS requirements should be partitioned into local (link) requirements and the motivation for this approach was extensively discussed. [19] discussed unicast connections and focused on loss rate guarantees. It considered a utility function, which is equivalent to (the minus of) our cost function. However, rather than maximizing the overall utility, [19] focused on the optimization of the bottleneck utility over the connection's path. That is, their goal was to partition the end-to-end loss rate requirement into link requirements over a given path, so as to maximize the minimal utility value over the path links. Specifically, [19] investigated the performance of a heuristic that equally partitioned the loss rate requirements over the links. By way of simulations, it was indicated that the performance of that heuristic was reasonable for paths with few (up to five) links and tight loss rate requirements; this finding was further supported by analysis. However, it was indicated that performance deteriorated when either the number of links became larger or when the connection was more tolerant to packet loss. It was concluded that for such cases, as well as for alternate QoS requirements (such as delay), the development of optimal QoS partition schemes is of interest. [4] considered multicast trees and a cost function that is a special case of ours. Each tree link was assigned an upper bound on the cost, and the goal was to partition the end-to-end QoS into local requirements, so that no link cost exceeds its bound. Specifically, [4] considered two heuristics, namely equal and proportional partitions, and investigated their performance by way of simulations. It was demonstrated that proportional partition offers better performance than equal partition, however it is not optimal. [4] too concluded that more complex (optimal) partition schemes should be investigated. These two studies provide interesting insights into our framework, and strongly motivate the optimization problems that we investigate. Another sequence of studied that is related to the present one is [9], [16]. These studies investigated QoS partitioning and routing for unicast connections, in networks with uncertain pa- rameters. Their goal was to select a path, and partition the QoS requirements along it, so as to maximize the probability of meeting the QoS requirements. As shall be shown, the link probability distribution functions considered in [9], [16] correspond to a special case of the cost functions considered in the present pa- per. The algorithms presented in [16] solve both the routing and the QoS partition problem for unicast connections, under certain assumptions. The present study offers an improved, less restric- tive, solution for unicast, and, more importantly, a generalized solution for multicast. The general resource allocation problem is a constraint optimization problem. Due to its simple structure, this problem is encountered in a variety of applications and has been studied extensively [12]. Optimal Partition of end-to-end QoS requirements over unicast paths is a special case of that problem, however the multicast version is not. Our main contribution is in solving the problem for multicast connections. We also present several algorithms for the unicast problem, emphasizing network related aspect, such as distributed implementation. The rest of this paper is structured as follows. Section II formulates the model and problems, and relates our framework to QoS network architectures. The optimal QoS partition problem for unicast connections is investigated in Section III. The optimal partition problem for multicast connections is discussed in Section IV. This problem is solved using a similar approach to that used for unicast, nonetheless the analysis and solution structure turn out to be much more complex. Section V applies these findings to unicast and multicast QoS routing. Finally, concluding remarks are presented in Section VI. Due to space limits, many technical details and proofs are omitted from this version and can be found in [15]. II. MODEL AND PROBLEMS In this section we present our framework and introduce the QoS partition problem. We assume that the connection topology is given, i.e., a path p for unicast, or a tree T for multicast. The problem of finding such topologies, namely QoS routing, is briefly discussed in Section V. For clarity, we detail here only the framework for unicast connections. The definitions and terminology for multicast trees are similar and are presented, together with the corresponding solution, in Section IV. A. QoS Requirements A QoS partition of an end-to-end QoS requirement Q, on a path p, is a vector l2p of local QoS requirements, which satisfies the end-to-end QoS requirement, Q. There are two fundamental classes of QoS parameters: bottleneck parameters, such as bandwidth, and additive parame- ters, such as delay and jitter. Each class induces a different form of our problem, and the complexities of the solutions are vastly different. For bottleneck parameters, we necessarily have determined by the bottle-neck link, i.e., allocating more than Q induces a higher cost, yet does not improve the overall QoS, the optimal partition is x . For additive QoS require- ments, a feasible partition, xp , must satisfy In this case, the optimal QoS partition problem is intractable [16], however we will show that, by restricting ourselves to convex cost functions, we can achieve an efficient (tractable) solution. Some QoS parameters, such as loss rate, are multiplicative, i.e., l2p (x l ). For instance, for a loss rate QoS requirement L, we have case too can be expressed as an additive requirement, by solving for (\Gamma log Q); indeed, the end- to-end requirement becomes (\Gamma log an additive requirement. There are QoS provision mechanisms, in which the (additive) delay bounds are determined by a (bottleneck) "rate". A notable example is the Guaranteed Service architecture for IP [23], which is based on rate-based schedulers [7], [22], [25], [26]. In some cases, such mechanisms may allow to translate a delay requirement on a given path into a bottleneck (rate) require- ment, hence the partitioning is straightforward. However, such a translation cannot be applied in general, e.g. due to complications created by topology aggregation and hierarchical routing. 2 Hence, our study focuses on the general partition problem of additive QoS requirements. B. Cost Functions As mentioned, we associate with each local QoS requirement value x l , a cost c l (x l ), and make the natural assumption that c l (x l ) is higher as x l is tighter. For instance, when x l stands for delay, c l (x l ) is a non-increasing function, whereas when x l stands for bandwidth, c l (x l ) is non-decreasing. The overall cost of a partition is the sum of the local costs, i.e., The cost may reflect the resources, such as bandwidth, needed to guarantee the QoS requirement. Alternatively, the cost may be the price that the user is required to pay to guarantee a specific QoS. The cost may be associated with either the set-up or the run-time phase of a connection. Also, it may be used for 1 This is also true for a multicast tree T. 2 Indeed, the ATM hierarchical QoS routing protocol [21], requires local (per cluster) QoS guarantees. network management to discourage the use of congested links, by assigning higher costs to those links. A particular form of cost evolves in models that consider uncertainty in the available parameters at the connection setup phase [9], [16], which we now briefly overview. In such mod- els, associated with each link is a probability of failure f l (x l ), when trying to set up a local QoS requirement of x l . The optimal QoS partition problem is then to find a QoS partition- that minimizes the probability of failure; that is, it minimizes the product we have log l2p log f l (x l ), we can restate this problem back as a summa- tion, namely we define a cost function for each link, c l log f l (x), and solve for these costs. C. Problem Formulation The optimal QoS partition problem is then defined as follows. Problem OPQ (Optimal Partition of QoS): Given a path p and an end-to-end QoS requirement Q, find a QoS partition x l g l2p , such that c x for any (other) QoS partition x 0 . This study focuses on the solution of Problem OPQ for additive QoS parameters, which, as mentioned, is considerably more complex than its bottleneck version. In Section III we solve the problem for unicast paths, and in Section IV we generalize the solution to multicast trees. For clarity, and without loss of gen- erality, we concretize the presentation on end-to-end delay requirements III. SOLUTION TO PROBLEM OPQ In this section we investigate the properties of optimal solutions to Problem OPQ for additive QoS parameters and present efficient algorithms. These results will be used in the next section to solve Problem MOPQ, i.e., the generalization of Problem OPQ to multicast trees. As mentioned, Problem OPQ is a specific case of the resource allocation problem. The fastest solution to this problem, [5], requires O(jpj log D=jpj). 3 In Section III-B we present a greedy pseudo-polynomial solution. This solution provides appealing advantages for distributed and dynamic implementations, as discussed in Section III-C. In Section III-D we present a polynomial solution that, albeit of slightly higher complexity than that of [5], provides the foundations of our solution to Problem MOPQ. Finally, in Section III-E we discuss special cases with lower complexity. As mentioned, we assume that the QoS parameter is end-to- delay. We further assume that all parameters are integers, and that the link cost functions are non-increasing with the delay and (weakly) convex. A. Notations xp (D) is a feasible partition of an end-to-end delay requirement D on the path p if it satisfies l2p x l - D. We omit the subscript p and/or the argument D when there is no ambigu- ity. x (D) denotes the optimal partition, namely the solution of Problem OPQ for an end-to-end delay requirement D and a path 3 jpj log(D=jpj) is also a lower bound for solving Problem OPQ [11], that is no fully polynomial algorithms exist. p. We denote by jx p j the norm l2p jx l j, hence x is feasible if The average ffi-increment gain for a link l is denoted by The average ffi-move gain is denoted by \Delta e!l (x; B. Pseudo-polynomial Solution Problem OPQ is a special case of the general resource allocation problem which has been extensively investigated [11], [12]. With the (weak) convexity assumption on the cost functions, it is a convex optimization problem with a simple constraint. It can be proved [12] that a greedy approach is applicable for such problems, namely it is possible to find an optimal solution by performing locally optimal decisions. GREEDY-ADD (D; ffi; c(\Delta); p): do 5 return x Fig. 1. Algorithm GREEDY-ADD Algorithm GREEDY-ADD (Figure 1) employs such a greedy approach. It starts from the zero allocation and adds the delay bit-by-bit, each time augmenting the link where the (negative) ffi-increment gain is minimal, namely where it most affects the cost. Using an efficient data structure (e.g. a heap), each iteration requires O(log jpj), which leads to an overall complexity of O( D log jpj). In [11] it is shown that the solution is ffi-optimal in the following sense: if x ffi is the output of the algorithm and x is the optimal solution, then jx Algorithm GREEDY-MOVE (Figure 2) is a modification of Algorithm GREEDY-ADD that, as shall be explained in Section III- C, has important practical advantages. The algorithm starts from any feasible allocation and modifies it until it reaches an optimal partition. Each iteration performs a greedy move, namely the move with minimal (negative) ffi-move gain. Fig. 2. Algorithm GREEDY-MOVE Let '(x) be the distance of a given partition from the optimal one, namely '(x) j jx \Gamma x j, where x is the optimal partition that is nearest to x. The next lemma implies that Algorithm GREEDY-MOVE indeed reaches a ffi-optimal solution. Lemma 1: Each iteration of Algorithm GREEDY-MOVE decreases '(x) by at least ffi, unless x is a ffi-optimal partition. Lemma 1 implies that Line 3 can be used as a (ffi-)optimality check. It also implies that the algorithm terminates with a ffi- optimal solution and that the number of iterations is proportional to '(x). Theorem 1 summarizes this discussion. 4 This also established the error due to our assumption of integer parameters. Theorem 1: Algorithm GREEDY-MOVE solves Problem OPQ in O( '(x) log jpj). Proof: By Lemma 1, there are at most '(x)=ffi iterations and a ffi-optimal solution is achieved. Each iteration can be implemented in O(log jpj) and the result follows. C. Distributed and Dynamic Implementation Algorithm GREEDY-MOVE can be employed in a distributed fashion. Each iteration can be implemented by a control message that traverses back and forth between the source and des- tination. At each traversal e; l of Line 2 are identified and the allocation change of the previous iteration is performed. This requires O(jpj'=ffi) end-to-end messages. Such a distributed implementation also exempts us from having to advertise the updated link cost functions. Algorithm GREEDY-MOVE can be used as a dynamic scheme that reacts to changes in the cost functions after an optimal partition has been established. Note that the complexity is proportional to the allocation modification implied by the cost changes, meaning that small allocation changes incur a small number of computations. D. Polynomial Solution In this section we present an improved algorithm of polynomial complexity in the input size (i.e., jpj and log D). In Section IV, we derive a solution to Problem MOPQ using a similar technique. Algorithm BINARY-OPQ (Figure finds optimal solutions for different values of ffi. The algorithm consecutively considers smaller values of ffi, until the minimal possible value is reached, at which point a (global) optimum is identified. 2 start from the partition 3 repeat Fig. 3. Algorithm BINARY-OPQ Obviously, the algorithm finds an optimal solution, since its last call to GREEDY-MOVE is with 1. The number of iterations is clearly of order O (log(D=jpj)). We need to bound the number of steps required to find the ffi-optimal partition at Line 4. Each iteration (except the first, for which starts from a 2ffi-optimal partition and employs greedy moves until it reaches a ffi-optimal partition. This bound is the same for all iterations, since it is a bound on the distance between a 2ffi-optimal partition and a ffi-optimal partition. Lemma 2: Let x GREEDY-MOVE(x; 2; c(\Delta); p). Then This lemma, proven in [15], resembles the proximity theorem presented in [11]. Theorem 2: Algorithm BINARY-OPQ solves Problem OPQ in O (jpj log jpj log(D=jpj)). Proof: By Lemma 2 and Theorem 1, each call to GREEDY-MOVE requires O(jpj log jpj). Since there are O(D=jpj) such calls, the result follows. E. Faster Solutions The following lemma, which is a different form of the optimality test in Algorithm GREEDY-MOVE, provides a useful threshold property of optimal partitions. Lemma 3: Let \Delta j min l2p \Delta l For all l 2 p, \Delta l (d; \Gamma1) is a non-increasing function (c l is convex) and (by definition) \Delta l (d the threshold \Delta relates to the optimal allocation as follows: l - d , \Delta l (d; This implies that an optimal solution to Problem OPQ can be found by selecting the D largest elements from the set For certain cost functions, this can be done analytically. For instance, in [16] we provide an O(jpj) solution for cost functions that correspond to delay uncertainty with uniform probability distributions. More generally, if the cost functions are strictly convex, then, given \Delta , one can use (1) to find an optimal solution in O(jpj). In [16], a binary search is employed for finding \Delta . Ac- cordingly, the resulting overall solution is of O(jpj log \Delta where pg. Note that log \Delta max is bounded by the complexity of representing a cost value. IV. SOLUTION TO MULTICAST OPQ (MOPQ) In this section we solve Problem OPQ for multicast trees. Specifically, given a multicast tree, we need to allocate the delay on each link, such that the end-to-end bound is satisfied on every path from the source to any member of the multicast group, and the cost associated with the whole multicast tree is minimized. We denote the source (root) of the multicast by s and the set of destinations, i.e., the multicast group, by M . A multicast tree is a set of edges T ' E such that for all there exists a path, p T (s; v), from s to v on links that belong to the tree T. We assume there is only one outgoing link from the source s, 5 and denote this link by r. all the outgoing links from v, i.e., all of l's neighbors; when N(l) is an empty set then we call l a leaf. T l is the whole sub-tree originating from l (including l itself). The branches of T are denoted by - frg. Observe that A feasible delay partition for a multicast tree T, is a set of link-requirements xT l2T such that We can now define Problem OPQ for multicast trees. Problem MOPQ (Multicast OPQ): Given a multicast tree T and an end-to-end delay requirement D, find a feasible partition x (D) , such that c (x (D)), for every (other) feasible partition xT (D). Remark 1: If there is more than one outgoing link from the source, then we can simply solve Problem MOPQ indepen- dently, for each tree T r i corresponding to an outgoing link r i 5 See Remark 1. 6 Again, when no ambiguity exists, we omit the sub-script T and/or the argument D. from s. Thus, our assumption, that there is only one outgoing link from r, does not limit the solution. We denote by MOPQ(T; d) the set of optimal partitions on a tree T with delay D. c T (d) denotes the tree cost function, i.e., the cost of (optimally) allocating a delay d on the tree T. In other words, c T A. Greedy Properties The general resource allocation problem can be stated with tree-structured constraints and solved in a greedy fashion [12]. An efficient O(jTj log jTj log D) algorithm is given in [11]. However, that "tree version" of the resource allocation problem has a different structure than Problem MOPQ. Indeed, the simple greedy approach, namely repeated augmentation of the link that most improves the overall cost, fails in our framework. However, as we show below, some greedy structure is main- tained, as follows: if at each iteration we augment the sub-tree that most improves the overall cost, then an optimal solution is achieved. The main difference of our framework is that the constraints are not on sub-trees, but rather on paths. The greedy approach fails because of the dependencies among paths. On the other hand, we note that the tree version of the resource allocation problem may be applicable to other multicast resource allocation problems, in which the constraints are also on sub-trees. For example, suppose a feasible allocation must recursively sat- isfy, for any sub-tree T e , some arbitrary (sub-tree) constraint. We proceed to establish the greedy structure of Problem MOPQ. First, we show that if all link cost functions are convex, then so is the tree cost function. Lemma 4: If fc l g l2T are convex then so is c T (d). By Lemma 4, we can replace T by an equivalent convex link. Any sub-tree T l , can also be replaced with an equivalent convex link, hence so can - T. However, these results apply only if the allocation on every sub-tree is optimal for the sub-tree. This property is sometimes referred to as the "optimal sub-structure" property [1], and is the hallmark of the applicability of both dynamic-programming and greedy methods. Lemma 5: Let x let the sub-partition x e (D e l g l2Te , where D r . Then x e Lemma 5 implies that, for any optimally partitioned trees, we can apply the greedy properties of Section III. That is, the partition on r and - T is a solution to Problem OPQ on the 2-link path T). This suggests that employing greedy moves between r and - T will solve Problem MOPQ, and this method can be applied recursively for the sub-trees of T. Indeed, this scheme is used by the algorithms presented in the next sections. B. Pseudo-polynomial Solution We employ greedy moves between r and - T. The major difficulty of this method is the fact that c - T (d) is unavailable. Computing c - T (d) for a specific d requires some x MOPQ(T; d). Fortunately, we can easily compute c - T (d given x (d). Since the greedy approach is applicable, we may simply perform a greedy augmentation and recompute the cost. Note that adding ffi to - adding ffi to all fT l g l2N(r) . In the worst case, this must be done recursively for all the sub-trees and O(jTj) links are augmented. Procedure TREE-ADD (Figure performs a ffi-augmentation on a tree T. We assume that for each sub-tree T l the value of (D T l ; ffi) for the current allocation is stored in the variable (ffi). At Line 1 it is decided if r or - T would be augmented. (D T l ) or \Delta r can be made by a simple comparison. If - T should be augmented then Procedure TREE-ADD is called recursively on its compo- nents. Finally, \Delta T (\Sigmaffi) is updated at Lines 6-7. 3 else 4 for each l 2 N(r) do (ffi)g a a If r is a leaf we define the sum to be 1. Fig. 4. Procedure TREE-ADD Algorithm BALANCE (Figure 5) is a dynamic algorithm that solves Problem MOPQ. It starts from any feasible tree partition and performs greedy moves between r and - T. The while loop at Line 7 computes \Delta r! - T (x; ffi). If it is negative then moving from r to - T reduces the overall cost. The augmentation of - is done by calling TREE-ADD on each of its components. The while loop at at Line 11 performs moves from - T to r in a similar way. To be able to check the while condition and for calling TREE- ADD, we must have (\Sigmaffi) for all l 2 T. This requires an optimal partition on each sub-tree. Algorithm BALANCE makes sure that this is indeed the case by recursively calling itself (Line 5 ) on the components of - T. Since any allocation to a leaf is an optimal partition on it, the recursion is stopped once we reach a leaf. After the tree is balanced the algorithm updates which is used by the calling iteration. if T is a leaf then 4 return 5 (else) for each l 2 N(r) do 7 while 9 for each l 2 N(r) do 13 for each l 2 N(r) do Fig. 5. Algorithm BALANCE We proceed to analyze the complexity of BALANCE. We first define a distance 'T (x) which is the tree version of the path distance defined in Section III-B. Let ' r (xT where x is the optimal partition nearest to T. Let x l x l e e2T l . We define '(xT Theorem 3: Algorithm BALANCE finds a ffi-optimal solution to Problem MOPQ in O(jTj('(x)=ffi) Proof: '(x)=ffi bounds the number of calls to Procedure TREE-ADD. At the worst case, TREE-ADD requires O(jTj) for each call. The recursive calls to BALANCE also require O(jTj). Remark 2: We can apply Algorithm BALANCE on the feasible partition rg. Clearly, '(x) - D in this case. Thus, Problem MOPQ can be solved in O(jTjD=ffi). C. Distributed and Dynamic Implementation Algorithm BALANCE can be readily applied in a distributed fashion. Each augmentation in Procedure TREE-ADD is propagated from the root to the leafs. A straightforward implementation requires O(t) time, 7 where t is the depth of the tree. At most O(t) recursive calls to BALANCE are performed se- quentially. Finally, the number of calls to TREE-ADD after the sub-trees are balanced, is bounded by ' max r (x)=ffi, where ). The overall complexity is, there- fore, O r (x)=ffi . Note that for balanced trees O(log jTj). Algorithm BALANCE (as is the case for Algorithm GREEDY- MOVE) can be used as a dynamic scheme that reacts to changes in the cost functions after an optimal partition is established. The complexity of Algorithm BALANCE is proportional to the distance from the new optimal allocation. Again, small changes (i.e., small '(x)) incur a small number of computations. D. Polynomial Solution We can now present a polynomial solution. Algorithm BINARY-MOPQ uses an approach that is identical to the one used for the solution of Problem OPQ. The algorithm consecutively calls BALANCE for smaller values of ffi, until the minimal possible value is reached, at which point an optimal partition is identified. 4 repeat Fig. 6. Algorithm BINARY-MOPQ We will show that, at each iteration of the algorithm, ' max r (x) is bounded by tffi. Therefore, '(x)=ffi - tjTj and the over-all complexity of this algorithm is O(jTj 2 t log(D=jTj)). The Lemma 6 is the equivalent of Lemma 2 for multicast. Lemma t. 7 assuming that traveling a link requires one time unit. denote the value of \Delta T (ffi) at the termination of T). Note that \Delta T (d; ffi) assumes a ffi-optimal partition on the tree, hence it is different from (c T (d which assumes the optimal partition. Theorem 4 states the complexity of Algorithm BINARY- MOPQ. Theorem 4: Algorithm BINARY-MOPQ finds a solution to Problem MOPQ in O(jTj 2 t log(D=jTj)). Comparing this result to the O(jTjD=ffi) complexity of Algorithm BALANCE (see Remark 2), indicates that Algorithm BALANCE is preferable when D=ffi ! jTjt log(D=jTj). Again, note that, for balanced trees we can implement Algorithm BINARY-MOPQ in a distributed fashion (as in Section IV-C), with an overall complexity of O(t 3 log(D=jTj). Remark 3: Algorithm BINARY-MOPQ starts from a coarse partition and improves the result by refining the partition at each iteration; this means that one may halt the computation once the result is good enough, albeit not optimal. Remark 4: It is possible to modify the algorithm to cope with heterogeneity in the QoS requirements of the multicast group members. In the worst case, the complexity of the solution grows by a factor of O (jM j), while the complexity of the pseudo-polynomial solution remains unchanged. The details of this extension are omitted here. V. ROUTING ISSUES In the previous sections, we addressed and solved optimal QoS partition problems for given topologies. These solutions have an obvious impact on the route selection process, as the quality of a route is determined by the cost of the eventual (opti- mal) QoS partition over it. Hence, the unicast partition problem OPQ induces a unicast routing problem, OPQ-R, which seeks a path on which the cost of the solution to Problem OPQ is min- imal. Similarly, Problem MOPQ induces a multicast routing problem MOPQ-R. In this section we briefly discuss some current and future work in the context of these routing problems. A. OPQ-R As was the case with Problem OPQ, with OPQ-R too there is a significant difference between bottleneck QoS requirements and additive ones. As explained in Section II, for a bottleneck QoS requirement, the end-to-end requirement determines Q Q for all links in the path (or tree), and a corresponding link cost, c l (Q). Therefore, the routing problem OPQ-R boils down to a "standard" shortest-path problem with link length c l (Q). As noted, in the context of Problem OPQ, providing delay requirements through rate-based schedulers [7], [22], [25], [26], translates the additive requirement into a (simpler) bottleneck requirement. However, in the context of Problem OPQ-R, such a translation is not possible anymore, since paths differ also in terms of constant (rate-independent) link delay components. Efficient solutions for Problem OPQ-R, under delay requirements and rate-based schedulers, have been presented in [9]. The general OPQ-R problem, under additive QoS require- ments, is much more complex, and has been found to be intractable [16]. Note that, even in a simpler, constant-cost frame- work, where each link is characterized by a delay-cost pair (rather than a complete delay-cost function), routing is an intractable problem [6]. In the latter case, optimal routing can be achieved through a pseudo-polynomial dynamic-programming scheme, while ffl-optimal solutions can be achieved in polynomial time [10]. The general OPQ-R problem, under the framework of the present study, was solved in [16]. The solution is based on dynamic-programming and assumes that the link cost functions are convex. An exact pseudo-polynomial solution, as well as an ffl-optimal polynomial solution, have been presented. We note that a single execution of those algorithms finds a unicast route from the source to every destination and for every end-to-end delay requirement. B. MOPQ-R As could be expected, finding optimal multicast trees under our framework is much more difficult than finding unicast paths. Even with bottleneck QoS requirements, MOPQ-R boils down to finding a Steiner tree, which is known to be an intractable problem [6]. We are currently working on solving MOPQ-R for additive QoS requirements. We established an efficient scheme for the fundamental problem of adding a new member to an existing multicast tree. This provides a useful building block for constructing multicast trees. Another important building block is the optimal sub-structure property (established in Section IV), which is an essential requirement for the application of greedy and dynamic programming solution approaches. Interestingly, the above problem, of adding members to multicast trees, may serve to illustrate the power of our framework over the simpler, constant-cost framework. In the latter, there is a single delay-cost pair for each link (rather than a complete delay-cost function), and the goal is to find a minimal cost tree that satisfies an end-to-end delay constraint. 8 Under that frame- work, it is often impossible to connect a new member to the "tree top", i.e., the leaves and their neighborhood. This is a consequence of cost minimization considerations, which usually result with the consumption of all (or most of) the available delay at the leaves. For example, consider the network of Figure 7. The source is S and the multicast group is fA; Bg; the end-to- end delay bound is 10 and the link delay-cost pairs are specified. Suppose we start with a tree for node A, i.e., the link (S; A). Since A exhausts all the available end-to-end delay, we cannot add B to the tree by extending A's branch with the (cheap) link rather, we have to use the (expensive) link (S; B). Note that we would get the same result even if there were an additional link from S to A with shorter delay, say 9, and slightly higher cost, say 11. Our framework allows a better solution, as it lets the link several delays and costs. For instance, it could advertise a delay of 10 with a cost of 10 and a delay of 9 with a cost of 11. When adding B to the tree, we can change the delay allocation on (S; to 9 (thus paying 11 instead of 10), which allows us to use the link (A; B) for adding B. The cost of the resulting tree is 12, as opposed to 20 in the previous solution (i.e., using link (S; B)). Note that, when adding B, one can 8 That framework was the subject of numerous studies on constrained multi-cast trees. 'j 'i `j 'i 'j 'i Fig. 7. Example: extending a multicast tree consider the "residual" cost for each link, i.e., the cost of tightening the delay bound on existing allocations. In our example, the residual cost function of link (S; A) is 0 for a delay of 10 (i.e., the current allocation) and 1 for a delay of 9 (i.e., the added cost for tightening the requirement). The last observation implies that adding a new member to an existing tree boils down to finding an optimal unicast path, with respect to the residual cost functions, from the new member to the source; i.e., an instance of Problem OPQ-R, for which efficient solutions have been established in [16]. VI. CONCLUSIONS We investigated a framework for allocating QoS resources on unicast paths and multicast trees, which is based on partitioning QoS requirements among the network components. The quality of a partition is quantified by link cost functions, which increase with the severity of the QoS requirement. We indicated that this framework is consistent with the major proposals for provisioning QoS on networks. Indeed, the problem of how to efficiently partition QoS requirements among path or tree links has been considered in previous studies, however till now only heuristic approaches have been addressed. The present study is the first to provide a general optimal solution, both for unicast paths and multicast trees. We demonstrated how the various classes of QoS requirements can be accommodated in within our framework. We showed that the partitioning problems are simple when dealing with bottleneck requirements, such as bandwidth, however they become intractable for additive (or multiplicative) requirements, such as delay, jitter and loss rate. Yet we established that, by introducing a mild assumption of weak convexity on the cost efficient solutions can be derived. We note that weak convexity essentially means that, as the QoS requirement weakens, the rate of decrease of the cost function diminishes. This is a reasonable property, as cost functions are lower-bounded, e.g. by zero. Moreover, it indeed makes sense for a cost function to strongly discourage severe QoS requirements, yet gradually become indifferent to weak (and, eventually, practically null) requirements. Hence, the scope of our solutions is broad and general. Specifically, we presented several greedy algorithms for the unicast problem (OPQ). Algorithm GREEDY-MOVE, is a pseudo-polynomial solution, which can be implemented in a distributed fashion. The complexity of this solution is O('(x) log jpj), where '(x) - D is the distance between the initial allocation, x, and the optimal one. It can also be applied as a dynamic scheme to modify an existing allocation. This is useful in dynamic environments where the cost of resources changes from time to time. Note that the complexity is proportional to '(x), meaning that small cost changes require a small number of computations to regain optimality. Algorithm BINARY-OPQ is a polynomial solution from which we later build our solution to the multicast problem (MOPQ). The complexity of this solution is O (jpj log jpj log(D=jpj)). Next, we addressed the multicast problem MOPQ. We began by showing that the fundamental properties of convexity and optimal sub-structure generalize to multicast trees. Then, we established that Problem MOPQ also bears a greedy struc- ture, although much more complex than its OPQ counterpart. Again, the greedy structure, together with the other established properties, provided the foundations for an efficient solu- tions. Algorithm BALANCE is a pseudo-polynomial algorithm which can be applied as a dynamic scheme. Its complexity again, the distance between the initial allocation and the optimal one. A distributed implementation requires O(t 2 ' max r (x)), where t is the depth of the tree and ' max r (x) is the maximal distance of any link's allocation from it optimal one. Note that for balanced trees Algorithm BINARY-MOPQ is a polynomial solution with a complexity of O(jTj 2 t log(D=jTj)). A distributed implementation of this algorithm requires O(t 3 log(D=jTj). We note that our solutions are applicable to heterogeneous multicast members, each with a different delay requirement. Lastly, we discussed the related routing problems, OPQ-R and MOPQ-R. Here, the goal is to select either a unicast path or multicast tree, so that, after the QoS requirements are optimally partitioned over it, the resulting cost would be minimized. Again, unicast proves to be much easier than multicast. In par- ticular, for bottleneck QoS requirements, OPQ-R boils down to a simple shortest-path problem. For additive requirements, OPQ- R is intractable, yet an efficient, ffl-optimal solution has been established in [16]. For multicast, all the various versions of MOPQ-R are intractable. We are currently investigating Problem MOPQ-R under additive requirements, and have obtained an efficient scheme for adding new members to a multicast tree. Several important issues are left for future work. One is multicast routing, i.e., Problem MOPQ-R, for which just initial (yet encouraging) results have been obtained thus far. Another important aspect is the actual implementation of our solutions in within practical network architectures. In this respect, it is important to note that a compromise with optimality might be called for. Indeed, while our solutions are of reasonable complexity, a sub-optimal solution that runs substantially faster might be preferable in practice. Relatedly, one should consider the impact of the chosen solution for QoS partitioning on the routing process. The latter has to consider the quality of a selection (i.e., path or tree) in terms of the eventual QoS parti- tion. This means that simpler partitions should result in simpler routing decisions, which provides further motivation for compromising optimality for the sake of simplicity. The optimal solutions established in this study provide the required starting point in the search of such compromises. Lastly, we believe that the framework investigated in this study, where QoS provisioning at network elements is characterized through cost functions, provides a powerful paradigm for dealing with QoS networking. We illustrated the potential benefits through an example of dynamic tree maintenance. Further study should consider the implications and potential advantages of our framework, when applied to the various problems and facets of QoS networking. --R Introduction to Algo- rithms A framework for QoS-based routing in the internet - RFC no Call admission and resource reservation for multicast sessions. The complexity of selection and ranking in X Computers and Intractability. Efficient network QoS provisioning based on per node traffic shaping. QoS routing mechanisms and OSPF extensions. Approximation schemes for the restricted shortest path prob- lem Lower and upper bounds for the alloction problem and other nonlinear optimization problems. Resource Allocation Problems. Multicast routing for multimedia communication. Optimal partition of QoS requirements on unicast paths and multicast trees. QoS routing in networks with uncertain pa- rameters A new approach to service provisioning in ATM networks. Pricing Congestable Network Re- sources Allocation of local quality of service constraints to meet end-to-end requirements Incentive pricing in multi-class communication networks Private network-network interface specification v1 A generalized processor sharing approach to flow control in integrated services networks: the multiple node case. Specification of guaranteed quality of service - RFC no Service disciplines for guaranteed performance service in packet-switching networks --TR Introduction to algorithms Approximation schemes for the restricted shortest path problem Multicast routing for multimedia communication A new approach to service provisioning in ATM networks Lower and upper bounds for the allocation problem and other nonlinear optimization problems A generalized processor sharing approach to flow control in integrated services networks Efficient network QoS provisioning based on per node traffic shaping QoS routing in networks with uncertain parameters QoS routing in networks with inaccurate information --CTR Ariel Orda , Alexander Sprintson, A scalable approach to the partition of QoS requirements in unicast and multicast, IEEE/ACM Transactions on Networking (TON), v.13 n.5, p.1146-1159, October 2005 Wen-Lin Yang, Optimal and heuristic algorithms for quality-of-service routing with multiple constraints, Performance Evaluation, v.57 n.3, p.261-278, July 2004 Wen-Lin Yang, A comparison of two optimal approaches for the MCOP problem, Journal of Network and Computer Applications, v.27 n.3, p.151-162, August 2004 Sun-Jin Kim , Mun-Kee Choi, Evolutionary algorithms for route selection and rate allocation in multirate multicast networks, Applied Intelligence, v.26 n.3, p.197-215, June 2007 Ariel Orda , Alexander Sprintson, Precomputation schemes for QoS routing, IEEE/ACM Transactions on Networking (TON), v.11 n.4, p.578-591, August Bin Xiao , Jiannong Cao , Zili Shao , Qingfeng Zhuge , Edwin H. -M. Sha, Analysis and algorithms design for the partition of large-scale adaptive mobile wireless networks, Computer Communications, v.30 n.8, p.1899-1912, June, 2007

routing;broadband networks;convex costs;QoS partitioning;QoS-dependent costs;multicast;unicast

506837

Generalized loop-back recovery in optical mesh networks.

Current means of providing loop-back recovery, which is widely used in SONET, rely on ring topologies, or on overlaying logical ring topologies upon physical meshes. Loop-back is desirable to provide rapid preplanned recovery of link or node failures in a bandwidth-efficient distributed manner. We introduce generalized loop-back, a novel scheme for performing loop-back in optical mesh networks. We present an algorithm to perform recovery for link failure and one to perform generalized loop-back recovery for node failure. We illustrate the operation of both algorithms, prove their validity, and present a network management protocol algorithm, which enables distributed operation for link or node failure. We present three different applications of generalized loop-back. First, we present heuristic algorithms for selecting recovery graphs, which maintain short maximum and average lengths of recovery paths. Second, we present WDM-based loop-back recovery for optical networks where wavelengths are used to back up other wavelengths. We compare, for WDM-based loop-back, the operation of generalized loop-back operation with known ring-based ways of providing loop-back recovery over mesh networks. Finally, we introduce the use of generalized loop-back to provide recovery in a way that allows dynamic choice of routes over preplanned directions.

Introduction For WDM networks to oer reliable high-bandwidth services, automatic self-healing capabilities, similar to those provided by SONET, are required. In particular, pre-planned, ultrafast restoration of service after failure of a link or node is required. As WDM networks mature and expand, the need has emerged for self-healing schemes which operate over a variety of network topologies and in a manner which is bandwidth e-cient. While SONET provides a known and robust means of providing recovery in high-speed networks, the techniques used for SONET are not always immediately applicable to WDM systems. Certain issues, such as wavelength assignment and wavelength changing, make WDM self-healing dierent from SONET self-healing. Our purpose is to present a method for service restoration in optical networks which has the following characteristics: Speed: we want the speed of recovery to be of the order of the speed of switching and require minimal processing overhead. Transparency: we seek a method of recovery which can be done at the optical layer, without regard for whatever protocol(s) may be running over the optical layer. Flexibility: our method should not constrain primary routings and should provide a large choice of back-up routes to satisfy such requirements as bounds on average or maximum back-up length. In this paper, we present an approach which altogether moves away from rings. The rationale behind our approach is that, while ring recovery makes sense over network topologies which are composed of interconnected rings, rings are not fundamental to network recovery over mesh networks. Indeed, even embedding rings over a given topology can have signicant implications for hardware costs ([BGSM94]). We present generalized loop-back, a new method of achieving loop-back recovery over arbitrary two-link-redundant and two-node-redundant networks to restore service after the failure or a link or a node, respectively. Loop-back recovery over mesh networks without the use of ring covers was rst introduced in [FMB98, MFB99]. We represent each network by a graph, with each node corresponding to a vertex and each two-ber link to an undirected edge. The graph corresponding to a link (node)-redundant network is edge (vertex)-redundant. The principle behind generalized loop-back is to create primary and secondary digraphs, so that, upon failure of a link or node, the secondary digraph can be used to carry back-up tra-c that provides loop-back to the primary graph. Each primary or secondary digraph may correspond to a wavelength on a ber or a full ber. The secondary digraph is the conjugate of the primary digraph. Each direction in a link is associated with a given primary graph. Our algorithms perform the choice of directions to establish our digraphs. Our approach meets our three goals: speed, transparency and exibility. Although we use preplanning of directions, our network management protocol determines, in real time, the back-up route that will be utilized. We do not however, require processing as in traditional dynamic recovery schemes. In eect, our network management protocol provides dynamic real-time discovery of routings along pre-planned directions determined by our direction selection algorithms. Since our protocol (see Section 2.3) requires very simple processing and the optical layer remains responsible for recovery (ensuring transparency), we have speed of recovery combined with exibility. In particular, depending on availability of links or wavelengths (which may be aected by congestion or failures in the network), dierent back-up routes may be selected, but the selection will be automatic and will not require active comparison, by the network management, of the dierent possible routes. In Section 1.1, we give an overview of relevant work in the area of of network protection and restoration. In Section 2.1, we discuss generalized loop-back recovery for link failure. In Section 2.2, we present our method for loop-back recovery for node failures in arbitrary vertex-redundant networks. A simple network protocol, presented in Section 2.3, allows for distributed operation of recovery from link or node failure. When we consider recovery from node failures, we must contend with the fact that a node may carry several tra-c ows, all of which are disrupted when the node fails. Section 3 of our paper considers a range of dierent applications for generalized loop-back: We address the goal of exibility in the choice of back-up routings. We present a means of selecting directions for generalized loop-back so as to avoid excessive path lengths. Our algorithm allows a choice among a large number of alternative directions. The choice of directions may greatly aect the length of back-up paths. To avoid excessive loss and jitter along recovery paths, we present heuristic algorithms that signicantly reduce the average and maximum length of recovery paths over random choices of directions. We may use generalized loop-back to perform wavelength-based recovery, which we term WDM-based loop-back recovery, instead of ber-based recovery in mesh networks. We illustrate why the method of cover of rings using double-cycle covers is not directly applicable to WDM loop-back recovery. Generalized loop-back can yield several choices of backup routes for a given set of directions. We brie y illustrate how generalized loop-back can be used to avoid the use of certain links. Finally, in Section 4, we present conclusions and directions for further research. 1.1 Background Methods commonly employed for link protection in high-speed networks can be classied as either dynamic or pre-planned, though some hybrids schemes also exist ([SOOH93]). The two types oer a tradeo between adaptive use of back-up (or \spare") capacity and speed of restoration restoration typically involves a search for a free path using back-up capacity ([HKSM94, GK93, Bak91]) through broadcasting of help messages ([CBMS93, FY94, Gro87, Wu94]). The performance of several algorithms is given in ([BCS93, CBMS93]). Overheads due to message passing and software processing render dynamic processing slow. For dynamic link restoration using digital cross-connect systems, a two second restoration time is a common goal for SONET ([FY94, Gro87, KA93, Sos94, Wu94, YH88]). Pre-planned methods depend mostly on look-up tables and switches or add-drop multiplexers. For optical networks, switches may operate in a matter of microseconds or nanoseconds. Thus, to meet our speed requirement, we consider pre-planned methods, even though pre-planned methods suer from poorer capacity utilization than dynamic systems, which use of real-time availability of back-up capacity. Within pre-planned methods, we may distinguish between path and link or node restoration. Path restoration refers to recovery applied to connections following a particular path across a network. Link or node restoration refers to recovery of all the tra-c across a failed link or node, respectively. Path restoration may be itself subdivided into two dierent types: live (dual-fed) back-up and event-triggered back-up. In the rst case, two live ows, a primary and a back-up, are transmitted. The two ows are link-disjoint if we seek to protect against link failure, or node-disjoint (except for the end nodes) if we seek to protect against node failure. Upon failure of a link or node on the primary ow, the receiver switches to receiving on the back-up. Recovery is thus extremely fast, requiring action only from the receiving node, but back-up capacity is not shared among connections. In the second case, event-triggered path restoration, the back-up path is only activated when a failure occurs on a link or node along the primary path. Backup capacity can be shared among dierent paths ([WLH97]), thus improving capacity utilization for back-up channels and allowing for judicious planning ([BPG92, HBU95, GBV91, HB94, GKS96, SNH90, VGM93, Fri97, NHS97]). However, recovery involves coordination between the sender and receiver after a failure event and action from nodes along the back-up path. These coordination eorts may lead to signicant delays and management overhead. Pre-planned link or node restoration can be viewed as a compromise between live and event-triggered path restoration. Pre-planned link restoration is not as capacity-e-cient as event-triggered path restoration ([CWD97, RIG88, LZL94]), but is more e-cient than live back-up path restoration, since sharing of back-up bandwidth is allowed. The tra-c along a failed link or node is recovered, without consideration for the end points of the tra-c carried by the link or node. Thus, only the two nodes adjacent to the failure need to engage in recovery. The back-up is not live, but triggered by a failure. A comparison of the trade-os between end-to-end recovery and patch of a segment (we assume a segment to be a single link or node), is given in [DW94]. An overview of the dierent types of protection and restoration methods is given in [RM99] and comparisons between path protection and event-triggered path protection are given in [RM99, RIG88, JVS95, JAH94, XM99]. Link or node restoration also benets from a further advantage, which makes it very attractive for pre-planned recovery: since it is not dependent upon specic tra-c patterns, it can be pre-planned once and for all. Thus, link or node restoration is particularly attractive at lower layers, where network management may not be aware, at all locations of the network, of the origination and destination, or of the format ([Wu94]) of all the tra-c being carried at that location. Therefore, in this paper we concentrate on pre-planned link and node restoration in order to satisfy our transparency requirement. Moreover, link restoration satises the rst part of our exibility goal, since restoration is done without consideration for primary routings. For pre-planned link restoration, the main approaches have been through the use of covers of rings and, more recently, through pre-planned cycles ([GS98]). The most direct approach is to design the network in term of rings. The building blocks of SONET networks are generally self-healing rings (SHRs) and diversity protection (DP) ([WCB91, Was91, WB90, SWC93, SGM93, SF96, STW95, HT92]). SHRs are unidirectional path-switched rings (UPSRs) or bi-directional line-switched rings (BLSRs), while DP refers to physical redundancy where a spare link (node) is assigned to one or several links (nodes) ([Wu92] pp. 315-32). In rings, such as BLSR, link or node restoration is simply implemented using loop-back. The waste of bandwidth due to back-hauling may be remedied by looping back at points other than the failure location ([Mag97, KTK94]). Using only DP and SHRs is a constraint which has cost implications for building and expanding networks ([WKC89]); see [Sto92] for an overview of the design of topologies under certain reliability constraints. However, rings are not necessary to construct survivable networks ([NV91, WH91]). Mesh-based topologies can also provide redundancy ([Sto92, JHC93, WKC88]). Ring-based architectures may be more expensive than meshes ([BGSM94]), and as nodes are added, or networks are interconnected, ring-based structure may cease to be maintained, thus limiting their scalability. Even if we constrain ourselves to always use ring-based architectures, such architectures may not easily bear changes and additions as the network grows ([WKC89, Wu92, WKC88]. For instance, adding a new node, connected to its two nearest node neighbors, will preserve mesh structure, but may not preserve ring structure. Our arguments indicate that, for reasons of cost and extensibility, mesh-based architectures are more promising than interconnected rings. Covering mesh topologies with rings is a means of providing both mesh topologies and distributed, ring-based restoration. There are several approaches to covers of rings for networks in order to ensure link restorability. One approach is to cover nodes in the network by rings ([Was91]). In this manner, a portion of links are covered by rings. If primary routing is restricted to the covered links, then link restoration can be eected on each ring in the same manner as in a traditional SHR, by routing the back-up tra-c around the ring in the opposite direction to the primary tra-c. Using such an approach, the uncovered links can be used to carry unprotected tra-c, i.e. tra-c which may not be restored if the link which carries it fails. To allow every link to carry protected tra-c, other ring-based approaches ensure every link is covered by a ring. One approach to selecting such covers is to cover a network with rings so that every link is part of at least one ring ([Gro92]). This approach suers from some capacity drawbacks. With ber-based restoration, every ring is a four-ber ring. A link covered by two rings requires eight bers; a link covered by n rings requires 4n bers. Alternatively, the logical bers can be physically routed through four physical bers, but only at the cost of signicant network management overhead. Minimizing the amount of ber required to obtain redundancy using ring covers is equivalent to nding the minimum cycle cover of a graph, an NP-complete problem ([Tho97, ILPR81]), although bounds on the total length of the cycle cover may be found ([Fan92]). A second approach to ring covers, intended to overcome the di-culties of the rst approach, is to cover every link with exactly two rings, each with two bers. The ability to perform loop-back style restoration over mesh topologies was rst introduced in [ESH97, ES96]. In particular, [ESH97] considers link failure restoration in optical networks with arbitrary two-link redundant arbitrary mesh topologies and bi-directional links. The approach is an application of the double-cycle ring cover ([Jae85, Sey79, Sze73]). For planar graphs, the problem can be solved in polynomial-time; for non-planar graphs, it is conjectured that double cycle covers exist, and a counterexample would have to obey certain properties ([God85]). Node recovery can be eected with double cycle ring covers, but such restoration requires cumbersome hopping among rings. In subsection 3.2, we consider double- cycle covers in the context of wavelength-based recovery. In order to avoid the limitations of ring covers, an approach using pre-congured cycles, or p- cycles, is given in [GS98]. A p-cycle is a cycle on a redundant mesh network. Links on the p-cycle are recovered by using the p-cycle as a conventional BLSR. Links not on the p-cycle are recovered by selecting, along the p-cycle, one the paths which connect the nodes which are the end-points of the failed link. We may note that some di-culty arises from the fact that several p-cycles may be required to cover a network, making management among p-cycles necessary. The fact that a single p-cycle may be insu-cient arises from the fact that a Hamiltonian might not exist, even in a two-connected graph. Even nding p-cycles which cover a large number of nodes may be di-cult. Some results ([Fou85, Jac80, ZLY85]) and conjectures ([HJ85, Woo75]) exist concerning the length of maximal cycles in two-connected graphs. The p-cycle approach is in eect a hybrid ring approach, which mixes path restoration (for links not on the p-cycle) with ring recovery (for links on the p-cycle). Generalized Loop-back 2.1 Generalized Loop-back for Recovery from Link Failures The gist of our approach is to eliminate the use of rings. Instead, a primary (secondary) digraph (corresponding to a set of unidirectional bers or wavelengths) is backed up by another, secondary (primary) digraph (corresponding to a set of unidirectional bers or wavelengths in the reverse direction of the primary (secondary) digraph). After a failure occurs, we broadcast the stream carried by the primary (secondary) digraph along the failed link onto the secondary (primary) digraph. We later show a protocol which ensures that only a single connection arrives to each node on the back-up path.When the back-up path reaches the node which lost its connection along the primary (secondary) digraph because of the failure, the tra-c is restored onto the primary (secondary) digraph. To illustrate our method, consider a simple sample network. Our algorithm works by assigning directions to each of the two bers on each link. Figure 1.b shows in dashed arrow lines the directions of the primary digraph for each link and in thin dashed lines the directions of the secondary digraph for each link. The topology of the network is shown in bold lines without arrows. A break in a link is shown by discontinued lines. The shortest back-up path is node 3 4. Node 3 eliminates a duplicate connection which arrives to it via node 6 ! node 5 ! node 4 ! node 3. Node 7 eliminates a duplicate connection which arrives to it via node 2 ! node 1 ! node 8 ! node 7. Note that back-haul need not always occur. For instance, in Figure 1.b, if the original connection went from node 4 to node 2 via node 3, then after recovery the connection would commence at node 4 and traverse, in order, nodes 3, 6, 7 en route to node 2. Not every assignment of directions provides the possibility for loop-back recovery. As an example, consider in Figure 1.a the same network topology as in Figure 1.b with dierent directions. The directions are provided in such a way that, when no failures are present, all nodes are reachable from each other on the primary wavelength on ber 1 and on the secondary wavelength on ber 2. However, the same link failure as in Figure 1.b is not recoverable. This example illustrates the importance of proper selection of the directions on the links. We may now formalize our approach. We dene an undirected graph E) to be a set of nodes N and edges E. With each edge [x; y] of an undirected graph, we associate two directed arcs (x; y) and (y; x). We assume that if an edge [x; y] fails, then arcs (x; y) and (y; x) both fail. A directed graph is a set of nodes N and a set of directed arcs A. Given a set of directed arcs, A, dene the reversal of A to be Ag. Similarly, given any directed graph to be the reversal of P . Let us consider that we have a two vertex (edge)-connected graph, or redundant graph i.e. removal of a vertex (edge) leaves the graph connected. Our method is based on the construction of a pair of directed spanning sub-graphs, each of which can be used for primary connections between any pair of nodes in the graph. In the event of a failure, connections on y are looped back around the failure using R. Similarly, connections on R are looped back around the failure using B. For instance, if G were a ring, then B and R would be the clockwise and counter-clockwise cycles around the ring. To see how loop-back operates in a general mesh network, consider rst the case where an edge fails. Assume (w; y) and (x; z) are arcs of R and that the shortest loop-back path around [x; y] is node x ! node z ! ! node w ! node y. We create two looping arcs, Bloop x;z and Rloop y;w . Bloop x;z is created at node x by attaching the tail of (z; x) 2 A to the head of (x; z) 2 A so that signals which arrive for transmission on (x; y) in B are now looped back at x to R. Similarly, Rloop y;w is created by attaching the tail of (w; y) 2 A to the head of (y; w) 2 A, so that any signal which arrives for transmission on (w; y) in R is looped back to B at y. Figure 2 illustrates our loop-back example. Edge [x; y] can be successfully bypassed as long as there exists a working path with su-cient capacity from x to y in R and a working path with su-cient capacity from y to x in B. Let us consider that we have an edge-redundant undirected graph E). We seek a directed spanning sub-graph G, and its associated reversal connected, i.e. there is a directed path in B from any node to any other node. condition 2: (i; A. Since R is connected i B is connected, condition 1 insures that any connection can be routed on B or R. Condition 1 also ensures that loop-back can be performed. Suppose edge [x; y] fails. Also suppose without loss of generality that (x; y) is an arc of B. In order to eect loop-back, we require that there exist a path from x to y in Rn(y; x) and a path from y to x in Bn(x; y). Such paths are guaranteed to exist because B and R are connected and such paths obviously do not traverse (x; y) or (y; x). Hence, connectivity is su-cient to guarantee loop-back connectivity in the event of an edge failure. Since condition 2 implies that (i; j) cannot be an arc of B and R, condition 2 ensures that loop-back connections on R do not travel over the same arc as primary connections on B, and vice-versa. Therefore, any algorithm which builds a graph B with properties 1 and 2 will su-ce. The algorithm presented below is one such algorithm. We start by choosing an arbitrary directed cycle G with at least 3 nodes (k 3). Such a cycle is guaranteed to exist if G is edge-redundant. If this cycle does not include all nodes in the graph, we then choose a new directed path or cycle that starts and ends on the cycle and passes through at least one node not on the cycle. If the new graph does not include all nodes of the graph, we again construct another directed path or cycle, starting on some node already included, passing through one or more nodes not included, and then ending on another already included node. The algorithm continues to add new nodes in this way until all nodes are included. We now formally present the algorithm followed by the proof of its correctness. ALGORITHM FOR SELECTING DIRECTIONS TO RECOVER FROM NODE FAILURES 1. 2. Choose any cycle (c in the graph with k 3. 3. 4. If N 5. 6. Choose a path or cycle pc ) such that x j;0 and such that the other vertices, x j;i , 1 i L j are chosen outside of N j 1 . For a path, we require x j;0 6= x j;L j and A. For a cycle, we require L j < 3 and x 7. 8. Go to step 4. We rst show that the algorithm for the edge-redundant case terminates if the graph is two edge- connected, i.e. edge-redundant. We shall proceed by contradiction. The algorithm would fail to terminate correctly i, at step 6, no new path or cycle pc j could be found but a vertex in N was not included in N j 1 . We therefore assume, for the sake of contradiction, that such a vertex exists. Because the graph is connected, there is an edge which connects some x in N j 1 to some y in NnN j 1 . Because the graph is edge-redundant, there exists a path between x and y which does not traverse e. Let be the last edge from which this path exits N . Note that w and x can be the same, but if y. Similarly, y and z may be the same, but then x 6= w. Now, there exists a path from x to w, passing through y, which would be selected at step 6. Thus, we have a contradiction. It is easy to see that Condition 2 is satised. Note that if (i; j) is already included in the directed sub-graph, then Step 6 ensures that (j; i) cannot be added. Therefore, all that remains to be shown is that B is connected. We use induction on the sub-graphs obviously connected. Indeed, B 1 is an unidirectional ring. Assume B j 1 is connected, for 2. We need to show for all x; y 2 N j , there is a directed path from x to y in B j . There are 4 cases: (1) x, y Case 1 follows from the induction hypothesis and the fact that A j is a superset of A j 1 . For case 2, we have that x; y 2 pc j . Pick vertices l and k such that x i.e. y comes after x on the path is a path from x to y in . If l < k, i.e. y comes before x on the path ), then there exists a path from x to x j;L j on pc j and a path from x j;0 to y on pc j . If x , then (x; x is a path from y to x in B j . If x j;0 6= x j;L j , then, by the induction hypothesis, there exists a path p(x j;L j from x j;L j to x j;0 in B j 1 and hence on B j . Therefore, (x; x a path from x to y. For case 3, we have x 2 pc j , y 2 N j 1 . Pick k such that x by the induction hypothesis, there exists a path from x j;L j to y. Vertex x is therefore connected to y since there is a path from x to x j;L j on pc j . For case 4, we have that y 2 pc j , x 2 N j 1 . There is a path from x to x j;0 by the induction hypothesis, and from x j;0 to y on pc j . A very simple network management protocol will enable recovery using our choice of directions created by the above algorithm. When recovering from an arc failure on the primary (secondary) digraph, the protocol need only broadcast on the secondary (primary) digraph. Each node retains only the rst copy of the broadcast and releases all unnecessary connections created by the broadcast. This simple concept is embedded in the protocol presented in Section 2.3. 2.2 Generalized Loop-back for Recovery from Node Failures While the previous section dealt with recovery from link failures, we consider in this section the event where a node fails. Note that the failure of a node entails the failure of all links incident upon the failed node. The failure of a node therefore requires dierent techniques than those used to deal with link failures. Let us rst overview the operation of loop-back in a mesh network when there is failure of a node. Each node connected by a link to the failed node, i.e. adjacent to the failed node, independently performs loop-back in the same manner as if the link connecting the failed node to the looping node had failed. We assume that only one primary connection per wavelength is incident upon each node but that there may be several outputs along one wavelength per node. Thus, we allow the use of multicasting at nodes. The purpose of our restriction on the connections through a node is to ensure that, after loop-back, there are no collisions in the back-up graph. Multicasting applications are particularly attractive for WDM networks, because splitting at optical nodes oers a simple and eective way of performing multicasting. Note that two types of tra-c are looped back: tra-c destined for the failed node and tra-c which only traversed the failed node. Let us rst consider the rst type of tra-c in the case where a node, say j, performs loop-back on the link between j and node k, the failed node. Node j receives on a back-up channel tra-c intended for node k. Only two cases are possible: either link [j; k] failed but node k is still operational or node k failed. Note that we have made no assumption regarding the ability of the network management system to distinguish between the failure of a node and the failure of a link. Indeed, the nodes may only be aware that links have ceased to function, without knowing whether the cause is a single link failure or a node failure. Since we have a node-redundant network, our loop-back mechanism can recover from failure of node j, which entails failure of link [j; k]. Hence, even if there has been failure of link [j; k] only, node j can eliminate all back-up tra-c destined to node k, because the back-up mechanism ensures that back-up tra-c destined for node k arrives to node k even after failure of node j. If node k failed, then eliminating back-up tra-c destined for k will prevent such back-up tra-c from recirculating in the network, since recirculation would cause collisions and congestion. Thus, regardless of whether a node failure or a link failure occurred, back-up tra-c destined for the failed node will be eliminated when a node adjacent to the failed node receives it. In SONET SHRs, a similar mechanism eliminates tra-c intended for a failed node. We may now illustrate our mechanism with a specic example applied to the network we have been considering. Figure 3 shows a sample set of directions which can be selected for generalized loop-back recovery from node failure. Let us rst consider the case where we have a primary connection along the full line from node 1 to node 3 via node 2 and node 2 fails. The shortest loop-back path is node 3. Let us now consider the case where we have a primary connection along the full line from node 1 to node 2 and node 2 fails. Then, the back-up path goes from node 8 to node 7, which eliminates the connection, because node 7 is adjacent to node 2. We model our network as a vertex-redundant undirected graph E). We seek a directed spanning sub-graph G, and its associated reversal Condition 1: B is connected, i.e. there is a directed path in B from any node to any other node. Condition 2: (i; A. Condition 3: For all x; n; y 2 N such that (x; n), (n; y) are arcs of B, there exists a directed path from x to y in R which does not pass through n. As in the edge-redundant case, Condition 1 insures that any connection can be routed on B or R. However, unlike the edge-redundant case, connectivity is insu-cient to guarantee loop-back connectivity after failure. Also as in the edge-redundant case, Condition 2 insures that loop-back connections on R do not travel over the same arc as primary connections on B, and vice-versa. Condition 3 insures that loop-back can be successfully performed and is equivalent to the statement that all 3 adjacent nodes in B, x, n, y are contained in a cycle of B. We perform loop-back for node failures in the same manner as described above for link failures. For instance, let us select two distinct nodes w and z. Let p 1 be the path in B, i.e. the path traversed over 1 on ber 1, from w to z and let n be a node other than w or z traversed by p 1 . We consider the nodes x and y such that (x; n) and (n; y) are traversed in that order in p 1 . Thus, (x; n), (n; y) are in A. Let p 2 be a path in R which does not include vertex n and which goes from vertex x to vertex y. We perform loop-back from w to z using paths p 1 , p 2 at node n by traversing the following directed circuit: from w to x, we use path at x, we loop-back from primary to secondary from x to y, we use path at y, we loop-back from secondary to primary from y to z, we use path p 1 . As discussed previously, this loop-back is more general than the type of loop-back used in a ring. In particular, the loop-back is not restricted to use a back-haul route traversing successively In order to guarantee loop-back, it is su-cient to select B and R so that, in the event of any vertex (edge) failure aecting B or R, there exists a working path around the failure on the other sub-graph. Any sub-graph with Conditions 1-3 is su-cient to perform loop-back as described above. The algorithm below guarantees these conditions by amending the algorithm for the edge-redundant case. The edge-redundant algorithm fails to insure Condition 3 for two reasons. The rst reason is that cycles are allowed in Step 6, i.e. pc possible in iteration j, and hence failure of node x j;0 would leave both B and R disconnected. The second and more fundamental reason is that the ordering of the nodes on the added paths in steps 6 and 7 is very unrestrictive. Our algorithm starts by choosing a directed cycle of at least 3 vertices containing some arbitrary s]. If this cycle does not include all nodes in the graph, we then choose a directed path that starts on some node in the cycle, passes through some set of nodes not on the cycle, and ends on another node on the cycle. If the cycle and path above do not include all vertices of the graph, we again construct another directed path, starting on some node already included, passing through one or more nodes not included, and then ending on another already included node. The algorithm continues to add new nodes in this way until all nodes are included. It is simple to show that, in a vertex-redundant graph, for any edge e, a cycle with 3 vertices must exist containing e. It can also be seen that, for any such cycle, a path can be added as above, and subsequent paths can be added, in arbitrary ways, until all nodes are included. It is less simple to choose the direction of the added paths, and hence the B and R directed sub-graphs. The technique we present relies in part on results presented [MFBG99, MFB97, FMB97, MFGB98]. We now present the algorithm followed by the proof of its correctness. ALGORITHM FOR SELECTING DIRECTIONS TO RECOVER FROM NODE FAILURES 1. 1. Pick an arbitrary edge 2. (a) Choose any cycle (s; c in the graph with k 2. (b) Order these nodes by assigning values such that 3. 4. If N 5. 6. (a) Choose a path 2, in the graph such that x ). The other vertices, x j;i , 1 i < L j , are chosen outside of (b) Order the new vertices by assigning values such that v(x j;0 7. 8. Go to step 4. Note in step 6b that v max v(x j;L j We rst show that the algorithm for the node-redundant case terminates if the graph is vertex-redundant. We shall proceed by contradiction. The algorithm would fail to terminate correctly, i at step 6 no new path p j could be found but a vertex in N was not included in N j 1 . We assume for the sake of contradiction that such a vertex exists. Because of the graph is connected, there is an edge [x; y] which connects some x in N j 1 to some y in NnN j 1 . Pick a vertex q 2 N j 1 , such that q 6= x. Because the graph is node-redundant, there exists a path between y and q which does not use x. Let be the last edge from which this path exits . Note that possible. Now, there exists a path from x to w, passing through y, which would be selected at step 6 in the algorithm. Therefore, we have a contradiction. We now prove that B satises Conditions 1-3. The fact that B is connected follows by induction on j using almost identical arguments as used in the proof for the link-redundant case. In particular, we can see by induction on j that there is a directed path in B j from x 2 N j to any y 2 N j . Since these properties hold for each j, they also hold for the nal directed sub-graph B. We may therefore state that B is connected. As in the edge-redundant case, Condition 2 is satised by the restrictions on adding new arcs. Finally, we prove that B satises Condition 3. We need to prove the fact that, for all x; n; y 2 N such that (x; n), (n; y) are arcs of B, there exists a directed path from x to y in R which does not pass through n. Since R is the reversal of B, we can prove the equivalent statement that there exists a directed path from y to x in B which does not pass through n. The key observation is to note that has a special property. In particular, it is the only arc in B for which the value of the originating node is lower than the value of the terminating node, i.e. all (i; From this property it immediately follows that all directed cycles in B contain (t; s). To see this, let x be a cycle and note that, if (t; s) were not traversed in this cycle, then v(x could not be an arc in B. Also, since B is connected, we also have that (t; s) is the unique arc into s in B for, otherwise, we could construct a cycle through s which did not pass through t. Only two cases need to be considered to prove the desired property: First consider is connected, there exists a path from y to x in B and this path need not include s, since the only way to reach s is through t. Now consider n 6= s. There exists paths n) from y to n and p(n; from n to x, both in B. is a cycle, it includes s. Similarly, (n; xm ; xm n) is a cycle and hence includes s. Therefore, there is a path starting at y, proceeding on p(y; n) until s (which is before n in p(y; n)), starting in p(n; x) at s (which is after n in p(n; x) ), and ending at x. 2.3 Protocol We now overview a protocol which ensures proper recovery, using generalized loopback for node or link recovery. Our protocol is more involved than that need to recover only from a link failure, since we must contend with the failure of all links adjacent to the failed node. However, our algorithm will also operate properly for link failures without node failures. Our protocol uses negative acknowledgements and labels, to establish correct rerouting. The signaling for the protocol may be performed over an out-of-band control channel, or any in-band control channel scheme, such as a subcarrier multiplexed signal. Consider the failure of a primary ber from x to y. Failure of the ber may be due to failure of the ber itself or of node y. When x detects the failure, it writes "y" into the failure label and loops the primary stream back into the back-up digraph, splitting it across all outgoing arcs in the back-up digraph. As the tra-c enters each new node, the node forwards the tra-c, again splitting it over all outgoing arcs. Backup bers leaving a node can be pre-congured to split an incoming stream, shortening the time required to ood failure information across outgoing links. For nodes with only one incoming stream, the route is fully pre-planned, and no tra-c is lost during the decision process. For nodes with more than one incoming stream, the rst of the streams to receive tra-c is chosen for forwarding. A stream that becomes active after the rst- typically owing to tra-c from the same failure arriving via a dierent route- is dropped, and a negative acknowledgement (NACK)is returned on the reverse back-up arc. A node that receives a NACK on an outgoing link ceases to forward tra-c on that link. If all outgoing links for a node are NACKed, the node propagates a NACK on its own incoming link, in eect releasing the connection on that link. If all outgoing links at x are NACKed, recovery has failed (possibly multi-failure scenarios or scenarios where several connections over the same wavelength were present at a failed node). The NACK-based protocol can be extended with hop-count and signal-power (splitting) restrictions to reduce the area over which a failure propagates, but such restrictions require more careful selection of the back-up digraph to guarantee recovery from all single failures and to prevent signicant degradation of multi-failure recovery possibilities. The use of NACKs serves to limit the use of back-up arcs to those necessary to recovery. Another approach to achieving this goal is to mark the successful route and forward tear-down messages down all other arcs. The NACK scheme is superior to this approach in two ways. First, tear-down messages must catch up with the leading edge of the tra-c, but cannot travel any faster. In the worst case, a route is torn down only to allow a cyclic route to recreate itself, resulting in long-term tra-c instabilities in the back-up digraph. To avoid this possibility, tear-down requires that nodes remember the existence of failure tra-c between the time that they tear down the route and the time that the network settles (a global phenomenon). A second point in favor of the NACK-based scheme is that it handles multicast (possibly important for node failures) naturally by discarding only unused routes. A tear-down scheme must know the number of routes to recover in advance or discover it dynamically; the rst option requires a node to know the routes through its downstream neighbors, while the second option is hard because of timing issues (when have all routes been recovered?). Meanwhile, y detects the loss of the stream and begins listening for tra-c with its name or x's name on the back-up stream. The second case handles failure of node x, which results in ooding of tra-c destined for x. Note that tra-c for x can also be generated owing to failure of a primary arc ending at x, but, in such a case, y does not detect any failure and does not listen for back-up tra-c. Once a stream with either name is received, it is spliced into the primary tra-c stream, completing recovery. Other paths are torn down through NACKs. Note that if a stream ends at a failed node x, no node listens for the back-up tra-c for node x, and all connections carrying back-up tra-c for node x are eventually torn down. While our protocol for node failure is more complicated than that for link failure, it is still relatively simple. Node failure in ring-based systems is a very complex operation whenever a node is on more than one ring. For double cycle cover, node recovery requires hopping among rings, and thus necessitates a centralized controller with global knowledge of the network. Even for simple double-homed SONET rings, node recovery involves the use of matched nodes. Proper operation of matched nodes requires signicant inter-ring signaling as well as dual-fed path protection between rings. Applications 3.1 Choice of Routings In this Section, we present heuristic algorithms for selecting directions in the back-up graph. We seek to select directions in such a way to avoid excessive length for back-up paths. We consider three dierent algorithms. The rst algorithm, which we term Heuristic 1, rst nds, for each link, a shortest loop which includes that link. A loop is a directed cycle, thus a shortest loop is a directed cycle with the minimum number of links. We order the shortest loops of all the links in ascending order of length. Shortest loops with equal lengths are ordered arbitrarily with respect to each other. Beginning with the rst shortest loop, in ascending order, we assign, whenever possible, directions according to the directions of the arcs along the shortest loop. The second algorithm, Heuristic 2, also relies on considering shortest loops but takes into account the fact that the choice of direction on a link may aect other links. We create a heuristic measure of this eect, which we call associate number (AN). The AN of link is the number of dierent shortest loops which pass through that link. In particular, the AN can help us distinguish among links with equal length shortest loops. We order the links in ascending order of AN. We begin, as for Heuristic 1, by nding, for each link, a shortest loop which includes that link. Links with equal ANs are ordered arbitrarily with respect to each other. Beginning with the rst link and progressing in ascending order we assign directions, whenever possible, according to the shortest loop of the link being considered. The last algorithm we consider is a random assignment of directions. While the number of possible directions is exponential in the number of links, we may signicantly reduce that number by requiring that the directions be feasible. We apply our algorithms to three networks, NJ LATA, LATAX and ARPANET, shown in gures 4, 5 and 6. We consider the maximum length of a back-up path and the average length. Table 1 shows the results obtained from running the dierent algorithms for the three networks we consider. times for each network and the best result was kept for the maximum and the average. Note that the same choice of directions did not always yield both the best maximum and the best average. Heuristic 2 was run in the same way as Heuristic 1. For the random algorithm, we limited ourselves to 52 runs for NJ LATA, 128 runs for runs for ARPANET. The best maximum and the best average were chosen in each case. Comparing the running time of running the times versus the above number of times for the random algorithm, we obtain that Heuristic 1 yielded a run time improvement of 72 %, 90 %, 88 % over random choice of directions for NJ LATA, LATAX and ARPANET, respectively. Heuristic 2 yielded a run time improvement of 73 %, 91 %, 90 % over random choice of directions for NJ LATA, LATAX and ARPANET, respectively. From our simulations, Heuristic 2 was slightly better than Heuristic 1 in terms of run time and average back-up length. Table 1: Comparison of the best results between heuristic algorithms and the method of selecting directions randomly Random Heuristic 1 Heuristic 2 Max. Avg. Max. Avg. Max. Avg. 3.2 WDM-based Loop-back Recovery In ber-based restoration, the entire tra-c carried by a ber is backed by another ber. In ber- based restoration it does not matter whether the system is a WDM system. If tra-c is allowed in both directions in a network, ber-based restoration relies on four bers, as illustrated in Figure 7. In WDM-based recovery, restoration is performed in a wavelength-by-wavelength basis. WDM-based recovery requires only two bers, although it is applicable to a higher number of bers. Figure 8 illustrates WDM-based recovery. A two-ber counter-propagating WDM system can be used for WDM-based restoration, even if tra-c is allowed in both directions. Note that WDM restoration as shown on Figure 8 does not require any change of wavelength. Thus, tra-c initially carried by 1 is backed up by the same wavelength. Obviating the need for wavelength changing is economical and e-cient in WDM networks. One could, of course, back up tra-c from 1 on ber ber 2, if there were advantages to such wavelength changing, for instance in terms of wavelength assignment for particular tra-c patterns. We can easily extend the model to a system with more bers, as long as the back-up for a certain wavelength on a certain ber is provided by some wavelength on another ber. Moreover, we may change the ber and/or wavelengths from one ber section to another. For instance, the back-up to 1 on ber 1 may be 1 on ber 2 on a two-ber section and 2 on ber 3 on another section with four bers. Note, also, that we could elect not to back up 1 on ber 1 and instead use 1 on ber 1 for primary tra-c. The extension to systems with more bers, inter-wavelength back-ups and heterogeneous back-ups among ber sections can be readily done. There are several advantages to WDM-based recovery systems over ber-based systems. The rst advantage is that, if bers are loaded with tra-c at half of total capacity or less, then only two bers rather than four are needed to provide recovery. Thus, a user need only lease two bers, rather than paying for unused bandwidth over four bers. On existing four-ber systems, bers could be leased by pairs rather than fours, allowing two leases of two bers each for a single four-ber system. The second advantage is that, in ber based-systems, certain wavelengths may be selectively given restoration capability. For instance, half the wavelengths on a ber may be assigned protection, while the rest may have no protection. Dierent wavelengths may thus aord dierent levels of restoration QoS, which can be re ected in pricing. In ber-based restoration, all the tra-c carried by a ber is restored via another ber. If each ber is less than half full, WDM-based loop-back can help avoid the use of counterpropagating wavelengths on the same ber. Counterpropagating wavelengths on the same ber are intended to enable duplex operation in situations which do not require a full ber's worth of capacity in each direction and which have scarce ber resources. However, counterpropagation on the same ber is onerous and reduces the number of wavelengths that a ber can carry with respect to unidirectional propagation. Our WDM-based loop-back may make using 2 unidirectional bers preferable to using counterpropagating bers, where one ber is a back-up for the other. We may now draw a comparison between generalized loop-back and double cycle covers for WDM-based loop-back recovery. The ability to perform restoration over mesh topologies was rst introduced in [ESH97, ES96]. In particular, [ESH97] considers link failure restoration in optical networks with arbitrary two-link redundant arbitrary mesh topologies and bi-directional links. The scheme relies on applying methods for double cycle covers to restoration. Let us rst discuss how double-cycle ring covers can be used to perform recovery. A double cycle ring cover covers a graph with cycles in such a way that each edge is covered by two cycles. Cycles can then be used as rings to perform restoration. Each cycle corresponds either to a primary or a secondary two-ber ring. For two-edge connected planar graphs, a polynomial-time algorithm exists for creating double cycle covers. For two-edge connected non-planar graphs, the fact that double cycle covers exist is a conjecture, thus no algorithm except for exhaustive search exists. In this subsection, we consider an example network and its possible double cycle covers. On the basis of these double cycle covers, we discuss whether double cycle covers can be used in the context of WDM loop-back, described in the previous section. Let us consider a link covered by two rings, rings 1 and 2. If we assign a direction to ring 1 and the opposite direction to ring 2, then ring-based recovery using the double cycle cover uses ring 2 to back up ring 1. In eect, this recovery is similar to recovery in conventional SHRs, except that the two rings which form four-ber SHRs are no longer co-located over their entire length. Figure 9 shows the two possible double cycle covers, shown in thin lines, for a certain ber topology, shown in bold lines. In the case of four ber systems, with two bers in the same direction per ring, we have ber-based recovery, because bers are backed up by bers. For the type of WDM-based loop-back we consider in this section, each ring is both primary for certain wavelengths and secondary for the remaining wavelengths. For simplicity, let us again consider just two wavelengths. Figure 10 and 11 show that we cannot use one ring to provide WDM-based loop-back back-up for another unless we perform wavelength changing and add signicant network management. We cannot assign primary and secondary wavelengths in such a way that a wavelength is secondary or primary over a whole ring. We may point out another drawback of the back-up paths aorded by double cycle covers. For both ring covers shown in Figure 9, some links are covered by rings which are not of the same length. For instance, in Figure 10, a break on a link may cause one direction to be backed up on ring 1, while another direction may be backed up on ring 4. Thus, the back-up may traverse only three links along ring 1 in one direction and seven links along 4 in the other direction. The two directions on a link will therefore have dierent delays in their restoration time and incur dierent timing jitters. Such asymmetry in the propagation for the back-up path does not occur in SHRs or in generalized loop-back, since the back-up paths for both directions traverse the same links. 3.3 Plurality of Back-up Routes for Generalized Loop-back We have mentioned that our algorithm can be used to perform recovery even when there is a change in the conditions of the networks. In this section, we give a brief example of how such exibility is aorded. Figure 12 illustrates our example. We have single network, with a recovery sub-graph built for link failure restoration. In case of failure of the link between node 2 and node 3, the recovery back-up path is shown by a curving directed line. The shortest back-up path uses the link between nodes 9 and 10, as shown on Figure 12. Suppose that the link between nodes 9 and 10 becomes unusable for the back-up path, for instance because the link has failed or because all the wavelengths on that link are used to accommodate extra tra-c. Then, the back-up path for a failure between nodes 2 and 3 can be the path node 3 shown by a curved line on Figure 12.b. Thus, the same back-up sub-graph can be used both when the link between nodes 9 and 10 is available and when it is not available. Note that not all links can be allowed to become unavailable. If the link between nodes 3 and 10 becomes unavailable, the restoration after failure of the link between nodes 2 and 3 is not possible. However, it is possible to determine whether certain links are necessary for recovery in the case of failure of other links. Since there are two paths in the back-up sub-graph from node 10 to node 9, the link between nodes 9 and 10 is not necessary and that link can be freed up to carry extra tra-c, if the need arises. Recent results [MLT00] have shown that signicant savings, of the order of 25 %, can be achieved using generalized loop-back over a variety of networks without sacricing the length of the longest back-up and the ability to recover from double failures. 4 Summary and Conclusions We have presented generalized loop-back, a novel way of implementing loop-back on mesh networks without using ring-based schemes. We have established routings and protocols to ensure recovery after a link or a node failure. Our method is guaranteed to work in polynomial time regardless of whether the graph representation of the network is planar or not, whereas double cycle covers have polynomial-time solution only for planar graphs. The gist of our method is to assign directions to bers and the wavelengths traveling through them. Our method allows great exibility in planning the conguration of a network, as long as it has redundancy, while providing the bandwidth utilization advantages typically associated with loop-back protection in SHRs. Recovery, as in SONET BLSR, is performed by the nodes adjacent to the link or node failure. Moreover, our loop-back recovery method does not require the nodes performing loop-back to distinguish between a node and a link failure. We have shown that simple heuristic algorithms yield satisfactory results in terms of average and maximum length of back-up paths. We have compared our method to the previously known method for loop-back for link failure on mesh networks. That method ([ESH97]) is based upon double cycle covers and we have shown that such a method may not be applied to WDM-based loop-back systems. Moreover, we have shown by a simple example that generalized loop-back allows recovery to be performed in a bandwidth-e-cient manner. There are several areas of further work. One of them is considering the issue of wavelength assignment jointly with back-up considerations, whether the back-up be loop-back, APS or hybrid. Another issue is the determination of the back-up path. Broadcasting, or ooding, in the back-up wavelength causes that wavelength to be unavailable for other uses in parts of the network which are not required to provide back-up. Some methods for determining back-up paths are presented in [FMB98]. Another area for further research is the use of generalized loop-back to perform bandwidth-e-cient recovery. As we discussed in Section 2.1, link and node restoration generally are less e-cient, in terms of capacity utilization, than even-triggered path restoration. However, our scheme allows recovery of links which are not included in the back-up sub-graph, as long as the end nodes are included in the back-up sub-graph. This operation can be viewed as being akin to p-cycles, but with greater exibility in the choice of the back-up sub-graph. Eliminating links from the back-up sub-graph is economical in bandwidth but entails some degradation in terms of other performance metrics, such as length of back-up path or recovery from two failures. Preliminary results show that signicant savings in terms of bandwidth utilization can be achieved without appreciable aecting other performance metrics ([MLT00]). --R A distributed link restoration algorithm with robust preplanning. Performance analysis of fast distributed network restoration algorithms. An architecture for e-cient survivable networks Protection planning in transmission networks. A multi-layer restoration strategy for recon gurable networks A fast distributed network restoration algorithm. Spare capacity assignment for di Comparison of capacity e-ciency of DCS network restoration routing techniques Automatic protection switching for link failures in optical networks with bi-directional links Link failure restoration in optical networks with arbitrary mesh topologies and bi-directional links Covering graphs by cycles. Longest cycles in 2-connected graphs of independence number Optimal spare capacity design for various protection switching methods in ATM networks. Double search self-healing algorithm and its characteristics Near optimal spare capacity placement in a mesh restorable network. Techniques for Dynamic bandwidth-allocation and path- restoration in SONET self-healing networks A girth requirement for the double cycle cover conjecture. The selfhealing TM network. Case studies of survivable ring An optimal spare-capacity assignment model for survivable networks with hop limits The hop-limit approach for spare-capacity assignment in survivable networks Dynamic recon Covering graphs with simple circuits. Hamilton cycles in regular 2-connected graphs A survey of the double cycle cover conjecture. Topologocial optimization of a communication network subject to a reliability constraint. Veerasamy and J. Distributed control algorithms for dynamic restoration in DCS mesh networks: Performance evaluation. An ATM VP-based self-healing ring ATM virtual path self-healing based on a new path restoration protocol A bandwidth e-cient self-healing ring for B-ISDN A dynamic recon Design of survivable communication networks under performance constraints. Survivable WDM mesh networks Sums of circuits. Interconnection of self-healing rings An algorithm for survivable network design employing multiple self-healing rings Distributed self-healing control in SONET Service application for SONET DCS distributed restoration. Design of Survivable Networks. Survivable network planning methods and tools in taiwan. A capacity comparison for SONET self-healing ring networks Polyhedral decomposition of cubic graphs. On the complexity of Two strategies for spare capacity placement in mesh restorable networks. An algorithm for designing rings for survivable Feasibility study of a high-speed SONET self-healing ring architecture in future intero-ce networks A multi-period design model for survivable network architecture selection for SDH/SONET intero-ce networks Strategies and technologies for planning a cost-eective survivable network architecture using optical switches Survivable network architectures for broad-band ber optic networks: Model and performance comparison Backup vp preplanning strategies for survivable multicast ATM. Maximal circuits of graphs ii. Fiber Network Service Survivability. A passive protected self-healing mesh network architecture and applications A novel passive protected SONET bidirectional self-healing ring architecture Restoration strategies and spare capacity requirements in self-healing ATM networks Fitness: Failure immunization technology for network service survivability. An improvement of jackson's result on hamilton cycles in 2-connected graphs --TR Covering graphs by cycles A passive protected self-healing mesh network architecture and applications On the Complexity of Finding a Minimum Cycle Cover of a Graph Optimal capacity placement for path restoration in STM or ATM mesh-survivable networks Restoration strategies and spare capacity requirements in self-healing ATM networks Redundant trees for preplanned recovery in arbitrary vertex-redundant or edge-redundant graphs Fiber Network Service Survivability Spare Capacity Assignment in Telecom Networks Using Path Restoration --CTR Timothy Y. Chow , Fabian Chudak , Anthony M. Ffrench, Fast optical layer mesh protection using pre-cross-connected trails, IEEE/ACM Transactions on Networking (TON), v.12 n.3, p.539-548, June 2004 Mansoor Alicherry , Randeep Bhatia, Simple pre-provisioning scheme to enable fast restoration, IEEE/ACM Transactions on Networking (TON), v.15 n.2, p.400-412, April 2007 Canhui Ou , Laxman H. Sahasrabuddhe , Keyao Zhu , Charles U. Martel , Biswanath Mukherjee, Survivable virtual concatenation for data over SONET/SDH in optical transport networks, IEEE/ACM Transactions on Networking (TON), v.14 n.1, p.218-231, February 2006

mesh networks;WDM;loop-back;network restoration

506843

A mutual exclusion algorithm for ad hoc mobile networks.

A fault-tolerant distributed mutual exclusion algorithm that adjusts to node mobility is presented, along with proof of correctness and simulation results. The algorithm requires nodes to communicate with only their current neighbors, making it well-suited to the ad hoc environment. Experimental results indicate that adaptation to mobility can improve performance over that of similar non-adaptive algorithms when nodes are mobile.

Introduction A mobile ad hoc network is a network wherein a pair of nodes communicates by sending messages either over a direct wireless link, or over a sequence of wireless links including one or more intermediate nodes. Direct communication is possible only between pairs of nodes that lie within one another's transmission radius. Wireless link \failures" occur when previously communicating nodes move such that they are no longer within transmission range of each other. Like- wise, wireless link \formation" occurs when nodes that were too far separated to This is an extended version of the paper presented at the Dial M for Mobility Workshop, Dallas TX, Oct. 30, 1998. Supported by GE Faculty of the Future and Dept. of Education GAANN fellowships. Supported in part by NSF PYI grant CCR-9396098 and NSF grant CCR-9972235. Supported in part by Texas Advanced Technology Program grant 010115-248 and NSF grants CDA-9529442 and CCR-9972235. communicate move such that they are within transmission range of each other. Characteristics that distinguish ad hoc networks from existing distributed networks include frequent and unpredictable topology changes and highly variable message delays. These characteristics make ad hoc networks challenging environments in which to implement distributed algorithms. Past work on modifying existing distributed algorithms for ad hoc networks includes numerous routing protocols (e.g., [8,9,11,13,16,18,19,22{24]), wireless channel allocation algorithms (e.g., [14]), and protocols for broadcasting and multicasting (e.g., [8,12,21,26]). Dynamic networks are xed wired networks that share some characteristics of ad hoc networks, since failure and repair of nodes and links is unpredictable in both cases. Research on dynamic networks has focused on total ordering [17], end-to-end communication, and routing (e.g., [1,2]). Existing distributed algorithms will run correctly on top of ad hoc routing protocols, since these protocols are designed to hide the dynamic nature of the network topology from higher layers in the protocol stack (see Figure 1(a)). Routing algorithms on ad hoc networks provide the ability to send messages from any node to any other node. However, our contention is that eciency can be gained by developing a core set of distributed algorithms, or primitives, that are aware of the underlying mobility in the network, as shown in Figure 1(b). In this paper, we present a mobility aware distributed mutual exclusion algorithm to illustrate the layering approach in Figure 1(b). User Applications Distributed Primitives Routing Protocol Distributed Primitives Routing Protocol Ad Hoc Network User Applications Ad Hoc Network (b) (a) Figure 1. Two possible approaches for implementing distributed primitives The mutual exclusion problem involves a group of processes, each of which intermittently requires access to a resource or a piece of code called the critical section (CS). At most one process may be in the CS at any given time. Providing shared access to resources through mutual exclusion is a fundamental problem in computer science, and is worth considering for the ad hoc environment, where stripped-down mobile nodes may need to share resources. Distributed mutual exclusion algorithms that rely on the maintenance of a logical structure to provide order and eciency (e.g., [20,25]) may be inecient when run in a mobile environment, where the topology can potentially change with every node movement. Badrinath et al.[3] solve this problem on cellular mobile networks, where the bulk of the computation can be run on wired portions of the network. We present a mutual exclusion algorithm that induces a logical directed acyclic graph (DAG) on the network, dynamically modifying the logical structure to adapt to the changing physical topology in the ad hoc environment. We then present simulation results comparing the performance of this algorithm to a static distributed mutual exclusion algorithm running on top of an ad hoc routing protocol. Simulation results indicate that our algorithm has better average waiting time per CS entry and message complexity per CS entry no greater than the cost incurred by a static mutual exclusion algorithm running on top of an ad hoc routing algorithm. The next section discusses related work. In Section 3, we describe our system assumptions and dene the problem in more detail. Section 4 presents our mutual exclusion algorithm. We present a proof of correctness and discuss the simulation results in Sections 5 and 6, respectively. Section 7 presents our conclusions. 2. Related Work Token based mutual exclusion algorithms provide access to the CS through the maintenance of a single token that cannot simultaneously be present at more than one node in the system. Requests for CS entry are typically directed to whichever node is the current token holder. Raymond [25] introduced a token based mutual exclusion algorithm in which requests are sent, over a static spanning tree of the network, toward the token this algorithm is resilient to non-adjacent node crashes and recoveries, but is not resilient to link failures. Chang et al.[7] extend Raymond's algorithm by imposing a logical direction on a sucient number of links to induce a token oriented DAG in which, for every node i, there exists a directed path originating at i and terminating at the token holder. Allowing request messages to be sent over all links of the DAG provides resilience to link and site failures. However, this algorithm does not consider link recovery, an essential feature in a system of mobile nodes. Dhamdhere and Kulkarni [10] show that the algorithm of [7] can suer from deadlock and solve this problem by assigning a dynamically changing sequence number to each node, forming a total ordering of nodes in the system. The token holder always has the highest sequence number, and, by dening links to point from a node with lower to higher sequence number, a token oriented DAG is maintained. Due to link failures, a node i that wants to send a request for the token may nd itself with no outgoing links to the token holder. In this situation, oods the network with messages to build a temporary spanning tree. Once the token holder becomes part of such a spanning tree, the token is passed directly to node i along the tree, bypassing other requests. Since priority is given to nodes that lose a path to the token holder, it seems likely that other requesting nodes could be starved as long as link failures continue. Also, ooding in response to link failures and storing messages for delivery after link recovery make this algorithm ill-suited to the highly dynamic ad hoc environment. Our token based algorithm combines ideas from several papers. The partial reversal technique from [13], used to maintain a destination oriented DAG in a packet radio network when the destination is static, is used in our algorithm to maintain a token oriented DAG with a dynamic destination. Like the algorithms of [25], [7], and [10], each node in our algorithm maintains a request queue containing the identiers of neighboring nodes from which it has received requests for the token. Like [10], our algorithm totally orders nodes. The lowest node is always the current token holder, making it a \sink" toward which all requests are sent. Our algorithm also includes some new features. Each node dynamically chooses its lowest neighbor as its preferred link to the token holder. Nodes sense link changes to immediate neighbors and reroute requests based on the status of the previous preferred link to the token holder and the current contents of the local request queue. All requests reaching the token holder are treated symmetri- cally, so that requests are continually serviced while the DAG is being re-oriented and blocked requests are being rerouted. 3. Denitions The system contains a set of n independent mobile nodes, communicating by message passing over a wireless network. Each mobile node runs an application process and a mutual exclusion process that communicate with each other to ensure that the node cycles between its REMAINDER section (not interested in the CS), its WAITING section (waiting for access to the CS), and its CRITICAL section. Assumptions 1 on the mobile nodes and network are: 1. the nodes have unique node identiers, 2. node failures do not occur, 3. communication links are bidirectional and FIFO, 4. a link-level protocol ensures that each node is aware of the set of nodes with which it can currently directly communicate by providing indications of link formations and failures, 5. incipient link failures are detectable, providing reliable communication on a per-hop basis, and 6. partitions of the network do not occur. The rest of this section contains our formal denitions. We explicitly model only the mutual exclusion process at each node. Constraints on the behavior of the application processes and the network appear as conditions on executions. The system architecture is shown in Figure 2. We assume the node identiers are 0; 1. Each node has a (mutual exclusion) process, modeled as a state machine, with the usual set of states, some of which are initial states, and a transition function. Each state contains a local variable that holds the node identier and a local variable that holds the current neighbors of the node. The transition function is described in more detail shortly. Application Process Mutual Exclusion Process Network node i RequestCS EnterCS Figure 2. System Architecture Section 7 for a discussion of relaxing assumption 6. A conguration describes the instantaneous state of the whole system; for- mally, it is a set of n states, one for each process. In an initial conguration, each state is an initial state and the neighbor variables describe a connected undirected graph. A step of the process at node i is triggered by the occurrence of an input event. Input events are: the application process on node i requests access to the CS, entering its WAITING section. the application process on node i releases its access to the CS, entering its REMAINDER section. Recv i (j; m): node i receives message m from node j. receives notication that the link l incident on i is now up. LinkDown i (l): node i receives notication that the link l incident on i is now down. The eect of a step is to apply the process' transition function, taking as input the current state of the process and the input event, and producing as output a (possibly empty) set of output events and a new state for the process. Output events are: the mutual exclusion process on node i informs its application process that it can enter the CRITICAL section. Send i (j; m): node i sends message m to node j. The only constraint on the state produced by the transition function is that the neighbor set variable of i must be properly updated in response to a LinkUp or LinkDown event. are called application events, while are called network events. An execution is a sequence of the form C where the C k 's are congurations, the in k 's are input events, and the out k 's are sets of output events. An execution must end in a conguration if it is nite. A positive real number is associated with each in i , representing the time at which that event occurs. An execution must satisfy a number of additional conditions, which we now list. The rst set of conditions are basic \syntactic" ones. C 0 is an initial conguration. If in k occurs at node i, then out k and i's state in C k are correct according to i's transition function operating on in k and i's state in C k 1 . The times assigned to the steps must be nondecreasing. If the execution is innite, then the times must increase without bound. At most one step by each process can occur at a given time. The next set of conditions require the application process to interact properly with the mutual exclusion process. If in k is RequestCS i , then the previous application event at node i (if any) is If in k is ReleaseCS i , then the previous application event at node i must be The remaining conditions constrain the behavior of the network to match the informal description given above. First, we consider the mobility notication. occurs at time t if and only if LinkUp j (l) occurs at time t, where l joins i and j. Furthermore, LinkUp i (l) only occurs if j is currently not a neighbor of i (according to i's neighbor variable). The analogous condition holds for LinkDown. A LinkDown never disconnects the graph. Finally, we consider message delivery. There must exist a one-to-one and onto correspondence between the occurrences of Send j (i; m) and Recv i (j; m), for all i, j and m. This requirement implies that every message sent is received and the network does not duplicate or corrupt messages nor deliver spurious messages. Furthermore, the correspondence must satisfy the following: If Send i (j; m) occurs at some time t, then the corresponding Recv j (i; m) occurs at some later time t 0 , and the link connecting i and j is continuously up between t and t 0 . This implies that a LinkDown event for link l cannot occur if any messages are in transit on l. Now we can state the problem formally. In every execution, the following must hold: If in k is EnterCS i , then the previous application event at node i must be RequestCS i . I.e., CS access is only given to requesting nodes. Mutual Exclusion: If in k is EnterCS i , then any previous EnterCS j event must be followed by a ReleaseCS j prior to in k . there are only a nite number of LinkUp i and LinkDown i events, then if in k is RequestCS i , then there is a following EnterCS i . For the last condition, the hypothesis that link changes cease is needed because an adversarial pattern of link changes can cause starvation. 4. Reverse Link (RL) Mutual Exclusion Algorithm In this section we rst present the data structures maintained at each node in the system, followed by an overview of the algorithm, the algorithm pseudocode, and examples of algorithm operation. Throughout this section, data structures are described for node i, 0 i n 1. Subscripts on data structures to indicate the node are only included when needed. 4.1. Data Structures status: Indicates whether node is in the WAITING, CRITICAL, or REMAINDER section. Initially, status = REMAINDER. The set of all nodes in direct wireless contact with node i. Initially, N contains all of node i's neighbors. myHeight: A three-tuple (h1,h2,i) representing the height of node i. Links are considered to be directed from nodes with higher height toward nodes with lower height, based on lexicographic ordering. E.g., if myHeight and myHeight and the link between these nodes would be directed from node 1 to node 2. Initially at node 0, initialized so that the directed links form a DAG in which every node has a directed path to node 0. height[j]: An array of tuples representing node i's view of myHeight j for all Initially, In node i's viewpoint, then the link between i and j is incoming to node i if height[j] > myHeight, and outgoing from node i if height[j] < myHeight. tokenHolder: Flag set to true if node holds token and set to false otherwise. Initially, next: When node i holds the token, next = i, otherwise next is the node on an outgoing link. Initially, next next is an outgoing neighbor otherwise. Q: Queue containing identiers of requesting neighbors. Operations on Q include Enqueue(), which enqueues an item only if it is not already on Q, De- queue() with the usual FIFO semantics, and Delete(), which removes a specied item from Q, regardless of its location. Initially, receivedLI[j]: Boolean array indicating whether LinkInfo message has been received from node j, to which a Token message was recently sent. Any height information received at node i from a node j for which receivedLI[j] is false will not be recorded in height[j]. Initially, receivedLI i forming[j]: Boolean array set to true when link to node j has been detected as forming and reset to false when rst LinkInfo message arrives from node j. Initially, forming i formHeight[j]: An array of tuples storing value of myHeight when new link to rst detected. Initially, formHeight i 4.2. Overview of the RL Algorithm The mutual exclusion algorithm is event-driven. An event at a node i consists of receiving a message from another node j 6= i, or an indication of link failure or formation from the link layer, or an input from the application on node i to request or release the CS. Each message sent includes the current value of my- Height at the sender. Modules are assumed to be executed atomically. First, we describe the pseudocode triggered by events and then we describe the pseudocode for procedures. Requesting and releasing the CS: When node i requests access to the CS, it enqueues its own identier on Q and sets status to WAITING. If node i does not currently hold the token and i has a single element on its queue, it calls ForwardRequest() to send a Request message. If node i does hold the token, i can set status to CRITICAL and enter the CS, since it will be at the head of Q. When node i releases the CS, it calls GiveTokenToNext() to send a Token message if Q is non-empty, and sets status to REMAINDER. Request messages: When a Request message sent by a neighboring node j is received at node i, i ignores the Request if receivedLI[j] is false. Otherwise, i changes height[j], and enqueues j on Q if the link between i and j is incoming at i. If Q is non-empty, and status = REMAINDER, i calls GiveTokenToNext(), holds the token. Non-token holding node i calls RaiseHeight() if the link to j is now incoming and i has no outgoing links or i calls ForwardRequest() non-empty and the link to next has reversed. Token messages: When node i receives a Token message from some neighbor j, lowers its height to be lower than that of the last token holder, node j, informs all its outgoing neighbors of its new height by sending LinkInfo messages, and calls GiveTokenToNext(). Node i also informs j of its new height so that j will know that i received the token. messages: If receivedLI[j] is true when a LinkInfo message is received at node i from node j, j's height is saved in height[j]. If receivedLI[j] is false, checks if the height of j in the message is what it was when i sent the Token message to j. If so, i sets receivedLI[j] to true. If forming[j] is true, the current value of myHeight is compared to the value of myHeight when the link to j was rst detected, formHeight[j]. If myHeight and formHeight[j] are dierent, then a LinkInfo message is sent to j. Identier j is added to N and forming[j] is set to false. If j is an element of Q and j is an outgoing link, then j is deleted from Q. If node i has no outgoing links and is not the token holder, i calls RaiseHeight() so that an outgoing link will be formed. Otherwise, if Q is non-empty, and the link to next has reversed, i calls ForwardRequest() since it must send another Request for the token. Link failures: When node i senses the failure of a link to a neighboring node j, it removes j from N , sets receivedLI[j] to true, and, if j is an element of Q, deletes j from Q. Then, if i is not the token holder and i has no outgoing links, i calls RaiseHeight(). If node i is not the token holder, Q is non-empty, and the link to next has failed, i calls ForwardRequest() since it must send another Request for the token. Link formation: When node i detects a new link to node j, i sends a LinkInfo message to j with myHeight, sets forming[j] to true, and sets myHeight. Procedure ForwardRequest: Selects node i's lowest height neighbor to be next. Sends a Request message to next. Procedure GiveTokenToNext: Node i dequeues the rst node on Q and sets next equal to this value. If next = i, i enters the CS. If next 6= i, i lowers height[next] to (myHeight.h1, myHeight.h2 1; next), so any incoming Request messages will be sent to next, sets tokenHolder = false, sets receivedLI[next] to false, and then sends a Token message to next. If Q is non-empty after sending a Token message to next, a Request message is sent to next immediately following the Token message so the token will eventually be returned to i. Procedure RaiseHeight: Called at non-token holding node i when i loses its last outgoing link. Node i raises its height (in lines 1-3) using the partial reversal method of [13] and informs all its neighbors of its height change with LinkInfo messages. All nodes on Q to which links are now outgoing are deleted from Q. If Q is not empty at this point, ForwardRequest() is called since i must send another Request for the token. 4.3. The RL Algorithm When node i requests access to the CS: 1. status := WAITING 2. Enqueue(Q; i) 3. If (not tokenHolder) then 4. If 5. Else GiveTokenToNext() When node i releases the CS: 1. If (jQj > 0) then GiveTokenToNext() 2. status := REMAINDER When Request(h) received at node i from node j: // h denotes j's height when message was sent 1. If (receivedLI[j]) then 2. height[j] := h // set i's view of j's height 3. If (myHeight < height[j]) then Enqueue(Q; 4. If (tokenHolder) then 5. If ((status = REMAINDER) and (jQj > 0)) then GiveTokenToNext() 6. Else // not tokenHolder 7. If (myHeight < height[k], 8 k 2 N) then RaiseHeight() 8. Else if 9. ForwardRequest() // reroute request When Token(h) received at node i from node j: // h denotes j's height when message was sent 1. tokenHolder := true 2. height[j] := h 3. Send LinkInfo(h.h1, h.h2 1; i) to all outgoing k 2 N and to j 4. myHeight.h1 := h.h1 5. myHeight.h2 := h.h2 - 1 // lower my height 6. If (jQj > 0) then GiveTokenToNext() Else next := i When LinkInfo(h) received at node i from node j: // h denotes j's height when message was sent 1. N := N [ fjg 2. If ((forming[j]) and (myHeight 6= formHeight[j])) then 3. Send LinkInfo(myHeight) to j 4. forming[j] := false 5. If (receivedLI[j]) then height[j] := h 6. Else if 7. If (myHeight > height[j]) then Delete(Q; 8. If ((myHeight < height[k], 8k 2 N) and (not tokenHolder)) then RaiseHeight() // reroute request 9. Else if ((jQj > 0) and (myHeight < height[next])) then ForwardRequest() When failure of link to j detected at node i: 1. N := N fjg 2. Delete(Q; 3. receivedLI[j] := true 4. If (not tokenHolder) then 5. If (myHeight < height[k], 8k 2 N) then RaiseHeight() // reroute request 6. Else if ((jQj > 0) and (next 62 N)) then ForwardRequest() When formation of link to j detected at node i: 1. Send LinkInfo(myHeight) to j 2. forming[j] := true 3. formHeight[j] := myHeight Procedure 1. next := l 2. Send Request(myHeight) to next Procedure GiveTokenToNext(): // only called when jQj > 0 1. next := Dequeue(Q) 2. If (next 6= i) then 3. tokenHolder := false 4. height[next] := (myHeight.h1, myHeight.h2 1, next) 5. receivedLI[next] := false 6. Send Token(myHeight) to next 7. If (jQj > 0) then Send Request(myHeight) to next 8. Else // next 9. status := CRITICAL 10. Enter CS Procedure 1. 2. S := 3. If (S 6= ;) then myHeight.h2 := min l2S fheight[l].h2g 1 4. Send LinkInfo(myHeight) to all k 2 N Raising own height can cause some links to become outgoing 5. For (all k 2 N such that myHeight > height[k]) do Delete(Q; Must reroute request if queue non-empty, since just had no outgoing links 6. If (jQj > 0) then ForwardRequest() 4.4. Examples of Algorithm Operation We rst discuss the case of a static network, followed by a dynamic network. An illustration of the algorithm on a static network (in which links do not fail or form) is depicted in Figure 3. Snapshots of the system conguration during algorithm execution are shown, with time increasing from 3(a) to 3(e). The direct wireless links are shown as dashed lines connecting circular nodes. The arrow on each wireless link points from the higher height node to the lower height node. The request queue at each node is depicted as a rectangle, the height is shown as a 3-tuple, and the token holder as a shaded circle. The next pointers are shown as solid arrows. Note that when a node holds the token, its next pointer is directed towards itself. In Figure nodes 2 and 3 have requested access to the CS (note that nodes 2 and 3 have enqueued themselves on Q 2 and Q 3 ) and have sent Request messages to node 0, which enqueued them on Q 0 in the order in which the Request messages were received. Part (b) depicts the system at a later time, where node 1 has requested access to the CS, and has sent a Request message to node 3 (note that 1 is enqueued on Q 1 and Q 3 ). Figure 3(c) shows the system conguration after node 0 has released the CS and has sent a Token message to node 3, followed by a Request sent by node 0 on behalf of node 2. Observe that the logical direction (d) (b) (c) Figure 3. Operation of Reverse Link Mutual Exclusion Algorithm on Static Network of the link between node 0 and node 3 changes from being directed away from node 3 in part (b), to being directed toward node 3 in part (c), when node 3 receives the Token message and lowers its height. Notice also the next pointers of nodes 0 and 3 change from both nodes having next pointers directed toward node 0 in part (b) to both nodes having next pointers directed toward node 3 in part (c). Part (d) shows the system conguration after node 3 sent a Token message to node 1, followed by a Request message. The Request message was sent because node 3 received the Request message from node 0. Notice that the items at the head of the nodes' request queues in part (d) form a path from the token holder, node 1, to the sole remaining requester, node 2. Part (e) depicts the system conguration after Token messages have been passed from node 1 to node 3 to 0, and from node 0 to 2. Observe that the middle element, h2, of each node's myHeight tuple decreases by 1 for every hop the token travels, so that the token holder is always the lowest height node in the system. We now consider the execution of the RL algorithm on a dynamic network. The height information allows each node i to keep track of the current logical direction of links to neighboring nodes, particularly to the node chosen to be next. If the link to next changes and jQj > 0, node i must reroute its request by calling ForwardRequest(). Figure 4(a) shows the same snapshot of the system execution as is shown in Figure 3(a), with time increasing from 4(a) to 4(e). Figure 4(b) depicts the system conguration after node 3 has moved in relation to the other nodes in (a) (b) (c) (d) (e) Figure 4. Operation of Reverse Link Mutual Exclusion Algorithm on Dynamic Network the system, resulting in a network that is temporarily not token oriented, since node 3 has no outgoing links. Node 0 has adapted to the lost link to node 3 by removing 3 from its request queue. Node 2 takes no action as a result of the loss of its link to node 3, since the link to next 2 was not aected and node 2 still has one outgoing link. In part (c), node 3 has adapted to the loss of its link to node 0 by raising its height and has sent a Request message to node 1 (that has not yet arrived at node 1). Part (d) shows the system conguration after node 1 has received the Request message from node 3, has enqueued 3 on Q 1 , and has raised its height due to the loss of its last outgoing link. In part (e), node 1 has propagated the Request received from node 3 by sending a Request to node 2, also informing node 2 of the change in its height. Node 2 subsequently enqueued 1 on Q 2 , but did not raise its own height or send a Request, because node 2 has an intact link to next 2 , node 0, to which it already sent an unfullled request. 5. Correctness of Reverse Link Algorithm The following theorem holds because there is only one token in the system at any time. Theorem 1. The algorithm ensures mutual exclusion. To prove no starvation, we rst show that, after link changes cease, eventually the system reaches a \good" conguration, and then we apply a variant function argument. We will show that after link changes cease, the logical directions on the links imparted by height values will eventually form a \token oriented" DAG. Since the height values of the nodes are totally ordered, there cannot be any cycles in the logical graph, and thus it is a DAG. The hard part is showing that this DAG is token oriented, dened next. Denition 1. A node i is the token holder in a conguration if tokenHolder true or if a Token message is in transit from node i to next i . Denition 2. The DAG is token oriented in a conguration if for every node there exists a directed path originating at node i and terminating at the token holder. To prove Lemma 3, that the DAG is eventually token oriented, we rst show, in Lemma 1, that this condition is equivalent to the absence of \sink" nodes [13], as dened below. We then show, in Lemma 2, that eventually there are no more calls to RaiseHeight(). Throughout, we assume that eventually link changes cease. Denition 3. A node i is a sink in a conguration if Lemma 1. In every conguration of every execution, the DAG is token oriented if and only if there are no sinks. Proof: The only-if direction follows from the denition of a token oriented DAG. The if direction is proved by contradiction. Assume, in contradiction, that there exists a node i in a conguration such that tokenHolder false and for which there is no directed path starting at i and ending at the token holder. Since there are no sinks, i must have at least one outgoing link that is incoming at some other node. Since the number of nodes is nite, the network is connected, and all links are logically directed such that no logical path can form a cycle, there must exist a directed path from i to the token holder, a contradiction. To show that eventually there are no sinks (Lemma 3), we show that there are only a nite number of calls to RaiseHeight(). Lemma 2. In every execution with a nite number of link changes, there exists a nite number of calls to RaiseHeight(). Proof: In contradiction, consider an execution with a nite number of link changes but an innite number of calls to RaiseHeight(). Then, after link changes cease, some node calls RaiseHeight() innitely often. We rst note that if one node calls RaiseHeight() innitely often, then every node calls RaiseHeight() innitely often. To see this, consider that a node i would call RaiseHeight() innitely often only if it lost all its outgoing links innitely often. But this would happen innitely often at node i only if a neighboring node j raised its height innitely often, and neighboring node j would only call RaiseHeight() innitely often if its neighbor k raised its height innitely often, and so on. However, Claim 1 shows that at least one node calls RaiseHeight() only a nite number of times. 1. No node that holds the token after the last link change ever calls subsequently. Proof: Suppose the claim is false, and some node that holds the token after the last link change calls RaiseHeight() subsequently. Let i be the rst node to do so. By the code, node i does not hold the token when it calls RaiseHeight(). Suppose that node i sends the token to neighboring node j at time t 1 , setting its view of j to be outgoing, and at a later time, t 3 , node i calls RaiseHeight(). The reason i calls RaiseHeight() at time t 3 is that it lost its last outgoing link. Thus, at time t 2 between time t 1 and t 3 , the link between i and j has reversed direction in i's view from outgoing to incoming. By the code, the direction change at node i must be due to the receipt of a LinkInfo or Request message from node j. We discuss these cases separately below. Case 1: The direction change at node i is due to the receipt of a LinkInfo message from node j at time t 2 . By the code, when i sends the token to j at t 1 , it sets receivedLI[j] to false. Therefore, when the LinkInfo message is received at i from j at time t 2 , node i must have already reset receivedLI[j] to true or i would still see the link to j as outgoing and would not call RaiseHeight() at time t 2 . Since called after receiving the LinkInfo message from j at time t 2 , i must have received the LinkInfo message node j sent when it received the token from i before time t 2 , by the FIFO assumption on message delivery. Then node must have received the token and sent it to another node, k 6= i, after which j raised its height and sent the LinkInfo message that node i received at time t 2 . However, this violates our assumption that i is the rst node to call RaiseHeight() after the last link change, a contradiction. Case 2: The direction change at node i is due to the receipt of a Request message from node j at time t 2 . By a similar argument to case 1, any Request received from node j would be ignored at node i as long as receivedLI[j] is false. But this means that node j must have called RaiseHeight() after it received the token from node i and subsequently sent the Request received by i at time t 2 . Again, this violates the assumption that i is the rst node to call RaiseHeight() after the last link change, a contradiction. Therefore, node i will not call RaiseHeight() at time t 2 and the claim is true. Therefore, by Claim 1, there is only a nite number of calls to RaiseHeight() in any execution with a nite number of link changes. Lemma 3 follows from Lemma 2, since if a node becomes a sink, it will eventually be informed via LinkInfo messages and will then call RaiseHeight(). Lemma 3. Once link changes cease, the logical direction on links imparted by height values will eventually always form a token oriented DAG. Consider a node that is WAITING in an execution at some point after link changes and calls to RaiseHeight() have ceased. We rst dene the \request chain" of a node to be the path along which its request has propagated. Then we modify the variant function argument in [25] to show that the node eventually gets to enter the CS. Denition 4. Given a conguration, a request chain for any node l with a non-empty request queue is the maximal length list of node identiers queue is not empty, the link between p i 1 and p i is outgoing at p i 1 and incoming at p i , no Request message is in transit from p i 1 to p i , and no Token message is in transit from p i to p i 1 . Lemma 4 gives useful information about what is going on at the end of a request chain: Lemma 4. The following is true in every conguration: Let l be a node with a non-empty request queue and let request chain. Then (a) l is in Q l i l is WAITING, (c) either p j is the token holder, or a Token message is in transit to p j , or a Request message is in transit from p j to next p j , or a LinkInfo message is in transit from next p j to p j with next p j higher than or next p j sees the link to p j as failed. Proof: By induction on the execution. Property (a) can easily be shown to hold, since a node enqueues its own identier when its application requests access to the CS, at which point it changes its status to WAITING. By the code, at no point will a node dequeue its own identier until just before it enters the CS and sets its status to CRITICAL. Properties (b) and (c) are vacuously true in the initial conguration, since no node has a non-empty queue. Suppose (b) and (c) are true in the (t 1) st conguration, C t 1 , of the execution. It is possible to show these properties are true in the t th conguration, by considering in turn every possibility for the t th event. Most of the events applied to C t 1 are easily shown to yield a conguration C t in which properties (b) and (c) are true. Here we discuss the events for which the outcome is less clear by presenting the problematic cases that can appear to disrupt a request chain. We note that, in the following cases, non-token holding nodes are often required to nd an outgoing link due to link reversals or failures. It is not hard to show that a node i that is not the token holder can always nd an outgoing link due to the performance of RaiseHeight(). Case 1: Node i receives a Request(h) from node j and does not enqueue j on its request queue. To ensure that j's Request is not overlooked, causing possible starvation, we show that either a LinkInfo or a Token message is sent to j from a Request from j is received at i and j is not enqueued. Case 1.1: receivedLI[j] is false at i. It must be that i sent the token to j in some previous conguration and i has not yet received the LinkInfo message that j must send to i upon receipt of the token. If the token is not in transit from i to j or held by j in C t 1 , then earlier j had the token and passed it on. The Request received by i was sent before the LinkInfo message that j must send to i upon receipt of the token. So if j is WAITING in C t 1 , it has already sent a newer Request and properties (b) and (c) hold for this request chain in C t by the inductive hypothesis. Case 1.2: receivedLI[j] is true at i. Then if j is not enqueued on i's request queue, it must be that myHeight i > h. Since viewed i as outgoing when it sent the Request, node i must have either called in N i or the relative heights of i and j changed between the time link (i; was rst detected and before j was added to N i . In either case, node j must eventually receive a Linkinfo message from i and see that its link to next j has reversed, in which case j will take action resulting in the eventual sending of another Request. Case 2: Node i receives an input causing it to delete identier j from its request queue. To ensure that j's Request is not forgotten when i calls Delete(Q; j), we show that either node j received a Token message prior to the deletion, in which case j's Request is satised, or node j is notied that the link to i failed, in which case j will take the appropriate action to reroute the request chain. Case 2.1: Node i calls Delete(Q; receives a LinkInfo message from j indicating that i's link to j has become outgoing at i. Then, since i enqueued j, it must be that in some earlier conguration i saw the link to j as incoming. Since the receipt of the LinkInfo message from j caused the link to change from incoming to outgoing in i's view, it must be that the LinkInfo was sent by j when j received the token and lowered its height. If the token is not held by j in C t 1 , then earlier j had the token and passed it on. If j is WAITING in C t 1 , it has already sent a newer Request and properties (b) and (c) hold for this request chain in C t by the inductive hypothesis. Case 2.2: Node i calls Delete(Q; received an indication that link must receive the same indication, in which case it can take appropriate action to advance any request chains. Case 3: Node i receives an input which makes it see the link to next i as incoming or failed. In this case, any request chains including node i in C t 1 end at i in C t . We show that node i takes the correct action to propagate these request chains by sending either a new Request or a LinkInfo message. Case 3.1: Node i receives a LinkInfo message from neighbor indicating that i's link to j has become incoming at i. If the link to j was i's last outgoing link, then in C t i will call RaiseHeight(). Node i will delete the identiers of any nodes on outgoing links from its request queue. Node i will send a LinkInfo message to each neighbor, including nodes whose identiers were removed from i's request queue. If i's request queue is non-empty it will call ForwardRequest() and send a Request message to the node chosen as next i in Case 3.2: Node i receives an indication that the link to next i has failed. In C t , i will take the same actions as it did in case 3.1, when its link to next i reversed. Therefore, no action taken by node i can make properties (b) and (c) false and the lemma holds. Lemma 5. Once link changes and calls to RaiseHeight() cease, for every cong- uration in which a node l's request chain does not include the token holder, then there is a later conguration in which l's request chain does include the token holder. Proof: By Lemma 3, after link changes cease, eventually a token oriented DAG will be formed. Consider a conguration after link changes and calls to RaiseHeight() cease in which the DAG is token oriented, meaning that all LinkInfo messages generated when nodes raise their heights have been delivered. The proof is by contradiction. Assume node l's request chain never includes the token holder. So the token can only be held by or be in transit to nodes that are not in l's request chain. By our assumption on the execution, no LinkInfo messages caused by a call to RaiseHeight() will be in transit to a node in l's request chain, nor will any node in l's request chain detect a failed link to a neighboring node. Therefore, by Lemma 4(c), a Request message must be in transit from a node in l's request chain to a node that is not in l's request chain, and the number of nodes in l's request chain will increase when the Request message is received. At this point, l's request chain will either include the token holder, another Request message will be in transit from a node in l's request chain to a node that is not in l's request chain, or l's request chain will have joined the request chain of some other node. While the number of nodes in l's request chain increases, the number of nodes not in l's request chain decreases, since there are a nite number of nodes in the system. So eventually l's request chain includes all nodes. Therefore, if the token is not eventually contained in l's request chain, it is not in the system, a contradiction. Let l be a node that is WAITING after link changes and calls to Raise- cease. Given a conguration s in the execution, a function V l for l is dened to be the following vector of positive integers. Let l's request chain. V l (s) has either depending on whether a Request message is in transit from p m or not. In either case, v 1 is the position of p 1 (= l) in Q l , and for 1 < j m, v j is the position of p j 1 in (Positions are numbered in ascending order with 1 being the head of the queue.) If a Request message is in transit, then V l l (s) has only m elements. These vectors are compared lexicographically. Lemma 6. V l is a variant function. Proof: The key points to prove are: l never has more than n entries and every entry is between 1 and n the range of V l is well-founded. (2) Most events can be easily seen not to increase V l . Here we discuss the remaining events. When the Request message at the end of l's request chain is received by node j from node p m , l's request chain increases in length to m decreases from m+1 is p m 's position in Q j after the Request message is received. When a Token message is received by the node p m at the end of l's request chain, it is either - kept at p m , so V l decreases from hv - or sent toward l, so V l decreases from hv - or sent away from l, followed by a Request message, so V l decreases from (3) To see that the events that cause V l to decrease will continue to occur, consider the following two cases: Case 1: The token holder is not in l's request chain. By Lemma 5, eventually the token holder will be in l's request chain. Case 2: The token holder is in l's request chain. Since no node stays in the CS forever, at some later time the token will be sent and received, decreasing the value of V l , by part (2) of this proof. Once V l equals h1i, l enters the CS. We have: Theorem 2. If link changes cease, then every request is eventually satised. 6. Simulation Results In this section we discuss the static and dynamic performance of the Reverse Link (RL) algorithm compared to a mutual exclusion algorithm designed to operate on a static network. We simulated Raymond's token based mutual exclusion algorithm [25] as if it were running on top of a \routing" layer that always provided shortest path routes between nodes. In this section, we will refer to this simulation as \Raymond's with routing" (RR). Raymond's algorithm was used because it is the static algorithm from which the RL algorithm was adapted and because it does not provide for link failures and recovery and must rely on the routing layer to maintain logical paths if run in a dynamic network. In order to make our results more generally applicable, we made best-case assumptions about the underlying routing protocol used with Raymond's algorithm: that it always provides shortest paths and its time and message complexity are zero. If our simulation shows that the RL algorithm is better than the RR combination in some scenario, then the RL algorithm will also be better than Raymond's algorithm in that scenario when any real ad hoc routing algorithm is used. If our simulation shows that the RL algorithm is worse than the RR combination in some scenario, then it might or might not be worse in an actual situation, depending on how much worse it is in the simulation and what are the costs of the routing algorithm. We simulated a node system under various scenarios. We chose to simulate on a system because for networks larger than nodes the time needed for simulation was very high. Also, we envision ad hoc networks to be much smaller scale than wired networks like the Internet. Typical numbers of nodes used for simulations of ad hoc networks range from 10 to 50 [4{6,15,18,26]. In all our experiments, each CS execution took one time unit and each message delay was one time unit. Requests for the CS were modeled as a Poisson process with arrival rate req . Thus the time delay between when a node left the CS and made its next request to enter the CS is an exponential random variable with mean 1 time units. Link changes were modeled as a Poisson process with arrival rate mob . Thus the time delay between each change to the graph is an exponential random variable with mean 1 mob time units. Each change to the graph consisted of the deletion of a link chosen at random (whose loss did not disconnect the graph) and the formation of a link chosen at random. In each execution, we measured the average waiting time for CS entry, that is, the average number of time units that nodes spent in their WAITING sections. We also measured the average number of messages sent per CS entry. We varied the load on the system ( req ), the degree of mobility ( mob ), and the \connectivity" of the graph. Connectivity was measured as the percentage of possible links that were present in the graph. Note that a clique on nodes has 435 (undirected) links. In the graphs of our results, each plotted point represents the average of six repetitions of the simulation. Thus in plots of average time per CS entry, each point is the average of the averages from six executions, and similarly for plots of average number of messages per CS entry. For the RR simulations, we initially formed a random connected graph with the desired number of links and then used breadth-rst search to form a spanning tree of the graph to play the part of the static virtual spanning tree over which nodes communicate in Raymond's algorithm. After the spanning tree was formed, we randomly permuted the graph while maintaining the desired connectivity and then calculated the shortest paths from all nodes to their neighbors in the virtual spanning tree. After this, we started the mutual exclusion algorithm and began counting messages and waiting time per CS entry. When link changes occurred, we did not measure the time or messages needed to recalculate shortest path routes in the modied graph. We did measure any added time and distance that the application messages traveled due to route changes, charging one message per link traversed. For simulations of RL, we formed a random connected graph with the desired number of links, initialized the node heights and link directions, and then started the algorithm and performance measurements. When link changes occurred, the time and messages needed to nd new routes between nodes were included in the overall cost of performance. In this section, part (a) of each gure displays results when the graph is static, part (b) when (low mobility), and part (c) when Our choice for the value of the low mobility parameter corresponds to the situation where nodes remain stationary for a few tens of seconds after moving and prior to making another move. Our choice for the value of the high mobility parameter represents a much more volatile network, where nodes remain static for only a few seconds between moves. 6.1. Average waiting time per CS entry Load (Request Arrival Rate) (a) Load (Request Arrival Rate)2060100140180.0001 (b) Time Units/CS Entry Time Units/CS Entry Load (Request Arrival Rate) Time Units/CS Entry RR, 20% Connectivity RR, 80% Connectivity RL, 80% Connectivity RL, 20% Connectivity Figure 5. Load vs. Time/CS entry for (a) zero, (b) low, and (c) high mobility Figure 5 plots the average number of time units elapsed between host request and subsequent entry to the CS against values of req increasing from 10 4 (the mean time units between requests is requests is 1) from left to right along the x axis. We chose the high load value of req because at this rate each node would have a request pending almost all the time. The low load value of req represents a much less busy network, with requests rarely pending at all nodes at the same time. Plots are shown for runs with 20% (87 links) and 80% (348 links) connectivity for both the RL and RR simulations. Figure 5 indicates that RL has better performance than RR in terms of average waiting time per CS entry, up to a factor of six. The reason is that Raymond's algorithm sends application messages over a static virtual spanning tree; when a message is sent from a node to one of its neighbors in the virtual spanning tree, it may actually be routed over a long distance, thus increasing the time delay. In contrast, the RL algorithm uses accurate information about the actual topology, resulting in less delay between each request and subsequent CS entry. Both algorithms show an increase in average waiting time per CS entry from low to high load in Figure 5. The higher the load, the larger is the number of other nodes that precede a given node into the CS. The average waiting time for each CS entry reaches its peak for the RL simulation at around 75 time units per CS entry under the highest load. This is caused by an essentially round robin pattern of token traversal. However, the average waiting time for the RL simulation in Figure 5(c) at the highest load actually decreases under high mobility. This phenomenon may be due to the fact that, at high loads, frequent link failures break the fair pattern in which the token is received, causing some nodes to get the token more frequently. Figure 5 also shows that the waiting time advantage of RL over RR increases with increasing load and increasing mobility. The increased waiting time of RR with increased load when the network connectivity is low is due to longer average route lengths. In the simulation trials, the average route length roughly doubled when the connectivity decreased from 80% to 20%. The performance gap between waiting time for RL and RR is seen to a lesser degree at high connectivity, when average route length in RR is lower. However, it is apparent that the RR simulation suers from the combined eects of higher contention and imposed static spanning tree communication paths at high loads, while RL is mainly aected by contention for the CS at high loads. Finally, Figure 5 suggests that connectivity in the range tested is immaterial to the behavior of the RL algorithm at high load, whereas a larger connectivity is better for RR than a smaller connectivity at all loads. In order to further study the eect of connectivity, we ran the experiments shown in Figure 6: the average number of time units elapsed between host request and subsequent entry to the CS is plotted against network connectivity increasing from 10% (43 links) to 100% (435 links) along the x axis. Curves are plotted for low load, where (the mean time unit between requests is load, where mean time unit between requests) for both the RL and RR simulations. RL, Low Load RR, High Load RR, Low Load101000 X101000 (a) Connectivity Connectivity (b) Time Units/CS Entry Time Units/CS Entry Time Units/CS Entry Connectivity Figure 6. Connectivity vs. Time/CS entry for (a) zero, (b) low, and (c) high mobility Figure 6 conrms that connectivity does not aect the waiting time per CS entry in the RL simulation at high load. At high load, the RL algorithm does not exploit connectivity. When load is high, the RL simulation always sends request messages over the path last traveled by the token, even if there is a shorter path to the token when the request is made. At low load in RL, connectivity does aect the waiting time per CS entry because request messages are not always sent over the path last traveled by the token. This is because with lower load there is sucient time between requests for token movement to change link direction in the vicinity of the token holder, an eect that increases with higher connectivity, shortening request paths. The waiting time for the RR algorithm decreases with increasing connec- tivity, since the path lengths between neighbors in the virtual spanning tree approach one. However, even with a clique, when shortest path lengths are all one, the time for RR does not match that for RL. The reason is that the spanning tree used by RR for all communication might have a relatively large diameter, whereas in RL neighboring nodes are always in direct communication. The results of the simulations in this section are summarized in Table 1. Zero Mobility Low Mobility High Mobility Connectivity 20% 80% 20% 80% 20% 80% RR high load 185 107 185 140 294 290 RL high load RR low load 17 8 RL low load 7 4 Table 1: Summary of time per CS entry. 6.2. Average number of messages per CS entry The RR algorithm sends request and token messages along the virtual spanning tree. Each message from a node to its virtual neighbor is converted into a sequence of actual messages, that traverse the (current) shortest path from the sender to the recipient. The RL algorithm sends Request and Token messages along the actual token oriented DAG. In addition, as the token traverses a path, each node on that path sends LinkInfo messages to all its outgoing neighbors. Additional LinkInfo messages are sent, and propagated, when a link failure causes a node to lose its last outgoing link. Our experimental results re ect the relative number of routing messages for RR vs. LinkInfo messages for RL. When interpreting these results, it is important to remember that the simulation of the RR algorithm is not charged for messages needed to recalculate the routes due to topology changes. Thus, if RL is better than RR in some situation, it will certainly be better when routing messages are charged to it, even if they are prorated. Also, if RR is better than RL in another Load (Request Arrival Rate) (a)10152535100010Load (Request Arrival Rate)X Entry Entry Entry Load (Request Arrival Rate) RR, 20% Connectivity RR, 80% Connectivity RL, 80% Connectivity RL, 20% Connectivity Figure 7. Load vs. Messages/CS Entry for (a) zero, (b) low, and (c) high mobility situation, depending on how much better it is, RL might be comparable or even better than RR when routing messages are charged to RR. Figure 7 plots the average number of messages received per CS execution against values of req ranging from 10 4 (the mean time units between requests (the mean time units between requests is 1) from left to right along the x axis. Plots are shown for runs with 20% (87 links) and 80% (348 links) connectivity for both the RL and RR simulations. Figure 7(b) and (c) show that the RR algorithm sends fewer messages per CS entry than the RL algorithm in all simulation trials with mobility, although as load increases the message advantage of RR decreases markedly. In all situations studied, except the RL simulation in the static case with high connectivity, the number of messages per CS entry tends to decrease as load increases. The reason is that, although the overall number of messages increases with load in both algorithms, due to the additional token and request messages, it increases less than linearly with the number of requests, and hence less than linearly with the number of CS entries. In the extreme, at very high load, every time the token moves, it is likely to cause a CS entry. In the static case with high connectivity, the RL algorithm experiences a threshold eect around load of .01: when load is less than .01, the number of messages per CS entry is roughly constant at a lower value, and when the load is above .01, the number of messages per CS entry is roughly constant at a higher value. The threshold eect becomes less pronounced as connectivity decreases. We conjecture that some qualitative behavior of the algorithm on a 30 node graph changes when load increases from .001 to .01. This change may be attributed to the observation that token movement more eectively shortens request path length at high connectivity with low load. This is because at low load there is sucient time between requests for nodes to receive LinkInfo messages sent as the token moves, causing nodes to send requests over direct links to the token holder rather than over the last link on which they sent the token. This eect is amplied at high connectivity because each node is more likely to be directly connected to the token holder. The RL algorithm sends more messages per CS entry than the RR algorithm when mobility causes link changes, and the number of messages sent in the RL algorithm grows very large under low loads, as can be observed in Figure 7(b) and (c). When links fail and form, the RL algorithm sends many LinkInfo messages to maintain the token oriented DAG, resulting in a higher message to CS entry ratio at low loads when the degree of mobility remains constant. However, when interpreting these results, it is important to note that the RL algorithm is being charged for the cost of routing in the simulations with mobility, while the RR simulation is not charged for routing. Figure 8 shows the results of experiments designed to understand the eect of connectivity on the number of messages per CS entry. In the gure, the average number of messages per CS entry is plotted against network connectivity increasing from 10% (43 links) to 100% (435 links) from left to right on the x axis. Curves are plotted for low load, where between requests is load, where between requests is 1) for both the RL and RR simulations. In the static case, the number of RL messages per CS entry increases linearly with connectivity, for a xed load. As connectivity increases, the number of neighbors per node increases, resulting in more LinkInfo messages being sent as RR, Low Load RR, High Load RL, Low Load Connectivity Connectivity (b) Entry Entry Entry Connectivity (c) Figure 8. Connectivity vs. Messages/CS Entry for (a) zero, (b) low, and (c) high mobility the token travels. However, the number of RR messages per CS entry decreases (less than linearly) with connectivity, since the shortest path lengths between neighbors in the virtual spanning tree decrease. In fact, our results for RR at 100% connectivity (when the virtual spanning tree is an actual spanning tree) and high load match the performance of approximately 4 messages per CS entry cited by Raymond [25] at high load. Part (a) of Figure 8 shows that in the static case the RL algorithm uses fewer messages per CS entry below 25% connectivity for high load and below 60% connectivity for low load. Figure 8(b) and (c) show that, in the dynamic cases, the number of messages per CS entry is little aected by connectivity for a xed load. In the RL algorithm, there are two opposing trends with increasing connectivity that appear to cancel each other out: higher connectivity means more neighbors per node, which means more LinkInfo messages will be sent with each failure. On the other hand, more neighbors per node means that it is less likely for a link failure to be that of the last outgoing link, and thus LinkInfo messages due to failure will propagate less. For the RR case, the logarithmic scale on the y axis in Figure 8(c) hides the slight decrease in messages per CS entry, making both curves appear at. The results of the simulations in this section are summarized in Table 2. Zero Mobility Low Mobility High Mobility Connectivity 20% 80% 20% 80% 20% 80% RR high load 13 6 11 7 RR low load RL low load 13 17 189 180 1900 1825 Table 2: Summary of messages per CS entry. 7. Conclusion and Discussion We presented a distributed mutual exclusion algorithm designed to be aware of and adapt to node mobility, along with a proof of correctness, and simulation results comparing the performance of this algorithm to that of a static token based mutual exclusion algorithm running on top of an ideal ad hoc routing protocol. We assumed there were no partitions in the network throughout this paper for simplicity; partitions can be handled in our algorithm by using a method similar to that used in the TORA ad hoc routing protocol [22]. In [22], additional labels are used to represent the heights of nodes, allowing nodes to detect, by recognition of the originator of a chain of height increases, when a series of height changes has occurred at all reachable nodes without encountering the \destination". A similar partition detection mechanism could be encorporated into our mutual exclusion algorithm at the expense of slightly larger messages. Our algorithm compares favorably to the layered approach using an ad hoc routing protocol, providing better average waiting time per CS entry in all tested scenarios. Our simulation results indicate that in many situations the message complexity per CS entry of our algorithm would not be greater than the message cost incurred by a static mutual exclusion algorithm running on top of an ad hoc routing algorithm, when messages of both the mutual exclusion algorithm and the routing algorithm are counted. Acknowledgements We thank Savita Kini for many discussions on previous versions of the algo- rithm, Soma Chaudhuri for careful reading and helpful comments on the liveness proof, and Debra Elkins for helpful discussions. --R The slide mechanism with applications in dynamic networks. Polynomial end to end communication. Structuring distributed algorithms for mobile hosts. A distance routing e A performance comparison of multi-hop wireless ad hoc network routing protocols Query localization techniques for on-demand routing protocols in ad hoc networks A fault tolerant algorithm for distributed mutual exclusion. Routing and multicast in multihop A distributed routing algorithm for mobile wireless networks. A token based k-resilient mutual exclusion algorithm for distributed systems Signal stability based adaptive routing (SSA) for ad-hoc mobile networks Scheduling broadcasts in multihop radio networks. Distributed algorithms for generating loop-free routes in networks with frequently changing topology Dynamic source routing in ad hoc wireless networks. A cluster-based approach for routing in dynamic networks Reliable broadcast in mobile multihop packet networks. A highly adaptive distributed routing algorithm for mobile wireless networks. Highly dynamic destination-sequenced distance-vector routing for mobile computers Multicast operation of the ad-hoc on-demand distance vector routing protocol --TR A tree-based algorithm for distributed mutual exclusion The slide mechanism with applications in dynamic networks A token based <italic>k</italic>-resilient mutual exclusion algorithm for distributed systems Highly dynamic Destination-Sequenced Distance-Vector routing (DSDV) for mobile computers A distributed routing algorithm for mobile wireless networks Efficient message ordering in dynamic networks Reliable broadcast in mobile multihop packet networks A cluster-based approach for routing in dynamic networks Multicluster, mobile, multimedia radio network Location-aided routing (LAR) in mobile ad hoc networks A distance routing effect algorithm for mobility (DREAM) A performance comparison of multi-hop wireless ad hoc network routing protocols Query localization techniques for on-demand routing protocols in ad hoc networks Scenario-based performance analysis of routing protocols for mobile ad-hoc networks Multicast operation of the ad-hoc on-demand distance vector routing protocol Ad-hoc On-Demand Distance Vector Routing A Highly Adaptive Distributed Routing Algorithm for Mobile Wireless Networks --CTR Chen , Jennifer L. Welch, Self-stabilizing mutual exclusion using tokens in mobile ad hoc networks, Proceedings of the 6th international workshop on Discrete algorithms and methods for mobile computing and communications, September 28-28, 2002, Atlanta, Georgia, USA Chen , Jennifer L. Welch, Self-stabilizing dynamic mutual exclusion for mobile ad hoc networks, Journal of Parallel and Distributed Computing, v.65 n.9, p.1072-1089, September 2005 Djibo Karimou , Jean Frdric Myoupo, An Application of an Initialization Protocol to Permutation Routing in a Single-Hop Mobile Ad Hoc Networks, The Journal of Supercomputing, v.31 n.3, p.215-226, March 2005 Emmanuelle Anceaume , Ajoy K. Datta , Maria Gradinariu , Gwendal Simon, Publish/subscribe scheme for mobile networks, Proceedings of the second ACM international workshop on Principles of mobile computing, October 30-31, 2002, Toulouse, France Gruia-Catalin Roman , Jamie Payton, A Termination Detection Protocol for Use in Mobile Ad Hoc Networks, Automated Software Engineering, v.12 n.1, p.81-99, January 2005 M. Benchaba , A. Bouabdallah , N. Badache , M. Ahmed-Nacer, Distributed mutual exclusion algorithms in mobile ad hoc networks: an overview, ACM SIGOPS Operating Systems Review, v.38 n.1, p.74-89, January 2004

mobile computing;mutual exclusion;distributed algorithm;ad hoc network

506896

Replication requirements in mobile environments.

Replication is extremely important in mobile environments because nomadic users require local copies of important data. However, today's replication systems are not "mobile-ready". Instead of improving the mobile user's environment, the replication system actually hinders mobility and complicates mobile operation. Designed for stationary environments, the replication services do not and cannot provide mobile users with the capabilities they require. Replication in mobile environments requires fundamentally different solutions than those previously proposed, because nomadicity presents a fundamentally new and different computing paradigm. Here we outline the requirements that mobility places on the replication service, and briefly describe ROAM, a system designed to meet those requirements.

Introduction Mobile computing is rapidly becoming standard in all types of environments: academic, com- mercial, and private. Widespread mobility impacts multiple arenas, but one of particular importance is data replication. Replication is especially important in mobile environments, since disconnected or poorly connected machines must rely primarily on local resources. The monetary costs of communication when mobile, combined with the lower bandwidth, higher latency, and reduced availability, eectively require that important data be stored locally on the mobile machine. In the case of shared data, between multiple mobile users or between mobile and stationary ma- chines, replication is often the best and sometimes the only viable approach. Many replication solutions [4,16] assume a static infrastructure; that is, the connections This work was sponsored by the Advanced Research Projects Agency under contract DABT63-94-C-0080. Gerald Popek is also a-liated with CarsDirect.com. themselves may be transient but the connection location and the set of possible synchronization partners always remain the same. However, mobile users are by denition not static, and a replication service that forces them to adjust to a static infrastructure hinders mobility rather than enables it. Extraordinary actions, such as long distance telephone calls over low-bandwidth links, are necessary for users to conform to the underlying static model, costing additional time and money while providing a degraded service. Additionally, mobile users have di-culty inter-operating with other mobile users, because communication patterns and topologies are typically predened according to the underlying infrastruc- ture. Often, direct synchronization between mobile users is simply not permitted. Other systems [2,14,18] have simply traded the above communication problem for another one: scaling. They provide the ability for any-to- any synchronization, but their model suers from inherent scaling problems, limiting its usability in real environments. Good scaling behavior is s very important in the mobile scenario. Mobile users clearly require local replicas on their mobile machines. Yet, replicas must also be stored in the o-ce environment for reliability, intra- o-ce use by non-mobile personnel, and system- administration activities like backups. Addition- ally, typical methods for reducing replication fac- tors, such as local area network sharing tech- niques, are simply not feasible in the mobile con- text. Mobile users require local replicas of critical information, and in most cases desire local access to non-critical objects as well, for cost and performance reasons. The inability to scale well is as large an obstacle to the mobile user as the restriction of a static infrastructure discussed above. The main problem is that mobile users are replicating data using systems that were not designed for mobility. As such, instead of the replication system improving the state of mobile com- puting, it actually hinders mobility, as users nd themselves forced to adjust their physical motion and computing needs to better match what the system expects. This paper outlines the requirements of a replication service designed for the mobile context. We conclude with a description of Roam, a replication solution redesigned especially for mobile computing. Built using the Ward architecture [11], it enables rather than hinders mobil- ity, and provides a replication environment truly suited to mobile environments. 2. Replication Requirements Mobile users have special requirements above and beyond those of simple replication required by anyone wishing to share data. Here we discuss some of the requirements that are particular to mobile use: any-to-any communication, larger replication factors, detailed controls over replication behavior, and the lack of pre-motion actions. We omit discussion of well-understood ideas, such as the case for optimistic replication, discussed in [2,3,5,17]. 2.1. Any-to-any communication By denition, mobile users change their geographic location. As such, it cannot be predicted a priori what machines will be geographically collocated at any given time. Given that it is typically cheaper, faster, and more e-cient to communicate with a local partner rather than a remote one, mobile users want the ability to directly communicate and synchronize with whomever is \nearby." Consistency can be correctly maintained even if two machines cannot directly synchronize with each other, as demonstrated by systems based on the client-server model [4,16], but local synchronization increases usability and the level of functionality while decreasing the inherent synchronization cost. Users who are geographically collocated don't want updates to eventually propagate through a long-distance, sub-optimal path: the two machines are next to each other, and the synchronization should be instantaneous. Since users expect that nearby machines should synchronize with each other quickly and e- ciently, and it cannot be predicted which machines will be geographically collocated at any point in the future, a replication model capable of supporting any-to-any communication is required. That is, the model must allow any machine to communicate with any other machine|there can be no second-class clients in the system. Any-to-any communication is also required in other mobile arenas, such as in appliance mobility [6], the motion from device to device or system to system. For instance, given a desktop, a lap- top, and a palmtop, it is unlikely that one would want to impose a strict client-server relationship between the three; rather, one would want each to be able to communicate with any of the others. Providing any-to-any communication is equivalent to using a peer-to-peer replication model [10, 14,18]; if anyone can directly synchronize with anyone else, then everyone must by denition be equals, at least with respect to update-generation abilities. Some, however, have argued against peer models in mobile environments because of the relative insecurity regarding the physical devices themselves|for example, laptops are often stolen. The argument is that since mobile computers are physically less secure, they should be \second-class" citizens with respect to the highly secure servers located behind locked doors [15]. The class-based distinction is intended to provide improved security by limiting the potential security breach to only a second-class object. The argument is based on the assumption that security features must be encapsulated within the peer model, and therefore unauthorized access to any peer thwarts all security barriers and mecha- nisms. However, systems such as Truffles [13] have demonstrated that security policies can be modularized and logically situated around a peer replication framework while still remaining independent of the replication system. Truffles, an extension to the peer-based systems Ficus [2] and Rumor [14], incorporates encryption-based authentication and over-the-wire privacy and integrity services to increase a replica's condence in its peers. Truffles further supports protected denition and modication of security policies. For example, part of the security policy could be to only accept new le versions from specic (authenticated) replicas|which is eectively the degree of security provided by the \second-class" replicas mentioned above. With such an architecture, the problems caused by unauthorized access to a peer replica are no different from the unauthorized access of a client in a client-server model. Thus, the question of update- exchange topologies (any-to-any as compared to a more stylized, rigid structure as in client-server models) can be dealt with independently of the security issue and the question of how to enforce proper security controls. 2.2. Larger replication factors Most replication systems only provide for a handful of replicas of any given object. Ad- ditionally, peer algorithms have never traditionally scaled well. Finally, some have argued that peer solutions simply by their nature cannot scale well [15]. However, while mobile environments seem to require a peer-based solution (described above), they also seem to negate the assumption that a handful of replicas is enough. While we do not claim a need for thousands of writable copies, it does seem likely that the environments common today and envisioned for the near future will require larger replication factors than current systems allow. First and foremost, each mobile user requires a local replica on their laptop, doubling replication factors when data is stored both on the user's desktop and laptop. Additionally, although replication factors can often be minimized in o-ce environments due to LAN-style sharing and remote-access capabilities, such network-based le sharing cannot be utilized in mobile environments due to the frequency of network partitions and the wide range of available bandwidth and transfer latency. Second, consider the case of appliance mobility. The above discussion assumes that each user has one static machine and one mobile machine. The future will see the use of many more \smart" devices capable of storing replicated data. Palmtop computers are becoming more common, and there is even a wristwatch that can download calendar data from another machine. Researchers [19] have built systems that allow laptop and palmtop machines to share data dynamically and opportunis- tically. It is not di-cult to imagine other devices in the near future having the capability to store and conceivably update replicated data; such devices potentially increase replication factors dramatically Finally, some have argued the need for larger replication factors independent of the mobile sce- nario, such as in the case of air tra-c control [9]. Other scenarios possibly requiring larger replication factors include stock exchanges, network routing, airline reservation systems, and military command and control. Read-only strategies and other class-based techniques cannot adequately solve the scaling prob- slem, at least in the mobile scenario. Class-based solutions are not applicable to mobility, for the reasons described above (Section 2.1). Read-only strategies are not viable solutions because they force users to pre-select the writable replicas beforehand and limit the number of writable copies. In general one cannot predict which replicas require write-access and which ones do not. We must provide the ability for all replicas to generate updates, even though some may never do so. 2.3. Detailed replication controls By denition, a replication service provides users with some degree of replication control| a method of indicating what objects they want replicated. Many systems provide replication on a large-granularity basis, meaning that users requiring one portion of the container must locally replicate the entire container. Such systems are perhaps adequate in stationary environments, when users have access to large disk pools and network resources, but replication control becomes vastly more important to mobile users. Nomadic users do not in general have access to o-machine re- sources, and therefore objects that are not locally stored are eectively inaccessible. Everything the user requires must be replicated locally, which becomes problematic when the container is large. Replicating a large-granularity container means that some of the replicated objects will be deemed unimportant to the particular user. Unimportant data occupies otherwise usable disk space, which cannot be used for more critical objects. In the mobile context, where network disconnections are commonplace, important data that cannot be stored locally causes problems ranging from minor inconveniences to complete stoppages of work and productivity, as described by Kuenning [7]. Kuenning's studies of user behavior indicate that the set of required data can in fact be completely stored locally, but only if the underlying replication service provides the appropriate exibility to individually select objects for replication. Users and automated tools therefore require fairly detailed controls over what objects are replicated, because without them mobile users cannot adequately function. 2.4. Pre-motion actions One possible design point would have users \register" themselves as nomads for a specic time duration before becoming mobile. In doing so, the control structures and algorithms of the replication system could be greatly simplied; users would act as if they were stationary, and register their motion as the unusual case. For instance, suppose a user was taking a three-day trip from Los Angeles to New York. Before traveling, machines in Los Angeles and New York could exchange state to \re-congure" the user's portable to correctly interact with the machines in New York. Since replication requires underlying distributed algorithms, part of the reconguration process would require changing and saving the distributed state stored on the portable, to ensure correct algorithm execution. However, such a design policy drastically restricts the way in which mobility can occur, and does not match with the reality of mobile use. Mobility cannot always be predicted or scheduled. Often the chaos of real life causes unpredicted mo- bility: the car fails en route to work, freeway trafc causes unforeseeable delays, a child has to be picked up early from school, a family emergency occurs, or weather delays travel plans. Users are often forced to become mobile earlier or remain mobile longer than they had initially intended. In general, we cannot require that users know a priori either when they will become mobile or for how long. Additionally, this design policy makes underlying assumptions about the connectivity and accessibility of machines in the two aected geographic areas: Los Angeles and New York, in the above example. It assumes that before mobility occurs, the necessary machines are all accessible so the state-transformation operation can occur. Inaccessibility 5of any participant in this process blocks the user's mobility. Such a policy seems overly re- strictive, and does not match the reality of mobile use. Perhaps a user wants to change geographic locations precisely because a local machine is un- available, or perhaps a user needs to become mobile at an instant when connectivity is down between the multiple required sites. Since neither mobility nor connectivity can be predicted, one cannot make assumptions on the combination of the two. For these reasons, we believe that solutions that require \pre-motion" actions are not viable in the mobile scenario. Pre-motion actions force users to adapt to the system rather than having the system support the desired user behavior. Any real solution must provide the type of \get-up and go" functionality required by people for everyday use. 3. Roam Roam is a system designed to meet the above set of requirements. It is based on the Ward model [11] and is currently being implemented and tested at the University of California at Los Angeles 3.1. Ward model The Ward model combines classical elements of both the traditional peer-to-peer and client-server models, yielding a solution that scales well and provides replication exibility, allowing dynamic reconguration of the synchronization topology. The model's main grouping mechanism is the ward, or Wide Area Replication Domain. A ward is a collection of \nearby" machines, possibly only loosely connected. The denition of \nearby" depends on factors such as geographic location, expected network connectivity, bandwidth, latency, and cost; see [12] for a full discussion of these issues Wards are created as replicas are added to the system: each new replica chooses whether to join an existing ward or form a new one. We believe that it is possible to automate the assignment of ward membership, but as the issues involved are complex, we have avoided attempting to do so in the current system. Instead, this decision is controlled by a human, such as a system administrator or knowledgeable user. If necessary, the decision can be altered later by using ward-changing utilities. Although all members of the ward are equal peers, the ward has a designated ward master, similar to a server in a client-server model but with several important dierences: Since all ward members are peers, any two ward members can directly synchronize with one another. Typical client-server solutions do not allow client-to-client synchronization. Whether by design or by accident, mobile users will often encounter other mobile users; in such cases, direct access to the other ward member may be easier, cheaper and more e-cient than access to the ward master. Since all ward members are peers, any ward member can serve as the ward master. Automatic re-election and ward-master recongura- tion can occur should the ward master fail or become unavailable, and algorithms exist to re-solve multiple-master scenarios. Correctness is not aected by a transient ward master fail- ure, but the system maintains better consistency if the ward master is typically available and accessible. Since neither an inaccessible ward master nor multiple ward masters aects overall system correctness (see Section 3.2), the re-election problem is considerably easier than related distributed re-election problems. The ward master is not required to store actual data for all intra-ward objects, though it must be able to identify (i.e. name) the complete set. Most client-server strategies force the server to store a superset of each client's data. The ward master is the ward's only link with other wards; that is, only the ward master is aware of other replicas outside the ward. This is one manner in which the ward model achieves s wards.idraw Figure 1. The basic ward architecture. Overlapped members are a mobility feature (Section 3.4). good scaling|by limiting the amount of knowledge stored at individual replicas. Traditional peer models force every replica to learn about other replica's existence; in the ward model, replicas are only knowledgeable about the other replicas within their own ward. In fact, most replicas are completely unaware of the very existence of other wards. All ward masters belong to a higher-level ward, forming a two-level hierarchical model. 1 Ward masters act on their ward's behalf by bringing new updates into the ward, exporting others out of the ward, and gossiping about all known updates. Consistency is maintained across all replicas by having ward masters communicate directly with each other and allowing information to propagate independently within each ward. Figure 1 illustrates the basic architecture, as well as advanced features discussed in later sections. Wards are dynamically formed when replicas are created, and are dynamically maintained as suitable ward-member candidates change. Ward destruction occurs automatically when the last replica in a given ward is destroyed. 3.2. System correctness An important feature of the ward model is that system correctness does not depend on having precisely one master per ward. Even during recon- guration, updated les will ow between replicas without loss of information or other incorrect behavior. For most purposes, the ward master is simply another replica. Whether communicating within its own ward or with other wards, the master maintains consistency using the same algorithms as the non-master replicas. Thus, the propagation of information within and among wards follows from the correctness of these algorithms, 1 The rationale behind the two-level hierarchy and its impact on scaling is discussed in Section 3.5. rst described in [1]. If a ward master becomes temporarily unavail- able, information will continue to propagate between other ward members, due to the peer model. However, information will not usually propagate to other wards until the master returns. An exception to this rule will occur if a ward member temporarily or permanently moves to another ward, as described in Section 3.4, carrying data with it. If the master suers a permanent failure, a new master can be elected. We must then demonstrate that correctness will not suer during the transition to the new master. Correctness will be violated either if the failed master had some information that cannot be reconstructed, or if the failed master's participation is required for the completion of some distributed algorithm. The rst case can occur only if the lost information had been created at the master and had not yet propagated to another replica. In this case, the lost information cannot aect correctness because the situation is the same as if it had never existed. The second case is handled by distributed failure-recovery algorithms that are invoked when an administrator declares the old master as unrecoverable. If a new ward master is elected, there is a possibility of creating multiple masters. Correctness is not aected in this case because the master does not play any special role in the algorithms. The purpose of a ward master is not to coordinate the behavior of other ward members, but rather to serve as a conduit for information ow between wards. Multiple masters, like overlapped members (Section 3.4.2), will merely provide another communication path between wards. Since the peer-to-peer algorithms assume arbitrary communication patterns, correctness will not be aected by multiple ward masters. 3.3. Flexibility in the model Replication exibility is an important feature of the ward model. The set of data stored within each ward, called the ward set, is dynamically ad- justable, as is the set of ward members themselves. As ward members change their data demands and alter what replicated data they store locally, the ward set changes. Similarly, as mobile machines join or leave the ward, the set of ward participants changes. Both the ward set and ward membership are locally recorded and are replicated in an optimistic fashion. Additionally, each ward member, including the ward master, can locally store a dierent sub-set of the ward set. Such replication exibility, called selective replication [10] provides improved e-ciency and resource utilization: ward members locally store only those objects that they actively require. Replication decisions can be made manually or with automated tools [5,8]. Since the ward set varies dynamically, dier- ent wards might store dierent sets: not all ward sets will be equivalent. In essence, the model provides selective replication between wards them- selves. The reconciliation topologies and algorithms [10] apply equally well within a single ward and between ward masters. Brie y, the algorithms provide that machines communicate with multiple partners to ensure that each data object is synchronized directly with another replica. Addition- ally, the data synchronization algorithms support the reconciliation of non-local data via a third-party data-storage site, allowing the ward master to reconcile data that is not stored locally but is stored somewhere within the ward. 3.4. Support for mobility The model supports two types of mobility. Intra-ward mobility occurs when machines within the same ward become mobile within a limited geographic area; the machines encountered are all ward members. Since ward members are peers, direct communication is possible with any encountered machine. Intra-ward mobility might occur within a building, when traveling to a co-worker's house, or at a local coee shop. Perhaps more interesting, inter-ward mobility occurs when users travel (with their data) to another geographic region, encountering machines from another ward. Examples include businessmen traveling to remote o-ces and distant collaborators meeting at a common conference. Inter-ward mobility raises two main issues. First, recall that due to the model's replication exibility, two wards might not have identical ward sets. Thus, the mobile machine may store data objects not kept in the new ward, and vice- versa. Second, consider the typical patterns of mobility. Often users travel away from their \home location" for only a short time. The system would perform poorly if such transient mobile actions required global changes in data structures across multiple wards. On the other hand, mobile users occasionally spend long periods of time at other locations, either permanently or semi- permanently changing their denition of \home." In these scenarios, users should be provided with the same quality of service (in terms of local performance and time to synchronize data) as they experienced in their previous \home". Our solution resolves both issues by den- ing two types of inter-ward mobility|short-term (transient) and long-term (semi-permanent)|and providing the ability to transparently and automatically upgrade from the former to the latter. The two operations are called ward overlapping and ward changing respectively. Collectively, the two are called ward motion and enable peer-to- peer communication between any two replicas in the ward model, regardless of their ward membership 3.4.1. Ward changing Ward changing involves a long-term, perhaps permanent, change in ward membership. The moving replica physically changes its notion of its \home" ward, forgetting all information from the previous ward; similarly, the other participants in the old and new wards alter their notion of current membership. Ward membership information is maintained using the same optimistic algorithms that are used for replicating data, so that the problem of tracking membership in often- disconnected environments is straightforward. s The addition of a new ward member may change the ward set. Since the ward master is responsible for the inter-ward synchronization of all data in the ward set, the ward set must expand to properly encompass the replicated data stored at the moving replica. Similarly, the ward set at the old ward may shrink in size, as the ward set is dynamically and optimistically recalculated when ward membership changes. The ward-set changes propagate to other ward masters in an op- timistic, \need-to-know" fashion so that only the ward masters that care about the changes learn of them. Since both ward sets can potentially change, and these changes are eventually propagated to other ward masters, ward changing can be a heavyweight operation. However, users benet because all local data can be synchronized completely within the local ward, giving users the best possible quality of service and reconciliation performance 3.4.2. Ward overlapping In contrast, ward overlapping is intended as a very lightweight mechanism, and causes no global changes within the system. Only the new ward is aected by the operation. The localization of changes makes it a lightweight operation both to perform and to undo. Ward overlapping allows simultaneous multi- ward membership, enabling direct communication with the members of each ward. To make the mechanism lightweight, we avoid changing the ward sets by making the new replica an \over- lapped" member instead of a full-edged pant. Ward members (except for the ward master) cannot distinguish between real and overlapped members; the only dierence is in the management of the ward set. Instead of merging the existing ward set with the data stored on the mobile machine, the ward set remains unaltered. Data shared between the mobile machine and ward set can be reconciled locally with members of the new ward. However, data outside the new ward cannot be reconciled locally, and must either temporarily remain unsynchronized or else be reconciled with the original home ward. 3.4.3. Ward motion summary When a replica enters another ward, there are only two possibilities: the ward set can change or remain the same. The former creates a performance-improving but heavyweight solution; the latter causes a moderate performance degradation when synchronizing data not stored in the new ward but provides a very lightweight solution for transient mobile situations. Since both are operationally equivalent, the system can transparently upgrade from overlapping to changing if the motion seems more permanent than rst expected. Additionally, since ward formation is itself dy- namic, users can easily form mobile workgroups by identifying a set of mobile replicas as a new (possibly temporary) ward. By using ward over- lapping, mobile workgroups can be formed without leaving the old wards. Ward motion and dynamic ward formation and destruction allow easy and straightforward communication between any set of replicas in the entire system. 3.5. Scalability The scalability of the ward model is directly related to the degree of replication exibility. Ward sets can dynamically change in unpredictable ways; therefore the only method for a ward master to identify its ward set is to list each entry individually. The fully hierarchical generalization of the ward model to more than two levels faces scaling problems due to the physical problems of maintaining and indexing these lists of entries. Nevertheless, the proposed model scales well within its intended environment, and allows several hundred read-write replicas of any given ob- ject, meeting the demands of everyone from a single developer or a medium-sized committee to a large, international company. The model could be adapted to scale better by restricting the degree of replication freedom. For instance, if ward sets changed only in very regular fashions, they could be named as a unit instead of naming all members, dramatically improving scalability. However, we believe that replication exibility is an important design consideration in the targeted mobile envi- ronment, and one that users absolutely require, so we have chosen not to impose such regularity. 4. Performance 4.1. Disk space overhead Roam, like Rumor before it, stores its non-volatile data structures in lookaside databases within the volume but hidden from the user. From the user's perspective, anything other than his or her actual data is overhead and eectively shrinks the size of the disk. Minimal disk overhead is therefore an important and visible criterion for user satisfaction. Additionally, Roam is designed to be a scalable system. The Ward Model should support hundreds of replicas with minimal impact between wards. Specically, the creation of a new replica in ward X should not aect the disk overhead of the replicas in other wards. We therefore measured the disk overhead of Roam using two dierent volumes. The rst of these volumes was chosen as a typical representative of a user's personal subtree, while the second was chosen to stress Roam by storing small les that would exaggerate the system's space overhead After empirically measuring the overhead under dierent conditions, we tted equations to describe the overhead in terms of the number of les, types of les, number of replicas, and number of wards. These equations can be summarized as follows (full results are given in [12]): Each new directory costs 4.2KB object in the directory Each new le costs .24KB The rst replica within the ward, even without any user data, costs 57.36KB Each additional replica within the ward costs object stored at the replica Each new ward costs 6.44KB 4.2. Synchronization performance Since Roam's main task is the synchronization of data, we also measured the synchronization per- formance. We performed our experiments with two portable machines, in all cases minimizing extraneous processes to avoid non-repeatable ef- fects. One machine was a Dell Latitude XP with a 486DX4 running at 100MHz with 36MB of main memory, while the second was a TI TravelMate with a 133MHz Pentium and 64MB of main memory. Reconciliation was always performed by transferring data from the Dell machine to the TI machine. In other words, the reconciliation process always executed on the TI machine. Of course, reconciliation performance depends heavily on the sizes of the les that have been up- dated. Since Roam performs whole-le transfers, and any updated le must be transfered across the network in its entirety, we would expect reconciliation to take more time when more data has been updated. We therefore varied the amount of data updated from 0 to 100%, and within each trial we randomly selected the set of updated les. Since the les are selected at random, a given gure of X% is only an approximation of the amount of data updated, rather than an exact gure. In all mea- surements, we used the personal-subtree volume mentioned in Section 4.1, and performed at least seven trials at each data point. We performed ve dierent experiments under the above conditions. The rst two compared Roam and Rumor synchronization performance over a 10MB quiet Ethernet and WaveLAN wireless cards respectively. The third studied the effect of increasing numbers of replicas; the fourth studied the eect of increasing numbers of wards. The fth looked at the eects of selective replication [10] and dierent replication patterns on synchronization performance. These experiments showed that Roam is 10% s to 25% slower than Rumor when running with similar numbers of replicas. Most of the slow-down is due to Roam's more exible structure, which uses more processes and IPC to simplify the code and enhance scalability. Reconciliation of the 13.6MB volume under Roam takes from 46 to 206 seconds, depending on the transport mechanism, the number of les modied, and the number of replicas in the ward. We also studied the impact of multiple wards on the synchronization performance. We varied the number of wards from one, as in the previous ex- periments, to six. We placed three replicas within one of these wards, and measured the synchronization between two of them on the previously described portable machines. These experiments showed that at a 95% level of condence, adding wards has no impact on synchronization performance between two replicas. 4.3. Scalability We have already discussed some aspects of Roam's scalability, such as in disk space overhead (Section 4.1). However, another major aspect of scalability is the ability to create many replicas and still have the system perform well during synchronization. Synchronization performance includes two related issues. First, the reconciliation time for a given replica in a given ward should be largely unaected by the total number of replicas and wards. Second, the time to distribute an update from any replica to any other replica should presumably be faster in the Ward Model than in a standard approach (like Rumor), or else we have failed in our task. 4.3.1. Reconciliation time To measure the behavior of reconciliation time as the total number of replicas increases, we used a hybrid simulation. We created 64 replicas of our test volume, reducing the hardware requirements by using servers to store wards and replicas that were not actively participating in the experiments. Again, we found that at a 95% level of condence, the synchronization time does not change as the system conguration was varied from one ward with a total of three replicas to 7 wards with a total of 64 replicas. 4.3.2. Update distribution Another aspect of scalability concerns the distribution of updates to all replicas. A scalable system would presumably deliver updates to all replicas more quickly than a non-scalable system, at least at large numbers of replicas. Additionally, while it may not perform better at small numbers of replicas, a scalable system should at least not perform worse. Rather than measuring elapsed time, which depends on many complicated factors such as con- nectivity, network partitions, available machines, and reconciliation intervals, we considered the number of individual, pair-wise reconciliation ac- tions, and analytically developed equations that characterize the distribution of updates. We assume that there are M replicas; one of them, replica R, generates an update that must propagate to all other replicas. The following equations identify the number of separate reconciliation actions that must occur, both on the average and in the worst case, to propagate the update from R to some other replica S. In a non-ward system such as Rumor, since there are M replicas, M 1 of which do not yet have the update, and reconciliation uses a ring between all M replicas, we need M 1reconciliation actions on average. The worst case requires M 1 reconciliation actions. The analysis for Roam is a little more com- plicated. Assume that the M replicas are divided into N wards such that each ward has M=N mem- bers. Propagating an update from R to S requires rst sending it from R to R's ward master, then sending it from R's ward master to S's ward mas- ter, and then nally to S. Of course, if R and S are members of the same ward, then much of the expense is saved; however, we will solve the general problem rst before discussing the special case. Under the above conditions, we need 1( M reconciliation actions on average to distribute the update between a replica and its ward master, and2 actions on average between ward masters. From these building blocks, we calculate that, on average, Roam requires the following number of reconciliation actions:2 Note that when performance). Setting eliminates any benet from grouping. However, it is also interesting to note that when becomes M 1. Having only two wards does not improve the required time to distribute updates (although it does improve other aspects such as data structure size and network utilization). In general, Roam distributes updates faster than Rumor when 2 < N < M and M > 3; other- wise, Roam performs the same as Rumor (with respect to update distribution). From the two equations we calculate that the optimal number of wards for a given value of M is 2M. The above conditions yield a factor of three improvement at 50 replicas, and a factor of ve at 200 replicas. With a multi-level implementation, larger degrees of improvement are possible. The analysis for Roam also indicates that, in the worst case, Roam requires 2M reconciliation actions. As a special case, if R and S are in the same ward, only 1( M reconciliation actions are required on average, and M 1 in the worst case. 4.4. Ward motion Recall from Section 3.4 that Roam supports two dierent avors of ward motion: overlapping and changing. Overlapping is a lightweight, temporary form of motion that is easy to perform and undo. However, synchronization performance can become worse during overlapping. When the moving replica stores objects that are not part of the new ward, they must be synchronized with the original ward (or else remain unsynchronized during the presumably short time period). Changing is a more heavyweight, permanent form of motion that costs more but provides optimal synchronization performance in the new ward. We experimentally investigated the costs of both forms of ward motion, using the same 13.6MB volume used in the other tests. There are four types of costs involved in these operations: 1. Initial setup costs at the moving replica, 2. Disk overhead at the moving replica, 3. Costs imposed on other wards and ward mem- bers, and 4. Ongoing costs of synchronization. We summarize our results here; complete data is given in [12]. We found that setting up either type of motion took from 60 to 80 seconds, depending on the number of les stored on the local machine. Somewhat surprisingly, ward changing required only about 7% more elapsed time than overlapping. The disk overhead at the moving replica depends on the size of the destination ward. In essence, the other members of the destination ward must be tracked as if they were members of the replica's original ward, occupying 6.44KB plus 12 bytes per le (see Section 4.1). For ward over- lapping, this cost must be paid for both the original ward's members and the destination ward's members, while for ward changing, only the desti- nation's members must be tracked. In either case, these costs are insignicant compared to the space required by the volume itself. The costs imposed on other replicas and wards are minimal for ward overlapping, but for ward changing the old and new ward masters must change their ward sets, and these dierences will need to be propagated to all other ward masters by gossiping. However, the amount of information that must propagate is minimal (about 255 bytes per le that changes ward membership), so the additional network load is still quite low. s Finally, when wards are overlapped, synchronization costs increase because the moving replica must communicate with both the new ward and the old one. Synchronization with the (tempo- rary) new ward will take about the same amount of time as it would have in the original ward. However, to synchronize any les not available in the new ward, the original must be contacted. We measured these additional costs for various replication patterns in our test volume (described in [12]). To simulate the fact that communication with the original ward is probably long-distance and thus slower, we used a WaveLAN network for these experiments. We found that the synchronization time depends on both the number of les found only in the original ward, and on the number of modied les. When no les had been modied, the time was essentially constant at about 45 seconds. By contrast, when 100% of the (locally-stored) les had been modied, synchronization took from 78 to 169 seconds, depending on the exact set of les shared between the two ward masters. 5. Conclusion Replication is required for mobile computing, but today's replication services do not provide the features required by mobile users. Nomadicity requires a replication solution that provides any- to-any communication in a scalable fashion with su-ciently detailed control over the replication decisions. Roam was designed and implemented to meet these goals, paving the way not just to improved mobile computing but to new and better avenues of mobile research. Performance experiments have shown that Roam is indeed scalable and can handle the mobility patterns expected to be displayed by future users. --R Ficus: A Very Large Scale Reliable Distributed File System. The Little Work project. Disconnected operation in the Coda Presentation at the GloMo PI Meeting (February 4) at the University of Califor- nia Predictive File Hoarding for Disconnected Mobile Operation. Automated hoarding for mobile computers. Peer replication with selective control. The ward model: A scalable replication architecture for mobil- ity A Scalable Replication System for Mobile and Distributed Computing. The in uence of scale on distributed A highly available Experience with disconnected operation in a mobile computing environment. --TR Coda The Influence of Scale on Distributed File System Design FICUS: a very large scale reliable distributed file system Disconnected operation in the Coda File System Managing update conflicts in Bayou, a weakly connected replicated storage system Automated hoarding for mobile computers Seer Roam

mobile computing;file systems;replication

506906

Negotiation-based protocols for disseminating information in wireless sensor networks.

In this paper, we present a family of adaptive protocols, called SPIN (Sensor Protocols for Information via Negotiation), that efficiently disseminate information among sensors in an energy-constrained wireless sensor network. Nodes running a SPIN communication protocol name their data using high-level data descriptors, called meta-data. They use meta-data negotiations to eliminate the transmission of redundant data throughout the network. In addition, SPIN nodes can base their communication decisions both upon application-specific knowledge of the data and upon knowledge of the resources that are available to them. This allows the sensors to efficiently distribute data given a limited energy supply. We simulate and analyze the performance of four specific SPIN protocols: SPIN-PP and SPIN-EC, which are optimized for a point-to-point network, and SPIN-BC and SPIN-RL, which are optimized for a broadcast network. Comparing the SPIN protocols to other possible approaches, we find that the SPIN protocols can deliver 60% more data for a given amount of energy than conventional approaches in a point-to-point network and 80% more data for a given amount of energy in a broadcast network. We also find that, in terms of dissemination rate and energy usage, the SPIN protocols perform close to the theoretical optimum in both point-to-point and broadcast networks.

Introduction Wireless networks of sensors are likely to be widely deployed in the future because they greatly extend our ability to monitor and control the physical environment from remote lo- cations. Such networks can greatly improve the accuracy of information obtained via collaboration among sensor nodes and online information processing at those nodes. sensor networks improve sensing accuracy by providing distributed processing of vast quantities of sensing information (e.g., seismic data, acoustic data, high-resolution images, etc. When networked, sensors can aggregate such data to provide a rich, multi-dimensional view of the en- vironment. In addition, networked sensors can focus their Submitted to ACM Wireless Networks; an earlier version of this paper appeared in ACM MOBICOM '99. attention on critical events pointed out by other sensors in the network (e.g., an intruder entering a building). Finally, networked sensors can continue to function accurately in the face of failure of individual sensors; for example, if some sensors in a network lose a piece of crucial information, other sensors may come to the rescue by providing the missing data. sensor networks can also improve remote access to sensor data by providing sink nodes that connect them to other networks, such as the Internet, using wide-area wireless links. If the sensors share their observations and process these observations so that meaningful and useful information is available at the sink nodes, users can retrieve information from the sink nodes to monitor and control the environment from afar. We therefore envision a future in which collections of sensor nodes form ad hoc distributed processing networks that produce easily accessible and high-quality information about the physical environment. Each sensor node operates autonomously with no central point of control in the net- work, and each node bases its decisions on its mission, the information it currently has, and its knowledge of its com- puting, communication and energy resources. Compared to today's isolated sensors, tomorrow's networked sensors have the potential to perform with more accuracy, robustness and sophistication. Several obstacles need to be overcome before this vision can become a reality. These obstacles arise from the limited energy, computational power, and communication resources available to the sensors in the network. Energy: Because wireless sensors have a limited supply of energy, energy-conserving forms of communication and computation are essential to wireless sensor networks. Computation: Sensors have limited computing power and therefore may not be able to run sophisticated net-work protocols. Communication: The bandwidth of the wireless links connecting sensor nodes is often limited, on the order of a few hundred Kbps, further constraining inter- sensor communication. In this paper, we present SPIN (Sensor Protocols for Information via Negotiation), a family of negotiation-based information dissemination protocols suitable for wireless sensor networks. We designed SPIN to disseminate individual sensor observations to all sensors in a network, treating all (a) A (a) (a) (a) Figure 1: The implosion problem. In this graph, node A starts by ooding its data to all of its neighbors. Two copies of the data eventually arrive at node D. The system wastes energy and bandwidth in one unnecessary send and receive. sensors as potential sink nodes. SPIN thus provides a way of replicating complete views of the environment throughout an entire network. The design of SPIN grew out of our analysis of the different strengths and limitations of conventional protocols for disseminating data in a sensor network. Such protocols, which we characterize as classic ooding, start with a source node sending its data to all of its neighbors. Upon receiving a piece of data, each node then stores and sends a copy of the data to all of its neighbors. This is therefore a straightforward protocol requiring no protocol state at any node, and it disseminates data quickly in a network where bandwidth is not scarce and links are not loss-prone. Three deciencies of this simple approach render it inadequate as a protocol for sensor networks: Implosion: In classic ooding, a node always sends data to its neighbors, regardless of whether or not the neighbor has already received the data from another source. This leads to the implosion problem, illustrated in Figure 1. Here, node A starts out by ing data to its two neighbors, B and C. These nodes store the data from A and send a copy of it on to their neighbor D. The protocol thus wastes resources by sending two copies of the data to D. It is easy to see that implosion is linear in the degree of any node. Overlap: Sensor nodes often cover overlapping geographic areas, and nodes often gather overlapping pieces of sensor data. Figure 2 illustrates what happens when two nodes gather such overlapping data and then ood the data to their common neighbor (C). Again, the algorithm wastes energy and bandwidth sending two copies of a piece of data to the same node. Overlap is a harder problem to solve than the implosion problem|implosion is a function only of network topology, whereas overlap is a function of both topology and the mapping of observed data to sensor nodes. Resource blindness: In classic ooding, nodes do not modify their activities based on the amount of energy available to them at a given time. A network of embedded sensors can be \resource-aware" and adapt its communication and computation to the state of its energy resources. The SPIN family of protocols incorporates two key innovations that overcome these deciencies: negotiation and resource-adaptation. A s r Figure 2: The overlap problem. Two sensors cover an overlapping geographic region. When these sensors ood their data to node C, C receives two copies of the data marked r. To overcome the problems of implosion and overlap, SPIN nodes negotiate with each other before transmitting data. Negotiation helps ensure that only useful information will be transferred. To negotiate successfully, however, nodes must be able to describe or name the data they observe. We refer to the descriptors used in SPIN negotiations as meta-data. In SPIN, nodes poll their resources before data transmis- sion. Each sensor node has its own resource manager that keeps track of resource consumption; applications probe the manager before transmitting or processing data. This allows sensors to cut back on certain activities when energy is low, e.g., by being more prudent in forwarding third-party data. Together, these features overcome the three deciencies of classic ooding. The negotiation process that precedes actual data transmission eliminates implosion because it eliminates transmission of redundant data messages. The use of meta-data descriptors eliminates the possibility of overlap because it allows nodes to name the portion of the data that they are interested in obtaining. Being aware of local energy resources allows sensors to cut back on activities whenever their energy resources are low, thereby extending longevity. To assess the e-ciency of information dissemination via SPIN, we performed two studies of the SPIN approach based on two dierent wireless network models. In the rst study, we examined ve dierent protocols and their performance in a simple, point-to-point, wireless network where packets are never dropped and queueing delays never occur. Two of the protocols in this study are SPIN protocols (SPIN-PP and SPIN-EC). The other three protocols function as comparison protocols: (i) ooding, which we outlined above; (ii) gossiping, a variant on ooding that sends messages to random sets of neighboring nodes; and (iii) ideal, an idealized routing protocol in which each node has global knowledge of the status of all other nodes in the network, yielding the best possible performance. In the second study, we were interested in studying SPIN protocols in a more realistic wireless network model, where radios send packets over a single, unreliable, broadcast channel. SPIN-BC and SPIN- RL are two SPIN protocols that we designed specically for such networks, and we compare them to two other protocols, ooding and ideal. We evaluated each protocol under varying conditions by measuring the amount of data it transmitted and the amount of energy it used. The SPIN protocols disseminate information with low latency and conserve energy at the same time. Our results highlight the advantages of using meta-data to name data and negotiate data transmissions. SPIN-PP uses negotiation to solve the implosion and overlap problems; it reduces energy consumption by a factor of 3.6 compared to ooding, while disseminating data almost as quickly as theoretically possible. SPIN-EC, which additionally incorporates a threshold-based resource-awareness mechanism in addition to negotiation, disseminates 1.4 times more data per unit energy than ooding and in fact comes very close to the ideal amount of data that can be disseminated per unit energy. In a lossless, broadcast network with queueing delays, SPIN-BC reduces energy consumption by a factor of 1.6 and speeds up data dissemination by a factor of 1.8 compared to ooding. When the network loses packets, SPIN- RL is able to successfully recover from packet-losses, while still using half as much energy per unit data as ooding. The SPIN family of protocols rests upon two basic ideas. First, to operate e-ciently and to conserve energy, sensor applications need to communicate with each other about the data that they already have and the data they still need to obtain. Exchanging sensor data may be an expensive network operation, but exchanging data about sensor data need not be. Second, nodes in a network must monitor and adapt to changes in their own energy resources to extend the operating lifetime of the system. This section presents the individual features that make up the SPIN family of protocols. 2.1 Application-level Control Our design of the SPIN protocols is motivated in part by the principle of Application Level Framing (ALF) [4]. With ALF, network protocols must choose transmission units that are meaningful to applications, i.e., packetization is best done in terms of Application Data Units (ADUs). One of the important components of ALF-based protocols is the common data naming between the transmission protocol and application (e.g., [21]), which we follow in the design of our meta-data. We take ALF-like ideas one step further by arguing that routing decisions are also best made in application-controlled and application-specic ways, using knowledge of not just network topology but application data layout and the state of resources at each node. We believe that such integrated approaches to naming and routing are attractive to a large range of network situations, especially in mobile and wireless networks of devices and sensors. Because SPIN is an application-level approach to net-work communication, we intend to implement SPIN as middleware application libraries with a well dened API. These libraries will implement the basic SPIN message types, message handling routines, and resource-management functions. Sensor applications can then use these libraries to construct their own SPIN protocols. 2.2 Meta-Data Sensors use meta-data to succinctly and completely describe the data that they collect. If x is the meta-data descriptor for sensor data X, then the size of x in bytes must be shorter than the size of X, for SPIN to be benecial. If two pieces of actual data are distinguishable, then their corresponding meta-data should be distinguishable. Likewise, two pieces of indistinguishable data should share the same meta-data representation. SPIN does not specify a format for meta-data; this format is application-specic. Sensors that cover disjoint geographic regions may simply use their own unique IDs as meta-data. The meta-data x would then stand for \all the data gathered by sensor x". A camera sensor, in contrast, might use (x; is a geographic coordinate and is an orientation. Because each applica- tion's meta-data format may be dierent, SPIN relies on each application to interpret and synthesize its own meta- data. There are costs associated with the storage, retrieval, and general management of meta-data, but the benet of having a succinct representation for large data messages in SPIN far outweighs these costs. 2.3 SPIN Messages SPIN nodes use three types of messages to communicate: ADV { new data advertisement. When a SPIN node has data to share, it can advertise this fact by transmitting an ADV message containing meta-data. REQ { request for data. A SPIN node sends an REQ message when it wishes to receive some actual data. { data message. messages contain actual sensor data with a meta-data header. Because ADV and REQ messages contain only meta-data, they are smaller, and cheaper to send and receive, than their corresponding messages. 2.4 SPIN Resource Management SPIN applications are resource-aware and resource-adaptive. They can poll their system resources to nd out how much energy is available to them. They can also calculate the cost, in terms of energy, of performing computations and sending and receiving data over the network. With this informa- tion, SPIN nodes can make informed decisions about using their resources eectively. SPIN does not specify a particular energy management policy for its protocols. Rather, it species an interface that applications can use to probe their available resources. In this section, we present four protocols that follow the SPIN philosophy outlined in the previous section. Two of the protocols, SPIN-PP and SPIN-BC, tackle the basic problem of data transmission under ideal conditions, where energy is plentiful and packets are never lost. SPIN-PP solves this problem for networks using point-to-point transmission media, and SPIN-BC solves this problem for networks using broadcast media. The other two protocols, SPIN-EC and SPIN-RL, are modied versions of the rst two proto- cols, and they are meant to operate in networks that are not ideal. SPIN-EC, an energy-conserving version of SPIN-PP, reduces the number of messages it exchanges when energy in the system is low. SPIN-RL, a reliable version of SPIN-BC, recovers from losses in the network by selectively retransmitting messages. 3.1 SPIN-PP: A 3-Stage Handshake Protocol for Point- to-Point Media The rst SPIN protocol, SPIN-PP, is optimized for a networks using point-to-point transmission media, where it is possible for nodes A and B to communicate exclusively with each other without interfering with other nodes. In such a point-to-point wireless network, the cost of communicating with n neighbors in terms of time and energy is n times the cost of communicating with 1 neighbor. We start our study of SPIN protocols with a point-to-point network because of its relatively simple, linear cost model. The SPIN-PP protocol works in three stages (ADV-REQ- DATA), with each stage corresponding to one of the messages described above. The protocol starts when a node advertises new data that it is willing to disseminate. It does this by sending an ADV message to its neighbors, naming the new data (ADV stage). Upon receiving an ADV, the neighboring node checks to see whether it has already received or requested the advertised data. If not, it responds by sending an REQ message for the missing data back to the sender (REQ stage). The protocol completes when the initiator of the protocol responds to the REQ with a message, containing the missing data (DATA stage). Figure 3 shows an example of the protocol. Upon receiving an ADV packet from node A, node B checks to see whether it possesses all of the advertised data (1). If not, node B sends an REQ message back to A, listing all of the data that it would like to acquire (2). When node A receives the REQ packet, it retrieves the requested data and sends it back to node B as a DATA message (3). Node B, in turn, sends ADV messages advertising the new data it received from node A to all of its neighbors (4). It does not send an advertisement back to node A, because it knows that node A already has the data. These nodes then send advertisements of the new data to all of their neighbors, and the protocol continues. There are several important things to note about this example. First, if node B had its own data, it could aggregate this with the data of node A and send advertisements of the aggregated data to all of its neighbors (4). Second, nodes are not required to respond to every message in the protocol. In this example, one neighbor does not send an packet back to node B (5). This would occur if that node already possessed the data being advertised. Although this protocol has been designed for lossless net- works, it can easily be adapted to work in lossy or mobile networks. Here, nodes could compensate for lost ADV messages by re-advertising these messages periodically. Nodes can compensate for lost REQ and DATA messages by re-requesting data items that do not arrive within a xed time period. For mobile networks, changes in the local topology can trigger updates to a node's neighbor list. If a node notices that its neighbor list has changed, it can spontaneously re-advertise all of its data. This protocol's strength is its simplicity. Nodes using the protocol make very simple decisions when they receive new data, and they therefore waste little energy in compu- tation. Furthermore, each node only needs to know about its single-hop network neighbors. The fact that no other topology information is required to run the algorithm has some important consequences. First, SPIN-PP can be run in a completely uncongured network with a small startup cost to determine nearest neighbors. Second, if the topology of the network changes frequently, these changes only have to travel one hop before the nodes can continue running the algorithm. A A A ADV A A A (1) (2) Figure 3: The SPIN-PP Protocol. Node A starts by advertising its data to node B (1). Node B responds by sending a request to node A (2). After receiving the requested data (3), node B then sends out advertisements to its neighbors (4), who in turn send requests back to B (5,6). 3.2 SPIN-EC: SPIN-PP with a Low-Energy Threshold The SPIN-EC protocol adds a simple energy-conservation heuristic to the SPIN-PP protocol. When energy is plen- tiful, SPIN-EC nodes communicate using the same 3-stage protocol as SPIN-PP nodes. When a SPIN-EC node observes that its energy is approaching a low-energy threshold, it adapts by reducing its participation in the protocol. In general, a node will only participate in a stage of the protocol if it believes that it can complete all the other stages of the protocol without going below the low-energy threshold. This conservative approach implies that if a node receives some new data, it only initiates the three-stage protocol if it believes it has enough energy to participate in the full protocol with all of its neighbors. Similarly, if a node receives an advertisement, it does not send out a request if it does not have enough energy to transmit the request and receive the corresponding data. This approach does not prevent a node from receiving, and therefore expending energy on, ADV or messages below its low-energy threshold. It does, how- ever, prevent the node from ever handling a below this threshold. 3.3 SPIN-BC: A 3-Stage Handshake Protocol for Broadcast Media In broadcast transmission media, nodes in the network communicate using a single, shared channel. As a result, when a node sends out a message in a broadcast network, it is received by every node within a certain range of the sender 1 , regardless of the message's destination. If a node wishes to send a message and senses that the channel is currently in use, it must wait for the channel to become idle before This transmission range is determined by the power with which the sender transmitted the message and the sensitivity of the receiver. attempting to send the message. The disadvantage of such networks is that whenever a node sends out a message, all nodes within transmission range of that node must pay a price for that transmission, in terms of both time and en- ergy. However, the advantage of such networks is that, when a single node sends a message out to a broadcast address, this node can reach all of its neighbors using only one trans- mission. One-to-many communication is therefore 1=n times cheaper in a broadcast network than in a point-to-point net- work, where n is the number of neighbors for each node. improves upon SPIN-PP for broadcast networks by exclusively using cheap, one-to-many communi- cation. This means that all messages are sent to the broadcast address and thus processed by all nodes that are within transmission range of the sender. We justify this approach by noting that, since broadcast and unicast transmissions use the same amount of network resources in a broadcast network, SPIN-BC does not lose much e-ciency by using the broadcast address. Moreover, SPIN-BC nodes can coordinate their resource-conserving eorts more eectively because each node overhears all transactions that occur within its transmission range. For example, if two nodes A and B send requests for a piece of data to node C, C only needs to broadcast the requested data once in order to deliver the data to both A and B. Thus, only one node, either A or B, needs to send a request to C, and all other requests are redundant. If A and B address their requests directly to only C will hear the message, though all of the nodes within the transmission range of A and B will pay for two requests. However, if A and B address their requests to the broadcast address, all nodes within range will overhear these requests. Assuming that A and B are not perfectly synchro- nized, then either A will send its request rst or B will. The node who does not send rst will overhear the other node's request, realize that its own request is redundant, and suppress its own request. In this example, nodes that use the broadcast address can roughly halve their network resource consumption over nodes that do not. As we will illustrate shortly, this kind of approach, often called broadcast message-suppression, can be used to curtail the proliferation of redundant messages in the network. Like the SPIN-PP protocol, the SPIN-BC protocol has an ADV, REQ, and DATA stage, which serve the same purpose as they do in SPIN-PP. There are three central differences between between SPIN-PP and SPIN-BC. First, as mentioned above, all SPIN-BC nodes send their messages to the broadcast address, so that all nodes within transmission range will receive the messages. Second, SPIN-BC nodes do not immediately send out requests when they hear advertisements for data they need. Upon receiving an ADV, each node checks to see whether it has already received or requested the advertised data. If not, it sets a random timer to expire, uniformly chosen from a predetermined interval. When the timer expires, the node sends an REQ message out to the broadcast address, specifying the original advertiser in the header of the message. When nodes other than the original advertiser receive the REQ, they cancel their own request timers, and prevent themselves from sending out redundant copies of the same request. The nal difference between SPIN-PP and SPIN-BC is that a SPIN-PP node will send out the requested data to the broadcast address once and only once, as this is su-cient to get the data to all its neighbors (assuming a lossless network). It will not respond to multiple requests for the same piece of data. Figure 4 shows an example of the protocol. Upon receiving an ADV packet from node A, A's neighbors check to see A (1) A (2) A A F G ADV nodes without data nodes with data = transmission range nodes waiting to transmit REQ Figure 4: The SPIN-BC Protocol. Node A starts by advertising its data to all of its neighbors (1). Node C responds by broadcasting a request, specifying A as the originator of the advertisement (2), and suppressing the request from D. After receiving the requested data (3), E's request is also suppressed, and C, D, and E send advertisements out to their neighbors for the data that they received from A (4). whether they have received the advertised data (1). Three of neighbors, C, D, and E, do not have A's data, and enter request suppression mode for dierent, random amounts of time. C's timer expires rst, and C broadcasts a request for A's data (2), which in turn suppresses the duplicate request from D. Though several nodes receive the request, only A responds, because it is the originator of the original packet (3). After A sends out its data, E's request is suppressed, and C, D, and E all send out advertisements for their new data (4). 3.4 SPIN-RL: SPIN-BC for Lossy Networks SPIN-RL, a reliable version of SPIN-BC, can disseminate data e-ciently through a broadcast network, even if the net-work loses packets. The SPIN-RL protocol incorporates two adjustments to SPIN-BC to achieve reliability. First, each SPIN-RL node keeps track of which advertisements it hears from which nodes, and if it doesn't receive the data within a reasonable period of time following a request, the node re-requests the data. It lls out the originating-advertiser eld in the header of the REQ message with a destination, randomly picked from the list of neighbors that had advertised that specic piece of data. Second, SPIN-RL nodes limit the frequency with which they will resend data. If a SPIN-RL node sends out a message corresponding to a specic piece of data, it will wait a predetermined amount of time before responding to any more requests for that piece of data. A (a) (a) (a)(a) 4 Figure 5: Gossiping. At every step, each node only forwards data on to one neighbor, which it selects randomly. After node D receives the data, it must forward the data back to the sender (B), otherwise the data would never reach node C. 4 Other Data Dissemination Algorithms In this section, we describe the three dissemination algorithms against which we will compare the performance of SPIN. 4.1 Classic Flooding In classic ooding, a node wishing to disseminate a piece of data across the network starts by sending a copy of this data to all of its neighbors. Whenever a node receives new data, it makes copies of the data and sends the data to all of its neighbors, except the node from which it just received the data. The amount of time it takes a group of nodes to receive some data and then forward that data on to their neighbors is called a round. The algorithm nishes, or converges, when all the nodes in the network have received a copy of the data. Flooding converges in O(d) rounds, where d is the diameter of the network, because it takes at most d rounds for a piece of data to travel from one end of the network to the other. Although ooding exhibits the same appealing simplicity as SPIN-PP, it does not solve either the implosion or the overlap problem. 4.2 Gossiping Gossiping [9] is an alternative to the classic ooding approach that uses randomization to conserve energy. Instead of indiscriminately forwarding data to all its neighbors, a gossiping node only forwards data on to one randomly selected neighbor. If a gossiping node receives data from a given neighbor, it can forward data back to that neighbor if it randomly selects that neighbor. Figure 5 illustrates the reason that gossiping nodes forward data back to the sender. If node D never forwarded the data back to node B, node C would never receive the data. Whenever data travels to a node with high degree in a classic ooding network, more copies of the data start oating around the network. At some point, however, these copies may end up imploding. Gossiping avoids such implosion because it only makes one copy of each message at any node. The fewer copies made, the lower the likelihood that any of these copies will ever implode. While gossiping distributes information slowly, it dissipates energy at a slow rate as well. Consider the case where a single data source disseminates data using gossiping. Since A (c) (a,c) (a,c) (a) (c) (a) Figure dissemination of observed data a and c. Each node in the gure is marked with its initial data, and boxed numbers represent the order in which data is disseminated in the network. In ideal dissemination, both implosion, caused by B and C's common neighbor, and overlap, caused by A and C's overlapping initial data item, c, do not occur. the source sends to only one of its neighbors, and that neighbor sends to only one of its neighbors, the fastest rate at which gossiping distributes data is 1 node/round. Thus, if there are c data sources in the network, gossiping's fastest possible distribution rate is c nodes/round. Finally, we note that, although gossiping largely avoids implosion, it does not solve the overlap problem. 4.3 Ideal Dissemination Figure 6 depicts an example network where every node sends observed data along a shortest-path route and every node receives each piece of distinct data only once. We call this ideal dissemination because observed data a and c arrive at each node in the shortest possible amount of time. No energy is ever wasted transmitting and receiving useless data. Current networking solutions oer several possible approaches for dissemination using shortest-paths. One such approach is network-level multicast, such as IP multicast [5]. In this approach, the nodes in the network build and maintain distributed source-specic shortest-path trees and themselves act as multicast routers. To disseminate a new piece of data to all the other nodes in the network, a source would send the data to the network multicast group, thus ensuring that the data would reach all of the participants along shortest-path routes. In order to handle losses, the dissemination protocol would be modied to use reliable multicast. Unfortunately, multicast and particularly reliable multicast both rely upon complicated protocol machinery, much of which may be unnecessary for solving the specic problem of data dissemination in a sensor network. In many respects, SPIN may in fact be viewed as a form of application-level multicasting, where information about both the topology and data layout are incorporated into the distributed multicast trees. Since most existing approaches to shortest-path distribution trees would have to be modied to achieve ideal dis- semination, we will concentrate on comparing SPIN to the results of an ideal dissemination protocol, rather than its implementation. For point-to-point networks, it turns out that we can simulate the results of an ideal dissemination protocol using a modied version of SPIN-PP. We arrive at this simulation approach by noticing that if we trace the message history of the SPIN-PP protocol in a network, the messages in the network would match the history of an ideal dissemination protocol. Therefore, to simulate an RCApplication Resource Manager Network Interface RCAgent Network Neighbor Energy Link Link Link Meta-Data Data Meta-Data Data Resource-Adaptive Node Figure 7: Block diagram of a Resource-Adaptive Node. ideal dissemination protocol for point-to-point networks, we run the SPIN-PP protocol and eliminate any time and energy costs that ADV and REQ messages incur. Dening an ideal protocol for broadcast networks is more tricky. We approximate an ideal dissemination protocol for broadcast networks by running the SPIN-BC protocol on a lossless net-work and eliminating any time and energy costs that ADV and REQ messages would incur. 5 Point-to-Point Media Simulations In order to study the SPIN-PP and SPIN-EC approaches discussed in the previous sections, we developed a sensor net-work simulator by extending the functionality of the ns software package. Using this simulation framework, we compared SPIN-PP and SPIN-EC with classic ooding and gossiping and the ideal data distribution protocol. We found that SPIN-PP provides higher throughput than gossiping and the same order of throughput as ooding, while at the same time it uses substantially less energy than both these protocols. SPIN-EC is able to deliver even more data per unit energy than SPIN-PP and close to the ideal amount of data per unit energy by adapting to the limited energy of the network. We found that in all of our simulations, nodes with a higher degree tended to dissipate more energy than nodes with a lower degree, creating potential weak points in a battery-operated network. 5.1 ns Implementation ns [16] is an event-driven network simulator with extensive support for simulation of TCP, routing, and multicast pro- tocols. To implement the SPIN-PP and SPIN-EC protocols, we added several features to the ns simulator. The ns Node class was extended to create a Resource-Adaptive Node, as shown in Figure 7. The major components of a Resource-Adaptive Node are the Resources, the Resource Manager, the Resource-Constrained Application (RCApplication), the Resource-Constrained Agent (RCAgent) and the Network Interface. The Resource Manager provides a common interface between the application and the individual resources. The RCApplication, a subclass of ns's Application class, is responsible for updating the status of the node's resources through the Resource Manager. In addition, the RCApplica- tion implements the SPIN communication protocol and the resource-adaptive decision-making algorithms. The RCA- gent packetizes the data generated by the RCApplication and sends the packets to the Node's Network Interface for -5515Test Network Meters Meters Figure 8: Topology of the 25-node, wireless test network. The edges shown here signify communicating neighbors in a point-to-point wireless medium. transmission to one of the node's neighbors. For each point- to-point link that would exist between neighboring nodes in a wireless network, we created a wired link using ns's built-in link support. We made these wired links appear to be wireless by forcing them to consume the same amount of time and energy that would accompany real, wireless link communications. 5.2 Simulation Testbed For our simulations, we used the 25-node network shown in Figure 8. This network, which was randomly generated with the constraint that the graph be fully connected, has 59 edges, a degree of 4.7, a hop diameter of 8, and an average shortest path of 3.2 hops. The power of the sensor radio transmitter is set so that any node within a 10 meter radius is within communication range and is called a neighbor of the sensor. The radio speed (1 Mbps) and the power dissipation (600 mW in transmit mode, 200 mW in receive mode) were chosen based on data from currently available radios. The processing delay for transmitting a message is randomly chosen between 5 ms and 10 ms 2 . We initialized each node with 3 data items, chosen randomly from a set of 25 possible data items. This means there is overlap in the initial data of dierent sensors, as often occurs in sensor networks. The size of each data item was set to 500 bytes, and we gave each item a distinct, byte, meta-data name. Our test network assumes no net-work losses and no queuing delays. Table 1 summarizes these network characteristics. Using this network conguration, we ran each protocol and tracked its progress in terms of the rate of data distribution and energy usage. For each set of results, we ran the simulation 10 times and averaged the data distribution times and energy usage to account for the random processing delay. The results of these simulations are presented in the following sections. 5.3 Unlimited Energy Simulations For the rst set of simulations, we gave all the nodes a virtually innite supply of energy and simulated each data distribution protocol until it converged. Since energy is not lim- ited, SPIN-PP and SPIN-EC are identical protocols. There- 2 Note that these simulations do not account for any delay caused by accessing, comparing, and managing meta-data. Nodes Edges 59 Average degree 4.7 neighbors Diameter 8 hops Average shortest path 3.2 hops reach Radio propagation delay 3x10 8 m/s Processing delay 5-10 ms speed 1 Mbps Transmit cost 600 mW Receive cost 200 mW Data size 500 bytes Meta-data size 16 bytes Network losses None Queuing delays None Table 1: Characteristics of the 25-node wireless test net-work fore, the results in this section only compare SPIN-PP with ooding, gossiping, and the ideal data distribution protocol. 5.3.1 Data Acquired Over Time Figure 9 shows the amount of data acquired by the network over time for each of the protocols. These graphs clearly show that gossiping has the slowest rate of convergence. However, it is interesting to note that using gossiping, the system has acquired over 85% of the total data in a small amount of time; the majority of the time is spent distributing the last 15% of the data to the nodes. This is because a gossiping node sends all of the data it has to a randomly chosen neighbor. Because the nodes obtain a large amount of data, this transmission will be costly, and since it is very likely that the neighbor already has a large proportion of the data which is being transmitted, it will also be very wasteful. A gossiping protocol which kept some per-neighbor state, such as having each node keep track of the data it has already sent to each of its neighbors, would perform much better by reducing the amount of wasteful transmissions. Figure 9 shows that SPIN-PP takes 80 ms longer to converge than ooding, whereas ooding takes only ms longer to converge than ideal. Although it appears that SPIN-PP performs much worse than ooding in convergence time, this increase is actually a constant amount, regardless of the length of the simulation. Thus for longer simulations, the increase in convergence time for the SPIN-PP protocol will be negligible. Our experimental results showed that the data distribution curves were convex for all four protocols. We therefore speculated that these curves might generally be convex, regardless of the network topology. If we could predict the shape of these curves, we might be able to gain some intuition about the behavior of the protocols for dierent net-work topologies. To do this, we noted that the amount of data received by a node i at each round d depends only on the number of neighbors d hops away from this node, n i (d). However, since n i (d) is dierent for each node i and each distance d and is entirely dependent on the specic topol- ogy, we found that, in fact, no general conclusions can be drawn about the shape of these curves. Time Total Data Total Data Acquired in the Sensor Network Flooding Gossiping Total Data Total Data Acquired in the Sensor Network Flooding Gossiping Figure 9: Percent of total data acquired in the system over time for each protocol. (a) shows the entire time scale until all the protocols converge. (b) shows a blow-up of the rst seconds. 5.3.2 Energy Dissipated Over Time For the previous set of simulations, we also measured the energy dissipated by the network over time, as shown in Figure 10. These graphs show that gossiping again is the most costly protocol; it requires much more energy than the other two protocols to accomplish the same task. As stated before, adding a small amount of state to the gossiping protocol will dramatically reduce the total system energy usage. Figure also shows that SPIN-PP uses approximately a factor of 3.5 less energy than ooding. Thus, by sacric- ing a small, constant oset in convergence time, SPIN-PP achieves a dramatic reduction in system energy. SPIN-PP is able to achieve this large reduction in energy since there is no wasted transmission of the large 500-byte data items. We can see this advantage of the SPIN-PP protocol by looking at the message proles for the dierent protocols, shown in Figure 11. The rst three bars for each protocol show the number of data items transmitted throughout the network, the number of these data items that are redundant and thus represent wasteful transmission, and the number Energy Dissipated Total Energy Dissipated in the Sensor Network Flooding Gossiping Energy Dissipated Total Energy Dissipated in the Sensor Network Flooding Gossiping Figure 10: Total amount of energy dissipated in the system for each protocol. (a) shows the entire time scale until all the protocols converge. (b) shows a blow-up of the rst 0.22 seconds. of data items that are useful. The number of useful data transmissions is the same for each protocol since the data distribution is complete once every node has all the data. The last three bars for each protocol show the number of meta-data items transmitted and the number of these items that are redundant and useful. These bars have a height zero for ideal, ooding, and gossiping, since these protocols do not use meta-data transmissions. Note that the number of useful meta-data transmissions for the SPIN-PP protocol is three times the number of useful data transmissions, since each data transmission in the SPIN-PP protocol requires three messages with meta-data. Flooding and gossiping nodes send out many more data items than SPIN-PP nodes. Furthermore, 77% of these data items are redundant for ooding and 96% of the data items are redundant for gossiping, and these redundant messages come at the high cost of 500 bytes each. SPIN-PP nodes also send out a large number of redundant messages however, these redundant messages are meta-data messages. Meta-data messages come at a relatively low cost and come with an important benet: meta-data negotiation500015000 Redundant data Data items Meta-data items Useful meta-data items received items received sent/received sent/received Flooding Gossiping Useful data items received Redundant meta-data items received Protocol Number of Messages Figure 11: Message proles for the unlimited energy simu- lations. Notice that SPIN-PP does not send any redundant data messages. 90.080.120.16Number of neighbors Energy dissipated Energy Dissipated per Node Versus Number of Neighbors Flooding Figure 12: Energy dissipation versus node degree for unlimited energy simulations. keeps SPIN-PP nodes from sending out even a single redundant data-item. We plotted the average energy dissipated for each node of a certain degree, as shown in Figure 12. This gure shows that for all the protocols, the energy dissipated at each node depends upon its degree. The repercussions of this nding is that if a high-degree node happens to lie upon a critical path in the network, it may die out before other nodes and partition the network. We believe that handling such situations is an important area for improvement in all four protocols. The key results from these unlimited energy simulations are summarized in Table 2. 5.4 Limited Energy Simulations For this set of simulations, we limited the total energy in the system to 1.6 Joules to determine how eectively each protocol uses its available energy. Figure 13 shows the data acquisition rate for the SPIN-PP, SPIN-EC, ooding, gos- siping, and ideal protocols. This gure shows that SPIN- Performance Protocol Relative to Ideal SPIN-PP Flooding Gossiping Increase in Energy 1.25x 4.5x 25.5x Dissipation Increase in 90 ms 10 ms 3025 ms Convergence Time Slope of Energy 1.25x 5x 25x Dissipation vs. Node Degree Correlation Line % of Total Data 0 77% 96% Messages that are Redundant Table 2: Key results of the unlimited energy simulations for the SPIN-PP, ooding, and gossiping protocols compared with the ideal data distribution protocol. EC puts its available energy to best use and comes close to distributing the same amount of data as the ideal pro- tocol. SPIN-EC is able to distribute 73% of the total data as compared with the ideal protocol which distributes 85%. We note that SPIN-PP distributes 68%, ooding distributes 53%, and gossiping distributes only 38%. Figure 14 shows the rate of energy dissipation for this set of simulations. This plot shows that ooding uses all its energy very quickly, whereas gossiping, SPIN-PP, and SPIN-EC use the energy at a slower rate and thus are able to remain operational for a longer period of time. Figure 15 shows the number of data items acquired per unit energy for each of the protocols. If the system energy is limited to below 0.2 Joules, none of the protocols has enough energy to distribute any data. With 0.2 Joules, the gossiping protocol is able to distribute a small amount of data; with Joules, the SPIN protocols begins to distribute data; and with 1.1 Joules, the ooding protocol begins to distribute the data. This shows that if the energy is very limited, the gossiping protocol can accomplish the most data distribu- tion. However, if there is enough energy to get the ooding or one of the SPIN protocols started, these protocols deliver much more data per unit energy than gossiping. This graph also shows the advantage of SPIN-EC over SPIN-PP, which doesn't base any decisions on the current level of its resources. By making the communication decisions based on the current level of the energy available to each node, SPIN- EC is able to distribute 10% more data per unit energy than SPIN-PP and 60% more data per unit energy than ooding. 6 Broadcast Media Simulations For our second study, we examined the use of SPIN protocols in a single, shared-media channel. The nodes in this model use the 802.11 MAC layer protocol to gain access to the channel. Packets may be queued at the nodes themselves or may be lost due to transmission errors or channel collisions. We used this framework to compare the performance of SPIN-BC, SPIN-RL, ooding, and an ideal data distribution protocol. We found that SPIN-RL is able to use meta-data to successfully recover from packet losses, while acquiring twice as much data per unit energy as ooding. Because ooding does not have any built-in mechanisms for providing reliability, it can not recover from packet losses and never converges. Total Data Total Data Acquired in the Sensor Network Flooding Gossiping Figure 13: Percent of total data acquired in the system for each protocol when the total system energy is limited to 1.6 Joules. Time Energy Dissipated Total Energy Dissipated in the Sensor Network Flooding Gossiping Figure 14: Energy dissipated in the system for each protocol when the total system energy is limited to 1.6 Joules. 6.1 Simulation Implementation and Setup We used monarch, a variant of the ns simulator for all the simulations in this study. monarch [14] extends the functionality of ns to enable the simulation of realistic wireless com- munication. These extensions include a radio propagation model and a detailed simulation of the IEEE 802.11 DCF MAC protocol. We extended monarch's MobileNode class to create wireless Resource-Adaptive Nodes. The only difference between these Resource-Adaptive Nodes and those described in Section 5 is that we replaced the wired Network Interface shown in Figure 7 with a wireless 802.11 MAC in- terface. We also made several modications to monarch's built-in 802.11 MAC implementation in order to perform our simulations. First, we modied the MAC implementation to appropriately subtract energy from a node's Energy Resource whenever it sends and receives a packet. Second, we added a switch to the MAC layer that turns collisions and losses. The simulation testbed that we used in our second study is the same as the testbed used in our rst study. We used Dissipated (J) Total Data Total Data Acquired per Amount of Energy Flooding Gossiping Figure 15: Data acquired for a given amount of energy. SPIN-EC distributes 10% more data per unit energy than SPIN-PP and 60% more data per unit energy than ooding. the same topology and radio characteristics as those given in Figure 8 and in Table 1. The only dierences between these two studies are that packets in this study may experience queueing delays and, depending upon the test congu- ration, may also be lost due to multi-path fading or packet collisions. 6.2 Simulations without Packet Losses For the rst set of simulations, we gave all the nodes a virtually innite supply of energy, turned losses, and ran each data distribution protocol until it converged. 6.2.1 Data Acquired Over Time Figure shows the amount of data acquired by the net-work over time for each of the protocols. These graphs show that SPIN-BC converges faster than ooding, and almost as quickly as the ideal protocol. The dierence in convergence times between SPIN-BC and ooding can be explained by queueing delays in the network. Recall that in a broadcast network, each node must wait for the channel to become in order to send out a packet. When many nodes in a small area have packets to send, these nodes queue up their packets while waiting for access to the channel. If some of these packets are redundant, than they cause other, useful packets in the network to wait needlessly in queues. Flooding does not provide any mechanisms to circumvent implosion and overlap and therefore sends out many useless packets, as shown in Figure 17. These packets therefore cause unnecessary delays in the running time of the ooding algorithm. 6.2.2 Energy Dissipated Over Time For the previous set of simulations, we also measured the energy dissipated by the network over time, as shown in Figure 18. These gures show that SPIN-BC reduces energy consumption by a factor of 1.6 over ooding. We can see the advantage of the SPIN-BC protocol by examining the message proles for each protocol given in Figure 17. Because these protocols all use broadcast, some redundant data-transmissions are unavoidable, as illustrated by the Time Total Data Total Data Acquired in the Sensor Network Flooding Figure Percent of total data acquired in the system over time in a lossless broadcast-network. Performance Protocol relative to Ideal no losses losses Flooding SPIN-RL Increase in Energy 1.6x 2.4x 1.6x Dissipation Increase in 1.1x 2x 5x Convergence Time Slope of Energy .11x 1.67x 1.6x Dissipation vs. Node Degree Correlation Line Total Data 1x 2.2x .89x Messages received % of Total Data 1.1x 1.8x .96x Messages that are Redundant Table 3: Key results of the broadcast network simulations compared with the ideal data distribution protocol. ideal protocol's message prole. What this gure illustrates is that, by sacricing small amounts of energy sending meta-data messages, SPIN-BC achieves a dramatic reduction in wasted data messages and a corresponding reduction in system energy and convergence time. Figure 20 further reinforces these results, showing that SPIN-BC nodes acquire times more data per unit energy expended than ooding. The key results from these simulations are summarized in Table 3. 6.3 Simulations with Packet Losses For the second set of simulations, we gave all the nodes a virtually innite supply of energy and allowed the MAC layer to lose packets due to collisions and transmission errors. We compare SPIN-RL, our reliable protocol, to SPIN-BC and ooding. As a point of reference, we also compare SPIN- RL to the ideal protocol, run in a lossless network. We ran each protocol until it either converged or ceased to make any progress towards converging. Number of items received Data items Redundant data items Useful data items Meta-data items Redundant meta-data items Useful meta-data items Figure 17: Message proles for each protocol in a lossless broadcast-network. 6.3.1 Data Acquired Over Time Figure 21 shows the amount of data acquired by the net-work over time for each of the protocols. Only three of the protocols, namely SPIN-BC, SPIN-RL, and ooding were run on a lossy network. The ideal protocol was run on a lossless network, and is provided as a best-case reference point. Of all three of the protocols run on the lossy network, SPIN-RL is the only protocol that will retransmit lost pack- ets, and therefore is the only protocol that converges. It is interesting to note that, although SPIN-BC outperformed ooding in the lossless network, it does not perform as well as ooding in a lossy network. We can account for SPIN- BC's poor performance by the fact that SPIN-BC nodes must successfully send and receive three messages in order to move a piece of data over a hop in the network, whereas ooding nodes only have to send one. SPIN-BC's protocol is therefore three times more vulnerable to network losses than ooding, which explains the dierence in behavior we see between Figures 16 and 21. 6.3.2 Energy Dissipated Over Time For the previous set of simulations, we also measured the energy dissipated by the network over time, as shown in Figure 22. These gures show that, of all the protocols, SPIN-RL expends the most energy, only slightly more than ooding. We can account for the relative energy expenditure of each protocol by examining the message proles, given in Figure 25. Of all the protocols, SPIN-RL nodes receive the most data messages, as well as the most meta-data messages. This extra expenditure is well justied, however, if we look at how it is put to use. Figure 24 shows the amount of data acquired per unit energy for each protocol. Using almost the same amount of energy, SPIN-RL is able to acquire twice the amount of data as ooding. The key results from these simulations are summarized in Table 3. 7 Related Work Perhaps the most fundamental use of dissemination protocols in networking is in the context of routing table dissem- Time Energy Dissipated Total Energy Dissipated in the Sensor Network Flooding Figure Total amount of energy dissipated in the system for each protocol in a lossless broadcast-network. ination. For example, nodes in link-state protocols (such as OSPF [15]) periodically disseminate their view of the net-work topology to their neighbors, as discussed in [10, 25]. Such protocols closely mimic the classic ooding protocol we described earlier. There are generally two types of topologies used in wireless networks: centralized control and peer-to-peer communications [17]. The latter style is better suited for wireless sensor networks than the former, given the ad hoc, decentralized nature of such networks. Recently, mobile ad hoc routing protocols have become an active area of research [3, 11, 18, 20, 24]. While these protocols solve important problems, they are a dierent class of problems from the ones that arise in wireless sensor networks. In particular, we believe that sensor networks will benet from application-controlled negotiation-based dissemination protocols, such as SPIN. Routing protocols based on minimum-energy routing [12, 23] and other power-friendly algorithms have been proposed in the literature [13]. We believe that such protocols will be useful in wireless sensor networks, complementing SPIN and enabling better resource adaptation. Recent advances in operating system design [7] have made application-level approaches to resource adaptation such as SPIN a viable alternative to more traditional approaches. Using gossiping and broadcasting algorithms to disseminate information in distributed systems has been extensively explored in the literature, often as epidemic algorithms [6]. In [1, 6], gossiping is used to maintain database consistency, while in [19], gossiping is used as a mechanism to achieve fault tolerance. A theoretical analysis of gossiping is presented in [9]. Recently, such techniques have also been used for resource discovery in networks [8]. Close in philosophy to the negotiation-based approach of SPIN is the popular Network News Transfer Protocol (NNTP) for Usenet news distribution on the Internet [2]. Here, news servers form neighborhoods and disseminate new information between each other, using names and timestamps as meta-data to negotiate data dissemination. There has been a lot of recent interest in using IP multicast [5] as the underlying infrastructure to e-ciently and Number of neighbors Energy dissipated Energy Dissipated per Node Versus Number of Neighbors Flooding Figure 19: Energy dissipation versus node degree in a loss-less broadcast-network. Energy Dissipated (J) Total Data Total Data Acquired Per Unit Energy Floodiing Figure 20: Energy dissipated versus data acquired in a loss-less broadcast-network. Time Total Data Total Data Acquired in the Sensor Network Flooding Figure 21: Percent of total data acquired in the system over time for each protocol in a lossy broadcast-network. reliably disseminate data from a source to many receivers [22] on the Internet. However, for the reasons described in Section 4, we believe that enabling applications to control routing decisions is a less complex and better approach for wireless sensor networks. Conclusions In this paper, we introduced SPIN (Sensor Protocols for Information via Negotiation), a family of data dissemination protocols for wireless sensor networks. SPIN uses meta-data negotiation and resource-adaptation to overcome several deciencies in traditional dissemination approaches. Using meta-data names, nodes negotiate with each other about the data they possess. These negotiations ensure that nodes only transmit data when necessary and never waste energy on useless transmissions. Because they are resource-aware, nodes are able to cut back on their activities whenever their resources are low to increase their longevity. We have discussed the details of four specic SPIN pro- tocols, SPIN-PP and SPIN-EC for point-to-point networks, and SPIN-BC and SPIN-RL for broadcast networks. SPIN- PP is a 3-stage handshake protocol for disseminating data, and SPIN-EC is a version of SPIN-PP that backs o from communication at a low-energy threshold. SPIN-BC is a variant of SPIN-PP that takes advantage of cheap, MAC-layer broadcast, and SPIN-RL is a reliable version of SPIN- BC. Finally, we compared the SPIN-PP, SPIN-EC, SPIN- BC, and SPIN-RL protocols to ooding, gossiping, and ideal dissemination protocols using the ns simulation tool. After examining SPIN in this paper, both qualitatively and quantitatively, we arrive at the following conclusions: Naming data using meta-data descriptors and negotiating data transmissions using meta-data successfully solve the implosion and overlap problems described in Section 1. The SPIN protocols are simple and e-ciently disseminate data, while maintaining only local information about their nearest neighbors. These protocols are well-suited for an environment where the sensors are Time Energy Dissipated Total Energy Dissipated in the Sensor Network Flood Figure 22: Total amount of energy dissipated in the system for each protocol in a lossy broadcast-network. mobile because they base their forwarding decisions on local neighborhood information. In terms of time, SPIN-PP achieves comparable results to classic ooding protocols, and in some cases outperforms classic ooding. In terms of energy, SPIN-PP uses only about 25% as much energy as a classic ing protocol. SPIN-EC is able to distribute 60% more data per unit energy than ooding. In all of our ex- periments, SPIN-PP and SPIN-EC outperformed gos- siping. They also come close to an ideal dissemination protocol in terms of both time and energy under some conditions. Perhaps surprisingly, SPIN-BC and SPIN-RL are able to use one-to-many communications exclusively, while still acquiring data faster than ooding using less en- ergy. Not only can SPIN-RL converge in the presence of network packet losses, it is able to dissipate twice the amount of data per unit energy as ooding. In summary, SPIN protocols hold the promise of achieving high performance at a low cost in terms of complexity, en- ergy, computation, and communication. Although our initial work and results are promising, there is still work to be done in this area. Though we have discussed energy-conservation in terms of point-to-point media and reliability in terms of broadcast media, we would like to explore methods for combining these techniques for both kinds of networks, and we do not believe this would be di-cult to accomplish. We would also like to study SPIN protocols in a mobile wireless network model. We expect that these networks would challenge the speed and adaptiveness of SPIN protocols in a way that stationary networks do not. Finally, we would like to develop more sophisticated resource-adaptation protocols to use available energy well. In particular, we are interested in designing protocols that make adaptive decisions based not only on the cost of communicating data, but also the cost of synthesizing it. Such resource-adaptive approaches may hold the key to making compute-intensive sensor applications a reality in the future. 90.040.080.120.16Number of neighbors Energy dissipated Energy Dissipated per Node Versus Number of Neighbors Flooding Figure 23: Energy dissipation versus node degree for each protocol in a lossy broadcast-network. Energy Dissipated (J) Total Data Total Data Acquired Per Unit Energy Flooding Figure 24: Energy dissipated versus data acquired for each protocol in a lossy broadcast-network. The symbol in the second graph highlights the last data-point of the SPIN-BC line. Flooding50015002500Number of items received Data items received Redundant data items Useful data items Meta-data items(x.1) Redundant meta-data items (x.1) Useful meta-data items (x.1) Figure 25: Message proles for the each protocol in a lossy broadcast-network. Acknowledgments We thank Wei Shi, who participated in the initial design and evaluation of some of the work in this paper. We thank Anantha Chandrakasan for his helpful comments and suggestions throughout this work. We also thank Suchitra Raman and John Wroclawski for several useful comments and suggestions on earlier versions of this paper. This research was supported in part by a research grant from the NTT Corporation and in part by DARPA contract DAAN02-98- K-0003. Wendi Heinzelman is supported by a Kodak Fellowship --R Epidemic Algorithms in Replicated Databases. Network News Transport Protocol. A Performance Comparison of Multi-Hop Wireless Ad Hoc Netowrk Routing Protocols Architectural Consideration for a New Generation of Protocols. Multicast Routing in Datagram Internetworks and Extended LANs. Epidemic Algorithms for Replicated Database Maintenance. An operating system architecture for application-level resource management Resource Discovery in Distributed Networks. A Survey of Gossiping and Broadcasting in Communication Networks. Routing in the Internet. Routing in Ad Hoc Networks of Mobile Hosts. Spectral E-ciency Considerations for Packet Radio Distributed Network Protocols for Wireless Communication. Monarch Extensions to the ns-2 Network Simulator Information Networks. A Highly Adaptive Distributed Routing Algorithm for Mobile Wireless Ne- towrks Highly Dynamic Destination-Sequenced Distance-Vector Routing (DSDV) for Mobile Computers Scalable Data Naming for Application Level Framing in Reliable Multicast. Reliable Multicast Research Group. A Channel Access Scheme for Large Dense Packet Radio Networks. Routing in Communication Networks. --TR Epidemic algorithms for replicated database maintenance Multicast routing in datagram internetworks and extended LANs Architectural considerations for a new generation of protocols Highly dynamic Destination-Sequenced Distance-Vector routing (DSDV) for mobile computers Routing in the Internet information networks Routing in communications networks Exokernel A channel access scheme for large dense packet radio networks Epidemic algorithms in replicated databases (extended abstract) A performance comparison of multi-hop wireless ad hoc network routing protocols Scalable data naming for application level framing in reliable multicast Resource discovery in distributed networks A Highly Adaptive Distributed Routing Algorithm for Mobile Wireless Networks --CTR Kavitha Gundappachikkenahalli , Hesham H. Ali, ADPROC: an adaptive routing framework to provide QOS in wireless sensor networks, Proceedings of the 24th IASTED international conference on Parallel and distributed computing and networks, p.76-83, February 14-16, 2006, Innsbruck, Austria Mahesh (Umamaheswaran) Arumugam, Infuse: a TDMA based reprogramming service for sensor networks, Proceedings of the 2nd international conference on Embedded networked sensor systems, November 03-05, 2004, Baltimore, MD, USA Jonathan W. Hui , David Culler, The dynamic behavior of a data dissemination protocol for network programming at scale, Proceedings of the 2nd international conference on Embedded networked sensor systems, November 03-05, 2004, Baltimore, MD, USA Ossama Younis , Sonia Fahmy, HEED: A Hybrid, Energy-Efficient, Distributed Clustering Approach for Ad Hoc Sensor Networks, IEEE Transactions on Mobile Computing, v.3 n.4, p.366-379, October 2004 Chih-fan Hsin , Mingyan Liu, A distributed monitoring mechanism for wireless sensor networks, Proceedings of the 3rd ACM workshop on Wireless security, p.57-66, September 28-28, 2002, Atlanta, GA, USA Wei Liu , Yanchao Zhang , Wenjing Lou , Yuguang Fang, A robust and energy-efficient data dissemination framework for wireless sensor networks, Wireless Networks, v.12 n.4, p.465-479, July 2006 Constandinos X. Mavromoustakis , Helen D. Karatza, Handling Delay Sensitive Contents Using Adaptive Traffic-Based Control Method for Minimizing Energy Consumption in Wireless Devices, Proceedings of the 38th annual Symposium on Simulation, p.295-302, April 04-06, 2005 Constandinos X. Mavromoustakis , Helen D. Karatza, Adaptive Energy Conservation Model using Dynamic Caching for Wireless Devices, Proceedings of the 37th annual symposium on Simulation, p.257, April 18-22, 2004 Foad Lotfifar , Hadi Shahriar Shahhoseini, A mesh-based routing protocol for wireless ad- sensor networks, Proceeding of the 2006 international conference on Communications and mobile computing, July 03-06, 2006, Vancouver, British Columbia, Canada Christopher M. Sadler , Margaret Martonosi, Dali: a communication-centric data abstraction layer for energy-constrained devices in mobile sensor networks, Proceedings of the 5th international conference on Mobile systems, applications and services, June 11-13, 2007, San Juan, Puerto Rico Wang Lei , Chen Zhi-ping , Jiang Xin-hua, Researches on scheme of pairwise key establishment for distributed sensor networks, Proceedings of the 1st ACM workshop on Wireless multimedia networking and performance modeling, October 13-13, 2005, Montreal, Quebec, Canada Ramana Rao Kompella , Alex C. Snoeren, Practical lazy scheduling in sensor networks, Proceedings of the 1st international conference on Embedded networked sensor systems, November 05-07, 2003, Los Angeles, California, USA Christopher M. Sadler , Margaret Martonosi, Dali: a communication-centric data abstraction layer for energy-constrained devices in mobile sensor networks, Proceedings of the 5th international conference on Mobile systems, applications and services, June 11-13, 2007, San Juan, Puerto Rico M. H. Ali , Walid G. Aref , Cristina Nita-Rotaru, SPASS: scalable and energy-efficient data acquisition in sensor databases, Proceedings of the 4th ACM international workshop on Data engineering for wireless and mobile access, June 12-12, 2005, Baltimore, MD, USA Flvia Coimbra Delicato , Paulo F. Pires , Luci Pirmez , Luiz Fernando Carmo, A Service Approach for Architecting Application Independent Wireless Sensor Networks, Cluster Computing, v.8 n.2-3, p.211-221, July 2005 Alexandra Meliou , David Chu , Joseph Hellerstein , Carlos Guestrin , Wei Hong, Data gathering tours in sensor networks, Proceedings of the fifth international conference on Information processing in sensor networks, April 19-21, 2006, Nashville, Tennessee, USA Jae-Hwan Chang , Leandros Tassiulas, Maximum lifetime routing in wireless sensor networks, IEEE/ACM Transactions on Networking (TON), v.12 n.4, p.609-619, August 2004 Alexandros G. Dimakis , Vinod Prabhakaran , Kannan Ramchandran, Ubiquitous access to distributed data in large-scale sensor networks through decentralized erasure codes, Proceedings of the 4th international symposium on Information processing in sensor networks, April 24-27, 2005, Los Angeles, California Elena Fasolo , Christian Prehofer , Michele Rossi , Qing Wei , Jrg Widmer , Andrea Zanella , Michele Zorzi, Challenges and new approaches for efficient data gathering and dissemination in pervasive wireless networks, Proceedings of the first international conference on Integrated internet ad hoc and sensor networks, May 30-31, 2006, Nice, France Harshavardhan Sabbineni , Krishnendu Chakrabarty, Location-Aided Flooding: An Energy-Efficient Data Dissemination Protocol for Wireless Sensor Networks, IEEE Transactions on Computers, v.54 n.1, p.36-46, January 2005 A. Wadaa , S. Olariu , L. Wilson , M. Eltoweissy , K. Jones, Training a wireless sensor network, Mobile Networks and Applications, v.10 n.1-2, p.151-168, February 2005 C. -F. Chiasserini , R. Gaeta , M. Garetto , M. Gribaudo , D. Manini , M. Sereno, Fluid models for large-scale wireless sensor networks, Performance Evaluation, v.64 n.7-8, p.715-736, August, 2007

negotiation-based protocols;energy-efficient protocols;meta-data;wireless sensor networks;information dissemination

507059

The Impulse Memory Controller.

AbstractImpulse is a memory system architecture that adds an optional level of address indirection at the memory controller. Applications can use this level of indirection to remap their data structures in memory. As a result, they can control how their data is accessed and cached, which can improve cache and bus utilization. The Impulse design does not require any modification to processor, cache, or bus designs since all the functionality resides at the memory controller. As a result, Impulse can be adopted in conventional systems without major system changes. We describe the design of the Impulse architecture and how an Impulse memory system can be used in a variety of ways to improve the performance of memory-bound applications. Impulse can be used to dynamically create superpages cheaply, to dynamically recolor physical pages, to perform strided fetches, and to perform gathers and scatters through indirection vectors. Our performance results demonstrate the effectiveness of these optimizations in a variety of scenarios. Using Impulse can speed up a range of applications from 20 percent to over a factor of 5. Alternatively, Impulse can be used by the OS for dynamic superpage creation; the best policy for creating superpages using Impulse outperforms previously known superpage creation policies.

Introduction Since 1987, microprocessor performance has improved at a rate of 55% per year; in contrast, latencies have improved by only 7% per year, and DRAM bandwidths by only 15-20% per year [17]. the result is that the relative performance impact of memory accesses continues to grow. In addition, as instruction issue rates increase, the demand for memory bandwidth grows at least proportionately, possibly even superlinearly [8, 19]. Many important applications (e.g., sparse database, signal processing, multimedia, and CAD applications) do not exhibit sufficient Contact information: Prof. John Carter, School of Computing, 50 S Central Campus Drive, Room 3190, University of Utah, SLC, UT 84112-9205. retrac@cs.utah.edu. Voice: 801-585-5474. Fax: 801-581-5843. locality of reference to make effective use of the on-chip cache hierarchy. For such applications, the growing processor/memory performance gap makes it more and more difficult to effectively exploit the tremendous processing power of modern microprocessors. In the Impulse project, we are attacking this problem by designing and building a memory controller that is more powerful than conventional ones. Impulse introduces an optional level of address translation at the memory controller. the key insight that this feature exploits is that "unused" physical addresses can be translated to "real" physical addresses at the memory controller. An unused physical address is a legitimate address that is not backed by DRAM. For example, in a conventional system with 4GB of physical address space and only 1GB of installed DRAM, 3GB of the physical address space remains unused. We call these unused addresses shadow addresses, and they constitute a shadow address space that the Impulse controller maps to physical memory. By giving applications control (mediated by the OS) over the use of shadow addresses, Impulse supports application-specific optimizations that restructure data. Using Impulse requires software modifications to applications (or compilers) and operating systems, but requires no hardware modifications to processors, caches, or buses. As a simple example of how Impulse's memory remapping can be used, consider a program that accesses the diagonal elements of a large, dense matrix A. the physical layout of part of the data structure A is shown on the right-hand side of Figure 1. On a conventional memory system, each time the processor accesses a new diagonal element (A[i][i]), it requests a full cache line of contiguous physical memory (typically 32-128 bytes of data on modern systems). the program accesses only a single word of each of these cache lines. Such an access is shown in the top half of Figure 1. Using Impulse, an application can configure the memory controller to export a dense, shadow- space alias that contains just the diagonal elements, and can have the OS map a new set of virtual addresses to this shadow memory. the application can then access the diagonal elements via the new virtual alias. Such an access is shown in the bottom half of Figure 1. Remapping the array diagonal to a dense alias yields several performance benefits. First, the program enjoys a higher cache hit rate because several diagonal elements are loaded into the caches at once. Second, the program consumes less bus bandwidth because non-diagonal elements are not sent over the bus. Third, the program makes more effective use of cache space because the diagonal elements now have contiguous shadow addresses. In general, Impulse's flexibility allows applications to customize addressing to fit their needs. Section 2 describes the Impulse architecture. It describes the organization of the memory controller itself, as well as the system call interface that applications use to control it. the operating system must mediate use of the memory controller to prevent applications from accessing each other's physical memory. Section 3 describes the types of optimizations that Impulse supports. Many of the optimizations that we describe are not new, but Impulse is the first system that provides hardware support for them all in general-purpose computer systems. the optimizations include transposing matrices in memory without copying, creating superpages without copying, and doing scatter/gather through an indirection vector. Section 4 presents the results of a simulation study of Impulse, and shows that these optimizations can benefit a wide range of applications. Various applications see speedups ranging from 20% to a factor of 5. OS policies for dynamic superpage creation using Impulse have around 20% better speedup than those from prior work. Section 5 describes related work. A great deal of work has been done in the compiler and operating systems communities on related optimizations. the contribution of Impulse is that it provides hardware support for many optimizations that previously had to be performed purely in software. As a result, the tradeoffs for performing these optimizations are different. Section 6 summarizes our conclusions, and describes future work. Impulse Architecture Impulse expands the traditional virtual memory hierarchy by adding address translation hardware to the memory controller. This optional extra level of remapping is enabled by the fact that not all physical addresses in a traditional virtual memory system typically map to valid memory locations. the unused physical addresses constitute a shadow address space. the technology trend is putting more and more bits into physical addresses. For example, more and more 64-bit systems are coming out. One result of this trend is that the shadow space is getting larger and larger. Impulse allows software to configure the memory controller to interpret shadow addresses. Virtualizing unused physical addresses in this way can improve the efficiency of on-chip caches and TLBs, since hot data can be dynamically segregated from cold data. Data items whose physical DRAM addresses are not contiguous can be mapped to contiguous shadow addresses. In response to a cache line fetch of a shadow address, the memory controller fetches and compacts sparse data into dense cache lines before returning the data to the proces- sor. To determine where the data associated with these compacted shadow cache lines reside in physical memory, Impulse first recovers their offsets within the original data structure, which we call pseudo-virtual addresses. It then translates these pseudo-virtual addresses to physical DRAM addresses. the pseudo-virtual address space page layout mirrors the virtual address space, allowing Impulse to remap data structures that lie across non-contiguous physical pages. the shadow!pseudo-virtual!physical mappings all take place within the memory controller. the operating system manages all the resources in the expanded memory hierarchy and provides an interface for the application to specify optimizations for particular data structures. 2.1 Software Interface and OS Support To exploit Impulse, appropriate system calls must be inserted into the application code to configure the memory controller. the Architecture and Language Implementation group at the University of Massachusetts is developing compiler technology for Impulse. In response to an Impulse system call, the OS allocates a range of contiguous virtual addresses large enough to map the elements of the new (synthetic) data structure. the OS then maps the new data structure through shadow memory to the corresponding physical data elements. It does so by allocating a contiguous range of shadow addresses and downloading two pieces of information to the MMC: (i) a function that the MMC should use to perform the mapping from shadow to pseudo-virtual space and (ii) a set of page table entries that can be used to translate pseudo-virtual to physical DRAM addresses. As an example, consider remapping the diagonal of an n n matrix A[]. Figure 2 depicts the memory translations for both the matrix A[] and the remapped image of its diagonal. Upon seeing an access to a shadow address in the synthetic diagonal data structure, the memory controller gathers the corresponding diagonal elements from the original array, packs them into a dense cache line, and returns this cache line to the processor. the OS interface allows alignment and offset characteristics of the remapped data structure to be specified, which gives the application some control over L1 cache behavior. In the current Impulse design, coherence is maintained in software: the OS or the application programmer must keep aliased data consistent by explicitly flushing the cache. 2.2 Hardware Organization the organization of the Impulse controller architecture is depicted in Figure 3. the critical component of the Impulse MMC is the shadow engine, which processes all shadow accesses. the shadow engine contains a small SRAM Assembly Buffer, which is used to scatter/gather cache lines in the shadow address space; some shadow descriptors to store remapping configuration information; an wasted bus bandwidth Conventional Memory System Impulse Memory System Impulse Controller Cache Physical pages Figure 1: Using Impulse to remap the diagonal of a dense matrix into a dense cache line. the black boxes represent data on the diagonal, whereas the gray boxes represent non-diagonal data. Diagonal virtual memory shadow memory pseudo-virtual memory physical memory Page table Conventional Memory System Impulse Memory System A Remapping function Figure 2: Accessing the (sparse) diagonal elements of an array via a dense diagonal variable in Impulse. CPU MMU system memory bus Shadow Address DRAM Interface Normal Address Shadow engine AddrCalc Assembly buffer Descriptors Impulse MMC Figure 3: Impulse memory controller organization. ALU unit (AddrCalc) to translate shadow addresses to pseudo-virtual addresses; and a Memory Controller Translation Lookaside Buffer (MTLB) to cache recently used translations from pseudo- virtual addresses to physical addresses. the shadow engine contains eight shadow descriptors, each of which is is capable of saving all configuration settings for one remapping. All shadow descriptors share the same ALU unit and the same MTLB. Since the extra level of address translation is optional, addresses appearing on the memory bus may be in the physical (backed by DRAM) or shadow memory spaces. Valid physical addresses pass untranslated to the DRAM interface. Shadow addresses must be converted to physical addresses before being presented to the DRAM. To do so, the shadow engine first determines which shadow descriptor to use and passes its contents to the AddrCalc unit. the output of the AddrCalc will be a series of offsets for the individual sparse elements that need to be fetched. These offsets are passed through the MTLB to compute the physical addresses that need to be fetched. To hide some of the latency of fetching remapped data, each shadow descriptor can be configured to prefetch the remapped cache line following the currently accessed one. Depending on how Impulse is used to access a particular data structure, the shadow address translations can take three forms: direct, strided, or scatter/gather. Direct mapping translates a shadow address directly to a physical DRAM address. This mapping can be used to recolor physical pages without copying or to construct superpages dynamically. Strided mapping creates dense cache lines from array elements that are not contiguous. the mapping function maps an address soffset in shadow space to pseudo-virtual address pvaddr stride soffset, where pvaddr is the starting address of the data structure's pseudo-virtual image. pvaddr is assigned by the OS upon configuration. Scatter/gather mapping uses an indirection vector vec to translate an address soffset in shadow space to pseudo-virtual address pvaddr 3 Impulse Optimizations Impulse remappings can be used to enable a wide variety of optimizations. We first describe how Impulse's ability to pack data into cache lines (either using stride or scatter/gather remapping) can be used. We examine two scientific application kernels-sparse matrix-vector multiply (SMVP) and dense matrix-matrix product (DMMP)-and three image processing algorithms-image filter- ing, image rotation, and ray tracing. We then show how Impulse's ability to remap pages can be used to automatically improve TLB behavior through dynamic superpage creation. Some of these results have been published in prior conference papers [9, 13, 44, 49]. 3.1 Sparse Matrix-Vector Product Sparse matrix-vector product (SMVP) is an irregular computational kernel that is critical to many large scientific algorithms. For example, most of the time in conjugate gradient [3] or in the simulations [33] is spent performing SMVP. To avoid wasting memory, sparse matrices are generally compacted so that only non-zero elements and corresponding index arrays are stored. For example, the Class A input matrix for the NAS conjugate gradient kernel (CG-A) is 14,000 by 14,000, and contains only 1.85 million non-zeroes. Although sparse encodings save tremendous amounts of memory, sparse matrix codes tend to suffer from poor memory performance because data must be accessed through indirection vectors. CG-A on an SGI Origin 2000 processor (which has a 2-way, 32K L1 cache and a 2-way, exhibits L1 and L2 cache hit rates of only 63% and 92%, respectively. the inner loop of the sparse matrix-vector product in CG is roughly: for to n do sum := 0 for j := ROWS[i] to ROWS[i+1]-1 do sum += the code and data structures for SMVP are illustrated in Figure 4. Each iteration multiplies a row of the sparse matrix A with the dense vector x. the accesses to x are indirect (via the COLUMN index vector) and sparse, making this code perform poorly on conventional memory systems. Whenever x is accessed, a conventional memory system fetches a cache line of data, of which only one element is used. the large sizes of x, COLUMN, and DATA and the sparse nature of accesses to x inhibit data reuse in the L1 cache. Each element of COLUMN or DATA is used only once, and almost every access to x results in an L1 cache miss. A large L2 cache can enable reuse of x, if physical data layouts can be managed to prevent L2 cache conflicts between A and x. Unfortunately, conventional systems do not typically provide mechanisms for managing physical layout. the Impulse memory controller supports scatter/gather of physical addresses through indirection vectors. Vector machines such as the CDC STAR-100 [18] provided scatter/gather capabilities in hardware within the processor. Impulse allows conventional CPUs to take advantage of scat- ter/gather functionality by implementing the operations at the memory, which reduces memory traffic over the bus. To exploit Impulse, CG's SMVP code can be modified as follows: impulse remap(x, x', N, COLUMN, INDIRECT, .) for to n do sum := 0 for j := ROWS[i] to ROWS[i+1]-1 do sum += DATA[j] * x'[j] the impulse remap operation asks the operating system to 1) allocate a new region of shadow space, 2) map x' to that shadow region, and 3) instruct the memory controller to map the elements of the shadow region x'[k] to the physical memory for x[COLUMN[k]]. After the remapped array has been set up, the code accesses the remapped version of the gathered structure rather than the original structure (x). This optimization improves the performance of SMVP in two ways. First, spatial locality is improved in the L1 cache. Since the memory controller packs the gathered elements into cache lines, each cache line contain 100% useful data, rather than only one useful element. Second, sum := 0 for j := ROWS[i] to ROWS[i+1]-1 do sum += DATA[j] * x[COLUMN[j]]; for to n do A F2 C D Y U COLUMN ROWS Figure 4: Conjugate gradient's sparse matrix-vector product. the matrix A is encoded using three dense arrays: DATA, ROWS, and COLUMN. the contents of A are in DATA. ROWS[i] indicates where the i th row begins in DATA. COLUMN[i] indicates which column of A the element stored in DATA[i] comes from. the processor issues fewer memory instructions, since the read of the indirection vector COLUMN occurs at the memory controller. Note that the use of scatter/gather at the memory controller reduces temporal locality in the L2 cache. the remapped elements of x' cannot be reused, since all of the elements have different addresses. An alternative to scatter/gather is dynamic physical page recoloring through direct remapping of physical pages. Physical page recoloring changes the physical addresses of pages so that reusable data is mapped to a different part of a physically-addressed cache than non-reused data. By performing page recoloring, conflict misses can be eliminated. On a conventional machine, physical page recoloring is expensive: the only way to change the physical address of data is to copy the data between physical pages. Impulse allows physical pages to be recolored without copying. Virtual page recoloring has been explored by other authors [6]. For SMVP, the x vector is reused within an iteration, while elements of the DATA, ROW, and COLUMN vectors are used only once in each iteration. As an alternative to scatter/gather of x at the memory controller, Impulse can be used to physically recolor pages so that x does not conflict with the other data structures in the L2 cache. For example, in the CG-A benchmark, x is over 100K bytes: it would not fit in most L1 caches, but would fit in many L2 caches. Impulse can remap x to pages that occupy most of the physically-indexed L2 cache, and can remap DATA, ROWS, and COLUMNS to a small number of pages that do not conflict with x. In our experiments, we color the vectors x, DATA, and COLUMN so that they do not conflict in the L2 cache. the multiplicand vector x is heavily reused, so we color it to occupy the first half of the L2 cache. To keep the large DATA and COLUMN structures from conflicting, we divide the second half of the L2 cache into two quarters, and then color DATA and COLUMN so they each occupy one quarter of the cache. In effect, we use pieces of the L2 cache as a set of virtual stream buffers [29] for DATA, ROWS, and COLUMNS. 3.2 Tiled Matrix Algorithms Dense matrix algorithms form an important class of scientific kernels. For example, LU decomposition and dense Cholesky factorization are dense matrix computational kernels. Such algorithms are tiled (or blocked) to increase their efficiency. That is, the iterations of tiled algorithms are reordered to improve their memory performance. the difficulty with using tiled algorithms lies in choosing an appropriate tile size [27]. Because tiles are non-contiguous in the virtual address space, it is difficult to keep them from conflicting with each other or with themselves in cache. To avoid conflicts, either tile sizes must be kept small, which makes inefficient use of the cache, or tiles must be copied into non-conflicting regions of memory, which is expensive. Impulse provides an alternative method of removing cache conflicts for tiles. We use the simplest tiled algorithm, dense matrix-matrix product (DMMP), as an example of how Impulse can improve the behavior of tiled matrix algorithms. Assume that we are computing want to keep the current tile of the C matrix in the L1 cache as we compute it. In addition, since the same row of the A matrix is used multiple times to compute a row of the C matrix, we would like to keep the active row of A in the L2 cache. Impulse allows base-stride remapping of the tiles from non-contiguous portions of memory into contiguous tiles of shadow space. As a result, Impulse makes it easy for the OS to virtually remap the tiles, since the physical footprint of a tile will match its size. If we use the OS to remap the virtual address of a matrix tile to its new shadow alias, we can then eliminate interference in a virtually-indexed L1 cache. First, we divide the L1 cache into three segments. In each segment we keep a tile: the current output tile from C, and the input tiles from A and B. When we finish with one tile, we use Impulse to remap the virtual tile to the next physical tile. To maintain cache consistency, we must purge the A and B tiles and flush the C tiles from the caches whenever they are remapped. As Section 4.1.2 shows, these costs are minor. 3.3 Image Filtering Image filtering applies a numerical filter function to an image to modify its appearance. Image filtering may be used to attenuate high-frequency components caused by noise in a sampled image, to adjust an image to different geometry, to detect or enhance edges within an image, or to create various special effects. Box, Bartlett, Gaussian, and binomial filters are common in practice. Each modifies the input image in a different way, but all share similar computational characteristics. We concentrate on a representative class of filters, binomial filters [15], in which each pixel in the output image is computed by applying a two-dimensional "mask" to the input image. Binomial filtering is computationally similar to a single step of a successive over-relaxation algorithm for solving differential equations: the filtered pixel value is calculated as a linear function of the neighboring pixel values of the original image and the corresponding mask values. For example, for an order-5 binomial filter, the value of pixel (i; j) in the output image will be256 To avoid edge effects, the original image boundaries must be extended before applying the masking function. Figure 5 illustrates a black- and-white sample image before and after the application of a small binomial filter. In practice, many filter functions, including binomial, are "separable," meaning that they are symmetric and can be decomposed into a pair of orthogonal linear filters. For example, a two-dimensional mask can be decomposed into two, one-dimensional, linear masks ([ - the two-dimensional mask is simply the outer product of this one-dimensional mask with its transpose. the process of applying the mask to the input image can be performed by sweeping first along the rows and then the columns, calculating a partial sum at each step. Each pixel in the original image is used only for a short time, which makes filtering a pure streaming application. Impulse can transpose both the input and output image arrays without copying, which gives the column sweep much better cache behavior. 3.4 Image Rotation Image warping refers to any algorithm that performs an image-to-image transformation. Separable image warps are those that can be decomposed into multiple one-dimensional transformations [10]. For separable warps, Impulse can be used to improve the cache and TLB performance of one-dimensional traversals orthogonal to the image layout in memory. the three-shear image rotation algorithm is an example of a separable image warp. This algorithm rotates a 2-dimensional image around its center in three stages, each of which performs a "shear" operation on the image, as Figure 5: Example of binomial image filtering. the original image is on the left, and the filtered image is on the right. illustrated in Figure 6. the algorithm is simpler to write, faster to run, and has fewer visual artifacts than a direct rotation. the underlying math is straightforward. Rotation through an angle can be expressed as matrix multiplication:@ y 0A =@ cos sin sin cos A@ x the rotation matrix can be broken into three shears as follows:@ cos sin sin cos A =@ 1 0 tan 1A@ 1 sin tan 1A None of the shears requires scaling (since the determinant of each matrix is 1), so each involves just a shift of rows or columns. Not only is this algorithm simple to understand and implement, it is robust in that it is defined over all rotation values from 0 - to 90 - . Two-shear rotations fail for angles near 90 - . We assume a simple image representation of an array of pixel values. the second shear operation (along the y axis) walks along the column of the image matrix, which gives rise to poor memory performance for large images. Impulse improves both cache and TLB performance by transposing the matrix without copying, so that walking along columns in the image is replaced by walking along rows in a transposed matrix. Isosurface Rendering Using Ray Tracing Our isosurface rendering benchmark is based on the technique demonstrated by Parker et al. [37]. This benchmark generates an image of an isosurface in a volume from a specific point of view. In Figure Three-shear rotation of an image counter-clockwise through one radian. the original image (upper left) is first sheared horizontally (upper right). That image is sheared upwards (lower right). the final rotated image (lower left) is generated via one final horizontal shift. contrast to other volume visualization methods, this method does not generate an explicit representation of the isosurface and render it with a z-buffer, but instead uses brute-force ray tracing to perform interactive isosurfacing. For each ray, the first isosurface intersected determines the value of the corresponding pixel. the approach has a high intrinsic computational cost, but its simplicity and scalability make it ideal for large data sets on current high-end systems. Traditionally, ray tracing has not been used for volume visualization because it suffers from poor memory behavior when rays do not travel along the direction that data is stored. Each ray must be traced through a potentially large fraction of the volume, giving rise to two problems. First, many memory pages may need to be touched, which results in high TLB pressure. Second, a ray with a high angle of incidence may visit only one volume element (voxel) per cache line, in which case bus bandwidth will be wasted loading unnecessary data that pollutes the cache. By carefully hand-optimizing their ray tracer's memory access patterns, Parker et al. achieve acceptable performance for interactive rendering (about 10 frames per second). They improve data locality by organizing the data set into a multi-level spatial hierarchy of tiles, each composed of smaller cells. the smaller cells provide good cache-line utilization. "Macro cells" are created to cache the minimum and maximum data values from the cells of each tile. These macro cells enable a simple min/max comparison to detect whether a ray intersects an isosurface within the tile. Empty macro cells need not be traversed. Careful hand-tiling of the volume data set can yield much better memory performance, but choosing the optimal number of levels in the spatial hierarchy and sizes for the tiles at each level is difficult, and the resulting code is hard to understand and maintain. Impulse can deliver better performance than hand-tiling at a lower programming cost. There is no need to preprocess the volume data set for good memory performance: the Impulse memory controller can remap it dynamically. screen volume Figure 7: Isosurface rendering using ray tracing. the picture on the left shows rays perpendicular to the viewing screen being traced through a volume. the one on the right illustrates how each ray visits a sequence of voxels in the volume; Impulse optimizes voxel fetches from memory via indirection vectors representing the voxel sequences for each ray. In addition, the source code retains its readability and modifiability. Like many real-world visualization systems, our benchmark uses an orthographic tracer whose rays all intersect the screen surface at right angles, producing images that lack perspective and appear far away, but are relatively simple to compute. We use Impulse to extract the voxels that a ray potentially intersects when traversing the vol- ume. the right-hand side of Figure 7 illustrates how each ray visits a certain sequence of voxels in the volume. Instead of fetching cache lines full of unnecessary voxels, Impulse can remap a ray to the voxels it requires so that only useful voxels will be fetched. 3.6 Online Superpage Promotion Impulse can be used to improve TLB performance automatically, by having the operating system automatically create superpages dynamically. Superpages are supported by the translation lookaside buffers (TLBs) on almost all modern processors; they are groups of contiguous virtual memory pages that can be mapped with a single TLB entry [12, 30, 43]. Using superpages makes more efficient use of a TLB, but the physical pages that back a superpage must be contiguous and properly aligned. Dynamically coalescing smaller pages into a superpage thus requires that all the pages be be coincidentally adjacent and aligned (which is unlikely), or that they be copied so that they become so. the overhead of promoting superpages by copying includes both direct and indirect costs. the direct costs come from copying the pages and changing the mappings. Indi- 0Physical Addresses 0x06155000 0x40138000 0x20285000 0x80243000 0x00005000 0x00004000 Virtual Addresses Shadow Addresses 0x00007000 00004 004physical size Processor TLB virtual Memory controller Figure 8: An Example of Creating Superpages Using Shadow Space rect costs include the increased number of instructions executed on each TLB miss (due to the new decision-making code in the miss handler) and the increased contention in the cache hierarchy (due to the code and data used in the promotion process). When deciding whether to create superpages, all costs must be balanced against the improvements in TLB performance. Romer et al. [40] study several different policies for dynamically creating superpages. Their trace-driven simulations and analysis show how a policy that balances potential performance benefits and promotion overheads can improve performance in some TLB-bound applications by about 50%. Our work extends that of Romer et al. by showing how Impulse changes the design of a dynamic superpage promotion policy. the Impulse memory controller maintains its own page tables for shadow memory mappings. Building superpages from base pages that are not physically contiguous entails simply remapping the virtual pages to properly aligned shadow pages. the memory controller then maps the shadow pages to the original physical pages. the processor's TLB is not affected by the extra level of translation that takes place at the controller. Figure 8 illustrates how superpage mapping works on Impulse. In this example, the OS has mapped a contiguous 16KB virtual address range to a single shadow superpage at "physical" page frame 0x80240. When an address in the shadow physical range is placed on the system memory bus, the memory controller detects that this "physical" address needs to be retranslated using its local shadow-to-physical translation tables. In the example in Figure 8, the processor translates an access to virtual address 0x00004080 to shadow physical address 0x80240080, which the controller, in turn, translates to real physical address 0x40138080. Performance We performed a series of detailed simulations to evaluate the performance impact of the optimizations described in Section 3. Our studies use the URSIM [48] execution-driven simulator, which is derived from RSIM [35]. URSIM models a microarchitecture close to MIPS R10000 micro-processor [30] with a 64-entry instruction window. We configured it to issue four instructions per cycle. We model a 64-kilobyte L1 data cache that is non-blocking, write-back, virtually in- dexed, physically tagged, direct-mapped, and has 32-byte lines. the 512-kilobyte L2 data cache is non-blocking, write-back, physically indexed, physically tagged, two-way associative, and has 128-byte lines. L1 cache hits take one cycle, and L2 cache hits take eight cycles. URSIM models a split-transaction MIPS R10000 cluster bus with a snoopy coherence pro- tocol. the bus multiplexes addresses and data, is eight bytes wide, has a three-cycle arbitration delay and a one-cycle turn-around time. We model two memory controllers: a conventional high-performance MMC based on the one in the SGI O200 server and the Impulse MMC. the system bus, memory controller, and DRAMs have the same clock rate, which is one third of the CPU clock's. the memory system supports critical word first, i.e., a stalled memory instruction resumes execution after the first quad-word returns. the load latency of the first quad-word is cycles. the unified TLB is single-cycle, fully associative, software-managed, and combined instruction and data. It employs a least-recently-used replacement policy. the base page size is 4096 bytes. Superpages are built in power-of-two multiples of the base page size, and the biggest superpage that the TLB can map contains 2048 base pages. We model a 128-entry TLB. In the remainder of this section we examine the simulated performance of Impulse on the examples given in Section 3. Our calculation of "L2 cache hit ratio" and "mem (memory) hit ratio" uses the total number of loads executed (not the total number of L2 cache accesses) as the divisor for both ratios. This formulation makes it easier to compare the effects of the L1 and L2 caches on memory accesses: the sum of the L1 cache, L2 cache, and memory hit ratios equals 100%. 4.1 Fine-Grained Remapping the first set of experiments exploit Impulse's fine-grained remapping capabilities to create synthetic data structures with better locality than in the original programs. 4.1.1 Sparse Matrix-Vector Product Table 4.1.1 shows how Impulse can be used to improve the performance of the NAS Class A Conjugate Gradient (CG-A) benchmark. the first column gives results from running CG-A on a non-Impulse system. the second and third columns give results from running CG-A on an Impulse system. the second column numbers come from using the Impulse memory controller to perform scatter/gather; the third column numbers come from using it to perform physical page coloring. On the conventional memory system, CG-A suffers many cache misses: nearly 17% of accesses go to the memory. the inner loop of CG-A is very small, so it can generate cache misses quickly, which leads to there being a large number of cache misses outstanding at any given time. the large number of outstanding memory operations causes heavy contention for the system bus, memory controller, and DRAMs; for the baseline version of CG-A, bus utilization reaches 88.5%. As a result, the average latency of a memory operation reaches 163 cycles for the baseline version of CG-A. This behavior combined with the high cache miss rates causes the average load in CG-A to take 47.6 cycles compared to only 1 cycle for L1 cache hits. Scatter/gather remapping on CG-A improves performance by over a factor of 3, largely due to the increase in the L1 cache hit ratio and the decrease in the number of loads/stores that go to memory. Each main memory access for the remapped vector x' loads the cache with several useful elements from the original vector x, which increases the L1 cache hit rate. In other words, retrieving elements from the remapped array x' improves the spatial locality of CG-A. Scatter/gather remapping reduces the total number of loads executed by the program from 493 million from 353 million. In the original program, two loads are issued to compute x[COLUMN[j]]. In the scatter/gather version of the program, only one load is issued by the processor, because the load of the indirection vector occurs at the memory controller. This reduction more than compensates for the scatter/gather's increase in the average cost of a load, and accounts for almost one-third of the cycles saved. To provide another example of how useful Impulse can be, we use it to recolor the pages of the major data structures in CG-A. Page recoloring consistently reduces the cost of memory accesses by eliminating conflict misses in the L2 cache and increasing the L2 cache hit ratio from 19.7% to 22.0%. As a result, fewer loads go to memory, and performance is improved by 17%. Although page recoloring improves performance on CG-A, it is not nearly as effective as scat- ter/gather. the difference is primarily because page recoloring does not achieve the two major improvements that scatter/gather provides: improving the locality of CG-A and reducing the number of loads executed. This comparison does not mean that page recoloring is not a useful optimiza- tion. Although the speedup for page recoloring on CG-A is substantially less than scatter/gather, page recoloring is more broadly applicable. 4.1.2 Dense Matrix-Matrix Product This section examines the performance benefits of tile remapping for DMMP, and compares the results to software tile copying. Impulse's alignment restrictions require that remapped tiles be aligned to L2 cache line boundaries, which adds the following constraints to our matrices: Tile sizes must be a multiple of a cache line. In our experiments, this size is 128 bytes. This constraint is not overly limiting, especially since it makes the most efficient use of cache space. Arrays must be padded so that tiles are aligned to 128 bytes. Compilers can easily support this constraint: similar padding techniques have been explored in the context of vector processors [7]. Table 4.1.2 illustrates the results of our tiling experiments. the baseline is the conventional no-copy tiling. Software tile copying outperforms the baseline code by almost 10%; Impulse tile remapping outperforms it by more than 20%. the improvement in performance for both is primarily due to the difference in cache behavior. Both copying and remapping more than double the L1 cache hit rate, and they reduce the average number of cycles for a load to less than two. Impulse has a higher L1 cache hit ratio than software copying, since copying tiles can incur cache misses: the number of loads that go to memory is reduced by two-thirds. In addition, the cost of copying the tiles is greater than the overhead of using Impulse to remap tiles. As a result, using Impulse provides twice as much speedup. This comparison between conventional and Impulse copying schemes is conservative for several reasons. Copying works particularly well on DMMP: the number of operations performed on a tile of size O(n 2 ) is O(n 3 ), which means the overhead of copying is relatively low. For algorithms where the reuse of the data is lower, the relative overhead of copying is greater. Likewise, as caches (and therefore tiles) grow larger, the cost of copying grows, whereas the (low) cost of Impulse's tile remapping remains fixed. Finally, some authors have found that the performance of copying can vary greatly with matrix size, tile size, and cache size [45], but Impulse should be insensitive to cross-interference between tiles. Conventional Scatter/Gather Page Coloring Time 5.48B 1.77B 4.67B L2 hit ratio 19.7% 15.9% 22.0% mem hit ratio 16.9% 6.3% 14.3% avg load time 47.6 23.2 38.7 loads 493M 353M 493M speedup - 3.10 1.17 Table 1: Simulated results for the NAS Class A conjugate gradient benchmark, using two different opti- mizations. Times are in processor cycles. Conventional Software copying Impulse Time 664M 610M 547M L1 hit ratio 49.6% 98.6% 99.5% L2 hit ratio 48.7% 1.1% 0.4% mem hit ratio 1.7% 0.3% 0.1% avg load time 6.68 1.71 1.46 Table 2: Simulated results for tiled matrix-matrix product. Times are in millions of cycles. the matrices are 512 by 512, with 32 by tiles. Tiled Impulse Time 459M 237M L1 hit ratio 98.95% 99.7% mem hit ratio 0.24% 0.05% avg load time 1.57 1.16 issued instructions (total) 725M 290M graduated instructions (total) 435M 280M issued instructions (TLB) 256M 7.8M graduated instructions (TLB) 134M 3.3M Table 3: Simulated results for image filtering with various memory system configurations. Times are in processor cycles. TLB misses are the number of user data misses. 4.1.3 Image Filtering Table 3 presents the results of order-129 binomial filter on a 321024 color image. the Impulse version of the code pads each pixel to four bytes. Performance differences between the hand- tiled and Impulse versions of the algorithm arise from the vertical pass over the data. the tiled version suffers more than 3.5 times as many L1 cache misses and 40 times as many TLB faults, and executed 134 million instructions in TLB miss handlers. the indirect impact of the high TLB miss rate is even more dramatic - in the baseline filtering program, almost 300 million instructions are issued but not graduated. In contrast, the Impulse version of the algorithm executes only 3.3 million instructions handling TLB misses and only 10 million instructions are issued but not graduated. Compared to these dramatic performance improvements, the less than 1 million cycles spent setting up Impulse remapping are a negligible overhead. Although both versions of the algorithm touch each data element the same number of times, Impulse improves the memory behavior of the image filtering code in two ways. When the original algorithm performs the vertical filtering pass, it touches more pages per iteration than the processor TLB can hold, yielding the high kernel overhead observed in these runs. Image cache lines conflicting within the L1 cache further degrade performance. Since the Impulse version of the code (what appear to the processor to be) contiguous addresses, it suffers very few TLB faults and has near-perfect temporal and spatial locality in the L1 cache. 4.1.4 Three-Shear Image Rotation Table 4 illustrates performance results for rotating a color image clockwise through one radian. the image contains 24 bits of color information, as in a ".ppm" file. We measure three versions of this benchmark: the original version, adapted from Wolberg [47]; a hand-tiled version of the code, in which the vertical shear's traversal is blocked; and a version adapted to Impulse, in which the matrices are transposed at the memory controller. the Impulse version requires that each pixel be padded to four bytes, since Impulse operates on power-of-two object sizes. To quantify the performance effect of padding, we measure the results for the non-Impulse versions of the code using both three-byte and four-byte pixels. the performance differences among the different versions are entirely due to cycles saved during the vertical shear. the horizontal shears exhibit good memory behavior (in row-major layout), and so are not a performance bottleneck. Impulse increases the cache hit rate from roughly 95% to 98.5% and reduces the number of TLB misses by two orders of magnitude. This reduction Original Original Tiled Tiled Impulse padded padded Time 572M 576M 284M 278M 215M L1 hit ratio 95.0% 94.8% 98.1% 97.6% 98.5% L2 hit ratio 1.5% 1.6% 1.1% 1.5% 1.1% mem hit ratio 3.5% 3.6% 0.8% 0.9% 0.4% avg load time 3.85 3.85 1.81 2.19 1.50 issued instructions (total) 476M 477M 300M 294M 232M graduated instructions (total) 346M 346M 262M 262M 229M issued instructions (TLB) 212M 215M 52M 51M 0.81M graduated instructions (TLB) 103M 104M 24M 24M 0.42M Table 4: Simulation results for performing a 3-shear rotation of a 1k-by-1k 24-bit color image. Times are in processor cycles. TLB misses are user data misses. in the TLB miss rate eliminates 99 million TLB miss handling instructions and reduces the number of issued but not graduated instructions by over 100 million. These two effects constitute most of Impulse's benefit. the tiled version walks through all columns 32 pixels at a time, which yields a hit rate higher than the original program's, but lower than Impulse's. the tiles in the source matrix are sheared in the destination matrix, so even though cache performance for the source is nearly perfect, it suffers for the destination. For the same reason, the decrease in TLB misses for the tiled code is not as great as that for the Impulse code. the Impulse code requires 33% more memory to store a 24-bit color image. We also measured the performance impact of using padded 32-bit pixels with each of the non-Impulse codes. In the original program, padding causes each cache line fetch to load useless pad bytes, which degrades the performance of a program that is already memory-bound. In contrast, for the tiled program, the increase in memory traffic is balanced by the reduction in load, shift, and mask operations: manipulating word-aligned pixels is faster than manipulating byte-aligned pixels. the padded, tiled version of the rotation code is still slower than Impulse. the tiled version of the shear uses more cycles recomputing (or saving and restoring) each column's displacement when traversing the tiles. For our input image, this displacement is computed 1024= 32 times, since the column length is 1024 and the tile height is 32. In contrast, the Impulse code (which is not tiled) only computes each column's displacement once, since each column is completely traversed when it is visited. Isosurface Rendering Using Ray Tracing For simplicity, our benchmark assumes that the screen plane is parallel to the volume's z axis. As a result, we can compute an entire plane's worth of indirection vector at once, and we do not need to remap addresses for every ray. This assumption is not a large restriction: it assumes the use of a volume rendering algorithm like Lacroute's [26], which transforms arbitrary viewing angles into angles that have better memory performance. the isosurface in the volume is on one edge of the surface, parallel to the x-z plane. the measurements we present are for two particular viewing angles. Table 5(A) shows results when the screen is parallel to the y-z plane, so that the rays exactly follow the layout of voxels in memory (we assume an x-y-z layout order). Table 5(B) shows results when the screen is parallel to the x-z plane, where the rays exhibit the worst possible cache and TLB behavior when traversing the x-y planes. These two sets of data points represent the extremes in memory performance for the ray tracer. In our data, the measurements labeled "Original" are of a ray tracer that uses macro-cells to reduce the number of voxels traversed, but that does not tile the volume. the macro-cells are 444 voxels in size. the results labeled "Indirection" use macro-cells and address voxels through an indirection vector. the indirection vector stores precomputed voxel offsets of each x-y plane. Finally, the results labeled "Impulse" use Impulse to perform the indirection lookup at the memory controller. In Table 5(A), where the rays are parallel to the array layout, Impulse delivers a substantial performance gain. Precomputing the voxel offsets reduces execution time by approximately 9 million cycles. the experiment reported in the Indirection column exchanges the computation of voxels offsets for the accesses to the indirection vector. Although it increases the number of memory loads, it still achieves positive speedup because most of those accesses are cache hits. With Impulse, the accesses to the indirection vector are performed only within the memory controller, which hides the access latencies. Consequently, Impulse obtained a higher speedup. Compared to the original version, Impulse saves the computation of voxels offsets. In Table 5(B), where the rays are perpendicular to the voxel array layout, Impulse yields a much larger performance gain - a speedup of 5.49. Reducing the number of TLB misses saves approximately million graduated instructions while reducing the number of issued but not graduated instructions by approximately 120 million. Increasing the cache hit ratio by loading no useless voxels into the cache saves the remaining quarter-billion cycles. the Indirection version executes about 3% slower than the original one. With rays perpendicular to the voxel array, accessing voxels Original Indirection Impulse Time 74.2M 65.0M 61.4M L1 hit ratio 95.1% 90.8% 91.8% L2 hit ratio 3.7% 7.3% 6.3% mem hit ratio 1.2% 1.9% 1.9% avg load time 1.8 2.8 2.5 loads 21.6M 17.2M 13M issued instructions (total) 131M 71.4M 57.7M graduated instructions (total) 128M 69.3M 55.5M issued instructions (TLB) 0.68M 1.14M 0.18M graduated instructions (TLB) 0.35M 0.50M 0.15M speedup - 1.14 1.21 Original Indirection Impulse Time 383M 397M 69.7M L2 hit ratio 0.6% 2.2% 5.1% mem hit ratio 12.3% 15.2% 1.6% avg load time 8.2 10.3 2.4 loads 32M 27M 16M issued instructions (total) 348M 318M 76M graduated instructions (total) 218M 148M 68M issued instructions (TLB) 126M 156M 0.18M graduated instructions (TLB) 59M 60M 0.15M Table 5: Results for isosurface rendering. Times are in processor cycles. TLB misses are user data misses. In (A), the rays follow the memory layout of the image; in (B), they are perpendicular to the memory layout. generates lots of cache misses and frequently loads new data into the cache. These loads can evict the indirection vector from the cache and bring down the cache hit ratio of the indirection vector accesses. As a result, the overhead of accessing the indirection vector outweighs the benefit of saving the computation of voxel offsets and slows down execution. 4.2 Online Superpage Promotion To evaluate the performance of Impulse's support for inexpensive superpage promotion, we reevaluated Romer et al.'s work on dynamic superpage promotion algorithms [40] in the context of Im- pulse. Our system model differs from theirs in several significant ways. They employ a form of trace-driven simulation with ATOM [42], a binary rewriting tool. That is, they rewrite their applications using ATOM to monitor memory references, and the modified applications are used to do on-the-fly "simulation" of TLB behavior. Their simulated system has two 32-entry, fully-associative TLBs (one for instructions and one for data), uses LRU replacement on TLB entries, and has a base page size of 4096 bytes. To better understand how TLB size may affect the perfor- mance, we model two TLB sizes: 64 and 128 entries. Romer et al. combine the results of their trace-driven simulation with measured baseline performance results to calculate effective speedup on their benchmarks. They execute their benchmarks on a DEC Alpha 3000/700 running DEC OSF/1 2.1. the processor in that system is a dual-issue, in-order, 225 MHz Alpha 21064. the system has two megabytes of off-chip cache and 160 megabytes of main memory. For their simulations, they assume the following fixed costs, which do not take cache effects into account: each 1Kbyte copied is assigned a 3000-cycle cost; the asap policy is charged cycles for each TLB miss; and the approx-online policy is charged 130 cycles for each TLB miss. the performance results presented here are obtained through complete simulation of the bench- marks. We measure both kernel and application time, the direct overhead of implementing the superpage promotion algorithms and the resulting effects on the system, including the expanded TLB miss handlers, cache effects due to accessing the page tables and maintaining prefetch coun- ters, and the overhead associated with promoting and using superpages with Impulse. We present comparative performance results for our application benchmark suite. 4.2.1 Application Results To evaluate the different superpage promotion approaches on larger problems, we use eight programs from a mix of sources. Our benchmark suite includes three SPEC95 benchmarks (com- press, gcc, and vortex), the three image processing benchmarks described earlier (iso- rotate, and filter), one scientific benchmark (adi), and one benchmark from the DIS benchmark suite (dm) [28]. All benchmarks were compiled with Sun cc Workshop Compiler 4.2 and optimization level "-xO4". Compress is the SPEC95 data compression program run on an input of ten million characters. To avoid overestimating the efficacy of superpages, the compression algorithm was run only once, instead of the default 25 times. gcc is the cc1 pass of the version 2.5.3 gcc compiler (for SPARC architectures) used to compile the 306-kilobyte file "1cp-dec1.c". vortex is an object-oriented database program measured with the SPEC95 "test" input. isosurf is the interactive isosurfacing volume renderer described in Section 4.1.5. filter performs an order-129 binomial filter on a 321024 color image. rotate turns a 10241024 color image clockwise through one radian. adi implements algorithm alternative direction integration. dm is a data management program using input file "dm07.in". Two of these benchmarks, gcc and compress, are also included in Romer et al.'s benchmark suite, although we use SPEC95 versions, whereas they used SPEC92 versions. We do not use the other SPEC92 applications from that study, due to the benchmarks' obsolescence. Several of Romer et al.'s remaining benchmarks were based on tools used in the research environment at the University of Washington, and were not readily available to us. Table 6 lists the characteristics of the baseline run of each benchmark with a four-way issue superscalar processor, where no superpage promotion occurs. TLB miss time is the total time spent in the data TLB miss handler. These benchmarks demonstrate varying sensitivity to TLB performance: on the system with the smaller TLB, between 9.2% and 35.1% of their execution time is lost due to TLB miss costs. the percentage of time spent handling TLB misses falls to between less than 1% and 33.4% on the system with a 128-entry TLB. Figures show the normalized speedups of the different combinations of promotion policies (asap and approx-online) and mechanisms (remapping and copying) compared to the baseline instance of each benchmark. In our experiments we found that the best approx-online threshold for a two-page superpage is 16 on a conventional system and is 4 on an Impulse system. These are also the thresholds used in our full-application tests. Figure 9 gives results with a 64- entry results with a 128-entry TLB. Online superpage promotion can improve Total Cache TLB TLB Benchmark cycles misses misses miss 64-entry TLB compress 632 3455 4845 27.9% gcc 628 1555 2103 10.3% vortex filter 425 241 4798 35.1% rotate 547 3570 3807 17.9% dm 233 129 771 9.2% 128-entry TLB compress 426 3619 36 0.6% gcc 533 1526 332 2.0% vortex 423 763 1047 8.1% isosurf 93 989 548 17.4% filter 417 240 4544 33.4% rotate 545 3569 3702 16.9% dm 211 143 250 3.3% Table Characteristics of each baseline run speedup Impulse+asap Impulse+approx_online copying+asap copying+approx_online compress gcc vortex 1.07 1.060.93 1.06 0.85 filter 1.76 1.74 1.70 1.68 rotate dm 1.13 1.09 1.11 1.05 Figure 9: Normalized speedups for each of two promotion policies on a 4-issue system with a 64-entry TLB. performance by up to a factor of two (on adi with remapping asap), but it also can decrease performance by a similar factor (when using the copying version of asap on isosurf). We can make two orthogonal comparisons from these figures: remapping versus copying, and asap versus approx-online. 4.2.2 Asap vs. Approx-online We first compare the two promotion algorithms, asap and approx-online, using the results from Figures 9 and 10. the relative performance of the two algorithms is strongly influenced by the choice of promotion mechanism, remapping or copying. Using remapping , asap slightly out-performs approx-online in the average case. It exceeds the performance of approx-online in 14 of the 16 experiments, and trails the performance of approx-online in only one case (on vortex with a 64-entry tlb). the differences in performance range from asap+remap outperforming aol+remap by 32% for adi with a 64-entry TLB, to aol+remap outperforming asap+remap by 6% for vortex with a 64-entry TLB. In general, however, performance differences between the two policies are small: asap is on average 7% better with a 64-entry TLB, and 6% better with a 128-entry TLB. the results change noticeably when we employ a copying promotion mechanism: approx-online outperforms asap in nine of the 16 experiments, while the policies perform almost identically in three of the other seven cases. the magnitude of the disparity between approx-online and asap results is also dramatically larger. the differences in performance range from asap outperforming approx-online by 20% for vortex with a 64-entry TLB, to approx-online outperforming asap by 45% for isosurf with a 64-entry TLB. Overall, our results confirm those of Romer et al.: the best promotion policy to use when creating superpages via copying is approx-online. Taking the arithmetic mean of the performance differences reveals that asap is, on average, 6% better with a 64-entry TLB, and 4% better with a 128-entry TLB. speedup Impulse+asap Impulse+approx_online copying+asap copying+approx_online compress gcc vortex 1.07 1.03 0.84 0.82 1.06 1.060.92 filter 1.73 1.69 1.67 1.65 rotate dm Figure 10: Normalized speedups for each of two promotion policies on a 4-issue system with a 128-entry TLB. the relative performance of the asap and approx-online promotion policies changes when we employ different promotion mechanisms because asap tends to create superpages more aggressively than approx-online. the design assumption underlying the approx-online algorithm (and the reason that it performs better than asap when copying is used) is that superpages should not be created until the cost of TLB misses equals the cost of creating the superpages. Given that remapping has a much lower cost for creating superpages than copying, it is not surprising that the more aggressive asap policy performs relatively better with it than approx-online does. 4.2.3 Remapping vs. Copying When we compare the two superpage creation mechanisms, remapping is the clear winner, but by highly varying margins. the differences in performance between the best overall remapping-based algorithm (asap+remap) and the best copying-based algorithm (aonline+copying) is as large as 97% in the case of adi on both a 64-entry and 128-entry TLB. Overall, asap+remap outperforms aonline+copying by more than 10% in eleven of the sixteen experiments, averaging 33% better with a 64-entry TLB, and 22% better with a 128-entry TLB. 4.2.4 Discussion Romer et al. show that approx-online is generally superior to asap when copying is used. When remapping is used to build superpages, though, we find that the reverse is true. Using Impulse-style remapping results in larger speedups and consumes much less physical memory. Since superpage promotion is cheaper with Impulse, we can also afford to promote pages more aggressively. Romer et al.'s trace-based simulation does not model any cache interference between the application and the TLB miss handler; instead, that study assumes that each superpage promotion costs a total of 3000 cycles per kilobyte copied [40]. Table 7 shows our measured per-kilobyte cost (in CPU cycles) to promote pages by copying for four representative benchmarks. (Note that cycles per 1K average baseline bytes promoted cache hit ratio cache hit ratio filter 5,966 99.80% 99.80% raytrace 10,352 96.50% 87.20% dm 6,534 99.80% 99.86% Table 7: Average copy costs (in cycles) for approx-online policy. we also assume a relatively faster processor.) We measure this bound by subtracting the execution time of aol+remap from that of aol+copy and dividing by the number of kilobytes copied. For our simulation platform and benchmark suite, copying is at least twice as expensive as Romer et al. assumed. For gcc and raytrace, superpage promotion costs more than three times the cost charged in the trace-driven study. Part of these differences are due to the cache effects that copying incurs. We find that when copying is used to promote pages, approx-online performs better with a lower (more aggressive) threshold than used by Romer et al. Specifically, the best thresholds that our experiments revealed varied from four to 16, while their study used a fixed threshold of 100. This difference in thresholds has a significant impact on performance. For example, when we run the adi benchmark using a threshold of 32, approx-online with copying slows performance by 10% with a 128-entry TLB. In contrast, when we run approx-online with copying using the best threshold of 16, performance improves by 9%. In general, we find that even the copying-based promotion algorithms need to be more aggressive about creating superpages than was suggested by Romer et al. Given that our cost of promoting pages is much higher than the 3000 cycles estimated in their study, one might expect that the best thresholds would be higher than Romer's. However, the cost of a TLB miss far outweighs the greater copying costs; our TLB miss costs are about an order of magnitude greater than those assumed in his study. 5 Related Work A number of projects have proposed modifications to conventional CPU or DRAM designs to improve memory system performance, including supporting massive multithreading [2], moving processing power on to DRAM chips [25], or developing configurable architectures [50]. While these projects show promise, it is now almost impossible to prototype non-traditional CPU or cache designs that can perform as well as commodity processors. In addition, the performance of processor-in-memory approaches are handicapped by the optimization of DRAM processes for capacity (to increase bit density) rather than speed. the Morph architecture [50] is almost entirely configurable: programmable logic is embedded in virtually every datapath in the system, enabling optimizations similar to those described here. the primary difference between Impulse and Morph is that Impulse is a simpler design that can be used in current systems. the RADram project at UC Davis is building a memory system that lets the memory perform computation [34]. RADram is a PIM, or processor-in-memory, project similar to IRAM [25]. the RAW project at MIT [46] is an even more radical idea, where each IRAM element is almost entirely reconfigurable. In contrast to these projects, Impulse does not seek to put an entire processor in memory, since DRAM processes are substantially slower than logic processes. Many others have investigated memory hierarchies that incorporate stream buffers. Most of these focus on non-programmable buffers to perform hardware prefetching of consecutive cache lines, such as the prefetch buffers introduced by Jouppi [23]. Even though such stream buffers are intended to be transparent to the programmer, careful coding is required to ensure good memory performance. Palacharla and Kessler [36] investigate the use of similar stream buffers to replace the L2 cache, and Farkas et al. [14] identify performance trends and relationships among the various components of the memory hierarchy (including stream buffers) in a dynamically scheduled processor. Both studies find that dynamically reactive stream buffers can yield significant performance increases. the Imagine media processor is a stream-based architecture with a bandwidth-efficient stream register file [38]. the streaming model of computation exposes parallelism and locality in applica- tions, which makes such systems an attractive domain for intelligent DRAM scheduling. Competitive algorithms perform online cost/benefit analyses to make decisions that guarantee performance within a constant factor of an optimal offline algorithm. Romer et al. [40] adapt this approach to TLB management, and employ a competitive strategy to decide when to perform dynamic superpage promotion. They also investigate online software policies for dynamically remapping pages to improve cache performance [6, 39]. Competitive algorithms have been used to help increase the efficiency of other operating system functions and resources, including paging, synchronization, and file cache management. Chen et al. [11] report on the performance effects of various TLB organizations and sizes. Their results indicate that the most important factor for minimizing the overhead induced by TLB misses is reach, the amount of address space that the TLB can map at any instant in time. Even though the SPEC benchmarks they study have relatively small memory requirements, they find that TLB misses increase the effective CPI (cycles per instruction) by up to a factor of five. Jacob and Mudge [22] compare five virtual memory designs, including combinations of hierarchical and inverted page tables for both hardware-managed and software-managed TLBs. They find that large TLBs are necessary for good performance, and that TLB miss handling accounts for much of the memory-management overhead. They also project that individual costs of TLB miss traps will increase in future microprocessors. Proposed solutions to this growing TLB performance bottleneck range from changing the TLB structure to retain more of the working set (e.g., multi-level TLB hierarchies [1, 16]), to implementing better management policies (in software [21] or hardware [20]), to masking TLB miss latency by prefetching entries (again, in software [4] or hardware [41]). All of these approaches can be improved by exploiting superpages. Most commercial TLBs support superpages, and have for several years [30, 43], but more research is needed into how best to make general use of them. Khalidi [24] and Mogul [31] discuss the benefits of systems that support superpages, and advocate static allocation via compiler or programmer hints. Talluri et al. [32] report on many of the difficulties attendant upon general utilization of superpages, most of which result from the requirement that superpages map physical memory regions that are contiguous and aligned. 6 Conclusions the Impulse project attacks the memory bottleneck by designing and building a smarter memory controller. Impulse requires no modifications to the CPU, caches, or DRAMs. It has one special form of "smarts": the controller supports application-specific physical address remapping. This paper demonstrates how several simple remapping functions can be used in different ways to improve the performance of two important scientific application kernels. Flexible remapping support in the Impulse controller can be used to implement a variety of optimizations. Our experimental results show that Impulse's fine-grained remappings can result in substantial program speedups. Using the scatter/gather through an indirection vector remapping mechanism improves the NAS conjugate gradient benchmark performance by 210% and the volume rendering benchmark performance by 449%; using strided remapping improves performance of image filtering, image rotation, and dense matrix-matrix product applications by 94%, 166%, and 21%, respectively. Impulse's direct remappings are also effective for a range of programs. They can be used to dynamically build superpages without copying, and thereby reduce the frequency of TLB faults. Our simulations show that this optimization speeds up eight programs from a variety of sources by up to a factor of 2.03, which is 25% better than prior work. Page-level remapping to perform cache coloring improves performance of conjugate gradient by 17%. the optimizations that we describe should be applicable across a variety of memory-bound ap- plications. In particular, Impulse should be useful in improving system-wide performance. For ex- ample, Impulse can speed up messaging and interprocess communication (IPC). Impulse's support for scatter/gather can remove the software overhead of gathering IPC message data from multiple user buffers and protocol headers. the ability to use Impulse to construct contiguous shadow pages from non-contiguous pages means that network interfaces need not perform complex and expensive address translations. Finally, fast local IPC mechanisms like LRPC [5] use shared memory to map buffers into sender and receiver address spaces, and Impulse could be used to support fast, no-copy scatter/gather into shared shadow address spaces. Acknowledgments We thank Al Davis, Bharat Chandramouli, Krishna Mohan, Bob Devine, Mark Swanson, Arjun Dutt, Ali Ibrahim, Shuhuan Yu, Michael Abbott, Sean Cardwell, and Yeshwant Kolla for their contributions to the Impulse project. --R AMD Athlon processor technical brief. the Tera computer system. the NAS parallel benchmarks. Software prefetching and caching for translation buffers. Lightweight remote procedure call. Avoiding conflict misses dynamically in large direct-mapped caches the organization and use of parallel memories. Memory bandwidth limitations of future microprocessors. A simulation based study of TLB performance. Compaq Computer Corporation. Revisiting superpage promotion with hardware support. Image Processing for Computer Graphics. HAL Computer Systems Inc. Computer Architecture: A Quantitative Approach. Control Data STAR-100 processor design the intrinsic bandwidth requirements of ordinary programs. Pentium Pro Family Developer's Manual A look at several memory management units Improving direct-mapped cache performance by the addition of a small fully associative cache and prefetch buffers Virtual memory support for multiple page sizes. Scalable processors in the billion-transistor era: IRAM Fast Volume the cache performance and optimizations of blocked algorithms. Access ordering and memory-conscious cache utilization Big memories on the desktop. Surpassing the TLB performance of superpages with less operating system support. Sparse matrix kernels for shared memory and message passing systems. Active pages: A model of computation for intelligent memory. RSIM reference manual Evaluating stream buffers as a secondary cache replacement. Interactive ray tracing for isosurface rendering. A bandwidth-efficient architecture for media processing Using Virtual Memory to Improve Cache and TLB Performance. Reducing TLB and memory overhead using online superpage promotion. ATOM: A system for building customized program analysis tools. Increasing TLB reach using superpages backed by shadow memory. To copy or not to copy: A compile-time technique for assessing when data copying should be used to eliminate cache conflicts Digital Image Warping. URSIM reference manual. Memory system support for image processing. Architectural adaptation for application-specific locality optimizations --TR Lightweight remote procedure call The cache performance and optimizations of blocked algorithms A simulation based study of TLB performance To copy or not to copy ATOM Evaluating stream buffers as a secondary cache replacement Avoiding conflict misses dynamically in large direct-mapped caches Surpassing the TLB performance of superpages with less operating system support Reducing TLB and memory overhead using online superpage promotion Memory bandwidth limitations of future microprocessors The intrinsic bandwidth requirements of ordinary programs Fast volume rendering using a shear-warp factorization of the viewing transformation The Tera computer system Memory-system design considerations for dynamically-scheduled processors Active pages Increasing TLB reach using superpages backed by shadow memory Computer architecture (2nd ed.) Interactive ray tracing for isosurface rendering A bandwidth-efficient architecture for media processing A look at several memory management units, TLB-refill mechanisms, and page table organizations Improving direct-mapped cache performance by the addition of a small fully-associative cache and prefetch buffers Recency-based TLB preloading Digital Image Warping Image Processing for Computer Graphics Scalable Processors in the Billion-Transistor Era Baring It All to Software 3-D transformations of images in scanline order Access ordering and memory-conscious cache utilization Software-Managed Address Translation Impulse Memory System Support for Image Processing Architectural adaptation for application-specific locality optimizations Reevaluating Online Superpage Promotion with Hardware Support Using virtual memory to improve cache and tlb performance --CTR Jun Miyazaki, Hardware supported memory access for high performance main memory databases, Proceedings of the 1st international workshop on Data management on new hardware, June 12-12, 2005, Baltimore, Maryland McCorkle, Programmable bus/memory controllers in modern computer architecture, Proceedings of the 43rd annual southeast regional conference, March 18-20, 2005, Kennesaw, Georgia Hur , Calvin Lin, Memory Prefetching Using Adaptive Stream Detection, Proceedings of the 39th Annual IEEE/ACM International Symposium on Microarchitecture, p.397-408, December 09-13, 2006 Lixin Zhang , Mike Parker , John Carter, Efficient address remapping in distributed shared-memory systems, ACM Transactions on Architecture and Code Optimization (TACO), v.3 n.2, p.209-229, June 2006 Daehyun Kim , Mainak Chaudhuri , Mark Heinrich , Evan Speight, Architectural Support for Uniprocessor and Multiprocessor Active Memory Systems, IEEE Transactions on Computers, v.53 n.3, p.288-307, March 2004 Xipeng Shen , Yutao Zhong , Chen Ding, Locality phase prediction, ACM SIGOPS Operating Systems Review, v.38 n.5, December 2004 Eero Aho , Jarno Vanne , Timo D. Hmlinen, Configurable data memory for multimedia processing, Journal of Signal Processing Systems, v.50 n.2, p.231-249, February 2008 Xiaodong Li , Zhenmin Li , Yuanyuan Zhou , Sarita Adve, Performance directed energy management for main memory and disks, ACM Transactions on Storage (TOS), v.1 n.3, p.346-380, August 2005 Xiaodong Li , Zhenmin Li , Francis David , Pin Zhou , Yuanyuan Zhou , Sarita Adve , Sanjeev Kumar, Performance directed energy management for main memory and disks, ACM SIGARCH Computer Architecture News, v.32 n.5, December 2004 Zhen Fang , Lixin Zhang , John B. Carter , Ali Ibrahim , Michael A. Parker, Active memory operations, Proceedings of the 21st annual international conference on Supercomputing, June 17-21, 2007, Seattle, Washington Bruno Diniz , Dorgival Guedes , Wagner Meira, Jr. , Ricardo Bianchini, Limiting the power consumption of main memory, ACM SIGARCH Computer Architecture News, v.35 n.2, May 2007 Enric Gibert , Jesus Sanchez , Antonio Gonzalez, Distributed Data Cache Designs for Clustered VLIW Processors, IEEE Transactions on Computers, v.54 n.10, p.1227-1241, October 2005 Xipeng Shen , Yutao Zhong , Chen Ding, Predicting locality phases for dynamic memory optimization, Journal of Parallel and Distributed Computing, v.67 n.7, p.783-796, July, 2007 Ruchira Sasanka , Man-Lap Li , Sarita V. Adve , Yen-Kuang Chen , Eric Debes, ALP: Efficient support for all levels of parallelism for complex media applications, ACM Transactions on Architecture and Code Optimization (TACO), v.4 n.1, p.3-es, March 2007 Martin Hirzel, Data layouts for object-oriented programs, ACM SIGMETRICS Performance Evaluation Review, v.35 n.1, June 2007

computer architecture;memory systems

507198

Worst and Best Irredundant Sum-of-Products Expressions.

AbstractIn an irredundant sum-of-products expression (ISOP), each product is a prime implicant (PI) and no product can be deleted without changing the function. Among the ISOPs for some function $f$, a worst ISOP (WSOP) is an ISOP with the largest number of PIs and a minimum ISOP (MSOP) is one with the smallest number. We show a class of functions for which the Minato-Morreale ISOP algorithm produces WSOPs. Since the ratio of the size of the WSOP to the size of the MSOP is arbitrarily large when $n$, the number of variables, is unbounded, the Minato-Morreale algorithm can produce results that are very far from minimum. We present a class of multiple-output functions whose WSOP size is also much larger than its MSOP size. For a set of benchmark functions, we show the distribution of ISOPs to the number of PIs. Among this set are functions where the MSOPs have almost as many PIs as do the WSOPs. These functions are known to be easy to minimize. Also, there are benchmark functions where the fraction of ISOPs that are MSOPs is small and MSOPs have many fewer PIs than the WSOPs. Such functions are known to be hard to minimize. For one class of functions, we show that the fraction of ISOPs that are MSOPs approaches 0 as $n$ approaches infinity, suggesting that such functions are hard to minimize.

2. MSOP : STn; k n . 3. Proof. See the Appendix. tu In the proof of Lemma 2.1, we showed that which is six more than the lower bound given in 3 of Theorem 3.1. Therefore, for n 7 and k 3, the lower bound is not tight. However, for k 1, the lower bound is exact. That is, Theorem 3.2. 2: Proof. See the proof of Theorem 7.3 in the Appendix. tu SASAO AND BUTLER: WORST AND BEST IRREDUNDANT SUM-OF-PRODUCTS EXPRESSIONS 937 A special case of these theorems occurs when n 3 and TABLE 1 k 1. and for STn; 1 versus n Example 3.3. ST3; 1x1 _ x2 _ x3x1 _ x2 _ x3 has the following properties: 2. MSOP : ST3; 1 3. 3. Fig. 1a and Fig. 1b show the MSOP and WSOP of respectively. Interestingly, the ISOP generator of Minato [19], which is based on Morreale's [20] algorithm produces a WSOP for ST3; 1 instead of an MSOP. This will be discussed in more detail later. Definition 3.2. The redundancy ratio of a function f is The normalized redundancy ratio of an n-variable function f is are the sizes of WSOPs and MSOPs. If this ratio is small, any logic minimization algorithm will do well since, even if a WSOP is generated, it is not much worse than an MSOP. On the other hand, a large ratio suggests that care should be exercised. The normalized redundancy ratio is normalized with respect to the number of variables. It is a convenience; it allows one to compare the redundancy ratio of two functions with a different number of variables. From the expressions for MSOP : STn; k and given in Theorem 3.1, we can state: Theorem 3.3. s From the expressions for MSOP : STn; 1 and Theorems 3.1 and 3.2, respec- tively, we can state: Theorem 3.4.STn; 1 2 ; r Table 1 shows the values of and for STn; 1, where 8. It can be seen that takes its maximum value when That is, as n increases above 2, first increases, peaking at 4, and then it continually decreases. From Theorem 3.4, is monotone increasing with an upper limit of 2. Thus, for STn; 1 functions, the number of PIs in a WSOP is never more than two times the number of PIs in an MSOP. An important question is whether there exist functions where is larger than 2. Indeed, we show a class of functions in which increases without bound as n increases. This has important consequences for heuristics that produce ISOPs. For such heuristics there is the prospect of generating an ISOP whose size is much larger than the minimum. We consider this topic now. Definition 3.3. Let STm; kr be the n m r-variable function ^r where r is the AND (product) of r functions. Theorem 3.5. STm; kr has the following properties: r r mr 2. MSOP : STm; k . 3. Proof. See the Appendix. tu For k 1, we have: Theorem 3.6. Example 3.4. For m 3 and k 1, ST3; 1r has 6r PIs, We have: Theorem 3.7. "#r Example 3.5. For m 4 and k 1, we have From this, it can be seen that becomes arbitrarily large as r approaches infinity. In this example, there are n 4r 938 IEEE TRANSACTIONS ON COMPUTERS, VOL. 50, NO. 9, SEPTEMBER 2001 variables. This represents a class of functions for which grows without bound as the number of variables grows. 4EXTENSION TO MULTIPLE-OUTPUT FUNCTIONS In the case of multiple-output functions, minimization of two-level networks or programmable logic arrays (PLAs) can be done using characteristic functions [26], [27], [29]. Definition 4.1. For an n-variable function with m output values, form an n 1-variable two-valued single output function Fx1;x2; .;xn;Xn1, where xi is a binary valued variable, for 1 i n and Xn1 takes m values such that Fx1;x2; .;xn;j1 iff fjx1;x2; .;xn1 represents all and only the permitted combinations of inputs and nonzero output values of f. F is called the characteristic function (for nonzero outputs). The significance of the characteristic function is seen in Theorem 4.1 below. Definition 4.2. XS is a literal, where X takes a value in 1g such that XS 1 if X a 2 S and XS 0, otherwise. A logical product of literals that contains at most one literal for each variable is a product term. Products combined with OR operators form a sum-of-products expression (SOP). A Prime implicant (PI), irredundant sum-of-products expression (ISOP), worst ISOP (WSOP), and minimum SOP (MSOP) are defined in a manner similar to the two-valued case. Theorem 4.1 [15], [27], [29]. The number of AND gates in the minimum AND-OR two-level network for the function f0;f1; .;fm1 is equal to the number of PIs in the MSOP for the characteristic function F. Definition 4.3. An n-bit decoder has n inputs x1;x2; .;xn, and 2n outputs f0;f1; .;f2n1, where fi 0 iff the binary number representation of x1x2; .;xn is i. Example 4.1. The 4-bit decoder has 16 outputs, as follows: . . . Definition 4.4. DEC_n is the characteristic function of an n-bit decoder. Example 4.2. DEC_4 is shown in positional cube notation in the upper table of Fig. 2. That is, each entry in this table is a prime implicant of DEC_4, where xi appears as xi, xi, or don't care (absent) if the corresponding entry is 10, 01, or 11, respectively. For X5, the entry 0111111111111111 is the literal X , etc. Therefore, thefirst corresponds to Fig. 2. Positional cubes for two ISOPs of DEC_4. the prime implicant x1x2x3x4X5 . Collectively, the 16 entries in the upper table of Fig. 2 represent an ISOP of DEC_4 with PIs. An ISOP for DEC_4 with only eight PIs exists, as shown in the lower table of Fig. 2. The observations of Example 4.2 can be generalized as follows: Theorem 4.2. The function DEC_n has a WSOP that requires at least 2n PIs and an MSOP that requires at most 2n PIs. The above theorem proves the existence of an n-variable 2n-output variable function, where the sizes of the MSOP and the WSOP are at most 2n and at least 2n, respectively. The upper ISOP for DEC_4 shown in Fig. 2 is not a WSOP since an ISOP with 20 PIs has been found for DEC_4. 5DERIVATION OF ALL ISOPs Very little is known about the distribution of the sizes for ISOPs. For example, even for single-output functions, we know of no study that shows how many ISOPs exist with various number of product terms. Although various methods to generate all the ISOPs for a logic function are known [22], [12], [21], [6], [35], [25], no experimental results have been reported. Experiments are computationally intensive even for functions with a small number of variables. However, we can obtain the statistical properties of ISOPs for some interesting functions. Before showing the complete algorithm, consider the SASAO AND BUTLER: WORST AND BEST IRREDUNDANT SUM-OF-PRODUCTS EXPRESSIONS 939 has two MSOPs with three PIs and four WSOPs with four PIs. Fig. 3. fx1;x2;x3ST3; 1. Example 5.1. f ST3; 1x1 _ x2 _ x3x1 _ x2 _ x3 has six minterms (Fig. Fig. 4 is the covering table for ST3; 1. It shows the following relations: To cover m1;p1 _ p6 is necessary: To cover m3;p1 _ p2 is necessary: To cover m2;p2 _ p3 is necessary: To cover m6;p3 _ p4 is necessary: To cover m4;p4 _ p5 is necessary: To cover m5;p5 _ p6 is necessary: To satisfy all the conditions at the same time, we have Pf1, where Pf is called the Petrick function [22]. By expanding Pf into SOPs, we have p1p3p5 _ p2p3p5p6 _ p1p2p4p5 _ p2p4p5p6 _ p1p3p4p6 _ p2p3p4p6 _ p1p2p4p6 _ p2p4p6: Note that each product with an underline is covered by another product having fewer literals. Such products are redundant. Deleting these products, we have _ p2p3p4p6 _ p2p4p6: Pf consists of all the PIs of the Petrick function [23] and each PI of Pf corresponds to an ISOP for f. Furthermore, each literal pi in the PI of Pf corresponds to a PI for f. For example, p1p3p5 corresponds to the ISOP x1x3 _ x2x3 _ x1x2. Note that there are six ISOPs; two have three PIs, while four have four PIs. Thus, ST3; 1 Fig. 4. Covering table of ST3;1. In this way, all the ISOPs are obtained. For general functions, the number of minterms and PIs are very large. Thus, we use an ROBDD (reduced ordered binary decision diagram) to represent the function and a Prime_TDD (Ternary decision diagram) [31] to represent the set of all the PIs. In the Prime_TDD for f, each path from the root node to the constant 1 node corresponds to a PI for f.We also use an ROBDD to represent the Petrick function. While there are many ways to generate all the ISOPs of a given function f, we use the following algorithm: Algorithm 5.1 (Generation of all ISOPs for a function f). 1. Generate all the PIs for f by using the Prime_TDD (the ternary decision diagram representing PIs) of f. 2. From the set of PIs and the set of minterms for f, generate the Petrick function Pf (which represents the covering table [22]). 3. Generate the Prime_TDD (which represents all the PIs) of Pf. 4. Generate the 1-paths of the Prime_TDD and, for each 1-path, generate the corresponding ISOP. In the Prime_TDD in Step 4, each path from the root node to the constant 1 corresponds to a PI for Pf and to an ISOP for f. Each 1 edge has weight 1 and each 0 edge has weight 0. The total sum of weights from the root node to the constant 1 nodes is the number of PIs in the ISOP. Note that the shortest path corresponds to an MSOP and the longest path corresponds to a WSOP. 6.1 STn; k Functions Using Algorithm 5.1, we compare the number of PIs in STn; kr for different n, k, and r. Table 2 shows the number of PIs in the MSOP and the WSOP of STn; kr, as well as the total number of PIs. Shown also are the results of the Minato-Morreale algorithm. The 9SYM (or SYM9) [11], [15] function shown in [3, p. 165] is identical to ST9; 3. It has 1,680 PIs, POP [9], a PRESTO-type [4], [33] logic minimization algorithm, produced an ISOP with 148 products. CAMP [1] produced an ISOP with 130 PIs, while MINI [15] did well, producing 85 PIs. Table 3 shows the distribution of ISOPs to the number of PIs in an ISOP for STn; 1 for 3 n 7. This data was obtained by Algorithm 5.1. It can be seen that the set of MSOPs is small compared to the set of all ISOPs. 6.2 Other Functions We also applied Algorithm 5.1 to compare the number of PIs for multiple-output functions. Table 4 shows the distribution of the number of PIs in ISOPs for various functions [32]. INCn is an n-input n 1 output function such that the value of the output is x 1, where x is the value of the 940 IEEE TRANSACTIONS ON COMPUTERS, VOL. 50, NO. 9, SEPTEMBER 2001 Number of PIs and Redundancy Ratio for Various Functions input; WGT5 is the same as RD53, a 5-input 3-output function, where the output is a binary number whose value is the number of 1s on the inputs; ROT6 computes the square root of a 6-bit integer; LOG5 computes the logarithm of the 5-bit integer; ADR3 is a 3-bit adder; and SQR5 computes the square of the 5-bit input. Note that all the ISOPs for INC6 have the same number of PIs. This means any logic minimizer obtains an exact minimum solution. This is also is true for WGT5. For ADR3, most of the ISOPs have 31 PIs or 33 PIs. This is consistent with the observation that the logic minimization of ADR3 is relatively easy. For SQR5, the distribution is very wide. The MSOPs have 27 PIs, while WSOPs have 37 PIs. This is consistent with the observation that the minimization of SQR5 is more difficult. Note that SQR5 is a 10 output binary function. The data shown is for all outputs. Although we could not obtain the distribution for SQR6 due to the memory overflow, we conjecture that the distribution of number of PIs for SQR6 is also wide. We also developed WIRR, a heuristic algorithm to obtain ISOPs with many products. For SQR5, SQR6, and 9SYM, the numbers of PIs in the solutions are shown in Table 5. 7DISTRIBUTION OF ISOPs-AN ANALYTIC APPROACH The distribution of ISOPs to the number of PIs is a way to represent the search space a heuristic algorithm must traverse in a minimization of an expression. For the case of STn; 1 functions, we can show a part of this distribu- tion; a graph representation of the set of PIs allows this. Definition 7.1. Let F be an ISOP of STn; 1.Inthegraph representation GF of F 1. GF has nodes x1;x2; ., and xn, and 2. GF has an edge from xi to xj iff xixj is a PI in F. SASAO AND BUTLER: WORST AND BEST IRREDUNDANT SUM-OF-PRODUCTS EXPRESSIONS 941 Distribution of ISOPs in STn; 1 Functions Example 7.1. Fig. 5 shows the graph representations of the MSOP and WSOP for ST3; 1 (shown in Fig. 1). We show that the graph representation of an ISOP of F has a special property. Definition 7.2. A directed graph G is strongly connected iff for every pair of vertices a; b in G, there is a path from a to b and from b to a. A directed graph G is minimally strongly connected iff it is strongly connected and the removal of any edge causes G not to be strongly connected. Theorem 7.1. Let GF be a graph representation of F. F is an ISOP of STn; 1 iff GF is minimally strongly connected. Proof. See the Appendix. tu The graph representations of the MSOP and WSOP of shown in Fig. 5, are both strongly connected, as they should be by Theorem 7.1. Since each edge represents a prime implicant, an MSOP has a graph representation with the fewest edges. This observation facilitates the enumeration of MSOPs. Theorem 7.2. The number of MSOPs for STn; 1 is n 1!. Proof. See the Appendix. tu The graph representation allows a characterization of ISOPs. Specifically, complementing all variables in an ISOP of STn; 1 is equivalent to reversing the direction of all edges in the graph representation GF of F.IfGF is minimally strongly connected, then the graph obtained from GF by reversing the direction of all edges is also minimally strongly connected. This proves: Lemma 7.1. If F is an ISOP of STn; 1, then the SOP derived from F by complementing all variables is an ISOP of STn; 1. Example 7.2. When all variables are complemented, the graph representations of the ISOPs shown in Fig. 5 produce the graphs in Fig. 6, which also represent an MSOP and a WSOP. It is important to note the difference between changing an ISOP F and changing the function realized by F. That is, an function is unchanged by a complementation of variables, i.e. it is a self-anti-dual function [32]. However, an ISOP for an STn; k function may or may not be changed when all variables are comple- mented. For example, F x1x2 _ x2x3 _ x3x1 is an ISOP for ST3; 1. Complementing all variables in F yields, F x1x2 _ x2x3 _ x3x1, a different ISOP. It is interesting that the WSOP for ST3; 1 is unchanged by a complementation of all variables, as can be seen by comparing Fig. 5b with Fig. 6b. The invariance of an ISOP with respect to complementation of all literals is a unique characteristic of WSOPs, as shown in the next result. Lemma 7.2. Let F be an ISOP of STn; 1. F is a WSOP iff complementing all variables in F leaves F unchanged. Proof. See the Appendix. tu Distribution of ISOPs in Arithmetic Functions 942 IEEE TRANSACTIONS ON COMPUTERS, VOL. 50, NO. 9, SEPTEMBER 2001 Number of PIs Produced by Various Algorithms on Three Benchmark Functions It is interesting that Lemma 7.2 does not generalize to k. Specifically, for ST5; 2, the ISOP F x1x2x3x5 _ x3x5x1x2 _ x1x3x2x4 _ x2x4x1x3 _ x1x4x2x5 _ x2x5x1x4 _ x1x5x3x4 _ x3x4x1x5 _ x2x3x4x5 _ x4x5x2x3 is invariant with respect to complementation of all variables. However, it is an MSOP and not a WSOP. We can also enumerate WSOPs as follows: Theorem 7.3. The number of WSOPs for STn; 1 is nn2. Proof. See the Appendix. tu The graph representation allows the enumeration of other classes of ISOPs. For example, we can enumerate ISOPs that have one more PI than is in the MSOP. Specifically, Theorem 7.4. The number of ISOPs for STn; 1 with n 1 PIs is Proof. See the Appendix. tu By comparing the number of MSOPs with either the number of WSOPs or the number of ISOPs with one more PI than in the MSOP, we find that the former is much less than either of the latter for large n. That is, as n approaches infinity, the ratio of MSOPs to WSOPs approaches 0 (use Stirling's formula to replace n 1! in the expression for the number of MSOPs). This proves the following: Theorem 7.5. The fraction of ISOPs for STn; 1 that are MSOPs approaches 0 as n approaches infinity. It is interesting that the ratio of the number of ISOPs with than is in an MSOP) to the number of WSOPs also approaches 0 as n approaches infinity. This suggests that WSOPs are much more common than minimal or near-minimal ISOPs. The existence of an algorithm that finds the worst sum-of- products expression for a class of functions is surprising. It counters our expectation that a heuristic algorithm should perform reasonably well. Also, the large difference between the size of the worst and the best expression is especially compelling since such an algorithm will perform very poorly. It is, therefore, an interesting question of whether there are other algorithms and other functions that exhibit the same characteristics. Fig. 6. Graph representations of Fig. 5 with all variables complemented. Fig. 5. Graph representations of the MSOP and WSOP for ST3;1. (a) MSOP. (b) WSOP. (a) MSOP. (b) WSOP. SASAO AND BUTLER: WORST AND BEST IRREDUNDANT SUM-OF-PRODUCTS EXPRESSIONS 943 We show a multiple-output function where the worst and the best ISOPs differ greatly in size. Specifically, a decoder with 2n outputs and n inputs realizes a function where a WSOP has at least 2n PIs and an MSOP has at most 2n PIs. Since this is a commonly used logic function, disparity in the size of WSOPs and MSOPs cannot be viewed as a characteristic of contrived functions only. Although computationally intensive, enumeration of the ISOPs for representative functions gives needed insight into the problem. We show an algorithm to compute all ISOPs of a given function. We apply it to benchmark functions and show there are significant differences in the distributions of ISOPs. That is, some functions have a narrow distribution, where the WSOP is nearly or exactly the same size as the MSOP. These tend to be easy to minimize. For example, for unate functions [17] and parity functions, there is exactly one ISOP. Such functions are classified as trivial in the Berkeley PLA Benchmark Set (e.g., ALU1, BCD, DIV3, CLP1, CO14, MAX46, NEWPLA2, NEWBYTE, NEWTAG, and RYY6) [26]. Other functions display a wide range and tend to be hard to minimize. For example, 9SYM or SYM9 (ST9; has a wide range, i.e., the number of PIs in a WSOP and an MSOP is 148 and 84 PIs, respectively. This function is known to be hard to minimize. For a class of functions, we provide an analysis showing that the number of MSOPs is significantly smaller than the number of WSOPs. That is, by showing a correlation with directed graphs, we enumerate all MSOPs and all WSOPs of the class and show that the number of MSOPs and WSOPs is n 1! and nn2, respectively. As n increases, the ratio of PIs in a WSOP to the PIs in an MSOP grows without bound. This suggests such functions are hard to minimize. A complete understanding of the minimization process will require knowledge of the search space and how various algorithms progress through it. However, such an understanding is not likely to be achieved in the near future. Our research suggests that there is merit to understanding the correlation between the degree of difficulty in minimizing a function and the distribution of its ISOPs. Lemma 2.1. Proof. There are two steps. In the first step, we prove that an ISOP with 70 PIs exists for this function. In the second step, we show that it is a WSOP. For the first step, it is convenient to view the symmetric function as having three parts. Specifically, f0;1;3;4;6;7g f0;1g f3;4g f6;7g A WSOP is obtained by finding a WSOP of each of the three parts separately. Consider the 7-bit Hamming code shown in Table 6. For each code word, create a PI that covers two minterms by replacing one of the most abundant bits in A 7-Bit Hamming Code the code word by a don't care. In the case of code word, 0000000; this creates seven PIs, each of which covers the minterm with all variables 0 and one minterm with exactly one 1. This covers all minterms of S7 . Similarly, seven PIs generated from code word 1111111 cover all minterms of S7 . All of the remaining 14 code words have either four 0s and three 1s or four 1s and three 0s. For each, create four PIs by changing one of the four logic values in the majority to a don't care. Collectively, the four PIs cover the original code word and four words that are a distance one away from the code word. Because the distance between any pair of code words is at least three, a change in a single bit of a code word in the Hamming code creates a word that is not a code word that is distinct from either another code word or a word that is one bit different from another code word. This implies that those minterms a distance one away from a code word are covered by at most one PI. It follows that each PI is irredundant. Since each of the 14 code words corresponds to a set of PIs that cover five distinct minterms, there are 14 5 70 minterms total. On the otherhand,thenumberofmintermsforS7 is f3;4g It follows that these PIs cover all the minterms of S7 . In all, 7 56 7 70 PIs cover all of the 8 minterms of the function. It follows that this set of PIs is a cover for S7 . Further, it is an irredundant cover and we have an ISOP. We have proven that an ISOP with 70 PIs exists for S7 . We show that this is a WSOP by showing that no more than seven, 56, and seven PIs can cover the minterms in S7 , S7 , and S7 , respectively. Since f0;1g f3;4g f6;7g S7 and S7 are monotone functions, their ISOPs are unique. Each consists of seven PIs. For S7 , the ISOP of above covers the 70 minterms associated with this function. On the contrary, assume that the proposed ISOP is not a WSOP. Thus, there is a set of p>56 PIs that forms an ISOP for these minterms. Each PI covers exactly two minterms, for a total of 2p>112 instances of a PI covering a minterm. Let m1 and m>1 be the number of minterms covered by one and more than one PI, 944 IEEE TRANSACTIONS ON COMPUTERS, VOL. 50, NO. 9, SEPTEMBER 2001 respectively. m>1 70 m1. Since the set of PIs is irredundant, each PI covers at least one minterm that is not covered by any other PI. Thus, m1 p>56.It follows that 2p 3m1 > 280. Further, 2p m1 > 280 4m1 470 m1 and we can write > 4: A:1 Here, the numerator is the number of instances in which a PI covers a minterm that is covered by more than one PI, while the denominator represents the number of minterms covered by more than one PI. Since this ratio exceeds four, by the Pigeonhole Principle, there is at least one minterm covered by at least five PIs. But, this is impossible; each minterm is covered by no more than four PIs (i.e., each code word is covered by a PI derived from a code word by converting one of the four most abundant variables, 0 or 1, to a don't care). Thus, it must be that the proposed ISOP is a WSOP. tu Theorem 3.1. 2. MSOP : STn; k n . 3. Proof. 1. An implicant of STn; k has the form xi1 xi2 xik xink1 xink2 xin : Thatis,forthisimplicanttobe1,atleast k variables (xi1 ;xi2 ; ., and xik ) must be 1 and at least k variables must be 0 (xink1 ;xink2 ; ., and xin ), where 2k n.Thisimplicantisprime; deleting a literal creates an implicant that is 1 when STn; k should be 0. Specifically, deleting creates an implicant that is 1 when less than n variables are 1, while deleting creates an implicant that is 1 when more than n k variables are 1. The number of such PIs is the number of ways 3. to separate n variables into three parts, where order withina part is not important. This is the multinomial n n! . 2. First, we show that n is a lower bound on the number of PIs of STn; k. Then, we show a set of n PIs that covers all and only minterms in the function STn; k. It follows that the OR of all PIs in is an MSOP for STn; k. Consider the set MI of n minterms of the form xi1 xi2 .xik xik1 xik2 .xin , where ij 2 of minterms that are 1 when exactly k of the n variables are 0. No PI for STn; k covers two or more minterms in MI. As such, MI is a set of independent minterms and at least n PIs are needed. is formed as follows: For each minterm mt in MI, apply Algorithm 1.1 below, producing Pmt,a PI that covers mt. Add Pmt to . Since noPI covers two or more minterms in MI, has n distinct PIs. Since Pmt has exactly k 0s and k 1s, it covers only minterms in STn; k. Next, we show that covers all minterms in STn; k by applying Algorithm 1.1 to an arbitrary minterm mt0 of producing Pmt0 , a PI that covers mt0. from Pmt0 by setting all -s to 1s. Applying Algorithm 1.1 to mt00 yields Pmt00 that is identical to Pmt0 , from which we can conclude Pmt0 2 . Algorithm 1.1 (Produce a PI that covers a given minterm). Input: Minterm mt mt0mt1 .mtn 1 Output: Prime implicant Pmt Pmt0Pmt1 .Pmtn 1 (Initially, Pmtifor all i where 1. ZeroOnePairs 0 2. Repeat until ZeroOnePairs k do { if and mtimti s01 then and ZeroOnePairs ZeroOnePairs 1g; where index addition is mod n (in this algorithm, we assume that subscript indices range from 0 to n 1, i.e., the variables are It is straightforward to show that Algorithm 1.1 produces a PI in if the minterm input has at least k 1s and k 0s. This PI may be different depending on the values of i chosen in each repetitive step. Assume We can form an ISOP of STn; k as follows: F xnF1 _ xnF2 _ F3; where a. F1 is an SOP such that each product term is formed as the AND of i) a set X1 of fxng, and of ii) k uncomplemented variables from X fxngX1,wherethe indices of the uncomplemented variables are all as small as possible (given the choice of X1). Because X1 can be chosen in nk1 ways, F1 has nk1 product terms. b. F2 is an SOP such that each product term is formed as the AND of i) a set X2 of k 1 uncomplemented variables, SASAO AND BUTLER: WORST AND BEST IRREDUNDANT SUM-OF-PRODUCTS EXPRESSIONS 945 where X2 X fxng and of ii) k complemented variables from X fxngX2, where the indices of the complemented variables are all as small aspossible. Because X2 can be chosen in nk1 ways, F2 has nk1 product terms. c. F3 is one of the WSOPs for STn 1;k. Fromthe inductive hypothesis, F3 has F is an expression for STn; k as follows: Consider a minterm m in STn; k.Ifm has at least k 1s and at least k 0s, regardless of the value of xn, then m is covered by F3.Ifm has exactly k 0s, including a 0 value for xn, and at least k 1s, then it is covered by xnF1.Ifm has exactly k 1s, including a 1 value for xn, and at least k 0s, it is covered by xnF2. Thus, F covers all minterms in STn; k. Because each PI in xnF1 and xnF2 (and also in F3) has exactly k uncomplemented and exactly complemented variables, F covers only minterms with at least k 0s and at least k 1s, i.e., only minterms in STn; k. It follows that F is an SOP for STn; k. If F has no redundant PIs, its n1 k1 k1 an ISOP for STn; k. Thus, Next, we show that F has no redundant PIs. First, each PI in xnF1 covers a minterm m0 having k 0s and n k 1s that is not covered by any combination of PIs from xnF2 and F3. Thus, no PI in xnF1 is redundant. By a similar argument, no PI in xnF2 is redundant. Second, no PI in F3 is redundant, as follows: Since F3 is a WSOP for STn 1;k,noPIin F3 is covered by the OR of one or more PIs in F3. If the OR of PIs from xnF1 and xnF2 covers aPIP from F3, then it follows that at least one product term P1 in F1, when ANDed with at least one product term P2 in F2, yields a non-0 result. Let and If P1P2 0,nosi is the same as a uj and no tpis the same as a vq. But, t1;t2; ., and tk were chosen to be as small as possible without overlapping s1;s2; .,andsk1, while v1;v2; ., and vk were chosen to be as small as possible without overlapping u1;u2; ., and uk1. Consider the indices I f1; 2; .;kg. The smallest index in I that appears neither in S fs1;s2; .;sk1g nor in U fu1;u2; .;uk1g appears in both T ft1;t2; .;tkg and V fv1;v2; .;vkg, causing P1P2 0, a contradiction. tu Theorem 3.5. STm; kr has the following properties: r r mr 2. MSOP : STm; k . 3. Proof. Items 1, 2, and 3 follow from the observation that a minterm in the product function STm; kr can be viewed as the AND of a minterm from each of the factor functions STm; k. Thus, a PI of STm; kr can be viewed as the product of a PI from each STm; k and directly. Also, r mr r follows directly. That is, m is an upper bound on the number of PIs in an MSOP for STm; kr, as an ISOP for STm; kr can be formed as the AND of PIs from the MSOPs of STm; k. As is shown by Voight and Wegener [36], certain product functions can have fewer PIs in their MSOPs than the product of the number of PIs in the MSOPs of the factor functions. However, when the factor functions are STm; k, we can observe the following: Let M be the set of minterms covered by STm; k, in which exactly k variables are uncomplemented. Since the PIs of exactly k uncomplemented and k complemented variables, none cover two or more minterms in M. It follows that at least m PIs are needed to cover STm; k. It follows that at least PIs are needed to cover the minterms in STm; kr that are the product of minterms in the set M for each factor function. Thus, r mr A WSOP for STm; kr can be formed as the product of WSOPs for each factor function. Since 2 m 2k is a lower bound on the number of PIs in each factor directly. tu Theorem 7.1. Let GF be the graph representation of F. F is an ISOP of STn; 1 iff GF is minimally strongly connected. Proof. (if) Consider a minimally strongly connected digraph GF , where F is its corresponding SOP. Thus, for every edge xj;xi in GF , there is an implicant xjxi in F.Onthe contrary, assume that F is not an ISOP for STn; 1. That is, either 1) F does not cover STn; 1 or 2) F covers STn; 1 but has a redundant PI. Consider 1). If F does not cover STn; 1, then there is a minterm mt such that mt either a) has no complemented variables or no uncomplemented variables, but F is 1 for this assignment or b) has at least one uncomplemented variable and at least one uncomplemented variable, but F is 0 for this assignment. The first part, a), is not possible; all PIs cover only minterms with at least one 946 IEEE TRANSACTIONS ON COMPUTERS, VOL. 50, NO. 9, SEPTEMBER 2001 complemented variable and at least one uncomplemented variable. The second part, b), is also not possible, as follows: Because GF is strongly connected, there is a path xj xk1 ;xk2 ; ;xkm xi, from xj to xi. Since xj 0 and xi 1, the assignment of values to xk1 ;xk2 ; ., and xkm corresponding to mt has the property that there is an s such that xks 0 and xks1 1. The corresponding PI xks xks1 is in F and is 1 for the assignment associated with mt. Thus, F covers mt. Consider 2). If F covers STn; 1, but has a redundant PI, xjxi, then F0, which is F with xjxi removed, also covers STn; 1. But, GF0 is GF with one edge, xjxi, removed. It follows that GF is not minimally strongly connected. (only if) Let GF be a graph representation of an ISOP F of STn; 1. Assume, on the contrary, that GF is not minimally strongly connected. That is, either 1) GF is not strongly connected or 2) GF is strongly connected, but is not minimal. If GF is not strongly connected, there are two nodes xj and xi such that no path exists from xj to xi. Let Sucxj be the set of all nodes for which there is a path from xj, i.e., all successors of xj. Let Prexi be the set of all nodes for which there exists a path to xi, i.e., all predecessors of xi. Consider a minterm mt that is 0 for xj and all variables associated with nodes in Sucxj and is 1 for xi and all variables associated with nodes in Prexi. Since there is no path from xj to xi, Sucxj\Prexi and such an assignment assigns exactly one value to nodes in Sucxj[Prexi[fxj;xig. Choose the values of all other variables to be 1 (or 0). No edge in GF has a 0 at its tail and a 1 at its head. Thus, all PIs are 0 and F is not an SOP for STn; 1. It is, thus, not an ISOP, a contradiction If GF is strongly connected, but is not minimal, there is at least one edge xj;xi that can be removed without affecting the connectedness of GF . It follows that GF0 , where F0 is F with xjxi removed, is a graph representation of F0, an SOP for the same function as F. Thus, F is an SOP, but not an ISOP, contradicting the assumption.tu Theorem 7.2. The number of MSOPs for STn; 1 is n 1!. Proof. From Theorem 7.1, an MSOP for STn; 1 corresponds to a directed graph with the fewest edges which is strongly connected, but is not strongly connected when any edge is removed. Such a graph is a directed cycle of arcs through all variables. As such, it represents a cyclic permutation on the variables. The number of such permutations is n 1!. tu Lemma 7.2. Let F be an ISOP of STn; 1. F is WSOP iff complementing all variables in F leaves F unchanged. Proof. (if) Let F be an ISOP that is unchanged by a complementation of all variables. This implies that if xixj is a PI of F, then so also is xjxi. It follows that if xi;xj is an edge in GF , then xj;xi is an edge in GF . That is, all edges between nodes occur in pairs, one going one way and the other going the other way. In such a graph, there are n 1 pairs or 2n 2 edges in all. (Replace each pair by an undirected edge. If the directed graph is strongly connected, the undirected graph must be connected. From Harary [14], there must be n 1 edges.) Since there are 2n 2 edges in GF , there are 2n 2 PIs in F and, thus, F is a WSOP. (only if) Let F be a WSOP of STn; 1. We show that GF consists of cycles of length 2 only. Thus, if xixj is a PI of F, so also is xjxi. It follows that complementing all variables of F leaves F unchanged. Suppose that GF contains a cycle of length m, where m>2. Such a cycle represents a strongly connected subgraph of GF in which there are m edges. However, the cycle can be replaced by a minimally strongly connected graph with more edges (e.g., where all edges occur in pairs). The result is a strongly connected graph, where the deletion of an edge leaves it unconnected, which has more edges than the original version. This contradicts the statement that F is a WSOP. tu Theorem 7.3. The number of WSOPs for STn; 1 is nn2. Proof. From Theorem 7.1, a WSOP for STn; 1 corresponds to a minimally strongly connected graph with the largest number of edges. We show that this graph consists of cycles of length 2 exclusively, as follows: Suppose, on the contrary, the graph has a cycle of length m>2. There are m edges in this cycle. However, this subgraph can be replaced by a subgraph with more edges, 2m 1.It follows that the original graph does not represent a WSOP. Each cycle of length 2 connects two nodes by edges in the two directions. Replace each pair of edges by an undirected edge, forming an undirected tree with n 1 edges. Thus, there are 2n 1 PIs in a WSOP for STn; 1. This proves Theorem 3.2. It follows that the number of WSOPs is the number of undirected trees on n labeled nodes. Cayley [5] in 1889 showed that this number is nn2. tu Theorem 7.4. The number of ISOPs for STn; 1 with n 1 PIs is Proof. From Theorem 7.1, an ISOP of STn; 1 that has n 1 PIs corresponds to a minimally strongly connected graph with edges. All graphs with this property have two cycles of nodes, of which i are common, where 2. We can represent each instance as a permutation of nodes that has been divided into three nonempty sets, the i common nodes, nodes N1, in one cycle only, and nodes N2 in the other cycle only. N1 and N2 must be nonempty since an empty set corresponds to a redundant edge. There are n! ways to permute the nodes and n1 ways to divide them into threenonempty sets. However, this double counts graphs since interchanging N1, and N2 does not change the graph. Thus, the total number of graphs with n 1 edges is 1 n1 n!. tu SASAO AND BUTLER: WORST AND BEST IRREDUNDANT SUM-OF-PRODUCTS EXPRESSIONS 947 ACKNOWLEDGMENTS This work was supported in part by a Grant in Aid for Scientific Research of the Ministry of Education, Culture, Sports, Science, and Technology of Japan and in part by NPS Direct Funded Grant UHCD1. Discussions with Dr. Shin-ichi Minato were quite useful. Mr. Munehiro Matsuura and Mr. Shigeyuki Fukuzawa did part of the experiments. The authors acknowledge the contributions of two referees. This paper is an extended version of T. Sasao and J.T. Butler, Comparison of the Worst and Best Sum-of- Products Expressions for Multiple-Valued Functions, Proceedings of the IEEE International Symposium on Multiple-Valued Logic, pp. 55-60, May 1997. --R Logic Minimization Algorithms for VLSI Synthesis. Logic Design of Digital Systems Introduction to Switching and Automata Theory. Graph Theory. Logic Design and Switching Theory. Representation of Discrete Functions Switching Theory for Logic Synthesis Advanced Logical Circuit Design Techniques. The TTL Data Book for Design Engineers --TR PALMINIMYAMPERSANDmdash;fast Boolean minimizer for personal computers Two-level logic minimization: an overview Switching Theory for Logic Synthesis Logic Design and Switching Theory Logic Design of Digital Systems Logic Minimization Algorithms for VLSI Synthesis A Remark on Minimal Polynomials of Boolean Functions A State-Machine SynthesizerMYAMPERSANDmdash;SMS An application of multiple-valued logic to a design of programmable logic arrays --CTR Alan Mishchenko , Tsutomu Sasao, Large-scale SOP minimization using decomposition and functional properties, Proceedings of the 40th conference on Design automation, June 02-06, 2003, Anaheim, CA, USA Olivier Coudert , Tsutomu Sasao, Two-level logic minimization, Logic Synthesis and Verification, Kluwer Academic Publishers, Norwell, MA, 2001

symmetric functions;worst sum-of-products expressions;prime implicants;logic minimization;multiple-output functions;minimum sum-of-products expressions;graph enumeration;irredundant sum-of-products;heuristic minimization;complete sum-of-products expressions;minimally strongly connected digraphs

507237

A logical foundation for deductive object-oriented databases.

Over the past decade, a large number of deductive object-oriented database languages have been proposed. The earliest of these languages had few object-oriented features, and more and more features have systematically been incorporated in successive languages. However, a language with a clean logical semantics that naturally accounts for all the key object-oriented features, is still missing from the literature. This article takes us another step towards solving this problem. Two features that are currently missing are the encapsulation of rule-based methods in classes, and nonmonotonic structural and behavioral inheritance with overriding, conflict resolution and blocking. This article introduces the syntax of a language with these features. The language is restricted in the sense that we have omitted other object-oriented and deductive features that are now well understood, in order to make our contribution clearer. It then defines a class of databases, called well-defined databases, that have an intuitive meaning and develops a direct logical semantics for this class of databases. The semantics is based on the well-founded semantics from logic programming. The work presented in this article establishes a firm logical foundation for deductive object-oriented databases.

Introduction The objective of deductive object-oriented databases is to combine the best of the deductive and object-oriented ap- proaches, namely to combine the logical foundation of the deductive approach with the modeling capabilities of the object-oriented approach. Based on the deductive object-oriented database language proposals as well as the work in object-oriented programming languages and data models, it is becoming clear that the key object-oriented features in deductive object-oriented databases include object identity, complex objects, typing, rule-based methods, encapsulation of methods, overloading, late binding, polymorphism, class hierarchy, multiple behavioral inheritance with overriding, blocking, and conflict handling. However, a clean logical semantics that naturally accounts for all the features is still missing from the literature. In particular the encapsulation of rule-based methods in classes, and non-monotonic multiple behavioral inheritance have not been addressed properly so far. In object-oriented programming languages and data models, methods are defined using functions or procedures and are encapsulated in class definitions. They are invoked through instances of the classes. In deductive object-oriented databases, we use rules instead of functions and procedures. By analogy, methods in deductive object-oriented databases should be defined using rules and encapsulated in class definitions. Such methods should be invoked through instances of the classes as well. However, most existing deductive object-oriented database languages, including F-logic [9], IQL [1], Datalog meth [2], ROL [12], do not allow rule-based methods to be encapsulated in the class definitions. The main difficulty is that the logical semantics is based on programs that are sets of rules. If rules are encapsulated into classes, then it is not clear how to define their semantics. Several proposals such as Datalog meth and Datalog ++ provide encapsulation but use rewriting-based semantics which do not address the issue directly. The authors of [3] address encapsulation but do not include other important object-oriented features, like inheritance. Non-monotonic multiple behavioral inheritance is a fundamental feature of object-oriented data models such as O 2 [5] and Orion [10]. The user can explicitly redefine (or override) the inherited attributes or methods and stop (or block) the inheritance of attributes or methods from super- classes. Ambiguities may arise when an attribute or method is defined in two or more superclasses, and the conflicts need to be handled (or resolved). Unfortunately, a logical semantics for multiple inheritance with overriding, blocking and conflict-handling has not been defined. The main difficulty is that the inherited instances of a superclass may not be well-typed with respect to its type definition because of overriding and blocking. Most deductive object-oriented database languages, including F-logic 1 , LOGRES [4], LIV- 1 F-logic however supports indeterminate non-monotonic default value ING IN LATTICE [7], COMPLEX [6], only allow monotonic multiple structural inheritance, which is not powerful enough. Some deductive object-oriented languages such as Datalog meth only support non-monotonic single inheritance by allowing method overriding. One extreme case is IQL, which does not support multiple inheritance at the class level at all. Instead, it indirectly supports it at the instance level via the union type so that inherited instances of a superclass can still be well-typed with respect to its type definition which is the union of the type for its direct instances and the type for its non-direct instances. ROL has a semantics that accounts for non-monotonic multiple structural inheritance with overriding and conflict-handling in a limited context, but without blocking. Datalog ++ takes a quite different approach towards non-monotonic inheri- tance. It disallows the inheritance of conflicting attributes and methods, like in C++. It provides mechanisms for the user to block the inheritance of attributes and methods. However, it only provides an indirect, rewriting-based semantics for such non-monotonic inheritance. This paper provides a direct well-defined declarative semantics for a deductive object-oriented database language with encapsulated rule-based methods and non-monotonic behavioral inheritance with overriding, conflict resolution and blocking. In order to keep the setting simple, we omit some well understood features that don't affect the semantics described, e.g. set-valued attribute values, and we focus on a static database rather than a dynamic database, (see [13] for the semantics of updates to the database). In the language, methods are declared in the class defini- tions, and the methods are invoked through instances of the classes. We introduce a special class, none, to indicate that the inheritance of an attribute or method in a subclass is blocked i.e. it won't be inherited from its superclasses. We provide a very flexible approach to conflict resolution. Our mechanism consists of two parts. One part, the default part is similar to the method used in Orion, namely a subclass inherits from the classes in the order they are declared in the class definition. The other part allows the explicit naming of the class the attribute or method is to be inherited from. Therefore, a subclass can inherit attribute or method definitions from any superclasses. We then define a class of databases, called well-defined databases, that have an intuitive meaning and develop a direct logical semantics for this class of databases. The semantics naturally accounts for method encapsulation, multiple behavioral inheritance, overriding, conflict handling and blocking, and is based on the well-founded semantics [16] from logic programming. We define a transformation that has a limit, I for well-defined databases, and prove that I , if it is defined, is a minimal model of the database. inheritance. The value inherited depends on which inheritance step is done first at run time. This paper is organized as follows. We introduce the syntax and semantics of the language using an example in Section 2. In Section 3 the class of well-defined databases and the semantics of well-defined databases are defined, and the main results are presented. Section 4 concludes the paper, reiterating our results and comparing this work with related work. Due to space limitation, the paper is quite terse and we have omitted proofs. They are included in [14]. Example Our language in fact supports many of the important object-oriented features in a rule-based framework with a well-defined declarative semantics. In this section, we introduce and demonstrate concepts that are important in the paper. A more extensive description of the syntax can be found in [14]. The schema in Figure 1(a) defines four classes, person, employee, student, and wstudent (working student). The class person has three attributes, name, birthyear, and spouse, and two methods: married(person) and single(). The attribute birthyear has a default value of 1945. Method married(X) is true if the person the method is applied to has a spouse, X, and method single() is true if the person is not married. Notice, that the semantics of negation are defined using extended negation as failure [11]. The class employee inherits all attribute declarations, default values and method declarations from class person unless they are blocked or overridden in class employee. We say that class employee is a direct subclass of person and person is a direct superclass of employee. New attributes can also be declared in subclasses. The attribute declarations for name, birthyear, and spouse, and the method declarations for married(person), and single() are inherited but the default value of birthyear is overridden in employee, i.e., the default value for attribute birthyear is redefined to 1960. The class student also inherits from person. Two methods are declared in student, namely extrasupport() and support(). The class wstudent inherits from two classes, employee and student. With multiple inheritance, there can be conflicting declarations i.e. default values, attributes and methods may be declared in more than one superclass. There is one possible conflict to be resolved in wstudent, default value birthyear is defined on both employee and student. There are two ways that conflicts can be resolved. A conflict resolution declaration indicates explicitly which class a property is to be inherited from e.g. birthyear \Delta student indicates that the definition of birthyear and the default value 1970 are inherited from student. If there is a conflict and there is no conflict resolution declaration then the property is inherited from the superclasses in the order the are listed in the class declaration. Notice that Key declaration value declaration ffl! default value class person [ name birthyear birthyear ffl! 1945; spouse single() fsingle() :- :married(X)g class employee isa person [ birthyear ffl! 1960; class student isa person [ birthyear ffl! 1970; student X; :student X; class wstudent isa employee; student [ birthyear \Delta student; (a) Schema employee tom [name ! "T om"; birthyear ! 1963; spouse student sam [name ! "Sam"] wstudent pam [name ! "P am"; spouse ! tom] (b) Instance Figure 1. Sample Database the method support() is blocked in wstudent (i.e. its return type is none), and the method extrasupport() in wstudent overrides the method extrasupport() in student. A method declaration in a subclass overrides a method declaration in a superclass if the methods have the same signature, independent of their return values. A method has the same signature as another method if the method has the same method label and the same arguments, e.g. extrasupport() in student has the same signature as extrasupport() in wstu- dent. While classes employee and student are direct superclasses of wstudent, person is an indirect superclass of wstudent. The instance in Figure 1(b) contains three objects with oids tom, sam, and pam. In the database instance, each object is associated with a class and attributes are assigned values. For example, object tom is a direct instance of em- ployee, and the value of its attribute name is "Tom". The value of attribute birthyear is 1963, i.e. the default 1960 in employee is not used. The value of its attribute spouse is object identifier pam. We say that employee is the primary class of object tom, and object tom is an non-direct instance of person. The birthyear of sam is 1970, i.e. the default in class student is used because a value for attribute birthyear is not provided in object sam. The value of attribute birthyear is not given in object pam, nor in class wstudent. The default value 1970 is inherited from student because there is a conflict resolution declaration in wstu- dent. We can ask the following queries on the sample database in Figure 1. The queries demonstrate how methods are encapsulated in classes, i.e. a method is declared in a class and invoked through instances of the class. 1. Find the birthyear of Sam. ?- student O[name ! "Sam"; birthyear ! X] The default value of birthyear for instances in class student is returned, 2. Find what support Sam gets. ?- student O[name ! "Sam"; support() ! X] The support() method in class student invokes the extrasupport() method. The extrasupport() rules in turn invoke the married(person) and single() methods defined in class person. As Sam has no spouse, Sam is not married, so Sam is single, and the third rule for extrasupport() is used. The extrasupport() that receives is 100, so 3. Find what support Pam gets. wstudent This method support() is blocked on wstudent, an error message indicating that this method is undefined is returned. 4. Find all students whose extra support is not 500. ?- student O [extrasupport() ! X]; X !? 500 This query returns the oids of all the objects that belong to class student or subclasses of student whose value for method extrasupport is not 500. The answer is samg. We make the following observations. Two kinds of classes are distinguished: value classes and oid classes. There are two special value classes, none and void. Class none is used to indicate that the inheritance of an attribute or method from a superclass is blocked in a subclass. Class void has only one value, namely nil, which is returned by a method if no other value is returned. Like in C++ and Java, we have a special variable, This, that is used to refer to the current object. Variables are represented throughout the paper using uppercase alphabetic characters. A schema K is a set of class declarations, which can be represented abstractly as a tuple where C is a finite set of oid classes, isa is a finite set of superclass declarations, ff is a finite set of attribute decla- rations, ffi is a finite set of default value declarations, - is a finite set of method declarations, and - is a finite set of conflict resolution declarations. For simplicity, we assume that there is no abbreviation in ff, ffi , and -. We write ff(c), (c) for the sets of attribute, default value, method, attribute conflict resolution declarations, and method conflict resolution declarations in ff, ffi , -, and - for the class c respectively. We impose constraints on the schema to capture the intended semantics of multiple inheritance with overriding, blocking and conflict handling. An instance I is a set of object declarations, that can be represented as a tuple I = (-) where - is a set of ground oid membership expressions called oid assignments and - is a set of ground positive attribute expressions called attribute value assignments. A database DB consists of two parts: the schema K and the instance I, which can be represented abstractly as a tuple Figure 1. A query is a sequence of expressions prefixed with ?-. In this section, we define the semantics of a database and queries. First we give the meaning of the schema and instance of the database, then we identify a class of databases, called well-defined databases, and finally, we define the meaning of the rule based methods of well-defined databases, based on the meaning of the schema and in- stance. The semantics of a database is based on the well-founded semantics except in this case the semantics of the rule-based methods must take into account the meaning of the schema and the instance of the database. 3.1 Semantics of Schema and Instance Encapsulation is dealt with in this subsection; each at- tribute, default value and method that are applicable to a class are identified. In order to determine which attributes, default values, and methods are applicable to a class, it is necessary to consider inheritance with overriding, blocking and conflict handling. Recall that ff(c), ffi (c) and -(c) are the sets of attribute declarations, default value and method declarations respectively that are defined on c. In this sec- tion, we define ff (c), ffi (c), and - (c), the attribute decla- rations, default value and method declarations that are applicable to class c, taking inheritance, overriding, conflict resolution and blocking into account. In [14], we define difference operators that find the attribute declarations (default value declarations, method declarations respectively) that are defined on one class and not redefined on another class. Consider the database in Figure 1. The difference between the sets of attribute declarations for person and student is: person [name ) string]; person [birthyear ) integer]; person [spouse ) person], The result is the attribute declarations in person that are not redefined in student. In Figure 1 the difference between the default attribute declarations for person and student is: This is not surprising because the default value for birthyear is redefined in student. The following definition outlines an algorithm to find the applicable declarations ff (c), ffi (c), - (c), which are the sets of declarations that are implicitly or explicitly declared on c with the blocked declarations re- moved, and the name of the class to which they apply changed. For example, consider the class wstudent in Figure 1. The algorithm produces: ff wstudent[name wstudent[birthyear wstudent[spouse wstudent[birthyear ffl! 1970], Overriding with Conflict Handling and Blocking The semantics of multiple inheritance with overriding, conflict handling and blocking are defined using the difference operators as follows: 1. If a class does not have any superclasses, then there is no inheritance, overriding, conflict resolution or block- ing. The declarations in the class are the only ones that apply to the class. That is, if there does not exist a class c 0 such that 2. if c isa c 1 ; :::; c n , then (a) we extend the sets of declarations to include new declarations due to explicit conflict resolution declarations and c 00 [l and 9 c 00 [l ffl! and :9 c[l ffl! (c) and M 2 - bci (b) we extend the sets of declarations to include declarations that are inherited from both direct and indirect superclasses using the difference opera- tor are defined analogously. (c) we remove blocked declarations and change the class names in the sets of declarations and c 0 6= noneg and :9 c 00 [l the type of M is is obtained from M by substituting c for c 0 g The symbols ff bc (c); ff bci (c), etc. are used only in the definition of applicable declarations, and are not referred to anywhere else in this paper. Let (C; isa; ff; ffi; -) be a database. Then ff Cg. If ff ; ffi , and - are defined, then the semantics of the schema by ff ; ffi , and - . We have dealt with non-monotonic inheritance within a database schema. We now describe the semantics of inheritance within an instance of a database, by introducing the notions of isa , - , and - . We overload the isa notion so that if c isa c 1 isa as the reflexive transitive closure of isa, that captures the general inheritance hierarchy. Note that c isa c. We say that is a non-direct instance of class c, (denoted by c c is a value class, o is a value, and o is an element in the collection which c denotes, or c is an oid class, o is an oid, and there exists a c 0 such that c 0 isa c and c 0 -. The notion of - captures the semantics of instance inheritance; that is, if an oid is a direct instance of a subclass c, then it is a non-direct instance of c and the superclasses of c. In the case, where there is a default value declaration for an attribute in a class, the instances of the class inherits the default value for the attribute. We extend the notion - to - to capture such intended semantics: -g. be a database. If - and - are defined, then the semantics of the instance I = (-) is given by - ; - . It is possible to define a database that has no intuitive meaning. For example it is possible to define a database schema with a cycle in its class hierarchy or an attribute in a class that has two distinct default values, or a database instance where an object is an instance of more than one class, or an attribute has more than one value for an ob- ject. In [14], we discuss a number of constraints that can be used to guarantee an intended semantics of the database and queries on the database, we give properties that demonstrate that the set of expressions defined have the intended semantics, and define a well-defined database. In the following subsection, we are concerned only with well-defined databases, that is databases with an intuitive meaning. A database instance does not have an intuitive meaning if an object is a direct instance of more than one class; or if an attribute has more than one value for an object. 3.2 Semantics of Databases and Queries In this paper, we focus on static databases rather than dynamic databases i.e. databases where classes of oids and their attribute values remain the same. The semantics for dynamic databases can be found in [13]. The classes of oids and their attributes form our extensional database (EDB) in the traditional deductive database sense. The methods, however, are represented intensionally by method rules. They define our intensional database (IDB). In this section, we define the semantics of methods based on the well-founded semantics proposed in [16]. Our definition differs from [16] in the following ways. We are concerned with a typed language with methods rather than an untyped language with predicates. We introduce a well-typed concept and take typing into account when deducing new facts from methods. The definition of satisfaction of expressions is simple in [16] and more complex in this paper because of the many kinds of expressions. Our definition reflects the fact that our model effectively has two parts, an extensional database (EDB) that models oid membership and attribute expressions, and an intensional database (IDB) that models method expressions. The EDB is a 2-valued model while the IDB is a 3-valued model. In the EDB, expressions are true if they're in the model otherwise they are false. In the IDB, method expressions are true if they are in the model, false if their complement belongs to the model, otherwise they are undefined. When a method expression is undefined, either the method isn't defined on the invoking object, or it isn't possible to assign a truth value to that expression. Every well-defined program has a total model, unlike in the well-founded semantics, where a program may have a partial model. In fact we prove that every well-defined program has a minimal model. We first define terminology that is needed later in this section. Herbrand Base Let be a well-defined database. The Herbrand base BDB based on DB is the set of all ground simple positive method expressions formed using the method names in DB (without ab- breviations). The definition for compatible sets of expressions can be found in [16]. Consider the set Because ;, the set is incompatible. Ground method expressions are required to be well-typed with respect to the appropriate class declarations. Let be a well-defined database, and method expression. Then / is well-typed in DB iff the following hold: 1. there exists a class c such that c 2. there exists a method in - (c) with the method type A set of ground method expressions is well-typed in DB iff each ground method expression is well-typed in DB. Methods can return values. However, for the same argu- ments, a method should return only one value. We formalize this using the notion of consistency. A set of ground method expressions are consistent iff there do not exist such that r . Interpretation Let be a well-defined database. A partial interpretation of DB is a is a compatible, consistent, and well-typed set of method expressions in DB, and each atom in S is an element of the Herbrand base. A total interpretation is a partial interpretation that contains every well-typed method atom of the Herbrand base or its negation. For an interpretation I = (-; S), - and - form an extensional database whereas S forms an intensional database. Note that S contains both positive and negative expres- sions, and different interpretations of DB have the same extensional database but different intensional databases. A ground substitution ' is a mapping from variables to oids and values. It is extended to terms and expressions in the usual way. Satisfaction Let be a well-defined database and I = (-; S) an interpretation of DB. The notion of satisfaction of expressions, denoted by j=, and its negation, denoted by 6j=, are defined as follows. 1. The satisfaction of ground positive and negative oid membership expressions, ground positive and negative attribute expressions, and ground arithmetic comparison expressions are defined in the usual way. 2. For a ground positive method expression , I 3. For a ground negative method expression : , I 4. For a ground composite Vn ], ffl I ffl I 6j= iff I 6j= c o or I 6j= o:V i for some i with For a ground composite Vn ], ffl I ffl I 6j= iff I 6j= o:V i for some i with 5. For a method rule each ground substitution ', ffl I j= 'A; or ffl I 6j= 'A and for each ground method rule with head 'A there exists an L i with that I ffl there exists an L i with 1 - i - n such that neither I In other words, I j= means that is true in I; I 6j= means that is false in I; if neither I then is unknown in I. Model Let be a well-defined database and I = (-; S) an interpretation of DB. Then I is a model of DB if I satisfies every ground method rule in - . Consider the following database: class person [ spouse single()fsingle() :- :married()g person sam[spouse ! pam] person pam The following set is a model of this database: Due to the typing and compatibility constraints as in ROL [12], it is possible that a database has no models. Also, a well-defined database may have several models. Our intention is to select a proper minimal model as the intended semantics of the database. An unfounded set for a database with respect to an interpretation provides a basis for false method expressions in our semantics. The greatest unfounded set (GUS) is the set of all the expressions that are false in a database with respect to an interpretation and is used to provide the negative expressions when finding the model of a database. The definition for unfounded sets and greatest unfounded sets can be found in [16]. The greatest unfounded set is used in the definition of a model, i.e. a limit of the following transformation Transformation Let be a well-defined database. The transformation TDB of DB is a mapping from interpretation to interpretation defined as follows. (-; W(I)) if W(I) is well-typed and compatible undefined otherwise where is a method rule in DB and there exists a ground substitution ' such that I G is the GUS of DB with respect to I. Model For all countable ordinals h the tuple I h for database the limit of the transformation TDB is defined recursively by: 1. For limit ordinal h, I 2. For successor ordinal k Note that 0 is a limit ordinal, and I sequence reaches a limit I . We now prove that I is a model. Theorem 3.1 Let DB be a well-defined database. If I (-; S) is defined, then it is a model of DB. 2 Minimal model Let be a model of a database DB. We say that model M is minimal if there does not exist an expression / in S such that (-; S \Gamma /) is still a model. We now prove that for a well-defined database DB, I is a minimal model of DB if it is defined. Theorem 3.2 Let DB be a well-defined database. If I (-; S) is defined, then it is a minimal model of DB. 2 Semantics of Databases The semantics of a well-defined database represented by the limit I if it is defined. Semantics of Queries Let be a well-defined database, Q a query of the form substitution for variables of Q. Assume I is defined. Then the answer to Q based on DB is one of the following: 1. true if I 2. false if there exists an L i with 1 - i - n such that I 6j= 'L i , and 3. unknown otherwise. In other words, for a method expression /, if I then expression / is true, if I is false, and if I 6j= / and I 6j= :/ then expression / is undefined. Let us consider an example with unknown answers. Consider the following database: class person [ spouse single()fsingle() :- :married()g person sam[spouse ! pam] person pam Then I pam]g; ;) is a three-valued model, in which the answers to the following queries are unknown. There are two reasons why I may be undefined. One is that the inferred set of method expressions is not well-typed. The other is that it is not consistent. For the first problem, we could define another constraint on method rules using type substitution as in [13] to constrain the database. For the second problem, run-time checking is necessary. Logical semantics have played an important role in database research. However, the object-oriented approach to databases was dominated by "grass-roots" activity where several systems were built without the accompanying theoretical progress. As a result, many researchers feel the area of object-oriented databases is misguided [9]. The deductive object-oriented database research, however, has taken quite a different approach. It has logical semantics as its main objective and started with a small set of simple features taken from the object-oriented paradigm such as F-logic [9], and gradually incorporates more and more difficult features that can be given a logical semantics such as ROL [12] and Datalog++ [8]. The main contribution of the paper is the addition of two outstanding object-oriented features to deductive object-oriented databases together with a direct logical semantics. The two outstanding features were rule-based methods and the encapsulation of these methods in classes, and multiple behavioral inheritance, with overriding, blocking, and conflict handling. We have shown that these object-oriented features which are believed to be difficult to address, can indeed be captured logically. We believe that the semantics given in this paper have a far reaching influence on the design of deductive object-oriented languages and even object-oriented languages in general. The language and semantics defined on the language form the theoretical basis for a practical query language. Indeed, the practical deductive object-oriented database language ROL2 [15] supports the theory discussed here. Our work differs from the work of others in many ways. Most existing deductive object-oriented database languages do not allow rule-based methods to be encapsulated in the class definitions. Those that do, do not address the issue directly. Also, most existing deductive object-oriented database languages do not allow non-monotonic multiple behavioral inheritance. ROL does, but deals with conflict handling only in a limited context and doesn't have block- ing. Datalog ++ provides blocking and disallows the inheritance of conflicting properties. F-logic supports monotonic structural inheritance and indeterminate non-monotonic default value inheritance by allowing a database to have multiple possible models. For a class, not only its subclasses but also its elements can inherit its properties. --R Object as a Query Language Primitive. Methods and Rules. A Logic for Encapsulation in Object Oriented Languages. COMPLEX: An Object-Oriented Logic Programming System The LIVING IN A LATTICE Rule Language. Implementing Abstract Objects with Inheritance in Datalog neg Logical Foundations of Object-Oriented and Frame-Based Languages Introduction to Object-Oriented Databases ROL: A Deductive Object Base Language. Incorporating Methods and Encapsulation into Deductive Object-Oriented Database Languages A Logic for Deductive Object-Oriented Databases A Real Deductive Object-Oriented Database Language The Well-Founded Semantics for General Logic Programs --TR C-logic of complex objects Integrating object-oriented data modelling with a rule-based programming paradigm LLO The well-founded semantics for general logic programs The O<subscrpt>2</subscrpt> system The ObjectStore database system The GemStone object database management system Introduction to object-oriented databases An overview of three commercial object-oriented database management systems The LIVING IN A LATTICE rule language Methods and rules A logic for programming with complex objects Design and implementation of ROCK MYAMPERSANDamp; ROLL On the declarative and procedural semantics of deductive object-oriented systems Logical foundations of object-oriented and frame-based languages ROL: a deductive object base language Object identity as a query language primitive A Formal Definition of the Chimera Object-Oriented Data Model The Story of O2 COMPLEX Implementing Abstract Objects with Inheritance in Datalogneg A Logic for Encapsulation in Object Oriented Languages Object Migration in ISA Hierarchies ROL2 Incorporating Methods and Encapsulation into Deductive Object-Oriented Database Languages Overview of the ROL2 Deductive Object-Oriented Database System

deductive databases;nonmonotonic multiple inheritance;declarative semantics;object-oriented databases;rule-based languages

507257

The regular viewpoint on PA-processes.

PA is the process algebra allowing non-determinism, sequential and parallel compositions, and recursion. We suggest viewing PA-processes astrees, and usingtree-automata techniques for verification problems on PA. Our main result is that the set of iterated predecessors of a regular set of PA-processes is a regular tree language, and similarly for iterated successors. Furthermore, the corresponding tree automata can be built effectively in polynomial time. This has many immediate applications to verification problems for PA-processes, among which a simple and general model-checking algorithm.

Introduction Veri-cation of In-nite State Processes is a very active -eld of research today in the concurrency-theory community. Of course, there has always been an active Petri-nets com- munity, but researchers involved in process algebra and model-checking really became interested into in-nite state processes after the proof that bisimulation was decidable for normed BPA-processes [BBK87]. This prompted several researchers to investigate decidability issues for BPP and BPA (with or without the normedness hypothesis) (see [CHM94, Mol96, BE97] for a partial survey). From BPA and BPP to PA: BPA is the inon-determinism recursionj fragment of process algebra. BPP is the inon-determinism parallel composition recursionj fragment. PA (from [BEH95]) combines both and is much less tractable. A few years ago, while more and more decidability results for BPP and BPA were presented, PA was still beyond the reach of the current techniques. Then Mayr showed the decidability of reachability for PA processes [May97c], and extended this into decidability of model-checking for PA w.r.t. the EF fragment of CTL [May97b]. This was an important breakthrough, allowing Mayr to successfully attack more powerful process algebras [May97a] while other decidability results for PA were presented by him and other researchers (e.g. [Ku#96, Ku#97, JKM98]). A -eld asking for new insights: The decidability proofs from [May97b] (and the following papers) are certainly not trivial. The constructions are quite complex and hard to check. It is not easy to see in which directions the results and/or the proofs could be adapted or generalized without too much trouble. Probably, this complexity cannot be avoided with the techniques currently available in the -eld. We believe we are at a point where it is more important to look for new insights, concepts and techniques that will simplify the -eld, rather than trying to further extend already existing results. Our contribution: In this paper, we show how tree-automata techniques greatly help dealing with PA. Our main results are two Regularity Theorems, stating that Post (L) (resp. Pre (L)) the set of con-gurations reachable from (resp. allowing to reach) a conguration in L is a regular tree language when L is, and giving simple polynomial-time constructions for the associated automata. Many important consequences follow directly, including a simple algorithm for model-checking PA-processes. Why does it work ? The regularity of Post (L) and Pre (L) could only be obtained after we had the combination of two main insights: 1. the tree-automata techniques that have been proved very powerful in several -elds (see [CKSV97]) are useful for the process-algebraic community as well. After all, PA is just a simple term-rewrite system with a special context-sensitive rewriting strategy, not unlike head-rewriting, in presence of the sequential composition operator. RR n-0123456789 2. the syntactic congruences used to simplify notations in simple process algebras help one get closer to the intended semantics of processes, but they break the regularity of the behavior. The decidability results are much simpler when one only introduces syntactic congruences at a later stage. (Besides, this is a more general approach.) Plan of the paper: We start by recalling the basics of tree-automata theory (# 1) before we introduce our de-nition for the PA process algebra (# 2). After we explain how sets of PA processes can be seen as tree languages (# 3) we give a simple proof showing how Post (t) and Pre (t) are regular tree languages and start listing applications to veri-cation problems. We then move on to Post (L) and Pre (L) for L a regular language (# 5). These are our main technical results and we devote # 6 to the important applications in model-checking. We end up with an extension to reachability and model-checking under constraints (# 7) and some simple but important techniques allowing to deal with PA processes modulo structural equivalence (# 8). Related work: Several recent works in the -eld use tree-automata to describe the behaviour of systems. We use them to describe set of con-gurations. The set of all reachable con-gurations of a pushdown automaton form a regular (word) language. This was proven in [B#c64] and extended in [Cau92]. Applications to the model-checking of pushdown automata have been proposed in [FWW97, BEM97]. The decidability of the -rst-order theory of the rewrite relation induced by a ground term rewrite system relies on ground tree transducers [DT90] (note that PA is de-ned by a conditional ground rewrite system). Among the applications we develop for our regularity theorems, several have been suggested by Mayr's work on PA [May97c, May97b] and/or our earlier work on RPPS [KS97a, KS97b]. Regular tree languages and tree automata We recall some basic facts on tree automata and regular tree languages. For more details, the reader is referred to any classical source (e.g. [CDG A ranked alphabet is a -nite set of symbols F together with an arity function This partitions F according to arities: the set of terms over F and call them -nite trees or just trees. A tree language over F is any subset of T (F). A (-nite, bottom-up) tree automaton A is a tuple hF ; Q; F; Ri where F is a ranked alphabet, is a -nite set of states, F ' Q is the subset of -nal states, and R is a -nite set of transition rules of the form f(q is the arity j(f) of symbol f 2 F . Tree automata with "-rules also allow some transition rules of the form The transition rules de-ne a rewrite relation on terms built on F [Q (seeing states from Q as nullary symbols). This works bottom-up. At -rst the nullary symbols at the leaves INRIA The Regular Viewpoint on PA-Processes 5 are replaced by states from Q, and then the quasi-leaf symbols immediately on top of leaves from Q are replaced by states from Q. We write t A can be rewritten (in some number of steps) to q 2 Q and say t is accepted by A if it can be rewritten into a -nal state of A. We write L(A) for the set of all terms accepted by A. Any tree language which coincide with L(A) for some A is a regular tree language. Regular tree languages are closed under complementation, union, etc. An example: Let F be given by F ffg. There is an automaton accepting the set of all occurs an even number of times in t. A is given by Q g. Let t be g(f(g(a); b)). A rewrites t as follows: g(f(g(a); b)) 7\Gamma! g(f(g(q 0 Hence t 7\Gamma! q 0 2 F so that t 2 L(A). If we replace R by we have an automaton accepting the set of all t where there is an even number of g's along every path from the root to a leaf. The size of a tree automaton, denoted by jAj, is the number of states of A augmented by the size of the rules of A where a rule f(q 2. In this paper, we shall never be more precise than counting jQj, the number of states of our automata. Notice that, for a -xed F where the largest arity is m - 2, jAj is in O(jQj m ). A tree automaton is deterministic if all transition rules have distinct left-hand sides (and there are no "-rule). Otherwise it is non-deterministic. Given a non-deterministic tree automaton, one can use the classical isubset constructionj and build a deterministic tree automaton accepting the same language, but this construction involves a potential exponential blow-up in size. Telling whether L(A) is empty for A a (non-necessarily deterministic) tree automaton can be done in time O(jAj). Telling whether a given tree t is accepted by a given (non-necessarily deterministic) A can be done in time polynomial in jAj A tree automaton is completely speci-ed (also complete) if for each f 2 Fn and q Q, there is a rule f(q adding a sink state and the obvious rules, any A can be extended into a complete one accepting the same language. 2 The PA process algebra For our presentation of PA, we explicitly refrain from writing terms modulo some simplication laws (e.g. the neutral laws for 0). Hence our use of the IsNil predicate (see below), inspired by [Chr93]. This viewpoint is in agreement with the earliest works on (general) process algebras like CCS, ACP, etc. It is a key condition for the results of the next section, and it clearly does not prevent considering terms modulo some structural congruence at a later stage, as we demonstrate in section 8. RR n-0123456789 6 D.Lugiez, Ph.Schnoebelen 2.1 Syntax is a set of action names. is a set of process variables. is the set of PA-terms, given by the following abstract syntax t; Given t 2 EPA , we write Var(t) the set of process variables occurring in t and Subterms(t) the set of all subterms of t (includes t). A guarded PA declaration is a -nite set a i ng of process rewrite rules. Note that the X i 's need not be distinct. We write Var (\Delta) for the set of process variables occurring in \Delta, and Subterms (\Delta) the union of all Subterms(t) for t a right- or a left-hand side of a rule in \Delta. there is a rule iX a ! tj in \Deltag and \Delta(X) is S a2Act \Delta a (X). Var ? is the set of variables for which \Delta provides no rewrite. In the following, we assume a -xed Var and \Delta. 2.2 Semantics A PA declaration \Delta de-nes a labeled transition relation ! \Delta ' EPA \Theta Act \Theta EPA . We always omit the \Delta subscript when no confusion is possible, and use the standard notations and abbreviations: t w inductively de-ned via the following SOS rules: a a a a (X a a a a a where the predicate is inductively de-ned by ae true if false otherwise. The IsNil predicate is a syntactic test for termination, and indeed Lemma 2.1. The following three properties are equivalent: 1. 2. t 6! (i.e. t is terminated), 3. INRIA The Regular Viewpoint on PA-Processes 7 Proof. (3 This derivation used has some X i a i induction over t to prove that is obvious from the de-nition. as a tree language We shall use tree automata to recognize sets of terms from EPA . This is possible because EPA is just a T (F) for F given by F kg. Of course, we shall keep using the usual in-x notation for terms built with i:j or ikj. We begin with one of the simplest languages in EPA : Proposition 3.1. For any t, the singleton tree language ftg is regular, and an automaton for ftg needs only have jtj states. Similarly, an immediate consequence of Lemma 2.1 is Corollary 3.2. L ? , the set of terminated processes, is a regular tree language, and an automaton for L ? needs only have one state. 4 Regularity of the reachability set For Post (t) def denote the set of iterated predecessors (resp. the set of iterated successors, also called the reachability set) of t. These notions do not take into account the sequences w 2 Act of action names allowing to move from some t to some t 0 in Post (t). Indeed, we will forget about action names until section 7 which is devoted to Pre [C](t) and Post [C](t) for C ' Act . Given two tree languages L; L 0 ' EPA , we let 4.1 Regularity of Pre We de-ne (L t ) t2EPA , an in-nite family of tree languages, as the least solution of the following set of recursive equations. The intuition is that these are quasi-regular equations satis-ed RR n-0123456789 by Pre (t). Y a Y a Y a Y a Y a (1) Observe that all equations de-ne L t as containing all LY 's for Y a process variable allowing a one step transition Y a t. Lemma 4.1. For any t 2 EPA , L Proof. (Sketch) The proof that u is an induction over the length of the transition sequence from u to t, then a case analysis of which SOS rule gave the last transition, and then an induction over the structure of t. The proof that u 2 L t implies u t is a -xpoint induction, followed by a case analysis over which summand of which equation is used, and relies on simple lemmas about reachability, such as it 1 The equations from (1) can easily be transformed into regular equations, just by introducing new variables for sets ftg in the de-nitions for the L t:t 0 's. Now, because any given L t only depends on a -nite number of L u 's and fug's, namely only for u's in Subterms(t) [ Subterms (\Delta), we have 1 Corollary 4.2. For any t 2 EPA , the set L t is a regular tree language. and the corresponding tree automaton has O(j\Deltaj states. This entails Theorem 4.3. For any t 2 EPA , Pre (t), Pre(t) and Pre + (t) are regular tree languages. 4.2 Regularity of Post We de-ne (L 0 two in-nite families of tree languages, as the least solution of the following set of recursive equations. Our aim is that L 0 should coincide with In section 5.1, we shall see that Corollary 4.2 holds even when \Delta is in-nite (but Var (\Delta) must be -nite). INRIA The Regular Viewpoint on PA-Processes 9 Post (t) (resp. Post X a X a (2) Again, these can easily be turned into regular equations. Again, any given L 0 t or L 00 only depends on a -nite number of L 0 u 's and fug's. Corollary 4.4. For any t 2 EPA , the sets L 0 t and L 00 are regular tree languages. and the corresponding tree automata have O(j\Deltaj states. As with Pre (t), we can easily show Lemma 4.5. For any t 2 EPA , L 0 Post (t) and L 00 hence the corollary Theorem 4.6. For any t 2 EPA , Post (t), Post(t) and Post + (t) are regular tree languages that can be constructed eoeectively. Theorems 4.3 and 4.6 will be generalized in sections 5 and 7. However, we found it enlightening to give simple proofs of the simplest variants of our regularity results. Already, Theorems 4.3 and 4.6 and the eoeective constructibility of the associated automata have many applications. 4.3 Some applications Theorem 4.7. The reachability problem iis t reachable from t 0 ?j is in P. Proof. Combine the cost of membership testing for non-deterministic tree automata and the regularity of Pre (t 0 ) or the regularity of Post (t). For a dioeerent presentation of PA and ! \Delta , [May97c] shows that the reachability problem is NP-complete. In section 8, we describe how to get his result as a byproduct of our approach. Many other problems are solved by simple application of Theorems 4.3 and 4.6: boundedness. Is Post (t) in-nite ? RR n-0123456789 covering. (a.k.a. control-state reachability). Can we reach a t 0 in which Y (resp. do not occur). inclusion. Are all states reachable from t 1 also reachable from t 2 ? Same question modulo a regularity preserving operation (e.g. projection). liveness. where a given \Delta live if, in all reachable states, at least one transition from \Delta 0 can be -red. 5 Regularity of Post (L) and Pre (L) for a regular language In this section we prove the regularity of Pre (L) and Post (L) for a regular language L. For notational simplicity, given two states q; q 0 of an automaton A, we denote by q k q 0 state q 00 such that q k q 0 A possibly using "-rules. 5.1 Regularity of Pre Ingredients for A Pre : Assume AL is an automaton recognizing L ' EPA . A Pre is a new automaton combining several ingredients: ffl A? is a completely speci-ed automaton accepting terminated processes (see Corollary 3.2). ffl AL is the automaton accepting L. ffl We also use a boolean to record whether some rewriting steps have been done. States of A Pre : A state of A Pre is a 3-tuple (q ? where Q ::: denotes the set of states of the relevant automaton. Transition rules of A Pre : The transition rules of A Pre are de-ned as follows: type 0: all rules of the form 0 7\Gamma! (q 7\Gamma! q L . type 1a: all rules of the form X 7\Gamma! (q ? ; q L ; true) s.t. there exists some u 7\Gamma! q L . type 1b: all rules of the form X 7\Gamma! (q 7\Gamma! q L . type 2: all rules of the form (q type 3a: all rules of the form (q state of A? . INRIA The Regular Viewpoint on PA-Processes 11 type 3b: all rules of the form (q Lemma 5.1. For any t 2 EPA , t A Pre there is some u 2 EPA and some p 2 N such that t p u, u A? Proof. By structural induction over t. There are three cases: 1. Because A Pre has no "-rules, we only have to observe that its rules of type 0, 1a and 1b exactly correspond to what the lemma requires. 2. required that, for A Pre there is a type 3 rule (q 1 The induction hypothesis entails there are t 1 corresponding to the rewrite of t 1 and t 2 by A Pre . Now if A Pre used a type 3b rule, then b . If we used a type 3a rule, then q 1 is a -nal state, therefore u 1 2 L? is a terminated process, hence t 1 :t 2 and Conversely, assume 7\Gamma! q L . Then u is some for In the -rst case the ind. hyp. entails t 1 A Pre AL A Pre Now we can use a type 3b rule to show t A Pre with u AL L . In the second case, AL ? a -nal state of A? . We can use a type 3a rule to show t A Pre 3. This case is similar to the previous one (actually it is simpler). If we now let the -nal states of A Pre be all states (q q L is a -nal state of AL , then t ! u for some u accepted by AL ioe A Pre accepts t (this is where we use the assumption that A? is completely speci-ed.) Theorem 5.2. (Regularity) (1) If L is a regular subset of EPA , then Pre (L) is regular. (2) Furthermore, from an automaton AL recognizing L, is it possible to construct (in polynomial time) an automaton A Pre recognizing Pre (L). If AL has k states, then A Pre needs only have 4k states. RR n-0123456789 Proof. (1) is an immediate consequence of Lemma 5.1. Observe that the result does not need the -niteness of \Delta (but Var (\Delta) must be -nite). (2) Building A Pre eoeectively requires an eoeective way of listing the type 1a rules. This can be done by computing a product of AX , an automaton for Post + (X), with A? and AL . Then there exists some u 7\Gamma! q L ioe the the language accepted by the -nal states f(q a -nal state of AX g is not-empty. This gives us the pairs q ? ; q L we need for type 1a rules. Observe that we need the -niteness of \Delta to build the AX 's. 5.2 Regularity of Post Ingredients for A Post : Assume AL is an automaton recognizing L ' EPA . A Post is a new automaton combining several ingredients: Automata A? and AL as in the previous construction, but this time we need to assume each of them is a completely speci-ed automata. A \Delta is a completely speci-ed automaton recognizing the subterms of \Delta. It has all states q s for s 2 Subterms (\Delta). We ensure it A \Delta by taking as transition rules belongs to Subterms (\Delta). In addition, the automaton has a sink state q ? and the obvious transitions so that it is a completely speci-ed automaton. ffl Again, we use a boolean b to record whether rewrite steps have occurred. States of A Post : The states of A Post are 4-uples (q ? Transition rules of A Post : The transition rules are: type 0: all rules of the form 0 7\Gamma! (q type 1: all rules of the form X 7\Gamma! (q 7\Gamma! q L , and type 2: all "-rules of the form (q s is a rule in \Delta with X AL 7\Gamma! q L . type 3: all rules of the form type 4a: all rules of the form INRIA The Regular Viewpoint on PA-Processes 13 type 4b: all rules of the form is a -nal state of A? . Lemma 5.3. For any t 2 EPA , t A Post there is some u 2 EPA and some such that u p t, u AL Proof. We -rst prove the ()) direction by induction over the length k of the rewrite t A Post We distinguish four cases: 1. and we used a type 0 or type 1 rule. Taking sati-es the requirements. 2. k ? 1 and the last rewrite step used a type 2 "-rule: Then the rewrite has the form z - true). By ind. hyp., there is a u 0 and a p 0 s.t. t. Now u The existence of the type 2 rule entails t. Taking the requirements. 3. k ? 1 and the last rewrite step used a type 4 rule: Then t is some t 1 :t 2 and the type 4 rule applied on top of two rewrite sequences t i 7\Gamma! (q i 2. The ind. hyp. gives us, for If the last rule was a type 4a rule, then b t. Taking the requirements. Otherwise the the last rule was a type 4b rule. Then q 1 ? is a -nal state and t 1 A? entails that t 1 is a terminated process. Hence t. Again, taking the requirements. 4. k ? 1 and the last rewrite step used a type 3 rule: This case is similar (actually simpler) to the previous one. For the (() direction, we assume u p t with the accompanying conditions (a.c.), and proceed by induction over the length of the transition sequence (i.e. over p), followed by structural induction over u. There are -ve cases: 1. and the a.c.'s ensure we can use a type 0 rule to show t A Post 2. 0: Like the previous case but with a type 1 rule. 3. 0: Then the sequence has the form X 1 t. Here the a.c.'s read Subterms (\Delta). If we now take a q 0 RR n-0123456789 14 D.Lugiez, Ph.Schnoebelen s.t. L (one such q 0 must exist) and let b 0 be false ioe the ind. hyp. gives us t A Post there must be a type 2 "-rule (q We use it to show t A Post 4. t is a combination of some u 1 with Additionally, if For 2, The rewrites t A? 7\Gamma! q L and u A \Delta used some t i A? AL L and u i A \Delta If we now de-ne b i according to p i , the ind. hyp. entails that, for A Post There are two cases. If t 1 2 L? then q 1 ? is a -nal state of A? and A Post has a type 4b rule (q 1 that we can use. If t 1 62 L? , then There is a type 4a rule that we can use. 5. Similar to the previous case (actually it is simpler). If we now let the -nal states of A Post be all states (q q L is a -nal state of AL , then A Post accepts a term t ioe u t for a u accepted by AL ioe t belongs to Post (L). Theorem 5.4. (Regularity) (1) If L is a regular subset of EPA , then Post (L) is regular. (2) Furthermore, from an automaton AL recognizing L, is it possible to construct (in polynomial time) an automaton A Post recognizing Post (L). If AL has k states, then A Pre needs only have O(k:j\Deltaj) states. Proof. Obvious from the previous construction. Our results relate t and Pre (t) (resp. Post (t)). A natural question is to ask if the relation i !j (i.e. f(t; u) j t ug is recognizable in some sense. The most relevant notion of recognizability related to our problem is linked to ground tree transducers, GTT's for short, details. Since it can be shown that the relation induced by a ground rewrite system is recognizable by a GTT, we tried to extend this result to our PA case where the rules are ground rewrite rules with simple left hand sides, but where there is a notion of pre-x rewriting. Unfortunately, this pre-x rewriting entails that our ! is not stable under contexts and the natural extensions of GTT that could handle such conditional rules are immediately able to recognize any recursively enumerable relation. 6 Model-checking PA processes In this section we show a simple approach to the model-checking problem solved in [May97b]. We see this as one more immediate application of our main regularity theorems. INRIA The Regular Viewpoint on PA-Processes 15 We consider a set P of atomic propositions. For Mod (P ) denotes the set of PA processes for which P holds. We only consider propositions P such that Mod (P ) is a regular tree-language. Thus P could be it can make an a-labeled step right nowj, ithere is at least two occurences of X inside tj, ithere is exactly one occurence of X in a non-frozen positionj, The logic EF has the following syntax: and semantics Thus EX' reads iit is possible to reach in one step a state s.t. 'j and EF' reads iit is possible to reach (via some sequence of steps) a state s.t. 'j. De-nition 6.1. The model-checking problem for EF over PA has as inputs: a given \Delta, a given t in EPA , a given ' in EF. The answer is yes ioe t If we now de-ne Mod(') def Mod Mod Theorem 6.2. (1) For any EF formula ', Mod (') is a regular tree language. (2) If we are given tree-automata AP 's recognizing the regular sets Mod (P ), then a tree- automaton A' recognizing Mod(') can be built eoeectively. Proof. A corollary of (3) and the regularity theorems. This gives us a decision procedure for the model-checking problem: build an automaton for Mod (') and check whether it accepts t. We can estimate the complexity of this approach in term of j'j and n alt ('). We de-ne n alt (') the number of alternation of negations and temporal connectives in ' as RR n-0123456789 Theorem 6.3. (Model-checking) An automaton for Mod (') can be computed in time2 j'jj\Deltaj2 O(j'jj\Deltaj) Proof. We assume all automata for the Mod (P )'s have size bounded by M (a constant). We construct an automaton for Mod (') by applying the usual automata-theoretic constructions for intersection, union, complementation of regular tree languages, and by invoking our regularity theorems for Pre and Pre . All constructions are polynomial except for complementation. With only polynomial constructions, we would have a 2 O(j'j) size for the resulting automaton. The negations involving complementation are the cause of the non-elementary blowup. Negations can be pushed inward except that they cannot cross the temporal connectives EF and EX. Here we have one exponential blowup for determinization at each level of alternation. This is repeated n alt (') times, yielding the given bound on the number of states hence the overall complexity. The procedure described in [May97b] is non-elementary and today the known lower bound is PSPACE-hard. Observe that computing a representation of Mod(') is more general than just telling whether a given t belongs to it. Observe also that our results allow model-checking approches based on combinations of forward and backward methods (while Theorem 6.2 only relied on the standard backward approach.) 7 Reachability under constraints In this section, we consider reachability under constraints. Let C ' Act be a (word) language over action names. We write t C that t 0 can be reached from t under the constraint C. We extend our notations and write Pre [C](L), Post with the obvious meaning. Observe that, even if we assume C is regular, the problem of telling whether t C !, i.e. whether Post [C](t) is not empty, is undecidable for the PA algebra. This can be proved by a reduction from the intersection problem for context-free languages as follows: Let \Sigma be an alphabet and # some distinguished symbol. We use two copies a; a of every letter a in \Sigma [ f#g. Context-free languages can be de-ned in BPA (PA without k), that is, for any context-free language L 1 (resp. L 2 ) on \Sigma, we can de-ne PA rules such that X 1 w:# w:# These rules don't overlap. We now introduce the regular an holds ioe is undecidable. INRIA The Regular Viewpoint on PA-Processes 17 In this section we give suOEcient conditions over C so that the problem becomes decidable (and so that we can compute the C-constrained Pre of a regular tree language). Recall that the shu-e w k w 0 of two -nite words is the set of all words one can obtain by interleaving w and w 0 in an arbitary way. De-nition 7.1. ffl m )g is a (-nite) seq-decomposition of C ioe for all w; w 0 2 Act we have m )g is a (-nite) paral-decomposition of C ioe for all w; w we have The crucial point of the de-nition is that a seq-decomposition of C must apply to all possible ways of splitting any word in C. It even applies to a decomposition w:w 0 with so that one of the C i 's (and one of the C 0 contains ''. Observe that the formal dioeerence between seq-decomposition and paral-decomposition comes from the fact that w k w 0 , the set of all shu-es of w and w 0 usually contains several elements. De-nition 7.2. A family Cng of languages over Act is a -nite decomposition system ioe every C 2 C admits a seq-decomposition and a paral-decomposition only using C i 's from C . Not all C ' Act admit -nite decompositions, even in the regular case. Consider (ab) and assume is a -nite paral-decomposition. Then for every k, there is a shu-e of a k and b k in C. Hence there must be a . Now if then there there must exist a shu-e w 00 of a k and b k 0 with w 00 2 C. This is only possible . Hence all i k 's are distinct, contradicting -niteness. A simple example of a -nite decomposition system is i.e. the set of all singleton languages with words shorter than k. Here the paral-decomposition of fwg is f(fw 1 g)g where the w i 's are all subwords 2 of w (and w 0 i is the corresponding remainder). This example shows that decomposability is not composability: not all pairs from C appears in the decomposition of some member of C. More generally, for any linear weight function ' of the form '(w) def with N, the sets C ('=k) belong to -nite decomposition system. Assume C is a -nite decomposition system. We shall show Theorem 7.3. (Regularity) For any regular L ' EPA and any C 2 C , Pre [C](L) and Post [C](L) are regular tree languages. A subword of w is any w 0 obtained by erasing letters from w at any position. RR n-0123456789 Ingredients for A Post [C] : We build A Post [C] in the same way as A Post but states contain a new C 2 C component. States of A Post [C] : The states of A Post [C] are 5-uples (q ? Transition rules of A Post [C] : The transition rules are: type 0: all rules of the form 0 7\Gamma! (q type 1: all rules of the form X 7\Gamma! (q type 2: all "-rules of the form (q s is a rule in \Delta with X AL appears in the seq-decomposition of C. type 3: all rules of the form appears in the paral-decomposition of C 00 . type 4a: all rules of the form type 4b: all rules of the form state of A? s.t. (C; C 0 ) appears in the seq-decomposition of C 00 . Lemma 7.4. For any t 2 EPA , t A Post [C] there is some u 2 EPA and some w 2 C such that u w t, u AL Proof. A Post [C] is A Post equipped with a new component and the proof follows exactly the lines of the proof of Lemma 5.3. We refer to this earlier proof and only explain how we deal with the new C components. The ()) direction is as in lemma 5.3. The new observations in the 4 cases are: 1. that we can take 2. k ? 1 and the last rewrite step used a type 2 "-rule: Use the fact that w 3. k ? 1 and the last rewrite step used a type 4 rule: Use the fact that C:C 0 ' C 00 . 4. k ? 1 and the last rewrite step used a type 3 rule: Use the fact that w 1 2 C and entail that there exists at least one shu-ing w of w 1 and w 2 s.t. w 2 C 00 . INRIA The Regular Viewpoint on PA-Processes 19 The (() direction is as in lemma 5.3. The new observations in the 5 cases are: 1. 0: The type 0 rules allow all C's containing ''. 2. 3. 0: Then the sequence has the form X a t. Now if there must be a (C in the seq-decomposition of C s.t. a 2 C 0 and w that there is a type 2 rule C) we can use. 4. and we have the type 4a rule we need. Otherwise there is a pair (C in the seq- decomposition of C s.t. w This pair gives us the type 4b rule we need. 5. some shu-e of w 1 and w 2 . Therefore there is a in the paral-decomposition of C s.t. w This pair gives us the type 3 rule we need. If we now let the -nal states of A Post [C] be all (q is a -nal state of AL , then A Post [C] accepts a term t ioe t 2 Post [C](t). (The set of -nal states can easily be adapted so that we recognize Post Ingredients for A Pre [C] : Same as in the construction of A Pre , with an additional C 2 C component. States of A Pre : A state of A Pre is a 4-tuple (q ? The -nal states are all (q is a -nal state of AL and C i the constraint to satisfy. Transition rules of A Pre : The transition rules of A Pre are de-ned as follows: type 0: all rules of the form 0 7\Gamma! (q type 1a: all rules of the form X 7\Gamma! (q ? ; q L ; true; C) s.t. there exists some u with u A? 7\Gamma! q L . type 1b: all rules of the form X 7\Gamma! (q type 2: all rules of the form appears in the paral-decomposition of C 00 . RR n-0123456789 type 3a: all rules of the form is a -nal state of A? and (C; C 0 ) appears in the seq-decomposition of C 00 type 3b: all rules of the form Lemma 7.5. For any t 2 EPA , t A Pre [C] there is some u 2 EPA and some u, u A? Proof. A Pre [C] is A Pre equipped with a new component and the proof follows exactly the lines of the proof of Lemma 5.3. We refer to this earlier proof and only explain how we deal with the new C components. 1. X: The conditions on the C component for the existence of rules of type 0, 1a and 1b agree with the statement of the lemma. 2. A Pre there is a type 3 rule C). Also, the ind. hyp. gives t i In the type 3b case, w 1 2 C. In the type 3a case, we use C Here we have either (1) In the -rst case we apply the induction hypothesis with C itself on t 1 and some C 0 containing " on t 2 , then we can use a type 3b rule. In the second case, there must be a in the seq-decomposition of C, with w and we just have to use the ind. hyp. and a type 3a rule. 3. This case is similar to the previous one. The (() direction uses the pair accouting for w in the paral-decomposition of C. The ()) direction uses the crucial fact that whenever t i shu-ing w of w 1 and w 2 , in particular for the w that C must contain. 7.1 Applications to model-checking The above results let us apply the model-checking method from section 6 to an extended EF logic where we now allow all hCi' formulas for decomposable C. The semantics is given by Mod (hCi') def INRIA The Regular Viewpoint on PA-Processes 21 Decomposability of C is a quite general condition. It excludes the undecidable situations that would exist in the general regular case and immediately includes the extensions proposed in [May97b]. Observe that it is possible to combine decomposable constraints already in the model-checking algorithm: when C 2 C and C are decomposable, we can deal with hC " directly (i.e. without constructing a -nite decomposition system containing C and C 0 ) because it is obvious how to extend the construction for A Pre [C] to some A Pre [C;C 0 ] where several C components are dealt with simultaneously. We can also deal with hC [ C 0 i' and hC:C 0 i' directly since Pre [C [ C 0 ](L) and Pre [C:C 0 ](L) are Pre [C](L) [ Pre [C 0 ](L) and Pre [C](Pre [C 0 ](L)) for any C; C 0 and L. 8 Structural equivalence of PA terms In this section we investigate the congruence j induced on PA terms by the following equations: This choice of equations is motivated by the fact that several recent works on PA (and extensions) only consider processes up-to this same congruence. Our techniques could deal with variants. It is useful to explain how our de-nition of PA compares with the de-nition used in [May97c, May97b]. We consider a transition system between terms from EPA . The terms Mayr considers for his transition system can be seen as equivalence classes, modulo j, of our EPA terms. Write [t] j for the set ft g. The transition relation used by Mayr coincides with a transition relation de-ned by a In the following, we speak of iPAj j when we mean the transition system one obtains with j-classes of terms as states, and transitions given by (4). Our approach is more general in the sense that we can de-ne the other approach in our framework. By contrast, if one reasons modulo j right from the start, one loses the information required to revert to the other approach. For example, the reachability problem ido we have t u ?j from Theorem 4.7 asks for a very precise form for u. The reachability problem solved in [May97c] asks for u modulo j. In our framework, this can be stated as igiven t and u, do we have t 0 RR n-0123456789 22 D.Lugiez, Ph.Schnoebelen and (see below). In the other framework, it is impossible to state our problem. (But of course, the -rst motivation for our framework is that it allows the two regularity theorems.) The rest of this section is devoted to some applications of our tree-automata approach to problems for PAj . The aim is not exhaustivity. Rather, we simply want to show that our framework allows solving (not just stating) problems from the other framework and its variants. 8.1 Structural equivalence and regularity are the associativity-commutativity axioms satis-ed by : and k. We call them the permutative axioms and write t =P u when t and t 0 are permutatively equivalent. are the axioms de-ning 0 as the neutral element of : and k. We call them the simpli-cation axioms and write t & u when u is a simpli-cation of t, i.e. u can be obtained by applying the simpli-cation axioms from left to right at some positions in t. Note that & is a (well-founded) partial ordering. We write . for (&) \Gamma1 . The simpli-cation normal form of t, written t#, is the unique u one obtains by simplifying t as much as possible (no permutation allowed). Such axioms are classical in rewriting and have been extensively studied [BN98]. j coincide with (=P because the permutative axioms commute with the simpli-cation axioms, we have This lets us decompose questions about j into questions about =P and questions about &. We start with =P . Lemma 8.1. For any t, the set [t] =P ug is a regular tree language, and an automaton for [t] =P needs only have m:(m=2)! states if Proof. (Sketch) This is because [t] =P is a -nite set with at most (m=2)! elements. (The exponential blowup cannot be avoided.) The simpli-cation axioms do not have the nice property that they only allow -nitely many combinations, but they behave better w.r.t. regularity. Write [L] & for fu j t & Lemma 8.2. For any regular L, the sets [L] . , [L] & , and [L]# are regular tree languages. From an automaton AL recognizing L, we can build automata of size O(jAj) for these three languages in polynomial time. Proof. 1. [L] . : u is in [L] . ioe u is some t 2 L with additional 0's that can be simpli-ed out. Hence an automaton accepting [L] . is obtained from AL by adding a new state q 0 for the subterms that will be simpli-ed. We also add rules 0 7\Gamma! q 0 , q 0 k q 0 7\Gamma! q 0 , and INRIA The Regular Viewpoint on PA-Processes 23 accepting these subterms, and, for any q in AL , rules q:q 0 7\Gamma! q, q 0 :q 7\Gamma! q, simulating simpli-cation. 2. have been simpli-ed. A simple way to obtain an automaton for [L] & is to synchronize the automaton AL accepting L with the complete automaton A 0 recognizing terms built with 0, : and k only. A 0 has only two states: q 0 and q 6=0 . Once the two automata are synchronized, we have t 7\Gamma! (q; q 0 7\Gamma! q and t A0 We simulate simpli-cation of nullable terms with additional "-rules. Namely, whenever there is a rule (q we add an "-rule (q add a symmetric rule if q 0 do the same for : instead of k. Now a routine induction on the length of derivations shows that s 7\Gamma! (q; q 0 ) ioe 9t 2 L 7\Gamma! q. 3. [L]#: The simplest way to see regularity is to note that [L]# is Note that for a regular L, [L] =P and [L] j are not necessarily regular [GD89]. However we have Proposition 8.3. For any t, the set [t] j is a regular tree language, and an automaton for needs only have m:(m=2)! states if Proof. Combine (5) with lemmas 8.1 and 8.2. 8.2 Structural equivalence and behaviour Seeing terms modulo j does not modify the observable behaviour because of the following standard result: Proposition 8.4. j is a bisimulation relation, i.e. for all t j t 0 and t a ! u there is a The proof is standard but tedious. We shall only give a proof sketch. Proof. For any single equation l = r in the de-nition of j, we show that the set f(loe; roe)g of all instances of the equation is a bisimulation relation. A complete proof of this for takes the better part of p 95 of the book [Mil89] and the other equations can be dealt with similarly, noting that IsNil() is compatible with j. Then there only remains to prove that the generated congruence is a bisimulation. This too is standard: the SOS rules for PA obey a format ensuring that the behaviour of a term depends on the behaviour of its subterms, not their syntax. We may now de-ne a new transition relation between terms: t a some u; u 0 . This amounts to the i[t] j a from (4) and is the simplest way to translate RR n-0123456789 problems for PAj into problems for our set of terms. We adopt the usual abbreviations Proposition 8.5. For any w 2 Act , t w Proof. By induction on the length of w, and using Proposition 8.4. 8.3 Reachability modulo j Now it is easy to prove decidability of the reachability problem modulo j: t Post (t)" Recall that [u] j and Post (t) are regular tree-languages one can build eoeectively. Hence it is decidable whether they have a non-empty intersection. This gives us a simple algorithm using exponential time (because of the size of [u] j ). Actually we can have a better result 3 : Theorem 8.6. The reachability problem in PAj , igiven t and u, do we have t in NP. Proof. NP-easiness is straightforward in the automata framework: We have t for some u 0 s.t. note that ju 00 j - juj. A simple algorithm is to compute u#, then guess non-deterministically a permutation u 00 , then build automata A 1 for [u 00 ] & and A 2 for Post (t). These automata have polynomial-size. There remains to checks whether A 1 and A 2 have a non-empty intersection to know whether the required u 0 exists. Corollary 8.7. The reachability problem in PAj is NP-complete. Proof. NP-hardness of reachability for BPP's is proved in [Esp97] and the proof idea can be reused in our framework. We reduce 3SAT to reachability in PAj . Consider an instance P of 3SAT. P has m variables and n clauses, so that it is some every \Gammag. We de-ne the following r i;j for (Note that j).) The (R1) rules pick a valuation v for the X r 's, the (R3) rules use v to list satis-ed clauses, the (R2) rules discard unnecessary elements. Finally Other applications are possible, e.g.: First proved in [May97c] INRIA The Regular Viewpoint on PA-Processes 25 Proposition 8.8. The boundedness problem in PAj is decidable in polynomial-time. Proof. [t] j can only reach a -nite number of states in PAj ioe t can only reach a -nite number of non-j terms in PA. Now because the permutative axioms only allow -nitely many variants of any given term, Post (L) contains a -nite number of non-j processes ioe [Post (L)]# is -nite. 8.4 Model-checking modulo j The model-checking problem solved in [May97b] considers the EF logic over PAj . Translated into our framework, this amounts to interpret the temporal connectives in terms of ) instead of !: if we write Mod j (') for the interpretation modulo j, we have Mod Additionally, we only consider atomic propositions P compatible with j, i.e. where t and t j u imply u Model-checking in PAj is as simple as model-checking in PA: Lemma 8.9. For any EF-formula ' we have Mod j Proof. By structural induction over ', using Prop. 8.5 and closure w.r.t. j for the hCi' case. The immediate corollary is that we can use exactly the same approach for model-checking in PA with or without j. Conclusion In this paper we showed how tree-automata techniques are a powerful tool for the analysis of the PA process algebra. Our main results are two general Regularity Theorems with numerous immediate applications, including model-checking of PA with an extended EF logic. The tree-automata viewpoint has many advantages. It gives simpler and more general proofs. It helps understand why some problems can be solved in P-time, some others in NP-time, etc. It is quite versatile and many variants of PA can be attacked with the same approach. Acknowledgments We thank H. Comon and R. Mayr for their numerous suggestions, remarks and questions about this work. RR n-0123456789 26 D.Lugiez, Ph.Schnoebelen --R Decidability of bisimulation equivalence for processes generating context-free languages More in-nite results Verifying in-nite state processes with sequential and parallel composition Reachability analysis of pushdown auto- mata: Application to model-checking Rewriting and All That. On the regular structure of pre-x rewriting Tree automata and their application Decidable subsets of CCS. Decidability and decomposition in process algebras. Applications of Tree Automata in Rewriting The theory of ground rewrite systems is decidable. Petri nets A direct symbolic approach to model checking pushdown systems (extended abstract). The reachability problem for ground TRS and some extensions. A model for recursive-parallel programs A formal framework for the analysis of recursive- parallel programs Combining Petri nets and PA-processes Model checking PA-processes Tableaux methods for PA-processes Communication and Concurrency. --TR Decidability of bisimulation equivalence for processes generating context-free languages Process algebra On the regular structure of prefix rewriting Petri nets, commutative context-free grammars, and basic parallel processes Tree languages rewriting and all that Efficient algorithms for pre* and post* on interprocedural parallel flow graphs An automata-theoretic approach to branching-time model checking Communication and Concurrency A Formal Framework for the Analysis of Recursive-Parallel Programs Combining Petri Nets and PA-Processes Bisimulation Equivanlence Is Decidable for Normed Process Algebra Decidable First-Order Transition Logics for PA-Processes Deciding Bisimulation-Like Equivalences with Finite-State Processes The Reachability Problem for Ground TRS and Some Extensions Reachability Analysis of Pushdown Automata How to Parallelize Sequential Processes Infinite Results Model Checking PA-Processes An Automata-Theoretic Approach to Interprocedural Data-Flow Analysis Regularity is Decidable for Normed PA Processes in Polynomial Time Tableau Methods for PA-Processes --CTR Ahmed Bouajjani , Javier Esparza , Tayssir Touili, A generic approach to the static analysis of concurrent programs with procedures, ACM SIGPLAN Notices, v.38 n.1, p.62-73, January Vineet Kahlon , Aarti Gupta, On the analysis of interacting pushdown systems, ACM SIGPLAN Notices, v.42 n.1, January 2007 Anne Labroue , Philippe Schnoebelen, An automata-theoretic approach to the reachability analysis of RPPS systems, Nordic Journal of Computing, v.9 n.2, p.118-144, Summer 2002 Markus Mller-Olm, Precise interprocedural dependence analysis of parallel programs, Theoretical Computer Science, v.311 n.1-3, p.325-388, 23 January 2004 Antonn Kuera , Philippe Schnoebelen, A general approach to comparing infinite-state systems with their finite-state specifications, Theoretical Computer Science, v.358 n.2, p.315-333, 7 August 2006 Denis Lugiez , Philippe Schnoebelen, Decidable first-order transition logics for PA-processes, Information and Computation, v.203 n.1, p.75-113, November 25, 2005 Kamal Lodaya, A regular viewpoint on processes and algebra, Acta Cybernetica, v.17 n.4, p.751-763, January 2006

verification of infinite-state systems;process algebra;tree automata

507259

Axioms for real-time logics.

This paper presents a complete axiomatization of two decidable propositional real-time linear temporal logics: Event Clock Logic (EventClockTL) and Metric Interval Temporal Logic with past (MetricIntervalTL). The completeness proof consists of an effective proof building procedure for EventClockTL. From this result we obtain a complete axiomatization of MetricIntervalTL by providing axioms translating formulae, the two logics being equally expressive. Our proof is structured to yield axiomatizations also for interesting fragments of these logics, such as the linear temporal logic of the real numbers (TLR).

Introduction Many real-time systems are safety-critical, and therefore deserve to be specified with mathematical precision. To this end, real-time linear temporal logics [5] have been proposed and served as the basis of specification languages. preliminary version of this paper appeared in the Proceedings of the Tenth International Conference on Concurrency Theory (CONCUR), Lecture Notes in Computer Science 1466, Springer-Verlag, 1998, pp. 219-236. ?? This work is supported in part by the ONR YIP award N00014-95-1-0520, the NSF CAREER award CCR-9501708, the NSF grant CCR-9504469, the DARPA/NASA grant NAG2-1214, the ARO MURI grant DAAH-04-96-1-0341, the Belgian National Fund for Scientific Research (FNRS), the European Commission under WGs Aspire and Fireworks, the Portuguese FCT under Praxis XXI, the Walloon region, and Belgacom. Preprint submitted to Elsevier Preprint 12 November 1999 They use real numbers for time, which has advantages for specification and compositionality. Several syntaxes are possible to deal with real time: freeze quantification [4,12], explicit clocks in a first-order temporal logic [11,21], integration over intervals [10], and time-bounded operators [17]. We study logics with time-bounded operators, because those logics are the only ones which have, under certain restrictions, a decidable satisfiability problem [5]. The logic extends the operators of temporal logic to allow the specification of time bounds on the scope of temporal operators. For example, the that "every p event is followed by some q event after exactly 1 time unit." It has been shown that the logic undecidable and even not recursively axiomatizable [4]. One reason for this undecidability result is the ability of to specify exact distances between events; these exact distance properties are called punctuality properties. The logic MetricIntervalTL is obtained from removing the ability to specify punctuality properties: all bounds appearing in temporal operators must be non-singular intervals. For example, the formula which expresses that "every p event is followed by some q event after at least 1 time unit and at most 2 time units," is a MetricIntervalTL formula, because the interval [1; 2] is non-singular. The logic MetricIntervalTL is decidable [3]. This decidability result allows program verification using automatic techniques. However, when the specification is large or when it contains first-order parts, a mixture of automatic and manual proof generation is more suitable. Unfortunately, the current automatic reasoning techniques (based on timed automata) do not provide explicit proofs. Secondly, an axiomatization provides deep insights into a logic. Third, a complete axiomatization serves as a yardstick for a definition of relative completeness for more expressive logics (such as first-order extensions) that are not completely axiomatizable, in the style of [16,20]. This is why the axiomatization of time-bounded operator logics is cited as an important open question in [5,17]. We provide a complete axiom system for decidable real-time logics, and a proof-building procedure. We build the axiom system by considering increasingly complex logics: LTR [6], EventClockTL with past clocks only, Event- with past and future clocks (also called with past and future operators. The method that we use to show the completeness of our axiomatization is standard: we show that it is possible to construct a model for each consistent formula. More specifically, our proof of completeness is an adaptation and an extension of the proof of completeness of the axiomatization of The handling of the real-time operators requires care and represents the core technical contribution of this paper. Some previous works presented axioms for real-time logics, but no true (versus relative) completeness result for dense real-time. In [12], completeness results are given for real-time logics with explicit clocks and time-bounded operators, but for time modeled by a discrete time domain, the natural numbers. In [9,7], a completeness result is presented for the qualitative (non real-time) part of the logics considered in this pa- per. There, the time domain considered is dense but the hypothesis of finite variability that we consider 1 is dropped and, as a consequence, different techniques have to be applied. In [17], axioms for real-time logics are proposed. These axioms are given for first-order extensions of our logics, but no relative completeness results are studied (note that no completeness result can be given for first-order temporal logics.) Finally, a relative completeness result is given for the duration calculus in [10]. The completeness is relative to the hypothesis that valid interval logic formulae are provable. Models and logics for real-time 2.1 Models As time domain T, we choose the nonnegative real numbers R 0g. This dense domain is natural and gives many advantages detailed elsewhere: compositionality [6], full abstractness [6], stuttering independence [1], easy refinement. These advantages, and the results of this paper, mainly depend on density: they can easily be adapted for the rational numbers Q, the real numbers R. To avoid Zeno's paradox, we add to our models the condition of finite variability [6] (condition (3) below): only finitely many state changes can occur in a finite amount of time. An interval I ' T is a convex subset of time. Given t 2 T, we freely use notations such as t + I for the interval ft 0 j 9t 00 2 I with t the constraint "t ? t 0 for all t 0 2 I ", # I for the interval ft ? 0j9t g. A bounded non-empty interval has an infimum (also called greatest lower bound, or left endpoint, or begin) and a supremum (also called least upper bound, or right endpoint, or end). Such an interval is thus usually written as e.g. (l; r], where l is the left endpoint, the rounded parenthesis in "(l" indicates that l is excluded from the interval, r is the right endpoint, and the square parenthesis in "r]" indicates that r is included in the interval. The interval is called left-open and right-closed. If we extend the notation, as usual, by allowing r to be 1, then any interval can be written in this form. Two intervals I and J are adjacent if the right endpoint of I, noted r(I), is equal to the left endpoint of J , noted l(J ), and either I is right-open and J is left-closed or I is right-closed and J is left-open. We say that a non-empty 1 In every finite interval of time, the interpretation of propositions can change only finitely many times. interval I is singular if r(I). In this case, we often use the rather than [t; t]. Similarly, ! l abbreviates (0; l), etc. An interval sequence is an infinite sequence of non-empty bounded intervals so that (1) the first interval I 0 is left-closed with left endpoint 0, (2) for all i - 0, the intervals I i and I i+1 are adjacent, and (3) for all t 2 T, there exists an Consequently, an interval sequence partitions time so that every bounded subset of T is covered by finitely many elements of the partition. Let P be a set of propositional symbols. A state s ' P is a set of propositions. A timed state sequence -s; - I) is a pair that consists of an infinite sequence - s of states and an interval sequence - I. Intuitively, it states the period I i during which the state was s i . Thus, a timed state sequence - can be viewed as a function from T to 2 P , indicating for each time t 2 T a state - 2.2 The Linear Temporal Logic of Real Numbers (LTR) The formulae of LTR [6] are built from propositional symbols, boolean con- nectives, the temporal "until" and "since" and are generated by the following where p is a proposition. The LTR formula OE holds at time t 2 T of the timed state sequence - , written according to the following definition, where we An LTR formula OE is satisfiable if there exists - and a time t such that (-; t) OE, an LTR formula OE is valid if for every - and every time t we have (-; t) This logic was shown to be expressively equivalent to the monadic first-order logic of the order over the reals [15]. Our operators U; S are slightly non-classical, but more intuitive: they do not require OE 2 to start in a left-closed interval. On the other hand, each of them is slightly weaker than its classical variant, but together they have the same expressive power, as we show by providing mutual translations below in sections 2.2.1 and 2.4.1. It is thus a simple matter of taste. We will note the classical until as - U. 2.2.1 Abbreviations In the sequel we use the following abbreviations: defined below). the "Until" reflexive for its first argument; the "Until" reflexive for its two arguments; meaning "just after in the future" or "for a short time in the future". The dual of fl is noted K + in [9], and it means thus "arbitrarily close in the future". We don't introduce it, since we will see that due to finite variability, fl is his own dual. meaning "eventually in the future"; meaning "always in the future"; ffl their reflexive counterparts: \Sigma meaning "unless in the future"; ffl its reflexive counterparts: W and the past counterpart of all those abbreviations: the "Since" reflexive for its first argument; the "Since" reflexive for its two arguments; meaning "just before in the past" or "arbitrarily close in the meaning "eventually in the past"; meaning "always in the past"; ffl their reflexive counterparts: meaning "unless in the past"; ffl its reflexive counterparts: Z 2.3 Event-Clock Temporal Logic The formulae of EventClockTL [22] are built from propositional symbols, boolean connectives, the temporal "until" and "since" operators, and two real-time op- erators: at any time t, the history operator / I OE asserts that OE was true last in the interval t \Gamma I, and the prediction operator . I OE asserts that OE will be true next in the interval t I. The formulae of EventClockTL are generated by the following I OE where p is a proposition and I is an interval which can be empty, singular and whose bounds are natural numbers (or infinite). The holds at time t 2 T of the timed state sequence - , written (-; t) according timevalue of the for timed trace sequence event tick reset event reset event reset undefined small big blocked small small blocked small tick event clock Fig. 1. A History clock evolving over time to the rules for LTR and the following additional clauses: t A . I OE formula can intuitively be seen as expressing a constraint on the value of a clock that measures the distance from now to the next time where the formula OE will be true. In the sequel, we use this analogy and call this clock a prediction clock for OE. Similarly, a / I OE formula can be seen as a constraint on the value of a clock that records the distance from now to the last time such that the formula OE was true. We call such a clock a history clock for OE. For a history (resp. prediction) clock about OE, ffl the next / =1 OE (resp. previous . =1 OE) is called its tick; ffl the point where OE held last (resp. will hold next) is called its event; ffl the point (if any) at which OE will hold again (resp. held last) is called its reset; ffl if OE is true at time t and was true just before t (resp. and will still be true just after t) then we say that the clock is blocked at time t; ffl if OE was never true before t (resp. will never be true after t) then the clock is undefined at time t. The main part of our axiomatization consists in describing the behavior and the relation of such clocks over time. For a more formal account on the relation between EventClockTL formulae and clocks, we refer the interested reader to [22]. We simply recall: Theorem 1 [22] The satisfiability problem for EventClockTL is complete for Pspace. which is the best result that can be expected, since any temporal logic has this complexity. Example 1 (p ! . =5 p) asserts that after every p state, the first subsequent p state is exactly 5 units later (so in between, p is false); the formula (/ =5 asserts that whenever the last p state is exactly 5 units ago, then q is true now (time-out). 2.4 Metric-Interval Temporal Logic restricts the power of MetricTLin an apparently different way from EventClockTL: here the real-time constraints are attached directly to the until, but cannot be punctual. The formulae of MetricIntervalTL [3] are built from propositional symbols, boolean connectives, and the time-bounded "until" and "since" operators: U I OE 2 j OE 1 S I OE 2 where p is a proposition and I is a nonsingular interval whose bounds are natural numbers or infinite. The holds at time of the timed state sequence - , written (-; t) according to the following definition (the propositional and boolean clauses are as for LTR): U I OE 2 iff 9t Here, we have used the classical until to respect the original definition, but this doesn't matter as explained in subsection 2.2.1. Theorem 2 [3] The satisfiability problem for MetricIntervalTL is complete for Expspace. So although the logics are equally expressive, their translation must be difficult enough to absorb the difference in complexity. Our translation, presented in section 5, indeed gives an exponential blowup of formulae. 2.4.1 Abbreviations In the sequel we use the following abbreviations: U (0;1) OE 2 , the untimed "Until" of MetricIntervalTL. UOE expresses that the next OE-interval is left-closed. U I OE 2 . U I OE, meaning "within I"; ffl I OE j :\Sigma I :OE, meaning "always within I"; and the past counterpart of all those abbreviations. The fact that we use the same notations as in the other logics is intentional and harmless, since the definitions are semantically equivalent. Furthermore, now that we have re-defined the basic operators of EventClockTL, we also use its abbreviations. asserts that every q state is preceded by a p state of time difference at most 5, which is right-closed, and all intermediate states are r states; the formula (p ! \Sigma [5;6) p) asserts that every p state is followed by a p state at a time difference of at least 5 and less than 6 time units. This is weaker than the EventClockTL example, since p might also hold in between, and of course because 5 units are not exactly required. 3 Axiomatization of EventClockTL In section 4, we will present a proof-building procedure for EventClockTL. In this section, we simply collect the axioms used in the procedure, and present their intuitive meaning. Our logics are symmetric for past and future (a duality that we call the "mirror principle"), except that time begins but does not end: therefore the axioms will be only written for the future, but with the understanding that their mirror images, obtained by replacing U by S, . by /, etc. are also axioms. This does not mean that we have an axiomatization of the future fragment of these logics: our axioms make past and future interact, and our proof technique makes this interaction is unavoidable, mainly in axiom (11). 3.1 Qualitative axioms (complete for LTR) We use the rule of inference of replacement of equivalent formulae: (1) All propositional tautologies (2) For the non-metric part, we use the following axioms and their mirror images: They mainly make use of the fl operator, because as we shall see, it corresponds to the transition relation of our structure. Axiom (3) is the usual necessitation or modal generalization rule, expressed as an axiom. Similarly, (4) is the usual weakening principle, expressed in a slightly non-classical form. (5), (6) allow to distribute fl with boolean operators. Note that the validity of (6) requires finite variability. (7), (8) describe how the U and S operators are transmitted over interval boundaries. (9) gives local consistency conditions over this transmission. (10) ensures eventuality when combined with (11). It can also be seen as weakening the left side of the U to ?. The induction axiom is essential to express finite variability: If a property is transmitted over interval boundaries, then it will be true at any point; said otherwise, any point is reached by crossing finitely many interval boundaries. The axioms below express that time begins (12) but has no end (13): We have written the other axioms so that they are independent of the begin or end axioms, in order to deal easily with other time domains (see subsection 4.4). This is why some apparently spurious fl? occur above, e.g. in (11): they are useful when the future is bounded. Remark 3 Theorem 21 shows that the axioms above form a complete axiomatization of the logic of the real numbers with finite variability, defined as LTR in [6]. The system proposed in [6] is unfortunately unsound, redundant and incomplete. Indeed, axiom F5 of [6] is unsound; axiom F7 can be deduced from axiom F8; and the system cannot derive the induction axiom (11). To see this last point, take the structure formed by R -0 followed by R, with finite variability: it satisfies the system of [6] (corrected according to [7]) but not the induction axiom. Thus this valid formula cannot be derived in their system. 3.2 Quantitative axioms For the real-time part, we first describe the static behavior; intersection, union of intervals can be translated into conjunction, disjunction due to the fact that there is a single next event: . I[J OE $ . I OE - . J OE (14) . I"J OE $ . I OE - . J OE (15) Since . is a strict future operator, the value 0 is never used: If we do not constrain the time of next occurrence, we simply require a future occurrence: Finally the addition corresponds to nesting: The next step of the proof is to describe how a single real-time . I OE evolves over time, using fl and -. We use (20) to reduce left-open events to the easier case of left-closed ones. These axioms are complete for formulae where the only real-time operators are prediction operators . I OE and they all track the same (qualitative) formula OE. For a single history tracked formula, we use the mirror of the axioms plus an axiom expressing that the future time is infinite, so that any bound will be exceeded: The description provided by these axioms are mostly expressed by the automaton of figure 2, showing the possible evolution of history predicates. This figure will receive a formal status in lemma 22. Most consequences of OE Fig. 2. The possible evolutions of a history clock these axioms can simply be read from this automaton: For instance, /?1 OE ! (/ ?1 OE - :OE)U - fl /!1 OE is checked by looking at paths starting from /?1 OE. As soon as several such formulae are present, we cannot just combine their individual behavior, because the .; / have to evolve synchronously (with the common implicit real time). We use a family of axioms (and their mirrors) to express this common speed. They express the properties of order and addition, but expressed with different clocks. Said otherwise, the ordering of the ticks should correspond to the ordering of their events. We use U (or W) to express the ordering: :pUq means that q will occur before (or at the same time as) any p. E.g. in (26), the antecedent / =1 OE states that OE ticks now, thus after of together with /. Then their events shall be in the same order: :OES/. Similarly, (30) says that if last OE was less than 1 ago, and / was even closer, than last was less than 1 ago as well. (. (. 3.3 Theorems We will use in the proof some derived rules of LTR (and thus EventClockTL): Lemma 4 The rules of modus ponens and modal generalization are derivable. OE Proof. ffl the rule of modus ponens (32) is derived from replacement (1) as follows: from OE we deduce propositionally OE $ ?; by (1) we replace OE by ? in propositionally /. ffl the rule of modal generalization (33) (also called necessitation) is derived similarly from (1) and (3): From OE, we deduce :OE $ ?. Replacing in (3), we obtain :(/U:OE). By taking / := ?, we get \LambdaOE. We'll also need some theorems: . I OE $ :\SigmaOE - . I OE (43) . I OE . I OE ! . J OE with (I ' J) (45) Proof. By (13), we can remove the condition fl? in the mirror of (6). We use (5) and duality through (34). Expanding the definition of -, we have to prove ?SOE ! ?S?. This results from the mirror of (4) with OE := ?; / := ?; / 0 := OE. From (36). So all - formulae are false at the beginning of time. (38) By (8). By (7). (40) By (13), (10). (41) Take (14) with I := ;; J := [0; 0]. By (16) we obtain (42) We'll prove its mirror. By (14), / I / ! /? 0/. By (17), \Sigma/. By (10), fl/. (43) By (15), (14), (17). (44) By (15), (14), (17). (45) By (15). (or by (14)). By (4). 4 Completeness of the axiomatic system for EventClockTL As usual, the soundness of the system of axioms can be proved by a simple inductive reasoning on the structure of the axioms. We concentrate here on the more difficult part: the completeness of the proposed axiomatic system. As usual with temporal logic, we only have weak completeness: for every valid formula of EventClockTL, there exists a finite formal derivation in our axiomatic system for that formula. So if As often, it is more convenient to prove the contrapositive: every consistent EventClockTL formula is satisfiable. Due to the mirror principle, most explanations will be given for the future only. Our proof is divided in steps, that prove the completeness for increasing fragments of EventClockTL. (1) We first deal with the qualitative part, without real-time. This part of the proof follows roughly the completeness proof of [19] for discrete-time logic. (a) We work with worlds that are built syntactically, by maximal consistent sets of formulae. (b) We identify the transition relation, and its syntactic counterpart: it was the "next" operator for discrete-time logic [19], here it is the fl, expressing the transition from a closed to an open interval, and -, expressing the transition from an open to a closed interval. (c) We impose axioms describing the possible transitions for each operator (d) We give an induction principle (11) that extends the properties of local transitions to global properties. (2) For the real-time part: (a) We give the statics of a clock; (b) We describe the transitions of a clock; (c) By further axioms, we force the clocks to evolve simultaneously. The completeness of these axioms is proved by showing that only realistic clock evolutions are allowed by the axioms. 4.1 Qualitative part Let us assume that the formula ff is consistent and let us prove that it is satisfiable. To simplify the presentation of the proof, we use the following lemma: Lemma 5 Every EventClockTL formula can be rewritten into an equivalent formula of (using only the constant 1). Proof. First by the use of the theorem . I OE . !I OE - . #I OE (44), every formula . I OE with l(I) 6= 0 can be rewritten as a conjunction of formulae with 0-bounded intervals. Using the axioms .-m+n OE $ .-m .-n OE (18) and every interval can be decomposed into a nesting of operators associated with intervals of length 1. \Xi In the sequel, we assume that the formula ff for which we want to construct a model is in EventClockTL 1 , as allowed by lemma 5. We now define the set C(ff) of formulae associated with ff: ffl Sub: the sub-formulae of ff. ffl The formulae of Sub subject to a future real-time constraint: Subg. We will say that a prediction clock is associated to these formulae. ffl For these formulae, we will also track flOE when the next occurrence of OE is left-open: this will simplify the notation. The information about OE will be reconstructed by axiom (20). Rg. ffl To select whether to track OE or flOE, we need the formulae giving the openness of next interval: ffl The formulae giving the current integer value of the clocks: I = f.!1 OE; . =1 OE; Jg. Thanks to our initial transformation, we only have to consider whether the integer value is below or above 1. Among these, the "tick" formulae will be used in F to determine the fractional parts of the clocks: Ig. ffl We also define the mirror sets. For instance, ffl The formulae giving the ordering of the fractional parts of the clocks, coded by the ordering of the ticks: g. ffl The eventualities: ffl The constant true ?, because -? will be used in lemma 14. We close the union of all sets above under :; fl; - to obtain the closure of ff, noted C(ff). This step preserves finiteness since we stop after adding just one of each of these operators. Theorems (39), (38) show that further addition would be semantically useless. For the past, we only have (6), (37). They also give the same result, since we only have two possible cases: if -? is true, we can move all negations outside and cancel them, except perhaps one. Otherwise, we know that all -/ are false by (4). In each case, at most one - or fl and one are needed. We use the notational convention to identify formulas with their simplified form. For example, we write OE 2 C(ff) $ flOE 2 C(ff) to mean is the simplification operator. Note that although we are in the qualitative part, we need already include the real-time formulae that will be used later. In this subsection they behave as simple propositions. A propositionally consistent structure A set of formulae F ae C(ff) is complete w.r.t. C(ff) if for all formulae OE 2 C(ff), either OE 2 F or :OE 2 F ; it is propositionally consistent if (i) for all We call such a set a propositional atom of C(ff). We define our first structure, which is a finite graph, the set of all propositional atoms of C(ff) and \Delta ' \Theta is the transition relation of the structure. \Delta is defined by considering two sub-relations: represents the transition from a right-closed to a left-open interval; represents the transition from a right-open to a left-closed interval. propositional atoms. We define The transition relation \Delta is the union of \Delta or Now we can define that the atom A is singular iff it contains a formula of the symmetrically OE - OE. Lemma 6 In the following, A and B are atoms: (1) A is singular iff it is irreflexive (i.e. (2) If A\Delta [ B, then A is not singular and (B is singular or A is not singular and (B is singular or singular, then there is at most one atom A such that a unique C such that B \Delta ] C. A is initial iff it contains : -?. It is then singular, since it contains ? - ?. A is monitored iff it contains ff, the formula of which we check floating satisfiability. Any atom A is exactly represented by the conjunction of the formulae that it contains, written - A. By propositional completeness, we have: Lemma 7 ' W A. For any relation \Delta, we define the formula \Delta(A) to be W B. The formula B can be simplified to V :flOE2A :OE, because in the propositional structure, all other members of a B are allowed to vary freely and thus cancel each other by the distribution rule. Lemma Proof. Dually, W B can be simplified to V OE2A -OE. Therefore: Lemma 9 ' - Now let \Delta + be the transitive closure of \Delta. Since \Delta Similarly, Lemma Using the disjunction rule for each reachable - A, we obtain: ' (A). Now we can use the induction axiom Using necessitation (33) and modus ponens (32), we obtain: Lemma An EventClockTL-consistent structure We say that an atom A is EventClockTL-consistent if it is propositionally consistent and consistent with the axioms and rules given in section 3. Now, we consider the structure - \Delta), where - is the subset of propositional atoms that are EventClockTL-consistent and - \Lambdag. Note that the lemmas above are still valid in the structure - \Pi as only inconsistent atoms are suppressed. We now investigate more deeply the properties of the structure - \Pi and show how we can prove from that structure that the consistent formula ff is satisfiable. We first have to define some notions. ffl A maximal strongly connected substructure (MSCS)\Omega is a non-empty set of of the structure - \Pi such that: (1) for all every atom can reach all atoms of \Omega\Gamma i.e., \Omega is strongly connected; (2) for all such that D 1 2\Omega then i.e.,\Omega is maximal. ffl A MSCS\Omega is called initial if for all D 1 \DeltaD 2 and D 2 2\Omega then D 1 2 \Omega\Gamma i.e. \Omega has no incoming edges. ffl A MSCS\Omega is called final if for all D 1 \DeltaD 2 and D 1 2\Omega then D 2 2 \Omega\Gamma i.e.\Omega has no outgoing edges. ffl A MSCS\Omega is called self-fulfilling if for every formula of the form OE 1 UOE 2 2 A with A 2 \Omega\Gamma there exists B 2\Omega such that OE 2 2 B. We now establish two properties of MSCS of our structure - \Pi. Every final MSCS\Omega of the structure - \Pi is self-fulfilling. Proof. Let us make the hypothesis that there exists OE 1 UOE 2 2 A with A 2\Omega and for all B 2 D, OE 2 62 B. By lemma 12 and as by hypothesis OE 2 62 B, for theorem (46) and a propositional reasoning, we conclude . Using the axiom (10) and the hypothesis that OE 1 we obtain ' - by definition of \Sigma, we obtain ' - contradiction with ' - which is impossible since A is, by hypothesis, consistent. \Xi Lemma 14 Every non-empty initial MSCS\Omega of the structure - \Pi contains an initial atom, i.e. there exists A 2\Omega such that -? 62 A. Proof. By definition of initial MSCS, we know that for all D 1 \DeltaD 2 and D 2 2 \Omega\Gamma then us make the hypothesis that for all A 2 \Omega\Gamma -? 2 A. By the mirror of lemma 12 we conclude, by a propositional reasoning and the hypothesis that -? 2 D for all D such that D - contradicts axiom (12), so A 62 - \Pi, thus\Omega is empty. \Xi Actually such initial MSCS are made of a single initial atom. In the sequel, we concentrate on particular paths, called runs, of the structure \Pi. A run of the structure - \Delta) is a is an infinite sequence of atoms and - is an infinite sequence of intervals such that: (1) Initiality: A 0 is an initial atom; (2) Consecution: for every i - 0, A i (3) Singularity: for every i - 0, if A i is a singular atom then I i is singular; Alternation: I 0 I alternates between singular and open intervals, i.e. for all i ? 0, I 2i is singular and I 2i+1 is open. (5) Eventuality: the set fA n ; :::; A n+mg is a final MSCS. Note that, for the moment, the timing information provided in - I is purely qualitative (singular or open); therefore any alternating sequence is adequate at this qualitative stage. Later, we will construct a specific sequence satisfying also the real-time constraints. In the sequel, given I), ae(t) denotes the atom A i such that t 2 I i . Lemma 15 The transition relation - \Delta of the structure - \Pi is total, i.e. for all atoms A 2 - , there exists an atom such that A - \DeltaB. Proof. We prove - is consistent and can thus be completed to form an atom B. Assume it is not: by definition We can replace ? in (13), giving ' fl: - \Phi. By \Phi. By (5), the set fflOEj fl OE 2 Ag [ ffl:OEj: fl OE 2 Ag is inconsistent. Using (34) again, the set fflOEjflOE 2 Ag[f:flOEj:flOE 2 Ag ' A is inconsistent, and thus A is inconsistent, contradicting A 2 - . \Xi Lemma 16 For every atom A of the structure - \Pi, there is a run ae that passes through A. Proof. (1) Initiality, i.e. every atom of - \Pi is either initial or can be reached by an initial atom. Let us consider an atom A, if A is initial then we are done, otherwise, let us make the hypothesis that it can not be reached by an initial atom, it means: for all B - so by propositional completeness -? 2 B. By lemma 12 and a propositional reasoning, we Using axiom (12) we obtain a contradiction in A. We use this path for the first part of the run. (2) Consecution, by construction. (3) Singularity: i.e., every odd atom is not singular. For the first and second part of the run, we can obtain this by taking a simple path (thus without self-loops). Since the first atom A 0 is initial, it is singular; from there on, non-singular and singular states will alternate by lemma 6. For the final repetition, this technique might not work when the MSCS is a single atom. Then we know that this single atom is non-singular, and thus Singularity is also verified. Alternation: we can choose any alternating interval sequence, since the timing information is irrelevant at this point. Eventuality, i.e. every atom of - \Pi can reach one of the final MSCS of - \Pi. It is a direct consequence of the fact that - \Delta is total and the fact that - is finite. We use this reaching path for the second part of the run, then an infinite repetition of this final MSCS. A run I) of the structure - \Pi has the qualitative Hintikka property if it respects the semantics of the qualitative temporal operators which is expressed by the following conditions (real-time operators will be treated in the following H1 if A i is singular then I i is singular; either I i is singular and there exists j ? i s.t. OE 2 2 A j and for all k s.t. or I i is not singular and (2) or there exists either I i is singular and there exists and for all k s.t. or I i is not singular and (2) or there exists We call such a run a qualitative Hintikka run. Next, we show properties of some additional properties of runs related to the Hintikka properties above: Lemma 17 For every run I) of the structure - \Pi, with - for every i - 0 such that \SigmaOE 2 A ffl either I i is singular and there exists ffl or I i is non-singular and there exists j - i such that OE 2 A j . Proof. First let us prove the following properties of the transition relation - \Delta: Recall that \SigmaOE j ?UOE, and by definition of - propositional reasoning, we obtain that definition of - the mirror of axiom (8) and a propositional reasoning, we obtain By the two properties above, we have that if \SigmaOE 2 A i then either OE appears in A j with j ? i if I i is singular (and thus right closed), j - i if I i is not singular (and thus an open interval) or OE is never true and \SigmaOE propagates for the rest of the run. But this last possibility is excluded by our definition of run: by clause (5), every run eventually loops into a final (thus self-fulfilling by lemma MSCS\Omega\Gamma Then either OE is realized before this looping or \SigmaOE 2\Omega and by 2\Omega and is thus eventually realized. \Xi Lemma For every run I) of the structure - \Pi, for every position i in the run if OE 1 UOE 2 2 A i then the right implication of property H2 is verified, i.e: ffl either A i is singular and there exists and for all k s.t. ffl or A i is not singular and (2) or there exists Proof. By hypothesis we know that OE 1 UOE 2 2 A i and we first treat the case where A i is singular. ffl By the axiom (10) and lemma 17, we know that there exists j ? i such that us make the hypothesis that A j is the first OE 2 -atom after A i . ffl It remains us to show that: for all k s.t. We reason by induction on the value of k. Base case: By hypothesis we have OE 1 UOE 2 2 A i and also A i (as A i is right closed) and thus for all flOE 2 A i ; OE 2 A i+1 by definition of - theorem (35) and axiom (5), and the fact that by hypothesis OE 2 62 A i+1 , (Prop) allows us to conclude that OE 1 2 A i+1 . Induction case: induction hypothesis, we know that OE 1 2 A k\Gamma1 and OE 1 as is the first position after i where OE 2 is verified). To establish the result, we reason by cases : I k is open and thus I k\Gamma1 is singular and right closed. We have A and thus for all flOE 2 C(ff); flOE 2 A i by definition of As OE 1 UOE 2 2 A k\Gamma1 by induction hypothesis and the axiom (7) we conclude that OE 1 UOE 2 2 A k . Using the axiom (9), theorem (35), axiom (5) and the fact that OE 2 62 A k , and (Prop), we conclude that OE 1 2 A k . (2) I k is closed which implies that I k\Gamma1 is right open and A k\Gamma1 . By definition of - have that for all -OE 2 C(ff); -OE 2 A k A we have :OE 2 2 A k . Using those properties and the mirror of axiom (8) we conclude that OE 1 - OE 1 UOE 2 2 A k . We now have to treat the case where A i is not singular. By the axiom and lemma 17 we know that there exists a later atom A j , i.e. j - i, such that and we are done. Otherwise j ? i, and we must prove that for all k s.t. this can be done by the reasoning above. \Xi We now prove the reverse, i.e. every time that OE 1 UOE 2 is verified in an atom along the run then OE 1 UOE 2 appears in that atom. This lemma is not necessary for qualitative completeness but we use this property in the lemmas over real-time operators. Lemma 19 For every run I) of the structure - \Pi, for every position ffl either A i is singular and there exists and for all k s.t. ffl or A i is not singular and (2) or there exists then Proof. We reason by considering the three following mutually exclusive cases: (1) A i is singular and there exists We reason by induction to show that OE 1 UOE 2 2 A j \Gammal for all l s.t. ffl Base case: l = 1. By hypothesis, we know that OE 2 2 A j . We now reason by cases: (a) if A j \Gamma1 is right closed then we have A j \Gamma1 by definition of . Using the axiom (9) we deduce by (Prop) that (b) if A j \Gamma1 is right open then we know that by hypothesis) and thus OE 1 2 A j \Gamma1 . Also as A j \Gamma1 Using the mirror of axiom (8) and a propositional reasoning, we by definition of - ffl Induction case: 1 - l we have established the result us show that we have the result for A j \Gammal . First note that by hypothesis, OE 1 2 A j \Gamma(l\Gamma1) . We again reason by cases: (a) I j \Gammal is right closed. Then we have A j \Gammal and by definition of - \Gammal and by axiom (7) we have that OE 1 UOE 2 2 A j \Gammal . (b) A j \Gammal is right open. Then we have A j \Gammal - and by definition of - \Gammal . We know that by hypothesis, OE 1 2 A j \Gammal as singular and I j \Gammal (by induction hypothesis). Using the mirror of axiom (8) and a propositional reasoning, we obtain -(OE 1 and by definition of - that OE 1 UOE 2 2 A j \Gammal . (2) A i is not singular and OE 2 2 A j . As A i is not singular, we have A i definition of - . By the axiom (9) and a proposition reasoning, we obtain the desired result: OE 1 UOE 2 2 A i . (3) A i is not singular, OE 2 62 A j , and there exists and for all This case is treated by an inductive reasoning similar to the first one above. We have also the two corresponding mirror lemmas for the S-operator. From the previous proved lemmas, it can be shown that the qualitative axioms of section 3 are complete for the qualitative fragment of EventClockTL, i.e. the logic LTR. Lemma 20 A run ae has the Hintikka property for LTR formulae: for every Proof. The Hintikka property was proved in the lemmas above, but expressed without reference to time t. It remains to prove that this implies the usual definition, by induction on formulae. We must prove 9t 0 ? from H2. Of course, we take t 0 somewhere in I j , so that t 0 can be divided in 3 parts: the part in I i , which is empty when I i is singular, the part in some I k (i j), the part in I j . Each of them (2) Conversely, the usual definition implies H2: First note that given t, if ae(t) is not singular but I i is singular, it means that A lemma 6. Thus we can merge I i ; I i+1 to ensure that I i is singular iff A i is singular without loss of generality. Let j be the first index where OE 2 , I i is singular, or else j - i. We can take t 0 ? t in I j without loss of generality. Since we need t 00 must (H3) is symmetric. \Xi Finally, we have the following theorem that expresses the completeness of the qualitative axioms for the logic LTR: Theorem 21 Every LTR formula that is consistent with the qualitative axioms is satisfiable. Proof. Let ff be a consistent LTR formula. We construct - \Delta). Let be an atom of the structure such that ff 2 - . Such an atom B exists as ff is consistent. By lemma 16, there exists a run I) such that for some i - 0. By lemma 20, we have (ae; t) thus ff is We now turn to the completeness of real-time axioms. 4.2 Quantitative part A run I) of the structure - \Pi has the timed Hintikka property if it respects the Hintikka properties defined previously and the two following additional properties: H4 . I OE 2 ae(t) iff there exists t 0 2 t+I such that OE 2 ae(t 0 ) and A run that respects those additional properties is called a well-timed run. In the sequel, we will show that for each run of the structure - \Pi, we can modify its sequence of intervals, using a procedure, in such a way that the modified run is well-timed. Recall that given a tracked formula OE 2 R, ffl . =1 OE is called its tick; called its event (note that the second case need not be considered thanks to the axioms (20), ffl (OE -:OE) - (:OE - flOE) is called its reset. The evolution of the real-time predicates is described by figure 2. We can now see the status of this drawing: Lemma 22 For any tracked formula OE 2 R, the projection of - \Pi (restricted to atoms containing the formulae COE) on OE; /!1 OE; / =1 OE; /?1 OE; \Sigma\GammaOE is contained in figure 2. Proof. It suffices to show that no further consistent atoms nor transitions can be added to the figure. ffl Atoms: from the axioms (15), (17), (14), (16). ffl Transitions: We simply take all missing arrows of the figure, and show that they cannot exist. As the proof is fairly long, we only show some excerpts. (1) Assume that an atom A containing OE; / =1 OE is linked to an atom B containing OE in this way: A - B, by axioms (14), (15), (16), we have : /!1 OE 2 B. Now by definition of - and by (34), : fl /!1 OE 2 A. Now the main step: we use the mirror of (23), negated on both sides. :fl? is impossible by (13), and thus we can conclude contradicting OE 2 A. (2) Now we show the only two transitions which are eliminated by the restriction to COE. The first one is A - contains /!1 OE; :OE; COE and contains /!1 OE; OE. We prove using (9). In more detail, COE abbreviates :OEU(OE -:OE). Applying (9) and unfolding U - , we obtain using first disjunct is impossible, by (5), (34), (38). On the other hand, by definition of - whence the contradiction (3) The second transition eliminated is A - contains /?1 OE; :OE; COE and B contains /!1 OE; :OE. By definition of - A. By axiom (22), contradicting /?1 OE 2 A. A constraint is a real-time formula of an atom A i . The begin of a constraint is the index e at which its previous event, tick or reset occurred. The end of a constraint is the index j at which its next event, tick or reset occurs. This vocabulary refers to the order of time only: the begin is always before the corresponding end, whether for history or prediction operators. Begins, ends, ticks, resets, events are always singular. We say that (the history clock of) OE is active between an event OE and the next reset of OE. It is small between its event and the next tick or reset. After this, it is big. When it is big, it doesn't give actual constraints, since it can stay big for any time, on one hand, and on the other hand because it has passed first through a tick, which is forced to be 1 time unit apart from the event. Thus the monotonicity of time will ensure that big constraints are indeed semantically true. We define the scope of a constraint as the interval between the event and the next tick or reset, or equivalently between its begin and its end. The same vocabulary applies symmetrically for prediction operators. Actual constraints are either equalities (the time spend in their scope must be 1), linking an event to a tick, or inequalities (the time spend in their scope must be less than 1). An inequality is always linked to a small clock. Constraints can be partially ordered by scope: it is enough to solve constraints of maximal scope, as we shall see. A constraint of maximal scope always owns indexes: they are found at the end of its scope. The scope of an inequality extends from an event to a reset. Whether an atom A i is in the scope of a constraint, and which, can be deduced from its contents. The table below shows the contents of an atom A i that is the end of an equality. We distinguish the prediction and history cases. The table is simplified by the fact that we can assume that events are closed. The begin atom is the closest one in the past to contain the indicated formulae. Table Equality constraints - ticking clocks begin end in A i . =1 OE (tick) OE; :OES . =1 OE (event) The table below shows the contents of an atom A i indicating that the clock is small. It is thus in the scope of a constraint, whose begin is before and whose end is after. The begin (resp. end) is the closest atoms with the indicated contents. Table Small clocks begin in A i end . (tick or reset) Note that the existence of the begin and ends is guaranteed by fig. 2: a clock cannot stay small forever. In this section, we furthermore enforce that it will not stay small more than 1 unit of time. The proof shows that these constraints can be solved iff they are compatible in the sense that the scope of an equality cannot be included in the scope of an inequality, nor strictly in the scope of another equality. The axioms for several clocks ensure this compatibility. The previous section has built a run I), where - I is irrelevant, that is qualitatively correct. From any such run I), we now build a well-timed run J) by attributing a well-chosen sequence of intervals to the atoms of the run, so as to satisfy the real-time constraints. Before, we introduce two lemmas on which the algorithm relies, that can also be read from fig. 2: Lemma 23 For every run I) of the structure - \Pi, we have that if Proof. This lemma is a direct consequence of the mirrors of axioms (14) and (17). \Xi Lemma 24 For every run I) of the structure - \Pi, we have that if -:/;/;:/S . =1 / 2 A i then there exists Proof. This lemma is a direct consequence of the mirror of axiom (10). \Xi The algorithm proceeds by induction along the run, attributing time points As a consequence, an open interval (t attributed when i is odd: we don't mention it, and just define t i for even i. i.e. we attribute the interval [0; 0] to the initial atom A 0 . (2) Induction: we identify and solve the tightest constraint containing i. We define b as the begin of this tightest constraint, by cases: (a) equality constraints: (i) If there is an / =1 / 2 A i there has been a last (singular) atom A b containing / before at time t b . (ii) Else, if -:/;/;:/S . =1 / 2 A i there has been a last atom A b containing . =1 / before A i , at time t b . We set t (b) If there are no equality constraints, we consider inequality constraints: (i) We compute the earliest begin b of the small clocks using table 2. t i has to be between t i\Gamma2 and t b + 1. We choose t 1)=2. (ii) Otherwise, we attribute (say) t i\Gamma2 + 1=2 to A i . The algorithm selects arbitrarily an equality constraint, but is still deterministic Lemma 25 If two equality constraints have the same end i, their begins are identical. Proof. Four combinations of equality constraints are possible: (1) The first constraint is / =1 OE (a) The second constraint is / =1 /: A i contains thus /-1 / by (14). We apply (26) to obtain :OES/. We repeat this with /; OE inverted to obtain :/SOE. These formulae imply by the mirror of Lemma 19 that / cannot occur before OE, and conversely, thus they occur in the same atom. (b) The second constraint is the event /; :-/ with :/S . =1 /: then A i contains /-1 OE by (14). We apply (29) to obtain : . =1 /SOE. Since A i contains :/U - / =1 OE since its eventuality / =1 OE is true now. We apply (28) to obtain :OEZ(. we know that the tick occurs first (perhaps ex-aequo) among the possibilities that end the Z. These formulae imply by Lemma 19 that . =1 / cannot occur before OE, and conversely, thus they occur in the same atom. (2) The first constraint is the event OE with :OES . =1 OE 2 A (a) The second constraint is / This case is simply the previous one, with OE; / inverted. (b) The second constraint is the event / with :/S . =1 /: A i contains since its eventuality OE is true now. We apply (27) to obtain By :/S . =1 /, the tick . =1 / occurred first. We repeat this with /; OE inverted. These formulae imply by Lemma 19 that . =1 / cannot occur before . =1 OE, and conversely, thus they occur in the same atom. Solving an equation at its end also solves current partial inequations: Lemma 26 If A i is in the scope of an inequation, and the end of an equation, then the begin A j of the inequation is after the begin A b of the equation (b ! j). Proof. There are 3 possible forms of inequations in A i (see table 4.2): its begin, i.e. =1 / 2 A j . We must show that b ! j. The equation can be: (a) thus . The first case is true as by hypothesis :/S must occur before / in the past), and gives b - j. (b) OE; :OES . =1 OE 2 A i and . =1 OE 2 A b : using (27), we obtain : . =1 OEZ(. . The first case is true, by hypothesis, and gives b - j. We cannot assume because the mirror of lemma 25 then gives its begin (its event), i.e. / 2 A j . We must show that b ! j. The equation can be: (a) We apply (26) to obtain :OES/, meaning by the mirror of lemma 19 that b - j. :/SOE 62 A i , for otherwise we apply (30) yielding /!1 OE 2 contradicting / =1 OE 2 A i by (15), so we conclude b ! j. (b) OE; :OES . =1 OE 2 A i and . =1 OE 2 A b : by j. We cannot have the reverse :/S . =1 OE, for otherwise we apply the mirror of (31) and deduce its begin (a reset). Either .!1 / 2 A i already, or if the event is in A i , we use axiom (23) to show . Since there is no intervening / between j and i, the fig.2 implies .!1 / 2 A j+1 and thus Because :(:/S . =1 /) 2 A i , we deduce .!1 / 2 A j . Now, we must show that b ! j. The equation can be: (a) / =1 OE 2 A i and its event OE 2 A b : As apply (28) to obtain :OEZ(. means b - j. Again because there are no intervening / between j and i, using lemma 19 we have :/U / =1 OE 2 A j . Using the mirror of (31), /!1 OE; :OE 2 A j , thus is impossible, since :OE 2 A j and . We conclude b ! j. (b) OE; :OES . =1 OE 2 A i and . =1 OE 2 A b : so :/U - OE 2 A i , and we use (27) to obtain : . =1 OEZ(. A i . The reset / occurs strictly before the tick, so the first case is using because there are no intervening / between positions j and i, we have :/U / =1 OE 2 A j . Using the mirror of (30), .!1 OE 2 A j . The second case is thus true, and means b - j. impossible, since . We conclude b ! j. We now show that the algorithm Attr assigns time bounds of intervals that are increasing. Lemma 27 The sequence t i built by Attr is increasing. Proof. In the notation of the definition, this amounts to prove t when b is defined, since t i is either (in the case of an equality) or the middle point of (in the case of an inequality). If b is not defined (no constraints) then it is trivially verified as we attribute t i\Gamma2 + 1=2 to t i . We prove the non trivial cases by induction on i: (1) base case: 2. Either: (a) no constraint is active, b is undefined; just have to prove 0 ! 1. (2) induction: We divide in cases according to the constraint selected at whose begin is called (a) an equality: by lemmas 25, 26, its begin was before, i.e., b i\Gamma2 ! b. By inductive hypothesis, t i is increasing: t b i\Gamma2 ! t b . Thus t (b) an inequality: Thus the begin b i\Gamma2 - b i , since it was obtained by sorting. By inductive hypothesis, t i is increasing: so t b i\Gamma2 . By inductive hypothesis, t +1. Thus t Furthermore, the algorithm Attr ensures that time increases beyond any bounds: Lemma 28 The sequence of intervals - J of J) built by our algorithm has finite variability: for all t there exists an i - 0 such that Proof. Although there is no lower bound on the duration of an interval, we show that the time spend in each passage through the final cycle of - is at least 1=2. Thus any real number t will be reached before index 2tc, where c is the number of atoms in the final cycle. We divide in cases: (1) If the cycle A n A contains an atom which is not in the scope of any constraint, the time spent there will be 1=2. (2) Else, the cycle contains constraints, and thus constraints of maximal scope. This scope, however, cannot be greater than one cycle. Let e the end of such a constraint. Thus e is in the scope of no other constraint with an earlier begin. The time spent in the scope of the constraint until i is at least 1=2: Let again b be the begin of the scope of the constraint. t e\Gamma2 - t b (since the begin and end are singular and distinct), thus our algorithm gives 1=2. Since the scope cannot be greater than one cycle, the time spent in a cycle is at least 1=2. This procedure correctly solves all constraints: Lemma 29 The interval attribution Attr transforms any run ae in a well-timed run Attr(ae). Proof. We show the two supplementary properties of a well-timed run: We must show that the next / occurs in t \Gamma I. / I / can be: (a) /?1 /: These constraints are automatically satisfied because: (i) the mirror of the eventuality rule (17) guarantees / has occurred. us take the first such j, which is the corresponding event. (ii) According to fig.2, / will stay false, and eventually we will reach (iii) the axiom (25) guarantees that satisfying the equality will entail satisfying the greater-than constraint, since they refer to the same tracked event, and since the equality is later. In formulae, for any t i 2 I i , (b) / =1 /: Since this is an equality constraint, the algorithm Attr must have chosen an equality constraint with begin b. Thus t 1. By lemma 25, the begin event OE is also in A b . (c) /!1 /: If i isn't even (singular), we know that the constraint will still be active in the next atom because the end of a constraint is always singular. By (22): ffl It might become an equality (the clock may tick), in which case it is treated as in the previous case (with i+1 instead of i). Then the monotonicity of time will ensure that I ffl If it is still the same inequality, it is treated below (with instead of i). Then the monotonicity of time will ensure that I Thus at this point we can assume that i is even. Let j ! i be the begin of the constraint, OE 2 A j . The constraint selected by Attr at i can be: (i) an equality: by lemma 26, its begin b ! j, so that t (ii) or the constraint chosen in A i is an inequality. The pair /!1 / 2 is also an inequality in A i : let f be its begin. The algorithm has selected the constraint with the earliest begin b. (2) Let . I / 2 Very similarly, we must show that the next / occurs I / can be: (a) .?1 /: These constraints are automatically satisfied because: (i) the eventuality rule (17) guarantees / will occur: 9j ? i A j . We take the first such j, which is the corresponding event. We can assume it is singular. Figure 2 guarantees that there is first a tick: 9k (iii) the reset rule (25) guarantees that satisfying the equality will entail satisfying the greater-than constraint, since they refer to the same end event, and since the equality is later. In formulae, for any t i 2 I i , (b) . =1 /: let A j contain the next event of /. Since this is an equality con- straint, the algorithm Attr must have chosen an equality constraint at A j . By lemma 25, its begin is i. Thus t (c) .!1 /: Let A j contain the next event of /. The constraint selected by Attr at j can be: (i) an equality: by lemma 26 its begin b ! i, so that t (ii) or the constraint chosen in A j is an inequality. The pair .!1 / 2 is also an inequality in A j : let f be its begin. The algorithm has selected the constraint with the earliest begin b. The reader now expects a proof for the converse implication. This is not needed thanks to (43). \Xi As a consequence of the last lemmas, we have: Lemma timed run built by Attr has the Hintikka property for Event- Finally, we obtain the desired theorem: Theorem 31 Every EventClockTL-consistent formula is satisfiable. Proof. If ff is a EventClockTL-consistent formula then there exists an ff-monitored atom A ff in - \Pi. By lemma 16, there exists a set of runs \Sigma that pass through A ff and by the properties of the procedure Attr, lemma 18, lemma 28 and lemma 29, at least one run ( - has the Hintikka property for Event- ClockTL. It is direct to see that ( - I) is a model for ff at time t 2 I ff (the interval of time associated to A ff in ( - I) ) and thus ff is satisfiable. \Xi Corollary 32 The rule (1) and axioms (2)-(31) form a complete axiomatization of EventClockTL. 4.3 Comparison with automata construction In spirit, the procedure given above can be considered as building an automaton corresponding to a formula. The known procedures [3] for deciding use a similar construction, first building a timed automaton and then its region automaton. We could not use this construction directly here, because it involves features of automata that have no counterpart in the logic, and thus could not be expressed by axioms. However, the main ideas are similar. The region automaton will record the integer value of each clock: we code this by formulae of the form .!1 . =1 ::: . =1 OE. It will also record the ordering of the fractional parts of the clocks: this is coded here by formulae of the form :. =1 :::. =1 OEU. =1 :::. =1 /. There are some small differences, however. For simplicity we maintain more information than needed. For instance we record the ordering of any two ticks, even if these ticks are not linked to the current value of the clock. This relationship is only inverted for a very special case: when a clock has no previous and no following tick, we need not and cannot maintain its fractional information. It is easy to build a more careful and more efficient tableau procedure, that only records the needed information. The structure of atoms constructed here treats the eventualities in a different spirit than automata: here, there may be invalid paths in the graph of atoms. It is immediate to add acceptance conditions to eliminate them and obtain a more classical automaton. But it is less obvious to design a class of automata that is as expressive as the logic: this is done in [14]. 4.4 Other time domains As we have already indicated incidentally, our proofs are written to adapt to other time domains Twith minimal change. We only consider totally ordered dense time, however. For instance, we could use as time domain: (1) The real numbers, T= R: We replace (12) by the mirror of (13). (2) The rational numbers, T= Q: If we force the bounds of an interval to be rational as well, nothing has to be changed. Otherwise, a transition from an open interval to an open interval is now possible, if the common bound is irrational. This defeats the induction axiom (11). We postpone the study of this case to a further paper, but the basic ideas of the proof still apply. (3) A bounded real interval: (a) closed For the qualitative part, we replace (13) by the mirror of (12). For the quantitative part, we first remove the axiom (25). If the duration of the interval stating that the beginning is at distance d from the end. Otherwise, we add the best approximation of this: (b) open For the qualitative part, we replace (13) by the mirror of (12): from a qualitative point of view, an open interval is indistinguishable from an infinite one. 5 Translating MetricIntervalTL into EventClockTL The logics have been designed from a different philosophical standpoint: MetricIn- restricts the undecidable logic MetricTL by "relaxing punctuality", i.e., forbidding to look at exact time values; EventClockTL, in contrast, forbids to look past the next event in the future. However, we have shown in [14] that, surprisingly, they have the same expressive power. The power given by nesting connectives allows to each logic to do some of its forbidden work. Here, we need more than a mere proof of expressiveness, we need a finite number of axioms expressing the translation between formulae of the two logics. We give below both the axioms and a procedure that use them to provide a proof of the equivalence. First, we suppress intervals containing 0: U I / U J /) with Then we replace bounded untils - U I with 0 62 I by simpler \Sigma I , provided U I / where l is the left endpoint of I, the intervals Ig. We suppress classical until using: For infinite intervals, we reduce the lower bound l ? 0 to 0 using For finite intervals with left bound equal to 0, we exclude it if needed with (49), and we use the . operator: Note that the formulae .!u OE and .-u OE can be reduced to formulae that only use constant 1 using the axioms (18) and (19). When the left bound of the interval is different from 0 and the right bound different from 1, we reduce the length of the interval to 1 using: Then we use the following rules recursively until the lower bound is reduced to 0: In this way, any MetricIntervalTL formula can be translated into a Event- formula where bounds are always 0 or 1. Actually, we used a very small part of EventClockTL; we can further eliminate .!1 OE: showing that the very basic operators . have the same expressive power as full MetricIntervalTL. The converse translation is much simpler: . I OE $ :\Sigma !I OE - \Sigma Inf0g OE (62) OEU/ 5.1 Axiomatization of MetricIntervalTL To obtain an axiom system for MetricIntervalTL, we simply translate the axioms of EventClockTL and add axioms expressing the translation. Indeed, we have translations in each direction: Therefore, to prove a MetricIntervalTL formula -, we translate it into Event- and prove it there using the procedure of section 4. The proof - can be translated back to MetricIntervalTL in T (-) proving T (S(-)). Indeed, each step is a replacement, and replacements are invariant under syntax-directed translation preserving equivalence: To finish the proof we only have to add T (S(-)) - . Actually the translation axioms above are stronger, stating T (S(-. In our case, T (defined by (62), (63)) is so simple that it can be considered as a mere shorthand. Thus the axioms (1)-(29) and (49)-(60) form a complete axiomatization of MetricIntervalTL, with . I ; U now understood as shorthands. Theorem 33 The rule (1), axioms (2)-(29), and axioms (49)-(60) form a complete axiomatization of MetricIntervalTL. 6 Conclusion The specification of real-time systems using dense time is natural, and has many semantical advantages, but discrete-time techniques (here proof techniques [8,18]) have to be generalized. The model-checking and decision techniques have been generalized in [2,3]. Unfortunately, the technique of [3] uses a translation to automata which are more powerful and complex than temporal logic, and thus is not suitable for building a completeness proof. This paper provides complete axiom systems and proof-building procedures for linear real time, extending the technique of [19]. This procedure can be used to automate the proof construction of propositional fragments of a larger first-order proof. Some possible extensions of this work are: ffl The proof rules are admittedly cumbersome, since they exactly reflect the layered structure of the proof: for instance, real-time axioms are clearly separated from the qualitative axioms. More intuitive rules can be devised if we this constraint. This paper provides an easy way to show their com- pleteness: it is enough to prove the axioms of this paper. This also explains why we have not generalized the axioms, even when obvious generalizations are possible: we prefer to stick to the axioms needed in the proof, to facilitate a later completeness proof using this technique. ffl The logics used in this paper assume that concrete values are given for real-time constraints. As demonstrated in the HyTech checker [13], it is often useful to mention parameters instead (symbolic constants), and derive the needed constraints on the parameters, instead of a simple yes/no answer. ffl The extension of the results of this paper to first-order variants of MetricIn- should be explored. However, completeness is often lost in first-order variants [23]. ffl The development of programs from specifications should be supported: the automaton produced by the proposed technique might be helpful as a program skeleton in the style of [24]. --R the existence of refinement mappings. Model checking in dense real time. the benefits of relaxing punctuality. A really temporal logic. Logics and models of real time: a survey. A really abstract concurrent model and its temporal logic. Basic tense logic. Automatic verification of finite-state concurrent systems using temporal-logic specifications An axiomatization of the temporal logic with Until and Since over the real numbers. Semantics and completeness of duration calculus. the next generation. the regular real-time languages Tense Logic and the Theory of Order. A complete proof systems for QPTL. Specifying message passing and time-critical systems with temporal logic Checking that finite-state concurrent programs satisfy their linear specification the glory of the past. the anchored version of the temporal framework. Temporal Logic of Real-time Systems State clock logic: a decidable real-time logic Incompleteness of first-order temporal logic with until Synthesis of Communicating Processes from Temporal-Logic Specifications --TR Automatic verification of finite-state concurrent systems using temporal logic specifications Incompleteness of first-order temporal logic with until Temporal logic for real time systems Half-order modal logic: how to prove real-time properties The existence of refinement mappings Model-checking in dense real-time The benefits of relaxing punctuality Checking that finite state concurrent programs satisfy their linear specification A really abstract concurrent model and its temporal logic Specifying Message Passing and Time-Critical Systems with Temporal Logic The Regular Real-Time Languages State Clock Logic The Glory of the Past The anchored version of the temporal framework Logics and Models of Real Time Semantics and Completeness of Duration Calculus A Complete Proof Systems for QPTL HYTECH Synthesis of communicating processes from temporal logic specifications --CTR Carsten Lutz , Dirk Walther , Frank Wolter, Quantitative temporal logics over the reals: PSpace and below, Information and Computation, v.205 n.1, p.99-123, January, 2007

axiomatization;completeness;real time;temporal logic

507260

From rewrite rules to bisimulation congruences.

The dynamics of many calculi can be most clearly defined by a reduction semantics. To work with a calculus, however, an understanding of operational congruences is fundamental; these can often be given tractable definitions or characterisations using a labelled transition semantics. This paper considers calculi with arbitrary reduction semantics of three simple classes, firstly ground term rewriting, then left-linear term rewriting, and then a class which is essentially the action calculi lacking substantive name binding. General definitions of labelled transitions are given in each case, uniformly in the set of rewrite rules, and without requiring the prescription of additional notions of observation. They give rise to bisimulation congruences. As a test of the theory it is shown that bisimulation for a fragment of CCS is recovered. The transitions generated for a fragment of the Ambient Calculus of Cardelli and Gordon, and for SKI combinators, are also discussed briefly.

Introduction The dynamic behaviour of many calculi can be defined most clearly by a reduction semantics, comprising a set of rewrite rules, a set of reduction contexts in which they may be applied, and a structural congruence. These define the atomic internal reduction steps of terms. To work with a calculus, however, a compositional understanding of the behaviour of arbitrary subterms, as given by some operational congruence relation, is usually required. The literature contains investigations of such congruences for a large number of particular calculi. They are often given tractable definitions or characterisations via labelled transition relations, capturing the potential external interactions between subterms and their environments. Defining labelled transitions that give rise to satisfactory operational congruences generally requires some mix of calculus-specific ingenuity and routine work. In this paper the problem is addressed for arbitrary calculi of certain simple forms. We give general definitions of labelled transitions that depend only on a reduction semantics, without requiring any additional observations to be prescribed. We first consider term rewriting, with ground or left-linear rules, over an arbitrary signature but without a structural congruence. We then consider calculi with arbitrary signatures containing symbols 0 and j, a structural congruence consisting of associativity, commutativity and unit, left-linear rules, and non-trivial sets of reduction contexts. This suffices, for example, to express CCS-style synchronisation. It is essentially the same as the Computer Laboratory, University of Cambridge. Email: Peter.Sewell@cl.cam.ac.uk INTRODUCTION class of Action Calculi in which all controls have some number of arguments of In each case we define labelled transitions, prove that bisimulation is a congruence and give some comparison results. Background: From reductions to labelled transitions to reductions. Definitions of the dynamics (or small-step operational semantics) of lambda calculi and sequential programming languages have commonly been given as reduction relations. The -calculus has the rewrite rule (-x:M)N \Gamma!M [N=x] of fi reduction, which can be applied in any context. For programming lan- guages, some control of the order of evaluation is usually required. This has been done with abstract machines, in which the states, and reductions between them, are ad-hoc mathematical objects. More elegantly, one can give definitions in the structural operational semantics (SOS) style of Plotkin [Plo81]; here the states are terms of the language (sometimes augmented by e.g. a store), the reductions are given by a syntax-directed inductive definition. Explicit reformulations using rewrite rules and reduction contexts were first given by Felleisen and Friedman [FF86]. (We neglect semantics in the big-step/evaluation/natural style.) In contrast, until recently, definitions of operational semantics for process calculi have been primarily given as labelled transition relations. The central reason for the difference is not mathe- matical, but that lambda and process terms have had quite different intended interpretations. The standard interpretation of lambda terms and functional programs is that they specify computations which may either not terminate, or terminate with some result that cannot reduce further. Confluence properties ensure that such result terms are unique if they exist; they can implicitly be examined, either up to equality or up to a coarser notion. The theory of processes, however, inherits from automata theory the view that process terms may both reduce internally and interact with their environments; labelled transitions allow these interactions to be expressed. Reductions may create or destroy potential interactions. Termination of processes is usually not a central concept, and the structure of terms, even of terms that cannot reduce, is not considered examinable. An additional, more technical, reason is that definitions of the reductions for a process calculus require either auxiliary labelled transition relations or a non-trivial structural congruence. For example, consider the CCS fragment below. Its standard semantics has reductions P \Gamma!Q but also labelled transitions P ff \Gamma!Q and P - ff \Gamma!Q. These represent the potentials that P has for synchronising on ff. They can be defined by an SOS Out ff \Gamma!P In \Gamma!P Com ff Par \Gamma!Q \Gamma!Q \Gamma!Q \Gamma! is either \Gamma!, ff \Gamma! or - ff \Gamma!. It has been noted by Berry and Boudol [BB92], following work of Ban-atre and Le M'etayer [BM86] on the \Gamma language, that semantic definitions of process calculi could be simplified by working modulo an equivalence that allows the parts of a redex to be brought syntactically adjacent. Their presentation is in terms of Chemical Abstract Machines; in a slight variation we give a reduction semantics for the CCS fragment above. It consists of the rewrite rule Q, the set of reduction contexts given by and the structural congruence j defined to be the least congruence satisfying and use of j on the right, this gives exactly the same reductions as before. For this toy calculus the two are of similar complexity. For the -calculus ([MPW92], building on [EN86]), however, Milner has given a reduction semantics that is much simpler that the rather delicate SOS definitions of - labelled transition systems [Mil92]. Following this, more recent name passing process calculi have often been defined by a reduction semantics in some form, e.g. the HO- [San93], ae [NM95], Join [FG96], Blue [Bou97], Spi [AG97], dpi [Sew98b], D- [RH98] and Ambient [CG98] Calculi. Turning to operational congruences, for confluent calculi the definition of an appropriate operational congruence is relatively straightforward, even in the (usual) case where the dynamics is expressed as a reduction relation. For example, for a simple eager functional programming language, with a base type Int of integers, terminated states of programs of type Int are clearly observable up to equality. These basic observations can be used to define a Morris-style operational congruence. Several authors have considered tractable characterisations of these congruences in terms of bisimulation - see e.g. [How89, AO93, Gor95] and the references therein, and [GR96] for related work on an object calculus. For non-confluent calculi the situation is more problematic - process calculi having labelled transition semantics have been equipped with a plethora of different operational equivalences, whereas rather few styles of definition have been proposed for those having reduction semantics. In the labelled transition case there are many more-or-less plausible notions of observation, differing e.g. in their treatment of linear/branching time, of internal reductions, of termination and divergence, etc. Some of the space is illustrated in the surveys of van Glabbeek [Gla90, Gla93]. The difficulty here is to select a notion that is appropriate for a particular application; one attempt is in [Sew97]. In the reduction case we have the converse problem - a reduction relation does not of itself seem to support any notion of observation that gives rise to a satisfactory operational congruence. This was explicitly addressed for CCS and -calculi by Milner and Sangiorgi in [MS92, San93], where barbed bisimulation equivalences are defined in terms of reductions and observations of barbs. These are vestigial labelled transitions, similar to the distinguished observable transitions in the tests of De Nicola and Hennessy [DH84]. The expressive power of their calculi suffices to recover early labelled transition bisimulations as the induced congruences. Related work of Honda and Yoshida [HY95] uses insensitivity as the basic observable. .to labelled transitions Summarizing, definitions of operational congruences, for calculi having reduction semantics, have generally been based either on observation of terminated states, in the confluent case, or on observation of some barbs, where a natural definition of these exists. In either case, characterisations of the congruences in terms of labelled transitions, involving as little quantification over contexts as possible, are desirable. Moreover, some reasonable calculi may not have a natural definition of barb that induces an appropriate congruence. In this paper we show that labelled transitions that give rise to bisimulation congruences can be defined purely from the reduction semantics of a calculus, without prescribing any additional observations. It is preliminary work, in that only simple classes of reduction semantics, not involving name or variable binding, will be considered. As a test of the definitions we show that they recover the usual bisimulation on the CCS fragment above. We also discuss term rewriting and a fragment of the Ambient calculus of Cardelli and Gordon. To directly express the semantics of more interesting calculi requires a richer framework. One must deal with binding, with rewrite rules involving term or name substitutions, with a structural congruence that allows scope mobility, and with more delicate sets of reduction contexts. The Action Calculi of Milner [Mil96] are a candidate framework that allows several of the calculi mentioned above to be defined cleanly; this work can be seen as a step towards understanding operational congruences for arbitrary action calculi. Labelled transitions intuitively capture the possible interactions between a term and a surrounding context. Here this is made explicit - the labels of transitions from a term s will be contexts that, when applied to s, create an occurrence of a rewrite rule. A similar approach has been followed by Jensen [Jen98], for a form of graph rewriting that idealizes action calculi. Bisimulation for a particular action calculus, representing a -calculus, has been studied by Mifsud [Mif96]. In the next three sections we develop the theory for ground term rewriting, then for left-linear term rewriting, and then with the addition of an AC1 structural congruence and reduction contexts. Section 5 contains some concluding remarks. Most proofs are omitted, but can be found in the technical report [Sew98a]. Ground term rewriting In this section we consider one of the simplest possible classes of reduction semantics, that of ground term rewriting. The definitions and proofs are here rather straightforward, but provide a guide to those in the following two sections. Reductions We take a signature consisting of a set \Sigma of function symbols, ranged over by oe, and an arity function j j from \Sigma to N. Context composition and application of contexts to (tuples of) terms are written A : B and A : s, the identity context as and tupling with +. We say an n-hole context is linear if it has exactly one occurrence of each of its holes. In this section a; b; l; range over terms, A; B; C; D;F; H range over linear unary contexts and E ranges over linear binary contexts. We take a set R of rewrite rules, each consisting of a pair hl; ri of terms. The reduction relation is then Labelled Transitions The transitions of a term s will be labelled by linear unary contexts. Transitions s\Gamma!t labelled by the identity context are simply reductions (or -transitions). Transitions \Gamma!t for F 6j indicate that applying F to s creates an instance of a rewrite rule, with target instance t. For example, given the rule we will have labelled transitions for all C and The labels are f F j 9hl; ri 2 R; s and the contextual labelled transition relations F \Gamma! are defined by: \Gamma!t ri Bisimulation Congruence Let - be strong bisimulation with respect to these transitions. The congruence proof is straightforward. It is given some detail as a guide to the more intricate corresponding proofs in the following two sections, which have the same structure. Three lemmas show how contexts in labels and in the sources of transitions interrelate; they are proved by case analysis using a dissection lemma which is standard folklore. then one of the following cases holds. 1. (b is in a) There exists D such that a 2. (a is properly in b) There exists D with D 6= such that D : 3. (a and b are disjoint) There exists E such that then one of the following holds: 1. There exists some H such that s. 2. There exists some - t, A 1 and A 2 such that t. Proof By the definition of reduction ri Applying the dissection lemma (Lemma 1) to A : l gives the following cases. 1. (l is in s) There exists B such that Taking the second clause holds. 2. (s is properly in l) There exists B with B 6= such that Taking the second clause holds. 3. (s and l are disjoint) There exists E such that Taking r) the first clause holds.Lemma 3 If A : s F \Gamma!t and F 6= then s F : A \Gamma!t. Proof By the definition of labelled transitions ri linear \Gamma!t. 2 Lemma \Gamma!t then A : s F \Gamma!t. Proof so the conclusion is immediate, otherwise by the definition of transitions ri One then has A : s F \Gamma!t by the definition of transitions, by cases for F 6= and Proposition 5 - is a congruence. Proof We show is a bisimulation. 1. Suppose A : s\Gamma!t. By Lemma 2 one of the following holds: (a) There exists some H such that s. Hence (b) There exists some - t, A 1 and A 2 such that t. By s - s 0 there exists - t 0 such that s 0 A2 t. By the definition of reduction 2. Suppose A : s F \gamma!t for F 6= . \gamma!t. By s - s 0 there exists t 0 such that s t. \gamma!t 0 . alternative approach would be to take transitions for unary linear contexts F . Note that these are defined using only the reduction relation, whereas the definition above involved the reduction rules. Let - alt be strong bisimulation with respect to these transitions. One can show that - alt is a congruence and moreover is unaffected by cutting down the label set to that considered above. In general - alt is strictly coarser than -. For an example of the non-inclusion, if the signature consists of constants ff; fi and a unary symbol fl with reduction rules ff\gamma!ff, fi \gamma!fi and fl(fi)\gamma!fi, then ff 6- fi whereas ff - alt fi. This insensitivity to the possible interactions of terms that have internal transitions suggests that the analogue of - alt , in more expressive settings, is unlikely to coincide with standard bisimulations for particular calculi. Indeed, one can show that applying the alternative definition to the fragment of CCS ff ff (with its usual reduction relation) gives an equivalence that identifies ff j - ff with fi. Remark In the proofs of Lemmas 2-4 the labelled transition exhibited for the conclusion involves the same rewrite rule as the transition in the premise. One could therefore take the finer transitions F annotated by rewrite rules, and still have a congruence result. In some cases this gives a finer bisimulation relation. Remark The labelled transition relation is linear in R, i.e. the labelled transitions generated by a union of sets of rewrite rules are just the union of the relations generated by R 1 and R 2 . rewriting with left-linear rules In this section the definitions are generalised to left-linear term rewriting, as a second step towards a framework expressive enough for simple process calculi. Notation In the next two sections we must consider more complex dissections of contexts and terms. It is convenient to treat contexts and terms uniformly, working with n-tuples of m-hole contexts for m;n - 0. Concretely, we work in the category C \Sigma that has the natural numbers as objects and morphisms The identity on m is id m composition is substitution, with an [b strictly associative binary products, written with +. If a : m! k and b : m! l we write a \Phi b for l. Angle brackets and domain subscripts will often be elided. We let a; b; e; q; range over 0 !m morphisms, i.e. m-tuples of terms, range over m! 1 morphisms, i.e. m-hole contexts, and - over projections and permutations. Say a morphism linear if it contains exactly one occurrence of each if it contains at most one occurrence of each. We sometimes abuse notation in examples, writing Remark Many slight variations of C \Sigma are possible. We have chosen to take the objects to be natural numbers, instead of finite sets of variables, to give a lighter notation for labels. The concrete syntax is chosen so that morphisms from 0 to 1 are exactly the standard terms over \Sigma, modulo elision of the angle brackets and subscript 0. Reductions The usual notion of left-linear term rewriting is now expressible as follows. We take a set R of rewrite rules, each consisting of a triple hn; L; Ri where n - 0, linear and 1. The reduction relation over f s is then defined by Labelled Transitions The labelled transitions of a term s again be of two forms, s\gamma!t, for internal reductions, and s F \gamma!T where F 6= is a context that, together with part of s, makes up the left hand side of a rewrite rule. For example, given the rule we will have labelled transitions for all terms s Labelled transitions in which the label contributes the whole of the left hand side of a rule would be redundant, so the definition will exclude e.g. s ffi(fl( \gamma! ffl(s). Now consider the rule As before there will be labelled transitions for all s. In addition, one can construct instances of the rule by placing the term ff in contexts suggesting labelled transitions ff oe( ;fl(t)) \gamma! ffl(t) for any t. Instead, to keep the label sets small, and to capture the uniformity in t, we allow both labels and targets of transitions to be parametric in un-instantiated arguments of the rewrite rule. In this case the definition will give In general, then, the contextual labelled transitions are of the form s F \gamma!T , for 1. The first argument of F is the hole in which s can be placed to create an instance of a rule L; the other n arguments are parameters of L that are not thereby instantiated. The transitions are defined as follows. , s\gamma!T . \gamma!T , for linear and not the identity, iff there exist permutation linear and not the identity such that The definition is illustrated in Figure 1. The restriction to L 1 6= id 1 excludes transitions where the label contributes the whole of L. The permutation - is required so that the parameters of L can be divided into the instantiated and uninstantiated. For example the rule F R nn s Figure 1: Contextual Labelled Transitions for Left-Linear Term Rewriting. Boxes with m input wires (on their right) and n output wires (on their left) represent n-tuples of m-hole contexts. Wires are ordered from top to bottom. will give rise to transitions (The last is redundant; it could be excluded by requiring - to be a monotone partition of m into Bisimulation Congruence A binary relation S over terms f a j a is lifted to a relation by A [S] A 0 def Say S is a bisimulation if for any s S s 0 and write - for the largest such. As before the congruence proof requires a simple dissection lemma and three lemmas relating contexts in sources and labels. Lemma 6 (Dissection) If A : then one of the following holds. 1. (a is not in any component of b) There exist linear and not the identity such that i.e. there are m 1 components of b in a and m 2 in A. 2. (a is in a component of b) m - 1 and there exist partition such that linear then one of the following holds. 1. There exists some 2. Lemma 8 If A : s F one of the following holds. 1. There exists H : 1+n! 1 such that id n ). 2. There exist such that s F Lemma 9 If s \gamma! T for linear then for all Theorem 1 - is a congruence. Proof We show S , where is a bisimulation. First note that for any A : To see this, take linear such that An Each A i is linear, so A We now show that if A \gamma!T then there exists T 0 such that 1. Suppose A : s\gamma!t. By Lemma 7 one of the following holds: (a) There exists some Hence (b) There exist By s - s 0 there exists T 0 such that s 0 F By the definition of reduction A : s 2. Suppose A : s F linear and F 6= id 1 . By Lemma 8 one of the following holds. (a) There exists Hence (b) There exist such that s F By s - s 0 there exists - Now if for A i linear and s F then by the above there exists Tn such that F definition reduces to that of Section 2 if all rules are ground. For open rules, instead of allowing parametric labels, one could simply close up the rewrite rules under instantiation, by apply the earlier definition. In general this would give a strictly coarser congruence. For an example of the non-inclusion, take a signature consisting of a nullary ff and a unary fl, with R consisting of the rules fl( )\gamma!fl( ) and fl(fl(ff))\gamma!fl(fl(ff)). We have g. The transitions are for m;n - 1, so fl(ff) 6- R fl(fl(ff)) but fl(ff) - Cl(R) fl(fl(ff)). Proposition Comparison Bisimulation as defined here is a congruence for arbitrary left-linear term rewriting systems. Much work on term rewriting deals with reduction relations that are confluent and ter- minating. In that setting terms have unique normal forms; the primary equivalence on terms is ', have the same normal form. This is easily proved to be a congruence. In general, it is incomparable with -. To see one non-inclusion, note that - is sensitive to atomic reduction steps; for the other that - is not sensitive to equality of terms - for example, with only nullary symbols ff; fi; fl, and rewrite rule fl \gamma!fi, we have ff - fi and fi ' fl, whereas ff 6' fi and fi 6- fl. One might address the second non-inclusion by fiat, adding, for any value v, a unary test operator H v and reduction rule H v (v)\gamma!v. For the first, one might move to a weak bisimulation, abstracting from reduction steps. The simplest alternative is to take - to be the largest relation S such that if s S s 0 then and symmetric clauses. Say the set R of rewrite rules is right-affine if the right hand side of each rule is affine. Under this condition - is a congruence; the result without it is left open. Theorem 2 If R is right-affine then - is a congruence. Example - Integer addition For some rewrite systems - coincides with '. Taking a signature comprising nullary z for each integer z and binary plus and ifzero, and rewrite rules for all integers x and z gives labelled transitions x plus( ;z) together with the reductions \gamma!. Here the normal forms are simply the integers; - and ' both coincide with integer equality. In general, however, - is still incomparable with '. For example, with unary ffi; fl, nullary ff, and rules fl(ff)\gamma!ff, ffi(ff)\gamma!ff, and ffi(fl( ))\gamma! , we have ff 6- fi(ff). This may be a pathological rule set; one would like to have conditions excluding it under which - and ' coincide. Example - SKI Combinators Taking a signature \Sigma comprising nullary I ; K;S and binary ffl, and rewrite rules gives labelled transitions I ffl 1 together with some permutation instances of these and the reductions \gamma!. The significance of - and - here is unclear. Note that the rules are not right-affine, so Theorem 2 does not guarantee that - is a congruence. It is quite intensional, being sensitive to the number of arguments that can be consumed immediately by a term. For example, K rewriting with left-linear rules, parallel and boxing In this section we extend the setting to one sufficiently expressive to define the reduction relations of simple process calculi. We suppose the signature \Sigma includes binary and nullary symbols j and 0, for parallel and nil, and take a structural congruence j generated by associativity, commutativity and identity axioms. Parallel will be written infix. The reduction rules R are as before. We now allow symbols to be boxing, i.e. to inhibit reduction in their arguments. For each oe 2 \Sigma we suppose given a set B(oe) ' defining the argument positions where reduction may take place. We require 2g. The reduction contexts C ' f C linear are generated by Formally, structural congruence is defined over all morphisms of C \Sigma as follows. It is a family of relations indexed by domain and codomain arities; the indexes will usually be elided. Reductions The reduction relation over f s is defined by s\gamma!t iff This class of calculi is essentially the same as the class of Action Calculi in which there is no substantive name binding, i.e. those in which all controls K have arity rules of the form (here the a i are actions, not morphisms from C \Sigma ). It includes simple process calculi. For example, the fragment of CCS in Section 1 can be specified by taking a signature \Sigma CCS consisting of unary ff: and - ff: for each ff 2 A, with 0 and j, and rewrite rules Notation For a context f : m!n and i 2 1::m say f is shallow in argument i if all occurrences of in f are not under any symbol except j. Say f is deep in argument i if any occurrence of i in f is under some symbol not equal to j. Say f is shallow (deep) if it is shallow (deep) in all i 2 1::m. Say f is i-separated if there are no occurrences of any j in parallel with an occurrence of i . Labelled Transitions The labelled transitions will be of the same form as in the previous section, with transitions s F non-trivial label F may either contribute a deep subcontext of the left hand side of a rewrite rule (analogous to the non-identity labels of the previous section) or a parallel component, respectively with F deep or shallow in its first argument. The cases must be treated differently. For example, the rule TERM REWRITING WITH LEFT-LINEAR RULES, PARALLEL AND BOXING will generate labelled transitions As before, transitions that contribute the whole of the left hand side of a rule, such as s j ff j fi are redundant and will be excluded. It is necessary to take labels to be subcontexts of left hand sides of rules up to structural congruence, not merely up to equality. For example, given the rule we need labelled transitions Finally, the existence of rules in which arguments occur in parallel with non-trivial terms means that we must deal with partially instantiated arguments. Consider the rule The term -) j ae could be placed in any context oe( j s; t) to create an instance of the left hand side, with - (from the term) instantiating 1 , t (from the context) instantiating 2 , and ae j s (from both) instantiating 3 . There will be a labelled transition parametric in two places but partially instantiating the second by ae. The general definition of transitions is given in Figure 2. It uses additional notation - we write par n for h 1 and ppar n for n!n. Some parts of the definition are illustrated in Figure 3, in which rectangles denote contexts and terms, triangles denote instances of par, and hatched triangles denote instances of ppar. To a first approximation, the definition for F deep in 1 states that s F \gamma!T iff there is a rule L\gamma!R such that L can be factored into L 2 (with m 2 arguments) enclosing L 1 (with m 1 arguments) in parallel with m 3 arguments. The source s is L 1 instantiated by u, in parallel with e; the label F is roughly the target T is R with instantiated by u and m 3 partially instantiated by e. It is worth noting that the non-identity labelled transitions do not depend on the set of reduction contexts. The intended intuition is that the labelled transition relations provide just enough information so that the reductions of a term A : s are determined by the labelled transitions of s and the structure of A, which is the main property required for a congruence proof. A precise result, showing that the labelled transitions provide no extraneous information, would be desirable. Bisimulation Congruence Bisimulation - is defined exactly as in the previous section. As before, the congruence proof requires dissection lemmas, analogous to Lemmas 1 and 6, lemmas showing that if A : s has a transition then s has a related transition, analogous to Lemmas 2,3 and 7,8, and partial converses to these, analogous to Lemmas 4 and 9. All except the main dissection lemma are omitted here, but can be found in the long version. Lemma 11 (Dissection) If m - 0, with A and B linear, and A : a one of the following hold Transitions s F \gamma!T , for are defined by: ffl For F j id ffl For F deep in argument 1: s F \gamma!T iff there exist permutation linear and deep linear, deep in argument 1 and 1-separated such that ffl For F shallow in argument 1 and F 6j id \gamma!T iff there exist permutation linear and deep linear and deep such that Figure 2: Contextual Labelled Transitions em 3 e e R s F e R Deep Shallow Figure 3: Contextual Labelled Transitions Illustrated 1. (a is not deeply in any component of b) There exist linear and 1-separated linear and deep such that a j par 1+m3 There are m 1 of the b in a, m 2 of the b in A and m 3 of the b potentially overlapping A and a. The latter are split into e 1 , in a, and e 2 , in A. 2. (a is deeply in a component of b) m - 1 and there exist partition linear and deep such that The first clause of the lemma is illustrated in Figure 4. For example, consider A : a Clause 1 of the lemma holds, with This dissection should give rise to a transition Theorem 3 - is a congruence. Remark The definitions allow only rather crude specifications of the set C of reduction contexts. They ensure that C has a number of closure properties. Some reduction semantics require more delicate sets of reduction contexts. For example, for a list cons constructor one might want to allow is taken from some given set of values. This would require a non-trivial generalisation of the theory. A a Figure 4: Clause 1 of Dissection Lemma Example - CCS synchronization For our CCS fragment the definition gives \gamma! \gamma! together with structurally congruent transitions , i.e. those generated by and the reductions. Proposition 12 - coincides with bisimulation over the labelled transitions of Section 1. Proof Write - std for the standard bisimulation over the labelled transitions of Section 1. To show - std is a bisimulation for the contextual labelled transitions, suppose \gamma! T . There must exist u and r such that P j but then P ff so there exists Q 0 such that P 0 ff There must then exist u 0 and r 0 such that hence \gamma! . Using the fact that - std is a congruence we have so For the converse, suppose \gamma!Q. There must exist u and r such that P j ff:u j r and \gamma! so there exists T 0 such that P There must then exist u 0 and r 0 such that the definition of [ The standard transitions coincide (modulo structural congruence) with the contextual labelled transitions with their parameter instantiated by 0. One might look for general conditions on R under which bisimulation over such 0-instantiated transitions is already a congruence, and coincides with -. Example - Ambient movement The CCS fragment is degenerate in several respects - in the left hand side of the rewrite rule there are no nested non-parallel symbols and no parameters in parallel with any non-0 term, so there are no deep transitions and no partial instantiations. As a less degenerate example we consider a fragment of the Ambient Calculus [CG98] without binding. The signature \Sigma Amb has unary m[ ] (written outfix), in m:, out m: and open m:, for all m 2 A. Of these only the m[ ] allow reduction. The rewrite rules RAmb are open The definition gives the transitions below, together with structurally congruent transitions, permutation instances, and the reductions. in m:s j r n[ out open 5 Conclusion We have given general definitions of contextual labelled transitions, and bisimulation congruence results, for three simple classes of reduction semantics. It is preliminary work - the definitions may inform work on particular interesting calculi, but to directly apply the results they must be generalized to more expressive classes of reduction semantics. Several directions suggest themselves. Higher order rewriting Functional programming languages can generally be equipped with straightforward definitions of operational congruence, involving quantification over contexts. As discussed in the introduction, in several cases these have been given tractable characterisations in terms of bisimulation. One might generalise the term rewriting case of Section 3 to some notion of higher order rewriting [vR96] equipped with non-trivial sets of reduction contexts, to investigate the extent to which this can be done uniformly. Name binding To express calculi with mobile scopes, such as the -calculus and its descendants, one requires a syntax with name binding, and a structural congruence allowing scope extrusion. Generalising the definitions of Section 4 to the class of all non-higher-order action calculi would take in a number of examples, some of which currently lack satisfactory operational congruences, and should show how the indexed structure of - labelled transitions arises from the rewrite rules and structural congruence. Ultimately one would like to treat concurrent functional languages. In particluar cases it has been shown that one can define labelled transitions that give rise to bisimulation congruences, e.g. by Ferreira, Hennessy and Jeffrey for Core CML [FHJ96]. To express the reduction semantics of such languages would require both higher order rules and a rich structural congruence. Colouring The definition of labelled transitions in Section 4 is rather intricate - for tractable generalisations, to more expressive settings, one would like a more concise characterisation. A promising approach seems to be to work with coloured terms, in which each symbol except j and 0 is given a tag from a set of colours. This gives a notion of occurrence of a symbol in a term that is preserved by structural congruence and context application, and hence provides a different way of formalising the idea that the label of a transition s F \gamma!T must be part of a redex within F : s. Observational congruences We have focussed on strong bisimulation, which is a very intensional equivalence. It would be interesting to know the extent to which congruence proofs can be given uniformly for equivalences that abstract from branching time, internal reductions etc. More particularly, one would like to know whether Theorem 2 holds without the restriction to right-affine rewrite rules. One can define barbs for an arbitrary calculus by s # () 9F 6j id \gamma!T , so s # iff s has some potential interaction with a context. Conditions under which this barbed bisimulation congruence coincides with - could provide a useful test of the expressiveness of calculi. Structural operational semantics Our definitions of labelled transition relations are not inductive on term structure. Several authors have considered calculi equipped with labelled transitions defined by an SOS in some well-behaved format, e.g. [dS85, BIM95, GV92, GM98, TP97, Ber98]. The relationship between the two is unclear - one would like conditions on rewrite rules that ensure the labelled transitions of Section 4 are definable by a functorial operational semantics [TP97]. Conversely, one would like conditions on an SOS ensuring that it is characterised by a reduction semantics. Acknowledgements I would like to thank Philippa Gardner, Ole Jensen, S-ren Lassen, Jamey Leifer, Jean-Jacques L'evy, and Robin Milner, for many interesting discussions and comments on earlier drafts, and to acknowledge support from EPSRC grant GR/K 38403. --R A calculus for cryptographic protocols: The spi calculus. Full abstraction in the lazy lambda calculus. The chemical abstract machine. A congruence theorem for structured operational semantics of higher-order languages Bisimulation can't be traced. A new computational model and its discipline of programming. Mobile ambients. Testing equivalences for processes. A calculus of communicating systems with label-passing Control operators The reflexive CHAM and the join-calculus A theory of weak bisimulation for core CML. The linear time - branching time spectrum The linear time - branching time spectrum II The tile model. Bisimilarity as a theory of functional programming. Bisimilarity for a first-order calculus of objects with subtyping Structured operational semantics and bisimulation as a congruence. Equality in lazy computation systems. On reduction-based process semantics PhD thesis Control Structures. Functions as processes. Calculi for interaction. A calculus of mobile processes Barbed bisimulation. Constraints for free in concurrent computation. A structural approach to operational semantics. A typed language for distributed mobile processes. Expressing Mobility in Process Algebras: First-Order and Higher-Order Paradigms On implementations and semantics of a concurrent programming language. From rewrite rules to bisimulation congruences. Global/local subtyping and capability inference for a distributed Towards a mathematical operational semantics. Confluence and Normalisation for Higher-Order Rewriting --TR Equality in lazy computation systems The linear time-branching time spectrum (extended abstract) The chemical abstract machine Dynamic congruence vs. progressing bisimulation for CCS Structured operational semantics and bisimulation as a congruence A calculus of mobile processes, I Full abstraction in the lazy lambda calculus Turning SOS rules into equations Bisimulation can''t be traced On reduction-based process semantics A theory of weak bisimulation for core CML The reflexive CHAM and the join-calculus Bisimilarity for a first-order calculus of objects with subtyping The MYAMPERSANDpgr;-calculus in direct style A calculus for cryptographic protocols A typed language for distributed mobile processes (extended abstract) rewriting and all that The tile model Constraints for Free in Concurrent Computation Barbed Bisimulation Global/Local Subtyping and Capability Inference for a Distributed pi-calculus From Rewrite to Bisimulation Congruences A Categorical Axiomatics for Bisimulation The Linear Time - Branching Time Spectrum II On Implementations and Semantics of a Concurrent Programming Language Mobile Ambients Towards a Mathematical Operational Semantics A Congruence Theorem for Structured Operational Semantics of Higher-Order Languages --CTR Davide Grohmann , Marino Miculan, Directed Bigraphs, Electronic Notes in Theoretical Computer Science (ENTCS), 173, p.121-137, April, 2007 Henrik Pilegaard , Flemming Nielson , Hanne Riis Nielson, Active Evaluation Contexts for Reaction Semantics, Electronic Notes in Theoretical Computer Science (ENTCS), v.175 n.1, p.57-70, May, 2007 Vladimiro Sassone , Pawe Sobociski, Locating reaction with 2-categories, Theoretical Computer Science, v.333 n.1-2, p.297-327, 1 March 2005 Ole Hgh Jensen , Robin Milner, Bigraphs and transitions, ACM SIGPLAN Notices, v.38 n.1, p.38-49, January Hartmut Ehrig , Barbara Knig, Deriving bisimulation congruences in the DPO approach to graph rewriting with borrowed contexts, Mathematical Structures in Computer Science, v.16 n.6, p.1133-1163, December 2006 Massimo Merro , Francesco Zappa Nardelli, Behavioral theory for mobile ambients, Journal of the ACM (JACM), v.52 n.6, p.961-1023, November 2005

term rewriting;bisimuation;operational congruences;labelled transition systems;process calculi;operational sematics

507274

Communication complexity method for measuring nondeterminism in finite automata.

While deterministic finite automata seem to be well understood, surprisingly many important problems concerning nondeterministic finite automata (nfa's) remain open. One such problem area is the study of different measures of nondeterminism in finite automata and the estimation of the sizes of minimal nondeterministic finite automata. In this paper the concept of communication complexity is applied in order to achieve progress in this problem area. The main results are as follows:(1) Deterministic communication complexity provides lower bounds on the size of nfa's with bounded unambiguity. Applying this fact, the proofs of several results about nfa's with limited ambiguity can be simplified and presented in a uniform way. (2) There is a family of languages KONk2 with an exponential size gap between nfa's with polynomial leaf number/ambiguity and nfa's with ambiguity k. This partially provides an answer to the open problem posed by B. Ravikumar and O. Ibarra (1989, SIAM J. Comput. 18, 1263-1282) and H. Leung (1998, SIAM J. Comput. 27, 1073-1082).

Introduction In this paper the classical models of one-way nite automata (dfa's) and their nondeterministic counterparts (nfa's) [RS59] are investigated. While the structure and fundamental properties of dfa's are well understood, this is not the case for nfa's. For instance, we have ecient algorithms for constructing minimal dfa's, but the complexity of approximating the size of a minimal nfa is still unresolved (whereas nding a minimal nfa solves a PSPACE complete problem). Hromkovic, Seibert and Wilke [HSW97] proved that the gap between the length of regular expressions and the number of edges of corresponding nfa's is between n log 2 n and n log n, but the exact relation is unknown. Another principal open question is to determine whether there is an exponential gap between two-way deterministic nite automata and two-way nondeterministic ones. The last partially successful attack on this problem was done in the late seventies by Sipser [S80], who established an exponential gap between determinism and nondeterminism for so-called sweeping automata (the property of sweeping is essential [M80]). The largest known gap for the general case is quadratic [HS99]. Our main goal is to contribute to a better understanding of the power of nondeterminism in nite automata (see [RS59], [MF71], [Mo71], [Sc78] for very early papers on this topic). We focus on the following problems: 1. The best known method for proving lower bounds on the size of minimal nfa's is based on nondeterministic communication complexity [Hr97]. All other known methods are special cases of this method. Are there methods that provide better lower bounds at least for some languages? How can one prove lower bounds on the size of unambiguous nfa's (unfa's), that is nfa's which have at most one accepting computation for every word? 2. It is a well known fact [MF71], [Mo71] that there is an exponential gap between the sizes of minimal dfa's and nfa's for some regular languages. This is even known for dfa's and unfa's [Sc78], [SH85], [RI89], for unfa's and nfa's with constant ambiguity [Sc78], [RI89], and for ufa's with polynomial ambiguity and nfa's [HL98]. 1 But, it is open [RI89], [HL98] whether there exists an exponential gap between the sizes of minimal nfa's with constant ambiguity and nfa's with polynomial ambiguity. 3. The degree of nondeterminism is measured in the literature in three different ways. Let A be an nfa. The rst measure advice A (n) equals the number of advice bits for inputs of length n, i.e., the maximum number of nondeterministic guesses in computations for inputs of length n. The second measure leaf A (n) determines the maximum number of We apologize for claiming the above results as our contribution in the extended abstract of this paper [HKK00] instead of referring to [Sc78], [SH85], [RI89], [HL98] computations for inputs of length n. ambig A (n) as the third measure equals the maximum number of accepting computations for inputs of length at most n. Obviously the second and third measure may be exponential in the rst one. The question is whether the measures are more specically correlated. To attack these problems we establish some new bridges between automata theory and communication complexity. The communication complexity of two-party protocols was introduced by Yao [Y79] (and implicitly considered by Abelson [Ab78], too). The initial goal was to develop a method for proving lower bounds on the complexity of distributive and parallel computations (see, for instance, [Th79, Th80, Hr97, KN97]). Due to the well developed, nontrivial mathematical machinery for determining the communication complexity of concrete problems (see, for instance [AUY83, DHS96, Hr97, Hr00, KN97, L90, NW95, PS82]), communication complexity has established itself as a sub-area of complexity theory. The main contributions of the study of communication complexity lie especially in proving lower bounds on the complexity of specic problems, and in comparing the power of dierent modes of computation. Here, for the rst time, communication complexity is applied for the study of nondeterministic nite automata, with the emphasis on the tradeo between the size and the degree of nondeterminism of nfa's. Our procedure is mainly based on the following facts: (i) The theory of communication complexity contains deep results about the nature of nondeterminism (see, e.g. [KNSW94, HS96]) that use the combinatorial structure of the communication matrix as the computing problem representation. (ii) In [DHRS97, Hr97, HS00], the non-uniform model of communication protocols for computing nite functions was extended to a uniform model for recognizing languages in such a way that several results about communication complexity can be successfully applied for uniform computing models like automata. Combining (i) and (ii) with building of new bridges between communication complexity and nfa's we establish the following main results. 1. Let cc(L) resp. ncc(L) denote the deterministic resp. nondeterministic communication complexity of L. It is well known that 2 cc(L) and 2 ncc(L) are lower bounds on the sizes of the minimal dfa for L and a minimal nfa for L respectively. First we show that there are regular languages L for which there is an exponential gap between 2 ncc(L) and the minimal size of nfa's for L. This means, that the lower bound method based on communication complexity may be very weak. Then we show as a somewhat surprising result that 2 cc(L)=k 2 is a lower bound on the size of nfa's with ambiguity k for L. We furthermore show that Rank(M) 1=k 1 is a lower bound for the number of states for nfa's with ambiguity k, where M is a communication matrix associated with L. It is possible that this lower bound is always better than the rst one (see [KN97] for a discussion of the quality of the so-called rank lower bound on communication complexity). As a corollary we present a sequence of regular languages NIDm such that the size of a minimal nfa is linear in m, while the size of every unfa for NIDm is exponential in m. This substantially simplies the proofs of similar results in [Sc78], [SH85]. 2. We establish the relation advice A (n); ambig(n) A leaf A (n) O(advice A (n) ambig A (n)) for any minimal nfa A. Observe that the upper bound on leaf A (n) implies that minimal unambiguous nfa's may have at most O(advice A (n)) O(n) dierent computations on any input of size n, and an exponential gap between advice A (n) and leaf A (n) is possible only if the degree of ambiguity is exponential in n. Furthermore we show that leaf A (n) is always either bounded by a constant, or at least linear but polynomially bounded, or otherwise at least exponential in the input length. 3. We present another sequence of regular languages than in [HL98] with an exponential gap between the size of nfa's with exponential ambi- guity, and nfa's with polynomial ambiguity. This result is obtained by showing that small nfa's with polynomial ambiguity for the Kleene closure (L#) imply small unfa's that work correctly on a polynomial fraction of inputs. Our technique is more general than the proof method of Hing Leung [HL98] and provides an essentially shorter proof. Furthermore we describe a sequence of languages KON k 2 such that there is an exponential gap between the size of nfa's with polynomial ambiguity and nfa's with ambiguity k. This provides a partial answer to the open question [RI89], [HL98] whether there is an exponential gap between minimal nfa's with constant ambiguity and minimal nfa's with polynomial ambiguity. This paper is organized as follows. In section 2 we give the basic deni- tions and x the notation. In order to increase the readability of this paper for readers who are not familiar with communication complexity theory, we give more details about communication protocols and build the basic intuition of their relation to nite automata. Section 3 is devoted to the investigation of the relation between the size of nfa's and communication complexity. Section 4 studies the relation between dierent measures of nondeterminism in nite automata, and presents the remaining results. Denitions and Preliminaries We consider the standard one-way models of nite automata (dfa's) and nondeterministic nite automata (nfa's). For every automaton A, L(A) denotes the language accepted by A. The number of states of A is called the size of A and denoted size A . For every regular language L we denote the size of the minimal dfa for L by s(L) and the size of minimal nfa's accepting L by ns(L). For every alphabet , ng and ng. For any nfa A and any input x we use the computation tree T A;x to computations of A on x. Obviously the number of leaves of T A;x is the number of dierent computations of A on x. The ambiguity of an nfa A on input x is the number of accepting computations of A on x, i.e., the number of accepting leaves of T A;x . If the nfa A has ambiguity one for all inputs, then A is called an unambiguous nfa (unfa) and uns(L) denotes the size of a minimal unfa accepting L. More generally, if an nfa A has ambiguity at most k for all inputs, then A is called a k-ambiguous nfa and ns k (L) denotes the size of a minimal k-ambiguous nfa accepting L. For every nfa A we measure the degree of nondeterminism as follows. Let denote the alphabet of A. For every input x 2 and for every computation C of A on x we dene advice(C) as the number of nondeterministic choices during the computation C, i.e., the number of nodes on the path of C in T A;x which have more than one successor. Then advice A is a computation of A on xg and advice A For every x 2 we dene leaf A (x) as the number of leaves of T A;x and set leaf A For every x 2 we dene ambig A (x) as the number of accepting leaves of T A;x and set ambig A Since a language need not contain words of all lengths we dene ambiguity over all words of length at most n which makes the measure monotone. Observe that the leaf and advice measures are monotone as well. Note that dierent denitions have been used by other authors; see e.g. [GLW92], where the number of advice bits is maximized over all inputs and minimized over all accepting computations on those inputs. In this case there are nfa's which use more than constant but less than linear (in the input length) advice bits, but this behavior is not known to be possible for minimal nfa's. To prove lower bounds on the size of nite automata we shall use two-party communication complexity. This widely studied measure was introduced by Yao [Y79] and is the subject of two monographs [Hr97], [KN97]. First, we introduce the standard, non-uniform model of (communica- tion) protocols for computing nite functions. A (two-party communi- consists of two computers C I and C II of unbounded computational power (sometimes called Alice and Bob in the literature) and a communication link between them. P computes a nite function in the following way. At the beginning C I gets an input obtains an input 2 V . Then C I and C II communicate according to the rules of the protocol by exchanging binary messages until one of them knows f(; ). C I and C II may be viewed as functions in this communication, where the arguments of C I (C II ) are its input () and the whole previous communication history (the sequence c 1 of all messages exchanged between C I and C II up until now), and the output is the new message submitted. We also assume that C I (C II ) completely knows the behavior of C II (C I ) in all situations (for all arguments). Another important assumption is that every protocol has the prex-freeness property. This means, that for any ; any communication history , the message C I (; of C I ( prex of C II ( ))]. Informally, this means that the messages are self-delimiting and we do not need any special symbol marking the end of the message. Formally, the computation of a protocol (C I ; C II ) on an input is a sequence are the messages and 2 Z is the result of the computation. The communication complexity of the computation of P on an input (; ) is the sum of the lengths of all messages exchanged in the communication. The communication complexity of the protocol P , cc(P ), is the maximum of the communication complexities over all inputs from U V . Due to the prex-freeness property of messages we have that if, for two computations and then and a protocol allows m dierent computations, then its communication complexity must be at least dlog 2 me 1. The communication complexity of f , cc(f), is the communication complexity of the best protocol for f , i.e., The protocols whose computations consist of one message only (i.e., C I sends a message to C II and then C II must compute the result) are called one-way protocols. For every nite function f , is a one-way protocol computing fg is the one-way communication complexity of f . The representation of a nite function f : UV ! f0; 1g by the so-called communication matrix is very helpful for investigating the communication complexity of f . The communication matrix of f is the jU jjV j Boolean matrix M f [u; v] dened by for all u 2 U and v 2 V . So, M f [u; v] consists of jU j rows and jV j columns. If one wants to x this representation (which is not necessary for the relation to the communication complexity of f ), one can consider some kind of lexicographical order for elements in U and V . But, the special order of rows and columns does not matter for our applications. Figure 1 presents the communication matrix M f for the Boolean function dened by where is addition modulo 2. Denition 1. Let be two sets and . For every 2 U , the row of in M f is row For every 2 V , the column of in M f is is the number of dierent rows of M f . A submatrix of M f is any intersection of a non-empty set of rows with a non-empty set of columns. A -monochromatic submatrix, 2 f0; 1g of M f is any submatrix of M f whose elements are all equal to (Figure 1 depicts the 1-monochromatic submatrix that is the intersections of rows 001, 010, 100 and 111 with the columns 000, 011, 101 and 110). Figure be a set of monochromatic submatrices of a Boolean matrix M f . We say that S is a cover of M f if, for every element a of M f , there exists an m 2 kg such that a is an element of Mm . We say that S is an exact cover of M f if S is a cover of M f and kg. The tiling complexity of M f is is an exact cover of M f g ut The work of a protocol (C I ; C II ) for f can be viewed as a game on the communication matrix M f . C I with input knows the row row , C II with input knows the column column , and they have to determine f(; ). 2 A communication message c 1 submitted from C I to C II can be viewed as the reduction of M f to a submatrix M f consisting of rows for which C I sends c 1 because C II knows the behavior of C I . Similarly the second message 2 sent from C II to C I restricts M f of the columns of M f (c 1 ) for which C II with the second argument c 1 sends 2 Note that they do not need to estimate the coordinates of the intersection of row and column . knows the result. So, every computation of (C I ; C II ) that nishes with 1 (0) denes a 1-monochromatic (0-monochromatic) submatrix of M f . This means that all inputs (; ) contained in this monochromatic submatrix have the same computation of the protocol C I and C II . So, (C I ; C II ) unambiguously determine an exact cover of M f by monochromatic subma- trices. More precisely, a protocol with k dierent computations determines an exact cover of cardinality k. The immediate consequence is: Fact 1. For every nite function Another important consequence is the following fact. Fact 2. For every nite function (Row (M f ))e: Proof: For no two dierent rows row and row , a one-way protocol computing f can send the same message c because C II cannot determine the result for any such that column has dierent values on the intersections with row and row . On the other hand, Row dierent messages are enough (one message for a group of identical rows) to construct a one-way protocol for f . ut Since the number of 1-monochromatic matrices in any exact cover of all ones in M f is a trivial upper bound on the rank of M f , Fact 1 implies: Fact 3. For every nite function every eld F with neutral elements cc(f) dlog 2 (Rank F (M f ))e: Let Q be the set of rational numbers. Since it is well-known that is a eld with neutral elements 0 and 1g we formulate Fact 3 as for every nite function f . Now, we consider nondeterministic communication complexity and its relation to some combinatorial properties of M f . A nondeterministic protocol P computing a nite function consists of two nondeterministic computers C I and C II that have a nondeterministic choice from a nite number of messages for every input argument. For any input we say that that P accepts (; there exists a computation of P on (; ) that ends with the result 1. So, computes 0 for an input (; ) (rejects (; )) if all computations of P on (; ) end with the result 0. The nondeterministic communication complexity of P , denoted ncc(P ), is the maximum of the communication complexities of all accepting computations of P . The nondeterministic communication complexity of f is is a nondeterministic protocol computing fg Let ncc 1 (f) denote the one-way nondeterministic communication complexity of f . Similarly as in the deterministic case, every accepting computation of P for f unambiguously determines a 1-monochromatic submatrix of M f and the union of all such 1-monochromatic submatrices must cover all the 1's of M f but no 0 of M f . The dierence to the deterministic case is that these 1-monochromatic submatrices may overlap, which corresponds to the fact that P may have several dierent accepting computations on a given input. Denition 2. Let M f be a Boolean matrix, and let be a set of 1-monochromatic submatrices of M f . We say that S is a 1-cover of M f if every 1 of M f is contained in at least one of the 1-submatrices of S. We dene is a 1-cover of M f g: ut Fact 4. For every nite function Proof: The above consideration showing that a nondeterministic protocol with m accepting computations determines a 1-cover of M f of cardinality implies Since ncc(f) ncc 1 (f) for every f , it is sucient to prove ncc 1 (f) be a 1-cover of M f . A one-way nondeterministic protocol (C I ; C II ) can work on an input (; ) as follows. C I with input nondeterministically chooses one of the matrices of S with a non-empty intersection with row and sends the binary code of its index i to C II . If column has a non-empty intersection with M i , then C II accepts. Since dlog 2 me message length suces to code m dierent messages, The rst trivial bridge [Hr86] between automata and communication complexity says that for every regular language L and every positive integer n, where L. The argument for this lower bound is very simple. Let A be a dfa (nfa) accepting L with s(L) (ns(L)) states. Then a one-way protocol can compute f 2n;L as follows. For an input , C I simulates the work of A on and sends the name of the state q reached by A after reading to C II . C II continues in the simulation of the sux from the state q. If A accepts , then (C I ; C II ) accepts (; ). Unfortunately, the lower bound (1) may be arbitrarily bad for both s(L) and ns(L) because this non-uniform approach cannot completely capture the complexity of the uniform acceptance of L. We shall overcome this diculty in the next section. 3 Communication Complexity and Finite Automata To improve lower bounds on s(L) and ns(L) by communication complexity, Duris, Hromkovic, Rolim, and Schnitger [DHRS97] (see also [Hr86, HS00]) introduced uniform protocols and communication matrices of regular languages as follows. For every regular language L , we dene the innite Boolean matrix a Since every regular language has a nite index (Myhill-Nerode theorem), the number of dierent rows of ML is nite. So, we can again use the protocols as nite devices for accepting L. Denition 3. Let be an alphabet and let L . A one-way uniform protocol over is a pair (C I ; C II ), where is a function with the prex freeness property, and fC I () j 2 g is a nite set, and rejectg is a function. We say that The message complexity of the protocol D is (i.e., the number of the messages used by D), and the message complexity of L is is a one-way uniform protocol accepting Lg: The communication complexity of D is and the one-way communication complexity of L is is a one-way uniform protocol accepting Lg: ut If one wants to give a formal denition of a one-way nondeterministic protocol over , it is sucient to consider C I as a function from to a nite subset of f0; 1g . The acceptance criterion of L changes to I () such that accept 2 C II (; c)) , 2 L: denote the one-way nondeterministic message [communication] complexity of L. We observe that the main dierence between uniform protocols and (standard) protocols is the way the input is partitioned between C I and C II . If a protocol D computes a Boolean one can view this as the partition of inputs of f (from f0; 1g r+s ) into the prex of r bits and a sux of s bits (i.e. assigning the rst r bits to C I and the rest to C II ), and a communication between C I and C II in order to compute the value of f . A uniform protocol over considers, for every input partitions of for each of these partitions it must accept (reject) if 2 L ( 62 L). This means, that the matrices are special Boolean matrices with a ;1 ::: and a uniform protocol D for L must recognize the membership of to L for every partition of between C I and C II . The following result from [DHRS97, HS00] shows in fact that one-way uniform protocols are nothing else but deterministic nite automata. Fact 5. Let be an alphabet. For every regular language L , The idea of Proof: just a reformulation of the Myhill-Nerode theorem. In Section 2 we have already observed that Row (ML ) is exactly the number of dierent messages used by an optimal one-way protocol. 3 ut Following the idea of the simulation of a nite automaton by a protocol in the nondeterministic case, we have the following obvious fact [Hr97]. Fact 6. For every alphabet and every regular language L , Fact 6 provides the best known lower bound proof technique on the size of minimal nfa's. All previously known techniques like the fooling set approach are special cases of this approach. Moreover the fooling set method, which covers all previous eorts in proving lower bounds on ns(L), can (for some languages) provide exponentially smaller lower bounds than the method based on nondeterministic communication complexity [DHS96]. The rst question is therefore whether nmc(L) can be used to approximate ns(L). Unfortunately this is not possible. Note that a result similar to Lemma 1 was also independently established by Jiraskova [Ji99]. Lemma 1. There exists a sequence of regular languages fPART n g 1 n=1 such that Proof: Let PART For the next considerations it is important to observe that the condition equivalent to the condition x 6= z _ First we describe a nondeterministic uniform protocol (C I ; C II ) for PART n which uses O(n 2 ) messages. Players C I and C II compute the lengths l I ; l II of their inputs. C I communicates l I and C II rejects when l I l II 6= 3n. So we assume that l I in the following. Case 1: l I n. C I chooses a position 1 i l I and communicates I . C II accepts, accepts if and only if Observe that if x 6= z, then there is an accepting computation because there exists i such that x i 6= z i . If however that is i y. Case 2: n < l I 2n. C I chooses a position 1 i n and communicates I . Furthermore, C I compares x I n and sends the bit 1, if the strings are equal and the bit 0 if the strings are dierent. C II accepts if x i 6= z i . Otherwise (if x 3 The fact that ML is innite does not matter because ML has a nite number of dierent rows. Moreover, it would work for an innite number of dierent rows (i.e., for automata with an innite number of states), too [Eil74]. If the two strings are equal and the bit 1 was received, then C II accepts and rejects otherwise. Note that if x 6= z then there is an accepting computation. If not, then C II accepts if and only if Case 3: 2n < l I 3n. C I chooses a position l I 2n < i n and communicates I . Furthermore C I compares x with y. If then C I accepts. Otherwise C II accepts if and only if x i 6= z i . The protocol uses O(n 2 ) messages, so nmc(PART n Now, we prove that ns(PART N . Obviously, every nfa B accepting must have the following properties: there is an accepting computation of B on every word xxx or x 2 f0; 1g n , and i.e. there is no accepting computation of B on any word xyx with x 6= We prove that every nfa satisfying (i) and (ii) must have at least 2 nstates. Let us assume the opposite. Let A be a nfa with fewer than 2 nstates that satises (i) and (ii). Since L 1 L(B), there exists an accepting computation C x on xxx for every x 2 f0; 1g n . Let P attern(C x where p is the state of C x after reading x and q is the state of C x after reading xx. Since the number of states is smaller than 2 n 2 , the number of dierent patterns is smaller than 2 . So, there exist two words v, such that P attern(C u some states This means that starting to work from r on u as well as on v one can reach s after reading u or v. The immediate consequence is that there are accepting computations of B on uvu and vuv as well. Since u 6= v, uvu and vuv belong to L 2 , a contradiction with condition (ii). ut To nd lower bound methods for ns(L) that provide results at most polynomially smaller than ns(L) is one of the central open problems on nite automata. In the following, we concentrate on lower bounds for nfa's with constant ambiguity. Even for unambiguous automata no nontrivial general method for proving lower bounds has been known up to now. To introduce our method for proving lower bounds on nfa's with bounded ambiguity we have to work with the communication matrices for regular languages. In Fact 5 we have observed that every matrix ML has a - nite number of dierent rows, which is the index s(L) of the regular language L (this means that there exists a s(L) s(L) (nite) submatrix M of ML such that Row every eld F with neutral elements and Thus, instead of introducing the general two-way uniform communication protocols), we dene the communication complexity of L, denoted cc(L), as the communication complexity of the best protocol for the communication matrix ML . Because of the denition of ML , this approach covers the requirement that the protocol correctly decides membership of any input to L for any prex-sux partition of the input. Before formulating the main result of this section we build our intuition about the connection between cc(L) and uns(L). If one simulates an unambiguous automaton by a nondeterministic one-way protocol in the standard way described above, then the resulting protocol is unambiguous, too. This means that every 1 in ML is covered by exactly one accepting computation, i.e., the unfa A determines an exact cover of all 1's in ML of cardinality size A . The similarity to the deterministic communication complexity is that any such protocol determines an exact cover of all elements of the communication matrix by monochromatic submatrices. Some nontrivial results from communication complexity theory [KNSW94] are needed to relate cc(L) and uns(L) via the outlined connection. Theorem 1. For every regular language L , a) uns(L) RankQ (ML ), ns k (L) RankQ (ML ) 1=k 1, c) ns k (L) 2 cc(L))=k 2. Proof: Let A be an optimal unfa for L. A can be simulated by a one-way nondeterministic protocol as follows: C I simulates A on its input and communicates the obtained state. C II continues the simulation and accepts/rejects accordingly. Obviously the number of messages is equal to size A and the protocol works with unambiguous nondeterminism. It is easy to see that the messages of the protocol correspond to size A many submatrices of the matrix ML covering all ones exactly once. Hence the rank is at most size A and we have shown a), which is the rank lower bound on communication complexity [MS82] (see Fact 3 in Section 2). For b) observe that the above simulation induces a cover of the ones in ML so that each one is covered at most k times. By the following fact from [KNSW94] we are done: Fact 7. Let r (M) denote the minimal size of a set of submatrices covering the ones of a Boolean matrix M so that each is covered at most r times. Then For the other claim again simulate A by a one-way k-ambiguous nondeterministic protocol with size A messages. The results of [KNSW94] (see also [L90], [Y91]) imply that a k-ambigu- ous nondeterministic one-way protocol with m messages can be simulated by a deterministic two-way protocol with communication log(m log(m cc(L) log(size k and c) follow. ut Before giving an application of the lower bound method we point out that neither 2 nor RankQ (ML ) is a lower bound method capable of proving polynomially tight lower bounds on the minimal size of unfa's for all languages. In the rst case this is trivial, in the second case it follows from a modication of a result separating rank from communication complexity (see [KN97]). But the gap between RankQ (ML ) and uns(L) may be bounded by a pseudo-polynomial function. Now we apply Theorem 1 in order to present an exponential gap between ns(L) and uns(L) for a specic regular language. Let, for every positive integer m; g. Theorem 2. For every positive integer m (i) NIDm can be recognized by an nfa A with ambiguity O(m) and size O(m) (ii) Any nfa with ambiguity k for NIDm has size at least 2 m=k 1, and in particular any unfa for NIDm must have states. log m) for NIDm has polynomial size in m. Proof: (i) First the nfa guesses a residue i modulo m, and then checks whether there is a position p (ii) Observe that the submatrix spanned by all words u and v with is the \complement" of the 2 m 2 m identity matrix. The result now follows from the assertions a) and b) of Theorem 1. (iii) is an immediate consequence of (ii). ut We see that the proof of Theorem 2 is a substantial simplication of the proofs of similar results presented in [Sc78], [SH85]. 4 Degrees of Nondeterminism in Finite Automata It is easy to see that advice A (n) leaf A (n) 2 O(advice A (n)) and also that ambig A (n) leaf A (n) for every nfa A. The aim of this section is to investigate whether stronger relations between these measures hold. Lemma 2. For all nfa A either a) advice A (n) size A and leaf A (n) size size A A or advice A (n) n=size A 1 and leaf A (n) n=size A 1. @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ Figure Proof: If some reachable state q of A belongs to a cycle in A and if q has two edges with the same label originating from it such that one of these edges belongs to the cycle, then advice A (n) (n size A )=size A n=size A 1. Otherwise for all words all states with a nondeterministic decision are traversed at most once. ut Our next lemma relates the leaf function to ambiguity. The initial idea is that a computation tree of any minimal unfa A on any input w could look like the tree from Figure 2. There is exactly one path P from the root to a leaf (a computation) with several nondeterministic guesses and all paths having only one vertex in common with P do not contain any nondeterministic branching. In other words, if a computation branches into two computations P 1 and P 2 , then at least one of P 1 and P 2 should be completely deterministic. We are not able to verify this nice structure, but the next result shows that any computation tree of a minimal unfa A is very thin because every level of this tree can contain at most size A + 1 dierent computations. In what follows a state q of an nfa A is called terminally rejecting, if there is no word and no computation of A, such that A accepts when starting in q, i.e., (q; v) contains no accepting state for any word v. Clearly there is at most one terminally rejecting state in a minimal automaton, because otherwise these states can be joined reducing the size. Call all other states of A undecided. Lemma 3. Every nfa A with at most one terminally rejecting state satises leaf A (x) ambig A (jxj for all x. size A ). If the computation tree consists only of nodes marked with the terminally rejecting state, then the tree has just one leaf and the claim is trivial. For the general case, consider a level of the computation tree of A on x that is not the root level. Assume that the level contains more that k size A nodes labeled with undecided states (called undecided nodes). Then one undecided state q must appear at least times on this level. There are k computations of A on a prex of x such that q is reached. If q is accepting, then the prex of x is accepted a contradiction, since ambig A is monotone. If q is rejecting, but undecided, then there is a word v of length at most size A such that v is accepted by some computation of A starting in q. But then the prex of x concatenated with v is accepted by at least k a contradiction. Thus each level of the tree that is not the root level contains at most k size A undecided nodes. Overall there are at most jxj k size A undecided nodes. Observe that each node has at most one terminally rejecting child. Thus the number of terminally rejecting leaves is equal to the number of undecided nodes that have a terminally rejecting child. Hence the number of terminally rejecting leaves is at most the number of undecided nodes minus the number of undecided leaves. Thus the overall number of leaves is at most the number of terminally rejecting leaves plus the number of undecided leaves which is at most the number of undecided nodes. So overall there are at most leaves. ut Theorem 3. Every nfa A with at most one terminally rejecting state sat- ises advice A (n); ambig A (n) leaf A (n) O(ambig A (n) advice A (n)): Especially for any such unfa: advice A Proof: Observe that for all n: ambig A ambig A (n since ambig A is monotone and at most exponential. ut Next we further investigate the growth of the leaf function. Lemma 4 is a variation of a result in [IR86]. Lemma 4. For every nfa A, either leaf A (n) (nsize A ) size A or leaf A (n) n) . Proof: Assume that an nfa A contains some state q, such that q can be reentered on two dierent paths starting in q, where each path is labeled with the same word w. It is not hard to show that in this case there are two dierent paths from q to q labeled with a word w of length size 2 A 1. Then the computation tree of uw m (where u leads from the starting state to q) has at least 2 m 2 (n size A )=size 2 A leaves, where Now assume that A does not contain such a state. Then, for each non-deterministic state q (i.e., a state with more than one successor for the same letter) and any computation tree, the following holds: If q is the label of a vertex v, then q appears in each level of the subtree of v at most once. We prove by induction on the number k (k size A ) of dierent nondeterministic states in a computation tree that the number of leaves is at most (n size A ) k . The claim is certainly true if there are no nondeterministic states. Assume that there are k nondeterministic states, with some state q 1 appearing rst in the tree. Observe that no level in the entire computation tree contains q 1 more than once. For each occurrence of q 1 in the computation tree x some child, so that the overall number of leaves is maximized. We get a tree with one nondeterministic state less, and by the inductive hypothesis this tree has at most (n size A ) k 1 leaves. appears at most once on each level and since there are at most size A children of q 1 on each level, there are at most (n size A ) k leaves. ut Lemmas 2 and 4 give us Theorem 4. For every nfa A: leaf A (n) is either bounded by a constant, or in between linear and polynomial in n, or otherwise 2 (n) . Now, we consider the dierence between polynomial and exponential ambiguity resp. polynomial and exponential leaf number. We show that languages which have small automata of polynomial ambiguity are related to the concatenation of languages having small unfa's. If the language is a Kleene closure, then one unfa accepts a large subset. Compare this to closures are shown to be recognizable as ecient by nfa's with constant advice as by dfa's. Theorem 5. a) Let L be an innite regular language and A some nfa for L with polynomial ambiguity. Then there are d size A languages L i such that L 1 L d L, L i is recognizable by an unfa with O(size A ) states, and =for innitely many n. for a regular language K not using the letter # and let A be some nfa for L with polynomial ambiguity. Then, for all m, there is an unfa A 0 with O(size A ) states that decides L 0 L such that for innitely many n Proof: a) Dene the ambiguity graph of A in the following way: the nodes are the (reachable) states of A and there is an edge from q i to q j if there are two paths from q i to q j in A, with the same label sequence. Note that the ambiguity graph is acyclic i the ambiguity of A is polynomially bounded as we have seen in the proof of Lemma 4. Now we construct a unfa A i;j;k which accepts those words that lead in A from q i to q j and then via one edge to q k . Here, we assume that the longest path from q i to q k in the ambiguity graph consists of one edge and q j is reachable from q i in A, but not in the ambiguity graph. Moreover, we demand that there is an edge in A from q j to q k . The states of A i;j;k are the states reachable in A from q i , but not reachable in the ambiguity graph from q i , plus the state q k . The edges are as in A except that the only edges to q k come from q j . q i is the start. Accepting state is q k . L i;j;k is the language accepted by A i;j;k . Now consider the words w 2 L \ n . Each such word is accepted on some path in A leading from q 0 to some accepting state q a . Fix one such accepting state so that a constant fraction of all words w is accepted and make the other accepting states rejecting. On an accepting path for w the states appear without violating the topological ordering of the ambiguity graph. So, we may x a sequence of states q a such that . Since there are only nitely many such sequences we are done. b) Similar to a), we get k languages L decidable by small unfa's A i , such that =for innitely many n. A partition of the letters of words in ( m #) n is given by mapping the nm letters to the k unfa's. There are at most n possible partitions. So some partition must be consistent with accepting paths for a fraction of 1=poly (n) of (( m \ K)#) n . Fix one such partition. Then for each words w an unfa is responsible for some prex u, followed by a concatenation of words of the form # m , and nally a word of the form #v. For all i we x a prex u i , a sux v i , and states q entered when reading the rst and nal occurrence of #, such that as many words from (( m \ K)#) n as possible are accepted under this xing. At least a fraction of size k =2 1=poly (n) of (( m \K)#) n has accepting paths consistent with this xing. If any A i accepts less than a polynomial fraction (compared to the projection of (( m \K)#) n to the responsibility region of A i ) then overall less than a polynomial fraction is accepted. Hence one A i can be found, where from q i a polynomial fraction of words in ( m \ K)#) n=k leads to non- terminally rejecting states in A i . Making one non-terminally rejecting state reached by a # edge accepting and removing the original accepting states yields an unfa that accepts the desired subset for innitely many n. ut Applying Theorem 5 we can prove an exponential gap between nfa's and nfa's with polynomial ambiguity. This proof is also substantially simpler 4 than the proof of an exponential gap between polynomial ambiguity and exponential ambiguity for the language (0 Theorem 6. There is a family of languages KLm such that KLm can be recognized by an nfa with advice (n), leaf 2 (n) and size poly(m), while every nfa with polynomial leaf number/ambiguity needs size at least 2 m) to recognize KLm . Proof: Let LNDISJ from a size m 32 universe and the sets [ non-triviallyg. Moreover, let Given a polynomial ambiguity nfa for KLm , we get an unfa accepting a fraction of 1=poly(n) of (LNDISJ m#) n for innitely many n by Theorem 4b). Then we simulate the unfa by a nondeterministic communication protocol, where player C I receives all x and player C II all y inputs. The protocol needs O(n log size A ) bits to work correctly on a 1=poly(n) fraction of (LNDISJ m#) n and has unambiguous nondeterminism. A result from [HS96] implies that this task needs communication nm) and thus size A 2 m) . ut Thus, we have another strong separation between the size of automata with polynomial ambiguity and the size of automata with exponential ambi- guity. The situation seems to be more complicated, if one compares constant and polynomial ambiguity. Ravikumar and Ibarra [RI89] and Hing Leung [HL98] considered it as the central open problem related to the degree of ambiguity of nfa's. Here, we can only show that there is a family KON m of languages with small size nfa's of polynomial ambiguity, while nfa's of am- biguity are exponentially larger. In the following theorem we describe a candidate for a language that has ecient nfa's only when ambiguity is poly- nomial. Furthermore the language exhibits an almost optimal gap between the size of unfa's and polynomial ambiguity nfa's. In the proof the rank of the communication matrix of KON m is shown to be large by a reduction from the disjointness problem. Theorem 7. Let KON contains all words in f0; 1g with a number of 1's that is divisible by m. KON m can be recognized by an nfa A with ambig A (n); leaf A while any nfa with ambiguity k for KON m needs at least 2 (m 1)=k 2 states. Proof: Since the upper bound of theorem 7 is obvious, we focus on proving the lower bound. Consider the communication problem for the complement of the disjointness predicate NDISJ l . The inputs are of the form x; y 2 f0; 1g l , where x 4 If the known results about communication complexity are for free (i.e., not included in the measurement of the proof diculty). and y are interpreted as incidence vectors of subsets of a size l universe. The goal is to nd out, whether the two sets have a nontrivial intersection. Note that the rank of the communication matrix M NDISJ l is 2 l 1. We reduce NDISJ m 1 to KON m , i.e., identify a submatrix of MKONm that is the communication matrix M NDISJ m 1 Consider inputs to KON m of the form 01 and addition over ZZ m . For any subset s 1g one can nd such an input x s . These inputs correspond to the rows of our submatrix. For each subset x an input y s of the g. These 2 m 1 inputs correspond to the columns of our submatrix. Now consider the obtained submatrix: if s and r intersect non-trivially, then x s y r 2 KON m . On the other hand, if s and r are disjoint, then there is no sub-word which has a number of 1's divisible by m. r is not in KON m . We have identied a submatrix of rank 2 m 1 1. Applying Theorem 1(b) we obtain our lower bound. ut For every constant m, the language KON m 2 of Theorem 7 can be recognized with size O(m 2 ), leaf number and ambiguity (n), and advice (n), while every m ambiguous nfa has size 2 m) . Jurdzinski [Ju00] observed that KON m 2 can be computed by nfa's with constant ambiguity and size poly(m). Therefore the analysis of Theorem 7 cannot be improved substan- tially. Jurdzinski's observation also applies to the language f0; 1g 0 k f0; 1g which was proposed in [RI89] for separating constant from polynomial ambiguity 5 Conclusions and Open Problems We have shown that communication complexity can be used to prove lower bounds on the size of nfa's with small ambiguity. This approach is limited, because for nontrivial bounds ambiguity has to be smaller than the size of a minimal nfa. Is it possible to prove lower bounds for automata with arbitrarily large, but constant ambiguity, when equivalent automata of small size and polynomial ambiguity exist? In this context it would be also of interest to investigate the ne structure of languages with regard to constant ambiguity. At best one could show exponential dierences between the number of states for ambiguity k and the number of states for ambiguity k + 1. Observe however, that such an increase in power is impossible provided that the size of unfa's does not increase substantially under complementation [K00]. Analogous questions apply to polynomial and exponential ambiguity. Are there automata with non-constant but sub-linear ambiguity? A negative answer establishes Theorem 3 also for ambiguity as complexity measure. Other questions concern the quality of communication as a lower bound method. How far can Rank resp. 2 cc(L) be from the actual size of minimal unfa's? Note that the bounds are not polynomially tight. Are there alternative lower bound methods? Finally, what is the complexity of approximating the minimal number of states of an nfa? --R Lower bounds on information transfer in distributed computations. On notions of informations transfer in VLSI circuits. On measuring nondeterminism in regular languages. On the relation between ambiguity and nondeterminism in Separating exponentially amgigous On sparseness personal communication. Lower bounds for computation with limited nonde- terminism On automata with constant ambiguity. Communication Complexity. Las Vegas is better than determinism in VLSI and distributed computing. Economy of description by au- tomata On the bounds for state-set size in the proofs of equivalence between deterministic On ranks vs. communication com- plexity Communication complexity. Communication Complexity. Relating the type of ambiguity of Finite automata and their decision prob- lems Lower Bounds on the Size of Sweeping Automata. Succinctness of descriptions of context-free On the equivalence and containment problems for unambiguous regular expressions A complexity theory for VLSI. Expressing combinatorial optimization problems by linear programs. Some complexity questions related to distributed com- puting --TR On sparseness, ambiguity and other decision problems for acceptors and transducers Communication complexity hierarchy Relating the type of ambiguity of finite automata to the succinctness of their representation On measuring nondeterminism in regular languages On the relation between ambiguity and nondeterminism in finite automata Non-deterministic communication complexity with few witnesses Nondeterministic communication with a limited number of advice bits A comparison of two lower-bound methods for communication complexity Communication complexity and parallel computing Communication complexity Separating Exponentially Ambiguous Finite Automata from Polynomially Ambiguous Finite Automata On the power of Las Vegas for one-way communication complexity, OBDDs, and finite automata Automata, Languages, and Machines On the Power of Las Vegas II. Two-Way Finite Automata Measures of Nondeterminism in Finite Automata Las Vegas Versus Determinism for One-way Communication Complexity, Finite Automata, and Polynomial-time Computations Translating Regular Expressions into Small epsilon-Free Nondeterministic Finite Automata Communication complexity Las Vegas is better than determinism in VLSI and distributed computing (Extended Abstract) Area-time complexity for VLSI Some complexity questions related to distributive computing(Preliminary Report) On notions of information transfer in VLSI circuits Succinctness of descriptions of context-free, regular and finite languages. A complexity theory for vlsi --CTR Martin Kutrib , Andreas Malcher, Context-dependent nondeterminism for pushdown automata, Theoretical Computer Science, v.376 n.1-2, p.101-111, May, 2007 Galina Jirskov, State complexity of some operations on binary regular languages, Theoretical Computer Science, v.330 n.2, p.287-298, 2 February 2005

nondeterminism;limited ambiguity;descriptional complexity;communication complexity;finite automata

507384

On first-order topological queries.

One important class of spatial database queries is the class of topological queries, that is, queries invariant under homeomorphisms. We study topological queries expressible in the standard query language on spatial databases, first-order logic with various amounts of arithmetic. Our main technical result is a combinatorial characterization of the expressive power of topological first-order logic on regular spatial databases.

Introduction The expressive power of first-order logic over finite relational databases is now well understood [AHV95, EF95]. Much less is known in spatial databases (also called constraint databases), where the relations are no longer finite but finitely represented [KLP99]. The notion of genericity (invariance of queries under isomorphisms), fundamental for the relational database model, can be generalized to spatial databases in various ways [PVV94]. Given a group G of transformations (trans- lations, affinities, isometries, similarities, homeomorphism- s, etc.), a query Q is G-generic if for all database instances I and each transformation FOG we denote the set of G-generic first-order queries. The genericity of a first-order query is undecidable [PVV94], but the expressive power of FO G can be understood via sound and complete (decidable) languages. A language is said to be sound for G if it contains only FO G queries. It is complete for G if it expresses all FO G queries. The choice of the group G depends on which information one is interested in. [GVV97] gives sound and complete languages for several natural groups of transformations (translations, affinities, isometries, similarities). The case of the group of homeomorphisms was left open. To appear in Proceedings of the 15th IEEE Symposium on Logic in Computer Science. c IEEE 2000. Queries invariant under homeomorphisms, which are also called topological queries, are of fundamental importance in various applications of spatial databases. For ex- ample, in geographical databases, queries like "Is region A adjacent to region B?", "Is there a road from A to B?", or "Is A an island?" come up very naturally. Therefore, topological queries have received a lot of attention in the literature (e.g. [KPV97, PSV99, SV98, KV99]). A basic result known about topological queries is that connectivity of a region is not expressible in first-order logic [GS99, GS97, BDLW96]. Thinking of geographical databases again, planar (or 2- dimensional) database instances, where all relations are embedded in the plane R 2 , are of particular importance. In [PSV99] it has been proven that all topological properties of a planar spatial database can be represented in a finite structure called the topological invariant of the instance. In [SV98] it has been shown how this topological invariant can be used to answer topological queries. In particular, [SV98] have proven that first-order topological queries on a spatial database can be automatically translated into fixpoint queries on the topological invariant. The translation of first-order topological queries on the spatial database into first-order queries on the topological invariant was proven possible only in the special case of a single relation representing a closed region. It was left open in [SV98] whether this translation could be extended to the case of several re- gions. We answer this question negatively. The idea of representing the topological information of a spatial database instance by the topological invariant has two important drawbacks: In a sense, the topological invariant contains too much information; ideally we would just want to store the information that is actually accessible by the query language (which is usually FO). Furthermore, the topological invariant has no straightforward generalization to higher dimensions. The issue of finding an invariant more suitable for FO (and computable in any dimension) was raised in [KPV97]. In the special case of one single relation representing a closed planar region, a cone structure was given in [KPV97] capturing precisely the first-order topological information. Intuitively, the cone structure is a finite set containing all the possible small neighborhoods of a point. The results of [KPV97] show that, in this context, first-order topological queries could express only local properties, which is a situation known to be true in the finite case. [KPV97] asked whether their results generalize to database instances with a region that is not necessarily closed; we give a negative answer to this question. For instances with one closed region that satisfy the additional technical condition of being fully two dimensional, [KV99] introduced a cone logic CL and proved that it is sound and complete for topological FO. They asked if their results generalize to instances with not necessarily closed regions or several regions, again we give a negative answer. [KPV97] introduced two local operations on spatial database instances that preserve the equivalence under first-order topological queries (called topological elementary e- quivalence). We call two instances -equivalent if they can be transformed into instances homeomorphic to each other by applying the operations of [KPV97] finitely often. Our main technical result, from which all the rest easily follows, is that on especially simple instances that we call regular, -equivalence and topological first-order equivalence coincide The paper is organized as follows: After recalling a few basic definitions on spatial databases in Section 2, in Section 3 we discuss the topology of planar spatial databases and the topological invariant in detail. In Section 4 we introduce topological first-order queries and review some results of [KPV97]. In Section 5, we prove that -equivalence is decidable in PSPACE. Our main result on regular instances is proved in Section 6. In Section 7, we derive that not all first-order topological queries can be translated to first-order queries on the topological invariant, and in Section 8 we briefly discuss the problem of finding a language that is sound and complete for topological FO. 2. Preliminaries Spatial databases. We fix an underlying structure R over the reals; either we let (R; <; +; 0; 1) or R = R poly be the vocabulary of R (i.e. either For a point an jja bjj < rg be the open ball with radius r around a. 1 As a matter of fact, we could let R be any o-minimal structure over the reals, and the main results would remain true. poly we may also let a 2 n , and we will assume we have done so in our figures - it just looks better. Since we are only interested in topological queries, this makes no difference. A subset S R n , for some n 1, is R-definable, if there is a first-order formula '(x) of vocabulary R such that A schema is a finite collection of region names. Let n 1. An n-dimensional spatial database instance I over associates an R-definable set R I R n with every R 2 . The sets R I are called the regions of I . Formally, we may interpret the region names as n-ary relation symbols and view an instance I over as a first-order structure of vocabulary R [ obtained by expanding the underlying structure R by the relations R I , for I 2 . In this paper, we only consider 2-dimensional (or pla- spatial database instances. For convenience, we also assume that all regions are bounded, i.e. that for every instance I over a schema and for every R 2 there exists a R such that jjajj b for all a 2 R I . The boundedness assumption is inessential and can easily be removed, but it occasionally simplifies matters. Queries. An n-ary query (for an n 0) of schema is a mapping Q that associates an R-definable subset Q(I) R n with every instance I over . Here we consider R 0 as a one point space and let R query is usually called Boolean query. As our basic query language we take first-order logic FO. of vocabulary R [ defines the n-ary query I 7! '(I) := f(a an an )g of schema . 3. The topology of planar instances R 2 is equipped with the usual topology. The interior of a set S R 2 is denoted by int(S), the closure by cl(S), and the boundary by bd(S). We say that a set S R 2 touches a point a 2 R 2 if a 2 cl(S). Two sets touch if S 2 touches a point a 2 S 1 or vice versa. Stratifications. Stratification is the fundamental fact that makes the topology of our instances easy to handle. A stratification of an instance I over is a finite partition S of R 2 such that (1) For all S 2 S, either S is a one point set, or S is homeomorphic to the open interval (0; 1), or S is homeomorphic to the open disk (2) For all cl(S) \ cl(S 0 ) is the union of elements of S. (3) For all R 2 and S 2 S we either have S R I or The following lemma follows from the fact that all regions of an instance are R-definable. A proof can be found in [vdD98]. Lemma 3.1. For every instance I there exists a stratification of I . Colors and cones. Let I be an instance over . The pre- color of a point a 2 R 2 is the mapping (a) fint; bdi; bde; extg defined by int if a 2 int(R I ); bde if a 2 bd(R I ) n R I ; ext if a 2 R 2 n cl(R I A pre-cell is a maximal connected set of points of the same pre-color. The cone of a point a 2 R 2 , denoted by cone(a), is the circular list of the pre-colors of all pre-cells touching a. Lemma 3.1 implies that cones are well-defined and finite. A point a 2 R 2 is regular if for every neighborhood U of a there is a point a 0 2 U such that Otherwise a is singular. It follows from Lemma 3.1 that an instance has only finitely many singular points. We call an instance regular if it has no singular points. The cones Figure 1. Two singular and two regular cones of regular (singular) points are also called regular (singular, resp.) (cf. Figure 1). The cone-type of I , denoted by ct(I), is a list of all cones appearing in I . Furthermore, for every singular cone this list also records how often it occurs. The color of a point a 2 R 2 is the pair . Cells. A cell of color of I is a maximal connected set of points of color . The color of a cell C is denoted by (C). Lemma 3.1 implies that there are only finitely many cells. Our assumption that all regions are bounded implies that there is precisely one unbounded cell, which we call the exterior of I . Lemma 3.1 implies that every cell has a well defined dimension, which is either 0, 1, or 2. The 0- dimensional cells are precisely the sets fag, where a is a singular point. Let C I be the set of all cells of an instance I . We define a binary adjacency relation E I on C I by letting two cells be adjacent if, and only if, they touch. We call the graph the cell graph of I . We can partition C I into three subsets C I 1 , and C I 2 consisting of the 0, 1, and 2- dimensional cells, respectively. Observe that the graph G I is tri-partite with partition (C I I is planar. Lemma 3.2. Let I be an instance and C 2 C I . (1) If C 2 C I either C is homeomorphic to the open disk (0), or there exists an m 1 such that C is homeomorphic to the open disk Dm with m holes. To be definite, we let Dm := D n cl then C is homeomorphic to the sphere or to the open interval (0; 1). Proof: This follows easily from Lemma 3.1. 2 The skeleton S I of an instance I is the set of all 0- dimensional cells and all 1-dimensional cells homeomorphic to (0; 1). Note that the skeleton of a regular instance is empty. Lemma 3.3. Let I be an instance. Then every connected component of the graph G I n S I is a tree. In particular, if I is regular then G I is a tree. Proof: Follows from the Jordan Curve Theorem and Lemma 3.2. 2 Figure 2 illustrates a typical connected component of an instance I after removing the skeleton. Note that every con- Figure 2. nected component of the graph G I n S I has a unique "ex- terior" cell which we may consider as the root of the tree. Having the tree directed by fixing this root, we may speak of the parent and the children of a node. The following observation will be useful later. Lemma 3.4. Let I ; I 0 be instances and C 2 C I I 0such that 2 that is adjacent to C there exists a 2 that is adjacent to C 0 such that The topological invariant. Two instances I ; J over a schema are homeomorphic if there is a homeomorphism such that for all a 2 R 2 and R 2 we have a 2 R I () h(a) 2 R J . The topological invariant of an instance I over is an expansion Y I of the cell graph that carries enough information to characterize an instance up to homeomorphism. The vocabulary of Y I is ^ R O is 8-ary, and R for R 2 are unary. The restriction of Y I to fEg is the cell graph G I of I . dim i consists of the i-dimensional cells, for 2. X only contains the exterior of I (the unique unbounded cell). For every R 2 , the unary relation ^ R consists of all cells that are subsets of R I . O gives the orientation. It is an equivalence relation on the quadruples (C; are adjacent to C. Two such quadruples are equivalent if either i in the clockwise order of the cells adjacent to C i for both i 2 f1; 2g or B 0 appears i in the anti-clockwise order of the cells adjacent to C i for both i 2 f1; 2g . 3 Note that O is empty in regular instances. It is proven in [PSV99] that I and J are homeomorphic iff Y I and Y J are isomorphic and that Y I is computable from I in time polynomial in the size of I . Since G I is a planar graph and canonization of planar graphs is in P- TIME, we can actually assume that Y I is canonical in the sense that for homeomorphic instances I and J we have In [SV98] it is proven that FO-queries over I can be translated in linear time into fixpoint+counting queries over Y I . Furthermore, if FO-queries over I can be translated to FO-queries over Y I on instances with just one closed region. (More precisely, this means that there is a recursive mapping that associates with every ' 2 FO of vocabulary f<; Rg a ' 0 2 FO of vocabulary d fRg such that for all instances I over fRg, where R I is a closed set, we have I The question was left open whether this result extends to instances with one arbitrary region or with several regions. In Section 6, we give a negative answer to this question. 3 There are various ways to define the orientation, ours is equivalent to [SV98]. 4 More precisely, there is a PTIME algorithm that, given an instance I, computes Y I and a one-to-one mapping I : C I (a canonical numbering) such that for homoemorphic instances I; J the mapping is an isomorphism from Y I to Y J . 4. Topological queries and topological elementary equivalence Topological queries. A query is topological if for every homeomorphism h of R 2 and for all instances I we have denotes the set of all first-oder formulas defining a topological query. It is well-known (and easy to see) that the set FO top is not decidable. The following lemma collects a few basic FO-queries. Its proof is an easy exercise. Lemma 4.1. (1) For every color there is a first-order (x) such that for every instance I and for every a 2 R 2 we have I (a) () (2) For every (y) 2 FO there is a formula ' bd( FO such that for every instance I we have ' bd( (3) There is a formula '1 (x) 2 FO such that for every instance I we have Note that for every color the formula ' is in FO top . Moreover, for 2 FO top the formula ' bd( ) is in FO top . In particular, this is the case for for an R 2 . On the other hand, the formula '1 is not in FO top . Topological elementary equivalence. Two instances I ; J are (topologically) elementary equivalent (denoted I t J) if they satisfy the same topological first-order sentences. It is proven in [KPV97] that if = fRg then for all instances I ; J in which R I , R J are closed sets we have: I t J We will see in the next section that this equivalence cannot be extended to instances with one arbitrary region or with several regions. To prove this result, [KPV97] introduced two simple local operations transforming an instance into an elementary equivalent one. Their straightforward extension to several regions is depicted in Figure 3, which is to be read as fol- lows: Suppose we have an instance I that contains an open subset O R 2 homeomorphic to one of the left hand sides of Figure 3. The different shades of grey display different colors. Then it can be replaced by the corresponding subset on the right hand side (cf. [KPV97] for details). Note that both operations are symmetric, we can go from the right to the left by applying the same operation again. the first and ! 2 the second of the two operations in Figure 3. For instances I and J we write I !! i J if I can be transformed into an instance homeomorphic to J by an application of ! i (for i 2 f1; 2g). We Figure 3. Operations preserving t . I J if I and J can be transformed into each other by a finite sequence of operations !!1 and !!2 . Then the proof of [KPV97] easily yields: Lemma 4.2. For all instances I , J we have: I I t J . It is an open question whether the converse of Lemma 4.2 holds. In particular, this is interesting because is not known whether t is decidable or not, whereas is decidable in PSPACE (this is Proposition 5.5 of the next sec- tion). [KPV97] have shown that that t and coincide on instances with only one closed region. We can extend their result to several regions, but only on instances with only regular cones. 5. Minimal instances Definition 5.1. An instance I is minimal if it satisfies the following two conditions: 1 is homeomorphic to S 1 and Iare adjacent to C and homeomorphic to Dm , for some are adjacent to B and homeomorphic to S 1 , then Lemma 5.2. There is a PTIME algorithm that associates with every instance I a minimal instance M(I) such that I M(I). This can be done in such a way that for homeomorphic instances I ; J we have Proof: Suppose first that I does not satisfy (M1). We show that I can be transformed to instance J with fewer cells violating (M1), by two applications of ! 1 . Let C be a 1-dimensional cell homeomorphic to S 1 such that both neighbors of C have the same color, but neither is homeomorphic to D. Then instance I locally looks like Figure 4(1). We apply ! 1 twice (to the dashed boxes) and obtain an instance that locally looks like Figure 4(3). We have obviously reduced the number of cells violating (M1). (1) (2) (3) Figure 4. Suppose now that I does not satisfy (M2), then it can be transformed to an instance J with fewer cells violating (M2) by an application of ! 1 without violating (M1). To see this, let B be a 2-dimensional cell in I that is adjacent to the cells C homeomorphic to S 1 of the same color. A cell homeomorphic to S 1 is adjacent to two 2- dimensional cells. Let be the other neighbors of the same color. We have to distinguish between two cases: Case 1: Both C 1 and C 2 are children of B. Then we can reduce the number of cells violating (M2) by an application of Figure 5). (1) (2) Figure 5. Case 2: C 1 is the parent of B and C 2 its child. Figure 6 shows how to proceed. (1) (2) Figure 6. A PTIME algorithm transforming a given instance I to a minimal instance M(I) may proceed as follows: Given I , the algorithm first computes the invariant Y I ; this is possible in PTIME [PSV99]. The operations ! 1 and ! 2 translate to simple local operations on Y I . Our algorithm first applies pairs of ! 1 until (M1) holds (as in figure 4), and then holds. This can be done by a simple greedy strategy. The result is a structure Y that is the toplogical invariant of a minimal instance M(I). It is shown in [PSV99] that, given an invariant Y , an instance J such that Y can be computed in polynomial time. Because Y I is canonical (cf. Page 4), this algorithm also guarantees that for homeomorphic instances I ; J we have The language of an instance. A fundamental curve in an instance I is an R-definable continuous mapping lim a!1 jjf (a)jj = 1. If f is a fundamental curve in I , then for every 2-dimensional cell the set f 1 (C) is a finite union of open intervals (one of which is of the form a) and one of the form (b; 1)), and for every 0- and 1-dimensional cell C the set f 1 (C) is a finite union of closed intervals (some of which may just be single points). This follows from the fact that f is R-definable. We are interested in the finite sequence of colors appearing on a fundamental curve, i.e. in a finite word W (f; I) over the alphabet I consisting of all colors appearing in I . We say that a word W 2 I is realized in I if there is a fundamental curve f such that W (f; . The language L(I) is the set of all words realized by I . Example 5.3. Let I be the instance with R I := fa 2 R 2 j (1=2) < jjajj 1g. I has five cells C 1 := fa - be the colors of C respectively, and note that C 5 has the same color as C 1 . Figure 7. a b a a d d Figure 8. Then for example the words and are realized in I (cf. Figure 7). It is not hard to see that L(I) can be described by the finite automaton displayed in Figure 8. -From this example it is easy to see that for every I the language L(I) is regular; it is accepted by an automaton that is essentially the cell graph. More formally, let I be an instance. We define a finite automaton A I (where 1 is a new symbol that does not denote a cell). := f(C; denotes the exterior. Note that the graph underlying A I is the cell graph G I extended by one additional vertex 1 that is only adjacent to the exterior. The proof of the following lemma is straight-forward Lemma 5.4. For every instance I , the automaton A I accepts L(I). Thus L(I) is a regular language. A walk in a graph E) is a sequence for 1 i < n. For a mapping . Then it is almost immediate that for every instance I we have wn walk in G I with where as usual EXT denotes the exterior. Proposition 5.5. is decidable in PSPACE. Proof: Given two instances I and J , we want to check in PSPACE whether it is possible to go from I to J using only homeomorphisms and operations in . Given an instance I it is possible to compute its invariant Y I in PTIME [PSV99]. Thus the problem reduces to checking in PSPACE whether Y J can be derived from Y I using operations from . There are at most polynomially many different ways to apply operators from to Y I (one just has to consider tuples of at most 5 cells and check if an ! i can be applied to this tuple). Let I) be the set of invariants that can be obtained from Y I by applying one operation from . The previous remarks show that I) contains at most polynomially many elements and can be computed in PTIME. It is therefore possible to enumerate all topological invariants that can be derived from Y I by applying operations in and check at each step whether it corresponds to Y J or not. The latter can be done in PTIME, because this is deciding whether two planar graphs are isomorphic. We now give a strategy that ensures that this process will stop. Because M(I) and M(J) are computable in PTIME (Lemma 5.2), we can assume I and J minimal. It is easy to see that operations in do not change the cones of the instances involved. Therefore I J implies that I and J have the same cone-type (cf. Page 3) and the same orientation relation O. This can be checked in PTIME. As the cones determine the number of 1-dimensional cells homeomorphic to (0; 1), this implies that all instances J such that I J are such that their respective skeletons verify jS I We may view the skeletons S I and S J as embedded planar graphs with the 0-dimensional cells as vertices and the 1-dimensional cells homeomorphic to (0; 1) as edges. Let F I and F J , resepectively, denote the faces of these embedded graphs. Note that these faces correspond precisely to the connected components of G I n S I and G J n S J , re- spectively. The number of connected components of S J is bounded by c I < jI j, the number of cones of I . Therefore the number of faces of S J is given by the Euler formula, we have jF J is the number of connected components of S J , and e; v its number of edges and vertices. As is bounded by jI j, so is jF J j. This gives a linear (in jI j) bound on the number of connected components of G J n S J . We would also like to bound the size of each connected component of G J n S J (or equivalently the number of 1-dimensional cells homeomorphic to S 1 ). Un- fortunately, repeated application of the operations 1 and may produce arbitrary large such components. Nevertheless, it is possible to decide in PSPACE whether I ! K for an instance K such that there is an isomorphism from S K to S J that preserves the coloring and orientation O. Furthermore, if there is such a K, it can be computed in PSPACE. The complexity is PSPACE because there is no need to produce extra 1-dimensional cells homeomorphic to S 1 during this process. This would only restrict the number of possible connections between the connected components of the skeleton. So without loss of generality we can now assume that I and J are minimal and have isomorphic skeletons. Recall that every face f 2 F I (f 2 F J ) corresponds to a connected component T f of G I n S I (G J n S J , respectively). By Lemma 3.3, T f is a tree with a canonical root exterior cell. To find out whether I J , for each isomorphism that preserves the coloring and orientation we check whether there are instances I 0 I and J 0 such that S I 0 for every face f 2 F I 0 we have T f denotes the face corresponding to f under the isomorphism i). It suffices to prove the following for every f 2 F I : Claim 1. Either I 6 J , or I and J can be transformed to instances I 0 and J 0 , respectively, with S I 0 such that in I 0 the component T f is isomorphic to T i(f) in ffg, the component T g in I 0 has remained the same as in I , and for all h 2 F J n fi(f)g, the component T h in J 0 has remained the same as in J . It can be decided in PSPACE which of the two alternatives holds, for the latter I 0 ; J 0 can also be computed in PSPACE. We can go through all isomorphism i in PSPACE, so we fix one. We also fix a face f 2 F I . Before going on we need the following definitions. A branch of a tree T is a minimal walk in T going from the root of T to one of its leaves. A walk in G I is said to be regular if it never goes through a singular cell. Let be the set of colors of I and J . For a word W 2 , we define M(W ) as the word of computed using rules in the same spirit as in Lemma 5.2. More precisely this means rewriting the word W using the rules ! for all colors ; . Fact 1. Let x be a regular walk in G I which starts in . Then it is possible to transform I in such a way that T f contains a branch t x , such that (x). This can be proved by a single loop on the length of starting from the end. Without loss of generality we can assume that xn is a 2-dimensional cell. We start by constructing in xn 2 a new ball vn 1 vn of colo This is easily done by applying ! 1 or ! 2 once. Assume next that we have constructed a subinstance whose cell graph is a path attached to x i . Again we apply ! 1 or ! 2 once in order to get v surrounding v i+1 vn . The same kind of induction (but starting from the beginning of the word this time) shows that the converse also holds: Fact 2. If it is possible to construct using operations in a new branch t in T f then there exists a walk x t in G I starting from E f such that M( (t)). We say that a walk x realizes a word W in I if It can be checked in PSPACE whether a given word is realized by a walk in G I starting in E f ; one way to see this is to reduce the problem to the question of whether two regular languages have a non-empty intersection Now we can prove Claim 1. We first transform I to I 0 . For each branch x of T i(f) starting from E i(f) , let W (x) and check whether there is a walk x 0 that realizes W x . If this is not the case, Fact 2 shows that I 6 J . If it is the case, use Fact 1 to construct the corresponding branch in I . To construct J 0 , do the same after reversing the role of I and J . It is clear that if the algorithm does not find out that J on its way, after minimizing the resulting instances they satisfy the claim.In the next section, it will be convenient to work with a slight simplification of the cell graph. We call a 2- dimensional cell B 6= EXT of an instance I inessential if B is homeomorphic to a disk D (and thus has precisely one neighbor in G I ), and the neighbor C of B in G I has another neighbor (B). Let H I be the graph obtained from G I by deleting all vertices that are inessential cells. We call H I the reduced cell graph of I . Then (5.1) actually holds with H I instead of G I : wn walk in H I with 6. Regular instances Recall that an instance I is regular if all points a 2 R 2 are regular. The main result of this section is that and t coincide on regular instances. As a corollary, we will see that the equivalence (4.1) does not extend beyond instances with one closed region. To illustrate where the problems are, let us start with a simple example: Example 6.1. Let consider the two instances I ; J with R I := fa 3 jjajj 1g, R J := S I , and S J := R I (cf. Figure 9). Obviously, I and J have the same cone-type. Figure 9. Let us try to find a sentence ' 2 FO top such that I and J 6j= '. At first glance this looks easy, just take the sentence saying "every horizontal line that intersects region R intersects region S before". Then every instance homeomorphic to I satisfies ', and every instance homeomorphic to J does not satisfy '. Unfortunately, ' 62 FO top . Figure 10 shows why: All three instances displayed are homeomorphic, but only the last one satisfies '. We will see later that there is a sentence ' 2 FO top that distinguishes I from J , but such a ' is quite complicated. For now, let us just note that I 6 J . Figure 10. Recall that by Lemma 3.3, the cell graph and thus the reduced cell graph of a regular instance is a tree. We think of this tree as being directed with the exterior as its root. The leaves are the 2-dimensional cells homeomorphic to the disk D. For instances I ; J we write H I H J if there is an embedding h of H I into H J that preserves . We define accordingly. Lemma 6.2. For all regular instances I ; J we have H I H J if, and only if, I and J are homeomorphic. Proof: The backward direction is trivial. For the forward direction, let I and J be regular instances with H I H J . Then it follows from Lemma 3.4 that G I G J . Since for regular instances the orientation O is empty, this implies Y I Thus I and J are homeomorphic by [PSV99].2 Recall the definition of the minimal instance M(I) associated with an instance I . An inspection the proof of Lemma 5.2 shows that for a regular instance I we have If M is a minimal regular instance, then the reduced cell graph H M has the following nice property: If C is a vertex of H M , then all neighbors of C in H M have different colors. (6.2) Lemma 6.3. Let M;N be minimal regular instances. Then Proof: This is an easy consequence of (5.2) and (6.2). 2 Recall that a regular language L is aperiodic if there exists an n 2 N such that for all u; v; w 2 such that uv n w 2 L we also have uv n+1 w 2 L. By a well-known theorem of McNaughton, Papert [MP71] and Sch-utzenberger [Sch65], the aperiodic languages are precisely the languages that are definable in first-order logic. 5 The following lemmas will be useful later. They are all based on (6.2) and the fact that the reduced cell graphs of regular instances are trees. 5 Furthermore, these are precisely the star-free regular languages. Lemma 6.4. Let M be a minimal regular instance. Then L(M) is aperiodic. Proof: Let M be a minimal regular instance. The crucial step is to prove the following claim: Let a walk in H M such that y Suppose for contradiction that (6.3) is wrong. Choose l minimal such that there is a v 2 l , an n 2, and a walk y Since adjacent vertices in a cell graph (and thus in a reduced cell graph) have different colors and (y l+1 ), we have l 2. For notational convenience, we let y 0 := y nl , y We choose an such that there is no nlg such that y j is in the subtree below y i . Then y is the parent of y i . Let := Then for 0 k n we have (y By (6.2), this implies y y kl+j+1 . If l = 2, this implies y tion. If l 3 we can define a walk y 0 from y by deleting y kl+j and y kl+j+1 , for 0 k n 1. Then (6.3) holds with in contradiction to the minimality of l. This proves (6.3). Now let n > jH M j and u; v; w 2 such that We shall prove that uv n+1 w 2 L(M). Let k; l; m be the length of u; v; w, respectively, and x := a walk in H M with there exist such that x Applying (6.3) to the walk But then is a walk with be a minimal regular instance and instances such that J Proof: Recall the definitions of the operations ! 1 and ! 2 from Figure 3. It suffices to prove the statement for instances L(M). Note that there is a word obtained from W by replacing some letters by subwords (if J !!1 J 0 ) or (if Suppose for contradiction that W 0 2 L(M). Then there is a walk x 0 in H M such that . By (6.2), whenever Hence n is a walk in H M with Similarly, if , we have x that a contradiction. This shows that W 0 Lemma 6.6. Let M be a minimal regular instance and J a regular instance with L(J) 6 L(M). Then there is a R such that for the curve f b : x 7! (x; b) we have Proof: We assume that J and M are instances over the same schema. Then, denoting by EXT and EXT 0 the exterior of J , M , respectively, we have For every walk define a sequence n as follows: We let x 0 i be the (by (6.2) unique) neighbor of x 0 1 that has the same color as x i , if x 0 and such a neighbor exists, and x 0 H J is a tree and H M satisfies (6.2), x j or x 0 m). It follows that for any walk or y 0 . Note that actually the sequence x 0 only depends on the word (x). This if then only if, x 0 Now suppose that with x 0 that xm+1 must be a child of xm in the tree H J . Choose b 2 R such that the curve f b intersects the cell xm+1 . Let l be a walk in H J that corresponds to the curve f b and let j < l be minimal such that y xm+1 . Then y because xm+1 is a child of xm . Then either y 0 m . In both case we have Theorem 6.7. Let I be a regular instance of schema . Then there is a sentence ' I 2 FO top of vocabulary f<g[ such that an instance J of the same schema satisfies ' I if, and only if, J I . Proof: Let be the set of colors appearing in I . For every aperiodic language L there is a formula L (y) 2 FO of vocabulary f<g [ such that a regular instance I satisfies L (b) for a b 2 R if, any only if, W (f b ; I) 2 L. This is an easy consequence of the theorem of McNaughton, Papert and Sch-utzenberger that the aperiodic languages are precisely the first-order definable languages. Let M := M(I) and L := L(M). By Lemma 6.4, L is aperiodic. Let Clearly, M satisfies M . We claim that for all instances J we have: To prove this claim, note first that every instance satisfying M realizes the same cones as I and thus is regular. Let J be a regular instance. Assume first that J Then by Lemma 6.6, L(J) L(M). Because J M(J), by Lemma 6.5 we have L(M(J)) L(M). Thus Conversely, if M(J) and thus J This proves (6.4). It follows easily that M 2 FO top . Indeed, assume that . For every J 0 homeomorphic to J we have We let N minimal regular instance with H N We shall prove that for any instance J of schema we have Suppose that J I . Then in M(J) I by (6.4). By Lemma 6.6, this implies L(M(J)) L(M) and minimal regular N with H N H M . By Lemma 6.3, this implies M(J) M . By Lemma 6.2, it follows that M and M(J) are homeomorphic. Thus J I . Conversely, suppose that J I . Then by Lemma 4.2, I . Thus J I , because ' I 2 FO top and I (the later follows easily from the previous paragraph and Corollary 6.8. For all regular instances I ; J we have I Finally, we are ready to prove that the equivalence (4.1) does not extend beyond instances with one closed region. Corollary 6.9. The two instances of Example 6.1 are not elementary equivalent. Neither are the instances I ; J over fRg defined by: R I := fa 7. Translating sentences to the topological in- variant Recall that it is proven in [SV98] that there is a recursive mapping that associates with every ' 2 FO of vocabulary f<; Rg a ' 0 2 FO of vocabulary d fRg such that for all instances I over fRg, where R I is a closed set, we have I The purpose of this section is to prove that this does not extend to arbitrary instances. Proposition 7.1. There is a sentence ' 2 FO top of vocabulary f<; Rg such that for every sentence ' 0 2 FO of vocabulary d fRg there is an instance I such that I G I 6j= ' 0 . Proof: Let I 0 ; J 0 be the instances defined by For n 1, let I n ; Jn be defined by R In := R I0 [ R In := R I0 [ a (cf. Figure 11). Figure 11. The instances I 3 and J 3 . Note that for all n; m 1 we have I n I m and Jn Jm , but I n 6 Jn . Corollary 6.8 implies that there is a sentence ' 2 FO top such that for all n 1 we have I n and Jn 6j= '. Let n 1. The graph G In is just a path with 5 tices, say, C . Denoting the colors by the colors on this path form the following sequence: | {z } times | {z } times G Jn is the same, except that . The other relations of the topological invariant are identical in Y In and Y Jn . dim 0 is empty, dim 1 consists of all C i with even i, and dim 2 consists of the C i with odd i. The orientation O is empty. Finally, g. Standard Ehrenfeucht-Fra-ss-e techniques show that for every sentence ' 0 2 FO there is an n 1 such that Y In ' if, and only if, Y Jn The statement of the proposition follows. 2 8. On completeness of languages An open problem that we have not considered so far is to find a (recursive) language that expresses precisely the first-order topological queries. Although this is certainly an interesting question, we doubt that, even if theoretically such a language exists, it would be a natural language that may serve as a practical query language. Our results show that first-order topological queries are not local; on the other hand, it is known that first-order logic fails to express the most natural non-local topological queries such as connectivity of a region. [KV99] have introduced a topological query language, the cone-logic CL, only expressing local properties. This language is a two tier language that allows to build first-order expressions whose atoms are again first-order expressions talking about the cones and colors of points. [KV99] have proven that CL captures precisely the first-order topological properties of instances with only one closed region that is "fully 2-dimensional". If the underlying structure is R< , it follows from [SV98] that the condition of being full 2-dimensional is not needed. Corollary 6.9 shows that this result does not extend to instances with several closed regions or one arbitrary region. We propose to extend CL by a path operator, as it has been introduced in [BGLS99]. Let us call the resulting topological query language PCL. The results of [BGLS99] show that this language has the basic properties expected from a reasonable spatial database query language. In addi- tion, it admits efficient query evaluation (the cost of evaluating a PCL-query is not substantially higher than the cost of evaluating a first-order query). An example of a PCL-query not expressible in FO is connectivity of regions. We conjecture that conversely every FO top -query is expressible in PCL. The idea behind this conjecture is that local properties (expressible in CL) together with the language of an instance, which can be described by the path-operator, seem to capture the first-order topological properties of an instance. As a first step towards proving this conjecture, let us remark that Corollary 6.8 implies that on regular instances every FO top -sentence is equivalent to a set of PCL-sentences. 9. Conclusions The results of this paper give a good understanding of first-order topological queries on regular instances. Of course one may argue that regular instances are completely irrelevant - look at any map and you will find singular points. However, we could use our results to answer several open questions concerning topological queries on arbitrary instances. As a matter of fact, we have shown that the previous understanding of first-order topological queries, viewing them as "local" in the sense that they can only speak about the colors of points, is insufficient; this may be our main contribution The problem of a characterization of topological elementary equivalence on arbitrary planar instances remains open. We conjecture that Corollary 6.8 generalizes, i.e. that t is the same as on arbitrary instances. If this was true, by Proposition 5.5 topological elementary equivalence would be decidable in PSPACE. Let us remark that we do not believe that the PSPACE-bound of Proposition 5.5 is optimal, we see no reason why should not be decidable in NP or even PTIME. --R Foundations of Databases. Relational expressive power of constraint query lan- guages Reachability and Connectivity Queries in Constraint Databases. Languages for relational databases over interpreted structures. Safe constraint queries. Finite Model Theory. Queries with arithmetical constraints. Finitely representable databases. Constraint Databases. Bart Kuijpers and Jan Van den Bussche. Topological queries in spatial databases. On finite monoids having only trivial subgroups. Querying spatial databases via topological invariants. --TR Towards a theory of spatial database queries (extended abstract) Queries with arithmetical constraints Finitely representable databases Relational expressive power of constraint query languages Topological queries in spatial databases Complete geometric query languages Reachability and connectivity queries in constraint databases Querying spatial databases via topological invariants Foundations of Databases Constraint Databases On Capturing First-Order Topological Properties of Planar Spatial Databases Fixed-Point Logics on Planar Graphs --CTR Michael Benedikt , Jan Van den Bussche , Christof Lding , Thomas Wilke, A characterization of first-order topological properties of planar spatial data, Proceedings of the twenty-third ACM SIGMOD-SIGACT-SIGART symposium on Principles of database systems, June 14-16, 2004, Paris, France Michael Benedikt , Bart Kuijpers , Christof Lding , Jan Van den Bussche , Thomas Wilke, A characterization of first-order topological properties of planar spatial data, Journal of the ACM (JACM), v.53 n.2, p.273-305, March 2006

constraint databases;first-order logic;topological queries

507385

Probabilistic game semantics.

A category of HO/N-style games and probabilistic strategies is developed where the possible choices of a strategy are quantified so as to give a measure of the likelihood of seeing a given play. A two-sided die is shown to be universal in this category, in the sense that any strategy breaks down into a composition between some deterministic strategy and that die. The interpretative power of the category is then demonstrated by delineating a Cartesian closed subcategory that provides a fully abstract model of a probabilistic extension of Idealized Algol.

Introduction Consider the fixpoint expression Y(x:(1 or x)). This can have two very different meanings attached. If, by or, it is meant that both sides can happen with, say, equal chances almost surely, we'll get the result 1. If, on the other hand, it is meant only that either can happen-but we don't know which-then the only consistent answer is to say that we may get 1 but we also may diverge. This latter meaning was investigated in a recent games model for nondeterminism by Harmer & McCusker [12]. In this paper, we construct a games model of the former by developing a suitable notion of probabilistic strategy. One motivation for doing this is to give a means of understanding quantitative aspects of uncertainty that might arise in the description of a process. Dually, inherently probabilistic processes-such as randomized optimization procedures, optimal strategies in classical game theory or more general adaptive behavior-can be accounted for in this framework. Another motivation is purely mathematical. Extending games semantics to the probabilistic case is a natural step to take, just as probabilistic strategies are extremely natural in VonNeumann matrix games. Indeed, our category smoothly extends the deterministic games world charted out by Hyland Ong [13] and Nickau [20] and further developed by McCusker [2] and Laird [15]. The category also fits in very nicely with the basic concepts of probability theory. The interaction of a strategy and a counter-strategy on a game gives rise to a Borelian sub- probability on the infinite plays. The category has a tensor product which corresponds to the classical product measure construction, and composition of strategies is defined using the image measure under the composition function (which in our case is always measurable). A factorization result is shown, saying that any probabilistic strategy can be broken down into the composition of a deterministic one and the 2- sided die. In that sense, the 2-sided die is shown to be uni- versal. This is a discrete version of the basic fact that any probability on [0; 1] can be reconstructed as the law of some random variable defined on the Lebesgue probability triple. In short, we give what seems the fundamentals of a higher-order probability theory in "interactive" form. A case study We eventually zero in on a Cartesian closed subcategory which we show provides a fully abstract model of a probabilistic extension of Idealized Algol. The exact choice of probabilistic primitive used to extend Algol is not overly important since, by our factorization result, any probabilistic process can be programmed using only a fair coin. The recent shifting of conditions from plays to strategies and other advances in the field allow, at this point, the use of a pretty standard machinery: we restrict to single-threaded strategies-those that are copyable and discardable with respect to the tensor product-and show that homsets admit a !-continuous CPO structure of which the basis is easily seen to be Algol-definable thanks to the factorization theorem and a previous theorem of Abramsky & McCusker [3]. For general reasons, this suffices to show full abstraction. Generality Our construction in no way depends on a particular games model: we give it in sufficient generality so that it can be adapted to whatever particular language we already know a deterministic games model of. And that is already quite a lot since, apart from the purely functional PCF, there are models for control, ground and general ref- erences. That games semantics concepts adapt in this way gives a sense of the homogeneity of this semantic universe, in due contrast with traditional denotational semantics. Related work The most obvious precursor to this work is the study of probabilistic powerdomains and their application to denotational semantics as in the early work of Saheb- Djahromi [22, 23] and of Jones & Plotkin [14] amongst oth- ers. More recent work [7, 8] has investigated bisimulation in a probabilistic context. It would be interesting to consider more closely the connection between that work and our own as this remains rather unclear for now. Another possible connection is with exact real arithmetic [9]. Intriguingly, several of the issues raised in our work, particularly those surrounding the factorization result, have a similar flavour to those raised in real arithmetic [21]. It would be interesting to pursue this line of investigation further and, indeed, to consider using game semantics to model a language such as Real PCF [10]. Future work One challenging application could be in probabilistic program analysis. Hankin and Malacaria's static analysis techniques derived from games semantics [16, 17, 18] could get on well with probabilities, and give tools in code optimization for instance. Less an application maybe, but still sort of practical, is the tempting question of whether the probabilistic observational equivalence can be shown to be decidable for our extension of Idealized Algol since, in that particular case, the so-called intrinsic equivalence in the intensional model has a simple concrete charac- terization. If solved to the positive, that could relate to formal methods in programming. A more theoretical question is that of relating our model to Harmer & McCusker's model [12]. There is an obvious embedding of theirs into ours, based on an equal chances interpretation of non-determinism, but it isn't fully abstract, as the opening example clearly shows. Another possibility is the investigation of a "timeless" version of our model in the spirit of the Baillot, Danos, Ehrhard & Regnier [6] construction starting from a symmetrized games model of Linear Logic and ending in a polarized variant of relational semantics. The concurrent games of Abramsky & Mel- lies [1] and their polarized version by Mellies could prove useful here, since they are an intermediate step between games and hypercoherences. Current investigation with Ehrhard shows that there are probabilistic coherence spaces that could bridge the gap, at least in the intuitionistic case. Game semantics represents computation with "execution traces" that describe the interaction between a System and its Environment. The System and Environment are modelled as the protagonists, respectively Player and Opponent, in a game; an execution trace by a sequence of moves played by the two protagonists. To make these ideas precise, we need a few bits of nota- tion. The set of strings over an alphabet \Sigma is written \Sigma ? . If s; t 2 \Sigma ? , we write s v t if s is a prefix of t, denote the length of s by jsj and, for 1 i jsj, denote the ith symbol of s by s i . The unique string of length 0 is written ". If for the restriction of s to \Sigma 0 , i.e. the subsequence obtained by erasing all symbols of s not in \Sigma 0 . 2.1 Arenas & legal plays The basic ground rules of a game are determined by the arena in which it is played. We formally define an arena as a triple hMA ; A ; 'A i where: ffl MA is a countable set of moves. Ag \Theta fI; ng is a labelling function designating each move m as an Opponent or a Player move, a Question or an Answer and as an Initial or non-initial move. We define OP A , QA A and In A as the first, second and third projections respectively from A . ffl 'A is a binary relation, known as enabling, on MA such A (m) 6= OP A (n) and In A (n) 6= I. In These conditions say that the protagonists enable each other's moves, not their own; that initial moves can't be en- abled; that only Opponent can have initial moves and, more- over, they must be Questions; and that Answers must be enabled by Questions. Flat arenas If X is a countable set, we define the flat arena over X by setting MX to be X making q 'X x for every x 2 X . We denote by C, B and N respectively the flat arenas over the one-point set fag, the two-point set ftt; ffg and the countably infinite set f0; Legal plays A string s A can be endowed with a "jus- tification structure" by equipping each non-initial move in s with a "pointer" pointing back to an earlier enabling move. This can be formalized with a function f s In then we say that then we call s i an initial occurrence. If f s then we say that s i is hereditarily justified by s jn . A string equipped with such an f s is called a justified string. A legal play is a justified string that strictly alternates between Opponent and Player: A (a) 6= OP A (b). We write LA for the set of all legal plays in the arena A; even A and L odd A denote the obvious subsets so that Player made the last move in s 2 L even A , etc. If s 2 LA , we write ie(s) for the set of all immediate extensions of s, i.e. those legal plays t such that s v t and 1. The current thread of sa 2 L odd A , written dsae, is defined as the subsequence of sa consisting of all moves hereditarily justified by the hereditary justifier of a. Product & arrow If A and B are arenas, we define their product A \Theta B by: the disjoint union. the copairing. This places A and B "side by side" with no chance of any interaction between them. The empty arena with sole legal play " is the unit for this constructor. Our other constructor is the arrow, defined by: OP In - In or In In A (n). In other words, the initial moves of A ) B are the initial moves of B, the r"oles of Opponent and Player are reversed in A and the (formerly) initial moves of A are now enabled by the (still) initial moves of B. 2.2 Strategies A strategy is a kind of "rule book" saying which moves may be made by Player and with what probability. First of all, a prestrategy oe on an arena A is just a function oe : L even The traces of oe, which we denote by T oe , are those even-length legal plays assigned non-zero weight by oe, i.e. The domain of oe, written dom(oe), is those odd-length plays that are "reachable" by oe, i.e. fsa 2 L odd g. Finally, given s 2 dom(oe), we define the range of oe at s, written rng oe (s), to be those immediate extensions of s that are in T oe , i.e. this can be empty. A prestrategy oe for A is a strategy iff A then oe(s) t2ie(sa) oe(t). Note that (p2) implies that oe is order reversing with respect to the prefix ordering on T oe and the usual ordering on [0; 1]. This means that T oe is even-length prefix-closed. Further- more, by (p1), we get oe(s) 1 for all s 2 LA . Basic constraints A strategy oe is deterministic iff for all Equivalently, oe takes values in f0; 1g. Note that (p2) asserts only an inequality. If oe is a strategy for which this is always in fact an equality, we say that it's a total strategy. This point is further discussed below. Local probabilities Given sa 2 dom(oe), define the conditional probability of sab given sa by: well-defined. By (p2), this gives a subprobability on rng oe (sa) for each sa 2 dom(oe). Intuitively speaking, oe(sab=sa) is the "die" that oe rolls for each sa 2 dom(oe). 2.3 Examples of strategies The fair coin The simplest example of a probabilistic strategy is probably the "fair coin" coin on the Boolean arena B. Its trace set is just the set of even-length sequences of the form qb 1 \Delta \Delta \Delta qb n where each b i is in ftt; ff g and is justified by the immediately preceding move. We assign a probability to such an s 2 T coin by e.g. etc. All local probabilities, such as oe(qffqtt=qffq), are 1=2. The polydie Another useful probabilistic strategy is polydie on N ) N. A typical play of polydie is: and the probability assigned to this play is 1=n. In other words, polydie asks for a number n and then rolls an n-sided die, returning something between 0 and 1 with uniform probability. Strategies on flat arenas as probability measures Plays in B of the form are in a 1-1 correspon- dence, denoted by B(\Delta), with finite binary strings over 1g. Given such a binary string w, we set xg. These Vw s are the basic open sets of the Cantor topology on infinite binary strings. Given a total strategy oe on B, we define p oe (V oe(s), which standardly extends to a unique probability measure on the Borelian oe-algebra B generated by the Vw s. We can now see that (p1) says that the measure of the whole space is 1, while (p2) says that p is countably additive (or subadditive if oe isn't assumed total). Conversely, any probability measure on B defines a total stategy on B. Such a probability measure could be any discrete-time stochastic process with values in f0; 1g and, in particular, it needn't be a Markov process. An example of this is the damped deviation die, denoted 3d, a strategy where the larger the deviation between trues and falses so far, the likelier it is to return that value which diminishes the deviation. For any play such as be the difference between the number of trues and falses in s, suppose d(s)+2 . Hence, 3d must be able to look arbitrarily far back in time. 2.4 Composing strategies Let u be a finite string of moves from arenas A, B and C equipped with candidate "justification pointers", i.e. a strictly decreasing function f 1g. We define uB;C to be the subsequence of u where we delete all moves from A and suitably "renor- f u to preserve the "pointers". We define u similarly. We define u A; C by removing all moves from and "pointers" to B except in the case where a 2 MA points to b 2 MB which, in turn, points to c 2 MC , whence we make a point directly to c in uA; C. A legal interaction of arenas A, B and C is such a u satisfying We write int(A; B; C) for the set of all legal interactions of A, B and C. If s 2 L even A)C , we define its set of witnesses by: i.e. all legal interactions playing s "on the outside". Given strategies oe and for A ) B and B ) C respectively, we define oe ; for A ) C by assigning a probability to with the following composition formula. We sometimes write oe k (u) for oe(u A; B) \Delta (uB; C). Note that oe and are "rolling" independent dice and that the above sum is always countable just because int(A; B; C) is. Well-defined composition It's clear that the composite of two prestrategies is itself a prestrategy (with the convention that also easy to see that (p1) is preserved. The lemma below does most of the rest of the work. Lemma 2.1 (probabilistic flow) If sa 2 L odd A)C and ua 2 C) such that u 2 wit(s), let W ua be the set of all ie(sa), i.e. all legal interactions starting like ua and witnessing an immediate extension of sa. If oe : A ) B and Proof We prove by induction that the interactions obtained by truncating W ua at any depth d, i.e. W ua dg, satisfy the inequality. Applying (p2) to a's component gives the base case. Other- wise, by the inductive hypothesis, oe k (u). Each u 0 2 W ua =d which hasn't yet reached the "outside" can have many extensions in W ua =d (p2) again, their sum cannot exceed the probability of u 0 . Hence Proposition 2.2 If oe and are strategies for A ) B and respectively then oe ; is a strategy for A ) C. Proof We just check (p2). If sa 2 L odd t2ie(sa) wit(t). For each u 2 wit(s), we have W ua as defined in lemma 2.1; clearly W ua ' W sa . Moreover, if W sa . We now apply lemma 2.1 to each witness u and its associated W ua so that and, since the desired result follows. \Xi Remark If both strategies are deterministic, i.e. with values in f0; 1g, then oe ; (s) is an integer which, by (p1) and (p2), can only be 0 or 1, so there can be at most one witness of s. In general, a play can have (infinitely) many witnesses. A category of arenas & strategies The above result tells us that we have a good notion of composition and it's routine to verify that this is associative. The usual copycat strategies are our identities: even s: sA where id A In summary, we have a category where objects are arenas and an arrow from A to B is a strategy for A ) B. We also have a subcategory of deterministic strategies since this class of strategies is closed under composition. 2.5 Examples of composing strategies Consider composing coin with the "wait for true" strategy defined as the semantics of Y(f: x: if x then skip else f(x)): Some typical interactions are: a ff a ff ff a The first witness of qa has probability 1=2, the second has probability 1=4, etc., so the overall probability of qa is 1=2+ If we use a biased coin instead-where the more time that's passed, the less likely it is to return true-the probability of becoming "trapped" in the loop can be non-zero. For instance, set coin 1+jsj=2 1. The overall probability of qa is now which can be anything between 0 and 1. Of course, for coin 0 to satisfy (p2) we must have, for all but this is the case since, by assumption, We can easily make coin 0 into a total strategy so that the class of total strategies can't be closed under composition: pieces of chance can get lost in infinite chatters. This is the reason why (p2) is an inequality. Another variation on this example uses the "count for true" strategy on B ) N defined as the semantics of Y(f: x: if x then 0 else 1+f(x)): In the composite with (the usual) coin, we get q0 with probability 1=2, q1 with probability 1=4, q2 with probability 1=8 and so on. In other words, the result is a geometric distribution on the natural numbers: we output n with probability Variations on a -scheme Consider the strategy mu 2 defined as the semantics of of type (Nat Com. This term takes an input f and feeds it successive arguments, beginning with 2, until it gives an output 0, at which point it just terminates. If we compose polydie with mu 2 , the behaviour is as fol- lows. Initially, a 2-sided die is rolled. If it turns up a 0, we're done; otherwise, a 3-sided die is rolled. If a 0 turns up this time, we're done; otherwise a 4-sided die is rolled, etc. The trace qa appears in the composite strategy with probability An interesting variation on this is to stop once a 0 has been rolled, but only continue with the larger die if a 1 was rolled; otherwise, we just give up. This can be done by replacing the 'if0' subterm with case f(x) of 0 7! skip If we compose polydie with the strategy rr associated with this modified -scheme term, the probability of qa in the composite is 1=2 0:718. This shows that the class of strategies where all probabilities are rational isn't closed under composition. 2.6 SMCC structure If oe and are strategies for A ) C and B ) D respectively then we define oe \Theta for (A\ThetaB) ) (C \ThetaD) in the following manner. Given s 2 T oe\Theta , we define its probability simply as the product of its probabilities in the two components: (oe \Theta Note that there is no interaction between oe and : they each have their own die. This makes \Theta into a bifunctor on our cat- egory. In fact, the evident copycat strategies provide us with a symmetric monoidal structure. In the usual way, a simple relabelling of moves gives rise to a natural isomorphism between oe on we have a symmetric monoidal closed category. The tensor is the terminal object. Finally, for an arena A, the unique strategy ! A on A ) 1 and the diagonal strategy \Delta A on A ) A with a cocommutative comonoid structure: for s 2 even (with probability 1) iff s: s means the subsequence of s 0 consisting of all moves hereditarily justified by an initial occurrence in A ' and s 0 r is defined similarly. 2.7 Ordering strategies Given strategies oe and for some arena A, we set oe 6 iff, for all s 2 LA , we have oe(s) (s). This is clearly a partial order. It's straightforward to show that composition is monotone with respect to this. CPO-enrichment In fact, we can show that each homset of our category is a CPO and that composition is continuous, i.e. our category "is" CPO-enriched. Proposition 2.3 Given a 6-directed set of strategies \Delta, if we define F \Delta by G \Deltag then it's the least upper bound (lub) of \Delta. Proof We show only that F \Delta is a well-defined strategy; that it is the lub follows easily. It's clear that F \Delta is a pre- strategy satisfying (p1). For (p2), it suffices to consider F \Delta). If rng F \Delta any n 2 N and ffl ? 0, we can find strategies such that i (sab the i s have an upper bound in \Delta so that G G G and the result follows. \Xi An argument with a similar flavour establishes that composition respects lubs of directed sets, i.e. if oe : A ) B, and \Delta is a directed set of strategies for then oe ; A countable basis For any arena A, the collection of strategies with finite trace set and where all probabilities are rational forms a basis for the CPO of all strategies. i.e. for any oe on A, the set \Delta oe of such strategies approximating oe is directed and has oe as its least upper bound. The only tricky point to check is that \Delta oe is indeed di- rected. Given oe with trace set the union of T oe 1 and T oe 2 , assigning probabilities as follows. If s is a maximal trace, we set oe 3 (s)g. Otherwise, set oe 3 (s) to be the maximum of oe 1 (s), oe 2 (s) and f g. Then, by con- struction, oe 3 is a valid strategy and is an upper bound for oe 1 and oe 2 in \Delta oe . This basis is always countable so that the set of all strategies for A is an !-continuous CPO. This contrasts with all previous HO-games models, including [12], where strategies form !-algebraic CPOs, the compact strategies being those with a finite number of traces. 2.8 Deterministic factorization The technique of factorization has become standard in game semantics. The underlying idea is similar to that of [12] where a nondeterministic strategy is reduced to a deterministic strategy that has access to a nondeterministic "oracle". In our case, the "oracle" strategy is a die. It suffices to use a 2-sided die but, for convenience, we use the polydie defined above. We first present a simplified factorization that works for basis strategies. Afterwards, we sketch a more general factorization that allows us to simulate any probabilistic strategy with a deterministic strategy and the polydie. Basis factorization Let oe be a basis strategy for arena A and consider sa 2 dom(oe). Since T oe is finite, we know that rng oe (sa) must be finite; we'll write it as rng oe g. Since all probabilities in oe are rational, the local probability oe(sab i =sa), for each sab i 2 rng oe (sa), is too. Theorem 2.4 (deterministic factorization) If oe is a basis strategy for the arena A then there exists a deterministic strategy Det(oe) for (N ) N) ) A such that polydie ; and Det(oe) is compact. Proof The above remarks imply that we can find a common denominator d for the local probabilities oe(sab i =sa). be the numerators such that p i =sa). The strategy Det(oe) proceeds in the following fashion. a d so on. In other words, we "slice up" the interval from 0 to d \Gamma 1 according to the numerators This clearly defines a deterministic strategy and it's easy to check that supplying the polydie as input achieves a correct simulation of the original strategy so that polydie ; obvious that, since oe has a finite trace set, Det(oe) must have too. \Xi Conversely, a small calculation with the composition formula shows that any strategy that can be factorized as polydie ; oe where oe is deterministic and compact must belong to the basis. General factorization To extend this idea to the general case, consider first a strategy oe all of whose probabilities are rational (but not necessarily having a finite trace set). If the local probabilities. If oe(sab 1 we roll a die. If the result is less than n 1 , we play b 1 ; other- wise, we renormalize the remaining probabilities by dividing through by then repeat this process for For example, if the local probabilities of b 1 and b 2 are 2=3 and 1=6 respectively, the following play produces b 2 . a We roll a 2-sided die the second time since the renormalized probability of b 2 is Finally, if a play in oe has irrational probability p, we can always find a sequence of rationals whose series of partial sums converges to p. So, for each point sab i in the range, we either have a single rational or a sequence of rationals. Using a standard trick (such as the one used to enumerate N \Theta N) we can reduce this to the previous case. The essential point of this is that a play with irrational probability is not factorized in "one go". In fact, it will have an infinite number of witnesses: the desired probability gradually builds up as we progress further and further down its sequence of rationals. 3 A fully abstract model of PA 3.1 Probabilistic Algol Our starting point is the language IA as defined in [3]. All we need to do is add in some probabilistic primitive. But should we add a coin or should we rather choose the polydie? Well, the polydie might be more handy to program with-indeed, the programminglanguageCAML has a "probabilistic func- tion" Random:int which exactly implements the polydie- but, for the contextual equivalence, it makes no difference whatsoever. In fact, the polydie can even be programmed inside Algol using only a coin. Informally it goes this way: given an n, choose a k such that 2 k n, then roll for a first round the 2-sided die k times. If the answer, which is a binary string of length k, codes for some integer in it; if not, go for another round, and so on, until you succeed. This idea can be programmed in a purely functional style. We give an example using CAML: let rec roll if k=0 then 0 else if coin () then (2* roll (k-1))+1 else (2* roll (k-1));; let rec polydie if res<n then res else polydie n;; where we assume 'findk n' returns dlog 2 ne and `coin ()' returns true or false with 50-50 odds. This coin can, in fact, also be written in CAML using the Random:int primitive mentioned above. The thunking is just because CAML is call-by-value. let coin if Random.int then true else false;; So, for the sake of simplicity, we just add a single term coin of type Bool. We call this extended language PA for probabilistic Algol. To complete the definition of the language, we need to give its operational interpretation. We do this in the usual "big step" style except that derivations must be decorated with probabilities. We have the obvious rules for the coin: The other rules follow naturally; we just give a few exam- ples. (For the sake of readability, we're eliding the use of stores; adding them poses no problem.) if M N L Because of coin, there might be (countably) many evaluations of M t be the sequence of their respective probabilities, it is easy to see that Finally, let M and N be closed terms of the same type. They are contextually equivalent iff, for all contexts C[\Gamma] of ground type, we have C[M . We denote this by M ' N . As an example, let Mn be the nth "Church numeral" 'h:h n (tt)' and N be `x: if coin ff x' of types (Bool ! We observe that \Gamman tt so that the context (\Gamma)N allows us to distinguish between all the Mn s. This demonstrates the discriminating power, even in the purely functional world, of probabilistic contexts, for it is easily seen that Mn and Mn+2 are contextually equivalent (for n 2) in PCF. 3.2 The model We now begin the process of building a model of PA by introducing the class of single-threaded strategies. Recall that the current thread of sa 2 L odd A is written as dsae. If we have some tab 2 L even A such that dtae and the justifier of b occurs in dtae, we denote by Match(tab; sa) the play sab where b is justified in the same way as in tab. We say that a strategy oe is single-threaded iff ffl for all sab 2 T oe , the justifier of b occurs in dsae. Note that this condition is expressed in terms of the local probabilities of oe. It is now a lengthy but routine matter to verify the following result. Proposition 3.1 A strategy oe for A ) B is single-threaded if, and only if, it's a comonoid homomorphism, i.e. For general reasons [11], this shows that the single-threaded strategies form a Cartesian closed category C. Fur- thermore, the CPO-enriched structure all restricts down to C, i.e. we have a CPO-enriched CCC. Visibility & bracketing The remaining constraints required to cut the model down to match well with Algol both rely on the following definition of the view, written psq, of a non-empty legal play s. is an O-move and justifies m; is a P-move. A strategy oe satisfies the visibility condition iff, for all sab 2 oe , the justifier of b occurs in psaq; and oe satisfies the bracketing condition iff, for all sab 2 T oe where b is an Answer, b is justified by the pending Question of psaq, i.e. the most recently asked but, as yet, unanswered Question in psaq. Both these constraints are preserved by composition. Our interpretation of PA is in the sub-CCC of C, denoted by C vb , of (single-threaded) strategies subject to both these constraints. The interpretation of IA, i.e. the deterministic fragment of the language, follows that of Abramsky & McCusker [3]; our probabilistic primitive coin is interpreted by the strategy coin. We have the following soundness result for our model of PA in C vb as a corollary to the usual consistency and adequacy results. Theorem 3.2 If M and N are closed terms of PA of type T such that [[M 3.3 Full abstraction With the soundness result in place, the path to full abstraction is relatively standard. First of all, we prove a definabil- ity result for basis strategies in the model. Theorem 3.3 If T is some type of PA and oe is a basis strategy on of PA. Proof By the factorization theorem, where Det(oe) is a compact deterministic strategy. By the definability theorem of Abramsky & McCusker [3], this strategy is definable in IA. But, since polydie is definable in PA, we're done. \Xi We define an equivalence relation on each homset of C vb as follows. Given f; iff for all "test where is the name of f , defined by currying The so-called intrinsic quotient category E vb obtained by quotienting C vb with this relation is also a CCC and, more- over, the above soundness result survives the quotient. The converse to soundness also holds there. Theorem 3.4 (full abstraction) If M and N are closed terms of type T then E Proof We only need to prove the right-to-left direction. Suppose that E [[M test ff, homsets in C vb are continuous CPOs, we can WLOG assume ff to be a basis element. Hence, by definability, it follows that this context distinguishes M and N . \Xi --R Concurrent games and full completeness. A fully abstract game semantics for general references. Linearity, sharing and state: a fully abstract game semantics for Idealized Algol with active expressions. Full abstraction for Idealized Algol with passive expressions. games. Bisimulation for labeled Markov processes. A logical characterization of bisimulation for labeled Markov processes. Domains for computation in mathematics PCF extended with real numbers. A fully abstract game semantics for finite nondeterminism. On full abstraction for PCF: I A probabilistic power- domain of evaluations Full abstraction for functional languages with control. Generalised flowcharts and games. A new approach to control flow analysis. Games and full abstraction for FPC. Hereditarily sequential functionals. Probabilistic LCF. CPOs of measures for nondeter- minism --TR Rational probability measures A probabilistic powerdomain of evaluations PCF extended with real numbers Full abstraction for idealized Algol with passive expressions On full abstraction for PCF Hereditarily Sequential Functionals Generalised Flowcharts and Games A New Approach to Control Flow Analysis Call-by-Value Games Reasoning about Idealized ALGOL Using Regular Languages Games and Full Abstraction for FPC Bisimulation for Labelled Markov Processes Full abstraction for functional languages with control Semantics of Exact Real Arithmetic A Logical Characterization of Bisimulation for Labeled Markov Processes A Fully Abstract Game Semantics for General References A Fully Abstract Game Semantics for Finite Nondeterminism Non-Deterministic Games and Program Analysis --CTR Pierre-Louis Curien, Definability and Full Abstraction, Electronic Notes in Theoretical Computer Science (ENTCS), 172, p.301-310, April, 2007

probabilistic Idealized Algol;games semantics

507388

Back and forth between guarded and modal logics.

Guarded fixed-point logic GF extends the guarded fragment by means of least and greatest fixed points, and thus plays the same role within the domain of guarded logics as the modal -calculus plays within the modal domain. We provide a semantic characterization of GF within an appropriate fragment of second-order logic, in terms of invariance under guarded bisimulation. The corresponding characterization of the modal -calculus, due to Janin and Walukiewicz, is lifted from the modal to the guarded domain by means of model theoretic translations. Guarded second-order logic, the fragment of second-order logic which is introduced in the context of our characterization theorem, captures a natural and robust level of expressiveness with several equivalent characterizations. For a wide range of issues in guarded logics it may take up a role similar to that of monadic second-order in relation to modal logics. At the more general methodological level, the translations between the guarded and modal domains make the intuitive analogy between guarded and modal logics available as a tool in the further analysis of the model theory of guarded logics.

Introduction Guarded logics generalise certain features and desirable model theoretic properties of modal logics to a much wider context. The concept of guarded quantification was introduced by Andr-eka, van Benthem, and N-emeti [1], who proposed and analysed the guarded fragment of first-order logic, GF. This fragment GF provides a very satisfactory basis for the explanation of quite some of the good behaviour of modal logics at the level of a rich fragment of classical first-order logic. Moreover, the robust decidability properties of modal logics with respect to natural extension mech- anisms, are also reflected in GF. Most notably, not only is GF decidable itself [1], but so is its canonical fixed point extension, -GF, [6]. -GF extends GF so as to render definable least (and greatest) fixed points of guardedly definable positive operations in arbitrary arities. In particular it extends the modal -calculus to the guarded domain. Unlike its modal companion, however, it no longer shares the finite model property, though it remains decidable. Another interesting feature, which again highlights the role of -GF as a high-level analogue of the -calculus, concerns its potential for model checking applications. The alternation-free fragment of -GF admits linear time model checking algorithms [4]. These and other results of recent research into guarded logics indicate that GF and its relatives provide a very interesting domain of logics, combining rich levels of expressiveness with a very good balance towards algorithmical issues. Recall how modal logics, in the broad model theoretic sense including extensions like CTL and the modal calculus, are characterised by their invariance under bisim- ulation: they cannot distinguish between bisimilar struc- tures. Moreover, invariance under bisimulation is at the root of many of the successful model theoretic tools available for modal logics, like e.g. the tree model property which paves the way towards the use of automata theory. The eminent role that bisimulation plays in the domain of modal log- ics, is in the guarded domain taken by a similar but much more wide-ranging and finer notion of equivalence induced by guarded bisimulation. The characteristic feature of modal quantification, which is also at the heart of bisimulation, is that one can only directly access nodes along basic edge relations. In the guarded scenario, this is generalised to simultaneous direct accessibility of all tuples that are covered (guarded) by some ground atom. These are what are called guarded tuples, and guarded quantification is quantification over guarded tuples. Unlike the modal case, the notion of guardedness does not impose any arity restrictions, so that guarded logics are just as meaningful over structures with higher-arity relations and can address properties of tuples. One of the underlying technical themes of this paper, also concerning a methodological point of wider interest, is the potential of having certain reductions from the richer scenario of guarded logics to the simpler and well understood scenario of modal logics, or from guarded bisimulation to ordinary bisimulation. Corresponding ideas have been initially explored in [5] in the context of the satisfiability problem for guarded fixed point logic. This method- # Mathematische Grundlagen der Informatik, RWTH Aachen, D-52065 Aachen, {graedel,hirsch}@informatik.rwth-aachen.de Department of Computer Science, University of Wales Swansea, SA2 8PP, United Kingdom, m.otto@swan.ac.uk ology is carried much further here and applied to a characterisation issue, concerning the semantic characterisation of guarded fixed point logic within a suitable second-order framework. Characterisation theorems of this type have a strong tradition in the field. They are of particular interest since they tighten the close connection between logics in a certain family and characteristic equivalences. At the level of basic modal logic and the guarded fragment of first-order logic, the corresponding characterisation theorems are the see [2] and [1]. Theorem 1.1 (van Benthem). A property of transition systems is definable in propositional modal logic if and only if it is first-order definable and invariant under bisimulation. Theorem 1.2 (Andr- eka, van Benthem, N- emeti). The guarded fragment GF can define precisely the model classes that are first-order definable and invariant under guarded bisimulation. For the modal scenario, a similar and highly non-trivial analogous characterisation was given for the associated fixed point logic in [7]. Theorem 1.3 (Janin, Walukiewicz). A property of transition systems is definable in the modal -calculus if and only if it is definable in monadic second-order logic and invariant under bisimulation. Johan van Benthem has raised the question whether -GF admits a similar characterisation, in terms of guarded bisimulation invariance, within some suitable framework of second-order logic. It should be noted that MSO is clearly not the right framework, since monadic second-order quantification does not suffice to simulate the fixed point definitions in -GF. Full second-order logic, on the other hand, is obviously too strong; there are simple examples showing that even for bisimulation invariant properties, full second-order logic goes far beyond the expressive power of -GF. The resulting fragment is, for instance, no longer decid- able. It turns out, however, that there is a natural analogue for MSO, which we call guarded second-order logic GSO. GSO is best characterised in semantic terms, as full second-order logic with a semantics that restricts second-order quantifiers to range over sets of guarded tuples, rather than over arbitrary relations. The precise definition will be given in section 3, where we also discuss several syntactic variants which turn out to have exactly the expressive power of GSO. And indeed we find the following. Main Theorem. Guarded fixed point logic -GF can define precisely the model classes that are definable in guarded second-order logic and are invariant under guarded bisimulation 2. Preliminaries Transition systems and bisimulation. Transition systems (or Kripke structures) are structures whose universe is a set of states (worlds), labelled by unary predicates (atomic propositions), and carrying binary transition relations labelled by actions (accessibility relations). We typically with state set V , based on a set B of atomic propositions and a set A of actions. Definition 2.1. A bisimulation between two transition systems respecting the P b in the sense that for all b # B and (v, v # Z, and satisfying the following back and forth conditions. for all (v, v # Z, a # A and every w such that (v, w) # E a , there exists a w # such that (v # , w # E # a and (w, w # Z. Back: for all (v, v # Z, a # A and every w # such that a , there exists a w such that (v, w) # E a and (w, w # Z. Two transition systems with distinguished nodes are bisimilar, K, u # K # , u # , if there is a bisimulation Z between them with (u, u # Z. We say that two trees are bisimilar if they are bisimilar at their roots # and # , and just Definition 2.2. The unravelling T (K, u) of a transition system K from node u is the tree of all paths through K that start at u. More formally, given (V, the unravelling is is the set of all sequences A such that v b contains the sequences v 0 a a contains the pairs (v, vav) in V T - V T . It is easy to see that each pair (K, u) is bisimilar to its unravelling: K, u # T (K, u), u. Hence, as far as bisimulation invariant properties are concerned, we can restrict our attention to trees. Bisimilar trees admit special minimal bisimulations, which may be constructed inductively, level by level starting from (# ) in such a way that at each new level we add a tuple (w, w # ) to Z only where this is required by a back or forth requirement from the previous level. It is clear that the resulting bisimulation is minimal in the sense that no proper subset would still be a bisimulation. Note that minimal bisimulations are in general not unique. They have, however, the following useful properties. Lemma 2.3. Let Z be a minimal bisimulation between T and T # . If (v, v # Z and if u and u # are the parent nodes of v and v # , respectively, then also (u, u # Z and (u, v) # E a iff (u # , v # E a . Therefore, a minimal bisimulation satisfies the back and forth requirements also with respect to the converses of the E a . The converse forth property w.r.t. E a , for instance, is the following. For every (v, v # Z and every u such that (u, v) # E a , there exists a u # such that (u # , v # E # a and (u, u # Z. Proof. It is straightforward to show that otherwise Z - would still be a bisimulation. It follows that w.r.t. a minimal bisimulation Z between trees T and T # and for (v, v # Z, any path from v in T can be lifted to a bisimilar path from v # in T # . In particular consider for any w in T the unique minimal connecting path (from v up to the first common ancestor of v and w, and from there down to w). Then there is a path (z # r labelled exactly the same as the original path, and such that (z i , z all i. 2.1. Modal logics We recall the definitions of propositional modal logic and the -calculus. The formulae of these logics are evaluated on Kripke structures at a particular state. Given a formula and a transition system K with state v, we write to denote that the formula holds in K at state v. Propositional modal logic. We describe propositional (multi-)modal logic ML for several transition rela- tions, i.e., for reasoning about transition systems (V, may have more than one element. In the literature on modal logic, this system is sometimes called Kn (where is the number of actions or 'modalities'). Syntax of ML. The formulae of ML are defined by the following rules. . Each atomic proposition P b is a formula. . If and # are formulae of ML, then so are ( #), ( #) and - . . If is a formula of ML and a # A is an action, then #a# and [a] are formulae of ML. If there is only one transition relation, simply writes # and # for [a] and #a#, respectively. Semantics of ML. Let # be a formula of ML, (V, state. In the case of atomic propositions, Boolean connectives are treated in the natural way. Finally for the semantics of the modal operators we put The -calculus L- . The propositional -calculus L - is propositional modal logic augmented with least and greatest fixed points. It subsumes almost all of the commonly used modal logics, in particular LTL, CTL, CTL # , PDL and also many logics used in other areas of computer science, for instance description logics. Syntax of L- . The -calculus extends propositional modal logic ML (including propositional variables X,Y, . , also viewed as monadic second-order variables) by the following rule for building fixed point formulae. . If # is a formula in L- , and X is a propositional variable that does not occur negatively in #, then -X.# and #X.# are L- formulae. Semantics of L- . The semantics of the -calculus is given as follows. A formula #(X) with propositional variable X defines on every transition system K (with state set V , and with interpretations for other free second-order variables that # may have besides X) an operator P(V ) on the powerset P(V ) of V assigning to every set the set As X occurs only positively in # K is monotone for every K, and therefore has a least and a greatest fixed point. Now we put is an element of the least fixed point of the operator # K . Similarly K, v |= #X.# iff v is an element of the greatest fixed point of # K . Remark. By the well known duality between least and greatest fixed points, #X.# is equivalent to Hence we could eliminate greatest fixed points. However, it will be more convenient to keep least and greatest fixed points and to work with formulae in negation normal form, where negations are applied only to atomic propositions. There is a variant of L - which admits systems of simultaneous fixed points. These do not increase the expressive power but sometimes allow for more straightforward formalisations. Here one associates with any tuple all positive in the X i , a new formula -X. The semantics of # is induced by the least fixed point of the monotone operator # K mapping X to X # where X # precisely, K, v |= # iff v is an element of the first component of the least fixed point of the above operator. Similar conventions apply w.r.t. simultaneous greatest fixed point. It is well known that simultaneous fixed points can be uniformly eliminated in favour of indi- vidual, nested fixed points. It is easy to see that all formulae of ML and L - are invariant under bisimulation. 3. Guarded first and second-order logics Definition 3.1. Let B be a structure with universe B and vocabulary # . (i) A set X # B is guarded in B if there exists a ground atom is guarded in B if (b 1 , . , b n is a guarded list in B if its components are pairwise distinct and {b 1 , . , b k } is a guarded set. We admit the empty list # as a guarded list. is guarded if it only consists of guarded tuples. Note that a singleton set guarded by the atom b. The cardinality of guarded sets in B is bounded by the maximal arity of the relations in # , the width of # . Guarded tuples, however, can have any length. Guarded lists will be of technical interest later as succinct tuple representations of guarded subsets. The guarded fragment GF. The guarded fragment extends modal logic to a richer fragment of first-order logic. Its characteristic feature is a relativised pattern of quantifica- tion, which generalises modal quantification. Syntax of GF. The formulae of GF (in vocabulary # ) are defined by the following rules. are formulae of GF. (ii) If # and # are formulae of GF, then so are (#), (#) and -#. (iii) If #(x, y), with free variables among those listed, is a formula of GF and #(x, y) is a #-atom in which all displayed variables actually occur, then #y #(x, y) # are formulae of GF. The atoms # relativising GF quantifications are called guards. We shall often also use the more intuitive notation (#y . # and (#y . # as shorthand for correspondingly relativised first-order quantification. Semantics of GF. The semantics of GF is the usual one for first-order formulae. As one simple example of a formula in GF, which will be useful later, consider the formula guarded(x) in variables #-structures the set of all guarded k-tuples, cf. Definition 3.1. For any complete equality type on {1, . , k} specified by a quantifier-free formula #(x) in the language of just = , let x # be the sub- tuple of x comprising precisely one variable from each =- class specified by #. Let #(y, x # ) be a #-atom in which all variables in x # actually occur, the y new, i.e. disjoint form x. Put # (x) := #(x) #y#(y, x # ). For the degenerate case of # 0 , specifying the equality type of a singleton tuple It is easily checked that the disjunction over all these formulae # (x) is as desired. Note that first-order quantification over an individual single free variable is always admissible in GF, since singletons are guarded (by an =-atom): Guarded fixed point logic -GF. Guarded fixed point logic -GF as introduced in [6] is the natural extension of GF by means of least and greatest fixed points (or corresponding systems of simultaneous fixed points). Syntax of -GF. Starting from GF, with second-order variables X,Y, Z, . that are treated like predicates in # but may not be used in guards, we augment the syntax of GF by the following rule for least and greatest fixed points. . If #(X, x) is a formula of -GF, which is positive in X , X k-ary and (first-order) variables of # are among these x i , then -X.# and #X.# are also formulae of -GF. Semantics of -GF. The semantics of -X.# is the natural one associated with the least fixed point of the monotone . More precisely, is an element of the least fixed point of the operator # B . Similarly for #X.# and the greatest fixed point of # B . One may also admit simultaneous least (and great- est) fixed points w.r.t. tuples of formulae the X i have only positive occurrences, and the x i contain match the arities of the X i . Then we obtain new whose semantics is the natural one associated with the first component X 1 of the least or greatest fixed point of the corresponding monotone operator. As with the -calculus, one finds that simultaneous fixed points can be eliminated in -GF, too. Guarded second-order logic. We introduce the natural second-order extension of the guarded fragment, or the guarded fragment of second-order logic, which relative to GF and -GF occupies a role analogous to that of MSO relative to ML and L - . The naturalness of this logic is further demonstrated below, where we show that the corresponding level of expressiveness is surprisingly robust under a number of changes in the actual formalisation and syntax. Indeed, we shall show that three natural candidates for a second-order guarded logic all have the same expressive power. It should be stressed that for all considerations, guardedness (of sets, tuples, or relations) always refers to guardedness w.r.t. the underlying vocabulary # ; at no point will second-order variables be admitted as guards. In our preferred definition of guarded second-order logic we simply use the syntax of ordinary second-order logic, but restrict it semantically by the stipulation that all second-order quantifiers range just over guarded relations, rather than over arbitrary relations. We refer to this semantic restriction as guarded semantics for the second-order quanti- fiers. It is clear, however, that this stipulation may alternatively be captured purely syntactically, by only allowing occurrences of atoms Xx in conjunction with the GF-formula guarded(x), which says that x is a guarded tuple, thus effectively restricting X to its guarded part. Definition 3.2. Guarded second-order logic GSO is second-order logic with guarded semantics for the second-order quantifiers. Note that GSO includes full first-order logic. Hence GSO is undecidable and, unlike GF and -GF, not invariant under guarded bisimulation (cf. Definition 4.2). Also note that, as singletons are always guarded, the monadic version of guarded second-order logic coincides with full MSO. Consequently, since MSO is strictly more expressive than FO, the same is true for GSO. Furthermore, we shall see in Lemma 3.5 below that GSO collapses to MSO over words. The robustness of GSO, and its place properly in between MSO and full second-order SO, is underlined by the following. Lemma 3.3. The following fragments of second-order logic are equally expressive (with respect to sentences): (1) The extension of GF by full second-order quantification (2) The extension of GF by second-order quantification with guarded semantics. (3) Guarded second-order logic GSO. Proof. It suffices to argue for translations from (1) and (3) into (2). For (1) # (2) consider a second-order variable X in a formula according to (1), which is meant to range over arbitrary rather than guarded relations. Any atom Xx occurring in a GF sentence necessarily is in the scope of a guarded quantification (Qy . #(y, z))# where the occurrence of x in Xx is free in #, whence the x all occur in #. It follows that the truth value of Xx for non-guarded tuples has no impact on the truth value of #. For (3) # (2) it suffices to show that unrestricted first-order quantification can be simulated by guarded (in fact: monadic) second-order quantification over GF. To this end, each element variable x is replaced by a set variable X , and we use the following rules for translating formulae. Rx #x . Rx) # i where singleton(X) is a formula stating that X contains exactly one element: Note that these translations and in particular singleton(X) are in GF, since first-order quantification over a single free first-order variable is always guarded. The following two lemmas show that GSO lies strictly between MSO and SO. Lemma 3.4. GSO is strictly more expressive than MSO. Proof. We show that the existence of a Hamiltonian cycle in an undirected graph can be expressed in GSO. It is known, however, that Hamiltonicity is not expressible in MSO, see e.g. [3]. Here is a GSO-sentence expressing Hamiltonicity: (#x#y(Hxy #zHxz # Lemma 3.5. SO is strictly more expressive than GSO. In particular GSO collapses to MSO over words. Proof. We can show that every guarded set over words can be encoded into a series of monadic predicates. The edge or successor relation is the predicate of maximal arity occuring in structures encoding words. Hence any guarded set contains at most two subsequent nodes. Due to the directedness of the edges we can simply choose the element a to represent the guarded set {a, b}, where b is the successor of a. The encoding uses one monadic predicate for each second-order predicate and each possibility of forming a tuple of appropriate arity out of two variables. Hence GSO is not more expressive than MSO over words, i.e. able to define exactly the regular languages. On the other hand full second-order logic is known to capture the polynomial-time hierarchy. To summarise, we have the following hierarchy of logics 3.1. A normal form for guarded logics. We present a normal form for GF and GSO that will be useful in the following. Let . } be two disjoint sets of variables. Let Z stand for either X or Y . GFX and GF Y are defined inductively as follows. (1) Every relational atomic formula # with belongs to GFZ . (2) A boolean combination of formulae in GFZ also belongs to GFZ . be any partial bijection between {1, . , n} and {1, . , m}. Then, for every guard 1 #(y 1 , . , yn ) and #(y 1 , . , yn #y. #(i)=j y is in GFX . By interchanging x- and y-variables we obtain an analogous rule for GF Y . It should be noted that the formulae in GFX and GF Y are syntactically not in GF. It is clear, however, that these are logically equivalent to guarded ones. Let These syntactic stipulations extend from GF to GSO in the obvious way. We let GSO 0 be the extension of GF 0 by second-order quantification over guarded relations. In the sequel, relativised quantifications of the type (#y. #(i)=j y as used in GF 0 will be abbreviated as (#= y . #(x, y) #(y)). Proposition 3.6. Every sentence in GF is equivalent to a sentence in GF 0 . Corollary 3.7. Every sentence in GSO is equivalent to a sentence in GSO 0 . The proof is not difficult but a bit technical. We just explain the idea. First, it is well known, that first-order sentences can be reformulated so that in all subformulae distinct variables are always to be interpreted by distinct elements. One way to make this precise is to use the quantifier #= rather than # where means that there exists an x that is distinct from z 1 , . , z n such that #(x, z 1 , . , z n ) holds. Obviously, this does not change the expressive power of first-order sentences, since #x #(x, z 1 , . , z n ) is equivalent to # n combine this with the idea that any quantification over a part of the free variables in a formula should come with a complete renaming of all variables, so that we move, say, from x-variables to y-variables. Consider a formula of the form We then replace this by the equivalent formula (#= y 1 - y 4 . y (Here the notation #= y means that the quantified y-variables must assume distinct values, but the values of a y-variable can be the same as for an x-variable.) To see why this might be useful (beyond the applications in this paper actually) let us consider another simple exam- ple. In the evaluation game for the sentence the verifier first has to pick distinct elements a 1 , . , a 4 and is then challenged by the falsifier to justify one of the three conjuncts. Suppose that she has to justify the subformula This means that she has to keep a 2 and extend it by some new a 1 and a 5 so that she can win the evaluation game for #(a 1 , a 5 , a 2 ) holds. Similarly, if she is challenged to verify the third conjunct she has to produce a new a 3 so that she wins the game for #(a 3 , a 4 , a 3 ). This becomes more transparent if we reformulate in GF 0 as Indeed this formulation makes it apparent that the verifier moves from a tuple to a new tuple (y 1 , y 2 , y 3 or (y 1 , y 2 ) subject to explicitly given equality constraints. 4. Guarded bisimulation and tree representation Guarded bisimulation is for GF what bisimulation is for ML. Definition 4.1. A guarded bisimulation between two # - structures A and B is a non-empty set I of finite partial from A to B, where X # A and are guarded sets, such that the following back and forth conditions are satisfied. For every We require that all y i occur in #, however order and multiplicity is arbitrary! for every guarded set X # A there exists a partial in I such that f and g agree on Back: for every guarded set Y # B there exists a partial in I such that f -1 and g -1 agree on Y # Y # . Two #-structures A and B are guarded bisimilar (in sym- bols: A # g B) if there exists a guarded bisimulation between them. We refer to the relation # g , which obviously is an equivalence relation between structures, as guarded bisimulation equivalence. Definition 4.2. We say that a sentence # is invariant under guarded bisimulation if it does not distinguish between guarded bisimilar structures, i.e. if A # g B and A |= # then also B |= #. A logic L is invariant under guarded bisimulation if all its sentences are. Proposition 4.3. GF and -GF are invariant under guarded bisimulation. The guarded Ehrenfeucht-Fra-ss- e game. It is well understood how guarded bisimulation equivalence may be described by means of an associated Ehrenfeucht-Fra-ss-e game. We indicate the characteristic features of the guarded game, in a version directly relating to our GF 0 normal form. Two players, player I and player II, take turns to place and relocate two groups of corresponding, labelled pebbles on elements of the underlying structures A and B, respectively. The rules of the game are such that after each round the groups of pebbles positioned in A and B, respectively, label guarded lists a and b in such a way that the correspondence a # b is a partial isomorphism. In each round, player I has the choice in which of the two structures to play. In the chosen structure, player I then may leave some pebbles fixed, remove some from the board, and re-locate some oth- ers. The only restriction is that the new pebble positions again have to label a guarded list. Player II then has to re-position the corresponding pebbles in the other structure so as to produce a locally isomorphic configuration. Player II loses if no such response is available. Now A and B are guarded bisimilar iff player II can always respond indefi- nitely, i.e. iff there is a winning strategy for player II in the infinite guarded game. It is apparent that a guarded bisimulation in the sense of Definition 4.1 is a formalisation of a (non-deterministic) strategy for player II. We now use the intuition behind the game to introduce the following structural transformations. One which abstracts from a given #-structure B a tree representation T (B), which fully describes as a transition system the behaviour of B in the guarded game, and thus characterises B up to guarded bisimulation equivalence. Another one which, conversely, associates with a tree T a #-structure D(T ), such that the game behaviour specified by T is realised in the guarded game on D(T ). The guiding idea behind these transformations is that guarded bisimulations at the level of the #-structures lift to ordinary bisimulations at the level of the abstracted transition systems. In particular, B # := D(T (B)) # g B will be a tree-like variant of B, intuitively corresponding to a guarded unravelling of B, as considered e.g. in [1] and [5]. From structures to trees. Recall from Definition 3.1 that a guarded list in B is a tuple (b 1 , . , b k ) of distinct elements such that {b 1 , . , b k } is a guarded set. We regard such guarded lists as descriptions of positions over B in the restricted guarded game. The nodes of the induced tree are associated with guarded lists over B, and we symmetrise the description of the game so as to allow re-labellings (permu- tations) of the guarded list in conjunction with every move in the game. The information to be recorded in each node precisely the isomorphism type of the induced substructure on {b 1 , . , b k } in B. The information to be recorded concerning possible moves describes the constraints imposed on a move from l ) to some choice of elements that remain fixed (according to player I's choice). For a #-structure B we work with the following vocabulary - # for the associated tree. If m is the width (maximal arity of predicates) of # , let S be the set of all #-structures A with universe {1, . , k}, admitted F be the set of all pairs (k, #), where is a partial bijection from {1, . , k} to {1, . , m}. Then - # has monadic predicates PA for all A # S, and binary predicates E k # for all (k, # F . Let G be the set of all guarded lists over B. Define V as the set of all sequences all have | and for all j # dom(#) the j-th element at g i+1 is the same as the #(j)-th element at g i . A node . gn of T (B) is naturally associated with the guarded list in B, in which the sequence terminates. In particular we set |v| := |g n and let A v be the unique A # S that is isomorphic with . Then the tree associated with B is # . The following is a direct consequence of the fact that the transition system T (B) is precisely set up to capture the behaviour of B w.r.t. the guarded game. Lemma 4.4. For all #-structures B and From trees to structures. Conversely, we would like to associate with a tree T of type - # a #-structure D for which . This is clearly not possible for arbitrary T . Definition 4.5. Let T be any -tree. We call T consistent if the following are satisfied. (a) For each node v there is a unique A # S such that T |= PA (v), denoted A v . A #. (b) If (u, v) # then the partial bijection # is an isomorphism between A v # dom(#) and A u # im(#), and for each A v from some structure A # S to A v there exists a node w such that and (u, w) # . Lemma 4.6. Within the class of -trees, the class of consistent trees is bisimulation invariant and first-order definable. Every tree T (B) is consistent. Conversely, if consistent -tree, then we may define an associated #-structure D(T ) as follows. Let # be the reflexive and symmetric transitive closure of the following relation on U : (v, The universe of D(T ) is D := U/# , the set of #- equivalence classes [v, i] in U . We say that an equivalence class lives at node u if [v, We observe that if an equivalence class d lives at nodes u and v in T then it must also live at every node of the unique shortest path connecting u to v in T . It follows from consistency condition (b), that we may consistently interpret every k-ary predicate in # over D(T ) by putting Note that guarded list in D(T ) if and only if of size |w| = k. We say that v represents the guarded list d. Lemma 4.7. For all #-structures B: D(T (B)) # g B. Proof. Let that T is the unravelling of K, whence any node v of T may be associated with a unique node v 0 in K (the last node in the unravelled path that gives rise to v). Further recall that being a node in K, is a guarded list in B. For any guarded list d in D and any node v of T representing this guarded list, let f d,v be the bijection which maps d to the guarded list v 0 in B. Note that as a map f d,v only depends on the guarded set {d 1 , . , d k } and not on the order of components or on the chosen representative v. If v and v # represent the same guarded set, they are linked in T by a path along which all components always live to- gether. The corresponding path in K similarly links v 0 to matching permutations along that path. We claim that the set of all these f d,v is a guarded bisimulation between D and B. It is obvious from the construction that the f d,v are partial isomorphisms whose domains are guarded sets. The forth property is immediate from the construction. We indicate the argument for the back prop- erty. Let d and d # be guarded lists in D, f d,v : d # v 0 . Let c be the tuple of common components in d and d # . Let v # be any representative of the guarded list d # . The tuple c lives along the unique shortest path from v to v # in T . Projecting this path down to K we see that f d # ,v # and f d,v agree on their common domain. may be regarded as a guarded unravelling of B, analogous to the standard unravelling of transition systems. Indeed, the resulting guarded bisimilar structure B # is also tree-like in that it has a tree decomposition of width m- 1, where m is the width of # . The naked tree T (B), stripped of its labels and regarded as a directed graph, together with the natural association of a node v of with the guarded set it represents in B # , induces a tree decomposition of B # in the usual sense. Tree unrav- ellings, and the corresponding generalised tree model property for GF and some of its relatives have been put to use e.g. in [1] and [5]. The following proposition extends the intuition that the bisimulation type of T (B) captures the guarded bisimulation type of B to the setting of all consistent trees. The proof is via a canonical lift of tree bisimulations. Proposition 4.8. For any consistent - -trees T and Proof. Let Z # V -V # be a minimal bisimulation between T and T # , cf. Lemma 2.3. For each pair (v, v # Z let f vv # be the function that maps the guarded list represented by v to that represented by v # . Clearly, f vv # is a partial isomorphism, since A For # and # , the maps f uu # and f vv # agree on their common domain. This follows from the construction of D(T ) and D(T # ), because elements represented at u and v are identified in D(T ) via # in exactly the same way as the corresponding elements at are identified in D(T # ). We claim that the set of all f vv # , for (v, v # Z, is a guarded bisimulation between D(T ) and D(T # ). To verify, for instance, the forth condition, let c and d be guarded lists in D(T ), represented by u and v in T , respec- tively. Let (u, u # Z, f We need to find such that (v, v # Z and such that f vv # agrees with f uu # on their common domain. Let X be the set of common elements in c and d. Consider the unique shortest path from u to v in T . As Z is minimal, this path gives rise to a bisimilar path from u # to some v # in T . Now f vv # is as desired: the elements of X live at all nodes on the path from u to v, whence all the intermediate mappings fww # along the path, and in particular f vv # respect f uu #X . 5. Back and forth Recall the characterisation of the modal -calculus from Theorem 1.3. We want to apply this characterisation in restriction to trees, and therefore refer to the following vari- ant, which is proved en route to Theorem 1.3 in [7]. Theorem 5.1 (Janin, Walukiewicz). A class of trees is definable in the modal -calculus if and only if it is definable in monadic second-order logic and closed under bisimulation within the class of all trees. Towards a reduction of our main theorem to Theorem 5.1, we define a "forth" translation that maps every sentence # GSO[# ] to a formula # (x) # MSO[-# ] with translation that maps every to a sentence # -GF[# ]. These translation will be such that (1) If T is a consistent tree with root #, then T |= #) iff D(T ) |= #. (2) If B is a #-structure and # is the root of T (B), then It follows from (1) and Proposition 4.8 that GSO sentences that are invariant under guarded bisimulation are mapped to MSO formulae that are bisimulation invariant on consistent trees. Before giving the formal definitions, we informally discuss the main problems arising from the differences between the guarded and the modal viewpoint. We wish to translate GSO sentences to MSO formulae and L- formulae back to -GF sentences. We want a modal formula to hold at some node v if and only if the corresponding guarded formula holds of the guarded list represented by v. For second-order variables we need to map sets of guarded tuples to sets of nodes and vice versa. For each node v of a consistent tree T , the associated structure A v has in general many different guarded tuples that can occur in any single guarded set. Therefore a second-order variable Z in a GSO sentence will be translated into a sequence Z # of set variables Z i 1 ,.,i r , one for each (lo- cal) choice of elements. In the other direction we have to deal with monadic second-order variables, which in general range over sets of nodes of arbitrary size. Consequently a monadic second-order variable X is translated into a sequence that each X i ranges over guarded tuples of length i. 5.1. From guarded to monadic second-order logic. Without loss of generality we restrict attention to GSO sentences in GSO 0 , cf. Proposition 3.7. Let m be the width of # , i.e. the maximal arity of relations in # . Definition 5.2. Let T be a consistent -tree, and let D(T ) be the associated #-structure. If Z is an r-ary second-order variable, then Z # := (Z is the corresponding sequence of monadic predicates. A tuple J # of monadic predicates J on T encodes an r-ary guarded relation J on D(T ) iff J Not all sequences J of monadic predicates over T do indeed encode a guarded relation over D(T ). To do so, they have to satisfy the following correctness conditions. (a) only contains nodes v where all i j are in A v . (b) J is consistent on tuples living at different nodes, i.e. if in D(T ) a tuple (d 1 , . , d r ) is represented by (i 1 , . , i r ) at node u and by (j 1 , . , j r ) at node v, then Lemma 5.3. For each r # m there exists a first-order that expresses the correctness conditions These conditions are necessary and sufficient in the sense that a tuple J encodes a guarded relation on D(T ), if and only if T |= correct(J). Proof. Note that it suffices to express condition (b) for adjacent nodes to enforce it globally. Thus the consistency requirement for Z # can be expressed by a first order formula states condition (a) in an obvious way and con- tains, for each pair (k, # F and each tuple (i 1 , . , i r ) a clause The proof of the adequacy claim is straightforward. First-order quantifiers also require a special treatment. Let us first consider the case where T is a tree representation (B). By construction such trees satisfy a strong homogeneity condition, namely for all nodes u, u # and all successors v of u there is a successor v # of u # such that the subtrees with roots v and v # are isomorphic. To put it differently, an indistinguishable copy of any guarded list anywhere in D(T ) is available locally, at some child of the current node. Therefore guarded first-order quantifications over D(T ) can be simulated over T by moving to an immediate successor of the current node (i.e. by a modal quantifier # or #). However, if T is an arbitrary consistent - # -tree, this is no longer the case. To verify a formula of the form (#= y . #(x, y) #(y))#(y) we want to move from the current tuple x to a new tuple y, guarded by #, such that #(y) is true and the overlap conditions for x and y as stated by #(x, y) are satisfied. In arbitrary consistent trees, such a witness need not exist locally, but may only occur as a remote node, which is linked to the current node by a path along which the common components according to # are kept. To capture this situation in MSO we use a sequence W of monadic predicates W k # , one for each (k, # F , that partition the set of nodes according to their size and their overlap with the current node. The proof of the following is then straightforward. Lemma 5.4. There is a first-order formula F-part(z, W ) expressing the following correctness conditions on partitions with respect to node z. For every consistent tree T , every sequence predicates on T , we have T |= F-part(u, W ) iff W k |Aw Proof. As in the previous lemma, it suffices to express the required condition at the current node and for every edge. The formula F-part(z, W ) states that (a) The sets W form a partition of the universe. (b) The node z itself belongs to W k id , where (c) If y # W k (d) For all (k, #) and then x # W # # . By induction on the distance from z one easily shows that F-part(z, W ) expresses the right property. The translation. Recall that formulae in GSO 0 either belong to GSOX and have all their free first-order variables in X , or they belong to GSO Y and have all their are two disjoint sets of variables. Further, formulae in GSOX that start with a quantifier are of the form #(x) := (#= y . #(x, y) # #(y))#(y) with #(y), #(y) # GSO Y . We inductively translate every GSOX [# into an MSO[-# r ), with a single free first-order variable x and sequences of monadic second-order variables i that correspond to the second-order variables Z i . Sim- ilarly, formulae in GSO Y are translated into formulae # y. We just present the translation for formulae in GSOX : r-ary relation variable (x). (#W . F-part(x, W ))(#y . W |y| (5) For #Z# with r-ary relation variable Z, let Theorem 5.5. Let T be a consistent -tree with root #, let D(T ) be the associated #-structure, and let # be a sentence in GSO[# ]. Then D(T Proof. This theorem is a consequence of the following more general statement. Consider any formula second-order variables as displayed. Its translation into MSO[-# guarded list in D(T ), and let v be a node of T such that be sets of guarded tuples in D(T ), and let J # s be their representations according to Definition 5.2. Note that for sentences (i.e. the claim implies the theorem. The claim itself is established inductively. The cases corresponding to (1) - (3) are immediate. (#W . F-part(x, W ))(#y . W |y| As second-order variables play no role for this case, we suppress them. Suppose that D(T ) |= #(d). Then there exists a guarded list that D(T ) |= #(e) #(e) and e e is guarded, there exists a node w such that all e i live together at w. Actually, due to the last of the conditions for consistent trees, we can assume that and |Aw | = #. We know that there exists a W satisfying that for all (i, [v, j]. It follows that T |= W # # (w) and, by induction hy- pothesis, T |= (# (w). Therefore T |= # (v). Conversely, suppose that T |= # (v). Hence for the (unique) tuple W satisfying F-part(v, W ), there exists a node w # W # # such that T |= (# (w). For e i := [w, i] we find that e and, by induction hy- pothesis, D(T ) |= (#)(e). Therefore D(T ) |= #(d). For (5), the claim is immediate from the induction hypothesis and from Lemma 5.3. Corollary 5.6. If # GSO[# ] is a sentence that is invariant under guarded bisimulation, then # (x) is bisimulation invariant on consistent -trees. Proof. Let T , T # be two bisimilar, consistent -trees. Then Proposition 4.8. It follows that T |= 5.2. From the -calculus to guarded fixed point logic. The translation back from the modal into the guarded world again requires some preliminary discussion. Every formula # of the -calculus, evaluated on a tree defines the set # T of all nodes v such that T , v |= #. Recall how each node v of T represents a guarded list g v in B. So the idea is to translate # into a guarded formula #(x) defining in B the set # should be equal to {g v #}. The main problem is that guarded lists do not have a fixed length, so that must actually be translated into a tuple Definition 5.7. Let X be a monadic second-order variable. is an i- ary second-order variable. Let N be a set of nodes in a tree T (B). The representation of N in B is N := Note that, since the root # is the only node of size 0, either we say that or N In the sequel we assume w.l.o.g. that L - formulae are written without #-operators, that in any formula -X.# the fixed point variable X occurs in # only inside the scope of modal operators, and even that in each # L - , the fixed point formulae -X.#(X) themselves only occur inside the scope of modal operators, cf. [8]. The translation. For every formula # L -# ] we now define one for each m, in which each monadic second-order variable X of # is represented by a tuple of second-order variables X # (of arities 0, 1, . , m). := false if |A| #= k, and be the conjuction over all atomic and negated atomic #-formulae # such that that for is the following simultaneous least fixed point, which can be simulated by nested individual fixed points in the standard way: Theorem 5.8. Let B be a #-structure, let # be the root of # be a formula in L -# T (B), #. Proof. Again we prove a more general statement involving variables. W.l.o.g., consider the case of just one monadic second-order variable Y . Let #(Y ) be a formula in L-# ], N a set of nodes in Claim. is the representation of #(N) T in B. The claim is proved inductively. The cases corresponding to (1) - (3) are trivial. Consider (4) and let #. Suppose that T , v |= #. Then there is a node w such that (v, w) # E # # and for all (i, and, by induction hypothesis, B |= Conversely, suppose that B |= # k (b 1 , . , b k ). This means that there exists a guarded list By the construction of T (B), there exists a node w such that (v, w) # E # # and The induction hypothesis implies that T , w |= #. It follows that T , v |= #. For (5) finally let Consider the stages of the fixed point induction on # 0 := false , #+1 := # for limit ordinals #. Let similarly, for the simultaneous fixed point induction on wise) in limits. By induction hypothesis, if M # is the representation of M in B, then the representation of #(M) T . By induction on # (together with the representation of the set defined by # in T . Hence the same is true for the least fixed points. We are now in a position to prove our main theorem. Theorem 5.9. Every sentence in GSO that is invariant under guarded bisimulation is equivalent to a sentence in -GF. Proof. Let # GSO[# ] be invariant under guarded bisimulation and let # (z) be its translation into MSO[-# ]. By Lemma 5.6 # (z) is bisimulation-invariant on consistent trees. Recall that the consistency condition for trees can be formulated by a monadic second-order sentence #, which is bisimulation invariant with respect to all trees. As a conse- quence, the formula (# )(x) is bisimulation invariant on arbitrary trees. By the Janin-Walukiewicz Theorem, Theorem 5.1 above, there exists an equivalent formula # in the -calculus. Let # 0 be its translation into -GF[# ]. Putting everything together, we have The first equivalence uses guarded bisimulation invariance of # and Lemma 4.7; the second one is an application of Theorem 5.5; the third equivalence reflects the input from the Janin Walukiewicz Theorem; the fourth is an application of Theorem 5.8. --R Modal languages and bounded fragments of predicate logic. Modal Logic and Classical Logic. Finite Model Theory. lite. The decidability of guarded fixed point logic. Guarded fixed point logic. On the expressive completeness of the propositional mu-calculus with respect to monadic second order logic Games for the mu-calculus --TR Automata on infinite objects Games for the MYAMPERSANDmgr;-calculus Reasoning in description logics Why are modal logics so robustly decidable? Dynamic Logic Modal Logic over Finite Structures Games and Model Checking for Guarded Logics Modal and Guarded Characterisation Theorems over Finite Transition Systems On the Expressive Completeness of the Propositional mu-Calculus with Respect to Monadic Second Order Logic The monadic second-order logic of graphs XIV Guarded Fixed Point Logic Back and Forth between Guarded and Modal Logics --CTR Dirk Leinders , Jan Van den Bussche, On the complexity of division and set joins in the relational algebra, Journal of Computer and System Sciences, v.73 n.4, p.538-549, June, 2007 Dirk Leinders , Jan Van den Bussche, On the complexity of division and set joins in the relational algebra, Proceedings of the twenty-fourth ACM SIGMOD-SIGACT-SIGART symposium on Principles of database systems, June 13-15, 2005, Baltimore, Maryland Dirk Leinders , Maarten Marx , Jerzy Tyszkiewicz , Jan Bussche, The Semijoin Algebra and the Guarded Fragment, Journal of Logic, Language and Information, v.14 n.3, p.331-343, June 2005 Erich Grdel , Wolfgang Thomas , Thomas Wilke, Literature, Automata logics, and infinite games: a guide to current research, Springer-Verlag New York, Inc., New York, NY, 2002

modal logic;bisimulation;model theory;guarded logic

507449

On the Quality of Service of Failure Detectors.

Editor's Note: This paper unfortunately contains some errors which led to the paper being reprinted in the May 2002 issue. Please see IEEE Transactions on Computers, vol. 51, no. 5, May 2002, pp. 561-580 for the correct paper (available without subscription).We study the quality of service (QoS) of failure detectors. By QoS, we mean a specification that quantifies 1) how fast the failure detector detects actual failures and 2) how well it avoids false detections. We first propose a set of QoS metrics to specify failure detectors for systems with probabilistic behaviors, i.e., for systems where message delays and message losses follow some probability distributions. We then give a new failure detector algorithm and analyze its QoS in terms of the proposed metrics. We show that, among a large class of failure detectors, the new algorithm is optimal with respect to some of these QoS metrics. Given a set of failure detector QoS requirements, we show how to compute the parameters of our algorithm so that it satisfies these requirements and we show how this can be done even if the probabilistic behavior of the system is not known. We then present some simulation results that show that the new failure detector algorithm provides a better QoS than an algorithm that is commonly used in practice. Finally, we suggest some ways to make our failure detector adaptiveto changes in the probabilistic behavior of the network.

Introduction Fault-tolerant distributed systems are designed to provide reliable and continuous service despite the failures of some of their components. A basic building block of such systems is the failure detector. Failure detectors are used in a wide variety of settings, such as network communication protocols [10], computer cluster management [23], group membership protocols [5, 9, 7, 27, 22, 21], etc. Roughly speaking, a failure detector provides some information on which processes have crashed. This information, typically given in the form of a list of suspects, is not always up-to-date or correct: Research partially supported by NSF grant CCR-9711403 and an Olin Fellowship. y Oracle Corporation, One Oracle Drive, Nashua, NH 03062, USA. Email: Wei.Chen@oracle.com z DIX Departement d'Informatique, Ecole Polytechnique, 91128 Palaiseau Cedex, France. Email: sam@dix.polytechnique.fr x Compaq Systems Research Center, 130 Lytton Avenue, Palo Alto, CA 94301-1044, USA. Email:aguilera@pa.dec.com a failure detector may take a long time to start suspecting a process that has crashed, and it may erroneously suspect a process that has not crashed (in practice this can be due to message losses and delays). Chandra and Toueg [12] provide the first formal specification of unreliable failure detectors and show that they can be used to solve some fundamental problems in distributed computing, namely, consensus and atomic broadcast. This approach was later used and generalized in other works, e.g., [20, 16, 17, 1, 3, 2]. In all of the above works, failure detectors are specified in terms of their eventual behavior (e.g., a process that crashes is eventually suspected). Such specifications are appropriate for asynchronous systems, in which there is no timing assumption whatsoever. 1 Many applications, however, have some timing constraints, and for such applications, failure detectors with eventual guarantees are not suf- ficient. For example, a failure detector that starts suspecting a process one hour after it crashed can be used to solve asynchronous consensus, but it is useless to an application that needs to solve many instances of consensus per minute. Applications that have timing constraints require failure detectors that provide a quality of service (QoS) with some quantitative timeliness guarantees. In this paper, we study the QoS of failure detectors in systems where message delays and message losses follow some probability distributions. We first propose a set of metrics that can be used to specify the QoS of a failure detector; these QoS metrics quantify (a) how fast it detects actual failures, and (b) how well it avoids false detections. We then give a new failure detector algorithm and analyze its QoS in terms of the proposed metrics. We show that, among a large class of failure detectors, the new algorithm is optimal with respect to some of these QoS metrics. Given a set of failure detector QoS requirements, we show how to compute the parameters of our algorithm so that it satisfies these requirements, and we show how this can be done even if the probabilistic behavior of the system is not known. Finally, we give simulation results showing that the new failure detector algorithm provides a better QoS than an algorithm that is commonly used in practice. The QoS specification and the analysis of our failure detector algorithm is based on the theory of stochastic processes. To the best of our knowledge, this work is the first comprehensive and systematic study of the QoS of failure detectors using probability theory. 1.1 On the QoS Specification of Failure Detectors We consider message-passing distributed systems in which processes may fail by crashing, and messages may be delayed or dropped by communication links. 2 A failure detector can be slow, i.e., it may take a long time to suspect a process that has crashed, and it can make mistakes, i.e., it may erroneously suspect some processes that are actually up (such a mistake is not necessarily permanent: the failure Even though the fail-aware failure detector of [17] is implemented in the "timed asynchronous" model, its specification is for the asynchronous model. We assume that process crashes are permanent, or, equivalently, that a process that recovers from a crash assumes a new identity. up trust suspect trust suspect FD at q Figure 1: Detection time T D detector may later stop suspecting this process). To be useful, a failure detector has to be reasonably fast and accurate. In this paper, we propose a set of metrics for the QoS specification of failure detectors. In general, these QoS metrics should be able to describe the failure detector's speed (how fast it detects crashes) and its accuracy (how well it avoids mistakes). Note that speed is with respect to processes that crash, while accuracy is with respect to processes that do not crash. A failure detector's speed is easy to measure: this is simply the time that elapses from the moment when a process p crashes to the time when the failure detector starts suspecting p permanently. This QoS metric, called detection time, is illustrated in Fig. 1. How do we measure a failure detector's accuracy? It turns out that determining a good set of accuracy metrics is a delicate task. To illustrate some of the subtleties involved, consider a system of two processes p and q connected by a lossy communication link, and suppose that the failure detector at q monitors process p. The output of the failure detector at q is either "I suspect that p has crashed" or "I trust that p is up", and it may alternate between these two outputs from time to time. For the purpose of measuring the accuracy of the failure detector at q, suppose that p does not crash. Consider an application that queries q's failure detector at random times. For such an application, a natural measure of accuracy is the probability that, when queried at a random time, the failure detector at q indicates correctly that p is up. This QoS metric is the query accuracy probability. For example, in Fig. 2, the query accuracy probability of FD 1 at q is 12=(12 The query accuracy probability, however, is not sufficient to fully describe the accuracy of a failure detector. To see this, we show in Fig. 2 two failure detectors FD 1 and FD 2 such that (a) they have the same query accuracy probability, but (b) FD 2 makes mistakes more frequently than FD 1 . 3 In some applications, every mistake causes a costly interrupt, and for such applications the mistake rate is an important accuracy metric. Note, however, that the mistake rate alone is not sufficient to characterize accuracy: as shown in Fig. 3, two failure detectors can have the same mistake rate, but different query accuracy probabilities. 3 The failure detector makes a mistake each time its output changes from "trust" to "suspect" while p is actually up. 1 . up Figure 2: FD 1 and FD 2 have the same query accuracy probability of :75, but the mistake rate of FD 2 is four times that of FD 1 up Figure 3: FD 1 and FD 2 have the same mistake rate 1=16, but the query accuracy probabilities of FD 1 and FD 2 are :75 and :50, respectively Even when used together, the above two accuracy metrics are still not sufficient. In fact, it is easy to find two failure detectors FD 1 and FD 2 , such that (a) FD 1 is better than FD 2 in both measures (i.e., it has a higher query accuracy probability and a lower mistake rate), but (b) FD 2 is better than FD 1 in another respect: specifically, whenever FD 2 makes a mistake, it corrects this mistake faster than FD in other words, the mistake durations in FD 2 are smaller than in FD 1 . Having small mistake durations may be important to some applications. As it can be seen from the above, there are several different aspects of accuracy that may be important to different applications, and each aspect has a corresponding accuracy metric. In this paper, we identify six accuracy metrics (since the behavior of a failure detector is probabilistic, most of these metrics are random variables). We then use the theory of stochastic processes to quantify the relation between these metrics. This analysis allows us to select two accuracy metrics as the primary ones in the sense that: (a) they are not redundant (one cannot be derived from the other), and (b) together, they can be used to derive the other four accuracy metrics. In summary, we show that the QoS specification of failure detectors can be given in terms of three basic metrics, namely, the detection time and the two primary accuracy metrics that we identified. Taken together, these metrics can be used to characterize and compare the QoS of failure detectors. 1.2 The Design and Analysis of a New Failure Detector Algorithm In this paper, we consider a simple system of two processes p and q, connected through a communication link. Process p may fail by crashing, and the link between p and q may delay or drop messages. Message delays and message losses follow some probabilistic distributions. Process q has a failure detector that monitors p and outputs either "I suspect that p has crashed" or "I trust that p is up" ("suspect p" and "trust p" in short, respectively). A Common Failure Detection Algorithm and its Drawbacks. A simple failure detection algorithm, commonly used in practice, works as follows: at regular time intervals, process p sends a heartbeat message to q; when q receives a heartbeat message, it trusts p and starts a timer with a fixed timeout value if the timer expires before q receives a newer heartbeat message from p, then q starts suspecting p. This algorithm has two undesirable characteristics; one regards its accuracy and the other its detection time, as we now explain. Consider the i-th heartbeat message m i . Intuitively, the probability of a premature timeout on m i should depend solely on m i , and in particular on m i 's delay. With the simple algorithm, however, the probability of a premature timeout on m i also depends on the heartbeat m that precedes In fact, the timer for m i is started upon the receipt of m i\Gamma1 , and so if m i\Gamma1 is "fast", the timer for m i starts early and this increases the probability of a premature timeout on m i . This dependency on past heartbeats is undesirable. To see the second problem, suppose p sends a heartbeat just before it crashes, and let d be the delay of this last heartbeat. In the simple algorithm, q would permanently suspect p only d+TO time units after crashes. Thus, the worst-case detection time for this algorithm is the maximum message delay plus TO . This is impractical because in many systems the maximum message delay is orders of magnitude larger than the average message delay. The source of the above problems is that even though the heartbeats are sent at regular intervals, the timers to "catch" them expire at irregular times, namely the receipt times of the heartbeats plus a fixed TO . The algorithm that we propose eliminates this problem. As a result, the probability of a premature timeout on heartbeat m i does not depend on the behavior of the heartbeats that precede m i , and the detection time does not depend on the maximum message delay. A New Algorithm and its QoS Analysis. In the new algorithm, process p sends heartbeat messages periodically every j time units (just as in the simple algorithm). To determine whether to suspect p, q uses a sequence - called freshness points, obtained by shifting the sending time of the heartbeat messages by a fixed parameter ffi. More precisely, - where oe i is the time when m i is sent. For any time t, let i be so that t 2 [- trusts p at time t if and only if q has received heartbeat m i or higher. Given the probabilistic behavior of the system (i.e., the probability of message losses and the distribution of message delays), and the parameters j and ffi of the algorithm, we determine the QoS of the new algorithm using the theory of stochastic processes. Simulation results given in Section 7 are consistent with our QoS analysis, and they show that the new algorithm performs better than the common one. In contrast to the common algorithm, the new algorithm guarantees an upper bound on the detection time. Moreover, the new algorithm is optimal in the sense that it has the best possible query accuracy probability with respect to any given bound on the detection time. More precisely, we show that among all failure detectors that send heartbeats at the same rate (they use the same network bandwidth) and satisfy the same upper bound on the detection time, the new algorithm has the best query accuracy probability. The first version of our algorithm (described above) assumes that p and q have synchronized clocks. This assumption is not unrealistic, even in large networks. For example, GPS and Cesium clocks are becoming accessible, and they can provide clocks that are very closely synchronized (see, e.g., [29]). When synchronized clocks are not available, we propose a modification to this algorithm that performs equally well in practice, as shown by our simulations. The basic idea is to use past heartbeat messages to obtain accurate estimates of the expected arrival times of future heartbeats, and then use these estimates to find the freshness points. This is explained in Section 6. Configuring our Algorithm to Meet the Failure Detector Requirements of an Application. Given a set of failure detector QoS requirements (provided by an application), we show how to compute the parameters of our algorithm to achieve these requirements. We first do so assuming that one knows the probabilistic behavior of the system (i.e., the probability distributions of message delays and message losses). We then drop this assumption, and show how to configure the failure detector to meet the QoS requirements of an application even when the probabilistic behavior of the system is not known. 1.3 Related Work In [19], Gouda and McGuire measure the performance of some failure detector protocols under the assumption that the protocol stops as soon as some process is suspected to have crashed (even if this suspicion is a mistake). This class of failure detectors is less general than the one that we studied here: in our work, a failure detector can alternate between suspicion and trust many times. In [28], van Renesse et. al. propose a scalable gossip-style randomized failure detector protocol. They measure the accuracy of this protocol in terms of the probability of premature timeouts. 4 The probability of premature timeouts, however, is not an appropriate metric for the specification of failure detectors in general: it is implementation-specific and it cannot be used to compare failure detectors that use timeouts in different ways. This point is further explained at the end of Section 2.3. In [24], Raynal and Tronel present an algorithm that detects member failures in a group: if some process detects a failure in the group (perhaps a false detection), then all processes report a group failure and the protocol terminates. The algorithm uses heartbeat-style protocol, and its timeout mechanism is the same as the simple algorithm that we described in Section 1.2. 4 This is called "the probability of mistakes" in [28]. In [29], Ver-ssimo and Raynal study QoS failure detectors - these are detectors that indicate when a service does not meet its quality-of-service requirements. In contrast, this paper studies the QoS of failure detectors, i.e., how well a failure detector works. Heartbeat-style failure detectors are commonly used in practice. To keep both good detection time and good accuracy, many implementations rely on special features of the operating system and communication system to try to ensure that heartbeat messages are received at regular intervals (see discussion in Section 12.9 of [23]). This is not easy even for closely-connected computer clusters, and it is very hard in wide-area networks. The probabilistic network model used in this paper is similar to the ones used in [14, 6] for probabilistic clock synchronization. The method of estimating the expected arrival times of heartbeat messages is close to the method of remote clock reading of [6]. The rest of the paper is organized as follows. In Section 2, we propose a set of metrics to specify the QoS of failure detectors. In Section 3, we describe a new failure detector algorithm and analyze its QoS in terms of these metrics; we also present an optimality result. We then explain how to set the algorithm's parameters to meet some given QoS requirements - first in the case when we know the probabilistic behavior of messages (Section 4), and then in the case when this is not known (Sec- tion 5). In Section 6 we deal with unsynchronized clocks. We present the results of some simulations in Section 7, and we conclude the paper with some discussion in Section 8. Appendix A lists the main symbols used in the paper, and Appendices B to D give the proofs of the main theorems. More detailed proofs can be found in [13]. 2 On the QoS Specification of Failure Detectors We consider a system of two processes p and q. We assume that the failure detector at q monitors p, and that q does not crash. Henceforth, real time is continuous and ranges from 0 to 1. 2.1 The Failure Detector Model The output of the failure detector at q at time t is either S or T , which means that q suspects or trusts p at time t, respectively. A transition occurs when the output of the failure detector at q changes: An S-transition occurs when the output at q changes from T to S; a T-transition occurs when the output at q changes from S to T . We assume that there are only a finite number of transitions during any finite time interval. Since the behavior of the system is probabilistic, the precise definition of our model and of our QoS metrics uses the theory of stochastic processes. To keep our presentation at an intuitive level, we omit the technical details related to this theory (they can be found in [13]). We consider only failure detectors whose behavior eventually reaches steady state, as we now explain informally. When a failure detector starts running, and for a while after, its behavior depends on the trust suspect suspect up FD at q Figure 4: Mistake duration T M , good period duration T G , and mistake recurrence time T MR initial condition (such as whether initially q suspects p or not) and on how long it has been running. Typically, as time passes the effect of the initial condition gradually diminishes and its behavior no longer depends on how long it has been running - i.e., eventually the failure detector behavior reaches equilibrium, or steady state. In steady state, the probability law governing the behavior of the failure detector does not change over time. The QoS metrics that we propose refer to the behavior of a failure detector after it reaches steady state. Most of these metrics are random variables. 2.2 Primary Metrics We propose three primary metrics for the QoS specification of failure detectors. The first one measures the speed of a failure detector. It is defined with respect to the runs in which p crashes. Detection time (T D Informally, T D is the time that elapses from p's crash to the time when q starts suspecting p permanently. More precisely, T D is a random variable representing the time that elapses from the time that p crashes to the time when the final S-transition (of the failure detector at q) occurs and there are no transitions afterwards (Fig. 1). If there is no such final S-transition, then T such an S-transition occurs before p crashes, then T We next define some metrics that are used to specify the accuracy of a failure-detector. Throughout the paper, all accuracy metrics are defined with respect to failure-free runs, i.e., runs in which p does not crash. 6 There are two primary accuracy metrics: Mistake recurrence time (T MR this measures the time between two consecutive mistakes. More precisely, T MR is a random variable representing the time that elapses from an S-transition to the next one (Fig. 4). Mistake duration (T M this measures the time it takes the failure detector to correct a mistake. More precisely, T M is a random variable representing the time that elapses from an S-transition to the next T-transition (Fig. 4). As we discussed in the introduction, there are many aspects of failure detector accuracy that may be important to applications. Thus, in addition to T MR and T M , we propose four other accuracy metrics in the next section. We selected T MR and T M as the primary metrics because given these two, one can 5 We omit the boundary cases of other metrics since they can be similarly defined. 6 As explained in [13], these metrics are also meaningful for runs in which p crashes. compute the other four (this will be shown in Section 2.4). 2.3 Derived Metrics We propose four additional accuracy metrics: Average mistake rate (- M this measures the rate at which a failure detector make mistakes, i.e., it is the average number of S-transitions per time unit. This metric is important to long-lived applications where each failure detector mistake (each S-transition) results in a costly interrupt. This is the case for applications such as group membership and cluster management. Query accuracy probability (P A this is the probability that the failure detector's output is correct at a random time. This metric is important to applications that interact with the failure detector by querying it at random times. Many applications can make progress only during good periods - periods in which the failure detector makes no mistakes. This observation leads to the following two metrics. Good period duration (T G this measures the length of a good period. More precisely, T G is a random variable representing the time that elapses from a T-transition to the next S-transition (Fig. 4). For short-lived applications, however, a closely related metric may be more relevant. Suppose that an application is started at a random time in a good period. If the remaining part of the good period is long enough, the short-lived application will be able to complete its task. The metric that measures the remaining part of the good period is: Forward good period duration (T FG this is a random variable representing the time that elapses from a random time at which q trusts p, to the time of the next S-transition. At first sight, it may seem that, on the average, T FG is just half of T G (the length of a good period). But this is incorrect, and in Section 2.4 we give the actual relation between T FG and T G . An important remark is now in order. For timeout-based failure detectors, the probability of premature timeouts has sometimes been used as the accuracy measure: this is the probability that when the timer is set, it will prematurely timeout on a process that is actually up. The measure, however, is not appropriate because: (a) it is implementation-specific, and (b) it is not useful to applications unless it is given together with other implementation-specific measures, e.g., how often timers are started, whether the timers are started at regular or variable intervals, whether the timeout periods are fixed or variable, etc. (many such variations exist in practice [10, 19, 28]). Thus, the probability of premature timeouts is not a good metric for the specification of failure detectors, e.g., it cannot be used to compare the QoS of failure detectors that use timeouts in different ways. The six accuracy metrics that we identified in this paper do not refer to implementation-specific features, in particular, they do not refer to timeouts at all. 2.4 How the Accuracy Metrics are Related Theorem 1 below explains how our six accuracy metrics are related. We then use this theorem to justify our choice of the primary accuracy metrics. Henceforth, Pr(A) denotes the probability of event A; E(X), E(X k ), and V(X) denote the expected value (or mean), the k-th moment, and the variance of random variable X , respectively. Parts (2) and (3) of Theorem 1 assume that in failure-free runs, the probabilistic distribution of failure detector histories is ergodic. Roughly speaking, this means that in failure-free runs, the failure detector slowly "forgets" its past history: from any given time on, its future behavior may depend only on its recent behavior. We call failure detectors satisfying this ergodicity condition ergodic failure detectors. Ergodicity is a basic concept in the theory of stochastic processes [26], but the technical details are substantial and outside the scope of this paper. We have also determined the relations between our accuracy metrics in the case that ergodicity does not hold. The resulting expressions are more complex (they are generalized versions of those given below) and can be found in [13]. Theorem 1 For any ergodic failure detector, the following results hold: (1) T . (2) If is always 0. If R x In particular, (3c) E(T FG The fact that T holds is immediate by definition. The proofs of parts (2) and (3) use the theory of stochastic processes. Part (2) is intuitive, while part (3), which relates T G and T FG , is more complex. In particular, part (3c) is counter-intuitive: one may think that E(T FG (3c) says that E(T FG ) is in general larger than E(T G )=2 (this is a version of the "waiting time paradox" in the theory of stochastic processes [4]). We now explain how Theorem 1 guided our selection of the primary accuracy metrics. Parts (2) and (3) show that - M , P A and T FG can be derived from T MR This suggests that the primary metrics should be selected among T MR , T M and T G . Moreover, since T it is clear that given the joint distribution of any two of them, one can derive the remaining one. Thus, two of T MR T M and T G should be selected as the primary metrics, but which two? By choosing T MR and T M as our primary metrics, we get the following convenient property that helps to compare failure detectors: if FD 1 is better than FD 2 in terms of both E(T MR ) and E(T M ) (the expected values of the primary metrics) then we can be sure that FD 1 is also better than FD 2 in terms of E(T G ) (the expected values of the other metric). We would not get this useful property if T G were selected as one of the primary metrics. 7 7 For example, FD 1 may be better than FD 2 in terms of both E(T G ) and E(T M ), but worse than FD 2 in terms of E(T MR ). 3 The Design and QoS Analysis of a New Failure Detector Algorith 3.1 The Probabilistic Network Model We assume that processes p and q are connected by a link that does not create or duplicate messages, 8 but may delay or drop messages. Note that the link here represents an end-to-end connection and does not necessarily correspond to a physical link. We assume that the message loss and message delay behavior of any message sent through the link is probabilistic, and is characterized by the following two parameters: (a) message loss probability which is the probability that a message is dropped by the link; and (b) message delay D, which is a random variable with range (0; 1) representing the delay from the time a message is sent to the time it is received, under the condition that the message is not dropped by the link. We assume that the expected value E(D) and the variance V(D) of D are finite. Note that our model does not assume that the message delay time D follows any particular distribution, and thus it is applicable to many practical systems. have access to their own local clocks. For simplicity, we assume that there is no clock drift, i.e., local clocks run at the same speed as real time (our results can be easily generalized to the case where local clocks have bounded drifts). In Sections 3, 4 and 5, we further assume that clocks are synchronized. We explain how to remove this assumption in Section 6. For simplicity we assume that the probabilistic behavior of the network does not change over time. In Section 8, we explain how to modify the algorithm so that it dynamically adapts to changes in the probabilistic behavior of the system. We assume that crashes cannot be predicted, i.e., the state of the system at any given time has no information whatsoever on the occurrence of future crashes (this excludes a system with program-controlled crashes [11]). Moreover, the delay and loss behaviors of the messages that a process sends are independent of whether (and when) the process crashes. 3.2 The Algorithm The new algorithm works as follows. The monitored process p periodically sends heartbeat messages to q every j time units, where j is a parameter of the algorithm. Every heartbeat message m i is tagged with its sequence number i. Henceforth, oe i denotes the sending time of message . The monitoring process q shifts the oe i 's forward by ffi - the other parameter of the algorithm - to obtain the sequence of times uses the - i 's and the times it receives heartbeat messages, to determine whether to trust or suspect p, as follows. Consider 8 Message duplication can be easily taken care of: whenever we refer to a message being received, we change it to the first copy of the message being received. With this modification, all definitions and analyses in the paper go through, and in particular, our results remain correct without any change. (c) (b) (a) FD at q suspect trust trust suspect Figure 5: Three scenarios of the failure detector output in one interval [- Process p: 1 for all i - 1, at time oe Process q: 3 for all i - 1, at time - 4 if did not receive m j with j - i then output fsuspect p if no fresh message is receivedg 5 upon receive message m j at time t 2 [- ftrust p when some fresh message is receivedg Figure Failure detector algorithm NFD-S with parameters j and ffi (clocks are synchronized) time period [- checks whether it has received some message m j with j - i. If so, trusts p during the entire period [- starts suspecting p. If at some time before - i+1 , q receives some message m j with j - i then q starts trusting p from that time until - i+1 . (Fig. 5 (b)). If by time - i+1 , q has not received any message m j with j - i, then q suspects p during the entire period [- This procedure is repeated for every time period. The detailed algorithm with parameters j and ffi is denoted by NFD-S, and is given in Fig. 6. 9 Note that from time - i to - i+1 , only messages m j with j - i can affect the output of the failure detector. For this reason, - i is called a freshness point: from time - i to - i+1 are still fresh (useful). With this algorithm, q trusts p at time t if and only if q received a message that is still fresh at time t. 9 This version of the algorithm is convenient for illustrating the main idea and for performing the analysis. We have omitted some obvious optimizations. 3.3 The QoS Analysis of the Algorithm We now give the QoS of the algorithm (the analysis is given in Appendix B). We assume that the link from p to q satisfies the following message independence property: the behaviors of any two heartbeat messages sent by p are independent. 10 Henceforth, let - 0 (as in line 3 of the algorithm). We first formalize the intuition behind freshness points and fresh messages: Lemma 2 For all trusts p at time t if and only if q has received some message t. The following definitions are for runs where p does not crashes. (1) For any i - 1, let k be the smallest integer such that for all is sent at or after time (2) For any i - 1, let p j (x) be the probability that q does not receive message m i+j by time for every j - 0 and every x - 0; let (3) For any i - 2, let q 0 be the probability that q receives message m (4) For any i - 1, let u(x) be the probability that q suspects p at time (5) For any i - 2, let p S be the probability that an S-transition occurs at time - i . The above definitions are given in terms of i, a positive integer. Proposition 3, however, shows that they are actually independent of i. Proposition 3 (1) dffi=je. (2) For all By definition, if the probability that q receives m i by time - i is 1. Thus, Similarly, it is easy to see that if are degenerated cases of no interest. We henceforth assume that The following theorem summarizes our QoS analysis of the new failure detector algorithm. In practice, this holds only if consecutive heartbeats are sent more than some \Delta time units apart, where \Delta depends on the system. So assuming that the behavior of heartbeats are independent is equivalent to assuming that j ? \Delta. Theorem 4 Consider a system with synchronized clocks, where the probability of message losses is and the distribution of message delays is P (D - x). The failure detector NFD-S of Fig. 6 with parameters j and ffi has the following properties. (1) The detection time is bounded: (2) The average mistake recurrence time is: E(T MR (3) The average mistake duration is: From E(T MR ) and E(T M ) given in the theorem above, we can easily derive the other accuracy measures using Theorem 1. For example, we can get the query accuracy probability P Theorem 4 (1) shows an important property of the algorithm: the detection time is bounded, and the bound does not depend on the behavior of message delays and losses. In Sections 4, 5 and 6, we show how to use Theorem 4 to compute the failure detector parameters, so that the failure detector satisfies some QoS requirements (given by an application). 3.4 An Optimality Result Among all failure detectors that send heartbeats at the same rate and satisfy the same upper bound on the detection time, the new algorithm provides the best query accuracy probability. More precisely, let C be the class of failure detector algorithms A such that in every run of A, process p sends heartbeats to q every j time units and A satisfies T D - T U for some constant T U . Let A be the instance of the new failure detector algorithm NFD-S with parameters j and j. By part (1) of Theorem 4, we know that A 2 C. We can show that Theorem 5 For any A 2 C, let P A be the query accuracy probability of A. Let P A be the query accuracy probability of A . Then P A The theorem is a consequence of the following important property of algorithm A . Consider any algorithm A 2 C. Let r be any failure-free run of A , and r 0 be any failure-free run of A in which the heartbeat delays and losses are exactly as in r. We can show that if q suspects p at time t in r, then q also suspects p at time t in r 0 . With this property, it is easy to see that the probability that q trusts p at a random time in A must be at least as high as the probability that q trusts p at a random time in any A 2 C. The detailed proof is given in Appendix C. Failure Detector Configurator QoS Requirements U MR U Pr (D - x) Probabilistic Behavior of Heartbeats Figure 7: Meeting QoS requirements with NFD-S. The probabilistic behavior of heartbeats is given, and clocks are synchronized 4 Configuring the Failure Detector to Satisfy QoS Requirements Suppose we are given a set of failure detector QoS requirements (the QoS requirements could be given by the application that uses this failure detector). We now show how to compute the parameters j and ffi of our failure detector algorithm, so that these requirements are satisfied. We assume that (a) the local clocks of processes are synchronized, and (b) one knows the probabilistic behavior of the messages, i.e., the message loss probability p L and the distribution of message delays Pr(D - x). In Sections 5 and 6, we consider the cases when these assumptions do not hold. We assume that the QoS requirements are expressed using the primary metrics. More precisely, a set of QoS requirements is a tuple (T U MR of positive numbers, where T U is an upper bound on the detection time, T L MR is a lower bound on the average mistake recurrence time, and T U is an upper bound on the average mistake duration. In other words, the requirements are that: 11 MR Our goal, illustrated in Fig. 7, is to find a configuration procedure that takes as inputs (a) the QoS re- quirements, namely T U MR , and (b) the probabilistic behavior of the heartbeat messages, namely p L and Pr(D - x), and outputs the failure detector parameters j and ffi so that the failure detector satisfies the QoS requirements in (4.4). Furthermore, to minimize the network bandwidth taken by the failure detector, we want a configuration procedure that finds the largest intersending interval j that satisfy these QoS requirements. Using Theorem 4, our goal can be stated as a mathematical programming problem: 11 Note that the bounds on the primary metrics E(T MR ) and E(T M ) also impose bounds on the derived metrics, according to Theorem 1. More precisely, we have - M - 1=T L MR MR MR MR M , and E(T FG MR )=2. subject to MR where the values of u(x) and p S are given by Proposition 3. Solving this problem is hard, so instead we show how to find some j and ffi that satisfy (4.5)-(4.7) (but the j that we find may not be the largest possible). To do so, we replace (4.7) with a simpler and stronger constraint, and then compute the optimal solution of this modified problem (see Appendix D). We obtain the following procedure to find j and ffi: . If j "QoS cannot be achieved" and stop; else continue. ffl Step 2: Let Find the largest j - j max such that MR . Such an j always exists. To find such an j, we can use a simple numerical method, such as binary search (this works because when j decreases, increases exponentially fast). ffl Step 3: Set Theorem 6 Consider a system in which clocks are synchronized, and the probabilistic behavior of messages is known. Suppose we are given a set of QoS requirements as in (4.4). The above procedure has two possible outcomes: (1) It outputs j and ffi. In this case, with parameters j and ffi the failure detector NFD-S of Fig. 6 satisfies the given QoS requirements. (2) It outputs "QoS cannot be achieved". In this case, no failure detector can achieve the given QoS requirements. As an example of the configuration procedure of the failure detector, suppose we have the following QoS requirements: (a) a crash failure is detected within s; (b) on average, the failure detector makes at most one mistake per month, i.e., T L s; (c) on average, the failure detector corrects its mistakes within one minute, i.e. T U Assume that the message loss probability is p 0:01, the distribution of message delay D is exponential, and the average message delay E(D) is 0:02 s. By inputting these numbers into the configuration procedure, we get these parameters, our failure detector satisfies the given QoS requirements. Failure Detector Configurator Estimator of the Probabilistic Behavior of Heartbeats E(D) QoS Requirements U MR U Figure 8: Meeting QoS requirements with NFD-S. The probabilistic behavior of heartbeats is not known, and clocks are synchronized 5 Dealing with Unknown Message Behavior In Section 4, our procedure to compute the parameters j and ffi of NFD-S to meet some QoS requirements assumed that one knows the probability p L of message loss and the distribution Pr(D - x) of message delays. This assumption is not unrealistic, but in some systems the probabilistic behavior of messages may not be known. In that case, it is still possible to compute j and ffi, as we now explain. We proceed in two steps: (1) we first show how to compute j and ffi using only p L , E(D) and V(D) (recall that E(D) and V (D) are the expected value and variance of message delays, respectively); (2) we then show how to estimate p L , E(D) and V(D). In this section we still assume that local clocks are synchronized (we drop this assumption in the next section). See Fig. 8. Computing Failure Detector Parameters j and ffi Using p L , E(D) and V(D). With E(D) and V(D), we can bound Pr(D ? t) using the following One-Sided Inequality of probability theory (e.g., see [4], p.79): For any random variable D with a finite expected value and a finite variance, With this, we can derive the following bounds on the QoS metrics of algorithm NFD-S. Theorem 7 Consider a system with synchronized clocks and assume ffi ? E(D). For algorithm NFD-S, we have E(T MR Y and Note that in Theorem 7 we assume that ffi ? E(D), where ffi is a parameter of NFD-S. This assumption is reasonable because if ffi - E(D) then NFD-S would generate a false suspicion every time the heartbeat message is delayed by more than the average message delay. But then, NFD-S would make too many mistakes to be a useful failure detector. Theorem 7 can be used to compute the parameters j and ffi of the failure detector NFD-S, so that it satisfies the QoS requirements given in (4.4). Recall that these QoS requirements are given as a tuple MR is an upper bound on the worst-case detection time, T L MR is a lower bound on the average mistake recurrence time, and T U is an upper bound on the average mistake duration. The configuration procedure is given below. This procedure assumes that T U E(D), i.e., the required detection time is greater than the average message delay (a reasonable assumption). cannot be achieved" and stop; else continue. ffl Step 2: Let Y Find the largest j - j max such that MR . Such an j always exists. ffl Step 3: Set Notice that the above procedure does not use the distribution Pr(D - x) of message delays; it only uses p L , E(D) and V(D). Theorem 8 Consider a system in which clocks are synchronized, and the probabilistic behavior of messages is not known. Suppose we are given a set of QoS requirements as in (4.4), and suppose E(D). The above procedure has two possible outcomes: (1) It outputs j and ffi. In this case, with parameters j and ffi the failure detector NFD-S of Fig. 6 satisfies the given QoS requirements. (2) It outputs "QoS cannot be achieved". In this case, no failure detector can achieve the given QoS requirements. The above configuration procedure works when the distribution of the message delay D is not known (only E(D) and V (D) are known). To illustrate this procedure, we take the same example as in Section 4, except that we do not assume that the distribution of D is exponential. Specifically, suppose that the failure detector QoS requirements are that: (a) a crash failure is detected within onds, i.e., T U s; (b) on average, the failure detector makes at most one mistake per month, i.e., T L MR s; (c) on average, the failure detector corrects its mistakes within one minute, i.e. T U Assume that the message loss probability is p 0:01, the average message delay E(D) is 0:02 s, and the variance V (D) is also 0:02. By inputting these numbers into the configuration procedure, we get these parameters, failure detector NFD-S satisfies the given QoS requirements. Note that when we go from the case that the distribution of D is known (example of Section 4) to the case that D is not known, j decreases from 9:97 s to 9:71 s. This corresponds to a slight increase in the heartbeat sending rate (in order to achieve the same given QoS). Estimating p L , E(D) and V(D). It is easy to estimate p L , E(D) and V(D) using heartbeat messages. For example, to estimate p L , one can use the sequence numbers of the heartbeat messages to count the number of "missing" heartbeats, and then divide this count by the highest sequence number received so far. To estimate E(D) and V(D), we use the synchronized clocks as follows: When p sends a heartbeat with the sending time S, and when q receives m, q records the receipt time A. In this way, A \Gamma S is the delay of m. We then compute the average and variance of A \Gamma S for multiple past heartbeat messages, and thus obtain accurate estimates for E(D) and V(D). 6 Dealing with Unknown Message Behavior and Unsynchronized Clocks So far, we assumed that the clocks of p and q are synchronized. More precisely, in the algorithm NFD-S of Fig. 6, q sets the freshness points - i 's by shifting the sending times of heartbeats by a constant. When clocks are not synchronized, the local sending times of heartbeats at p cannot be used by q to set the - i 's, and thus q needs to do it in a different way. The basic idea is that q sets the - i 's by shifting the expected arrival times of the heartbeats, and q estimates the expected arrival times accurately (to compute these estimates, q does not need synchronized clocks). 6.1 NFD-U: an Algorithm that Uses Expected Arrival Times We now present NFD-U, a new failure detector algorithm for systems with unsynchronized clocks. The new algorithm is very similar to NFD-S; the only difference is that q now sets the - i 's by shifting the expected arrival times of the heartbeats, rather than the sending times of heartbeats. We assume that local clocks do not drift with respect to real time, i.e., they accurately measure time intervals (our algorithm and results can be easily generalized to the case where local clocks have small bounded drifts with respect to real time). Let oe i denote the sending time of m i with respect to q's local clock. Then, the expected arrival time of m i at q is EA E(D) is the expected message delay. Assume that q knows the EA i 's (we will soon show how q can accurately estimate them). To set the shifts the EA i 's forward by ff time units (i.e., - ff is a new failure detector parameter that replaces ffi. The intuition here is that EA i is the time when m i is expected to be received, and ff is a slack added to EA i to accommodate the possible extra delay or loss of m i . Figure 9 shows the whole algorithm, denoted by NFD-U. We restructured the algorithm a little, to Process p: fusing p's local clockg 1 for all i - 1, at time i \Delta j, send heartbeat m i to q; Process q: fusing q's local clockg keeps the largest sequence number in all messages q received so farg the current time: fif the current time reaches - '+1 , then none of the messages received is still freshg 6 output / fsuspect p since no message received is still fresh at this timeg 7 upon receive message m j at time t: freceived a message with a higher sequence numberg fset the next freshness point - '+1 using the expected arrival time of m '+1 g ftrust p since m ' is still fresh at time tg Figure 9: Failure detector algorithm NFD-U with parameters j and ff (clocks are not synchronized, but the EA i 's are known) show explicitly when q uses the EA i 's. Variable ' keeps the largest heartbeat sequence number received so far, and - '+1 refers to the "next" freshness point. Note that when q updates ', it also changes - '+1 . If the local clock of q ever reaches time - '+1 (an event which might never happen), then at this time none of the heartbeats received is still fresh, and so q starts suspecting p (lines 5-6). When q receives m j , it checks whether this is a new heartbeat (j ? ') and in this case, (1) q updates ', (2) q sets the next freshness point - '+1 to EA '+1 + ff, and (3) q trusts p if the current time is less than - '+1 (lines 9-11). Note that this algorithm is identical to NFD-S, except in the way in which q sets the - i 's. In particular, for any time t, let i be so that t 2 [- trusts p at time t if and only if q has received heartbeat m i or higher. 6.2 Analysis and Configuration of NFD-U NFD-U and NFD-S differ only in the way they set the - i 's: in NFD-S, - while in NFD-U, last equality holds because EA E(D)). Thus, the QoS analysis of NFD-U is obtained by simply replacing ffi with E(D) + ff in Proposition 3, Theorem 4 and Theorem 7. To configure the parameters j and ff of NFD-U to meet some QoS requirements, we use a method similar to the one in Section 5. We proceed in two steps: (1) we first show how to compute j and ff using only p L and V(D) (note that E(D) is not used); (2) we then show how to estimate p L and V(D). Fig. 10. Computing Failure Detector Parameters j and ff using p L and V(D). By replacing ffi with E(D)+ ff in Theorem 7, we obtain the following bounds on the accuracy metrics of NFD-U: Failure Detector NFD-U are known) Configurator Estimator of the Probabilistic Behavior of Heartbeats h a QoS Requirements MR U Figure 10: Meeting QoS requirements with NFD-U. The probabilistic behavior of heartbeats is not known; clocks are not synchronized, but they are drift-free Theorem 9 Consider a system with drift-free clocks and assume ff ? 0. For algorithm NFD-U, we have E(T MR Y Note that the bounds given in Theorem 9 uses only p L and V(D); on the other hand, E(D) is not used. Theorem 9 can be used to compute the parameters j and ff of the failure detector NFD-U, so that it satisfies some QoS requirements. We assume the QoS requirements are given as a tuple (T u MR of positive numbers. The requirements are that: MR Note that the upper bound on the detection time T D is not T u plus the unknown average message delay E(D). So, the actual upper bound T U on the detection time is T u E(D). In other words, the QoS requirement on detection time is not absolute as in (4.4), but relative to E(D). This is justified as follows. Note that when local clocks are not synchronized and only one-way messages are used, an absolute bound T U on detection time cannot be enforced by any nontrivial failure detector. Moreover, it is reasonable to specify an upper bound requirement relative to the average delay E(D) of a heartbeat. In fact, a failure detector that guarantees to detect crashes faster than E(D) makes too many mistakes to be useful. The following is the configuration procedure for algorithm NFD-U, modified from the one in Section 5. cannot be achieved" and stop; else continue. ffl Step 2: Let dT u Y Find the largest j - j max such that MR . Such an j always exists. ffl Step 3: Set Theorem 10 Consider a system with unsynchronized, drift-free clocks, where the probabilistic behavior of messages is not known. Suppose we are given a set of QoS requirements as in (6.11). The above procedure has two possible outcomes: (1) It outputs j and ff. In this case, with parameters j and ff the failure detector NFD-U of Fig. 9 satisfies the given QoS requirements. (2) It outputs "QoS cannot be achieved". In this case, no failure detector can achieve the given QoS requirements. Estimating p L and V(D). When local clocks are not synchronized, we can estimate p L and V(D) using the procedure of Section 5. To estimate p L , this procedure did not use clocks, and so it works just as before. For V(D), the procedure did use clocks, but it works even though the clocks are not synchronized. To see why, recall that the procedure estimates V(D) by computing the variance of of multiple heartbeat messages, where A is the time (with respect to q's local clock time) when q receives a message m, and S is the time (with respect to p's local clock time) when p sends m. When clocks are not synchronized, A \Gamma S is not the actual delay of m, but rather the delay of m plus a constant, namely, the skew between the clocks of p and q. Thus the variance of A \Gamma S is the same as the variance V(D) of message delays. 6.3 NFD-E: an Algorithm that Uses Estimates of Expected Arrival Times Failure detector NFD-U (Fig. assumes that q knows the exact value of all the EA i 's (the expected arrival time of messages). In practice, q may not know such values, and needs to estimate them. To do so, every time q executes line 10 of algorithm NFD-U in Fig. 9, q considers the n most recent heartbeat messages (for some n), denoted m 0 be the sequence numbers of such messages and A 0 n be their receipt times according to q's local clock. Then EA '+1 can be estimated by: Intuitively, this formula first "normalizes" each A 0 i by shifting it backward in time by js i . Then it computes the average of the normalized A 0 Finally, it shifts forward the computed average by ('+1)j. It is easy to see that this is a good estimate of EA '+1 . We denote by NFD-E the algorithm obtained from Fig. 9 by replacing EA '+1 with this estimate. Our simulations show that NFD-E and NFD-U are practically indistinguishable for values of n as low as 30. Thus, for large values of n, the configuration procedure for NFD-U can also be used to configure NFD-E. See Fig. 11. Failure Detector Configurator Estimator of the Probabilistic Behavior of Heartbeats h a QoS Requirements MR U Figure Meeting QoS requirements with NFD-E (same as with NFD-U, except that the expected arrival times EA i 's of heartbeats are estimated) 7 Simulation Results We simulate both the new failure detector algorithm that we developed and the simple algorithm commonly used in practice (as described in Section 1.2). In particular, (a) we simulate the algorithm NFD-S (the one with synchronized clocks), and show that the simulation results validate our QoS analysis of NFD-S in Section 3.3; (b) we simulate the algorithm NFD-E (the one without synchronized clocks that estimates the expected arrival times), and show that it provides essentially the same QoS as NFD-S; and (c) we simulate the simple algorithm and compare it to the new algorithms NFD-S and NFD-E, and show that the new algorithms provide a much better accuracy than the simple algorithm. The settings of the simulations are as follows. For the purpose of comparison, we normalize the intersending time j of heartbeat messages in both the new algorithm and the simple algorithm to 1. The message loss probability p L is set to 0:01. The message delay D follows the exponential distribution (i.e., Pr(D - We choose the exponential distribution because of the following two reasons: first, it has the characteristic that a large portion of messages have fairly short delays while a small portion of messages have large delays, which is also the characteristic of message delays in many practical systems [14]; second, it has a simple analytical representation which allows us to compare the simulation results with the analytical results given in Theorem 4. The average message delay E(D) is set to 0:02, which is a small value compared to the intersending time j. This corresponds to a system in which message delays are in the order of tens of milliseconds (typical for messages transmitted over the Internet), while heartbeat messages are sent every few seconds. Note that since D follows an exponential distribution, the standard deviation is the variance is To compare the accuracy of different algorithms, we first set their parameters so that: (a) they send messages at the same rate (recall that they satisfy the same bound T U on the detection 2.5 3 3.5 required bound T U D on the worst-case detection time average mistake recurrence time obtained from the simulations analytical Figure 12: The average mistake recurrence times obtained by: (a) simulating the new algorithms NFD-S and NFD-E (shown by + and \Theta), (b) simulating the simple algorithm (shown by -ffi- and -\Pi-), and (c) plotting the analytical formula for E(T MR ) of the new algorithm NFD-S (shown by -). time. We simulated runs for values of T U ranging from 1 to 3.5, and for each value of T U , we measured the accuracy of the failure detectors in terms of the average mistake recurrence time E(T MR ) and the average mistake duration E(T M ). For each value of T U , we plotted E(T MR ) by considering a run with 500 mistake recurrence intervals, and computing the average length of these intervals. We do not show the plots for E(T M ) because the E(T M ) of all the algorithms were similar and bounded above by approximately 7.1 Simulation Results of NFD-S and NFD-E To ensure that NFD-S meets the given upper bound T U on the detection time, we set ffi to T U prescribed by Theorem 4 (1)). In algorithm NFD-E, we choose to estimate the expected arrival time using the most recent messages. To ensure NFD-E meets the given upper bound T U , we set In Fig. 12, we show the simulation results for algorithms NFD-S and NFD-E, together with the analytical formula of E(T MR ) derived in Section 3.3. These results show that: (a) the accuracy of algorithms NFD-S and NFD-E are very similar, and (b) the simulation results of both algorithms match the analytical formula for E(T MR ). 7.2 Simulation Results of the Simple Algorithm The simple algorithm has no upper bounds on the detection time. However, such an upper bound can be guaranteed with a simple modification: the general idea is to discard heartbeats which have very large delays. More precisely, the modified algorithm has another parameter, the cutoff time c, such that all heartbeats delayed by more than c time units, called slow heartbeats, are discarded. 12 With this modification, the detection time T D is at most T D - c + TO . Given a bound T U on the detection time, there is a tradeoff in setting the cutoff time c and the timeout the larger the cutoff time c, the smaller the number of slow heartbeats being discarded, but the shorter the timeout value TO , and vice versa. In our simulations, we choose two cutoff times times the average message delay, respectively. The timeout TO is set to T U c. The algorithm with c = 0:16 is denoted by SFD-L, and the one with by SFD-S. The simulation results on the average mistake recurrence times of SFD-L and SFD-S (Fig. 12) show that the accuracy of the new algorithms (with or without synchronized clocks) is better - sometimes by an order of magnitude - than the accuracy of the simple algorithm. Intuitively, this is because the use of a cutoff time to bound the detection time in the simple algorithm is detrimental to its accuracy: if the simple algorithm uses a large cutoff time, then it must use a small timeout value, and this decreases the accuracy of the failure detector; if it uses a small cutoff time, then it discards more heartbeats, and this is equivalent to an increase in the message loss probability; this in turn also decreases the accuracy of the failure detector (a detailed explanation of the simulation results can be found in [13]). Concluding Remarks An Adaptive Failure Detector. In this paper, we assumed that the probabilistic behavior of heartbeat messages does not change. In some networks, this may not be the case. For instance, a corporate net-work may have one behavior during working hours (when the message traffic is high), and a completely different behavior during lunch time or at night (when the system is mostly idle): During peak hours, the heartbeat messages may have a higher loss rate, a higher expected delay, and a higher variance of delay, than during off-peak hours. Such networks require a failure detector that adapts to the changing conditions, i.e., it dynamically reconfigures itself to meet some given QoS requirements. It turns out that our failure detectors can be made adaptive, as we now explain. For the case when clocks are synchronized, we make NFD-S adaptive by periodically reexecuting the configuration outlined in Fig. 8. The basic idea is to periodically run the estimator, which uses the n most recent heartbeats to estimate the current values of p L ; E(D) and V (D). These estimates are then fed into the configurator to recompute the new failure detector parameters j and ffi. This assumes that the algorithm can detect slow messages; this is not easy when local clocks are not synchronized, but a fail-aware datagram service [18] may be used. Similarly, when clocks are not synchronized, we can make NFD-E adaptive by periodically reexecuting the configuration outlined in Fig. 11. The only difference here is that the estimator also outputs EA i - the estimated arrival time of the next heartbeat - which is input into the failure detector NFD-E. The above adaptive algorithms form the core of a failure detection service that is currently being implemented and evaluated [15]. This service is intended to be shared among many different concurrent applications, each with a different set of QoS requirements. The failure detector in this architecture dynamically adapts itself not only to changes in the network condition, but also to changes in the current set of QoS demands (as new applications are started and old ones terminate). Acknowledgments We would like to thank Carole Delporte-Gallet, Hugues Fauconnier and anonymous referees of the conference version of this paper for their useful comments which helped us improve the paper. --R Using the heartbeat failure detector for quiescent reliable communication and consensus in partitionable networks. Failure detection and consensus in the crash-recovery model On quiescent reliable communication. Transis: a communication sub-system for high availability Probabilistic clock synchronization in distributed systems. Relacs: a communications infrastructure for constructing reliable applications in large-scale distributed systems Probability and Measure. Renesse, editors. Reliable Distributed Computing with the Isis Toolkit. Requirements for Internet Hosts-Communication Layers On the impossibility of group membership. Unreliable failure detectors for reliable distributed systems. On the Quality of Service of Failure Detectors. Probabilistic clock synchronization. Failure detector service for dependable computing (fast abstract). Failure detectors in omission failure environ- ments Accelerated heartbeat protocols. Non blocking atomic commitment with an unreliable failure detector. The Ensemble System. A fault-tolerant multicast group communication system In Search of Clusters. Group membership failure detection: a simple protocol and its probabilistic analysis. Stochastic Processes. --TR Probability, statistics, and queueing theory with computer science applications Unreliable failure detectors for reliable distributed systems Totem Horus On the impossibility of group membership In search of clusters (2nd ed.) Using the heartbeat failure detector for quiescent reliable communication and consensus in partitionable networks On Quiescent Reliable Communication Reliable Distributed Computing with the ISIS Toolkit Probabilistic Clock Synchronization in Distributed Systems Time in Distributed System Models and Algorithms Non blocking atomic commitment with an unreliable failure detector Fail-aware failure detectors A Fail-Aware Membership Service Accelerated Heartbeat Protocols Failure Detectors in Omission Failure Environments The ensemble system On the quality of service of failure detectors --CTR Xuanhua Shi , Hai Jin , Weizhong Qiang, ALTER: first step towards dependable grids, Proceedings of the 2006 ACM symposium on Applied computing, April 23-27, 2006, Dijon, France Tiejun Ma , Jane Hillston , Stuart Anderson, Evaluation of the QoS of crash-recovery failure detection, Proceedings of the 2007 ACM symposium on Applied computing, March 11-15, 2007, Seoul, Korea Marcos K. Aguilera , Carole Delporte-Gallet , Hugues Fauconnier , Sam Toueg, On implementing omega with weak reliability and synchrony assumptions, Proceedings of the twenty-second annual symposium on Principles of distributed computing, p.306-314, July 13-16, 2003, Boston, Massachusetts Y. Horita , K. Taura , T. Chikayama, A Scalable and Efficient Self-Organizing Failure Detector for Grid Applications, Proceedings of the 6th IEEE/ACM International Workshop on Grid Computing, p.202-210, November 13-14, 2005

probabilistic analysis;failure detectors;quality of service;fault tolerance;distributed algorithm

507478

Visual Input for Pen-Based Computers.

The design and implementation of a camera-based, human-computer interface for acquisition of handwriting is presented. The camera focuses on a standard sheet of paper and images a common pen; the trajectory of the tip of the pen is tracked and the contact with the paper is detected. The recovered trajectory is shown to have sufficient spatio-temporal resolution and accuracy to enable handwritten character recognition. More than 100 subjects have used the system and have provided a large and heterogeneous set of examples showing that the system is both convenient and accurate.

writing, and adopt the machine's ones: typing, mouse-clicking, knob-turning. Learning to use a keyboard eectively requires time and patience. Ditto for menu-based mouse inter- faces. Current interfaces were designed for habitual computer users and for a limited range of tasks. If the \computer revolution" is to reach and benet the majority of the world population, more intuitive interfaces have to be designed. If machines are to become our helpers, rather than ever more complicated tools, they must be designed to understand us, rather than us having to learn how to use them. The third shortcoming is inadequacy. Our machines and the rest of our world are not well integrated because machines lack a sensory system. A machine does not know what is happening in its neighborhood, rather it sits and waits for a human to approach it and touch skillfully some of its hardware. Our desktop, our white-board, the visitor in our o-ce are completely unknown to our o-ce PC. There are many tasks that a machine will simply not do because its interfaces are inadequate. One avenue towards improving human-machine interfaces is to imitate nature and develop 'senses' for machines. Take vision: cameras may be miniaturized, thus allowing the development of small and cheap hardware; humans can easily read the body language, sketches, and handwriting produced by other humans { if a machine could do the same this would provide a natural, friendly, and very eective vision-based interface. This interface would allow capturing much information that current interfaces ignore. The computer industry recognized the advantages of using handwriting as the human-machine communication modality. Pen-based interfaces provide convenience, exibility, and small size. After the unsuccessful introduction of the visionary Apple Newton in the early '90s, a new generation of pen-based PDAs has established itself in the market. These PDAs (e.g. the popular PalmPilot) represent an interesting compromise. Their input device is the computer screen: the screen must be as large as possible for convenience of use, and as small as possible for portability. The optimal size, as identied by the market (approximately 12x8cm), makes PDAs acceptable but denitely not excellent on both counts. Handwriting may also be captured using a video camera and computer vision techniques, rather than the traditional tablets and touch-sensitive screens. This is an attractive alternative because cameras may be miniaturized thus making the interface much smaller. Furthermore, a vision-based system would allow the user to write at will on any convenient e.g., write on a piece of paper with a normal pen, on a blackboard, etc., regardless of size and location. In this article, we present the rst fully on-line, vision-based interface for conveniently and accurately capturing both handwriting and sketching. The interface is designed to be small and simple to use. It is built with a single consumer-electronics video camera and captures handwriting at high temporal (60Hz) and spatial (about 6000x2500 samples) resolution without using a special writing instrument. It allows the user to write at normal speed within a large writing area (more than half a letter-size page) with an output quality that is su-cient for recognition. The input interface consists of a camera, a normal piece of paper, and a normal pen. The camera focuses on the sheet of paper and images the pen tip; computer analysis of the resulting images enables the trajectory of the pen to be tracked and contact of the pen with the paper to be detected. The paper is organized as follows. In section 1.1 we summarize previous work in this area. This will allow us to motivate our approach and design, which are described in section 2. Section 3 presents a number of experiments that explore the performance of the system. A few concluding observations, as well as themes for future research, are collected in section 4. 1.1 Previous work The literature on handwriting recognition (see [10, 22, 25] for very comprehensive surveys) is divided into two main areas of research: o-line and on-line systems. O-line systems deal with a static image in which the system looks for the handwritten words before doing recognition. On-line systems obtain the position of the pen as a function of time directly from the interface. On-line systems have better information for doing recognition since they have timing information and since they avoid the initial search step of their o-line counterparts. The most popular input devices for handwriting are electronic tablets for on-line capturing and optical scanners for o-line conversion. We are of course interested in building on-line human machine interfaces. The integration of the electronic and physical aspects of an o-ce has been explored by two ambitious experimental systems. The Digital Desk [30, 31] developed at Rank Xerox EuroPARC merges physical objects (paper documents and pencils) with their electronic counterparts using computer vision and video projection. A computer screen is projected onto a physical desk using a video projector, while a camera is set up to watch the workspace such that the surface of the projected image and the surface of the image area coincide. A tablet digitizer or a nger tracked by the camera, like the system developed at INPG, Grenoble [4, 5], are used to input mouse-type of information into the system allowing one to select or highlight words on paper documents, cut and paste portions of text, draw gures, etc. The Liveboard [6, 19] developed by Xerox is similar in concept to the digital desk. This device is the replacement for the pads of ip-chart paper used in meetings. A computer screen is projected onto a white-board and a cord-less pen is used as input. The same image could be displayed onto boards placed at dierent locations and the input from each of the boards overlaid on all of them, allowing in this way for remote collaboration. The Digital Desk and the Liveboard are steps towards the integration of paper documents into the computing environment; these systems motivate the development of human-computer interfaces that can collect and interpret sketches and handwriting, and that do not require special hardware such as tablets and instrumented pens. A few vision-based interfaces [2, 14, 15, 18, 32] for handwriting are described in the literature. The MEMO-PEN [18] consists of a special pen that carries a small CCD camera close to its tip, a stress sensor, a micro computer, and a memory. The camera captures a series of snapshots of the writing, while the stress sensor detects the pressure applied on the ballpoint to have a record of the pen-up/-down strokes. The images captured by the camera only include a partial portion of the writing, so, the whole handwritten trace is recovered by overlaying successive snapshots. This system is quasi on-line since timing information is provided by the causality of image collection; however, the corresponding recognizer would need to look for the ink trace on the images before doing recognition. Also, the user is forced to write with a special purpose stylus rather than with a common pen. Alternative approaches [2, 32] consist of a video camera aimed to a user writing on a piece of paper. The camera provides a sequence of images at a frequency of 19 Hz. The last image of the sequence is thresholded in order to segment out the written text. The temporal order of the handwriting is reconstructed with a batch process by detecting the trace of ink produced between each two successive images. This detection may be obtained by performing image dierencing between successive images at the location of the segmented text. The user is required to write with a white pen under carefully controlled lighting conditions [2]. This system provides the location of the pen tip on each image, but it still requires batch processing after all text have been written. Besides, the ink trace detection method is prone to errors due to changes in lighting conditions and small movements of the writing surface. In contrast with the mentioned systems, our approach [14, 15] is fully on-line. It obtains data from a xed video camera and it allows the user maximum exibility in choosing virtually any pen and writing surface. We track the pen tip in real time in order to reconstruct its trajectory accurately and independently of changes in lighting. As we show in section 2.5 (see also g. 5) our interface increases the spatial resolution of the interface by a factor of ten (as compared with the batch ink-trace approach [2, 32]), and improves robustness with respect to lighting and small motions of the writing surface. The pen tip is tracked continuously both when the user is writing and when the pen is traveling on top of the paper. The detection of the strokes corresponding to the ink trace is the added burden that our system pays for all the described improvements. Vision System for Pen Tracking Our design of the interface is subject to the following constraints: all components (camera, frame grabber, computer) must be cheap and readily available; the user has to be able to write in a comfortable position using a normal pen; the interface has to be simple and intuitive so that user's training time and eort is minimal; the acquired handwritten trajectory has to have su-cient spatio-temporal information to enable recognition. The rst premise constrains the selection of the video camera to commercial consumer electronics devices. Typical low-cost cameras have spatial resolution of 480x640 pixels (rows x cols) at a frequency of Most cameras are interlaced, so each frame is composed of two half-frames with a maximum resolution of 240x640 pixels at a frequency of Given that the cut-o temporal frequency of handwriting is below 20 Hz [12, 24, 29], we are well above the Nyquist frequency of handwriting by working at making sure that no frequency component of handwriting is lost. The spatial resolution of the interface should be such that it enables clear legibility of the acquired handwriting. Figure 1(c) presents one example image provided by a camera located above the writing hand, as shown on g. 1(b). The resulting acquired trajectory is a signature shown magnied in g. 1(d). This sequence approximately occupies 20 image pixels per centimeter of writing; the spatial accuracy of the interface is 0.1 pixels; thus, the resolution of the system is about 200 samples per centimeter. This signature as well as the other trajectories presented in the gure are easily readable, showing that this ratio of image pixels per centimeter of writing provides su-cient information for a human to perform recognition. All handwriting examples shown in this paper follow a similar ratio of pixels per centimeter of writing. In order to satisfy the other premises, our interface does not require calibration and provides the user with the exibility of arranging the relative positions of the camera and the piece of paper. Only two conditions are imposed onto the user, one is that the camera should be located so that it has a clear sight of the pen tip and the other is that the writing implement should have su-cient contrast with the piece of paper. Figure 1 shows the block diagram of the system, the experimental setup, an example of an image provided by the camera, and three pen tip trajectories captured with the interface along with their corresponding pen-down strokes. These examples show an important dierence between our interface and conventional handwriting capture devices: we obtain a continuous trajectory by tracking the position of the pen tip in each of the images in the sequence; for some applications this trajectory must be segmented into strokes corresponding to ink trace (pen-down strokes) and strokes corresponding to movement above the paper (pen-up strokes). The method developed to detect pen-up/-down strokes as well as the design and prototyping of the interface are the main contributions of this paper. 2.1 Initialization and preprocessing The detection and localization of the position of the pen tip in the rst frame and the selection of the template to be used for detection in subsequent frames is the rst problem to solve. There are two possible scenarios: (a) the user writes with a pen that is familiar to the system or (b) an unknown pen is used. The familiar-pen case is easy to handle: the system may use a previously stored template representing the pen tip and detect its position in the image by correlation. There are a number of methods to initialize the system when the pen is unknown. Our initialization method is a semi-automatic one that requires a small amount of user coop- eration. It is based on a few reasonable assumptions: we assume that the user is writing with a dark-colored pen on a light-colored piece of paper; we assume that the pen tip is conical in shape; and we assume that the edges between the pen tip and the paper have larger contrast than the edge between the pen tip and the nger (see g. 2(i)). The rst assumption restricts the pen to be used with the system to have a well dened contrast with the paper. Hence, transparent pen or pens without contrast could not be used. The restriction is not severe since pens come in all sort of colors and it is quite simple to get one that satisfy the requirement. The second assumption is true for most commercial pens. The third assumption restrict the paper to be lighter than human skin. The requirement is easily satised writing on a piece of a common white paper. We display the image captured by the camera on the screen of the computer. A rectangular box is overlaid on this image as shown in gure 2(a). The user is required to place the pen tip inside the displayed box, ready to start writing. The system watches for activity within this box, which is measured by image dierencing between frames. After the pen tip enters the box, the systems waits until there is no more activity within the box, meaning (a) Pen Tip Detector Pen up / down Classifier Ballpoint Detector Camera Initialization & Preprocessing Filter/ (b) (c) 100 200 300 400 500 600100200 (d) 50 100 150 200 250102030(g) 50 100 150 200 250102030(h) 50 100 150 200 25010203040 50 100 150 200 25010203040 Figure 1: Overview of the System. (a) Block Diagram of the system. The camera feeds a sequence of images to the preprocessing stage (sec. 2.1). This block initializes the algorithm, i.e., it nds the initial position of the pen, and selects the template (rectangular subregion of the image) corresponding to the pen tip. In subsequent frames, the preprocessing stage has only the function of cutting a piece of image around the predicted position of the pen tip and feeding it into the next block. The pen tip detector (sec. 2.2) has the task of nding the position of the pen tip in each frame of the sequence. The lter (sec. 2.3) is a recursive estimator that predicts the position of the tip in the next frame based on an estimate of the current position, velocity, and acceleration of the pen. The lter also estimates the most likely position of the pen tip for missing frames. The ballpoint detector (sec. 2.5) nds the position of the very end of the pen tip, i.e., the place where the pen is in contact with the paper when the user is writing. Finally, the last block of our system checks for the presence of ink on the paper at the positions where the ballpoint of the pen was detected (sec. 2.6). (b) Experimental setup. The system does not require any calibration. The user has the exibility of arranging the relative positions of the camera and the piece of paper in order to write comfortably as long as the system has a clear sight of the pen tip. (c) Image provided by the camera. The user has a writing area larger than half a letter-size page. This image is the last frame corresponding to the trajectory shown in (d). The pen tip is tracked continuously both when the user is writing (pen-down strokes) and when the pen is moving on top of the paper (pen-up strokes). The complete tracked trajectory is shown in (d). Pen-down strokes corresponding to trajectory (d). (f),(h) Two more examples of handwritten sequences acquired with the interface. (h),(i) Corresponding pen-down strokes. that the user has taken a comfortable position to start writing. When the activity within the box has returned to low for a period of time (bigger than 200 ms), the system acquires the pen tip template, sends an audible signal to the user, and starts tracking. Figure 2(i) shows the pen tip, whose conical shape projects onto the image plane as a triangle. One of the borders of this triangle corresponds to the edge between the pen tip and the user's nger and the two other boundaries correspond to the edges between the pen tip and the piece of paper. Detection and extraction of the pen tip template is reduced to nding the boundary points of the pen tip, computing the corresponding centroid, and selecting a portion of the image around the centroid. The edges between the pen tip and the paper have bigger contrast than the edge between the pen tip and the nger, thus, we only look for these two boundaries in the detection and extraction of the template. The boundaries of the pen tip are located using Canny's edge detector [3] as shown in gure 2(c). Since detection and extraction of the pen tip from a single frame is not very reliable due to changes in illumination, the system collects information about the pen tip for a few frames before extracting the template. The algorithm is summarized in gure 2. The selection of the pen tip template is performed only at the beginning of the acquisition. The function of the initialization and preprocessing module in subsequent frames is only to extract a region of interest centered around the predicted position of the pen tip. The region of interest is used by the following block of the system to detect the actual position of the centroid of the pen tip in the current image. 2.2 Pen Tip Detection The second module of the system has the task of detecting the position of the pen tip in the current frame of the sequence. The solution of this task is well known in the optimal signal detection literature [8, 26]. Assuming that the signal to be detected is known exactly except for additive white noise, the optimal detector is a matched lter, i.e., a linear lter that looks (a) (b) (c) (d) (e) 1st quadrant 2nd quadrant 3rd quadrant estimated edge most voted quad. Ballpoint Finger 2nd most voted quad. most voted quad. Ballpoint - Finger Figure 2: Tracking Initialization. (a) Image provided to the user; the white rectangle is the initialization box. (b) The user has to place the pen tip inside the box so that the system can acquire the tracking template. Image dierencing is used to detect when the pen tip gets inside the box. The gure shows the result of image dierencing when the pen enters the tip acquisition area. (c) The boundaries of the pen tip are extracted using Canny's edge detector inside the initialization box. Only pixels with high contrast are selected. The dots displays the boundary pixels and the cross indicates their centroid. Sub-pixel resolution in the location of edge elements is achieved by tting a parabolic cylinder to the contrast surface in the neighborhood of each pixel. (d) Orientation of the boundary edge elements obtained with Canny's detector. (e) The dierent boundaries of the pen tip are obtained by clustering the orientation of the edge elements into the four quadrants and interpolating lines through the corresponding clustered pixels. (f) In the case in which only one of the boundaries is reliably detected, the other pen tip boundary is obtained by searching the image brightness prole along lines perpendicular to the detected boundary. Points of maximum contrast on these proles dene the missing boundary. The detection of the boundaries of the pen tip is performed on a sequence of frames in order to increase the robustness of the template extraction. The nal centroid position is obtained as the mean of the location of the centroid in each individual frame. (g) The triangular model of the pen tip is completely specied with the location of the centroid of the tip, the orientation of the axis of the tip, and the positions of the nger and of the ballpoint. The pen tip axis is dened as the line passing through the centroid of the boundary pixels, whose orientation is the mean of the orientation of the boundary lines. (h) Image brightness prole across the estimated pen tip axis. The positions of the ballpoint and of the nger are extracted by performing a 1D edge detection on the prole. Subpixel accuracy is obtained by tting a parabola to the edge detection result. (i) Final template of the pen tip automatically extracted by the interface. Predicted position of the pen tip Predicted position of the pen tip The most likely position of the pen tip is given by the location of maximum correlation Location of maximum correlation Region of interest position of the pen tip centered on the predicted Pen tip template Correlation Figure 3: Pen Tip Detector. The detection of the pen tip is obtained in our system by locating the maximum of the normalized correlation between the pen tip template and a subimage centered on the predicted position of the pen tip. The system analyzes the values of maximum normalized correlation to detect whether the pen tip is within the predicted region of interest. If the value of maximum correlation is lower than a threshold, the system emits an audible signal and continues to look for the pen tip in the same place, waiting for the user to realize that tracking has been lost and that the pen tip must be returned to the region of interest. The system waits for a few frames; if the pen tip does not return to sight, then tracking stops. like the signal to be detected. In our case, the signal consists of the pixels that represent the pen tip and the noise has two components: one component is due to noise in the acquisition of the images; the other one is due to shadows, due to pen markings on the paper, and due to changes in the apparent size and orientation of the pen tip during the sequence of images. The acquisition noise is the result of a combination of many factors like changes in illumination due to light ickering or automatic gain of the camera, quantization noise, changes in gain of the frame grabber, etc., where not all these factors are additive. Changes in the apparent size and orientation of the pen while the user is writing signicantly distorts the pen tip image, as shown in gure 3. Clearly neither component of the noise strictly satises the additive white noise assumptions of the matched lter; however, as a rst approximation, we will assume that the pen tip can be detected in each frame using the matched lter. In our system, the nal localization of the pen tip is performed by tting a triangle to the image of the tip as described in section 2.5. 2.3 Filtering The lter predicts the most likely position of the pen tip on the following frame based on the current predicted position, velocity, and acceleration of the pen tip, and on the location of the pen tip given by the pen tip detector. The prediction provided by the lter allows the interface to reduce the search area, saving computations while still keeping a good pen tip detection accuracy. The measurements are acquired faster and the measured trajectory is smoothed by the noise rejection of the lter. A Kalman Filter [1, 9, 11] is a recursive estimation scheme that is suitable for this problem. We tested several dierent rst- and second-order models for the movement of the pen tip on the image plane. The model that provided the best performance with the easiest tuning was a simple random walk model for the acceleration of the pen tip on the image plane. The model is given by equation 1:> > > > > > > > < (1) where x(k), v(k), and a(k) are the two-dimensional components of the position, velocity, and acceleration of the pen tip, and n a are additive zero-mean, Gaussian, white noise processes. The output of the model y(k) is the position of the pen tip corrupted by additive noise. The lter parameters used in the real-time implementation of the system are listed on table 1. 2.4 Missing frames The algorithm described in section 2.2 detects the position of the pen tip in each frame of the sequence. Unfortunately, some intermediate frames could be missing due to problems in image acquisition, or, in the case of the real-time implementation, due to synchronization problems between the host computer and the frame grabber. It is desirable to sample the handwritten trajectory at a constant rate; hence, there is a need for estimating the most likely position of the pen tip for the missing frames. The Kalman smoother [1, 9] is the scheme used in our system to solve this estimation problem (for more information see [14]). 2.5 Ballpoint detection The pen tip detector nds the most likely position of the centroid of the pen tip, a point that will be close to the the center of gravity of the triangular model of the pen tip (see 2.1). The position of the ballpoint 1 is obtained using an algorithm similar to the one used in the initialization; the major dierence is that the pen is now in movement, so we need to compute one ballpoint position for each frame. Using Canny's edge detector, we nd the position and orientation of the boundary edges of the pen tip. The edge detector is only applied to small windows in order to save computations and to speed up the processing of the current frame. We calculate the expected position of the boundaries using the orientations of the boundaries in the previous frame, the distance from the ballpoint and the nger to the centroid of the pen tip, as well as the current detected position of the centroid. A few points on these boundaries (in the case of the real-time system, we use ve points) are chosen as the centers of the edge detection windows; we look for points in each window that have maximum contrast; the edges are found by interpolating lines through these points; the axis of the pen tip is computed as the mean line 1 The term ballpoint is loosely used to indicate the actual ballpoint of pens and the pencil lead of pencils. Ballpoint - Finger (b) Ballpoint - Finger (c) (d) Figure 4: Fine localization of the ballpoint. (a) Image of the pen tip displaying the elements used to detect the ballpoint. The cross '+' in the center of the image shows the centroid of the pen tip provided by the pen tip detector. The points marked with a star '*' show the places where the boundaries of the pen were found using edge detection. The lines on the sides of the pen tip are the boundary edges and the line in the middle is the pen tip axis. The other two crosses '+' show the estimated positions of the ballpoint and of the nger. (b) Brightness prole along the axis of the pen tip. The positions of the ballpoint and of the nger are obtained by performing a 1D edge detection on the prole. This 1D edge detection is computed by correlating the prole with a derivative of a Gaussian function. The spatial resolution of the interface is dened by the accuracy on the localization of the ballpoint. The desired locations are extracted with subpixel resolution by tting a parabola to the correlation peaks. (c) Result of correlating the image prole with a derivative of a Gaussian function. (d) Blow-up of the region between the dotted vertical lines in (c). The parabolic t of the peak identies the position of the ballpoint. The vertex of the parabola plotted with a cross 'x' corresponds to the estimated sub-pixel position of the ballpoint. dened by the pen boundary edges; the image brightness prole through the axis of the tip is extracted in order to nd the positions of the ballpoint and of the nger (see gure 4). 2.6 Pen up detection The trajectories obtained by tracking the ballpoint are not suitable for performing handwriting recognition using standard techniques; most of the recognition systems to date assume that their input is only formed by pen-down strokes, i.e., portions of the trajectory where the pen was in contact with the paper. Our interface has only one camera from which we cannot detect the 3D position of the ballpoint; therefore, contact has to be inferred indirectly. A stereo system would solve this problem at a cost in additional hardware, calibration, and visibility of the pen tip. The detection of the times when the pen is lifted and therefore, not writing, is accomplished in our system by using the additional information provided by the ink path on the paper. Given a particular position of the ballpoint, the system checks whether there is an (c) 1 2060100140180211 280 300 320 34095105(d) Figure 5: Di-culties in detecting the ink trace. (a) The plot shows one sequence acquired with the interface. The dots indicate the ballpoint position over time. (b) The image displays a portion of the last frame of the sequence (we can see part of the pen tip on the right side of the image) showing the corresponding ink trace deposited on the paper. (c) Recovered ballpoint trajectory overlaid on the image of the ink trace. The sample points land over the ink trace most of the time with the exception of points at the beginning of the sequence (shown on the left side of the image). This happens because there might have been a displacement of the paper generated by one of the strokes (probably the long horizontal stroke between samples 20 and 40). (d) Each column of the picture shows the brightness prole of the image along lines that pass through each sample point and are perpendicular to the direction of motion. Brightness is measured at the position of the ballpoint and on ve pixels on each side of the ballpoint along the mentioned perpendicular. We note that the ink trace is not always found at the ballpoint position (row 6 of the plot). We can see the ink trace being a few pixels o the ballpoint pixel at the beginning of the sequence (samples 1-20), then stabilizing on the ballpoint (samples 20-35) until the pen tip appears on the prole (samples and later disappearing because of a pen up stroke (samples 40-55). From this example, we observe that we cannot rely on the ink trace captured in last image of the sequence, but we should rather detect the presence of ink as the ballpoint trajectory is being acquired (see sections 2.6.1-2.6.4). ink trace on the paper at this place or not. The image brightness at any given place varies with illumination, writer's hand position, and camera gain; moreover, the image contrast could change from frame to frame due to light ickering and shadows. Hence, the detection of the ink trace on the paper using image brightness is quite di-cult, as illustrated by the example of gure 5. We can get several observations from the simple example of gure 5. The ink trace is narrow (1-2 pixels), so even a small error in locating the ballpoint could lead to a mismatch between the ballpoint and the ink trace. The handwritten strokes are quite distorted due to confidence measure U D Hidden Markov Model of being in state pen up (U) or pen down (D) that estimates the likelihood Ink absence Segmentation Classification Trajectory Trajectory Image state sequence Most likely Detection Figure Pen up/down classication. Block diagram of the pen-up/-down classication subsystem. We detect when the pen is up or down using a bottom-up approach. The brightness of each point in the trajectory is compared with the brightness of the surrounding pixels. This comparison provides a measure of the condence of ink absence. A Hidden Markov Model is used to model the transition of the condence measure between the states of pen up and pen down. Using the local condence measure and the estimated HMM state sequence, the system classies each point of the trajectory as pen up or pen down. The measure of ink absence is di-cult to obtain and prone to errors, so it is better to divide the trajectory of the ballpoint into strokes and aggregate the point-wise classication into a stroke-wise classication. the pixelization of the image, e.g., diagonal straight strokes present a staircase pattern. The value of brightness corresponding to the ink trace varies within the same image (and even more across images), so we need to detect the ink trace in a local and robust way. The local ink measurement should be performed as soon as possible since the paper might move in the course of the session. Working in this way, there would be a good t of the sample points on top of the ink trace and the system would provide pen-up/-down information as the writing is produced, in an on-line fashion. However, the measurement has to be done after the pen tip moves away; otherwise, the pen tip will obstruct the paper and the ink trace. Figure 6 shows a block diagram of the pen-up/-down detection subsystem. The following sections describe in more detail each of the blocks presented in the gure. (a) x y brightness samples Ink absence confidence value:0.075 Figure 7: Local ink absence detection. (a) Typical pen-down sample point. The center cross corresponds to the estimated position of the ballpoint of the pen. The detection of ink is performed locally by comparing the brightness at the ballpoint with the brightness of the pixels located on a circle centered at the ballpoint position. The brightness at each point is obtained by interpolation [13]. (b) Histogram of the brightness values measured on the circle and corresponding Gaussian p.d.f. estimated from these values. The vertical line shows the value of brightness corresponding to the ballpoint's position. The ink absence condence measure corresponds to the area below the Gaussian p.d.f. between 1 and the ballpoint's brightness. This condence measure is equal to 0.075 for this example indicating that ink is likely to be present. 2.6.1 Local ink detection The detection of the ink trace is performed locally for each point of the trajectory by comparing the brightness at the ballpoint with the brightness of surrounding pixels (see gure 7). A condence measure of ink absence is obtained in a probabilistic setting. The brightness of ink-less pixels is assumed to be a Gaussian-distributed random variable. The parameters of the corresponding Gaussian p.d.f. are estimated locally using the brightness of points located on a circle centered at the ballpoint. We assume that all these points correspond to ink-less pixels. The ink absence condence measure is computed as the probability of the brightness at the ballpoint pixel given that this pixel is ink-less. If there is ink present at the ballpoint pixel, this measure is low, close to zero; otherwise, the measure is high, close to one. The selection of this particular condence measure is very convenient since it provides automatic scaling between zero and one. The measurements of brightness cannot be obtained until the pen tip has left the measurement area; otherwise, the ink trace will be covered by the pen tip, by the hand of the user, or by both. The system assumes a simple cone-shaped model for the area of the image covered by the pen and the hand of the user. The ballpoint is located at the vertex of the cone, the axis of the pen tip denes the axis of the cone, the position of the nger has to be inside the cone, and the aperture of the cone is chosen to be 90 degrees. This simple model allows the system to determine if the user is left handed or right handed and whether a particular ballpoint position is within the cone. The system waits until the cone is su-ciently far away from the area of interest before doing any brightness measurements. Left-handed users are challenging since they usually cover with their hands the most recently written pen strokes; hence, the system has to wait much longer than for right-handed users in order to perform ink detection. We have acquired data from left-handed users for the experiment of section 3.3, but we haven't compared the accuracy of pen-up/-down detection for left- and right-handed users. 2.6.2 Local pen-up/-down modeling The ink absence condence measure described in the previous section could be used to decide whether a particular sample point corresponds to pen up or pen down. However, making hard decisions based on a single measurement is likely to fail due to noise and errors in brightness measurements. A soft-decision approach that estimates the probability of each individual point being a pen up or a pen down is more robust. A further improvement is provided by modeling the probability of transition between these two states (pen up or pen down), given the current measurement and the previous state. A Hidden Markov Model (HMM) with two states, one corresponding to pen up and the other corresponding to pen down, is a suitable scheme to estimate these probabilities. The HMM learns the probabilities of moving from one state to the other and the probabilities of rendering a particular value of condence measure from a set of examples, in an unsupervised fashion. The HMM used in our system has the topology presented in gure 8. We use the forward-backward algorithm [23] to train Confidence Measure Probability Confidence Measure Ink absence Confidence Measure up Figure 8: Local model of pen up/down. Hidden Markov Model that models the transitions between pen-up and pen-down states. The observation of the model is the ink absence condence measure, an intrinsically continuous variable as it was dened in section 2.6.1. The HMM output is a set of discrete symbols, so we need to quantize the value of the condence measure in order to dene the output symbols. The condence measure is a probability, so it is scaled between zero and one. The interval [0; 1] is divided into sixteen equal intervals to quantize each condence measure value and to translate it into observation symbols. The resulting HMM after training is shown in the gure. The bar plots displays the output probability distributions of each state, that are learned purely from the examples. the HMM using a training set of handwritten sequences collected with the system. The training set consists of examples of cursive handwriting, block letters, numbers, drawings, signatures, and mathematical formulas in order to sample the variability of the pen up/down transition for dierent types of writing. The most likely state of the system at each point in the handwritten trajectory is estimated using Viterbi's algorithm [7, 23]. 2.6.3 Trajectory segmentation The previous two sections describe local measures used to classify each sample of the hand-written trajectory as either pen up or pen down. The measurement of ink absence is subject to errors, so the performance may be improved by dividing the handwritten trajectory into dierent strokes and by aggregating the sample-wise classication into a stroke-wise classi- cation. The handwritten trajectory is segmented into strokes using two features, the curvilinear velocity of the pen tip and the curvature of the trajectory. Selection of these two features Figure 9: Trajectory segmentation. Several examples of trajectories acquired with the interface and the corresponding strokes obtained after segmentation. Successive strokes are indicated alternately with solid and dashed lines. The threshold in curvilinear velocity was chosen so that points that remain within the same pixel in two consecutive frames are discarded (see table 1). was inspired by the work of Viviani [27, 28] and Plamondon [20, 21], and also by the intuitive idea that on the limit points between two dierent handwriting strokes the velocity of the pen is very small, the curvature of the trajectory is very high, or both. The set of segmentation points is the result of applying a threshold on each of the mentioned features. These thresholds were obtained experimentally and their values are presented on table 1. Figure 9 shows several examples of trajectories and the corresponding segmented strokes. 2.6.4 Stroke classication Having divided the trajectory into strokes, we proceed to classify the strokes as either pen-up or pen-down. We experimented with two approaches, one based on the ink absence condence measure and the other using the state sequence provided by the HMM. In the rst approach, the mean of the ink absence condence measures for all points in the stroke was used as the stroke condence measure. In the second approach, a voting scheme was used to assess the likelihood of a particular stroke being a pen-up or pen-down, this likelihood provided the stroke condence measure. If needed, hard classication of each stroke as pen up or pen down can be obtained by applying a threshold on the stroke condence results. The hard classication, as well as the likelihood of pen up/down, are the stroke descriptors that our interface provides to a handwriting recognition system. 2.7 Stopping acquisition We have mentioned that the system automatically stops if the value of maximum correlation is very low since this would imply that the pen tip has moved outside the search window (or that there was such a change in illumination that the pen tip no longer matches the template). The user can exploit this behavior to stop the acquisition by taking the pen tip away from the search window. There is another stopping possibility oered to the user. The system checks whether the pen tip has moved at all between consecutive frames and counts the number of consecutive frames in which there is no movement; if this number reaches a predened threshold, the system stops the acquisition. Thus, if the user wants to nish the acquisition at the end of a desired piece of handwriting, he can hold the pen tip still and the system will stop the acquisition. 2.8 Real-time implementation The interface was implemented using a video camera, a frame grabber, and a Pentium II 230MHz PC. The camera was a commercial Flexcam ID, manufactured by Videolabs, equipped with manual gain control. It has a resolution of 480x640 pixels per interlaced image at 30Hz. The frame grabber was a PXC200 manufactured by Imagination. Figure 10 shows the graphical user interface (GUI) of the windows-based application that runs our system. Figure 10: Real-time application. This image shows the GUI (Graphical User Interface) of the windows-based application that implements our system. The biggest window is a Dialog Box that allows the user to input parameters and run commands. The top-left window displays the image captured by the camera in order to provide visual feedback to the user. The bottom-left window shows the acquired trajectory after having done point-wise pen up/down classication with a hard threshold. 3 Experimental Results 3.1 System specications Temporal and spatial acquisition resolutions are key parameters that dene the performance of the interface. The maximum working frequency provided by the camera is the temporal resolution of the system is at most 16.67ms. The system is able to work at maximum frame rate since the total processing time per frame is 14ms. However, some frames are missed due to a lack of synchronization between the CPU and the frame grabber. A component of the system (see 2.4) estimates the most likely state of the system in the case of missing frames. This scheme is useful if the number of missing frames is small, otherwise, the system would drift according to the dynamics of the model of equation 1. We have used the system for acquiring hundreds of handwritten sequences in real time, experiencing a missing frame rate of at most 1 out of every 200 frames. We have shown in references [14, 16, 17] the performance of a signature verication system in which signatures are captured in real-time with our interface. Signatures are written at higher speeds than normal handwriting and therefore, a bigger image neighborhood has to be searched in order to nd the pen tip. We acquired signature sequences by enlarging the search area and turning the pen-up detection block of the system. In these experiments, we experienced a missing frame rate of at most 1 out of every 400 frames. We observe that the system occasionally loses track of the pen tip when the subject produces an extremely fast stroke. This problem of losing track of the pen tip could be solved in the future by using a more powerful CPU or dedicated hardware (that is able to process a larger search area). Nevertheless, after a few trials, the user learns how to utilize the system without exceeding its limits. The spatial resolution of the system was estimated in static and dynamic conditions. We acquired a few sequences in which the pen tip was held xed at the same location, so any dierences in the acquired points were due to noise on the image acquisition and errors in pen tip localization. We repeated this experiment 10 times, placing the pen at dierent positions and using dierent illumination. The static resolution of the system was estimated by computing the average standard deviation of the points acquired in each of the sequences. We also acquired 10 sequences of a subject drawing lines of dierent orientations with the help of a ruler. The lines were carefully drawn to be straight, so any dierences from a straight line would be due to noise in the image acquisition and errors in ballpoint localization. We t a line through the acquired points and computed the distance between the points and the tting line. The dynamic resolution of the system was estimated by computing the average standard deviation of the mentioned distance in each of the sequences. Figure 11 shows two sequences used to compute the spatial resolution and summarizes the resolution of the system. We note that the vertical resolution is almost the same for the two experiments, but the horizontal resolution varies by a factor of two from one experiment to the other. This dierence is possibly due to the subject holding the pen mostly in a vertical writing position for the static resolution experiment. In any case, we observe that the system has quite a Sequence used for static resolution computation 100 150 200 2504080120Sequence used for dynamic resolution computation horiz. resolution (pixels) vert. resolution (pixels) static dynamic Figure 11: Spatial resolution of the interface. The rst gure corresponds to points acquired while the pen tip was kept still at a xed position. This sequence is used to estimate the static resolution of the system. The second gure shows a straight line drawn with a ruler that is used to estimate the dynamic resolution of the system. The standard deviations of the error from the ideal position are given in the table as the estimated static and dynamic resolution of the system. One could take two standard deviations (roughly 0.1 pixel) to obtain a more conservative value of the spatial resolution. good resolution of less than one tenth of a pixel. Table 1 summarizes all the parameters used in the implementation of the real-time system. Figure 12 shows several examples of complete handwritten sequences acquired in real time with our system. A few portions of one of the sequences are blown-up in order to depict the level of acquisition noise. 3.2 Pen up detection experiments Only the pen tracking and the local ink detection components of the system have been implemented in the real-time application. In order to evaluate the performance of the complete pen-up detection subsystem, we collected 20 sequences comprising various types of handwriting (cursive, block letters, printed letters, numbers, drawings, signatures, and mathematical formulas). We used half of these sequences for training the HMM and the other half for test- ing. We obtained ground truth by classifying by hand each of the points of the test sequences as a pen up or pen down. We also classied by hand each of the strokes in which the test Parameter Value Pen tip template size 25x25 pixels Initial dead time (given to the user to move the paper 2 sec. (120 frames) to nd a clean area where to write) Image dierence threshold 15 (3 bits of noise) Number of pixels required to detect movement 20 pixels Number of pixels required to detect lack of movement Time of no pen tip movement waited before acquiring 200 ms the pen tip template Time used to acquire information on the pen tip 1 sec. (60 frames) Edge detector scale 3 pixels Contrast threshold (used with Canny's edge detector) 0.7 Distance from parabolic cylinder axis to center of pixel 0.5 pixels threshold (used with Canny's edge detector) Correlation window size 15x15 pixels KF output noise covariance matrix (R) diag(10 KF state noise covariance matrix (Q) diag(0,0,0,0,10 4 KF initial estimation error covariance matrix (P 0 Maximum normalized correlation value considered as a match 0.75 Maximum velocity denoting pen not moving 0.5 pixels per frame Time waited before stopping 0.5 sec (30 frames) Minimum number of points in a sequence 150 samples Minimum velocity threshold used for trajectory segmentation 0.75 pixels per frame Maximum curvature threshold used for trajectory segmentation 0.05 pixels per frame 2 Table 1: System parameters. System parameters used in the real-time implementation. sequences were divided by the segmentation algorithm. Two types of error measurements were used to evaluate the performance of pen down detection: the false acceptance rate which measured the percentage of pen-up points (segments) that were classied as pen down by the system; and the false rejection rate (FRR) which provided the percentage of pen-down points (segments) that were classied as pen up by our system. The examples of gures 1, 9, and 12 were used for training the HMM. 3.2.1 Point-wise classication results All points in the test sequences were hard classied as either pen down or pen up in this experiment. Two dierent approaches were compared: the rst one used the value of the ink absence condence measure as the classication parameter; the second approach used the HMM to classify each point. The hard classication was provided by the most likely HMM Figure 12: Example sequences. The rst row shows examples of sequences captured with the real-time system. We collected examples of cursive writing, block letters, printed letters, drawings, and mathematical symbols. The second row displays enlargements of portions of the sequence \Maria Elena". The dots represent the actual samples acquired with the interface. The sequences present a very low acquisition noise. state sequence obtained with Viterbi's algorithm. Table 2 shows the resulting error rates. Figure 13 presents the results of these two approaches on three test sequences. 3.2.2 Stroke classication results All test sequences were segmented into strokes and each stroke was classied as either pen down or pen up in this experiment. Two classication approaches were compared: the rst one was based on the ink absence condence measure; the second one was based on the local measurements (%) HMM modeling (%) FAR 24.6 28.6 FRR 10.05 5.33 Table 2: Point-wise classication results. Comparison of the error rates of point-wise ink detection obtained using the ink absence condence measure and the HMM model. The classication threshold used for the ink absence condence measure is 0.4. We observe that none of the approaches is clearly better than the other. The HMM one has a lower FRR while the local measurements one has lower FAR. As we pointed out before, we have to wait until the pen tip is out of sight in order to measure brightness, so many pen-up points that correspond to a stroke that passes on top a segment of ink trace were misclassied as pen-down points. This is the main reason for the apparently large value of the FAR. Figure 13: Point-wise classication results. The rst row shows three test sequences. The gures of the second row display each segment of the sequences with a thickness that is proportional to the average of the condence of the endpoints. Most of the thicker segments corresponds to portions of the trajectory that should be classied as pen down. The third row shows only points of the trajectories that have been classied as pen down using the ink absence condence measure. The fourth row presents only points that have been classied as pen down by the HMM. We see that there are several segments that appear in areas where there should be no ink trace on the paper. This misclassication is due to a bad measurement of the condence of ink absence. From the plots of the third and fourth row, it seems that the HMM approach has lower FRR at the cost of a higher FAR. local measurements (%) HMM modeling (%) FAR 9.27 11.22 FRR 24.55 8.18 Table 3: Stroke classication results. Comparison of the error rates of stroke classication obtained using the ink absence condence measure and the HMM model. For the rst case, the average ink absence condence measure was used as the classication parameter. The classication threshold was set to 0.2 (strokes with stroke condence lower than the threshold were classied as pen down). For the second case, the percentage of points in the stroke classied as pen down by the HMM was used as the classication parameter. The classication threshold was also set to 0.2 in this case. We observe that the HMM has a much better FRR than the local measurements at the expense of a slightly worse FAR. We note that in most of the cases in which the stroke-up classication fails (re ected in the FAR), it is due to an incorrect segmentation, like the \C" in the sequence \PEDRO MUNICH" or the crossing stroke of the \x" in the mathematical formula of gure 14. These incorrectly segmented strokes were always classied as pen down in the ground truth. Leaving out these segments in the computation of the performance, we obtained a reduction in the FAR for both methods of approximately 1% (absolute error) while the FRR is unchanged. HMM. In the rst approach, the stroke condence measure was computed as the average of the ink absence condence measure of all points in the stroke. For the case of the HMM, the stroke condence measure was calculated using a voting scheme. The ratio between the number of points classied as pen down by the HMM and the number of points in the stroke provided the stroke condence measure. Table 3 shows the resulting error rates. Figure 14 presents the results of stroke classication on three test sequences. 3.3 Signature verication As mentioned before, the real-time interface was used as front-end for a signature verication system [14, 16, 17]. We acquired 25-30 true signatures and 10 forgeries from 105 subjects, adding to an approximate total of 4000 signature samples. We collected data over the course of a few months in which subjects would provide signatures at dierent times during the day. The interface was placed next to a window, so natural sunlight was used for capturing signatures at day time, while electric lighting was used for acquiring signatures during the night. The subjects were asked to provide data in three dierent sessions in order to sample their signature variability. Given the number of subjects involved in the experiment, the Figure 14: Stroke classication results. The corresponding strokes for the sequences of gure 13 are shown on the rst row. Successive strokes are plotted alternatively with solid or dashed lines. The gures of the second and third rows correspond to classication using the ink absence condence measure; the gures of the fourth and fth rows correspond to classication using HMM. The gures of the second and fourth row displays each stroke with a thickness that is proportional to the stroke condence measure. The plots of the third and fth rows shows only the strokes classied as pen down in each case. The classication based the HMM seems to provide better results than the one based on the ink condence measure. Signature s019010 90 100 110 120 130 140 1502060Signature s025005 Signature s041010 100 120 140 16010305070Signature s001010 Signature s026005 -551525Forgery s019001 Forgery s025000 100 120 140 1601030Forgery s041006 100 120 140 160 1801030Forgery s026000 Figure 15: Signature verication. The rst row of the gure shows examples of true signatures acquired with the interface. The second row presents examples of corresponding skilled forgeries also captured with our system. position and orientation of the camera was dierent from session to session and from subject to subject. Figure 15 shows some examples of acquired signatures. We achieved a verication error rate of less than 1.5% for skilled forgeries and a verication error rate of less than 0.25% for random forgeries. These rates correspond to the condition of equal false acceptance rate and false rejection rate. These results and the techniques used for verication will be reported in a forthcoming paper. 3.4 Discussion The examples presented in section 3 show that the interface is quite convenient and accurate for acquiring short handwritten trajectories. The system has not been tested for acquiring long sentences or even full pages of text. The main di-culty in this case would be perspective and radial distortion. This is not a problem for some applications, e.g., our signature verication algorithm which encodes handwriting in an a-ne-invariant parameterization. Perspective distortion of the image could be corrected easily if paper with a predened pattern of symbols, e.g., a set of crosses located at a known distance from each other, was used; however, this would make the interface less convenient and general. Besides signature verication, informal tests by human observers found the output of the interface well within the resolution limits for easy reading and interpretation. However, the interface has not been tested for handwriting recognition. The results of the pen-down detection experiments are encouraging. The stroke condence measure provides a soft classication of the pen-down and pen-up strokes that could be used in a handwriting recognizer. The usability of the interface has been tested by more than a hundred dierent subjects during the signature verication experiment. The acquisition of signatures took place under various lighting conditions and camera position, showing the robustness of the interface with respect to variability of the user's setup. Handwriting was captured at dierent scales with the interface. Changes of scale were introduced by the user when he adjusted the position and orientation of the camera to write more comfortably. These scale changes were small enough to be handled with a xed set of system parameters. Larger scale changes would require adaptation of the system parameters to the acquisition setup. An appropriate procedure may be designed for the user to help the system in this task. 4 Conclusion and further work The design and implementation of a novel human-computer interface for handwriting was presented. A camera is focused on the user writing on a piece of paper with a normal pen. We have shown that the handwriting trajectory is successfully recovered from its spatio-temporal representation given by the sequence of images. This trajectory is composed by handwritten strokes and pen movements between two strokes. The temporal resolution is su-cient. The spatial resolution is approximately a tenth of a pixel, which allows capturing handwriting at su-cient spatial resolution within an area corresponding to half a sheet of letter paper using a cheap 480x640 pixels camera. The spatial resolution approximately corresponds to 20 samples per millimeter of writing, resolution that is ve times lower than that of commercial tablets (100 lines per millimeter), but that is obtained with a much smaller and cheaper interface. The classication of pen-up and pen-down portions of the trajectory of the pen is obtained by using local measurements of the brightness of the image at the location in which the writing end of the pen was detected. Several modules of the interface are susceptible of improvement. We used only one pen tip template for the whole sequence acquisition. This template could be automatically updated once the peak value of correlation fell below a certain threshold. Since the information about the boundaries and the axis of the pen tip, as well as the position of the ballpoint and the nger are computed for each frame by the ballpoint detection module, the automatic extraction of a new pen tip template involves no extra computational cost. The region of interest used to detect the location of the pen tip has constant size in the current implementation of the system. The size of this region could be driven by the uncertainty on the predicted position of the pen tip, i.e., the size could depend on the covariance of the predicted location of the pen tip. Smaller regions would be required in cases of low uncertainty, reducing in this way the computational cost of performing correlation between the region of interest and the pen tip template. The ballpoint detection in the current frame of the sequence is based on the orientation of the axis and of the boundaries of the pen tip in the previous frame. We could improve the robustness of the ballpoint detection by modeling the change of axis and boundaries orientations from frame to frame. A recursive estimation scheme could be used to predict the desired orientations, allowing one to reduce the size of the windows used to perform edge detection and to decrease the number of computations. We used a Gaussian model for the brightness of ink-less pixels. The estimation of the model parameters was performed using the brightness of points lying on a circle centered at the ballpoint position, assuming that all the circle points are ink-less points. Clearly, this model is not strictly adequate for a random variable which takes values on the interval [0; 255], and the assumption is not completely valid since some circle points could correspond to the ink trace. This model could be improved by using a probability density function suitable for representing a random variable that takes values on a nite interval. However, as a rst order approximation we have shown that this model provides good results in pen-up/-down classication. The classication of strokes into pen-up strokes and pen-down strokes is based on local measurements of brightness. A few other local measurements such as the local orientation of the ink at the position of the ballpoint, the correlation of this orientation with the local direction of the trajectory of the pen tip, etc., could be used in order to improve the classi- cation rates. These local measurements of direction would decrease the FAR since a sample would be classied as \pen down" only if an ink trace with the corresponding direction is found at the location of the sample. These additional local measures could be naturally included in the system by increasing the dimensionality of the observation of the HMM. The set of examples used to estimate the HMM parameters and to evaluate the pen-up/- classication performance included examples of dierent types of writing provided by only one subject. More example sequences provided by dierent subjects should be acquired in order to estimate this performance in a writer-independent setting. Also, a bigger set of examples should be used to obtain a more accurate HMM for pen down detection. Acknowledgements We gratefully acknowledge support from the NSF Engineering Research Center on Neuromorphic Systems Engineering at Caltech (National Science Foundation (NSF) Cooperative Agreement No. EEC-9402726). --R Optimal Filtering. On line handwriting data acquisition using a video camera. A computational approach to edge detection. Vision for man-machine interaction Finger tracking as an input device for augmented reality. Liveboard: a large interactive display supporting group meetings The Viterbi algorithm. Statistical Pattern Recognition. Applied Optimal Estimation. Optical character recognition A new approach to linear Dynamic approaches to handwritten signature veri An iterative image registration technique with an application to stereo vision. Visual Input for Pen-based Computers Visual input for pen-based computers Visual signature veri A new input device. Interactive dynamic whiteboard for educational applications. An evaluation of motor models of handwriting. Fundamentals of Speech Recognition. The state of the art in on-line handwriting recognition The relation between linear extent and velocity in drawings movements. Trajectory determines movement dynamics. Analysis and synthesis of handwriting. Adaptative thresholding for the digitaldesk. calibration for digitaldesk. A new data tablet system for handwriting characters and drawing based on image processing. --TR A computational approach to edge detection The State of the Art in Online Handwriting Recognition Introduction to statistical pattern recognition (2nd ed.) Liveboard Fundamentals of speech recognition On-Line and Off-Line Handwriting Recognition Camera-Based ID Verification by Signature Tracking Online Handwriting Data Acquisition Using a Video Camera Visual Input for Pen-Based Computers Visual input for pen-based computers --CTR Mario E. Munich , Pietro Perona, Visual Identification by Signature Tracking, IEEE Transactions on Pattern Analysis and Machine Intelligence, v.25 n.2, p.200-217, February

pen-based interface;systems and applications;pen-based computing;active and real-time vision

507649

Generic validation of structural content with parametric modules.

We demonstrate a natural mapping from XML element types to ML module expressions. The mapping is inductive and definitions of common XML operations can be derived as the module expressions are composed. We show how to derive, in a generic way, the validation function, which checks an XML document for conformance to its DTD (Document Type Definition). One can view validation as assigning ML types to XML elements and the validation procedure a pre-requisite for typeful XML programming in ML. Our mapping uses the parametric module facility of ML in some contrived way. For example, in validating WML (WAP Markup Language) documents, we need to use 36ary type constructors, as well as higher-order modules that take in as many as 17 modules as input. That one can systematically model XML DTDs at the module level suggests ML-like languages are suitable for type-safe prototyping of DTD-aware XML applications.

INTRODUCTION &MOTIVATION XML (eXtensible Markup Language) is language for tagging documents for their structural content [2]. A XML document is tagged into a tree of nested elements. XML is extensible because each XML document can include a DTD (Doc- ument Type which lists the tags of the elements and specifies the tagging constraints. A central concept in document processing is validation. A XML document is valid if its content is tagged with the constraints specified Submitted to International Conference on Functional Pro- gramming, 2001, and available as Technical Report TR-IIS- 001-005, Institute of Information Science, Academia Sinica, Taipei, Taiwan (http://www.iis.sinica.edu.tw). Comments and suggestions are most welcome. by its DTD. A XML document is well-formed if each of its element is enclosed with matching start-tag and end-tag. A well-formed XML document is not necessarily valid. The following XML document contains a DTD that defines two element types folder and record. The document contains as a root a folder element, which has an empty record element as its only child. It is a valid XML document. !?xml version="1.0"?? (folder,(folder-record)+))? !!ELEMENT record EMPTY? The DTD in the above XML document models the structure where a record must contain no other element, and no folder is empty or contains just another folder. One may think of it modeling a tidy bookmark file. Of the following three elements, f3 is valid, but items f1 and f2 are not. Note that !record/? is a shorthand for !record?!/record?. The tag sequence !record?!folder?!/record?!/folder? is an example of not-well-formedness. To simplify discussion, we may say that each element type in the DTD is specified by its element content model (i.e., its tagging constraint) which is an unambiguous regular expression with element type names as symbols. The content model of an element type specifies what element sequences are allowed as the children of the element. Naturally, when coding XML programs, one need to map the element types in a DTD to the corresponding data types in the source programming language. A further requirement of the mapping is that content validation is translated into type correctness in the programming language, so that well-typed programs will always produce valid XML elements. Note that this goes beyond what is required of the so-called "validating XML processor", which need only report violations of element content models in the input XML document but need not impose restrictions on the output. There have been several directions in programming language support for writing XML applications. We can classify them into the following three categories. ADT for well-formed elements. Abstract data types and the accompanying library routines are designed to traverse and transform well-formed XML elements. The XML data is assumed to be validated in a separate phase, or its validation is a separate issue and may not even be required. Examples in this category include standard XML API in C++, Java, or other languages (e.g., Document Object Model, DOM [1]) and a combinator approach to writing XML processing functional programs [3, 18]. Type translation of DTD. A strongly typed language is used for XML programming, and the type system of the language is used to embed DTDs. The embedding is complete (every element type has a corresponding data type in the embedding language) and sound (an expression of the embedding language evaluates to a valid XML element if the expression is well-typed in the language). Examples in this category include HaXml [3, 18] and XMLambda [14]. If the strongly typed language is statically typed, then the soundness proof is done by the type checker at compile- time. Hence no type-correct program will produce invalid XML elements. One can also use constraint-based languages or logic programming languages to encode content models in a similar way [19]. The type translation approach is not completely satisfactory for two reasons. One is that the type translation may not be systematic and can be tedious if done manually. The other inconvenience is that code for generic XML processing operations need to be rewritten for every DTD because they are translated into different types. XML content validation, which check well-formed XML documents for conformance to their DTDs, is such a generic operation. Native language support of DTD. New languages are being designed with builtin XML support to help build XML-related applications. XDuce is a functional language with regular expression types, so as to allow direct representations of DTDs and processing of valid elements Expressions in the language are evaluated to valid XML elements, but variables must be annotated with their element types. The concept of validation is built into the language as type correct- ness, and programs are type-checked at compile-time. XDuce also provides regular expression patterns which further help write concise XML programs. XDuce, however, is currently a first-order and monomorphic language, and lacks some language features (e.g., a module system). In this paper, we show how to use parametric modules in ML-like languages to write XML-supporting program modules that are both expressive and generic. It is expressive because all XML DTDs can be constructed from the provided parametric modules. It is generic because common operations, including the validation function, are automatically generated. As such, our approach has the advantages of both the type translation approach and the native DTD support approach, but without their disadvantages. There is no need to recode generic operations, and no need to design new language. 2. AN ILLUSTRATING EXAMPLE For the tidy bookmark example described in Section 1, the following is the actual code we write in Objective Caml to specify the DTD, and to produce the validation functions for the two element types in the DTD. module struct type ('x0, Record of 'x1 let map (f0, module struct module module module Tag = BookmarkTag module In the above, module TidySys contains two modules F0 and F1, which are translations, word by word, in Objective Caml module language the XML element type declarations of folder and record. The higher-order module Alt is for "-", Seq for ",", Star for "*", and Plus for "+". Ideally, we would like to define the two XML element types as two mutually recursive ML modules T0 and T1 as the following. module and But Objective Caml, as most ML-like languages, does not support recursive modules. Instead we use two "place holder" modules P0 and P1 as the two parameters to higher-order modules (Alt, Seq, etc.), and use another higher-order module Mu (pronounced as -) to derive the two simultaneous fixed points. Module TidyDtd contains ffl module U, which defines the type for well-formed elements ffl module V, which contains modules T0 and T1 that each defines the type for valid folder and record elements, ffl functions validate and forget, which provide mappings between well-formed elements and valid elements. It also defines exception Invalid, which may be raised by function validate. Note that the following equations always hold forget validate The sample element f3 as shown in Section 1 can now be defined and validated by the following Objective Caml code (f3 u is well-formed and f3 v is In addition, the valid element returned by the validation function is parsed and typed in the sense that all of its sub-structures are given specific types and can be extracted by using ML pattern-matching. In this paper, we will use the above example to explain the idea and describe the construction. However, the idea and the construction can be systematically applied to DTDs with element types. One need just to define a n-ary fixed point module Mun that will take a system of n n-ary higher-order modules F0 , F1 produce the simultaneous fixed points. The definition of Mun is symmetric and is similar to Mu. We will later use WML (a markup language for wireless applications whose DTD defines 35 element types) as a benchmarking example to show the effectiveness of our approach. 3. GENERICPROGRAMMINGWITHPARA- The XML element types in the folder example can be translated into Objective Caml using a series of type definitions as shown below. type ('a, 'b) alt = L of 'a - R of 'b type ('a, 'b) list type 'a plus = One of 'a - More of 'a * 'a plus Folder of ((record, (folder, record) alt star) seq, (folder, (folder, record) alt plus) seq) alt One can abstract the right-hand-sides of the type equations for folder and record into two binary type constructors f0 and f1, and view folder and record as the least fixed points of f0 and f1. Functions folder and record are syntactic sugars, and can be defined by let folder ulist = BookmarkTag.Folder (TidyDtd.U.up ulist) let record ulist = BookmarkTag.Record (TidyDtd.U.up ulist) type ('a, 'b) ('a, ('a, 'b) alt plus) seq) alt type ('a, 'b) Folder of (folder, record) f0 and record = Record of (folder, record) f1 One can further rewrite f0 and f1 using the two projection functions p0 and p1, and the empty type constructor. type ('a, 'b) type ('a, 'b) type ((('a,'b)p1, (('a,'b)p0, ('a,'b)p1) alt star) seq, type At this point, it is clear that one can program in the module level, and define f0 and f1 as two module expressions using a predefined set of constant modules (for p0, p1, and empty), unary parametric modules (for star and plus), and binary parametric modules (for alt and seq). This is shown in Figure 1 where we also define the map function, inductively. All XML element types can be defined using a fixed set of parametric modules. We may say that modules F0 and F1 are objects in a functor category where each object has a type constructor t to map types to types, and a function map to map typed functions to typed functions. Parametric modules like Plus are arrows in the functor category, i.e., natural transformations. We view this definition of the map function a generic one, as each map instance is inductively indexed by its governing type expression. We will later show definitions of other generic values that are used in the definition of the validation function (which itself is generic as well). 4. PARAMETRICCONTENTMODELSAND SIMULTANEOUS FIXED POINTS In Figure modules F0 and F1 each defines a binary type constructor t, and the the two type constructors are used together to mutually define types folder and record. The code is reproduced below. module F0: module F1: Folder of (folder, record) F0.t and record = Record of (folder, record) F1.t The type constructors F0.t and F1.t are parametric content models in the sense that each maps a tuple of type instances to a content model. For example, given type instances folder and record, the type expression (folder, record) F0.t expands to ((record, (folder, record) alt star) seq, module type sig type ('a, 'b) t val map: ('a -? 'x) * ('b -? 'y) -? module type module type module Empty: struct type ('a, 'b) let map (f, g) module P0: struct type ('a, 'b) let map (f, g) module Plus: struct type ('a, 'b) One of ('a, 'b) F.t - More of ('a, 'b) F.t * ('a, 'b) t let rec map (f, g) match t with One s -? One (F.map (f, g) s) - More (v, w) -? More module struct type ('a, 'b) let map (f, g) (u, F1.map (f, g) v) module P1: module Star: module Alt: module F0: module F1: Folder of (folder, record) F0.t and record = Record of (folder, record) F1.t Figure 1: Inductive definitions of XML element types using parametric modules. Note: Module type annotations can be, and often are, omit- ted. W can take out the ": F2F" part in "module Plus: ", and at the same time expose the implementation of module Plus. The annotations are added for clarity and type-checking purposes. (folder, (folder, record) alt plus) seq) alt which is exactly the XML content model for element type folder. The main idea is to use type constructors as parametric content models, and view XML element types as simultaneous fixed points of a set of parametric content models. This viewpoint helps us develop primitive functions that are abstract and applicable to different content models (that is, the primitives are polymorphic). One of these primitives is the simultaneous induction operator - the fold function. We will later show that the validation procedure can be defined by using the fold function. We then model two recursively defined XML element types by two interdependent ML modules T0 and T1. Their signatures are the following. module T0: sig type ('x0, 'x1) cm val up: (T0.t, T1.t) cm -? T0.t val down T0.t -? (T0.t, T1.t) cm and module T1: sig type ('x0, 'x1) cm val up: (T0.t, T1.t) cm -? T1.t val down T1.t -? (T0.t, T1.t) cm In the above, type constructor ('x0, 'x1) cm is for the parametric content model, and type t is for the element type. Functions up and down map between an element and its content model, and together define their equivalence: Note that the above mutually defined signatures are not allowed in Objective Caml (as in most ML-like languages). However, one can use both auxiliary type names and additional type sharing constraints to overcome the problem. We can define a higher-order module MuValid that derives modules T0 and T1, when given a module that specifies the corresponding parametric content models and the tag set, see Figure 2. In Figure 2, modules F0 and F0 of the input module S specify the parametric content models, and module Tag specifies the tag set. Note that, in the module returned by MuValid, the type for all valid elements is simply defined as the disjoint sum of type T0.t and type T1.t: Also note that the simultaneous fold function has type val fold: (('a, 'b) T0.cm -? 'a) * (T0.t -? 'a) * (T1.t -? 'b) Function fold returns with two reduction functions (whose types are T0.t -? 'a and T1.t -? 'b) if given two properly typed induction functions as bases (whose types are ('a, 'b) T0.cm -? 'a and ('a, 'b) T1.cm -? 'b). Similarly, a higher-order module MuWF can be defined to derive a module for all well-formed elements; see Figure 3. In module MuWF, type constructor ('x0, 'x1) cm - the parametric content model for well-formed elements - is defined as a list of tagged values: list and type u - the type for well-formed elements - is defined as the fixed point of the parametric content model cm: type Note as well that type of all well-formed elements, type t, is defined as the disjoint sum of u and u, representing elements with two distinct tags. The definition of the simultaneous fold function is the same as that in module MuValid. In Figure 3, there are several functions in module U2V and V2U that are given their types but are left undefined. They are used to specify functions validate and forget. Function validate maps a well-formed element to a valid ele- ment, while forget is the inverse function. Let us look at functions cm0 and cm1 in module U2V first. Their types are the following val cm0: (V.T0.t, V.T1.t) U.cm -? (V.T0.t, V.T1.t) V.T0.cm val cm1: (V.T0.t, V.T1.t) U.cm -? (V.T0.t, V.T1.t) V.T1.cm Function cm0 maps a well-formed content, whose constituting parts are valid elements already, into a valid content. If function cm0 is composed with function V.T0.up, one gets a function that returns a valid element of type V.T0.t as result (we use $ as the function composition operator): Given these two functions as the inductive bases to the simultaneous fold function, one derives the validation functions for elements of types V.T0.t and V.T1.t. module type sig type ('x0, 'x1) t val map: ('x0 -?'y0) * ('x1 -? 'y1) -? module type sig module F0: FUN module F1: FUN module Tag: TAG module struct module Tag = S.Tag and module struct let up let down (V0 module struct let up let down (V0 let fold (f0, let rec (T0.down and (T1.down in Figure 2: Module MuValid derives element types as simultaneous fixed points of a set of parametric content models. module struct module list let map type let down (U let fold (f0, let rec (down and (down in module struct module module module exception Invalid module struct let cm0: (V.T0.t, V.T1.t) U.cm -? (V.T0.t, let cm1: (V.T0.t, V.T1.t) U.cm -? (V.T0.t, let (t0, t1): (U.u -? V.T0.t) * (U.u -? U.fold (V.T0.up let t: U.t -? module struct let cm0: (U.u, U.u) V.T0.cm -? let cm1: (U.u, U.u) V.T1.cm -? let (t0, t1): (V.T0.t -? U.u) * (V.T1.t -? V.fold (U.up let t: V.t -? Figure 3: Module MuWF derives the type for well-formed elements. Module Mu uses simultaneous fold to define the validation function. Note: Type annotations for functions are added for clarity purpose. U.fold (V.T0.up Recall that the types for all well-formed elements and all valid elements are defined by It follows that the validation function is defined by U.fold (V.T0.up As shown in Figure 3, one can define function forget in a similar way. It remains to be shown how functions like cm0 and cm1 are defined for all content models. This is shown next. 5. GENERIC VALIDATION OF CONTENT MODELS Recall that, in Figure 1, a map function is defined in a generic way for any module with signature FUN, as long as the module is generated with the predefined set of parametric modules (Empty, P0, P1, Star, etc. The vaildation and forgetting functions can be defined in a generic way as well. First we define the validation functions for the inductive bases. The validation function for any other content model can then be derived, automatically, as module expressions for the content are built. There are two remaining details. The first is that at the time of building the content model, one does not have access to the tag module. This tag module is of signature TAG, and defines the variant data type for tagging elements (e.g., module BookmarkTag in Section 2). Therefore the validation and forgetting functions must reside in a higher-order module that takes in a TAG module as input. One need also to maintain a nullable condition and a first set of element tags. A content model is nullable if it accepts the empty element sequence. The first set contains all tags that can appear at the first position of a valid sequence. It can be used to check if a content model is ambiguous, e.g., when the first sets of the two input modules to Alt overlap. When combined with a lookahead tag, it is used to implement a non-backtracking validation procedure as well. (More on this in Section 8.) Both nullable and first are generic values. The module signature FUN for parametric content model now consists of the following components. module type sig type ('x0, 'x1) t val map: ('x0 -? 'y0) * ('x1 -? 'y1) -? val nullable: bool val first: Natset.t module Content: functor (T: TAG) -? sig val validate: ('x0, 'x1) T.t list -? val forget: ('x0, 'x1) t -? ('x0, 'x1) T.t list Function validate takes a list of tagged values and turns it into a value of content model followed with the remaining list. Note that the type for the input, ('x0, 'x1) T.t list, is the same as the content model of well-formed element if the two share the same tag set. Figure 4 illustrates the construction by showing the implementations of modules P0 and Star. The validation and forgetting functions are wrapped in module Content. The definition of Content is inductive: It depends on the Content module in the input module F (see, e.g., the module expression module Star). We can view this as constituting a generic definition of the validation function, as each instance is systematically generated by its module expression. As evident in module Star, we adapt the longest prefix matching rule in validating the input element sequence against the "*" content model. This longest prefix matching rule is indeed required by XML. Validation functions for other modules, i.e., Empty, P0, P1, Plus, Seq, and Alt, can be similarly defined and are omitted here. Now we return to Figure 3 to complete the defintions of functions cm0 and cm1 in modules U2V and V2U. They are defined as the following. module struct module let cm0 ulist = match CM0.validate ulist with Some (v, []) -? v -? raise Invalid module struct module Function cm0 in module U2V need to validate the input sequence of tagged value with the content model of element type V.T0.t, using the current tag set. This can be accomplished by using the validation function in module Sys.F0.Content(Sys.Tag). The only difference is that, if there remains a non-empty sequence after a validated (longest) prefix, the entire sequence is not valid with respect to the content model V.T0.t. module P0: struct type ('x0, Natset.of-list [0] module Content = functor (T: TAG) -? struct let validate ulist = match ulist with T.fold ((fun x -? Some (x, t)), (fun x -? (* if success, return the untagged value along with the remaing list; otherwise returns None. *) let forget a = [T.x0 a] (* Tag with the first variant of type T.t *) module Star: struct type ('x0, F.first module Content = functor (T: TAG) -? struct module let rec validate ulist = match ulist with [] -? Some ([], ulist) if . h in first . then match CM.validate ulist with Some (u, t) -? (match validate t with Some (us, s) -? Some (u::us, s) else Some ([], ulist) let rec forget match t with Figure 4: Generic definition of the content validation functions. 6. TYPEFULXMLPROGRAMMINGINML One of the purposes of validation is to assign a type to an XML element. Programming with validated XML elements is now programming with typed values. Using a statically typed langauge for such programming allows one to detect type errors, hence expressions for invalid elements, at compile time. Our generic validation procedure gives types to valid ele- ments, and allows one to construct XML processors in a typeful way. In the following illustrating diagram, let U be the ML type for well-formed elements, and V and V 0 be the ML types that correspond to specific XML element types. U validate f forgetWe may say that functions in U ! U are untyped as they may produce invalid elements. However, functions in are typed as they always output valid elements. Whenever one is programming a function expects the output also to be valid, one can do so by programming a In Figure 5, we show some ML code fragment to illustrate the approach. The code maps a well-formed tidy bookmark to a well-formed flat bookmark (function tidy2flat u). Because the the mapping is composed from a typed conversion routine (function tidy2flat v), it will always output a valid element if the input element is valid. Note that the types for the functions below will be inferred by ML. The functions are annotated with their types in Figure 5 for clarity purpose only. 7. COMBING GENERICITY WITH POLY- MORPHISM The generic modeling of XML DTDs can be combined with type polymorphism for a better result. Indeed, we use both genericity and polymorphism to model XML element type declarations that are accompanied with attribute-list declarations. We can extend the previous folder example by requiring an optional subject attribute for each folder element, and a pair of title and url attributes for each record element. The following is a valid XML document with the newly extended DTD. !?xml version="1.0"?? (folder,(folder-record)+))? !!ELEMENT record EMPTY? subject !!ATTLIST record title module module struct module module module Tag = Tag module module module module module module let t2f-folder: fun fd -? match fd with . (* the case of a flat record r followed by a sequence t of flat records or folders *) . (* the case of a flat folder f followed by a non-empty sequence t of flat records or folders *) let t2f-record: let flatten-v: (TidyFolder.t, TidyRecord.t) Tag.t -? Tag.map (TidyDtd.V.fold FlatRecord.up let flatten-u: TidyDtd.U.t -? FlatDtd.forget Figure 5: An example of typeful XML programming Note: Type annotations for functions are added for clarity purpose. url !folder subject="Research Institutes"? !record title="Academia Sinica" url="http://www.sinica.edu.tw"/? The original definitions of folder and record (Figure 1, last two lines), Folder of (folder, record) F0.t and record = Record of (folder, record) F1.t can now be replaced by the following type ('u, 'v) Folder of and ('u, 'v) record = Record of 'v * (('u, 'v)folder, ('u, 'v)record) F1.t string option- string - In the above, attribute declarations are modeled at the type level. It can be lifted to the model level if needed. Further- more, the generic definition of the validation function can be modified accordingly to accommodate validation check for attribute formats and values. 8. MORE XML CONTENT VALIDATION XML requires content models in element type declarations be deterministic. Br-uggemann-Klein and Wood further clarified the requirement as meaning 1-unambiguity [7, 8]. A regular expression is 1-unambiguous if its sequence of symbols can be recognized deterministically, with one-symbol lookahead, by the corresponding nondeterministic finite-state machine. For example, the content model ((b, c)-(b, d)) is not 1-unambiguous, because given an initial b, one cannot know which b in the model is being matched without looking further ahead to see what follows b. However, the equivalent content model (b,(c-d)) is 1-unambiguous [2]. We can use the nullable predicate and the first set to check whether the content model as specified by a module expression is 1-unambiguous. The check is performed at module elaboration time so that an ambiguous content model is detected and an exception is raised as soon as possible. A content model may also contain epsilon ambiguity which is allowed by XML but demands additional work during validation. An example of epsilon ambiguity is (a*-b*), when the empty sequence is derivable from both a* and b*. Besides element content models (i.e., regular expressions on element type names), an XML element type may use other content specifications. For example, the element type may have EMPTY or ANY specification, or mixed content specifi- cation. These specifications impose no additional difficulty in the definition of the generic validation function. The ANY specification means that the sequence of child elements may contain elements of any declared element types, including text, in any order. The mixed content specification allows text data to be interspersed with elements of some prescribed types. One may think of ANY as a special case of mixed content. One can view text data, which is denoted as ("Parsed Character Data") in a mixed content specification, as elements enclosed within an pair of implicit !text? start-tag and !/text? end-tag. A Pcdata module, similar to the Empty module we already have, can be defined to help inductive definitions of mixed content specifications. For ex- ample, for DTDs with 2 element types, one can define an Any module as following by using a 3-ary alternative module Alt3: module Any: 9. EXPERIENCE WITH LARGER DTDS WML is a markup language for WAP applications. Its DTD consists of 35 element type definitions. We have applied the generic approach to validate WML documents. In order to do so, we need to produce ML modules that include and operate upon 36-ary type constructors (35 element types plus 1 for #PCDATA). We also need to construct higher-order modules that take in as many as 17 modules as input (one of the element type definitions needs a 17-ary Alt mod- ule). Our experience has been quite satisfactory: Our code is compiled without problem with Objective Caml, but the compilation time is not negligible (about 1 min. at a desktop workstation). The validation time is negligible how- ever, at least for the smallish examples we have tried (around 100 elements). We are working on both larger DTDs and documents, and are collecting more performance data. The size of the ML source code is quite large, however. Take the following ML module expression as an example. module One need a 10-ary module Seq10 to construct the required content model, which specifies a sequence of 10 elements, each of a different element type. Code for module Seq10 looks like the following: module struct type ('x0, 'x1, . , 'x35) t * ('x0, 'x1, . , 'x35) F1.t * . * ('x0, 'x1, . , 'x35) F9.t It is clear from the above that, for a DTD with n element types, the source for module Seq m will have code size O(mn). At the worst case, for a DTD of length n, our code will need O(n) unique type variables, will contain type sharing constraints of length O(n 2 ), and will have a overall code size of O(n 2 ). The source code of all the necessary ML modules for the 35-element WML DTD has a size of about 0.5 MB. When compiled, it produces a binary of size 175 KB (*.cmo file in Objective Caml), and an interface of size 2.3 MB (*.cmi file in Objective Caml). ML code for the WAP examples is accessible at the following URL: http://www.iis.sinica.edu.tw/~trc/x-dot-ml.html One can do a connected component analysis on the DTD so that the set of element types are partitioned into disjoint subsets where there is no type-dependency between the sub- sets. A subset with k element types need only use k-ary type constructors, and the overall code size for the modules used for the subset can be reduced. 10. RELATED WORK AND CONCLUSION In Section 1, we have introduced previous work that uses existing or new functional languages to model and program with XML DTDs. There is a wealth of research and system work that is related to XML content modeling but is not necessarily from the perspective of (functional) programming languages. We list just a few here. Br-uggemann-Klein and Wood addressed the problem of ambiguous XML (and SGML) content models, based on theory of regular languages and finite automata [7, 8]. In particu- lar, they showed that linear time suffices to decide whether a content model is ambiguous. It is showed that regular expressions in both "star normal form" and "epsilon normal are always unambiguous [9]. The Glushkov automaton that corresponds to a regular expression is used for checking ambiguity and, if not unambiguous, for validation as well. Murata has proposed a data model for XML document transformation that is based on forest-regular language theory [15, 16]. His model is a lightweight alternative to XML Schema and provides a framework for schema trans- formation. There is also work on type modeling for document transformation in a structured editing systems using data types [5]. However, none of the above work has used specific programming language as a modeling language. XML Schema is a maturing specification language for XML content that is being developed at World Wide Web Consortium [4]. XML Schema is more expressive than DTD and the specification language itself uses XML syntax. The difference between XML Schema and DTD seems to be XML Schema's ability to derive new types by extending or restricting the content models of existing types. XML Schema also provides a "substitution groups" mechanism to allow elements to be substituted for other elements. We are investigating whether ML-like module languages are expressive enough to model these mechanisms. Backhouse, Jansson, and Jeuring, and Meertens have written a detailed introduction to generic programming [6]. See also the introduction to fold/unfold by Meijer, Fokkinga, and Paterson [13], as well as work on using fold/unfold for structuring and reasoning about program semantics by Hutton [12]. Our extension of simple fold to simultaneous fold seems new. Most work about generic programming in the functional programming research community seems to rely on the mechanism of type class to derive type-specific instances of generic functions. The language of choice is often Haskell. We have shown in this paper that the parametric module mechanism in ML-like languages is suitable for generic programming as well. In fact, we think that parametric modules allow one to take finer control on the inductive derivations of generic values. More powerful module systems have been developed to allow mutually recursive modules, as well as modules that depend on values and types (see, e.g., Russo [17]). However, we showed here that the lack of recursive modules need not be a problem as long as the mutual dependency between the modules is only about interdependent type definitions. Viewed in the above context, our work can be thought to use the ML module facility to generate a deterministic automata that is specialized for the validation of elements for a specific DTD. Validation automata also gives types to the elements (and its parts). In additional, the construction of the validation automata is entirely generic and can be au- tomated. Our work also serves as a usage case of ML parametric modules, and can be used to stress test current ML implementations. It is a delight to see our contrived code of 36-ary type constructors and 17-ary higher-order modules is compiled and executed with no problem under Objective Caml. 11. --R XML Schema Part 0: Primer. Type modelling for document transformation in structured editing systems. Generic programming: An introduction. Anne Br-ugemann-Klein and Derick Wood Anne Br-uggemann-Klein XDuce: A typed XML processing language. Regular expression types for XML. Fold and unfold for program semantics. Functional programming with bananas Transformation of documents and schemas by patterns and contextual conditions. Data models for document transformation and assembly. Haskell and XML: Generic combinators or type-based translation? <Proceedings>In Proceedings of the International Conference on Functional Programming</Proceedings> A logic programming approach to supporting the entries of XML documents in an object database. --TR Functional programming with bananas, lenses, envelopes and barbed wire Regular expressions into finite automata One-unambiguous regular languages The under-appreciated unfold Fold and unfold for program semantics Haskell and XML Regular expression types for XML Recursive structures for standard ML A Logic Programming Approach to Supporting the Entries of XML Documents in an Object Database Out-of-Core Functional Programming with Type-Based Primitives Transformation of Documents and Schemas by Patterns and Contextual Conditions XDuce An Algebra for XML Query Data Model for Document Transformation and Assembly --CTR Tyng-Ruey Chuang , Jan-Li Lin, On modular transformation of structural content, Proceedings of the 2004 ACM symposium on Document engineering, October 28-30, 2004, Milwaukee, Wisconsin, USA

fixed points;functional programming;XML;validation;modules and interfaces

507673

Probabilistic congestion control for non-adaptable flows.

In this paper we present a TCP-friendly congestion control scheme for non-adaptable flows. The main characteristic of these flows is that their data rate is determined by an application and cannot be adapted to the current congestion situation of the network. Typical examples of non-adaptable flows are those produced by networked computer games or live audio and video transmissions where adaptation of the quality is not possible (e.g., since it is already at the lowest possible quality level). We propose to perform congestion control for non-adaptable flows by suspending them at appropriate times so that the aggregation of multiple non-adaptable flows behaves in a TCP-friendly manner. The decision whether or not a flow is to be suspended is based on random experiments. In order to allocate probabilities for these experiments, the data rate of the non-adaptable flow is compared to the rate that a TCP flow would achieve under the same conditions. We present a detailed discussion of the proposed scheme and evaluate it through extensive simulation with the network simulator ns-2.

Introduction C ONGESTION control is a vital element of computer networks such as the Internet. It has been widely discussed in the literature { and experienced in reality { that the lack of appropriate congestion control mechanisms will lead to undesirable situations such as a congestion collapse [1]. Under such conditions, the network capacity is almost exclusively used up by trac that never reaches its destination. In the current Internet, congestion control is primarily performed by TCP. During recent years, new congestion control schemes were devised, supporting networked applications that cannot use TCP. Typical examples of such applications are audio and video transmissions over the Internet. One prime aim that these congestion control schemes try to achieve is to share the available band-width in a fair manner with TCP-based applications, thus falling into the category of TCP-friendly congestion control mechanisms. TCP, as well as existing TCP-friendly congestion control algorithms, require that the data rate of an individual ow can be adapted to network conditions. Using TCP, it may take a variable amount of time to transmit a xed amount of data, or with TCP-friendly congestion control, the quality of an audio or video stream may be adapted to the available bandwidth. While for a large number of applications this not a limi- tation, there are cases where the data rate of an individual ow is determined by the application and cannot be adjusted to the network conditions. Networked computer games are a typical example, considering the fact that players are very reluctant to accept the delayed transmission of information about a remote player's actions. Live audio and video transmissions with a xed minimum qual- ity, below which reception is useless, fall into the same category. For this class of applications there are only two acceptable states: either a ow is on and the sender transmits at the data rate determined by the application or it is no data is transmitted at all. We call network ows produced by these applications non-adaptable ows. In this paper we describe a TCP-friendly end-to-end congestion control mechanism for non-adaptable unicast ows called Probabilistic Congestion Control (PCC). The main idea of PCC is to calculate a probability for the two possible states (on/o) so that the expected average rate of the ow is TCP-friendly, to perform a random experiment that succeeds with the above probability to determine the new state of the non-adaptable ow, and to repeat the previous steps continuously to account for changes in the network conditions. Through this mechanism it is ensured that the aggregate of multiple PCC ows behaves TCP-friendly. The remainder of this paper is structured as follows. Section II summarizes related work. In Section III we examine non-adaptable ows in more detail. A thorough description of the PCC mechanism is given in Section IV. The results of the simulation studies that were conducted are presented in Section V and we conclude the paper with a summary and an outlook on future work in Section VI. II. Related Work Much work has been done on TCP-friendly congestion control schemes for applications that cannot use TCP. Prominent examples of these schemes are PGMCC [2], TEAR [3], TFRC [4], and FLID-DL [5]. A discussion of such TCP-friendly congestion control mechanisms can be found in [?]. TCP, as well as all existing TCP-friendly congestion control schemes, requires that the bandwidth consumed by a ow be adapted to the level of congestion in the network. By denition, non-adaptable ows cannot use such congestion control mechanisms. It is conceivable to use reservation mechanisms such as Intserv/RSVP [7] or Diserv [8] for non-adaptable ows so as to prevent congestion altogether. However, these mechanisms require that the network supports the reservation of resources or provides dierent service classes. This is currently not the case for the Internet. In con- trast, PCC is an end-to-end mechanism that does not require support from the network. With PCC it is possible to \partly" admit a ow and to continuously adjust the number of ows to network conditions. We are not aware of any previous work that directly matches the category of probabilistic congestion control. III. Non-Adaptable Flows For the remainder of this paper, a non-adaptable ow is dened as a data ow with a sending rate that is determined by an application and cannot be adjusted to the level of congestion in the network. A non-adaptable ow has exactly two states: either it is in the state on, carrying data at the rate determined by the application, or it is o, meaning that no data is transmitted at all. Any data rate in between those two states is inecient, since the application is not able to utilize the oered rate. Examples of applications using non-adaptable ows are commercial network games such as Diablo II, Quake III, Ultima Online, and Everquest. These games typically employ a client-server architecture. The data rate of the ows between client and server is determined by the fact that the actions of the players must be transmitted instan- taneously. Similar restrictions hold for the ows between participants of distributed virtual environments without a centralized server. If a congestion control scheme delays the transmission of actions too long, the application quickly becomes unusable. This can easily be experienced by experimenting with a state-of-the-art TCP-based networked computer game during peak hours. For this rea- son, a number of applications resort to UDP and avoid congestion control altogether. A situation with either no congestion control at all or vastly reduced utility in the face of moderate congestion is not desirable. A much more preferable approach is to turn the ows of some participants o and to inform the application accordingly. All other participants do not need to react to the congestion. On average, all users should be able to participate in the session for a reason-able amount of time between o-periods to ensure utility of the application. At the same time, o-periods should be distributed fairly among all participants. Other examples of applications with non-adaptable ows are audio or video transmissions with a xed quality. There are two main reasons why it may not be possible to scale down a media ow: either the user does not accept a lower quality, or the quality is already at the lowest possible level. The second reason indicates that a congestion control mechanism for non-adaptable ows can complement congestion control schemes that adapt the rate of a ow to current network conditions. IV. Probabilistic Congestion Control The Probabilistic Congestion Control scheme (PCC) provides congestion control for non-adaptable unicast ows by suspending ows at appropriate times. PCC is an end-to-end mechanism and does not require the support of routers or other intermediate systems in the network. The key aspect of PCC is that { as long as there is a suciently high level of statistical multiplexing { it is not important that each single non-adaptable ow behave TCP-friendly at any specic point of time. What is important is that the aggregation of all non-adaptable ows on a given link behave as if the ows were TCP-friendly. Due to the law of large numbers this can be achieved if (a) each PCC ow has an expected average rate which is TCP-friendly and if (b) each link is traversed by a su- ciently large number of independent PCC ows. At rst glance (b) may be considered problematic, because it is possible that a link is traversed only by a small number of PCC ows. However, further re ection reveals that in this case the PCC ows will only be signicant in terms of network congestion if each individual PCC ow occupies a high percentage of the link's bandwidth. We therefore relax (b) to the following condition (c): a single PCC ow is expected to have a rate that is only a small fraction of the available bandwidth on any link that it crosses. Given the current development of available bandwidth in computer networks, this is a condition that is likely to hold true. A. Requirements There are a number of requirements that have to be fullled in order for PCC to be applicable: R1: High level of statistical multiplexing. Condition (c) discussed above is met. R2: No synchronization of PCC ows at startup. PCC ows start up independent of each other. R3: The average rate of a PCC ow can be predicted. In order for PCC to work, it must be possible to predict the average rate of a PCC R4: The average rate of a TCP ow under the same conditions can be estimated. We expect that there is a reasonably accurate method to estimate the average bandwidth that a TCP ow would have under the same network conditions. B. Architecture A simple overview of the PCC architecture is depicted in Figure 1. A PCC sender transmits data packets at the rate determined by the application, while the PCC receiver monitors the network conditions by estimating a TCP-friendly rate using a model of long-term TCP throughput. Whenever a PCC receiver observes a degradation in network conditions, it conducts a random experiment to determine whether or not the ow should be suspended. In the case of a negative result, a control packet is sent to notify the sender that it is temporarily required to stop. After a certain o-period, a sender may then resume data transmission. For PCC, we chose to allocate as much functionality to the receiver as possible to facilitate a future extension of PCC to multicast. (data, reflected timestamp, sequence number) data packets receiver sender control packets (flow state: on/off, timestamp) - start stop flow reflect timestamp on/off-probability calculation parameter measurements TCP-friendly rate and - random experiment Fig. 1. PCC Architecture While a ow is in the on-state, control packets are sent at certain time intervals. They allow to continuously measure the round-trip time required to determine the TCP-friendly rate and they serve as a backup mechanism in case of very heavy network congestion. In the absence of these periodic control messages, the sender stops sending, thus safeguarding against the loss of notications to stop. As long as the ow is in the on-state, the data packets are transmitted at the rate determined by the applica- tion. Each data packet includes the timestamp of the most recent control packet that the sender has received in order to be able to determine the round-trip time. Each 1 There are multiple ways in which this can be done, ranging from a constant bit-rate, ow where this prediction is trivial, to the usage of application level knowledge or prediction based on past samples of the data rate. data packet also contains a sequence number to allow the receiver to detect packet losses. For the remainder of this work we use the TCP through-put formula of Padhye et al. [9] to compute the TCP-friendly rate. In order to determine the parameters required for the formula, the current version of PCC uses the measurement mechanisms proposed for the TCP-Friendly Rate Control Protocol (TFRC) [4]. However, it is important to note that PCC is independent of the method used to estimate the throughput of a TCP ow for given network conditions. A possible alternative, for example, would be to use the rate calculation mechanism of TCP Emulation At Receivers (TEAR) [3]. C. Continuous Evaluation To determine the probability with which a PCC ow is allowed to send for a certain time interval T , it is necessary to compare the average rate r NA of PCC to the TCP-friendly rate r TCP . r NA (1) where p denotes the ratio of r NA to r TCP . When solving the equation, two outcomes are possible: 1: The non-adaptable ow consumes less than or the same amount of bandwidth that would be TCP-friendly and should therefore stay on. 1: The non-adaptable ow consumes more bandwidth than a comparable TCP-friendly ow. In this case, p is taken as a probability and the non- adaptable ow should be turned For uniformly distributed random number x is drawn from the interval (0; 1]. If x > p holds, the PCC ow is turned o for a time of T . After that time interval the ow may be turned on again. If x p, then the ow remains in the on-state. Since we require a sucient level of statistical multiplexing (R1) and because of the law of large numbers, the aggregation of all PCC ows behaves as if each of them were TCP friendly. T is an application-specic parameter that is crucial for the utility of the protocol and thus for the user acceptance of the congestion control mechanism. For example, if short news clips are transmitted T should be equal to the length of these clips. If a networked computer game is played, T should be determined so that in \normal" congestion situations the player is able to perform some meaningful tasks during the average time the ow stays on. If the network is designed to carry the required trac (i.e., congestion is low), then the average on-time will be a large multiple of T . Under the assumption of a relatively constant level of congestion, the further behavior of PCC is very simple. After a time of T , a ow that is in the on-state has to repeat the random experiment using the same r TCP . How- ever, in a real network the level of congestion is not constant but may change signicantly within a time frame much shorter than T . There are two cases to consider: network conditions may improve (increasing r TCP ) or the congestion may get worse. The rst case is not problematic since it does not endanger the overall performance of the network. PCC ows may be treated unfairly in that they are turned a higher probability than they should be. However, after a time of T the decision will be reevaluated with the correct probability and PCC will adjust to the new level of The second case is much more dangerous to the net- work. In order to prevent unfair treatment of competing adaptive ows or even a congestion collapse, it is very important that PCC ows respond quickly to an increase in congestion. Therefore, PCC continuously updates the value for p and performs further random experiments if necessary. Obviously, it is not acceptable to simply recalculate p without accounting for the fact that the ow could have been turned during one of the previous experiments. any adjustments, PCC would continue to perform the same random experiment again and again and the probability to survive those experiments would drop to 0. The general idea of how to avoid this drop-to-zero behavior is to adjust the rate used in the equations to represent the current expected average data rate of the ow. PCC modies the value r NA , taking into account the last random experiments that have been performed for the ow. To this end, PCC maintains a set P of the probabilities i with which the ow stayed on in the random experiments during the last T seconds. 2 The so-called effective rate r EFF is determined according to the following r NA r NA for For the continuous evaluation and the random experiments r EFF replaces r NA in Equation 1. D. Initialization At the initial startup and after a suspended ow restarts, a receiver does not have a valid estimate of the current condition of the network and thus is not able to instantaneously compute a meaningful TCP-friendly rate. To avoid unstable behavior, a ow will stay in the on-state for at least the protected time T 0 , where T 0 is the amount of time required to get the necessary number of 2 Note that p i the corresponding p 1. measurements to obtain a suciently accurate estimate of the network conditions. After T 0 , PCC determines whether it should cease to send or may continue. In order to take the data transmitted during the protected time into account, the probability of turning the ow increased during the rst interval of T so that the average amount of data transmitted during T 0 +T is equal to that carried by a competing NA denote the average rate of the non- adaptive ow during the protected time and r 0 TCP the average rate a TCP ow would have achieved during the same time. For the adjusted ratio p 0 can be calculated as Again, for 0 p 0 1 we use p 0 as the probability for the random experiment. If the ow is turned o, the application may resume sending after it has been a least T seconds, starting again with the initialization step. 3 If the ow is not turned o, then the ow will stay on for at least T more seconds, provided that the congestion situation of the network does not get worse. Note that it is now possible that p 0 0 if the non- adaptable ow transmits more data during T 0 than a TCP ow would carry during Obviously, in this case cannot be used as a probability for the random exper- iment. Instead, it is necessary to turn the ow o and to increase T , so that p Through the above mechanism the excess data transmitted during the protected time T 0 is distributed over a time span of T . At time T 0 , r 0 but in contrast to r 0 to be updated after T 0 . When a random experiment has to be conducted, it is necessary to calculate not only p 0 but also the corresponding p. Each is included in their respective set P 0 and P . As long as PCC is in the rst T slot and the protected time has to be accounted for, the values in P 0 are used to calculate the eective rate and thus the on-probability. Later on, the set P is used. It may be considered problematic to let a ow send at its full rate for T 0 as this violates the idea of exploring the available bandwidth as is done, e.g., by TCP slow-start. However, requirements level of statistical multi- plexing) and (no synchronization at startup) prevent 3 T can be adjusted by some random oset to prevent synchronization in case several ows with the same value for T were forced to cease sending simultaneously due to heavy congestion. this causing excessive congestion. In addition, the value of usually decrease the more congested the network is, since the actual measurement of the loss event rate makes up most of the time interval T 0 . Loss events become more frequent as congestion increases and therefore the estimate of the network conditions converges faster to the real value. While r TCP is determined, the receiver also calculates the average rate of the non-adaptable ow r NA . 4 Summing up, three important values are determined during initialization: r TCP , r NA , and T 0 . A nite state machine of a PCC receiver is depicted in Figure 2. timeout INIT set timer OFF T' over and p'<x/ set timer over and p'>=x/ ON p'>=x p>=x p<x/ set timer timeout set timer p'<x/ First T Fig. 2. Finite State Machine of a PCC Receiver The runtime of the timer used in this state machine is always T . F. FEC Since applications generating non-adaptable ows frequently have to obey real-time constraints, they benet from forward error correction to compensate for packet loss. However, packet loss typically signals congestion. Therefore it has long been considered unacceptable to compensate for congestion-based packet loss by increasing the data rate of a ow with redundant information for forward error correction. PCC supports the use of forward error correction in a straightforward fashion: When an application decides to employ forward error correction, the new r NA is simply set to the rate of the ow including the forward correction information. From the perspective of PCC this is equivalent to an application increasing its sending rate and thus needs no special treatment. Increasing r NA results in an 4 In our implementation, we use an exponentially weighted moving average of past PCC rates, but as noted in requirement R3, other options are possible. appropriate decrease of p and is therefore fair towards competing ows. G. Example of PCC Operation To provide a better understanding of the behavior of PCC, let us demonstrate how PCC operates by means of an example. As depicted in Figure 3, the sender starts transmitting at the rate determined by the application. After seconds the receiver arrives at an initial estimate of r Furthermore, let us assume that the application developer decided that seconds is a good value for the given application. Now p can be calculated as: s 100 KBit s The value of p is included in the set P and p 0 is calculated since we are in the rst T interval and have to make up for the data transmitted during the protected time. 10s (100 KBit s s 50s 100 KBit s rate time r TCP Fig. 3. Example of PCC Operation Now a random number is drawn from the interval (0; 1], deciding whether the ow will stay on or be turned o. Given a high level of statistical multiplexing, this will result in roughly 1 out of 4 PCC ows being turned o, with the aggregation of the remaining PCC ows using a fair, TCP-friendly share of the bandwidth. Let us assume that the random number drawn is smaller than p 0 and that the ow will stay in the on- state. As depicted in Figure 3, at some later point in time the bandwidth required by the application increases to r 200KBit=s. A new value for p is then calculated as follows: s 200 KBit s 0:8 This value for p is saved to the set P for later use. The adjusted probability p 0 has to be calculated based on the past value of p 0 . s 200 KBit s 0:76 10s (100 KBit s s 50s 200 KBit s 0:76 Let the random number drawn for this decision be below 0:5 so that the ow remains on. A few seconds after this decision the rate a TCP ow would have under the same conditions drops to r Consequently new values for p and p 0 are calculated: s 200 KBit s 0:8 0:5 s 200 KBit s 0:76 0:5 10s (100 KBit s 50s 200 KBit s 0:76 0:5 Again the value p is stored in P while the random number drawn is below p 0 . At rst, the data transmitted during the protected time need no longer be accounted for since PCC has made up for that during the past T interval. Therefore p 0 is no longer calculated. Second, the rst value within P times out and is removed from the set. If the network situation has not changed this will result in the following new value for p: s 200 KBit s 0:5 0:5 This time let the random number be larger than p. As a result the ow is suspended for the next T interval before it may start again with a protected time. It should be noted that this example was designed to demonstrate how PCC works. In reality, a situation where the rate of the non-adaptable ow is ve times the TCP-friendly rate indicates that the network resources are not sucient for this application. Extensions While the current version of PCC works as described above, there are a number of options and possible improvements that we have investigated. In the following we outline two modications that have not yet been incorporated into PCC. H.1 Probe While O ows on average may receive less bandwidth than competing TCP ows, since a ow that has been turned may resume only after a time of T , even if network conditions improve beforehand. This degrades PCC's perfor- mance, particularly if T is large. In order to improve average PCC throughput, ows that are o could monitor network congestion by sending probe packets at a very low data rate from the sender to the receiver. The data rate r OFF produced by the probe packets needs to be taken into account in the Equations 1 and 3 by including an additional factor (1 p) r OFF T . If the loss rate and the round-trip time of the probe packets signal that r TCP has improved, a ow that has been turned o may be turned on again immediately, without waiting for the remainder of the T to pass, and without performing an initialization step. This may be done only if, under the new network conditions, all experiments within the last T interval had been successful. If the congestion situation worsens later on, it must be checked whether any of the experiments during the last T interval had failed. If this is the case, the ow must be turned Only after the last entry in set P has timed out may the ow resume normal operation. For Probe While O to work correctly, it is of major importance that the estimate of the network parameters work independent of the data rate PCC is sending at. The current version of PCC does not include Probe While O, since it could lead to frequent changes between the states \on" and \o", which is likely to be distracting to the user of the application. Furthermore, probe packets waste bandwidth. Probe While O may be included in a later version of PCC as an option for the application. The mechanism can be improved by including a threshold, so that the ow is turned on again only if the available bandwidth increases signicantly. With this improvement, the number of state changes is reduced to improve stability. H.2 Probe Before On In PCC, a ow is turned on upon initialization. This has two drawbacks. First, it violates the idea of exploring the available bandwidth as in TCP slowstart. Second, the ow may be turned immediately after the initialization is complete, so that the user perceives only a brief moment where the application seems to work, before it is turned o. An alternative would be to send probe packets at an increasing rate before deciding whether or not to turn on the ow. Only after the parameters have been estimated and the random experiment has succeeded will real data for the ow be transmitted. The drawback to this method is that bandwidth is wasted by probe packets and that the initial startup of a ow is delayed. H.3 Loss Rate Monitoring ows do not take into account the impact of their actions on the network conditions. Assume that the random experiments of a number of PCC ows fail due to increased congestion, but that the congestion was largely caused by these PCC ows. Then too many ows will be suspended since it is impossible to include the expected improvement in the network conditions in the calculation of the on-probability. Similarly, when the bandwidth consumed by PCC ows during the protected time is a sig- nicant fraction of the bottleneck link bandwidth, severe congestion may be inevitable. Even after the protected time, the changes in network conditions caused by PCC ows that consume a large fraction of the bandwidth are undesirable. For these reasons it is vital that the condition of a sufcient level of statistical multiplexing holds and that the ows do not consume too large a fraction of the bandwidth of the bottleneck link. By continuously monitoring the packet loss rate (e.g., through probe packets) and correlating it with the on- and o-times of the PCC ow, it is possible to estimate the impact of the ow on the network conditions. If the PCC ow causes very large variations in the loss rate, the ow should be suspended permanently. With this extension it is possible to use PCC in environments where it is unclear whether the condition of a sucient level of statistical multiplexing is fullled. V. Simulations In this section, we use network simulations to analyze PCC's behavior. Simulations are based on the dumbbell topology (Figure since it is sucient to analyze PCC fairness and the results can be compared to those of other congestion control protocols evaluated with it. For the same reason, simulations were carried out with the ns-2 network simulator [10], commonly used to evaluate such protocols. Drop-tail queuing (with a buer size of 50 packets) was employed at the routers. We used the standard TCP implementation of ns for the ows competing with PCC. Receivers Senders Bottleneck Link Fig. 4. Simulation Topology A. TCP-Friendliness A typical example of PCC behavior is shown in Figure 5. For this simulation, 32 PCC ows and ows were run over the same bottleneck link with capacity. At an application sending rate of 750KBit/s, the PCC ows should ideally be in the on-state for two thirds of the time. In this example, T was set to 60s, leading to an expected average on-time of 120s. The graph depicts the throughput of one sample TCP ow and one sample PCC ow, as well as the average through-put of all ows. The starting time of the PCC ows is spread out over the rst 50s to avoid synchronization 20060010000 100 200 300 400 500 600 700 800 900 Throughput Time [s] Fig. 5. PCC and TCP throughput The TCP rate shows the usual oscillations around the fair rate of 500KBit/s. PCC's behavior is nearly perfect, with an average rate that closely matches the fair rate and an on-o ratio of two to one. Naturally, not all of the ows achieve exactly this ratio; some stay on for more, some for less time. B. Intra-Protocol Fairness Usually, it is desirable to evenly distribute the necessary o-times over all PCC ows instead of severely penalizing only few. To examine PCC's intra-protocol fairness, a simulation setup similar to the previous one was used, yet the number of concurrent PCC and TCP ows varied between 2 and 128. The probability density function of the throughput distribution from these simulations is shown in Figure 6. As expected, the throughput range is larger for PCC. The coecient of variation (standard deviation over mean) for PCC throughput is 15% compared to a TCP coecient of variation of only about 3%. This results from the time frame for changes in the states of the PCC ows being 60s instead of a few RTTs for TCP ows. There is a direct tradeo between the parameter T and the intra-protocol fairness. Longer on- times, achieved by a larger T , result at the expense of the ows that are suspended for a longer time, thus decreasing intra-protocol fairness. Taken to the extreme, for very Average Throughput (KBit/s) Fig. 6. Distribution of Flow Throughput large T ows may stay on for the whole duration of the session or are not permitted at all, leading to a type of admission control scheme. C. Responsiveness In addition to inter- and intra-protocol fairness, su- cient responsiveness of a ow to changes in the network conditions is important to ensure acceptable protocol be- havior. TCP adapts almost immediately to an increase in congestion (manifest in the form of packet loss). Through the continuous evaluation at timescales of less than T , as described in Section IV-C, PCC can react nearly as fast as TCP to increased congestion, however, it will react to improved network conditions on a timescale of T . Figure 7 depicts the average throughput of ows, again with parameter T set to 60s, and ows. A rather dynamic network environment was chosen where the loss rate increases abruptly from 2.5% to 5% from time 200s to 300s and from time 400s to 420s.100300500700100 150 200 250 300 350 400 450 500 Throughput Time [s] avg. TCP avg. PCC Fig. 7. Loss Bursts When the loss rate changes at time 200s, PCC does not adapt as fast as TCP but still achieves an overall average rate that is quite close to the TCP rate after only a few seconds. seconds later we can see a little spike in the average PCC rate, resulting from the PCC ows that reenter the protected time to probe for bandwidth once their o-time is over. Since the loss rate is still high, the average PCC rate settles at the appropriate TCP-friendly rate shortly thereafter. As soon as the loss rate is reduced to its original value, the probability that sus- pended ows reentering the protected time will immediately be suspended again (and the probability that the random experiment of ows in the on-state will fail) de- creases. Thus, after time 300s, the random experiments of more and more ows succeed until about 50 seconds later the TCP-friendly rate is reached again. Although PCC reacts more slowly than TCP, the average throughput of TCP and PCC up to time 350s is very similar. In contrast to long high-loss periods, short loss spikes hurt PCC performance much more than TCP performance. When the loss rate increases again at time 400s, suspended PCC ows will stay in the o-state for at least 60s, while the actual congestion persists for only 20s. From the time the congestion ends until the time the PCC ows are allowed to reenter the protected time, TCP throughput is considerably higher than PCC throughput. However, we can also see from the graph that during periods of congestion PCC throughput does not quite drop to the level of TCP throughput but remains slightly higher. In the following we will analyze this eect in more detail. D. PCC Throughput for Dierent Application Sending Rates Ideally, no PCC ows would be suspended as long as the PCC application sending rate is below the TCP-friendly rate. For higher application sending rates the average PCC rate should remain at exactly the fair rate through the use of the random experiments. From Figure 8 we take it that an average PCC rate of exactly the fair rate is not reached when the application sending rate equals the fair rate but for an application sending rate that is about 25% higher. The latter eect can be explained by PCC's susceptibility to dynamic network conditions. TCP's typical sawtooth-like sending rate results in variations in the network conditions which unduly cause suspension of PCC ows. When we compare the average PCC throughput to TCP throughput for high PCC application sending rates, we nd that PCC throughput and thus PCC's aggressiveness continues to increase with the application sending rate once the fair rate has been reached. The eect of increased aggressiveness at higher application sending rates can be attributed to the TCP model used by PCC. As stated in [9], the TCP model is based on the so-called loss event rate. A loss event occurs when one or more packets are lost within a round-trip time, and the loss event rate is consequently dened as the ratio of loss events to the number of packets sent. The denominator of the loss-event rate increases as more and more packets are sent during a round-trip time due to a higher application Average Rate [% of Fair PCC Application Sendrate [% of Fair Rate] Model Fig. 8. Comparison with Estimated TCP-Friendly Rate sending rate. At the same time, the number of loss events does not increase to the same extent since more and more lost packets are aggregated to a single loss event. An in-depth analysis of this eect can be found in [11]. When relating the estimated TCP-friendly rate at dierent application sending rates to the average PCC rate achieved in these simulations, it becomes obvious that PCC's aggressiveness is not caused by PCC's congestion control mechanism but by the dependence of the TCP model on the measurement of the loss event rate at sending rates close to the actual TCP rate (to ensure that for TCP and the TCP model the lost packets that constitute a loss event are the same). In addition to PCC's susceptibility to variations in the network conditions, the dierence between the TCP-friendly rate and the average PCC rate is also caused by taking into account only the rate estimates of ows in the on-state. E. PCC Fairness for Dierent Combinations of Flows Figure 9 shows the average throughput achieved by PCC for dierent combinations of PCC and TCP ows when the fair rate is 500KBit/s and the application sending rate is 750KBit/s. Generally, PCC throughput increases with the number of TCP ows, since the higher the level of statistical multiplexing, the lower the variations in the network conditions that degrade PCC per- formance. This eect is more pronounced, the lower the number PCC ows is. For a more detailed analysis of PCC and further net-work simulations we refer the reader to [12]. VI. Conclusions In this paper we presented a congestion control scheme for non-adaptable ows. This type of ow carries data at a rate determined by the application. It cannot be adapted to the level of congestion in the network in any way other than by suspending the entire ow. Existing41664 Number of PCC flows2832Number of TCP flows250750 Throughput [KBit/s] Fig. 9. Average PCC Throughput for Dierent Numbers of Flows congestion control approaches therefore are not viable for non-adaptable ows. We proposed to perform congestion control for such ows by suspending individual ows in such a way that the aggregation of all non-adaptable ows on a given link behaves in a TCP-friendly manner. The decision about suspending a given ow is made by means of random experiments In a series of simulations we have shown that PCC displays a TCP-friendly behavior under a wide range of network conditions. We identied the conditions under which PCC throughput does not correspond to the TCP-friendly rate. To some extent, these eects on the average PCC sending rate cancel each other out. Nevertheless, the may degrade PCC performance. We intend to include Probe While O as an optional element in PCC, which would improve PCC's behavior in highly dynamic network environments. Furthermore, we are currently investigating a method to perform a more accurate estimate of the fair TCP rate if the loss event rate is measured at a sending rate that diers considerably from the TCP-friendly rate. Finally, we plan to evaluate if and how PCC can complement congestion control for multicast transmissions. --R John Hei- demann --TR Promoting the use of end-to-end congestion control in the Internet Modeling TCP Reno performance pgmcc Equation-based congestion control for unicast applications FLID-DL Advances in Network Simulation Issues in Model-Based Flow Control

TCP-friendliness;non-adaptable flows;congestion control

507677

The design of a transport protocol for on-demand graphical rendering.

In recent years there have been significant advances in 3D scanning technologies that allow the creation of meshes with hundreds of millions of polygons. A number of algorithms have been proposed for the display of such large models. These advances, coupled with the steady growth in the speed of network links, have given rise to a new area of interest, the streaming transmission and visualization of three-dimensional data sets. We present a number of issues that must be addressed when designing a transport protocol for three-dimensional images as well as a method of transport. We then evaluate an implementation of an On-Demand Graphic Transport Protocol (OGP) that addresses these issues.

INTRODUCTION With the enormous increase in computing power of today's home computers, it is now possible for complex three-dimensional models to be used in home applications. The growth of the Internet introduces the possibility of using such three-dimensional models on websites and online applications. The size of the models, with the corresponding long download times, prohibits transferring the models completely before rendering begins; therefore, efficient methods for streaming three-dimensional models must be developed if the models are to be used in online applications. The use of a streaming protocol allows the user to begin viewing a portion of the model before transmission is completed. This type of Work funded in part by the NSF under grant CCR-0086094 Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. May 12-14, 2002, Miami, Florida, USA. streaming is, however, distinct from the problem of video stream- ing. Videos can be loss-tolerant and are linearly ordered. By con- trast, three-dimensional images are not linearly ordered. Given this information, we observe that while video streams have time- dependencies between video segments, three-dimensional images have space-dependencies between picture segments. This leads to radically different methods for optimal transport. A common scheme for rendering three-dimensional models is based on a bounding volume approach. The data structure containing the models is a tree with certain properties, such as loss- tolerance yielded by the partial ordering of the data structure, that can be leveraged to facilitate efficient streaming. The nodes of the tree represent bounding volumes of the model with the roots of the tree representing spheres that encompass the entire model. Nodes lower in the tree represent successively smaller volumes, the leaves being the smallest resolution points. The dependencies in the tree run along the branches. Each node is dependent only on its parent node and there is no ordering among siblings in the tree. This allows a certain flexibility in the reliable transport of the model because losses only affect subtrees rooted at the lost nodes. Another facet of the tree is that greater resolution is achieved by running down the tree, therefore it is possible to achieve a lower resolution version of the model by only streaming down to a certain level within the tree. The ability to send only part of the tree for rendering can be leveraged to control the quality of the model based on the bandwidth available to the user. A further property of the tree is that particular branches in the tree correspond to particular spatial sections of the picture. It is possible to take advantage of this fact and render exclusively, or more completely, the sections of the image that are of interest to the user, perhaps by allowing the user to mouse-over the portion of the model in which they are interested [14]. We therefore contribute an analysis of the issues involved in developing an efficient transport protocol for streaming three-dimensional models and the design of the On-Demand Graphic Transport Protocol (OGP). The standard Internet transport protocols are not appropriate for streaming such models. TCP's [11] full reliability can create unnecessary delays in the stream, but UDP [12] does not provide the rendering application the reliability it requires to maintain the partial ordering [3]. Therefore we present a protocol that maintains the partial ordering required by the rendering applications. We present a good node packing algorithm that can maintain properties such as loss-tolerance constant throughout transport, thereby allowing the transport protocol to continue to exploit these properties. We demonstrate that better performance can be obtained by a transport protocol by taking advantage of these properties. The rest of this paper is organized as follows. In Section 2, we describe approaches to streaming media in general, discussing their limitations. In Section 3, we present our analysis of the tree structured three-dimensional models and the design of an efficient streaming protocol for three-dimensional models. In Section 4, we present the parameters used in designing an efficient protocol for three-dimensional model transport. We then present the implementation of the On-Demand Graphic Transport Protocol (OGP). In Section 5 we present the results of the evaluation of OGP. In Section 6, we present some conclusions about three-dimensional model streaming. 2. STREAMING MEDIA The sheer size of media, such as entire videos and three-dimensional models, leads us to consider new representations that enable intelligent partitioning of the data into smaller application-layer data units [4]. Such partitioning supports the streaming of data, providing the benefits of pipelining data transmission with data pre- sentation. This new type of media is often progressive, with each additional piece of data adding to what has already been recieved. Protocols such as RTP [15] make use of application level framing. Many such protocols are integrated directly into the application, which would then sit on top of UDP. This allows the applications to be run without changes to the current Internet technologies. This research in streaming media has mainly focused on video as the media of choice. RTP, for example, is tailored to handle real-time requirements that are inherent in video. The streaming of three-dimensional models introduces different requirements for the transport of model data based on the structure of the data. 2.1 Video Since a video is the progressive representation of images, streaming video is naturally represented by a sequence of frames or groups of frames for optimization. These frames represent a fully linear se- quence. Loss cannot be tolerated within a frame, but each of these frames is an independent entity, enabling streaming video protocols to tolerate the loss of some frames. Essentially, these protocols are memory-less. For example, in MPEG encoding, the I-, P-, and B-frames must be in order to be usable but the stream can tolerate the loss of a B-frame affecting only the subsequent B-frames until the reception of the next I-frame [6]. There are also strict time considerations with video. Once the play-out time for the frame following a lost frame expires, it is too late to do anything with that lost frame. Such partitioning of video divides it into application data units that have temporal locality. Each group of pictures is related to the others in time, not space. The usefulness of designing image models that allow losses in particular parts of the image, thereby allowing the continued transfer and display of the rest of the image, with the lost piece appearing as a small blank has been shown to help intra-frame loss [16]. This can be seen as a motivating step toward the structuring of models using spatial partitioning as opposed to temporal partitioning. Three-dimensional modeling uses this concept extensively. 2.2 Three-Dimensional Models The design goal of three-dimensional models is to provide a representation of an object for presentation to the user. In order to provide a complete representation, all data must be available for presentation. In order to support more flexible presentation of data, these models have been designed in a progressive manner; the more of the data that is available, the better the quality of the presenta- tion. The three-dimensional models are divided into application data units that have spatial locality. Each node is related to the others with respect to their relative positions in the actual model. Such models inherently require memory of all available data. Model representations based on the bounding volume approach have a tree-based structure. This structure provides a partial ordering on the data. Each node that refines a particular spacial location on the model is dependent on all previous nodes that involve that location. If a node from one part of the model is lost, this does not affect other sections of the model. In this way, the data is non-linear. The partial ordering can be leveraged to make significant performance gains in the face of loss [2, 3, 8]. It is possible to maintain the stream efficiently by carefully keeping track of the dependencies existing between parts of the data structure [9]. When a loss is encounter in a model, any received children of the lost node cannot be rendered. If the rendering is on-demand, then the time delay in rendering caused by the transmission of nodes that cannot be immediately rendered will be perceived. In order to avoid this, nodes that are dependent on the loss should not be transmitted until the lost node is re-transmitted and received. This implies that rapid loss detection is essential to efficiently stream models for on-demand rendering. To this end, there has been some work in finding effective ways to detect losses early. Papadopoulos, et. al. [10] use gap-based loss detection as opposed to timer-based loss detection. This minimizes the latency for loss detection. While the time constraints for three-dimensional models are different than for video, it will still prove to be very useful to minimize the time it takes to detect a loss. By detecting losses early and transmitting only sections of the data structure that can be rendered, the door is opened for the possibility of not re-transmitting lost data immediately if it is not desired. This amounts to varying the level of reliability for the transfer of each packet. Varying the reliability levels dynamically during transmission of data has been shown to be useful in creating better efficiency in the transfer of data that can sustain some loss [7]. One approach to streaming three-dimensional models is to transmit data progressively with lower resolution data being transmitted first and then streaming higher-resolution data to fill in the details. One approach [13] uses the HTTP/1.0 protocol to request data from a standard web server, such as Apache, that would send requested bits of data to the client. We believe this approach is not optimal because it does not make use of the properties of the data structure. Since node size will not always be the same as packet size, it is important to consider how to pack the models into packets. The problem is essentially one of linearization. The tree in which the three-dimensional models is not linear, it is partially ordered, but transmission over the network requires linearization. If the data is not intelligently linearized, we can end up transmitting data that can not be used. 3. APPLICATION Today, the transport of data across the Internet is achieved primarily through the use of two transport protocols, TCP and UDP. TCP provides a fully reliable means of communication that guarantees in-order delivery of all packets sent without duplication. UDP, on the other hand, provides no reliability guarantees at all, packets may come out of order or even not at all. With the widespread use of these two transport protocols, it is tempting to take a very binary, one-dimensional view of transport service. The tendency is to either see the data transmission as reliable and fully ordered or not reliable and unordered. The main problem is that this cuts off a large portion of the possible space. Really, transport service should be seen as a two dimensional continuous function involving guarantee of delivery and ordering as in Figure 1 [2, 3, 7, 8]. UDP Ordering Reliability Full None 100% Partial Ordering Partial Reliability Figure 1: Transport Service Space 3.1 Exploring the Service Space Once the transport service space is expanded, the impact of using the incorrect transport service level can be analyzed. There are three possible relationships between the transport service level used in the transmission of data and the actual need of the application requiring the transmission. 1. The level of transport service is optimal for the application. 2. The level of transport service is too high for the application. 3. The level of transport service is too low for the application. For the second two relationships, there are two different aspects in which the transport service level could be seen as too high or low for an application. 1. The ordering requirements can be too strict or lax. 2. The loss tolerance can be too strict or lax. The various combinations of possibilities are derived directly from the expanded transport service space. The effects of the situations, however, must be discussed systematically. 3.2 Mapping Channel Service to Application Requirements If an optimal transport service level is chosen for the transport of an application's data, it is clear that there will be no wasted resources providing services that the application does not require. It is also clear that the application will be able to function since all of its service requirements would be met by the optimal matching. The interesting cases to study are the sub-optimal cases. If there is no significant impact on performance when a sub-optimal choice in transport service level is chosen, there would be no reason to consider alternative transport mechanisms to TCP. If the reliability level required by the application is not met, it will be impossible for the application to produce correct results. The actual effects on the application are scenario specific, but can range from total failure to incorrect functionality [9]. Similar effects can be seen from an inappropriate ordering constraint. If the application's ordering requirements are more strict than the transport layer can provide, then again the application will fail to function correctly. For three-dimensional modeling applications, it is not possible for the rendering engines to deal with loss or complete mis-ordering, therefore it is clear that UDP will not provide a sufficient level of transport service. Dependencies run down the branches Figure 2: Partially Ordered Tree If the transport service level required by the application is surpassed by the network, unnecessary delays and a loss of goodput is experienced. A similar result occurs when the ordering constraints of the network are too strict. If this happens, it is possible that some data will be ready to be sent but will have to wait due to false ordering constraints. In this way, both throughput and goodput will suffer [9]. 3.3 Maintaining Application Data Properties While there has been research into designing protocols that can handle partial orderings [2, 3, 8], the protocol to handle streaming bounding-sphere based three-dimensional models must be specially tailored. Current partial order protocols address the ordering of packets, but not how data will be organized into packets. Given that a node in a three-dimensional model may not fill an MTU, it may be necessary to pack multiple nodes in one packet. Such packing can greatly improve the performance and efficiency of the transmission of the model, but must be accomplished within the constraints of the original ordering. Also, due to the very specific nature of the tree and its partial ordering, a specialized transport scheme can enable on-demand improvement of specific, user-chosen parts of the model. We now analyze the tree used to store the models in bounding sphere methods of rendering. The only ordering constraint is that for any child received, its parent must have been received first. Es- sentially, there must be complete ordering only down the legs of the tree, as in Figure 2. The nodes of the tree represent the smallest data units that can be drawn, with each node relying on its parent. The leaves of the tree represent the individual pixels fully rendered. The nodes of the tree can vary between a few bytes and a few hundred bytes. Because of this small size, more than one node can be put in each packet. A decision must be made on the order in which to pack the nodes together. The main constraint is that transmission must not continue too far down any path in which one of the parent nodes has been lost or not yet been acknowledged. This problem can be avoided by only allowing new nodes to be sent after an acknowledgment has been received for the node's parent. Such information can be integrated into the flow control mechanism as discussed in Section 4.1. Node packing is very important in order to maintain the partial ordering throughout transmission of the model. We note that any subtree of the tree will have the same data structure properties as the whole tree. Therefore, in order to be able to leverage the properties that have been discussed throughout transport, nodes must be packed in such a way as to not break the subtree relationships across packets. Packing the nodes of the tree in a subtree-by-subtree order will effectively maintain the partial ordering of the tree throughout transmission. The details of node packing are discussed in Section 4.1. 4. ANON-DEMANDGRAPHICTRANSPORT There are fundamental design decisions that must be made in order to optimize a protocol for the transport of three-dimensional models. The inherent loss tolerance and partial ordering of the tree should be leveraged in order to efficiently stream the models. The models can sustain no loss along any branch. However, due to the fact that there are no dependencies between siblings in the tree, if a loss occurs, streaming of other branches in the tree may continue without interruption. The inherent loss tolerance of the data structure can be leveraged to allow the transfer of data to continue in the face of loss. The partial ordering of the data structure can be leveraged to allow the protocol to continue to transfer the data smoothly in the face of lost or reordered packets. The partial ordering can also be used to allow the protocol to take into account user focus information and to transfer only the part of the model that is currently in focus. The protocol must also contain some method of deciding when it is appropriate to retransmit lost packets and which packets to retransmit. Finally the protocol must package the nodes in such a way as to maintain the properties of the data structure throughout the transfer. In the following subsections, we present our implementation of a three-dimensional model streaming pro- tocol, OGP. We use OGP to illustrate how these issues in protocol design can be resolved. OGP is a self-clocked protocol that makes use of TCP congestion control algorithms [1]. Self clocking is achieved by allowing the receipt of an acknowledgment to trigger the transmission of new packets. In the current version of OGP, every packet is ac- knowledged, therefore, the loss of an acknowledgment implies the loss of a packet. This problem could be fixed by using selective- acknowledgments instead of single-packet acknowledgments. Congestion control is managed through the use of a window mechanism that monitors the number of outstanding packets, as discussed in Section 4.2. Because it is possible for multiple nodes to be sent in one packet, the sender maintains a mapping between nodes and packets. In this way, the receiver can acknowledge packets and the sender can then map these acknowledgments onto acknowledgments of nodes, which are used by the packing algorithm as described in Section 4.1. 4.1 Node Packing In order to take advantage of the properties of the tree, the partially ordered data structure must be linearized and sent across the network while maintaining the inherent properties. In order to do this, the partial ordering relations of the tree, as well as the spacial relations of the tree nodes, should be used. OGP begins by packing the largest possible subtree at the root. After this first packet has been acknowledged each node sent in that packet can now be used as a root of a new subtree to send. OGP continues by packing the children of the acknowledged nodes as roots of largest subtrees in breadth-first order [Figure 3]. Essentially, the largest possible next subtree is being packed in each successive packet. In the event of a loss, OGP avoids sending any of the children dependent on the lost nodes until the lost nodes have been successfully re-transmitted. The goal is to maintain a subtree-by-subtree transmission pattern in order to keep the partial ordering and reliability properties of the entire model constant throughout transfer. In order to keep track of what nodes have been sent, markers are added into the data structure that can be set as sent and ackd. As the tree is traversed breadth-first while packing nodes, the nodes are marked as sent. When a packet is acknowledged, each node contained in the packet is marked as ackd. Pointers are kept to the first node in the tree that is not acknowledged and the current node needing to be sent. In this way nodes needing to be marked and nodes needing to be re-transmitted can be found efficiently while still walking forward in the tree. The packing algorithm works as follows. Get Next Node takes the current sub tree as the root of a tree and increments node in breadth-first order using a standard tree walking algorithm. Get Next Subtree increments current sub tree to the next unmarked node in the tree using a standard breadth-first tree walking algorithm. while(node != end_of_tree) { { To illustrate the algorithm, consider an example of a three-dimensional model with node sizes of 150 bytes. The tree has a branching of 4. Consider a link over the Internet between a client and server with a delay of 20ms and a bandwidth of 256Kbits. The Ethernet frame size is 1500bytes. 10 nodes fit per packet and there can be at most 5 packets in flight between the last acknowledged and the last sent packets. Given this information, if the first nodes are packed in breadth-first order in the first packet, there will be 15 nodes that can be rendered but have not yet been sent. If nodes are continued to be packed breadth-first, there are possible cases where, after a loss, data will still be sent that cannot be rendered because the loss will not yet be noticed on the sender side. This problem will be exacerbated by smaller nodes sizes, such as Qsplat's 4 byte node sizes [14]. If however, the node packing is done intelligently, the branching factor of the tree can be leveraged to prevent the possibility of transmitting data that cannot be immediately rendered, see Figure 3. 4.2 Congestion Control The loss tolerant nature of the data structure allows OGP to continue transmitting data in the face of losses. The only invariant that must be maintained is that no node whose parent has not yet been received should be sent. The packing algorithm in Section 4.1 insures that the ordering properties of the tree will be maintained throughout transmission. A flow control algorithm is needed, how- ever, to insure that nodes are always available whose parents have already been acknowledged when it is time to send. OGP should also react to congestion appropriately so that it is TCP friendly. First Packet Further packets group subtrees moving across tree Figure 3: Intelligent Packing Order For congestion control, OGP reacts much the same way as TCP New-Reno, making use of the slow-start, congestion avoidance, fast recovery, and fast retransmit algorithms common to TCP [1]. Changes have been made to account for the fact that it does not need to actually retransmit a lost packet. OGP does require losses to be identified by the sender, however, so that nodes can be marked cor- rectly. OGP uses a TCP like sequence numbering scheme. The receiver always acknowledges the sequence number received and not the sequence number of the next packet expected. This is necessary because packets are not retransmitted so a lost sequence number will never be received. OGP has a congestion window, cwnd. The cwnd is started at 1 and is increase according to the TCP New-Reno algorithms. According to these algorithms, cwnd is never increased by more than one at any received acknowledgment. The cwnd is reduced due to loss according to the TCP New-Reno algorithms. A loss is detected when the acknowledgment received at the sender is not the next acknowledgment expected. In order to relate a packet loss to a node loss, the sender maintains a mapping between packets and nodes that are in flight. When a loss occurs, the sender refers to this mapping to decide which nodes were lost in the lost packet. The nodes in the lost packet are then marked unsent. It is possible that packets that are reordered will be considered lost. This is not a real problem, however, because OGP does not immediately retransmit lost packets, therefore when the packet is eventually received and acknowledged, the nodes will be marked ackd. OGP also has a time-out mechanism to allow it to recover in the face of losses of acknowledgments like TCP New-Reno. Because packets may be lost and never retransmitted, a standard sliding window, such as the one TCP uses, cannot be used. Instead only the number of packets still in flight need be considered, which can be calculated based on the last packet acknowledged and the last packet sent. This number is compared with the value of cwnd. When the number of packets in flight is less than the congestion window, OGP can send. Each time an acknowledgment is received, OGP will send as many packets as it can. An important issue is to make sure that there are always nodes whose parents have been acknowledged with which to build the next subtree to send in a packet. Consider that a packet is only sent after an acknowledgment is received. Also consider that the congestion window is never opened by more than 1. Then, in the face of no losses, there will never be more than 2 packets sent per acknowledged packet. Therefore, as long as the data structure has a branching factor of 2 and there is at least 1 node per packet, there will always be nodes to send. In the face of losses, it is possible that more than 2 nodes may be sent after an acknowledgment. How- ever, the congestion window is halved when a loss is encountered for each round-trip-time. Therefore, there will not be more packets to send until a number of new acknowledgments have arrived. Therefore, either the branching factor of the structure must be at least four, or there should be at least three nodes per packet. The largest node size that we have observed in the three-dimensional modeling schemes using bounding volumes is 300 bytes and the branching factors are typically 4. It is not unreasonable to assume that 4 nodes fit per packet, giving the needed 4 nodes to send per acknowledgment even assuming only a branching factor of 2. 4.3 Exploiting the Partial Ordering The partial ordering of the data structure combined with the node packing algorithm allows OGP to insure that the packets in flight are not dependent on one another. Because of this, OGP does not need a standard congestion-window and can focus on the number of nodes in flight. Therefore, OGP can continue to send in the face of packet reordering and loss. The only reason OGP should slow transmission is in response to congestion. Therefore, OGP should see more throughput than TCP. This result is achieved as shown in Section 5. The partial ordering of the data is also used to allow the protocol to send data that is relevant to the user's focus. Each node in the tree represents a particular bounding volume that covers a particular section of the model. The user's focus can be given to the protocol as the identity of the node that covers the section of interest. It is then possible, due to the partial ordering of the data, to only send the line of nodes down the branch that leads to the node covering the area of the model on which the user is focused. The subtree of the model rooted at this user focus defined node can then be sent using the normal transport algorithm. 4.4 Deciding What to Transmit Aside from sending without the retransmission of lost packets as described so far in the paper, the protocol also has the ability to use some its throughput to retransmit some or all of the lost data. We implemented full-reliability by marking lost nodes as lost in the data structure and first packing those nodes into packets to be sent. While these nodes will not normally reveal new nodes that can be sent, this is not a problem for the protocol as there were nodes that would have been sent that were not, had the protocol been functioning with no reliability. Another parameter to consider in transmission is user demand. The reliability level of a certain part of the tree can be changed depending on user focus. If there is a particular part of the model that the user is interested in, then OGP can immediately re-transmit any lost nodes from the corresponding section of the data structure. This mechanism works the same way that full-reliability does, except that only nodes in the part of the model that are of interest to the user are marked lost and are therefore retransmitted. User focus could also be used to partition out one part of the model to be filled in faster than the other parts, using all available bandwidth to concentrate on the refinement of relevant parts of the model if this is desired. Again, the part of the model that is of interest to the user can be viewed as a subtree of the data structure. Therefore, OGP begins by packing nodes in depth-first order to the point where the relevant data begins, and then uses the standard packing algorithm. We implemented this functionality by assuming that the user focus is represented by the node that surrounds the volume of the model that is of interest to the user. This node is then considered as the root of the subtree that should be sent. 5. EVALUATION In evaluating OGP, we found it useful to consider data-specific properties of the protocol, namely, node size and branching fac- tor. We also evaluated the performance of the protocol in terms of throughput and fairness to TCP. OGP New-Reno New-Reno New-Reno Queue Figure 4: Simulation Set-Up Packets Received Packets Dropped OGP 3898 135 Table 1: Simulation Results: queue 5.1 Simulation Set-Up In the simulations we used the ns2 [5] network simulator. The simulation layout we used is depicted in Figure 4. For the simulation runs, we used a node size of 200bytes and a branching factor of 4. The bottleneck queue size was varied between 5 and 40 packets. For the simulation runs, we had 3 TCP New-Reno flows and one OGP flow. The link speeds are given in Figure 4. 5.2 Data-Specific Evaluation We first evaluated the effect of node size on OGP. As mentioned in Section 4.1, as node sizes decrease and more nodes are included in each packet, a breadth-first packing algorithm will encounter more occurrences of packets being sent that cannot be rendered. however does not have this problem as it guarantees in-order, fully-reliable transport. We found however, that smaller node sizes merely increase the number of nodes that may be used as next sub-tree roots for each packet that is acknowledged. Therefore small node sizes have no real effect on OGP. We notice similar results from larger branching factors. We do not show these results here as they were all roughly the same. In our simulations, we found that OGP always achieved 100% goodput, meaning that all packets received can be rendered. TCP also achieves 100% goodput due to its full reliability. One interesting point to notice, however, is that the packets sent by OGP can be immediately rendered, there is never a need to wait for retransmissions or re-ordering. 5.3 Performance Evaluation The two main results that we wanted to evaluate were TCP fairness and good throughput. Our simulations show that these goals have been met. As the length of the simulations as well as the queue size on the bottleneck link were varied, we found that OGP5001500250035000 10 20 "OGP" Figure 5: Simulation Results: queue achieved very similar results to itself. The results for simulations with a queue size of 10 are displayed in Table 1 and Figure 5. OGP reacts quicker to congestion, and therefore has fewer packets in flight during a congestion period. This is due to the fact that OGP detects losses not by the number of duplicate acknowledgments but instead by the difference between the current acknowledgment and the expected acknowledgment, i.e. gap detection. This rapid detection of losses allows OGP to cut its congestion window more rapidly than TCP. OGP, however, gets more throughput due to the fact that it does not worry about re-ordered packets and it does not wait to re-send lost packets before opening its flow control window. TCP New-Reno and OGP did reach an equilibrium point where all flows were getting bandwidth. OGP achieves approximately 26% better throughput however, due to the mechanisms already mentioned. 6. CONCLUSION Given the greater processing power and better graphics cards in desktop computers, it is now possible for very complex three-dimensional models to be used in everyday applications. This leads to the desire to send such models across the network. Due to the potentially large size of the models, it is necessary to find a way to stream them efficiently over the network. In order to do this, the properties of the data structures used to store three-dimensional models must be exploited. The tree that is used in three-dimensional models has the property of being partially ordered, i.e. not completely linear. This led us to analyze the problems of having too narrow a view of the transport service space. Once we see the space as expanded, we can understand the problems caused by having too much reliability provided by the network. We therefore show which parameters must be considered in designing an efficient protocol for the transmission of on-demand three-dimensional models. We discussed the approach to linearize the non-linear partially ordered tree. We then exploit the approaches to design the On-Demand Graphic Transport Protocol (OGP). OGP is shown to gain performance benefits over TCP, making it clear that the inherent properties of the three-dimensional modeling data structures should be used to develop efficient transport mechanisms. Acknowledgments We would like to thank Mike Garland for his help in understanding three-dimensional modeling. We would also like to thank Prashant for his help in developing the packing methods used in the protocol. We would also like to thank the entire Mobius Group for their continued help and support. 7. --R Tcp congestion control. Architectural considerations for a new generation of protocols. Ns notes and documentation. Adaptive variation of reliability. Optimizing partially ordered transport services for multimedia applications. Transmission control protocol. User datagram protocol. Streaming qsplat: A viewer for networked visualization of large Qsplat: a multiresolution point rendering system for large meshes. A transport protocol for real-time applications Image transfer: An end-to-end design --TR Architectural considerations for a new generation of protocols Image transfer Partial-order transport service for multimedia and other applications Adaptive variation of reliability QSplat Streaming QSplat: a viewer for networked visualization of large, dense models

streaming;3-D models;partial order;partial reliability;transport protocol

507680

Impact of link failures on VoIP performance.

We use active and passive traffic measurements to identify the issues involved in the deployment of a voice service over a tier-1 IP backbone network. Our findings indicate that no specific handling of voice packets (i.e. QoS differentiation) is needed in the current backbone but new protocols and mechanisms need to be introduced to provide a better protection against link failures. We discover that link failures may be followed by long periods of routing instability, during which packets can be dropped because forwarded along invalid paths. We also identify the need for a new family of quality of service mechanisms based on fast protection of traffic and high availability of the service rather than performance in terms of delay and loss.

INTRODUCTION Recently, tier-1 Internet Service Providers (ISPs) have shown an ever increasing interest in providing voice and telephone services over their current Internet infrastructures. Voice-over-IP (VoIP) appears to be a very cost effective solution to provide alternative services to the traditional telephone networks. However, ISPs need to provide a comparable quality both in terms of voice quality and availability of the service. We can identify three major causes of potential degradation of performance for telephone services over the Internet: network congestion, link failures and routing instabilities. Our goal is to study the frequency of these events and to assess their impact on VoIP performance. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. NOSSDAV'02, May 12-14, 2002, Miami, Florida, USA. We use passive monitoring of backbone links to evaluate the occurrence and impact of network congestion on data traffic. Passive measurements carried over different locations in the U.S. Sprint IP backbone allow us to study the transmission delay of voice packets and to evaluate the degree of congestion. However, this kind of measurement cannot provide any information related to link failures or routing instabilities. For this purpose, we have deployed an active measurement infrastructure in two locations well connected to the backbone. We capture and timestamp the probe packets at both ends to quantify losses and observe the impact of route changes on the voice traffic. We performed many week-long experiments in order to observe different link failure scenarios. Given that all our measurements take place in the same Autonomous System (AS) we also complement our data with IS-IS routing information [9] collected in one of the backbone Points of Presence (POPs). This additional information give us a fairly complete view of the events that occur during our experiments. Indeed, active probes and routing information give us the capability of identifying precisely the links, the routers and even the interfaces that are responsible for failures or instabilities in the network. Our findings indicate that the Sprint IP backbone network is ready to provide a toll-quality voice service. The level of congestion in the backbone is always negligible and has no impact on the voice quality. On the other hand, link failures can impact the availability of VoIP services. We discovered that link failures may be followed by long periods of routing instability, during which packets can be dropped because forwarded along invalid paths. Such instabilities can last for tens of minutes resulting in the loss of reachability of a large set of end-hosts. The paper is structured as follows. Section 2 briefly presents some related work, while Section 3 provides detailed information on the measurement approaches followed in this study. Section 4 describes the model used to assess the subjective quality of voice calls from transport level measurable quantities. In Section 5 we finally discuss our findings, while Section 6 presents some concluding remarks. 2. RELATED WORK Past literature on end-to-end Internet measurements has often focused on the study of network loss patterns and delay characteristics [6, 8, 16, 26, 24]. For example, Kostas [18] studied the feasibility of real-time voice over the Internet and discussed measured delay and loss characteristics. In order to evaluate the quality of Internet Telephony, [14] provided network performance data (in terms of delay and losses) collected from a wide range of geographically distributed sites. All these studies were based on round-trip delay measurements. While information about delay and losses can give valuable insights about the quality of VoIP, they do not characterize the actual subjective quality experienced by VoIP users. In [11], Cole et al. propose a method for monitoring the quality of VoIP applications based upon a reduction of the E-model [3] to measurable transport level quantities (such as delay and losses). Markopoulou et al. [19] use subjective quality measures (also based on the E-model) to assess the ability of Internet backbones to support voice communications. That work uses a collection of GPS synchronized packet traces. Their results indicate that some backbones are able to provide toll quality VoIP, today. In addition, they report that even good paths exhibit occasional long loss periods that could be attributed to routing changes. However, they do not investigate the causes of network failures neither the impact they have on the voice traffic. 3. MEASUREMENTS In this section we describe the two measurement approaches used in our study, i.e. the passive measurement system deployed in the Sprint IP backbone network and the active measurement system that uses probe packets to study routing protocols stability and link failures. 3.1 Passive measurements The infrastructure developed to monitor the Sprint IP backbone consists of passive monitoring systems that collect packet traces on more than links located in three POPs of the network. Details on the passive monitoring infrastructure can be found in [13]. In this study, we use traces collected from various OC-12 intra- POP links on July 24th, 2001, September 5th, 2001 and November 8th, 2001. A packet trace contains the first 44 bytes of every IP packet that traverses the monitored link. Every packet record is also timestamped using a GPS reference signal to synchronize timing information on different systems [20]. We use the technique described in [23] to compute one-way delays across the Sprint backbone. The basic idea behind that technique is to identify those packets that enter the Sprint backbone in one of the monitored POPs and leave the network in another one. Once such packets are identified computing the delays simply requires to compute the difference between the recorded timestamps. 3.2 Active measurements Passive measurements provide valuable information about net-work characteristics, but the data collected depend on the traffic generated by other parties, which is completely out of our control. Moreover, given that we do not monitor all the links of the back-bone network, we are not able to measure jitter or loss rates through simple passive monitoring (packets may leave the network through not monitored links) [23]. Therefore, our passive measurements alone cannot provide results on the quality of the voice calls. These are the motivations behind the use of active measurements to complement the passive ones. In an active measurement environment we can perfectly control the amount and the characteristics of the traffic that we inject in the network and thus draw precise conclusions about the impact of the network on the monitored traffic. 3.2.1 Measurement infrastructure We deployed active measurement systems in two locations of the U.S. (Reston, VA and San Francisco, CA) well connected to the Sprint backbone, i.e. just one router away from the backbone network. Figure 1 shows the architecture of the testbed and the way the sites are connected through the Sprint network (the thick lines indicate the path followed by our traffic). Note that each access router in a POP is connected to two backbone routers for reliability and, usually, per-destination prefix load balancing is implemented. The access links to the backbone were chosen to be unloaded in order not to introduce additional delay. At the end of each experiment we verified that no packet losses were induced on the last hops of the paths. In each site, four systems running FreeBSD generate a traffic made of 200 byte UDP packets at a constant rate of 50 packets per second. We choose this rate so that the probes could be easily used to emulate a voice call compliant to the G.711 standard [2]. An additional system captures and timestamps the probe packets using a DAG3.2e card [10]. The DAG cards provide very accurate timestamping of packets synchronized using a GPS (or CDMA) receiver [20]. The probe packets are recorded and timestamped right before the access links of the two locations in both directions. In the experiment we discuss here, probes are sent from Reston (VA) to San Francisco (CA) for a duration of 2.5 days starting at 04.00 UTC on November 27th, 2001. We have run active measurements for several weeks but we have chosen that specific trace because it exhibits an interesting network failure event. In terms of delay, loss and voice call quality we have not measured, instead, any significant difference among the many different experiments. 3.2.2 Routing data We integrate our measurement data with IS-IS routing information collected in POP#2 (see Figure 1). We use an IS-IS listener [21] to record all routing messages exchanged during the ex- periment. IS-IS messages permit to correlate loss and delay events to changes in the routing information. In order to illustrate the kind of data that are collected by the listener, we give a brief description of the IS-IS protocol. IS-IS [22] is a link state routing protocol used for intra-domain routing. With IS-IS, each link in the network is assigned a metric value (weight). Every router 1 broadcasts information about its direct connectivity to other routers. This information is conveyed in messages called Link State PDUs (LSP). Each LSP contains information about the identity and the metric value of the adjacencies of the router that originated the LSP. In general, a router generates and transmits its LSPs periodically, but LSPs are also generated whenever the network topology changes (e.g. when a link or a router goes up or down). Thus, LSPs provide valuable information about the occurrence of events such as loss of connectivity, route changes, etc. Once a router has received path information from all other routers, it constructs its forwarding database using Dijkstra's Shortest Path First (SPF) algorithm to determine the best route to each destina- tion. This operation is called the decision process. In some transitory conditions (e.g. after rebooting), the decision process can take a considerable amount of time (several minutes) since it requires all the LSPs to be received in order to complete. During that transitory period, a router is responsible to make sure that other routers in the network do not forward packets towards itself. In order to do so, a router will generate and flood its own LSPs with the "Infinite Hippity Cost" bit set 2 . This way, other routers will not consider it as a valid node in the forwarding paths. 1 IS-IS has been designed within the ISO-OSI standardization effort using the OSI terminology. In this paper, we have instead decided to avoid the use of OSI terms. This bit is also referred to as the OverLoad (OL) bit. Figure 1: Topology of the active measurement systems (the thick lines indicate the primary path) 4. VOICE CALL RATING Even though active measurements may provide accurate information on network delay and losses, such statistics are not always appropriate to infer the quality of voice calls. In addition to mea- surements, we use a methodology to emulate voice calls from our packet traces and assess their quality using the E-model standard [3, 4, 5]. 4.1 A voice quality measure: the E-model The E-model predicts the subjective quality that will be experienced by an average listener combining the impairment caused by transmission parameters (such as loss and delay) into a single rat- ing. The rating can then be used to predict subjective user reactions, such as the Mean Opinion Score (MOS). According to ITU-T Recommendation G.107, every rating value corresponds to a speech transmission category, as shown in Table 1. A rating below unacceptable quality, while values above 70 correspond to PSTN quality (values above 90 corresponding to very good qual- ity). R-value range MOS Speech transmission quality 100 90 4.50-4.34 best 90 low very poor Table 1: Speech transmission quality classes and corresponding rating value ranges. The E-model rating R is given by: where R0 groups the effects of noise, Is represents impairment that occur simultaneously with the voice signal (quantization), is the impairment associated with the mouth-to-ear delay, and Ie is the impairment associated with signal distortion (caused by low bit rate codecs and packet losses). The advantage factor A is the deterioration that callers are willing to tolerate because of the 'access advantage' that certain systems have over traditional wire-bound telephony, e.g. the advantage factor for mobile telephony is assumed to be 10. Since no agreement has been reached for the case of VoIP services, we will drop the advantage factor in this study. 4.2 Reduction of the E-model to transport level quantities Although an analytical expression for is given in [4] and values for Ie are provided in Appendix A of [5] for different loss con- ditions, those standards do not give a fully analytical expression for the R-factor. In this work, we use a simplified analytic expression for the R-factor that was proposed in [11] and that describes the R-factor as a function of observable transport level quantities. In this section, we briefly describe the reduction of equation (1) to transport level quantities as proposed in [11] and we introduce the assumptions made about the VoIP connections under study. 4.2.1 Signal-to-noise impairment factors R0 and Is Both R0 (effect of background and circuit noise) and Is (effect of quantization) describe impairment that have to do with the signal itself. Since none of them depend on the underlying transport net- work, we rely upon the set of default values that are recommended in [4] for these parameters. Choosing these defaults values, the rating R can be reformulated as: 4.2.2 Delay impairment ITU-T Recommendation G.107 [4] gives a fully analytical expression for in terms of various delay measures (such as mouth- to-ear delay, delay from the receive side to the point where signal coupling occurs and delay in the four wire loop) and other parameters describing various circuit switched and packet switch inter-working scenarios. Since we focus, in this work, on pure VoIP scenarios, we make the following simplifications: i) the various delay measures collapse into a single one, the mouth-to-ear delay, and, ii) the default values proposed in [4] are used for all parameters in the expression of other than the delay itself. In particular, the influence of echo is supposed negligible. The curve obtained describing Id as a function of the mouth-to-ear delay can then be approximated by a piece-wise linear function [11]: where d is the mouth-to-ear delay and H is the Heavyside func- tion. d is composed of the encoding delay (algorithmic and packetization delay), the network delay (transmission, propagation and queuing delay) and the playout delay introduced by the playout buffer in order to cope with delay variations. The Heavyside function is defined as follows: 4.2.3 Equipment impairment Ie The impairment introduced by distortion are brought together in Ie . Currently, no analytical expression allows to compute Ie as a function of parameters such as the encoding rate or the packet loss rate. Estimates for Ie must be obtained through subjective measurements. A few values for Ie are given in Appendix A of [5] for several codecs (i.e. G.711, G.729,.) and several packet loss conditions. In this work, we focus on the G.711 coder which does not introduce any distortion due to compression (and hence leads to the smallest equipment impairment value in absence of losses). In ad- dition, we assume that the G.711 coder in use implements a packet loss concealment algorithm. In these conditions, the evolution of the equipment impairment factor Ie as a function of the average packet loss rate can be well approximated by a logarithmic func- tion. In particular, if we assume that we are in presence of random losses, the equipment impairment can be expressed as follows [11]: where e is the total loss probability (i.e., it encompasses the losses in the network and the losses due to the arrival of a packet after its playout time). In summary, the following expression will be used to compute the R-factor as a function of observable transport quantities: 0:024d where d is the mouth-to-ear delay, e is the total loss probability and H is the Heavyside function defined in equation (4). 4.3 Call generation and rating In order to assess the quality of voice calls placed at random times during the measurement period, we emulate the arrival of short business calls. We pick call arrival times according to a Poisson process with a mean inter-arrival time of 60 seconds. We draw the call durations according to an exponential distribution with a mean of 3.5 minutes [17]. The randomly generated calls are then applied to the packet traces for quality assessment. Since IP telephony applications often use silence suppression to reduce their sending rate, we simulate talkspurt and silence periods within each voice call using for both periods an exponential distribution with an average of 1.5s [15]. Packets belonging to a silence period are simply ignored. At the receiver end, we assume that a playout buffer is used to absorb the delay variations in the network. The playout delay is defined as the difference between the arrival and the playout time of the first packet of a talkspurt. Within a talkspurt, the playout times of the subsequent packets are scheduled at regular intervals following the playout time of the first one. Packets arriving after 28.35 28.4 28.45 28.5 28.55 28.6 28.65 28.71357Delay (msec) Frequency Figure 2: Passive measurements: distribution of the one-way transmission delay between East and West Coast of the U.S. their playout time are considered lost. A playout buffer can operate in a fixed or an adaptive mode. In a fixed mode, the playout delay is always constant while in an adaptive mode, it can be adjusted between talkspurts. In this work, we opt for the fixed playout strategy because the measured delays and jitters are very small and a fixed playout strategy would represent a worst case scenario. Thus, we implement a fixed playout delay of 75ms (which is quite high, but still leads to excellent results, as described in Section 5). The quality of the calls described above is then computed as fol- lows. For each talkspurt within a call, we compute the number of packet losses in the network and in the playback buffer. From these statistics, we deduce the total packet loss rate e for each talkspurt. In addition, we measure the mouth-to-ear delay d, which is the sum of the packetization delay (20ms, in our case), the network delay of the first packet of the talkspurt and the playout delay. In order to assess the quality of a call we apply equation (6) to each talkspurt and then we define the rating of a call as the average of the ratings of all its talkspurts. 5. RESULTS In this section we discuss our findings derived from the experiments and measurements. We first compare the results obtained via the passive and active measurements and then focus on the impact of link failures on VoIP traffic. We conclude with a discussion of the call rating using the methodology proposed in Section 4. 5.1 Delay measurements In Figure 2 we show the one-way delay between two Sprint POPs located on the East and West Coast of the United States. The data shown refers to a trace collected from the passive measurement system on July 24th 2001. However, we have systematically observed similar delay distributions on all the traces collected in the Sprint monitoring project [13]. The delay between the two POPs is around 28.50ms with a maximum delay variation of less than 200s. Such delay figures show that packets experience almost no queueing delay and that the element that dominates the transmission delay is the propagation over the optical fiber [23]. We performed the same delay measurements on the UDP packets sent every 20ms from Reston (VA) to San Francisco (CA) for a period of 2.5 days. Figure 3 shows the distribution of the one-way East Coast to West Coast Delay (ms) Empirical density function Figure 3: Active measurements: distribution of the one-way transmission delay from Reston (VA) to San Francisco (CA). transmission delay. The minimum delay is 30.95ms, the average delay is 31.38ms while the 99.9% of the probes experience a delay below 32.85ms. As we can see from the figures, the results obtained by the active measurements are consistent with the ones derived from passive measurements. Low delays are a direct result of the over-provisioning design strategies followed by most tier-1 ISPs. Most tier-1 backbones are designed in such a way that link utilization remains below 50% in the absence of link failures. Such strategy is dictated by the need for commercial ISPs to be highly resilient to network failures and to be always capable of handling short-term variations in the traffic demands. The delay distribution in Figure 3 shows also another interesting feature: a re-routing event has occurred during the experiment. The distribution shows two spikes that do not overlap and for which we can thus identify two minima (30.96ms and 31.46ms), that represent the propagation delays of the two routes 3 . While the difference between the two minima is relatively high (around 500s), the difference in router hops is just one (derived from the TTL values found in the IP packets). One additional router along the path cannot justify a 500s delay increase [23]. On the other hand, the Sprint backbone is engineered so that between each pair of POPs there are two IP routhat use too disjoint fiber paths. In our experiment, the 500s increase in the delay is introduced by a 100km difference in the fiber path between the POPs where the re-routing occurred. 5.2 Impact of failures on data traffic In this section we investigate further the re-routing event. To the best of our knowledge there is no experimental study on failures and their impact on traffic on an operational IP backbone network. It can be explained by the difficulties involved in collecting data on the traffic at the time of a failure. Within several weeks of experi- ments, our VoIP traffic has suffered a single failure. Nevertheless, we believe it is fundamental for researchers and practitioners to study such failure events in order to validate the behaviors and per- 3 The delay distribution derived from passive measurements also shows some spikes. In that case, however, we cannot distinguish between delays due to packet sizes [23] or due to routing, given that we lack the routing information that would let us unambiguously identify the cause of the peaks. formance of routing protocols, routing equipment and to identify appropriate traffic engineering practices to deal with failures. The failure perturbed the traffic during a 50 minutes period between and 07:20 UTC on November 28th, 2002. During that failure event, the traffic experienced various periods of 100% losses before being re-routed for the rest of the experiment (33 hours). We now provide an in-depth analysis of the series of events related to the failure and we identify the causes of loss periods. We complement our active measurements with the routing data collected by our IS-IS listener. Figure 4 shows the delay that voice probe packets experienced at the time of the failure. Each dot in the plot represents the average delay over a five-second interval. Figure 5 provides the average packet loss rate over the same five-second intervals. At time 06:34, a link failure is detected and packets are re-routed along an alternative path that results in a longer delay. It takes about 100ms to complete the re-routing during which all the packets sent are lost. Although the quality of a voice call would certainly be affected by the loss of 100ms worth of traffic, the total impact on the voice traffic is minimal given the short time needed for the re-routing (100ms) and the small jitter induced (about 500s). After about a minute, the original route is restored. A series of 100% loss periods follows, each of which lasts several seconds. Figure 6 shows the one-way delay experienced by all the packets during one of these 100% loss periods (the same behavior can be observed in all the other periods). As we can see from the figure, packets are not buffered by the routers during the short outages (packets do not experience long delays) but they are just dropped because forwarded along an invalid path. Figure 7 shows the sequence numbers of the packets as received by the end host on the West Coast. Again, no losses nor re-orderings occur during those periods. This is a clear indication that packet drops are not due to congestion events but due to some kind of interface or router failure At time 06:48, the traffic experiences 100% losses for a period of about 12 minutes. Surprisingly, during that period no alternative path is identified for the voice traffic. At time 07:04 a secondary path is found but there are still successive 100% loss periods. Fi- nally, at 07:19, the original path is operational again and at time 07:36, an alternative path is chosen and used for the remaining part of the experiment. The above analysis corresponds to what can be observed from the active measurements. The routing data can provide us more information on the cause of these events. Figure 8 illustrates the portion of the network topology with the routers involved in the failure. The routers (R1 to R5 ) are located in 2 different POPs. The solid arrows show the primary path used by the traffic. The dashed arrows show the alternative path used after the failure. Table summarizes all the messages that we have collected from the IS-IS listener during the experiment. The "Time" column indicates the time at which the LSPs are received by our listener, the central column ("IS-IS LSPs") describes the LSPs in the format while the third column describes the impact on the traffic of the event reported by IS-IS. At the time of the first re-routing, routers via IS-IS the loss of adjacency with R4 . The fact that all the links from R4 are signaled down is a strong indication that the failure is a router failure as opposed to link failure. As we said earlier, the network reacts to this first event as expected. In about 100ms, R5 routes the traffic along the alternative path through In the period between 06:35 and 06:59, the IS-IS listener receives several (periodic) LSPs from all the five routers reporting that all Time (UTC) Delay (ms) East Coast to West Coast Figure 4: Average delay during the failure. Each dot corresponds to a five-second interval Packet loss rate East Coast to West Coast Figure 5: Average packet loss rate during the failure computed over five-second intervals the links are fully operational. During that time, though, the traffic suffers successive 100% loss periods. For about 13 minutes, R4 oscillates between a normal operational state (i.e. it forwards the packets without loss or additional delay) and a "faulty" state during which all the traffic is dropped. However, such "faulty" state never lasts long enough to give a chance to the other routers to detect the failure. At time 06:48, R4 finally reboots. It then starts collecting LSP messages from all the routers in the network in order to build its own routing table. This operation is usually very CPU intensive for a network of the size of the Sprint backbone. It may require minutes to complete as the router has to collect the LSP messages that all the other routers periodically send. While collecting the routing information, R4 does not have a routing table and is therefore not capable of handling any packet. As we described in Section 3, a router is expected to send LSP messages with the "Infinity Hippity Cost" bit set. In our case R4 does not set that bit. R5 , having no other means to know that R4 is not ready to route packets, forwards the voice traffic to R4 , where it is dropped. Time (UTC) Delay (ms) East Coast to West Coast Figure One-way delay of voice packets during the first 100% loss period Sequence number East Coast to West Coast Figure 7: Sequence numbers of received voice packets during the first 100% loss period At time 06:59, R4 builds its first routing table and the traffic is partially restored but the links flapping resulting again in a succession of 100% loss periods. Note that the traffic is only restored along the alternative path (hence, the longer delays) because the link between R1 and R4 is reported to be down. We conjecture that the 100% loss periods are due to R5 forwarding traffic to R4 every time the link R4 R5 is up, although R4 does not have a route to R1 . Most likely the links are not flapping because of an hardware problem but because R4 is starting receiving the first BGP updates 4 force frequent re-computations of the routing table to add new destination prefixes. Finally, at time 07:17 all routers report that the links with R4 are up and the routing remains stable for the rest of the experiment. Traffic is however re-routed again along the alternative path after about even if the original path is operational. This is due to the fact that R5 modifies its load balancing policy over theR4 can setup the I-BGP sessions, that run over TCP, with its peers only once it has a valid routing table, i.e. it has received all LSP updates. Time IS-IS LSPs Impact on traffic Re-routed through link to R4 is down R3 in 100ms Re-routed adjacency with R4 recovered through R4 from loss periods. to 07:06 link to R4 "flaps" 7 times Re-routed through R3 from 07:00 loss periods. to 07:17 link to R4 "flaps" 5 times Re-routed through R3 from 07:04 R5 : 100% loss periods. to 07:17 link to R4 "flaps" 4 times Re-routed through R3 Re-routed link to R4 is down through restored link to R4 is definitely up on the original path Table 2: Summary of the events occurred during the failure event two equal cost paths (solid and dashed arrows in Figure 8). Routers that perform per-destination prefix load balancing (as R5 , in our case) can periodically modify their criteria (i.e., which flow follows which path) in order to avoid that specific traffic patterns defeat the load balancing (e.g., if most of the packets belong to few destination prefixes, one path may result more utilized than the other). In order to summarize our findings, we divide the failure we observed in two phases: The first phase from time 06:34 to 06:59 is characterized by instabilities in the packet forwarding on router R4 : only few LSPs are generated but the traffic experience periods of 100% packet loss. Such "flapping" phase is due to the particular type of failure that involved an entire router and most likely the operating system of the router. The effect on packet forwarding and routing is thus unpredictable and difficult to control protocol-wise. The second phase goes from time 06:48 to 07:19 and is instead characterized by a very long outage followed by some routing instabilities and periods of 100% loss. This phase was caused by router R4 that did not set the "Infinity Hip- pity Cost" bit. We cannot explain how this problem arised as resetting the hippity bit after the collection of all the BGP updates is a common engineering practice within the Sprint backbone network. It is important to observe that both the first and the second phase of Figure 8: Routers involved by the failure. The solid arrows indicate the primary path for our traffic. The dashed arrows indicate the alternative path through R3 . Time (UTC) Call Rating East Coast to West Coast Figure 9: Voice call ratings (excluding the failure event) the failure event are not due to the IS-IS routing protocol. There- fore, we do not expect that the use of a different routing protocol (e.g. "circuit-based" routing mechanisms such as MPLS [25]) would mitigate the impact of failures on traffic. Instead, it is our opinion that router vendors and ISPs should focus their efforts on the improvement of the reliability of routing equipment, intended both in terms of better hardware architectures and more stable software implementations. Another important direction of improvement is certainly the introduction of automatic validation tools for router configurations. However, such tools would require first to simplify the router configuration proce- dures. As a side note, introducing circuits or label-switched paths on top of the IP routing will not help in such simplification effort. 5.3 Voice quality This section is devoted to the study of the quality experienced by a VoIP user. Figure 9 shows the rating of the voice calls during the 2.5 days of the experiment. We did not place any call during the failure event (50 minutes out of the 2.5 days) because the E-model only applies to completed calls and does not capture the events of loss of connectivity. Figure shows the distribution of call quality for the 2.5 days of experiment. All these results were derived assuming a fixed playout buffer. One can notice that the quality of calls does not deviate much from its mean value which is fairly 90.27. Among the 3,364 calls that were placed, only one experiences a quality rating below 70, the lower threshold for toll-quality. We are currently in the process of investigating what caused the low quality of some calls. Moreover, with 99% of calls experiencing a quality above 84.68, our results confirm that the Sprint IP backbone can support a voice service with PSTN quality standards. The very good quality of voice traffic is a direct consequence of the low delays, jitter and loss rates that probes experience. Without taking into account the 50 minutes of failure, the average loss rate is 0.19%. We also studied the probability of having long bursts of losses. The goal was to verify that the assumptions on the distribution of packet losses (in Section 4 we assumed that the losses were not bursty) and on the performance of packet loss concealment techniques are well suited to our experiment. For this purpose, we define the loss burst length as the number of packets dropped between two packets correctly received by our Call Rating R Frequency East Coast to West Coast Figure 10: Distribution of voice call ratings (excluding the failure Loss burst length Frequency of occurence 4 and above 0.16% Table 3: Repartition of loss burst lengths (excluding the failure end hosts. Table 3 shows the repartition of burst length among the losses observed during the period of experiment. The vast majority of loss events have a burst length 1, while 99.84% of the events have a burst length less than 4. This tends to indicate that the packet loss process is not bursty. Moreover, with a large majority of isolated losses, we can conjecture that packet loss concealment techniques would be efficient in attenuating the impact of packet losses. The results shown in Table 3 are in line with previous work of Bolot et al. [7] and they suggest that the distribution of burst length is approximately geometric (at least for small loss burst lengths). Future work will include an in-depth study of the packet loss process. 6. CONCLUSION We have studied the feasibility of VoIP over a backbone network through active and passive measurements. We have run several weeks of experiments and we can derive the following conclusions. A PSTN quality voice service can be delivered on the Sprint IP backbone network. Delay and loss figures indicate that the quality of the voice calls would be comparable to that of traditional telephone networks. We have pointed out that voice quality is not the only metric of interest for evaluating the feasibility of a VoIP service. The availability of the service also covers a fundamental role. The major cause of quality degradation is currently link and router failures, even though failures do not occur very often inside the backbone. We have observed that despite careful IP route protec- tion, link failures can significantly, although unfrequently, impact a VoIP service. That impact is not due to the routing protocols (i.e. IS-IS or OSPF), but instead to the reliability of routing equip- ment. Therefore, as the network size increases in number of nodes and links, more reliable hardware architectures and software implementations are required as well as automatic validation tools for router configurations. Further investigation is needed to identify all the interactions between the various protocols (e.g. IS-IS, I-BGP and E-BGP) and define proper semantics for the validation tools. The introduction of circuit or label switching networks will not help in mitigating the impact of failures. The failure event we have described in Section 5 is a clear example of this. As long as the failure is reported in a consistent way by the routers in the network, the IS-IS protocol can efficiently identify alternative routes (the first re-routing event completed in 100ms). The MPLS Fast-ReRoute would provide the same recovery time. On the other hand, a failing and unstable router that sends invalid messages would cause MPLS to fail, in addition to any other routing protocol. Future work will involve more experiments. Through long-term measurements, we aim to evaluate the likelihood of link and node failures in a tier-1 IP backbone. We also intend to address the problem of VoIP traffic traversing multiple autonomous systems. Another important area will be the study of metrics to compare the telephone network availability with the Internet availability. On telephone networks, the notion of availability is based on the down-time of individual switches or access lines. The objective of such metric is to measure the impact of network outages on customers. The Federal Communications Commission requires telephone operators to report any outage that affects more that 90,000 lines for at least minutes. Such rule is however difficult to apply to the Internet for a few reasons: i) there is no definition of "line" that can be applied; ii) it is very difficult to count how many customers have been affected by a failure; iii) from a customer standpoint there is no difference between outages due to the network or due to the servers (e.g. DNS servers, web servers, etc. 7. --R Advanced topics in MPLS-TE deployment transmission and quality aspects (STQ) Provisional planning values for equipment impairment factor Ie. The case for FEC-based error control for packet audio in the internet Measurement and interpretation of internet packet loss. Use of OSI IS-IS for routing in TCP/IP and dual environments Design principles for accurate passive measurements. Voice over IP performance Measuring Internet Telephony quality: where are we today? Qos measurement of internet real-time multimedia services Modeling of packet loss and delay and their effect on real-time multimedia service quality An Engineering Approach to Computer Networking. Assessment of VoIP quality over Internet backbones. Precision timestamping of network packets. Python routeing toolkit Analysis of measured single-hop delay from an operational backbone network Multiprotocol Label Switching Architecture. Measurement and modelling of temporal dependence in packet loss. --TR End-to-end packet delay and loss behavior in the internet An engineering approach to computer networking End-to-end routing behavior in the Internet Precision timestepping of network packets Voice over IP performance monitoring --CTR Baek-Young Choi , Sue Moon , Rene Cruz , Zhi-Li Zhang , Christophe Diot, Quantile sampling for practical delay monitoring in Internet backbone networks, Computer Networks: The International Journal of Computer and Telecommunications Networking, v.51 n.10, p.2701-2716, July, 2007 Baek-Young Choi , Sue Moon , Rene Cruz , Zhi-Li Zhang , Christophe Diot, Practical delay monitoring for ISPs, Proceedings of the 2005 ACM conference on Emerging network experiment and technology, October 24-27, 2005, Toulouse, France Feng Wang , Zhuoqing Morley Mao , Jia Wang , Lixin Gao , Randy Bush, A measurement study on the impact of routing events on end-to-end internet path performance, ACM SIGCOMM Computer Communication Review, v.36 n.4, October 2006 Adjusting forward error correction with temporal scaling for TCP-friendly streaming MPEG, ACM Transactions on Multimedia Computing, Communications, and Applications (TOMCCAP), v.1 n.4, p.315-337, November 2005 Nate Kushman , Srikanth Kandula , Dina Katabi, Can you hear me now?!: it must be BGP, ACM SIGCOMM Computer Communication Review, v.37 n.2, April 2007 Athina P. Markopoulou , Fouad A. Tobagi , Mansour J. Karam, Assessing the quality of voice communications over internet backbones, IEEE/ACM Transactions on Networking (TON), v.11 n.5, p.747-760, October

traffic measurements;routing protocols

507688

Topology-aware overlay networks for group communication.

We propose an application level multicast approach, Topology Aware Grouping (TAG), which exploits underlying network topology information to build efficient overlay networks among multicast group members. TAG uses information about path overlap among members to construct a tree that reduces the overlay relative delay penalty, and reduces the number of duplicate copies of a packet on the same link. We study the properties of TAG, and model and experiment with its economies of scale factor to quantify its benefits compared to unicast and IP multicast. We also compare the TAG approach with the ESM approach in a variety of simulation configurations including a number of real Internet topologies and generated topologies. Our results indicate the effectiveness of the algorithm in reducing delays and duplicate packets, with reasonable algorithm time and space complexities.

INTRODUCTION A variety of issues, both technical and commercial, have hampered the widespread deployment of IP multicast in the global Internet [14, 15]. Application-level multicast approaches using over- Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. NOSSDAV'02, May 12-14, 2002, Miami, Florida, USA. lay networks [9, 11, 12, 18, 24, 32, 42] have been recently proposed as a viable alternative to IP multicast. In particular, End System Multicast (ESM) [11, 12] has gained considerable attention due to its success in conferencing applications. The main idea of ESM (and its Narada protocol) is that end systems exclusively handle group management, routing information exchange, and overlay forwarding tree construction. The efficiency of large-scale overlay multicast trees, in terms of both performance and scalability, is the primary subject of this paper. We investigate a simple heuristic, which we call Topology Aware Grouping (TAG), to exploit underlying network topology data in constructing efficient overlays for application-level multicast. Our heuristic works well when underlying routes are of good quality, e.g., in intra-domain environments, and when final hop delays are small. Each new member of a multicast session first determines the path (route) from the root (primary sender) of the session to itself. The overlap among this path and other paths from the root is used to partially traverse the overlay data delivery tree, and determine the best parent and children for the new member. The constructed overlay network has a low delay penalty and limited duplicate packets sent on the same link. TAG nodes maintain a small amount of state information- IP addresses and paths of only their parent and children nodes. Unlike ESM, the TAG heuristic caters to applications with a large number of members, which join the session at different times. TAG works best with applications which regard delay as a primary performance metric and bandwidth as a secondary metric. For exam- ple, in a limited bandwidth streaming application or multi-player on-line game, latency is an important performance measure. TAG constructs its overlay tree based on delay (as used by current Internet routing protocols), but uses bandwidth as a loose constraint. Bandwidth is also used to break ties among paths with similar de- lays. We investigate the properties of TAG and model its economies of scale factor, compared to unicast and IP multicast. We also demonstrate via simulations the effectiveness of TAG in terms of delay, number of identical packets, and available bandwidth in a number of large-scale configurations. The remainder of this paper is organized as follows. Section 2 describes the basic algorithm, its extensions, and some design con- siderations. Section 3 analyzes the properties of TAG and its economies of scale factor. Section 4 simulates the proposed algorithm and compares it to ESM using real and generated Internet topologies. Section 5 discusses related work. Finally, section 6 summarizes our conclusions and discusses future work. 2. TOPOLOGY AWARE OVERLAYS Although overlay multicast has emerged as a practical alterna- Figure 1: Example of topology aware overlay networks tive to IP multicast, overlay network performance in terms of delay penalty and number of identical packets (referred to as "link stress") in large groups have been important concerns. Moreover, exchange of overlay end-to-end routing and group management information limits the scalability of the overlay multicast approach. Most current overlay multicast proposals employ two basic mecha- nisms: (1) a protocol for collecting end-to-end measurements among members; and (2) a protocol for building an overlay graph or tree using these measurements. We propose to exploit the underlying network topology information for building efficient overlay networks, assuming the underlying routes are of good quality. By "underlying network topology," we mean the shortest path information that IP routers maintain. The definition of "shortest" depends on the particular routing protocol employed, but usually denotes shortest in terms of delay or number of hops, or according to administrative policies. Using topology information is illustrated in figure 1. In the figure, source S (the root node) and destinations D1 to D4 are end systems that belong to the multicast group, and R1 to R5 are routers. Thick solid lines denote the current data delivery tree from S to D1 D4. The dashed lines denote the shortest paths to a new node D5 from S and D1. If D5 wishes to join the delivery tree, which member is the best parent node to D5? If D1 becomes the parent of D5, a relay path from D1 to D5 is consistent with the shortest path from S to D5. More- over, no duplicate packet copies are necessary in the sub-path S to (packets in one direction are counted separately from packets in the reverse direction). This heuristic is similar to determining if a friend is, more or less, on your way to work, so giving him/her a ride will not excessively delay you, and you can reduce overall traffic by car pooling. If he/she is out of your way, however, you decide to drive separately. In addition, this heuristic is subject to both capacity (e.g., space in your car) and latency (car pooling will not make the journey excessively long) constraints. Of course, it is difficult to determine the shortest path and the number of identical packets, in the absence of any knowledge of the underlying network topology. If network topology information can be obtained by the multicast participant (as discussed in section 2.6), nodes need not exchange complete end-to-end mea- surements, and topology information can be exploited to construct high quality overlay trees. Therefore, our heuristic is: a TAG destination selects as a parent the destination whose shortest path from the source has maximal overlap with its own path from the source. This heuristic minimizes the increase in number of hops (and hence delay if we assume low delay of the last hop(s)) over the shortest unicast path. We also use loose bandwidth constraints. As with all overlay multicast approaches, TAG does not require class D addressing, or multicast router support. A TAG session can be identified by (root IP addr, root port), where "root" denotes the primary sender in a session. The primary sender serves as the root of the multicast delivery tree. The case of multiple senders will be discussed in section 2.8. 2.1 Assumptions TAG makes a number of assumptions: 1. TAG is used for single-source multicast or core-based mul- ticast: The source node or a selected core node is the root of the multicast forwarding tree (similar to single-source multi-cast [21] or core-based multicast [6] for IP multicast). 2. Route discovery methods exist: TAG can obtain the shortest path for a sender-receiver pair on the underlying network. A number of route discovery tools are discussed in section 2.6. 3. All end systems are reachable: Any pair of end systems on the overlay network can communicate using the underlying network. Recent studies, however, indicate that some Internet routes are unavailable for certain durations of time [31, 26, 7]. 4. Underlying routes are of good quality (in terms of delay): Intra-domain routing protocols typically compute the shortest path in terms of delay for a sender-receiver pair. Recent studies, however, indicate that the current Internet demonstrates a significant percentage of routing pathologies [31, 36]. Many of these arise because of policy routing techniques employed for inter-domain routing. TAG is best suited for well-optimized routing domains. 5. The last hop(s) to end systems exhibit low delay: A long latency last hop to an end system, e.g., a satellite link, adversely affects TAG performance. TAG works best with low-delay a last hop to an end system (or last few hops for the partial path matching flavor of TAG, as discussed in section 2.5). 2.2 Definitions We define the following terms, which will be used throughout the paper. DEFINITION 1. A path from node A to node B in TAG, denoted by P (A; B), is a sequence of routers comprising the shortest path from node A to node B according to the underlying routing protocol. P (S; will be referred to as the spath of A where S is the root of the tree. The length of a path P or len(P ) is the number of routers in the path. DEFINITION 2. A B if P (S; A) is a prefix of P (S; B), where S is the root of the tree. For example, the path from S to D5 (or spath of D5) in figure 1 is A TAG node maintains a family table (FT) defining parent-child relationships for this node. One FT entry is designated for the parent node, and the remaining entries are for the children. As seen in figure 2, an FT entry consists of a tuple (address, spath). The address is the IP address of the node, and the spath is the shortest path from the root to this node. 2.3 Complete Path Matching Algorithm The path matching algorithm traverses the overlay data delivery tree to determine the best parent (and possibly children) for a new node. The best parent for a new node N in a tree rooted at S is: <IPaddr(Bi), P(S,Bi)> <IPaddr(A), P(S,A)> Children Parent FT Figure 2: Family table New member of New member subscribes here (a) (b) FT New member subscribes here (c) Children FT FT New Member New Member New Member Children Figure 3: The three conditions for complete path matching A node C (C 6= N ) in the tree, such that C N and len(P (S; C)) len(P (S; R)) for all nodes R N in the tree. The algorithm considers three mutually exclusive conditions, as depicted in figure 3. Let N be a new member wishing to join a current session. Let C be the node being examined, and S be the root of the tree. If possible, we select a node A such that A is a child of and continue traversing the sub-tree rooted at A (figure 3(a)). Oth- erwise, if there are children A i of C such that N A i for some becomes a child of C with A i as its children (figure 3(b)). In case no child of C satisfying the first or second conditions ex- ists, N becomes a child of C (figure 3(c)). Note that no more than one child satisfying the first condition can exist. The complete path matching algorithm is presented in figure 4. In the algorithm, N denotes a new member; C is the node currently being examined by target is the next node which N will probe, if necessary. There are two reasons for selecting a node (in the first and second whose spath is the longest prefix of the spath of the new member. First, the path from the source to the new member will be consistent with the shortest path determined by routing algorithms. This reduces the additional delay introduced by overlays. Second, sharing the longest prefix curtails the number of identical packet copies in the overlay network, since a single packet is generated over the shared part. 2.4 Tree Management In this section, we discuss the multicast tree management proto- including member join and member leave operations, and fault resilience issues. 2.4.1 Member Join A new member joining a session sends a JOIN message to the primary sender S of the session (the root of the tree). Upon the receipt of a JOIN, S computes the spath to the new member, and executes the path matching algorithm. If the new member becomes a child of S, the FT of S is updated accordingly. Otherwise, S propagates a FIND message to its child that shares the longest spath prefix with the new member spath. The FIND message carries the IP address and the spath of the new member. The FIND is processed by executing path matching and either updating the FT, or propagating the FIND. The propagation of FIND messages continues until the new member finds a parent. The process is depicted in figure 5. An example is illustrated in figure 6. Source S is the root of the multicast tree, through R4 are routers, and D1 through D5 are destinations. The thick arrows denote the multicast forwarding tree computed by TAG. The FT of each member (showing only the children) is given next to it. The destinations join the session in the order D1 to D5. Upon the receipt of a JOIN message from proc ch := first child of C ; f lag := condition(3); while (ch is NOT NULL) do if (ch N ) then target := ch; f lag := condition(1); fi; if (N ch ) then add ch to children(N) ; f lag := condition(2); fi; if (flag is NOT then ch := next child of C ; else ch if (flag is then else add N to children(C) ; currently being examined new node joining the group ch : a child of C target : next node N will examine Figure 4: Complete path matching algorithm creates an entry for D1 in its FT (figure 6(a)). S computes the shortest path for a destination upon receiving the JOIN message of that destination. When D2 joins the session (figure 6(b)), S executes the path matching algorithm with the spath of D2. S determines that D1 is a better parent for D2 than itself, and sends a FIND message to D1 which takes D2 as its child. D3 similarly determines D2 to be its parent (figure 6(c)). When D4 joins the session, D4 determines D1 to be its parent and takes D2 and D3 as its children. The FTs of D1 to D4 are updated accordingly (figure 6(d)). Finally, D5 joins the session as a child of D4 (figure 6(e)). Figure 6(e) depicts the final state of the multicast forwarding tree and the FT at each node. 2.4.2 Member Leave A member can leave the session by sending a LEAVE message to its parent. For example, if D4 wishes to leave the session (fig- Root New Member Path Matching FIND FIND JOIN Request/Reply Request/Reply Request/Reply Figure 5: A member join process in TAG FT (c) (b) (a) (d) FT D3: (R1,R2,R4) FT FT D4: (R1, R2) FT FT FT FT D5: (R1,R2,R3) D4: (R1,R2) D3: (R1,R2,R4) FT FT FT FT D3: (R1, R2, R4) FT FT Figure Member join in TAG ure 7(a)), D4 sends a LEAVE message to its parent D1. A LEAVE message includes the FT of the leaving member. Upon receiving LEAVE from D4, D1 removes D4 from its FT and adds FT entries for the children of D4 (D2 and D5 in this case). The updated multicast forwarding tree is illustrated in figure 7(b). 2.4.3 Fault Resilience Failures in end systems participating in a multicast session (not an uncommon occurrence) affect all members of the subtree rooted at the failing node. To detect failures, a parent and its children periodically exchange reachability messages in the absence of data. When a child failure is detected, the parent simply discards the child from its FT, but when a parent failure is detected, the child must rejoin the session. 2.5 Bandwidth Considerations Since TAG targets delay-sensitive applications, delay (as defined by the underlying IP routing protocol) is the primary metric used in path matching. The complete path matching algorithm presented in figure 4 can reduce the delay from source to destination, and reduce the total number of identical packets. However, high link stress [12] (and limited bandwidth to each child) may be experienced near a few high degree or limited bandwidth nodes in the constructed tree. To alleviate this problem, we loosen the path matching rule when a node is searching for a parent. The new rule allows a node B to (a) (b) D5: (R1, R2, R3) D3: (R1, R2, R4) D4: (R1, R2) D5: (R1, R2, R3) D3: (R1, R2, R4) FT FT FT FT FT FT FT Figure 7: Member leave in TAG attach to a node A as a child, if A has a common spath prefix of length len(P (S; A)) k with B (S is the root of the tree), even if the remaining k elements of the spath of A do not match the spath of B. We call this method partial path matching or minus-k path matching. We use the symbol partial(k) to denote minus-k path matching. Minus-k path matching allows children of a bandwidth-constrained node to take on new nodes as their children, mitigating the high stress and limited bandwidth near the constrained node. When the available bandwidth at a given node falls below a threshold bwthresh, minus-k path matching is activated. The threshold bwthresh does not give a strict guarantee, but it gives an indication that alternate paths should be explored. The k parameter controls the deviation allowed in the path matching. A large value of k may increase the delay to a node, but it reduces the maximum stress a node may experience, therefore increasing available bandwidth to the node. With minus-k matching, a new member can examine several delay-based paths while traversing the tree, and select the path which maximizes bandwidth. When a node C is being probed by the new member, all children of C which are eligible to be potential ancestors of the new member (by the minus-k path matching algo- rithm) constitute a set of nodes to examine next. The node which gives the maximum bandwidth among these nodes is the one se- lected. The partial path matching algorithm is presented in figure 8. The member leave operation also employs minus-k path matching. When the parent of a leaving member receives a LEAVE message, the parent first removes the leaving member from its FT. Children of the leaving member (included in the LEAVE message) then execute minus-k path matching at the parent of the leaving member to find their new parents. Figure 9 illustrates an example of member join and leave with new member D3 takes D1 as its parent since D1 provides more bandwidth than D2. When D1 leaves, D4 becomes a child of D3. D3 maximizes bandwidth to D4 among the children of D0 (D2 and D3). Table 1 shows the tradeoff between the delay (mean RDP as defined in section 4), total link stress on the tree, and maximum link stress in the tree, for a variety of bwthresh values (in kbps). The simulation setup used will be discussed in section 4. The configuration used here is TAG-TS2 with 1000 members. We use a fixed these simulations, though we are currently investigating dynamic adaptation of k. As seen in the table, as bwthresh increases, minus-k path matching is activated more often. Conse- quently, a larger bwthresh value reduces the total number of identical packets and maximum stress, but increases the RDP value. If TAG does not use minus-k path matching (bwthresh=0), TAG trees suffer from a large number of identical packets and high maximum stress, yielding little bandwidth for many connections. leave Member Leave (k=1) Member Join (k=1) D0-D4: 100 kbps D1-D4: 800 kbps D2-D4: 300 kbps D3-D4: 600 kbps Figure 9: Member join and leave with partial path matching Table 1: Tradeoffs with different bwthresh values (in kbps) bwthresh (in kbps) RDP Total stress Max. stress 300 1.559278 1928 104 200 1.526061 2142 105 100 1.384193 2680 146 2.6 Obtaining Topology and Bandwidth Data When a new member joins a multicast session in TAG, the root must obtain the path from itself to the new member. We propose two possible approaches for performing this. The first approach is to use a network path-finding tool such as traceroute. Traceroute has been extensively used for network topology discovery [19, 31]. Some routers, however, do not send ICMP Time-Exceeded messages when Time-To-Live (TTL) reaches zero for several rea- sons, the most important of which is security. Recently, we conducted simple experiments where we used traceroute for 50 sites in different continents at different times and on different days. Approximately 90% of the routers responded. The average time taken to obtain and print the entire path information was 5.2 sec- onds, with a maximum of 42.6 seconds and a minimum of 0.2 sec- onds. 5 8% traceroute failures were reported in [31]. A recent study [20] indicates that router ICMP generation delays are generally in the sub-millisecond range (< 500 secs). This shows that only a few routers in today's Internet are slow in generating ICMP Time-Exceeded messages. The second option is to exploit topology servers. For example, an OSPF topology server [38] can track intra-domain topology, either by listening to OSPF link state advertisements, or by pushing and pulling information from routers via SNMP. Network topology can also be obtained from periodic dumps of router configuration files [17], from MPLS traffic engineering frameworks [4], and from policy-based OSPF monitoring frameworks [5]. Internet topology discovery projects [16, 8, 10] can also supply topology information to TAG when a new member joins or changes occur. Topology servers may, however, only include partial or coarse-grained (e.g., AS-level) information [29, 16, 8, 10]. Partial information can still be exploited by TAG for partial path matching of longest common subsequences. Bandwidth estimation tools are important for TAG, in conjunction with in-band measurements, to estimate the available band-width between nodes under dynamic network conditions. Tools similar to pathchar [22] estimate available bandwidth, delay, average queue, and loss rate of every hop between any source and destination on the Internet. Nettimer [27] is useful for low-overhead measurement of per-link available bandwidth. Other bandwidth or throughput measurement tools are linked through [1]. 2.7 Adaptivity and Scalability If network conditions change and the overlay tree becomes inefficient (e.g., when a mobile host moves or paths fail), TAG must adapt the overlay tree to the new network conditions. An accurate adaptation would entail that the root probe every destination periodically to determine if the paths have changed. When path changes are detected, the root initiates a rejoin process for the destinations affected. This mechanism, however, introduces scalability problems in that the root is over-burdened and many potential probe packets are generated. We propose three mechanisms to mitigate these scalability prob- lems. First, intermediate nodes (non-leaf nodes) participate in periodic probing, alleviating the burden on the root. The intermediate nodes only probe the paths to their children. Second, path-based aggregation of destinations can substantially reduce the number of hops and destinations probed. Destinations are aggregated if they have the same spath. Only one destination in a destination group is examined every round. During the next round, another member of the group is inspected. When changes are detected for a certain all members of that group are updated. Third, when changes in part of the spath are detected for a destination, not only the destination being probed, but also all the destinations in the same group and all the destinations in groups with overlapping spaths, are up-dated Although these TAG reorganizing mechanisms help reduce over- head, the root is likely to experience a high load when a large number of members join or rejoin simultaneously. The root is also a single failure point. To address these limitations, mechanisms similar to those used in Overcast [24] can be used. For example, requests from members to the root are redirected to less-burdened replicated roots. Another important consideration is the quality of the TAG tree in terms of both delay and bandwidth. TAG aligns overlay routes with underlying routes, assuming underlying routes are of good quality (fourth assumption in section 2.1). Unfortunately, today's Internet routing, particularly inter-domain routing, exhibits pathologies, such as slow convergence and node failures. Savage et al. [36, 35] found that there is an alternate path with significantly superior quality to the IP path in 30 80% of the cases (30 55% for latency, 75 85% for loss rate, and 70 80% for bandwidth). Intra-domain network-level routes, however, are generally of good qual- ity, and, in the future, research on inter-domain routing and traffic proc PathMatch (C ; N) ch := first child of C ; f lag := condition(3); then minus-k path matching in condition(1) activated ; while (ch is NOT NULL) do if (ch partial(k) N && bandwidth(ch) > maxbw) then target := ch; f lag := condition(1); fi; if (N ch&& flag is NOT then add ch to children(N) ; f lag := condition(2); fi; ch := next child of C ; if (flag is then else add N to children(C) ; currently being examined new node joining the group ch : a child of C target : next node N will examine among potential parents of N Figure 8: Partial path matching algorithm, with bandwidth as a secondary metric. The bandwidth() function gives the available bandwidth between a node and the new member engineering may improve route quality in general. Another important consideration is that a long latency last hop(s) to a TAG parent node may yield high delay to its children (section 2.1). A delay-constrained overlay tree construction mechanism can be combined with the TAG heuristic to prevent such high delay paths. As for bandwidth, it is only considered as a secondary metric (as a tie breaker among equal delay paths) in tree construction (sec- tion 2.5). Minus-k path matching does not guarantee bandwidth- it simply explores more potential paths when bandwidth is scarce. A bandwidth-constrained tree construction mechanism can be incorporated into TAG if bandwidth guarantees are required. 2.8 Multiple Sender Groups In the current version of TAG, a sender other than the root of the tree must first relay its data to the root. The root then multicasts the data to all members of the session. This approach is suitable for mostly single-sender applications, where the primary sender is selected as the root, and other group members may occasionally transmit. In applications where all members transmit with approximately equal probabilities, the root of the tree should be carefully selected. This is similar to the core selection problem in core-based tree approaches for multicast routing [6]. Multiple (backup) roots are also important for fault tolerance. 3. ANALYSIS OF TAG In this section, we investigate the properties of TAG, and study its bandwidth penalty compared to IP multicast. For simplicity, we use TAG with complete path matching (figure 4) in our analysis, except for the complexity analysis where we analyze both complete and mins-k path matching. 3.1 Properties of TAG In this section, we study the conditions used in the path matching algorithm, and the properties of the trees constructed by TAG. LEMMA 1. Node A is an ancestor of node B in the TAG tree iff A B. Proof: We first show that if node X is the parent of node Y , denoted by . In the path matching algorithm, X can become the parent of Y by the second or by the third path matching conditions. Both cases guarantee X Y . Then, we generalize to the case when node A is an ancestor (not necessarily the parent) of node B. In this case, there must be n (n > 0) nodes such that holds, according to the previous case. Similarly, M2 M1 B. Transitively, (: This follows from conditions 1 and 2 in the path matching algorithm. 2 In figure 1, node D1 is an ancestor of node D5 because P (S; D1) is a prefix of P (S; In contrast, the fact that P (S; is not a prefix of implies that node D2 is not an ancestor of D5. We now investigate the conditions of the path matching algorithm. LEMMA 2. The three conditions in the TAG path matching algorithm (given in figure are mutually exclusive (no two of the three conditions can occur simultaneously) and complete (no other case exists). Proof: We first prove mutual exclusion. To show mutual exclusion is equivalent to proving no two conditions can hold simultaneously. The first and the third conditions, and the second and the third conditions cannot co-exist by definition. Therefore, we need to show that the first and second conditions cannot both hold at the same time. Suppose the first and the second conditions occur simultaneously for a node C that is being examined. A new member N selects B, a child of C, such that B N for further probing by the first condition. By the second condition, there must exist a node another child of C, such that N B 0 , and B and B 0 are sib- lings. However, in this case, the path matching algorithm would have previously ensured that B 0 is a descendant of B, not a child of C, by lemma 1, since B N B 0 . This is a contradiction. Since the third condition includes the complement set of the first and the second conditions, the conditions are complete. 2 Now we study the number of trees TAG can construct. LEMMA 3. TAG constructs a unique tree if all members have distinct spaths, regardless of the order of joins. If there are at least two members with the same spaths, the order of joins alters the constructed tree. Proof: By lemma 1, a unique relationship among every two nodes (i.e., parent, ancestor, child, descendant, or none) is established R R R more hops Figure 10: Bound on the number of hops in TAG among every two nodes which have different spaths, independent of the order of joins. If two members have the same spath, one must be an ancestor of the other. Therefore, n! distinct trees can be constructed by TAG if n group members have the same spaths (according to the order of their joins). 2 We now study the properties of a parent node. LEMMA 4. For all i, the spath of the parent of nodes A i has the longest prefix of the spath P (S; A i ), where "longest" denotes longest in comparison to the spaths of all members in a session and S is the root of the tree. Proof: Consider two nodes B and C where B is the parent of C. By lemma 1, B C, i.e., P (S; B) is a prefix of P (S; C). Suppose there exists a node A such that P (S; A) is a prefix of are prefixes of the same path P (S; C) and len(P (S; A)) > len(P (S; B)), then P (S; B) is a prefix of P (S; A). By definition A. Therefore, A must be a descendant of B according to lemma 1. The path matching algorithm, however, would make C a child of A, instead of B, since B A and A C. This is a contradiction. Hence, P (S; B) must be the longest prefix of is the parent of C. 2 Finally, we give a bound for the number of hops on the path from the root to each member. LEMMA 5. For every destination i in a TAG tree, SPD(i) E(i) 3SPD(i) 2, where SPD(i) is the number of hops on the shortest path from root S to i, and E(i) is the actual number of hops from S to i in the TAG tree. Proof: Consider P (S; i), the path from root S to i. By the definition of len, since SPD(i) is the number of hops on the path P (S; i). The fact that SPD(i) is the number of hops on the shortest path P (S; i) ensures SPD(i) E(i). The maximum E(i) occurs when i has as many ancestors as len(P (S; i)). This situation is depicted in figure 10. In the fig- ure, R denotes a router; i is a destination, and nodes M are all ancestors of i. For every M , 2 hops are added to the path. Thus, are added to SPD(i). Therefore, the maximum E(i) is 3 SPD(i) 2. 2 3.2 Time Complexity In this section, we analyze the time complexities of both complete path matching and minus-k path matching. For simplicity, we assume that an overlay multicast tree has n end systems, and that each end system has an average of m children. We also assume that an average of v routers exist over the link between a parent and its child. After the source discovers the path to a new member, the member join process requires: (1) operations for each node, and (2) tree traversal. For (1), suppose that the new member has matched its path to a node at level i 1 and is searching for a matching node at level i in the overlay multicast tree, where 0 < i < log m n. (The height of the tree is log m n, which is assumed to be an inte- ger.) In complete path matching, operations in (1) require mvi operations for the new member at to search for the next matching node at level i, in the worst case. The new member can find the next matching node at level i by examining at most vi routers per child for each of the m children of . Operation (2) is of order log m n from the root to a leaf node. Therefore, the time complexity of member join is: log n) (1) Member leave requires one deletion and m additions of FT en- tries. Each entry requires v log m n operations for the worst case path length. Thus, the time complexity of member leave is O(mv log m n). In minus-k path matching, (1) requires m(k operations if k +vi v log m n. Otherwise, (1) requires mv log m n operations. A new member matched with node examines k more routers than in the complete path matching case. Hence, k + vi routers are examined by the new member per child to find the next matching node routers are examined (the maximum path length). This is performed for each of the m children of . Since operation (2) requires log m n operations and assuming that k is small, the time complexity of member join is: log c) n) (2) As discussed in section 2.5, children of a leaving member re-join the session starting from the parent of the leaving member in minus-k path matching. In this process, each of m children of the leaving member requires O(mv log 2 n) operations for the rejoin. Hence, the time complexity of member leave is O(m 2 v log 2 m n) in this case. 3.3 Modeling the Economies of Scale Factor Two important questions to answer about an overlay multicast tree are: (1) how much bandwidth it saves compared to naive uni- cast; and (2) how much additional bandwidth it consumes compared to IP multicast. IP multicast savings over naive unicast have been studied in [2, 13, 33]. Chuang and Sirbu [13] investigated the cost of IP multicast for a variety of real and generated network topologies. Their work was motivated by multicast pricing. They found that L(m) / m 0:8 , where L(m) is the ratio between total number of multicast links and average unicast path length, and m is the number of distinct routers to which end systems are connected. They also found that the cost of a multicast tree saturates when the number of subscribing end systems exceeds a certain value. Based on these results, they suggested membership-based pricing until m reaches the saturation point, and flat-rate pricing beyond that point. In this section, we quantify the network resources consumed by TAG. We derive a bound for the function LTAG (n), which denotes the sum of link stress values on all router-to-router links, for a multicast group of size n. Although the number of distinct routers to which end systems are connected is used in [13], we use Router End system Figure 11: A k-ary Tree Model d level (2) l Figure 12: TAG Trees n, the number of end systems in a multicast group. As discussed in [2], using the number of end systems is intuitively appealing and makes the analysis simpler. Note that m can be approximated by is the total number of possible routers to which end systems can be connected. m n when M 1. For simplicity, we assume a k-ary data dissemination tree in which tree nodes denote routers (as in [2, 33]), as depicted in figure 11. The height of the tree is H and all nodes except the leaves have degree k. We assume that no unary nodes (nodes at which no branching occurs) exist. Therefore, our results are approximate. An end system can be connected to any router (node in the tree). Suppose that n end systems join the multicast session. The probability that at least one end system is connected to a given node is: 1 is the number of possible locations for the subnet of an end system, which is equal to the number of nodes in the tree. We now evaluate the cost of transmission at each level of the tree. In figure 12, B l indicates the cost over the link between node - at level l and its parent at level l 1, and B l+1 (a) denotes the cost over the links between node - at level l and its children at level considering two different cases: when at least one end system is connected to node -, and when no end system is connected to -. Let B 1 l be the cost in the first case and B 2 l be the cost in the second case. TAG enforces that the first case costs one, for transmission between node - and its parent. This is because node - sends packets from the parent to the children (B l = 1). In the second case, however, since no system relays the packets at -, the cost over outgoing links of - towards the leaves is equal to the cost over the link between - and its parent. Therefore, B 1 We assume the end systems are uniformly distributed to tree nodes. This assumption implies that E[B l+1 l Normalized Overlay Tree Cost Number of Members Figure 13: TAG tree cost versus group size Hence, E[B l ] is defined as follows: l l Solving the recurrence in (4) and (5), we obtain: l E[B l The cost LTAG (n) is given by (6): Figure 13 plots the normalized overlay tree cost of TAG for a variety of k and H values on a log-log scale. The normalized overlay tree cost LTAG (n)=^u is defined as LTAG (n), the cost of an overlay with n members, divided by the average number of hops (only counting router-to-router links) for a unicast path from the source to receivers, ^ u. Since we assume end systems are uniformly distributed at nodes, ^ u is the average number of hops from the root to a node on the overlay tree: All curves stabilize for group sizes exceeding 1000 5000 mem- bers. The slope decreases because as group size grows, more end systems can share the links yielding more bandwidth savings. This is an important advantage of TAG over unicast. The figure shows that, approximately, LTAG (n)=^u / n 0:95 before the curves sta- bilize. The factor 0.95 is smaller than unicast, but larger than the factor for IP multicast (L IPmulticast (n)=^u / n 0:8 ), where replication at the routers, together with good multicast routing algorithms yield additional savings. We will verify these results via simulations in section 4.3. 4. PERFORMANCE EVALUATION We first discuss the simulation setup and metrics, and then analyze the results. 4.1 Simulation Setup and Metrics We have implemented session-level (not packet-level) simulators for both TAG and ESM [12] to evaluate and compare their performance. The simulators model propagation delays, but not queuing delays and packet losses. Two sets of simulations were performed on different topologies. The first set uses Transit-Stub topologies generated by GT-ITM [41]. The Transit-Stub model generates router-level Internet topologies. We also simulate AS-level topologies, specifically the actual Internet AS topologies from NLANR [29] and topologies generated by Inet [25]. The Transit-Stub model generates a two-level hierarchy: inter-connected higher level (transit) domains and lower level (stub) do- mains. We use three different topologies with different numbers of nodes: 492, 984, and 1640. When the total is 492 nodes, there are nodes per transit domain, 5 stub domains per transit node, and 8 nodes per stub domain. Similar distributions are used when the total number of nodes is 984 and 1640. We label the 3 transit-stub topologies TS1, TS2, and TS3 respectively, e.g., label "TAG-TS1" denotes the results of TAG on the Transit-Stub topology with 492 nodes. Multicast group members are assigned to stub nodes randomly. The multicast group size ranges from 60 to 5000 members. GT-ITM generates symmetric link delays ranging from 1 to 55 ms for transit-transit or transit-stub links. We use 1 to ms delays within a stub. We randomly assign bandwidth ranging between 100 Mbps and 1 Gbps to backbone links. We use 500 kbps to 10 Mbps for the links from edge routers to end systems. The AS topologies from NLANR and Inet give AS-level connec- tivity. AS-level maps have been shown to exhibit a power-law [16]. This means that a few nodes have high-degree connectivities, while most other nodes exhibit low-degree connectivities. We use the 1997 and 1998 NLANR data sets, named AS97 and AS98, respec- tively. We also use the 1997 Inet data set (named Inet97) and the 1998 Inet data set (Inet98), which have the same number of ASes as the NLANR data sets: 3015 and 3878. We have 4000 (for 1997) and 5000 (for 1998) members in a multicast session, and assign members to ASes randomly. Link delays and bandwidths in the same ranges as the Transit-Stub configuration are used for the AS configurations. The link delays are asymmetric. We assume that the IP layer routing algorithm uses delay as a metric for finding shortest paths. The routing algorithm for the mesh in ESM uses discretized levels of available bandwidth (in 200 kbps increments) as the primary metric, and delay as a tie breaker. The minus-k path matching algorithm is used in TAG with a fixed specified. We use the same parameters for ESM used in the simulations in [11, 12] (lower degree bound = 3, upper degree high delay penalty = 3), except for delay-related parameters (close neighbor delay = 85 ms) since we assign a wider range of delays to the links. We use the following performance metrics [11, 12] for evaluating TAG and ESM: 1. Mean Relative Delay Penalty (RDP): RDP is the relative increase in delay between the source and a receiver in TAG against unicast delay between the source and the same re- ceiver. The RDP from source s to receiver dr is the ratio latency(s;dr delay(s;dr . The latency latency(s; dr ) from s to dr is defined to be delay(s; d0 assuming s delivers data to dr via the sequence of end systems denotes the end- to-end delay from d i to d i+1 . We compute the mean RDP of all receivers. 2. Link Stress: Link stress is the total number of identical copies of a packet over a physical link. We compute the total stress for all tree links. We also compute the maximum value of link stress among all links.This is clearly a network-level2610 Mean Group Size Figure 14: Mean RDP: TAG versus ESM5000150002500035000 Total Stress Group Size Figure 15: Total stress: TAG versus ESM metric and is not of importance to the application user. 3. Mean Available Bandwidth (in kbps): This is the mean of the available bottleneck bandwidth between the source and all receivers. The TAG tree cost is also computed in section 4.3, and compared to IP multicast and unicast cost. 4.2 Performance Results 4.2.1 Transit-Stub Topologies The mean RDP values for TAG and ESM on the three different Transit-Stub topologies (TS1, TS2, TS3) are plotted in figure 14. From the figure, TAG exhibits lower mean RDP values than ESM for different group sizes in all 3 topologies. The mean RDP values for TAG-TS1, TS2, and TS3 are all in the range of 1 to 2, while the mean RDP values for ESM range from 2 to 6. This is because TAG considers delay as the primary metric while ESM uses delay only as a tie breaker. Although TS3 is a larger scale topology than TS2, and TS2 is larger than TS1, the mean RDP values for TAG are similar for the different topologies. Mean RDP values for TAG increase with the increase in group size. As more end systems join in TAG, the mean RDP values increase due to the bandwidth constraint in partial path matching (even though lower latency paths may become available). We observe that, unlike TAG, the mean RDP values for ESM do not always increase with the increase in group size. Figure 15 illustrates the total stress of TAG and ESM for the three different topologies. For all group sizes and topologies, TAG Cumulative Percentage Stress (log-scale) Figure Cumulative distribution of stress20060010000 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 Stress Group Size Figure 17: Maximum stress: TAG versus ESM total stress is below 8000. In contrast, ESM exhibits higher stress. The total stress for all 6 configurations increases in proportion to the group size, since more identical packets traverse physical links when more end systems join the session. TAG-TS1 and TAG- TS2 exhibit the lowest total stress. TAG can avoid some duplicate packets since TAG aligns overlay and underlying routes, subject to bandwidth availability. Packets from the source to a receiver are not duplicated along the path from the source to the parent of that re- ceiver. Figure 16 depicts the cumulative distribution of total stress. The figure shows that the three TAG configurations have slightly more low-stress links than the three ESM configurations. Figure 17 illustrates that TAG, with the correct parameters, can reduce the maximum stress value as well. With the complete path matching algorithm, a strategically located end system attached to a high-degree router can be the parent of numerous nodes, which severely constrains the bandwidth available to each of these nodes, and increases the stress at this end system. The minus-k path matching algorithm remedies this weakness, as shown in the figure. The mean bandwidth, depicted in figure 18, denotes the average of the available bottleneck bandwidths from the source to all mem- bers. The available bottleneck bandwidth with ESM is high (from 600 to 1600 kbps) for up to 500 members, and then stabilizes at approximately 600 kbps for groups exceeding 500 members. TAG gives 200 800 kbps bandwidth for up to 500 members. For larger groups, the bandwidth rapidly drops to under 200 kbps. The bottle-neck bandwidth given by TAG continues to decrease as the group size increases. TAG bottleneck bandwidth is very sensitive to the number of members, and to the bandwidth threshold 200 kbps here). This and the fact that ESM optimizes bandwidth200600100014000 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 Mean Bandwidth Group Size Figure Mean Mean Upper Degree Bound (UDB) Figure 19: Mean RDP for different UDB values in ESM explain why ESM performs better than TAG. An important point to note is that in the ESM algorithm, the lower and upper degree bounds (LDB and UDB, respectively) for each group member play a key role. The two parameters control the number of neighbors that each member communicates with. In particular, the upper degree bound is significant, as it impacts both protocol scalability and performance. In our simulations, we observe that increasing the upper degree bound for ESM reduces the delay penalty in some cases, but not always. Figure 19 plots the mean RDP values of ESM versus different UDB values on the 3 topologies TS1, TS2, and TS3 for 1000 members. In the figure, the mean RDP values decrease as UDB increases for ESM-TS1 and ESM-TS3 (except for an increase between UDB=10 and UDB=15). The mean RDP stabilizes beyond a certain UDB value. Increasing UDB generally helps a member find the best paths in terms of delay penalty and bandwidth. However, due to the discretized levels of available bandwidth used as a primary metric in ESM, changes using different UDB values are not substantial. A higher UDB, of course, increases the volume of routing information exchanged, which is detrimental to scalability. We have observed no significant change in the mean bandwidth, total stress, or maximum stress, with higher UDB values. The parameter choices for TAG, most significantly bwthresh (which should be tuned according to application bandwidth require- ments), significantly affect the results. For example, setting bwthresh to zero and using complete path matching dramatically improves the RDP values for TAG, at the expense of the maximum stress and bandwidth results, as discussed in section 2.5. 4.2.2 AS and Inet Topologies Table 2: Performance of TAG and ESM in AS and Inet Configuration Algorithm Mean RDP Total link stress Max. stress Mean parent-child bandwidth (kbps) Table 2 shows the performance of TAG and ESM on AS97, AS98, Inet97, and Inet98. We run three different versions of TAG with respect to bwthresh. TAG-50, TAG-100, and TAG-200 denote TAG with TAG-50 gives lower mean RDP than ESM over all configura- tions. In contrast, the mean RDPs of TAG-100 and TAG-200 are similar to, or even worse than, the mean RDP of ESM. All the TAG configurations exhibit lower mean parent-child bandwidth than ESM (in these simulations, we measure parent-child, not sender to re- ceiver, bottleneck bandwidth). Among the TAG configurations, achieves higher mean bandwidth than TAG-100, which gives higher mean bandwidth than TAG-50. Since TAG only considers bandwidth as a secondary metric, it does not consider band-width in its primary tree construction choices. This result shows that bwthresh in TAG must be chosen carefully. A tradeoff between RDP and bandwidth is clearly observed. In addition, note that fanout of nodes in AS-level topologies is higher than that in router-level topologies. The minus-k path matching algorithm in TAG increases the RDP of the nodes which can no longer take a high-degree node as a parent. The available bandwidth at the high-degree node is reduced by the possibly large number of children. TAG with a bwthresh of zero dramatically reduces the RDP (to 1.5 for AS97 and 1.6 for Inet97) at the expense of significant decrease in mean bandwidth. Total link stress values do not widely vary for TAG-50, TAG- 100, TAG-200, and ESM. However, the maximum stress of TAG- 50 is higher than the maximum stress of ESM on AS97 and Inet97. The maximum stress of TAG-50 and ESM on AS98 and Inet98 are similar. With a small bwthresh for TAG-50, TAG allows a node to have a large number of children, which results in a high maximum stress. The maximum stress decreases from TAG-50 to TAG-100 to TAG-200. As previously discussed, running the same simulations for TAG with larger bwthresh values reduces the maximum stress. 4.3 Economies of Scale Factor We compute overlay tree cost via simulations, in order to validate our analytical results from section 3.3. In order to compare results, we assume that one hop used by one point-to-point transfer represents a unit of bandwidth. We therefore add the total stress values for all router-to-router links, and use this quantity to denote101000100000 Normalized Overlay Tree Cost Number of Members Overlay Tree Cost Figure 20: Overlay tree cost versus group size (log-log scale) tree cost. We run three sets of simulations for unicast, TAG, and IP multicast on the TS2 configuration with 1280 end systems. The complete (not minus-k) path matching TAG is used. This is done to give a fair comparison of simulation results with the analytical results, which modeled complete path matching. The simulation results show that unicast, TAG, and IP multicast cost 16627, 4574, and 1265 respectively. We also plot the normalized overlay tree cost of TAG for a variety of group sizes (using the same methodology as in [13]) in figure 20. The normalized overlay tree cost LTAG (m)=^u is defined as in section 3.3. The figure shows that . The overlay tree cost stabilizes with tree saturation, as with IP multicast. This is consistent with our modeling results. 5. RELATED WORK End System Multicast (or Narada) [11, 12] is a clever overlay multicast protocol targeting sparse groups, such as audio and video conferencing groups. End systems in End System Multicast (ESM) exchange group membership information and routing information, build a mesh, and finally run a DVMRP-like protocol to construct a multicast forwarding tree. The authors show that it is important to consider both bandwidth (primarily) and latency, when constructing conferencing overlays. Other application-level multicast architectures include ScatterCast [9], Yoid [18], ALMI [32]. These architectures either optimize delay or optimize bandwidth. In par- ticular, Overcast [24] provides scalable and reliable single-source overlay multicast using bandwidth as a primary metric. More recently, Content-Addressable Network (CAN)-based multicast [34] was proposed to partition member nodes into bins using proximity information obtained from DNS and delay measure- ments. Node degree constraints and diameter bounds in the constructed overlay multicast network are employed in [39]. Liebeherr et al. investigate Delaunay triangulations for routing in overlay networks in [28]. A prefix-based overlay routing protocol is used in Bayeux [42]. Hierarchical approaches to improve scalability are also currently being investigated by several researchers. A protocol that was theoretically proven to build low diameter and low degree peer-to-peer networks was recently described in [30]. In addition to overlay multicast proposals, several recent studies are related to the TAG approach. A unicast-based protocol for multicast with limited router support (that includes some ideas that inspired TAG) is the REUNITE protocol [40]. Overlay networks that detect performance degradation of current routing paths and re-route through other end systems include Detour and RON [3]. Jagannathan and Almeroth [23] propose an algorithm which uses multicast tree topology information (similar to the manner in which we exploit path information in TAG) and loss reports from receivers for multicast congestion control. 6. CONCLUSIONS AND FUTURE WORK We have designed and studied a heuristic topology-aware application-level multicast protocol called TAG. TAG is single-source or core-based multicast protocol that uses network topology information to construct an overlay network with low delay penalty and a limited number of identical packets. Bandwidth is also considered in tree construction as a secondary metric. TAG, however, works best with high quality underlying routes, and assumes low delay on the last hop(s) to end systems. We have studied the properties of TAG, and analyzed its economies of scale factor, compared to both unicast and IP multicast. Simulation results on the Transit-Stub model (GT- ITM), Inet, and NLANR data indicate the effectiveness of TAG in building efficient trees for a large number of group members. We are currently extending TAG to incorporate a tight bandwidth constraint, and delay constrains. With dynamically varying values of the path deviation parameter k and the bandwidth threshold bwthresh, a new member can find a better parent, in terms of both latency and bandwidth. We are also considering a hierarchical approach for increasing adaptivity and scalability. This includes using partial topology in a subsequence matching algorithm. We will extend TAG to include other QoS parameters such as power availability in wireless nodes. In addition, we will incorporate TAG into two different applications (a multi-player online game and a video streaming application), and conduct experiments for evaluating the practical aspects and performance of a TAG implementation in the Internet. 7. ACKNOWLEDGMENTS The authors would like to thank the NOSSDAV 2002 reviewers for their valuable comments that helped improve the paper. This research is sponsored in part by the Purdue Research Foundation, and the Schlumberger Foundation technical merit award. 8. --R Performance Measurement Tools Taxonomy. Multicast Tree Structure and the Power Law. Resilient Overlay Networks. RATES: A server for MPLS traffic engineering. Monitoring OSPF routing. Core based trees (CBT): An architecture for scalable multicast routing. Towards Capturing Representative AS-Level Internet Topologies An Architecture for Internet Content Distribution as an Infrastructure Service. The Origin of Power Laws in Internet Topologies Revisited. Enabling Conferencing Applications on the Internet using an Overlay Multicast Architecture. A Case for End System Multicast. Pricing Multicast Communications: A Cost-Based Approach Multicast routing in datagram inter-networks and extended LANs Deployment Issues for the IP Multicast Service and Architecture. On Power-Law Relationships of the Internet Topology IP network configuration for traffic engineering. Yoid: Your Own Internet Distribution A Global Internet Host Distance Estimation Service. Estimating Router ICMP Generation Delays. Using Tree Topology for Multicast Congestion Control. Reliable multicasting with an overlay network. Inet: Internet Topology Generator. Internet Routing Instability. A Tool for Measuring Bottleneck Link Bandwidth. Building Low-Diameter P2P Networks ALMI: an Application Level Multicast Infrastructure. Scaling of multicast trees: Comments on the Chuang-Sirbu scaling law Detour: a Case for Informed Internet Routing and Transport. The End-to-End Effects of Internet Path Selection Network Address Translator (NAT)-Friendly Application Design Guidelines An OSPF Topology Server: Design and Evaluation. Routing in Overlay Multicast Networks. REUNITE: A Recursive Unicast Approach to Multicast. How to model an internetwork. An Architecture for Scalable and Fault-tolerant Wide-area Data Dissemination --TR Multicast routing in datagram internetworks and extended LANs Core based trees (CBT) End-to-end routing behavior in the Internet Scaling of multicast trees On power-law relationships of the Internet topology A case for end system multicast (keynote address) Bayeux Enabling conferencing applications on the internet using an overlay muilticast architecture Resilient overlay networks IDMaps Using Tree Topology for Multicast Congestion Control Building Low-Diameter P2P Networks --CTR Sonia Fahmy , Minseok Kwon, Characterizing overlay multicast networks and their costs, IEEE/ACM Transactions on Networking (TON), v.15 n.2, p.373-386, April 2007 K. K. To , Jack Y. B. Lee, Parallel overlays for high data-rate multicast data transfer, Computer Networks: The International Journal of Computer and Telecommunications Networking, v.51 n.1, p.31-42, 17 January 2007 Xing Jin , Kan-Leung Cheng , S.-H. Gary Chan, Scalable island multicast for peer-to-peer streaming, Advances in Multimedia, v.2007 n.1, p.10-10, January 2007 Yongjun Li , James Z. Wang, Cost analysis and optimization for IP multicast group management, Computer Communications, v.30 n.8, p.1721-1730, June, 2007 Yi Cui , Klara Nahrstedt, High-bandwidth routing in dynamic peer-to-peer streaming, Proceedings of the ACM workshop on Advances in peer-to-peer multimedia streaming, November 11-11, 2005, Hilton, Singapore Chao Gui , Prasant Mohapatra, Overlay multicast for MANETs using dynamic virtual mesh, Wireless Networks, v.13 n.1, p.77-91, January 2007 Chiping Tang , Philip K. McKinley, Topology-aware overlay path probing, Computer Communications, v.30 n.9, p.1994-2009, June, 2007 Rongmei Zhang , Y. Charlie Hu, Borg: a hybrid protocol for scalable application-level multicast in peer-to-peer networks, Proceedings of the 13th international workshop on Network and operating systems support for digital audio and video, June 01-03, 2003, Monterey, CA, USA Mojtaba Hosseini , Nicolas D. Georganas, Design of a multi-sender 3D videoconferencing application over an end system multicast protocol, Proceedings of the eleventh ACM international conference on Multimedia, November 02-08, 2003, Berkeley, CA, USA Amit Sehgal , Kenneth L. Calvert , James Griffioen, A flexible concast-based grouping service, Computer Networks: The International Journal of Computer and Telecommunications Networking, v.50 n.14, p.2532-2547, 5 October 2006 Tackling group-to-tree matching in large scale group communications, Computer Networks: The International Journal of Computer and Telecommunications Networking, v.51 n.11, p.3069-3089, August, 2007 Mojtaba Hosseini , Nicolas D. Georganas, End system multicast protocol for collaborative virtual environments, Presence: Teleoperators and Virtual Environments, v.13 n.3, p.263-278, June 2004 Minseok Kwon , Sonia Fahmy, Path-aware overlay multicast, Computer Networks: The International Journal of Computer and Telecommunications Networking, v.47 n.1, p.23-45, 14 January 2005

routing;overlay networks;application level multicast;topology;network

507695

Distributing streaming media content using cooperative networking.

In this paper, we discuss the problem of distributing streaming media content, both live and on-demand, to a large number of hosts in a scalable way. Our work is set in the context of the traditional client-server framework. Specifically, we consider the problem that arises when the server is overwhelmed by the volume of requests from its clients. As a solution, we propose Cooperative Networking (CoopNet), where clients cooperate to distribute content, thereby alleviating the load on the server. We discuss the proposed solution in some detail, pointing out the interesting research issues that arise, and present a preliminary evaluation using traces gathered at a busy news site during the flash crowd that occurred on September 11, 2001.

INTRODUCTION There has been much work in recent years on the topic of content distribution. This work has largely fallen into two cat- egories: (a) infrastructure-based content distribution, and (b) peer-to-peer content distribution. An infrastructure-based content distribution network (CDN) (e.g., Akamai) complements the server in the traditional client-server framework. It employs a dedicated set of machines to store and distribute content to clients on behalf of the server. The dedicated in- frastructure, including machines and networks links, is engineered to provide a high level of performance guarantees. On the other hand, peer-to-peer content distribution relies on clients to host content and distribute it to other clients. The P2P model replaces rather than complements the client-server framework. Typically, there is no central server that holds content. Examples of P2P content distribution systems include Napster and Gnutella. In this paper, we discuss Cooperative Networking (Coop- Net), an approach to content distribution that combines aspects of infrastructure-based and peer-to-peer content distri- bution. Our focus is on distributing streaming media content, both live and on-demand. Like infrastructure-based content distribution, we seek to complement rather than replace the traditional client-server framework. Specically, we consider the problem that arises when the server is overwhelmed by the volume of requests from its clients. For instance, a news site may be overwhelmed because of a large \ ash crowd" caused by an event of widespread interest, such as a sports event or an earthquake. A home computer that is webcasting a birthday For more information, please visit the CoopNet project Web page at http://www.research.microsoft.com/ e padmanab/pro- jects/CoopNet/. party live to friends and family might be overwhelmed even by a small number of clients because of its limited network bandwidth. In fact, the large volume of data and the relatively high bandwidth requirement associated with streaming media content increases the likelihood of the server being overwhelmed in general. Server overload can cause signicant degradation in the quality of the streaming media content received by clients. CoopNet addresses this problem by having clients cooperate with each other to distribute content, thereby alleviating the load on the server. In the case of on-demand content, clients cache audio/video clips that they viewed in the recent past. During a period of overload, the server redirects new clients to other clients that had downloaded the content previously. In the case of live streaming, the clients form a distribution tree rooted at the server. Clients that receive streaming content from the server in turn stream it out to one or more of their peers. The key distinction between CoopNet and pure P2P systems like Gnutella is that CoopNet complements rather than replaces the client-server framework of the Web. There is still a server that hosts content and (directly) serves it to clients. CoopNet is only invoked when the server is unable to handle the load imposed by clients. The presence of a central server simplies the task of locating content. In contrast, searching for content in a pure P2P system entails an often more expensive distributed search [20, 21, 24]. Individual clients may only participate in CoopNet for a short period of time, say just a few minutes, which is in contrast to the much longer participation times reported for systems such as Napster and Gnutella [23]. For instance, in the case of live streaming, a client may tune in for a few minutes during which time it may be willing to help distribute the con- tent. Once the client tunes out, it may no longer be willing to participate in CoopNet. This calls for a content distribution mechanism that is robust against interruptions caused by the frequent joining and leaving of individual peers. To address this problem, CoopNet employs multiple description coding (MDC). The streaming media content, whether live or on-demand, is divided into multiple sub-streams using MDC and each sub-stream is delivered to the requesting client via a dierent peer. This improves robustness and also helps balance load amongst peers. The rest of this paper is organized as follows. In Section 2, we discuss related work. In Section 3, we discuss the operation of CoopNet for live and on-demand content, and present an outline of multiple description coding. In Section 4, we use traces from the ash crowd that occurred on September 11, 2001 to evaluate how well CoopNet would have performed for live and on-demand content. We present our conclusions in Section 5. 2. RELATED WORK As noted in Section 1, two areas of related work are infrastructure-based CDNs and peer-to-peer systems. Infrastructure-based CDNs such as Akamai employ a dedicated network of thousands of machines in distributed locations, often with leased links inter-connecting them, to serve content on behalf of servers. When a client request arrives (be it for streaming media or other content), the CDN redirects the client to a nearby replica server. The main limitation of infrastructure-based CDNs is that their cost and scale is only appropriate for large commercial sites such as CNN and MSNBC. A second issue is that it is unclear how such a CDN would fare in the face of a large ash crowd that causes a simultaneous spike in tra-c at many or all of the sites hosted by the CDN. Peer-to-peer systems such as Napster and Gnutella depend on little or no dedicated infrastructure 1 . There is, however, the implicit assumption that the individual peers participate for a signicant length of time (for instance, [23] reports a median session duration of about an hour both for Napster and for Gnutella). In contrast, CoopNet seeks to operate in a highly dynamic situation such as a ash crowd where an individual client may only participate for a few minutes. The disruption that this might cause is especially challenging for streaming media compared to static le downloads, which is the primary focus of Napster and Gnutella. The short life-time of the individual nodes poses a challenge to distributed search schemes such as CAN [20], Chord [24], Pastry [21], and Tapestry [29]. Work on application-level multicast (e.g., ALMI [17], End System Multicast [3], Scattercast [2]) is directly relevant to the live streaming aspect of CoopNet. CoopNet could benet from the e-cient tree construction algorithms developed in previous work. Our focus here, however, is on using real traces to evaluate the e-cacy of CoopNet. Thus we view our work as complementing existing work on application-level multicast. We also consider the on-demand streaming case, which does not quite t in the application-level multicast framework. Existing work on distributed streaming (e.g., [13]) is also directly relevant to CoopNet. A key distinction of our work is that we focus on the distruption and packet loss caused by node arrivals and departures, which is likely to be signi- cant in a highly dynamic environment. Using traces from the September 11 ash crowd, we are able to evaluate this issue in a realistic setting. Systems such as SpreadIt [5], Allcast [31] and vTrails [33] are perhaps closest in spirit to our work. Like CoopNet, they attempt to deliver streaming content using a peer-to-peer ap- proach. SpreadIt diers from CoopNet is a couple of ways. First, it uses only a single distribution tree and hence is vulnerable to disruptions due to node departures. Second, the tree management algorithm is such that the nodes orphaned by the departure of their parent might be bounced around between multiple potential parents before settling on a new parent. In contrast, CoopNet uses a centralized protocol (Sec- tion 3.3), which enables much quicker repairs. It is hard for us to do a specic comparison with Allcast 1 Napster has central servers, but these only hold indices, not content. and vTrails, in the absence of published information. 3. COOPERATIVE NETWORKING (COOPNET) In this section, we present the details of CoopNet as it applies to the distribution of streaming media content. We rst consider the live streaming case, where we discuss and analyze multiple description coding (MDC) and distribution tree management. We then turn to the on-demand streaming case. 3.1 Live Streaming Live streaming refers to the synchronized distribution of streaming media content to one or more clients. (The content itself may either be truly live or pre-recorded.) Therefore multicast is a natural paradigm for distributing such content. Since IP multicast is not widely deployed, especially at the inter-domain level, CoopNet uses application-level multicast instead. A distribution tree rooted at the server is formed, with clients as its members. Each node in the tree transmits the received stream to each of its children using unicast. The out-degree of each node is constrained by the available outgoing bandwidth at the node. In general, the degree of the root node (i.e., the server) is likely to be much larger than that of the other nodes because the server is likely to have a much higher bandwidth than the individual client nodes. One issue is that the peers in CoopNet are far from being dedicated servers. Their ability and willingness to participate in CoopNet may uctuate with time. For instance, a client's participation may terminate when the user tunes out of the live stream. In fact, even while the user is tuned in to the live stream, CoopNet-related activity on his/her machine may be scaled down or stopped immediately when the user initiates other, unrelated network communication. Machines can also crash or become disconnected from the network. With a single distribution tree, the departure or reduced availability of a node has a severe impact on its descendants. The descendants may receive no stream at all until the tree has been repaired. This is especially problematic because node arrivals and departures may be quite frequent in ash crowd situations. To reduce the disruption caused by node departures, we advocate having multiple distribution trees spanning a given set of nodes and transmitting a dierent MDC description down each tree. This would diminish the chances of a node losing the entire stream (even temporarily) because of the departure of another node. We discuss this further in Section 3.2. The distribution trees need to be constantly maintained as new clients join and existing ones leave. In Section 3.3, we advocate a centralized approach to tree management, which exploits the availability of a resourceful server node, coupled with client cooperation, to greatly simplify the problem. 3.2 Multiple Description Coding (MDC) Multiple description coding is a method of encoding the audio and/or video signal into M > 1 separate streams, or descriptions, such that any subset of these descriptions can be received and decoded into a signal with distortion (with respect to the original signal) commensurate with the number of descriptions received; that is, the more descriptions re- ceived, the lower the distortion (i.e., the higher the quality) of encoder decoder encoder decoder base layer Figure 1: (a) Multiple description coding. (b) Layered coding. Bits Distortion R Packet 3 Packet 4 Packet M RS Bit stream Figure 2: Priority encoded packetization of a group of frames (GOF). Any m out of M packets can recover the initial Rm bits of the bit stream for the GOF. the reconstructed signal. This diers from layered coding 2 in that in MDC every subset of descriptions must be decodable, whereas in layered coding only a nested sequence of subsets must be decodable, as illustrated in Figure 1. For this extra exibility, MDC incurs a modest performance penalty relative to layered coding, which in turn incurs a slight performance penalty relative to single description coding. A simple MDC system for video might be the following. The original video picture sequence is demultiplexed into M subsequences, by putting every Mth picture, m into the mth subsequence, . The subsequences are independently encoded to form the M de- scriptions. Any subset of these M descriptions can be decoded and the pictures can be remultiplexed to reconstruct a video sequence whose frame rate is essentially proportional to the number of descriptions received. More sophisticated forms of multiple description coding have been investigated over the years; some highlights are [25, 26, 27, 6]. For an overview see [7]. A particularly e-cient and practical system is based on layered audio or video coding [18, Reed-Solomon coding [28], priority encoded transmission [1], and optimized bit allocation [4, 19, 11, 12]. In such a system the audio and/or video signal is partitioned into groups of frames (GOFs), each group having duration or so. Each GOF is then independently encoded, error protected, and packetized into M packets, as shown in Figure 2. If any m M packets are received, then the initial Rm bits of the bit stream for the GOF can be recovered, re- Layered coding is also known as embedded, progressive, or scalable coding. description description description GOF i-1 GOF i+1 . Figure 3: Construction of MDC streams from packetized GOFs. sulting in distortion D(Rm ), where and consequently D(R0 ) D(R1 ) D(RM ). Thus all M packets are equally important; only the number of received packets determines the reconstruction quality of the GOF. Further, the expected distortion is where p(m) is the probability that m out of M packets are re- ceived. Given p(m) and the operational distortion-rate function D(R), this expected distortion can be minimized using a simple procedure that adjusts the rate points subject to a constraint on the packet length [4, 19, 11, 12]. By sending the mth packet in each GOF to the mth descrip- tion, the entire audio and/or video signal is represented by M descriptions, where each description is a sequence of packets transmitted at rate 1 packet per GOF, as illustrated in Figure 3. It is a very simple matter to generate these optimized M descriptions on the y, assuming that the signal is already coded with a layered codec. 3.2.1 CoopNet Analysis: Quality During Multiple Failures Let us consider how multiple description coding achieves robustness in CoopNet. Suppose that the server encodes its AV signal into M descriptions as described above, and transmits the descriptions down M dierent distribution trees, each rooted at the server. Each of the distribution trees conveys its description to all N destination hosts. Ordinarily, all destination hosts receive all M descriptions. However, if any of the destination hosts fail (or leave the session), then all of the hosts that are descendents of the failed hosts in the mth distribution tree will not receive the mth descrip- tion. The number of descriptions that a particular host will receive depends on its location in each tree relative to the failed hosts. Specically, a host n will receive the mth description if none of its ancestors in the mth tree fail. This happens with probability (1 ) An , where An is the number of the host's ancestors and is the probability that a host fails (assuming independent failures). If hosts are placed at random sites in each tree, then the unconditional probability that any given host will receive its mth description is the average hosts in the tree. Thus the number of descriptions that a particular host will receive is randomly distributed according to a Binomial(M; N ) distri- bution, i.e., . Hence for large M , the fraction of descriptions received is approximately Gaussian with mean N and variance N (1 N ). This can be seen in Figure 4, which shows (in bars) the distribution p(m) for various values of In the gure, to compute N we assumed balanced binary trees with N nodes and probability of host failure that as N grows large, performance slowly degrades, because the depth of the tree (and hence 1 N ) grows like log 2 N . The distribution p(m) can be used to optimize the multiple description code by choosing the rate points R0 ; Figure 4: SNR in dB (line) and probabililty distribution (bars) as a function of the number of descriptions received, when the probability of host failure is to minimize the expected distortion P M subject to a packet length constraint. Figure 4 shows (in lines), the quality associated with each p(m), measured as SNR in as a function of the number of received descriptions, . In the gure, to compute the rate points R0 ; we assumed an operational distortion-rate function which is asymptotically typical for any source with variance 2 , where R is expressed in bits per symbol, and we assumed a packet length constraint given as 3.2.2 CoopNet Analysis: Quality During Single Failure The time it takes to repair the trees is called the repair time. If of the hosts fail during each repair time, then the average length of time that a host participates in the session is 1= repair times. When the number of hosts is small compared to 1=, then many repair times may pass between single failures. In this case, most of the time all hosts receive all descriptions, and quality is excellent. Degradation occurs only when a single host fails. Thus, it may be preferable to optimize the MDC system by minimizing the distortion expected during the repair interval in which the single host fails, rather than minimizing the expected distortion over all time. To analyze this case, suppose that a single host fails randomly. A remaining host n will not receive the mth description if the failed host is an ancestor of host n in the mth tree. This happens with probability An=(N 1), where An is the number of ancestors of host n. Since hosts are place at random sites in each tree, the unconditional probability that any given host will receive its mth description is the average 1)). Thus the number of descriptions that a particular host will receive is randomly distributed according to a Binomial(M; N ) distri- bution. Equivalently, the expected number of hosts that receive descriptions during the failure is (N 1)p(m), where This distribution can be used to optimize the multiple description code for the failure of a single host. Figure 5 illustrates this distribution and the corresponding optimized quality as a function of the number of descriptions received, for Note that as M increases, for xed N , the distribution again becomes Gaussian. One implication of this is that the expected number of hosts that receive 100% of the descriptions Figure 5: SNR in dB (line) and probabililty distribution (bars) as a function of the number of descriptions received during the failure of a single host. decreases. However it is also the case that the expected number of hosts that receive fewer than 50% of the descriptions decreases, resulting in an increase in quality on average. Fur- ther, as N increases, for xed M , performance becomes nearly perfect, since N 1 log 2 N=N , which goes to 1. However, for large N , it becomes increasingly di-cult to repair the trees before a second failure occurs. 3.2.3 Further Analyses These same analyses can be extended to d-ary trees. It is not di-cult to see that for d 2, a d-ary trees with log 2 d N nodes has the same height, and hence the same performance, as a binary tree with only N nodes. Thus when each node has a large out-degree, i.e., when each host has a large uplink bandwidth, much larger populations can be han- dled. Interestingly, the analysis also applies when if each host can devote only as much uplink bandwidth as its downlink video bandwidth (which is typically the case for modem users), then the descriptions can still be distributed peer-to-peer by arranging the hosts in a chain, like a bucket brigade. It can be shown that when the order of the hosts in the chain is random and independent for each description, then for a single failure the number of hosts receiving m out of M descriptions is binomially distributed with parameters M and N , where . Although this holds for any N , it is most suitable for smaller N . For larger N , it may not be possible to repair the chains before other failures occur. In fact, as N goes to innity, the probability that any host receives any descriptions goes to zero. In this section we have proposed optimizing the MDC system to the unconditional distribution p(m) derived by averaging over trees and hosts. Given any set of trees, however, the distribution of the number of received descriptions varies widely across the set of hosts as a function of their upstream connectivity. By optimizing the MDC system to the unconditional distribution p(m), we are not minimizing the expected distortion for any given host, but rather minimizing the sum of the expected distortions across all hosts, or equivalently, minimizing the expected sum of the distortions over all hosts. 3.3 Tree Management We now discuss the problem of constructing and maintaining the distribution trees in the face of frequent node arrivals and departures. There are many (sometimes con for the tree management algorithm: 1. Short and wide tree: The trees should be as short as possible so as to minimize the latency of the path from the root to the deepest leaf node and to minimize the probability of disruption due to the departure of an ancestor node. For it to be short, the tree should be balanced and as wide as possible, i.e., the out-degree of each node should be as much as its bandwidth will allow. However, making the out-degree large may leave little bandwidth for non-CoopNet (and higher priority) tra-c emanating from the node. Interference due to such tra-c could cause a high packet loss rate for the CoopNet streams. 2. E-ciency versus tree diversity: The distribution trees should be e-cient in that their structure should closely re ect the underlying network topology. So, for instance, if we wish to connect three nodes, one each located in New York (NY), San Francisco (SF), and Los Angeles (LA), the structure NY!SF!LA would likely be far more e-cient than SF!NY!LA (! denotes a parent-child relationship). However, striving for e-ciency may interfere with the equally important goal of having diverse distribution trees. The eectiveness of MDC-based distribution scheme described in Section 3.2 depends critically on the diversity of the distribution trees. 3. Quick join and leave: The processing of node joins and leaves should be quick. This would ensure that the interested nodes would receive the streaming content as quickly as possible (in the case of a join) and with minimal interruption (in the case of a leave). However, the quick processing of joins and leaves may interfere with the e-ciency and balanced tree goals listed above. 4. Scalability: The tree management algorithm should scale to a large number of nodes, with a correspondingly high rate of node arrivals and departures. For instance, in the extreme case of the ash crowd at MSNBC on September 11, the average rate of node arrivals and de- parturtes was 180 per second while the peak rate was about 1000 per second. With these requirements in mind, we now describe our approach to tree construction and management. We rst describe the basic protocol and then discuss optimizations. 3.3.1 Basic Protocol We exploit the presence of a resourceful server node to build a simple and e-cient protocol to process node joins and leaves. While it is centralized, we argue that this protocol can scale to work well in the face of extreme ash crowd situations such as the one that occurred on September 11. Despite the ash crowd, the server is not overloaded since the burden of distributing content is shared by all peers. Centralization also simplies the protocol greatly, and consequently makes joins and leaves quick. In general, a criticism of centralization is that it introduces a single point of failure. However, in the context of CoopNet, the point of centralization is the server, which is also the source of data. If the source (server) fails, it may not really matter that the tree management also breaks down. Also, recall from Section 1 that the goal of CoopNet is to complement, not replace, the client-server system. The server has full knowledge of the topology of all of the distribution trees. When a new node wishes to join the sys- tem, it rst contacts the server. The new node also informs the server of its available network bandwidth to serve furture downstream nodes. The server responds with a list of designated parent nodes, one per distribution tree. The designated parent node in each tree is chosen as follows. Starting at the server, we work our way down the tree until we get to a level where there are one or more nodes that have the necessary spare capacity (primarily network bandwidth) to serve as the parent of the new node. (The server could itself be the new parent if it has su-cient spare capacity, which it is likely to have during the early stages of tree construction.) The server then picks one such node at random to be the designated parent of the new node. This top-down procedure ensures a short and largely balanced tree. The randomization helps make the trees diverse. Upon receiving the server's message, the new node sends (concurrent) messages to the designated parent nodes to get linked up as a child in each distribution tree. In terms of messaging costs, the server receives one message and sends one. Each designated parent receives one message and sends one (an acknowledgement). The new node sends and receives is the number of MDC descriptions (and hence distribution trees) used. Node departures are of two kinds: graceful departures and node failures. In the former case, the departing node informs the server of its intention to leave. For each distribution tree, the server identies the children of the departing node and executes a join operation on each child (and implicitly the subtree rooted at the child) using the top-down procedure described above. The messaging cost for the server would at most be P receives, where d i is the number of children of the departing node in the ith distribution tree. (Note that the cost would be somewhat lower in general because a few of the children may be in common across multiple trees.) Each child sends and receives M messages. To reduce its messaging load, the server could make the determination of the designated parent for each child in each tree and then leave it to another node (such as the departing node, if it is still available) to convey the information to each child. In this case, the server would have to send and receive just one message. A node failure corresponds to the case where the departing node leaves suddenly and is unable to notify either the server or any other node of its departure. This may happen because of a computer crashing, being turned o, or becoming disconnected from the network. We present a general approach for dealing with quality degradation due to packet loss; node failure is a special case where the packet loss rate experienced by the descendants of the failed node is 100%. Each node monitors the packet loss rate it is experiencing in each distribution tree. When the packet loss rate reaches an unacceptable level (a threshold that needs to be ne-tuned based on further re- search), a node contacts its parent to check if the parent is experiencing the same problem. If so, the source of the problem (network congestion, node failure, etc.) is upstream of the parent and the node leaves it to the parent to deal with it. (The node also sets a su-ciently long timer to take action on its own in case its parent has not resolved the problem within a reasonable period of time.) If the parent is not experiencing a problem or it does not respond, the aected node will contact the server and execute a fresh join operation for it (and its subtree) to be moved to a new location in the distribution tree. 3.3.2 Optimizations We now discuss a few optimizations of the basic protocol. The rst optimization seeks to make the distribution trees e-cient, as discussed above. The basic idea here is to preferentially attach a new node as the child of an existing node that is \nearby" in terms of network distance (i.e., latency). The denition of \nearby" needs to be broad enough to accomodate signicant tree diversity. When trying to insert a new node, the server rst identies a (su-ciently large) sub-set of nodes that are close to the new node. Then using the randomized top-down procedure discussed in Section 3.3.1, it tries to nd a parent for the new node (in each tree) among the set of nearby nodes. Using this procedure, it is quite likely that many of the parents of the new node (on the the various distribution trees) will be in the same vicinity, which is ben- ecial from an e-ciency viewpoint. We argue that this also provides su-cient diversity since the primary failure mode we are concerned with is node departures and node failures. So it does not matter much that all of the parents may be located in the same vicinity (e.g., same metropolitan area). To determine the network distance between two nodes, we use a procedure based on previous work on network distance estimation [14], geographic location estimation [16], overlay construction [20], and nding nearby hosts [8]. Each node determines its network \coodinates" by measuring the network latency (say using ping) to a set of landmark hosts (about well-distributed landmark hosts should su-ce in practice). The coordinate of a node is the n-tuple (d1 n is the number of landmarks. The server keeps track of the coordinates of all nodes currently in the system (this information may need to be updated from time to time). When the new node contacts it, the server nds nearby nodes by comparing the coordinates of the new node with those of existing nodes. This comparison could involve computing the Eucledian distance between the coordinates of two nodes (as in [16]), computing a dierent distance metric such as the Manhattan distance, or simply comparing the relative ordering the various landmarks based on the measured latency (as in [20]). The second optimization is motivated by the observation that it would be benecial to have have more \stable" nodes close to the root of the tree. In this context, \stable" nodes are ones that are likely to participate on CoopNet for a long duration and have good network connectivity (e.g., few dis- truptions due to competing tra-c from other applications). Having such nodes close to the root of the tree would benet their many descendants. As a background process, the server could identify stable nodes by monitoring their past behavior and migrate them up the tree. Further research is needed to determine the feasibility of identifying stable nodes, the ben- ets of migrating such nodes up the tree, and the impact this might have on tree diversity. 3.3.3 Feasibility of the Centralized Protocol The main question regarding the feasibility of the centralized tree management protocol is whether the server can keep up. To answer this question, we consider the September 11 ash crowd at MSNBC, arguably an extreme ash crowd sit- uation. At its peak, there were 18,000 nodes in the system and the rate of node arrivals and departures was 1000 per second. 3 (The average numbers were 10000 nodes and 180 arrivals and departures per second.) In our calculations here, we assume that the number of distribution trees (i.e., the number of MDC descriptions) is 16 and that on average a node has 4 children in a tree. We consider various resources that could become a bottleneck at the server (we only focus on the impact of tree management on the server): Memory: To store the entire topology of one tree in memory, the server would need to store as many pointers as nodes in the system. Assuming a pointer size of 8 bytes (i.e., a 64-bit machine) and auxiliary data of 24 bytes per node, the memory requirement would be about 576 KB. Since there are 16 trees, the memory requirement for all trees would be 9.2 MB. In addition, for each node the server needs to store its network coordinates. Assuming this is a 10-dimensional vector of delay values bytes each), the additional memory requirement would be 360 KB. So the total memory requirement at the server would be under 10 MB, which is a trivial amount for any modern machine. Network departures are more expensive than node arrivals, so we focus on departures. The server needs to designate a new parent in each distribution tree for each child of the departing node. Assuming that nodes are identied by their IP addresses bytes assuming IPv6) and that there are 4 children per tree on average, the total amount of data that the server would need to send out is 1 KB. If there are 1000 departures per second, the bandwidth requirement would be 8 Mbps. This is likely to be a small fraction of the network bandwidth at a large server site such as MSNBC. CPU: Node departure involves nding a new set of parents for each child of the departing node. So the CPU cost is roughly equal to the number of children of the departing node times the cost of node insertion. To insert a node, the server has to scan the tree levels starting with the root until it reaches a level containing one or more nodes with the spare capacity to support a new child. The server picks one such node at random to be the new parent. Using a simple array data structure to keep track of the nodes in each level of the tree that have capacity, the cost of picking a parent at random can be made (a small) constant. Since the number of levels in the tree is about log(N ), where N is the number of nodes in the system, the node insertion cost (per tree) is O(log(N)). (With and an average of 4 children per node, the depth of the tree will be about 9.) A departure rate of 1000 per second would result in 64,000 insertions per second (1000 departures times 4 children per departing node times 16 trees). Given that memory speed by far lags CPU speed, we only focus on how many memory lookups we can do per insertion. Assuming a 40 ns memory cycle, we are allowed about memory accesses per insertion, which is likely to be more than su-cient. 3 One reason for the high rate of churn may be that users were discouraged by the degradation in audio/video quality caused by the ash crowd, and so did not stay for long. However, we are not in a position to conrm that this was the case. In general, the centralized approach can be scaled up (at least in terms of CPU and memory resources) by having a cluster of servers and partitioning the set of clients across the set of server nodes. We are in the process of benchmarking our implementation to conrm the rough calculations made above. 3.3.4 Distributed Protocol While the centralized tree management protocol appears to be adequate for large ash crowd situations such as that experienced by MSNBC on September 11, it is clear that there are limits to its scalability. For instance, in the future it is conceivable that ash crowds for streaming media content on the Web will in some cases be as large as television audiences during highly popular events | hundreds of millions or even billions of clients. A centralized solution may break down in such a situation, necessitating an alternative approach to tree management. We could leverage recent work on distributed hash tables (DHTs), such as CAN [20], Chord [24], Pastry [21], and Tapestry [29], to build construct and maintain the trees in a distributed fashion. Brie y, DHTs provide a scalable unicast routing framework for peer-to-peer systems. A multicast distribution tree can be constructed using reverse-path forwarding (as in systems such as Bayeux [30] and Scribe [22]). To construct multiple (and diverse) distribution trees, each node could be assigned multiple IDs, one per tree. There are a number of open research issues. First, while there exist algorithms to support node joins and leaves, the dynamic behavior of DHTs is poorly understood. Second, it is unclear how to incorporate constraints, such as limited node bandwidth, into the DHT framework. Some systems such as Pastry maintain multiple alternate routes at each hop. This should make it easier to construct multicast trees while accomodating node capacity constraints. 3.4 On-demand Streaming We now turn to on-demand streaming, which refers to the distribution of pre-recorded streaming media content on demand (e.g., when a user clicks on the corresponding link). As such, the streams corresponding to dierent users are not syn- chronized. When the server receives such a request, it starts streaming data in response if its current load condition per- mits. However, if the server is overloaded, say because of a ash crowd, it instead sends back a response including a short list of IP addresses of clients (peers) who have downloaded (part or all of) the requested stream and have expressed a willingness to participate in CoopNet. The requesting client then turns to one or more of these peers to download the desired content. Given the large volume of streaming media content, the burden on the server (in terms of CPU, disk, and network bandwidth) of doing this redirection is quite minimal compared to that of actually serving the content. So we believe that this redirection procedure will help reduce server load by several orders of magnitude. While the procedure described above is similar to one that might apply to static le content, there are a couple of important dierences arising from the streaming nature of the content. First, a peer may only have a part of the requested content because, for instance, the user may have stopped the stream halfway or skipped over portions. So in its initial handshake with a peer, a client nds out which part of the requested content is available at the peer and accordingly plans to make requests to other peers for the missing content, if any. A second issue is that, as with the live streaming case, peers may fail, depart, or scale back their participation in CoopNet at any time. In contrast with le download, the time-sensitive nature of streaming media content makes it especially susceptible to such disruptions. As a solution, we propose the use of distributed streaming where a stream is divided into a number of substreams, each of which may be served by a dierent peer. Each substream corresponds to a description created using MDC (Section 3.2). Distributed streaming improves robustness to disruptions caused by the untimely departure of peer nodes and/or network connectivity problems with respect to one or more peers. It also helps distribute load more evenly among peers. 4. PERFORMANCE EVALUATION We now present a performance evaluation of CoopNet based on simulations driven by traces of live and on-demand content served by MSNBC on September 11, 2001. 4.1 Live Streaming We evaluate the MDC-based live streaming design using traces of a 100kbps live stream. The trace started at 18:25 GMT (14:25 EST) and lasted for more than one hour (4000 seconds). 4.1.1 Trace Characteristics Figure 6 shows the time series of the number of clients simultaneously tuned in to the live stream. The peak number of simultaneous clients exceeds 17,000. On average, there are 84 clients departing every second. (We are unable to denitely explain the dip around the 1000-seond mark, but it is possibly due to a glitch in the logging process.) Figure 7 shows the distribution of client lifetimes. Over 70% of the clients remain tuned in to the live stream for less than a minute. We suspect that the short lifetimes could be because users were frustrated by the poor quality the video stream during the ash crowd. If the quality were improved (say using CoopNet to relieve the server), client lifetimes may well become longer. This, in turn, would increase the eectiveness of CoopNet. 4.1.2 Effectiveness of MDC We evaluate the impact of MDC-based distribution (Sec- tion 3.2) on the quality of the stream received by clients in the face of client departures. When there are no departures, all clients receive all of the MDC descriptions and hence perceive the full quality of the live stream. We have conducted two simulation experiments. In the rst experiment, we construct completely random distribution trees at the end of the repair interval following a client departure. We then analyze the stream quality received by the remaining clients. The random trees are likely to be diverse (i.e., uncorrelated), which improves the eectiveness of MDC-based distribution. In the second experiment, we simulate the tree management algorithm described in Section 3.3. Thus the distribution trees are evolved based on the node arrivals and departures recorded in the trace. We compare the results of these two experiments at the end of the section. In more detail, we conducted the random tree experiment as follows. For each repair interval, we construct M distribution trees (corresponding to the M descriptions of the MDC coder) spanning the N nodes in the system at the beginning of the interval. Based on the number of departing clients, d, Node Arrivals and Departures200060001000014000180000 500 1000 1500 2000 2500 3000 3500 4000 Time (seconds) of Nodes Figure Number of clients and departures. Duration Minutes Percentage of Nodes Figure 7: Duration distribution. Table 1: Random Tree Experiment: probability distribution of descriptions received vs. number of distribution trees recorded through the end of the repair interval, we randomly remove d nodes from the tree, and compute the number of descriptions received by the remaining nodes. The perceived quality of the stream at a client is determined by the fraction of descriptions received by that client. The set of distribution trees is characterized by three parameters: the number of trees (or, equivalently, descriptions), the maximum out-degree of nodes in each tree, and the out-degree of the root (i.e., the live streaming server). The out-degree of a node is typically a function of its bandwidth capacity. So the root (i.e., the server) tends to have a much larger out-degree than bandwidth-constrained clients. In our random tree construc- tion, each client is assigned a random degree subject to a maximum. We varied the degree of the root and the number of descriptions to study their impact on received stream qual- ity. We set the repair time to 1 second; we investigate the impact of repair time in Section 4.1.3. Table shows how the number of distribution trees, M , affects the fraction of descriptions received (expressed as a per- centage, P ). We compute the distribution of P by averaged across all client departures. We set the maximum out-degree Figure 8: Random Tree Experiment: SNR in dB (line) and probabililty distribution (bars) as a function of the number of descriptions received Quality (SNR in Time (seconds) Random Trees Multiple Descriptions (M=16) Single Description (M=1) Figure 9: Random Tree Experiment: The SNR over time for the MDC and SDC cases. At each time in- stant, we compute the average SNR over all clients. of a client to 4 and the root degree to 100. We vary the number of descriptions among 1, 2, 4, 8, or 16. Each column represents a range of values of P . For each pair of the range and number of descriptions, we list the average percentage of clients that receive at that level of quality. For example, the rst table entry indicates that when using 2 descriptions, 94.80% of clients receive 100% of the descriptions (i.e., both the descriptions). As the number of descriptions increases, the percentage of clients that receive the all of the descriptions (i.e., decreases. Nonetheless, the percentage of clients corresponding to small values of P decreases dramatically. With 8 de- scriptions, 96% (82.07% + 14.02%) of clients receive more than 87.5% of the descriptions. For both 8 and tions, all clients receive at least one description. Figure 8 shows the corresponding SNR. Figure 9 compares the SNR over time for the 16-description case and the single description case. MDC demonstrates a clear advantage over SDC. Table shows how the root degree aects the distribution of descriptions received. We set the number of descriptions to 8 and the maximum out-degree of a client to 4. As the root degree increases, the distribution shows an improvement. Figure shows the SNR and probability distribution. Compared to the case where all nodes (including the root) have the same degree d, a root degree of R shortens the tree by log d R. This means fewer ancestors for nodes in the tree, which as discussed in Section 3.2 increases the probability that a node will receive a particular description. R 100% [87.5,100) [75,87.5) [50,75) [25,50) 0 Table 2: Random Tree Experiment: probability distribution of the descriptions received vs. root degree Figure 10: Random Tree Experiment: SNR in dB (line) and probabililty distribution (bars) as a function of the number of descriptions received and the root degree Table 3: Evolving Tree Experiment: probability distribution of descriptions received vs. number of distribution trees In our second experiment, we evolved the distribution trees by simulating the tree management algorithm from Section 3.3. We set the root (i.e., server) out-degree to 100. The maximum out-degree of a client is set to 4. Table 3 shows the probability distribution of the descriptions received upon client departures. Figure 11 shows the corresponding SNR. The results are comparable to those of the random tree ex- periment. This suggests that our tree management algorithm is able to preserve signicant tree diversity even over a long period of time (more than an hour in this case). 4.1.3 Impact of Repair Time Finally, we evaluate the impact of the time it takes to repair the tree following a node departure. Clearly, the longer the repair time, the greater is the impact on the aected nodes. Also, a longer repair time increase the chances of other nodes departing before the repair is completed, thereby causing further disruption. We divide time into non-overlapping repair intervals and assume that all leaves happen at the beginning of each interval. We then compute fraction of descriptions received averaged over all nodes (this is the quantity N discussed in Section 3.2). As in Section 3.2, assume a balanced binary tree at all times. Figure 12 shows the average number of descriptions received as a function of time for four dierent settings of repair time: seconds. With a repair time of 1 second, clients would receive 90% of the descriptions on average. With a 10 second repair time, the fraction drops to 30%. We believe that these results are encouraging since in practice tree repair can be done very quickly, especially given that our tree management algorithm is centralized (Section 3.1). Even a 1-second Figure Evolving Tree Experiment: SNR in dB (line) and probabililty distribution (bars) as a function of the number of descriptions received0.10.30.50.70.90 500 1000 1500 2000 2500 3000 3500 4000 Time (seconds) Average fraction of descriptions received Figure 12: The average fraction of descriptions received for various repair times. repair interval would permit multiple round-trips between the server and the nodes aected by the repair (e.g., the children of the departed node). 4.2 On-Demand Streaming We now evaluate the potential of CoopNet in the case of on-demand streaming. The goals of our evaluation are to study the eects of client cooperation on: reduction in load at the server additional load incurred by cooperating peers amount of storage provided by cooperating peers likelihood of cooperating with proximate peers to improve performance. The cooperation protocol used in our simulations is based on server redirection as in [15]. The server maintains a xed- size list of IP addresses (per URL) of CoopNet clients that have recently contacted it. To get content, a client initially sends a request to the server. If the client is willing to co- operate, the server redirects the request by returning a short list of IP addresses of other CoopNet clients who have recently requested the same le. In turn, the client contacts these other CoopNet peers and arranges to retrieve the content directly from them. Each peer may have a dierent portion of the le, so it may be necessary to contact multiple peers for content. In order to select a peer (or a set of peers when using distributed streaming) to download content from, peers run a greedy algorithm that picks out the peer(s) with the longest portion of the le from the list returned by the server. If a client cannot retrieve content through any peer, it retrieves the entire content from the server. Note that the server only provides the means for discovering other CoopNet peers. Peers independently decide who they cooperate with. The server maintains a list of 100 IP addresses per URL, and returns a list of 10 IP addresses in the redirection messages in our simulations. We use traces collected at MSNBC during the ash crowd of Sep 11, 2001 for our evaluation. The ash crowd started at around 1:00 pm GMT (9:00 am EDT) and persisted for the rest of the day. The peak request rate was three orders of magnitudes more than the average. We report simulation results for the beginning of the ash crowd, between 1:00 pm to 3:00 pm GMT. There were over 300,000 requests during the 2-hour period. However, only 6% or 18,000 requests were Time of Day Average Bandwidth (bps) Server Server (CoopNet) Clients (CoopNet) (a) Average bandwidth at server and cooperating peers. Degree of Parallelism Average Bandwidth of CoopNet Clients (bps) (b) Average bandwidth at peers when using distributed streaming. 100.10.30.50.70.9Bandwidth of Active Peers (bps) Cumulative Distribution (Least Loaded) (c) Distribution of bandwidth at active peers. Figure 13: Performance of CoopNet for on-demand streaming. successfully served at an average rate of 20 Mbps with a mean session duration of 20 minutes. Unsuccessful requests were not used in the analysis because of the lack of content byte- range and session duration information. 4.2.1 Bandwidth Load In our evaluation, load is measured as bandwidth usage. We do not model available bandwidth between peers. We assume that peers can support the full bit rate (56 kbps, 100 kbps) of each encoded stream. We also do not place a bound on the number of concurrent connections at each peer. In practice, nding peers with su-cient available bandwidth and not overloading any one peer are important considerations, and we are investigating these issues in ongoing work. Figure 13(a) depicts the bandwidth usage during the 2-hour period for two systems: the traditional client-server system, and the CoopNet system. The vertical axis is average band-width and the horizontal axis is time. There are two peaks at around 1:40 pm and 2:10 pm, when two new streams were added to the server. In the client-server system, the server was distributing content at an average of 20 Mbps. However, client cooperation can reduce that bandwidth by orders of magnitude to an average of 300 kbps. As a result, the server is available to serve more client requests. The average band-width contribution that CoopNet clients need to make to the system is 45 kbps. Although the average bandwidth contribution is reasonably small, peers are not actively serving content all the time. We nd that typically less than 10% of peers are active at any second. The average bandwidth contribution that active CoopNet peers need to make to the system is as high as 465 kbps, where average bandwidth of active peers is computed as the total number of bits served over the total length of peers' active periods. To further reduce load at individual CoopNet clients, disjoint portions of the content can be retrieved in parallel from multiple peers using distributed streaming (Section 3.4). (The bandwidth requirement placed on each peer is correspondingly reduced.) Figure 13(b) depicts the average bandwidth contributed versus the degree of parallelism. The degree of parallelism is an upper-bound on the number of peers that can be used in parallel. For example, clients can retrieve content from up to 5 peers in parallel in a simulation with a degree of parallelism of 5. The actual number of peers used in parallel may be less than 5 depending on how many peers can provide content in the byte-range needed by the client. The load at each active peer is reduced as the degree of parallelism in- creases. When the degree of parallelism is 5, peers are serving content at only 35 kbps. However, the bandwidth of active peers (not depicted in this gure) is only slightly reduced to 400 kbps. This is because the large amount of bandwidth required to serve content during the two surges at 1:40 pm and 2:10 pm in uence the average bandwidth. The cumulative distribution of bandwidth contributed by active CoopNet peers, depicted in Figure 13(c), illustrates the impact of distributed streaming on bandwidth utiliza- tion. Each solid line represents the amount of bandwidth peers contribute when using 1, 5, and 10 degrees of paral- lelism. The median bandwidth requirement is 390 kbps when content is streamed from one peer, and only 66 bps for degrees of parallelism. The bandwidth requirement imposed on each peer is reduced as the degree of parallelism increases. Although this reduction is signicant, a small portion of peers still contribute more than 1 Mbps even when using 10 degrees of parallelism. We believe that the combination of the following two factors contribute to the wide range in bandwidth usage: the greedy algorithm a client uses to select peers and the algorithm the server uses to select a set of IP addresses to give to clients. For better load distribution, the server can run a load-aware algorithm that redirects clients to recently seen peers that are the least loaded (in terms of network bandwidth usage). In order to implement this algorithm, the server needs to know the load at individual peers. Therefore, peers constantly report their current load status to the server. We use a report interval of once every second in our simulations. Because the server caches a xed-size list of IP addresses, only those peers currently in the server's list need to send status up- dates. Given this information, the server then selects the 10 least loaded peers that have recently accessed the same URL as the requesting client to return in its redirection message. This algorithm replaces the one described earlier in this section where the server redirects clients to peers that were re-0.20.40.60.81 Storage Allocated At Each Peer (Bytes) Cumulative Distribution Figure 14: Storage requirement at CoopNet peers. cently seen. Clients, however, use the same greedy algorithm to select peers. We nd that using this new algorithm, active clients serve content at 385 kbps. The dashed line in Figure 13(c) depicts the cumulative distribution of bandwidth contributed by CoopNet clients when the load-aware algorithm is used at the server. In this simulation, clients stream content from at most one other peer (degree of parallelism of 1). For the most part, the distribution is similar to the one observed when the server redirects the request to recently seen peers. The dierence lies in the tail end of the distribution. About 6% of peers contributed more than 500 kbps of band-width when the server runs the original algorithm, compared to only 2% when the server runs the load-aware algorithm. In addition, the total number of active peers in the system doubles when the load-aware algorithm is used. We nd that client cooperation signicantly reduces server load, freeing up bandwidth to support more client connec- tions. In addition, the combination of distributed streaming and a load-aware algorithm used by the server further reduces the load on individual peers. 4.2.2 Storage Requirement In order to facilitate cooperation, clients also contribute storage for caching content. In our simulations, peers cache streams that they have downloaded for the entire duration of the simulation. Figure 14 depicts the cumulative distribution of the amount of storage each peer needs to provide. Storage sizes range from 200 B to 100 MB. Over half of the peers store less than 1 MB of content, and only 5% of peers store over 6MB of content. The storage requirement is reasonable for modern computers. 4.2.3 Nearby Peers Next, we look at the likelihood of cooperating with nearby peers. Finding nearby peers can greatly increase the e-ciency of peer-to-peer communications. In our evaluation, peers are close if they belong in the same BGP prex domain [9]. We cluster over 9,000 IP addresses of clients who successfully received content in the 2-hour trace based on BGP tables obtained from a BBNPlanet router [32] on Jan 24, 2001. The trace is sampled by randomly drawing ten 5-minute windows. We look the probability of nding at least n peers in the same AS domain, where n is the degree of parallelism, ranging from 1 to 10. The sampling is repeated for window sizes of 10 and minutes. For a window of 5 minutes, the probability of nding at least one peer who has requested the same content and belongs to the same BGP prex cluster is 12%. As the window size increases to 10 and 15 minutes, the probability slightly increases to 16% and 17%, accordingly. For distributed streaming, as the degree of parallelism increases, the probability of nding nearby peers decreases. Using a 10-minute window, the probability of nding at least 5 peers and 10 peers in the same BGP prex cluster are as low as 5% and 2%. To better understand whether the small number of IP addresses aects the probabilities of nding proximate peers, we also clustered over 90,000 IP addresses from the entire 2-hour trace, including unsuccessful requests. For the most part, the probabilities are the same or 1-2% higher than those those reported above for successful requests. Finding a proximate peer with su-cient available bandwidth is part of ongoing work. In summary, our initial results suggest that client cooperation can improve overall system performance. Distributed streaming and load-aware server are promising solutions to reduce load at individual peers while improving robustness. 5. CONCLUSIONS In this paper, we have presented CoopNet, a peer-to-peer content distribution scheme that helps servers tide over crisis situations such as ash crowds. We have focussed on the application of CoopNet to the distribution of streaming median content, both live and on-demand. One challenge is that clients may not participate in CoopNet for an extended length of time. CoopNet employs distributed streaming and multiple description coding to improve the robustness of the distributed streaming content in face of client departures. We have evaluated the feasibility and potential performance of CoopNet using traces gathered at MSNBC during the ash crowd that occurred on September 11, 2001. This was an extreme event even by ash crowd standards, so using these traces helps us stress test the CoopNet design. Our results suggest that CoopNet is able to reduce server load significantly without placing an unreasonable burden on clients. For live streams, using multiple independent distribution trees coupled with MDC improves robustness signicantly. We are currently building a prototype implementation of CoopNet for streaming media distribution. Acknowledgements We are grateful to Steven Lautenschlager, Ted McConville, and Dave Roth for providing us the MSNBC streaming media logs from September 11. We would also like to thank the anonymous NOSSDAV reviewers for their insightful comments 6. --R Joint source and channel coding for image transmission over lossy packet Multiple description coding: Compression meets the network. Unequal loss protection: Graceful degradation of image quality over packet erasure channels through forward error correction. Approximately optimal assignment for unequal loss protection. Embedded video subband coding with 3D SPIHT. Multiple description source coding through forward error correction codes. Design of multiple description scalar quantizers. Design of entropy-constrained multiple description scalar quantizers Optimal pairwise correlating transforms for multiple description coding. Control Systems for Digital Communication and Storage. BBNPlanet publically available route server --TR systems for digital communication and storage Enabling conferencing applications on the internet using an overlay muilticast architecture Chord A scalable content-addressable network An investigation of geographic mapping techniques for internet hosts Storage management and caching in PAST, a large-scale, persistent peer-to-peer storage utility Towards global network positioning The Case for Cooperative Networking Tapestry: An Infrastructure for Fault-tolerant Wide-area Location and Scattercast --CTR Xinyan Zhang , Jiangchuan Liu, Gossip based streaming, Proceedings of the 13th international World Wide Web conference on Alternate track papers & posters, May 19-21, 2004, New York, NY, USA Duc A. Tran , Kien A. Hua , Tai T. Do, Scalable media streaming in large peer-to-peer networks, Proceedings of the tenth ACM international conference on Multimedia, December 01-06, 2002, Juan-les-Pins, France Mubashar Mushtaq , Toufik Ahmed , Djamal-Eddine Meddour, Adaptive packet video streaming over P2P networks, Proceedings of the 1st international conference on Scalable information systems, May 30-June 01, 2006, Hong Kong Meng Zhang , Li Zhao , Yun Tang , Jian-Guang Luo , Shi-Qiang Yang, Large-scale live media streaming over peer-to-peer networks through global internet, Proceedings of the ACM workshop on Advances in peer-to-peer multimedia streaming, November 11-11, 2005, Hilton, Singapore Guang Tan , Stephen A. Jarvis , Xinuo Chen , Daniel P. Spooner, Performance Analysis and Improvement of Overlay Construction for Peer-to-Peer Live Streaming, Simulation, v.82 n.2, p.93-106, February 2006 Reza Rejaie , Antonio Ortega, PALS: peer-to-peer adaptive layered streaming, Proceedings of the 13th international workshop on Network and operating systems support for digital audio and video, June 01-03, 2003, Monterey, CA, USA Karthik Lakshminarayanan , Venkata N. Padmanabhan, Some findings on the network performance of broadband hosts, Proceedings of the 3rd ACM SIGCOMM conference on Internet measurement, October 27-29, 2003, Miami Beach, FL, USA Chuan Wu , Baochun Li, rStream: resilient peer-to-peer streaming with rateless codes, Proceedings of the 13th annual ACM international conference on Multimedia, November 06-11, 2005, Hilton, Singapore Zongpeng Li , Baochun Li , Lap Chi Lau, On achieving maximum multicast throughput in undirected networks, IEEE/ACM Transactions on Networking (TON), v.14 n.SI, p.2467-2485, June 2006 Sachin Agarwal , Jatinder Pal Singh , Shruti Dube, Analysis and implementation of Gossip-based P2P streaming with distributed incentive mechanisms for peer cooperation, Advances in Multimedia, v.2007 n.2, p.1-12, April 2007 Song Ye , Fillia Makedon, Collaboration-aware peer-to-peer media streaming, Proceedings of the 12th annual ACM international conference on Multimedia, October 10-16, 2004, New York, NY, USA Thorsten Strufe , Jens Wildhagen , Gnter Schfer, Towards the Construction of Attack Resistant and Efficient Overlay Streaming Topologies, Electronic Notes in Theoretical Computer Science (ENTCS), 179, p.111-121, July, 2007 Dan Rubenstein , Sambit Sahu, Can unstructured P2P protocols survive flash crowds?, IEEE/ACM Transactions on Networking (TON), v.13 n.3, p.501-512, June 2005 Chow-Sing Lin , Yi-Chi Cheng, P2MCMD: A scalable approach to VoD service over peer-to-peer networks, Journal of Parallel and Distributed Computing, v.67 n.8, p.903-921, August, 2007 Yi Cui , Klara Nahrstedt, Layered peer-to-peer streaming, Proceedings of the 13th international workshop on Network and operating systems support for digital audio and video, June 01-03, 2003, Monterey, CA, USA Raj Kumar Rajendran , Dan Rubenstein, Optimizing the quality of scalable video streams on P2P networks, Computer Networks: The International Journal of Computer and Telecommunications Networking, v.50 n.15, p.2641-2658, October 2006 Chuan Wu , Baochun Li, Optimal peer selection for minimum-delay peer-to-peer streaming with rateless codes, Proceedings of the ACM workshop on Advances in peer-to-peer multimedia streaming, November 11-11, 2005, Hilton, Singapore Chun-Chao Yeh , Lin Siong Pui, On the frame forwarding in peer-to-peer multimedia streaming, Proceedings of the ACM workshop on Advances in peer-to-peer multimedia streaming, November 11-11, 2005, Hilton, Singapore Yu-Wei Sung , Michael Bishop , Sanjay Rao, Enabling contribution awareness in an overlay broadcasting system, ACM SIGCOMM Computer Communication Review, v.36 n.4, October 2006 Yohei Okada , Masato Oguro , Jiro Katto , Sakae Okubo, A new approach for the construction of ALM trees using layered video coding, Proceedings of the ACM workshop on Advances in peer-to-peer multimedia streaming, November 11-11, 2005, Hilton, Singapore Yang Guo , Kyoungwon Suh , Jim Kurose , Don Towsley, P2Cast: peer-to-peer patching for video on demand service, Multimedia Tools and Applications, v.33 n.2, p.109-129, May 2007 Yi-Cheng Tu , Jianzhong Sun , Mohamed Hefeeda , Sunil Prabhakar, An analytical study of peer-to-peer media streaming systems, ACM Transactions on Multimedia Computing, Communications, and Applications (TOMCCAP), v.1 n.4, p.354-376, November 2005 Shiwen Mao , Xiaolin Cheng , Y. Thomas Hou , Hanif D. Sherali, Multiple description video multicast in wireless ad hoc networks, Mobile Networks and Applications, v.11 n.1, p.63-73, February 2006 Kunwadee Sripanidkulchai , Aditya Ganjam , Bruce Maggs , Hui Zhang, The feasibility of supporting large-scale live streaming applications with dynamic application end-points, ACM SIGCOMM Computer Communication Review, v.34 n.4, October 2004 Zhichen Xu , Chunqiang Tang , Sujata Banerjee , Sung-Ju Lee, RITA: receiver initiated just-in-time tree adaptation for rich media distribution, Proceedings of the 13th international workshop on Network and operating systems support for digital audio and video, June 01-03, 2003, Monterey, CA, USA Leonardo Bidese de Pinho , Claudio Luis de Amorim, Assessing the efficiency of stream reuse techniques in P2P video-on-demand systems, Journal of Network and Computer Applications, v.29 n.1, p.25-45, January 2006 Yi Cui , Baochun Li , Klara Nahrstedt, On achieving optimized capacity utilization in application overlay networks with multiple competing sessions, Proceedings of the sixteenth annual ACM symposium on Parallelism in algorithms and architectures, June 27-30, 2004, Barcelona, Spain Kunwadee Sripanidkulchai , Bruce Maggs , Hui Zhang, An analysis of live streaming workloads on the internet, Proceedings of the 4th ACM SIGCOMM conference on Internet measurement, October 25-27, 2004, Taormina, Sicily, Italy Yang Guo , Kyoungwon Suh , Jim Kurose , Don Towsley, P2Cast: peer-to-peer patching scheme for VoD service, Proceedings of the 12th international conference on World Wide Web, May 20-24, 2003, Budapest, Hungary Padmavathi Mundur , Poorva Arankalle, Optimal server allocations for streaming multimedia applications on the internet, Computer Networks: The International Journal of Computer and Telecommunications Networking, v.50 n.18, p.3608-3621, 21 December 2006 Daria Antonova , Arvind Krishnamurthy , Zheng Ma , Ravi Sundaram, Managing a portfolio of overlay paths, Proceedings of the 14th international workshop on Network and operating systems support for digital audio and video, June 16-18, 2004, Cork, Ireland Mohamed Hefeeda , Ahsan Habib , Boyan Botev , Dongyan Xu , Bharat Bhargava, PROMISE: peer-to-peer media streaming using CollectCast, Proceedings of the eleventh ACM international conference on Multimedia, November 02-08, 2003, Berkeley, CA, USA Mohamed M. Hefeeda , Bharat K. Bhargava , David K. Y. Yau, A hybrid architecture for cost-effective on-demand media streaming, Computer Networks: The International Journal of Computer and Telecommunications Networking, v.44 n.3, p.353-382, 20 February 2004 Zongpeng Li , Anirban Mahanti, A progressive flow auction approach for low-cost on-demand P2P media streaming, Proceedings of the 3rd international conference on Quality of service in heterogeneous wired/wireless networks, August 07-09, 2006, Waterloo, Ontario, Canada Ji Li , Karen Sollins , Dah-Yoh Lim, Implementing aggregation and broadcast over Distributed Hash Tables, ACM SIGCOMM Computer Communication Review, v.35 n.1, p.81-92, January 2005 Alan Kin Wah Yim , Rajkumar Buyya, Decentralized media streaming infrastructure (DeMSI): An adaptive and high-performance peer-to-peer content delivery network, Journal of Systems Architecture: the EUROMICRO Journal, v.52 n.12, p.737-772, December, 2006 Dejan Kosti , Adolfo Rodriguez , Jeannie Albrecht , Amin Vahdat, Bullet: high bandwidth data dissemination using an overlay mesh, Proceedings of the nineteenth ACM symposium on Operating systems principles, October 19-22, 2003, Bolton Landing, NY, USA Miguel Castro , Peter Druschel , Anne-Marie Kermarrec , Animesh Nandi , Antony Rowstron , Atul Singh, SplitStream: high-bandwidth multicast in cooperative environments, Proceedings of the nineteenth ACM symposium on Operating systems principles, October 19-22, 2003, Bolton Landing, NY, USA Konstantin Andreev , Bruce M. Maggs , Adam Meyerson , Ramesh K. Sitaraman, Designing overlay multicast networks for streaming, Proceedings of the fifteenth annual ACM symposium on Parallel algorithms and architectures, June 07-09, 2003, San Diego, California, USA Toufik Ahmed , Mubashar Mushtaq, P2P Object-based adaptivE Multimedia Streaming (POEMS), Journal of Network and Systems Management, v.15 n.3, p.289-310, September 2007 Karthik Lakshminarayanan , Ananth Rao , Ion Stoica , Scott Shenker, End-host controlled multicast routing, Computer Networks: The International Journal of Computer and Telecommunications Networking, v.50 n.6, p.807-825, 13 April 2006 Zongming Fei , Mengkun Yang, A proactive tree recovery mechanism for resilient overlay multicast, IEEE/ACM Transactions on Networking (TON), v.15 n.1, p.173-186, February 2007 Mojtaba Hosseini , Nicolas D. Georganas, End system multicast protocol for collaborative virtual environments, Presence: Teleoperators and Virtual Environments, v.13 n.3, p.263-278, June 2004 Ying Cai , Zhan Chen , Wallapak Tavanapong, Caching collaboration and cache allocation in peer-to-peer video systems, Multimedia Tools and Applications, v.37 n.2, p.117-134, April 2008

content distribution networks;multiple description coding;peer-to-peer networks;streaming media

507723

A cryptographic solution to implement access control in a hierarchy and more.

The need for access control in a hierarchy arises in several different contexts. One such context is managing the information of an organization where the users are divided into different security classes depending on who has access to what. Several cryptographic solutions have been proposed to address this problem --- the solutions are based on generating cryptographic keys for each security class such that the key for a lower level security class depends on the key for the security class that is higher up in the hierarchy. Most solutions use complex cryptographic techniques: integrating these into existing systems may not be trivial. Others have impractical requirement: if a user at a security level wants to access data at lower levels, then all intermediate nodes must be traversed. Moreover, if there is an access control policy that does not conform to the hierarchical structure, such policy cannot be handled by existing solutions. We propose a new solution that overcomes the above mentioned shortcomings. Our solution not only addresses the problem of access control in a hierarchy but also can be used for general cases. It is a scheme similar to the RSA cryptosystem and can be easily incorporated in existing systems.

Figure 1: Enterprise Wide Personnel Hierarchy 2. RELATED WORK A number of works [1, 2, 6, 9, 12, 14, 18, 23] relating to access control in a hierarchy have been proposed. In almost all these works, there is a relationship between the key assigned to a node and those assigned to its children. The difference between the related works lie mostly in the different cryptographic techniques employed for key generation. Some of these techniques [1, 6, 9, 12, 14] are extremely complex. Below we outline a few important works in this area. One of the early solutions to the hierarchical access control problem was proposed by Akl and Taylor [1, 2]. Their solution was based on the RSA cryptosystem [17]. In this work the authors choose the exponents in such a way that the key of a child node can be easily derived from the key of its parent. Mackinnon et al. [12] gave an optimal algorithm for selecting suitable exponents. One potential drawback of these schemes is that if a user at a node wishes to access information stored at a descendant node, he must traverse all the intermediate nodes between his node and the descendant node. This may not be very desirable for cases where the length of the hierarchy is large. Another drawback is that addition of a new node Ni to the leaf of the hierarchy results in key generation of all ancestors of Ni. Sandhu [18] proposed a key generation scheme for a tree hierar- chy. The solution was based on using different one-way functions to generate the key for each child node in the hierarchy. The one-way function was selected based on the name or identity of the child. When a new child is added, the keys for the ancestors do not have to be recomputed. This work, however, does not deal for the case of a general poset. Zheng et al. [23] proposed solutions for the general poset. The authors present two solutions ? one for indirect access to nodes (in which to access data at a lower node, the user has to traverse the intermediate nodes) and the other for direct access to nodes. 3. OVERVIEW OF OUR APPROACH Our approach is simple. We formulate a new encryption protocol that is used to encrypt the data stored in a database. The encryption ensures data integrity as well as data con?dentiality. The data is encrypted with appropriate keys at the same time it is generated. Different portions of the database are encrypted with different keys. A user, who has to retrieve information from the database will attempt to decrypt the entire database with a decrypting key that is assigned to him. However, the user is able to decrypt successfully only that portion(s) of the database for which the user is authorized. The remaining portion of the database is not decrypted successfully. Since the access control technology is based on encrypting with the appropriate key, we ?rst present the theory on which the key generation is based. 4. THEORY BEHIND THE CRYPTOGRAPHIC TECHNIQUES For the following exposition we use the term message to denote any piece of information that needs to be stored in the database. De?nition 1. The set of messages M is the set of non negative integers m that are less than an upper bound N, i.e. De?nition 2. Given an integer a and a positive integer N,the following relationship holds, a qN r where 0 r N and q a N (2) where x denotes the largest integer less than x.Thevalueqis referred to as the quotient and r is referred to as the remainder.The remainder r, denoted a mod N, is referred to as the least positive residue of a mod N. De?nition 3. For positive integers a, b and N,wesayais equivalent to b, modulo N, denoted by a b mod n,ifamod n b mod n. De?nition 4. Two integers a, b are said to be relatively prime if their only common divisor is 1, that is, gcd a b 1. De?nition 5. The integers n1,n2, ,nkare said to be pairwise relatively prime,ifgcdni nj 1 for i j. De?nition 6. The Euler's totient function f N is de?ned as the number of integers that are less than N and relatively prime to N. Below we give some properties of totient functions that are of importance 1. f N N1ifNis prime. 2. f N f N1 f N2 fNkif N N1N2 Nkand N1, N2, ,Nkare pairwise relatively prime. THEOREM 1. Euler's theorem states that for every a and N that are relatively prime, a 1modN PROOF. We omit the proof of Euler's theorem and refer the interested reader to any book on number theory [13] or cryptography [20]. COROLLARY 1. If 0 m N and N N1N2 Nkand N1,N2, ,Nk are primes, then mxf De?nition 7. A key K is de?ned to be the ordered pair e N , where N is a product of distinct primes, N M and e is relatively prime to f N ; e is the exponent and N is the base of the key K. De?nition 8. The encryption of a message m with the key K THEOREM 3. For any two messages m and m^ , such that m m^ e N , denoted as [m,K], is de?ned as N1 N2, De?nition 9. The inverse of a key K e N , denoted by K 1, is an ordered pair d N , satisfying ed 1modfN. THEOREM 2. For any message m. where K e N and K 1 d N . PROOF. We ?rst show that me mod N K 1 (Def. me mod N d mod N (Defs. 8, med mod N (mod arith) (Defs. 2, RHS By symmetry m K 1 K m. COROLLARY 2. An encryption, m K ,isone-to-one if it satis- ?es the relation De?nition 10. Two keys K1 e1 N1 and K2 e2 N2 are said to be compatible if e1 e2 and N1 and N2 are relatively prime. De?nition 11. If two keys K1 e N1 and K2 e N2 are compatible, then the product key, K1 K2,isde?nedas eN1N2 . LEMMA 1. For positive integers a, N1 and N2, a mod N1N2 a mod N1 PROOF. Let a N1 N2 x are integers. a mod N1N2 mod N1 N1N2x N1y z N1 N2Nx1N21 y z N1N2 mod N1 N1y z mod N1 z RHS a mod N1 N1N2x N1y z mod N1 z Hence the proof. where K1 is the key e N1 ,K2is the key e N2 and K1 K2 is the product key e N1 N2 . PROOF. The proof for (6) is the same as that for (5). We just consider the proof for (5). [If part] Given m m^ we have to prove that or me mod N1N2 mod N1 (Def. 8 and Def. 11) me mod N1 (subs. me for a lemma 1) R H S m^e mod N1 mod N1 (idempotency of mod op.) me mod N1 (since m m^,given) [Only If part] Given m K1 K2 m^ K1 mod N1,wehavetoprovem m^ me mod N1N2 m^e mod N1 (Defs. 8, 11) me mod N1 m^e mod N1 mod N1 (lem. 1) (encryption is one-to-one) 5. KEY GENERATION Most often in an organization the access requirements match its hierarchical structure. That is, a person higher up in the hierarchy will usually have access to the documents that can be accessed by persons lower in the hierarchy. So we present the key generation technique for this case ?rst. Later on, we describe how to accomodate the cases where the access requirement does not follow this hierarchical pattern. An organization is usually structured in the form of a hierarchy. Thus, we de?ne a hierarchy of levels as follows. De?nition 12. We consider an organization structure that can be represented as a partially ordered set (poset), L .Lis the set of levels of the organization and is the dominance relation between the levels. 1. If Li and Lj are two levels such that Li Lj, then we say that strictly dominates Li and Li is strictly dominated by Lj. 2. If Li and Lj are two levels such that Li Lj and Lj Li,then we say that the two levels Li and Lj are equal and denote this by Li Lj. 3. If levels Li and Lj are such that either Li Lj or Li Lj, then we say that Li is dominated by Lj or Lj dominates Li and denote this by Li Lj. 4. Two levels Li and Lj are said to be incomparable if nether 5. We say Ly is a parent of Lx if Lx Ly and there is no element Lz such that Lx Lz Ly.IfLx Ly,thenLxis said to be a descendant of Ly and Ly is said to be an ancestor of Lx. Since the organizational structure is a poset, it can be depicted in the form of a Hasse diagram [19]. De?nition 13. Each level Li in the organizational hierarchy isassociated with a unique pair of keys, (Ki Ki ) which we term the default keys for level Li. Ki is the default encryption key of level Li and Ki 1 is the default decryption key of level Li. If the access requirements match the hierarchical structure of the organization, then a person at level Li uses the default encryption key Ki for encrypting documents and the key K 1 for decrypting the encrypted documents. A document encrypted with key Ki possesses the following property 1. it can be decrypted by key K 1 where K 1 is the default decryption key of level Li, 2. it can be decrypted by key K 1 where K 1 is the default decryption key of level Lj such that Lj Li. 3. it cannot be decrypted by key K 1 where K 1 is the default decryption key of level Lk and Lk Li or Lk is incomparable with Li. 5.1 Determining the Default Encryption Key and Default Decryption Key ALGORITHM 1. Default keys generation Input: (i) L - the set of levels and (ii) - the dominance relation Output: (i) K - the set containing the default encryption key for each level and (ii) K 1 - the set containing the default decryption key for each level. choose exponent e obeying the requirements in Def. 7 for each Li Ldo begin K: K Ki return K The default encryption and decryption keys can adequately provide access control for cases where the access requirements match the hierarchical structure of the organization. 5.2 Example 1: Student Transcript Informatio To illustrate our approach we consider an academic organization ? the College of Engineering at some hypothetical university. The College of Engineering is headed by a Dean. Below the Dean are the Department Chairpersons who are responsible for managing the individual departments. Each Faculty (except for cases of joint appointment) is overseen by the respective Department Chair. The Students are at the very bottom of this hierarchy. Each student is advised by one or more faculty members. The organizational hierarchy is shown in ?gure 2. Dean CS Chair ECE Chair Procedure GenerateDefaultKeys((L, begin for each Li Ldo CS Faculty 1 CS Faculty 2 ECE Faculty 1 ECE Faculty 2 while F Ldo begin choose a maximal element Li in L F; for each parent Lj of Li do Student 1 Student 2 Student 3 choose random Ni where gcd Ni factori 1 Figure 2: Access Requirements matching Enterprise Wide Per- Ni Ni factori sonnel Hierarchy // Remove factors that have been included multiple times. We need to maintain the information about the transcripts of in- dividual students. Since this is sensitive information, we need to while D Ldo protect it. The following are the access requirements. begin 1. Each student is able to view his or her transcript. choose a maximal element Li in L D; for each ancestor Lk of Li do 2. A faculty advising a student is able to see his transcript. begin c := no. of distinct paths from Li to Lk 3. The chair of the department in which a student is majoring is Ni Ni Nk c 1 able to view the student's transcript. 4. The dean is able to view the student's transcript. 3. For this example, we consider the case for three students, namely, Student1, Student2, Student3. Each of these students have a faculty advisor who monitors and advises the student. Student1's advisor is CS Faculty1. Student2 is co-advised by CS Faculty2 and ECE Faculty1. Student3's advisor is ECE Faculty2. This is illustrated in ?gure 2. These are the access requirements for this example: 4. 1. Student1, CS Faculty1, CS Chair, Dean can view Student1's transcript. 2. Student2, CS Faculty2, ECE Faculty1, CSChair, ECE Chair, Dean can view Student2's transcript. 3. Student3, ECE Faculty2, ECE Chair, Dean can view Student3's transcript. 5. Note that, in this case the access requirements match the organizational hierarchical structure. That is, if a person X has access to some information, then a person Y at a higher level in the hierarchy will also have access to that information. 5.2.1 Access Control for Example 1 The access control requirement for Example 1 follows the hierarchical structure of the organization. Thus, using the default encryption keys the access can be appropriately restricted. Consider 6. the hierarchy shown in ?gure 2. The keys for the various people are as follows. 1. The default encryption, decryption keys for Studenti are denoted respectively by KSi , KSi 1 where 1 i 3. 2. The default encryption, decryption keys for CS Facultyi are denoted respectively by KCFi , KCF1i where i 1or2. 7. 3. The default encryption, decryption keys for ECE Facultyi are denoted respectively by KEFi, KEF1i where i 1or2. 4. The default encryption, decryption keys for CS Chair are denoted respectively by KCChair , KCC1hair . 5. The default encryption, decryption keys for the ECE Chair are denoted respectively by KEChair , KEC1hair 6. The default encryption, decryption keys for the Dean are de- 8. noted respectively by KDean, KDe1an. The key KDean is chosen as per de?nition 7. Once the Dean's key has been ?xed, the other keys as generated by Algorithm 1 are as follows: 1. Encryption Key of Dean : 9. KDean e NDean Decryption Key of Dean :KDean dDean NDean , where e dDean 1modfNDean 2. Default Encryption Key of CS Chair : KCChair KDean KC , Chair where KC is compatible with KDean Chair that is, KCChair e NDean NCChair 10. Decryption Key of CS Chair : KCC1hair dCChair NCChair , where e dCChair 1modfNChair Encryption Key of ECE Chair : KEChair KDean KEChair , where KEChair is compatible with KDean that is, KEChair e NDean NEChair Decryption Key of ECE Chair : KEC1hair dEChair NEChair , where e dEChair 1modfNEChair Encryption Key of CS Faculty1 : KCF1 KCChair KCF1 , where KCF1 is compatible with KCChair that is, KCF1 e NDean NCChair NCF1 Decryption Key of CS Faculty1 :KCF1 dCF1 NCF1 , where e dCF1 1modfNCF1 Encryption Key of CS Faculty2 : KCF2 KCChair KCF2 , where KCF2 is compatible with KCChair that is, KCF2 e NDean NCChair NCF2 Decryption Key of CS Faculty2 :KCF2 dCF2 NCF2 , where e dCF2 1modfNCF2 Encryption Key of ECE Faculty1 : KEF1 KEChair KEF1, where KEF1 is compatible with KEChair that is, KEF1 e NDean NEChair NEF1 Decryption Key of ECE Faculty1 :KEF1 dEF1 NEF1 , where e dEF1 1modfNEF1 Encryption Key of ECE Faculty2 : KCF2 KEChair KEF2, where KEF2 is compatible with KEChair that is, KCF2 e NDean NEChair KEF2 Decryption Key of ECE Faculty2 :KEF2 dEF2 NEF2 , where e dEF2 1modfNEF2 Encryption Key of Student1 : KS1 KCF1 KS1 , where KS1 is compatible with KCF1 that is, KS1 e NDean NCChair NCF1 NS1 Decryption Key of Student1 :KS1 dS1 NS1 , where e dS1 1modfNS1 Encryption Key of Student2 : KS2 KCF2 KEF1 KS2, where KS1 is compatible with KCF1 and KEF1 that is, KS2 e NDean NCChair NEChair NCF2 NEF1 NS2 Decryption Key of Student2 :KS2 dS2 NS2 , where e dS2 1modfNS2 Encryption Key of Student3 : KS3 KEF2 KS3, where KS3 is compatible with KEF2 that is, KS3 e NDean NEChair NEF2N Decryption Key of Student3 : KS31 dS3 NS3 , where e dS3 1modfNS3 Consider Student1's transcript. If this transcript is encrypted with key KS1 , then any of the keys KDean, KCChair , KCF1 , KS1 can be used to decrypt it. Thus all the personnel higher up in the hierarchy can decrypt this transcript using their own default decryption key. 5.3 Determining a Customized Encryption Key As illustrated by Examples 2 and 3, sometimes the access requirement does not match the organizational structure and then a different encryption key, which we term customized encryption key, must be used to protect this information. The decryption key, how- ever, remains the same. ALGORITHM 2. Customized encryption key generation Input: (i) L - the set of levels, (ii) - the dominance relation, (iii) - the level for which the default encryption key is being gener- ated, (iv) A - the set of levels that do not dominate Li but who are to be given access, (v) D - the set of levels that dominate Li and who are to be denied access, (vi) K - the set of default encryption keys. Output: Kci - the customized key generated for level Li. Procedure GenerateCustomEncKey(L, ,Li, D, A, K) begin while A allow do begin for each Lj A allow that is minimal do begin // Deny access to ancestors of Lj for each parent Lk of Lj do allow : allow Lj while D deny do begin for each Lj D deny that is minimal do begin Give access to ancestors of Lj for each parent Lk of Lj do // If Nk has been included multiple times: for each parent Lk of Lj do begin c := no. of paths from Lk to Lj Kci e Nci return Kci Dean CS Chair ECE Chair Student 1 Student 1 (CS Figure 3: Access Requirements for Individual Courses 5.4 Example 2: Individual Course Grade Informatio The organizational hierarchy in this case is the same as before (please refer to ?gure 2). We need to maintain information about the grades each student receives in each of the courses he/she has taken. The access requirements are complicated in this case: 1. Each student is able to view his or her grades for the courses he or she has taken. 2. The faculty offering the course is able to see the grades for all students who have taken the course under the faculty. 3. The chair of the department which has offered the course can view the grades of the students who have taken that course. 4. A student's advisor is able to see his grades in all the courses. 5. The student's department chair can view the student's grades for all the courses. 6. The dean is able to view the grades of all students that have taken a course offered by a department in the college. Suppose Student1 takes two courses: (i) CS 350 offered by the Computer Science Department and taught by CS Faculty2 and (ii) offered by the ECE Department and taught by ECE Faculty1. This is shown in ?gure 3. Note that ?gure 3 shows only the relevant portion of the hierarchy given in ?gure 2. The access requirements are as follows: 1. Student1's CS 350 grade can be viewed by Student1, CS Faculty1, CS Faculty2, CS Chair, Dean. 2. Student1's ECE 373 grade can be viewed by Student1, CS Faculty1, ECE Faculty1, ECE Chair, CS Chair, Dean. Note that, in this case the access requirements do not match the organizational hierarchy given in ?gure 2. More access is required than permitted by the organizational hierarchy - for exam- ple, ECE Faculty1 must be given access to the ECE 373 grades of Student1. 5.4.1 Access Control for Example 2 In this case the access patterns do not match the organizational structure. For example, CS Faculty2 must have access to the CS 350 grade of Student1 even though CS Faculty2 is not his advisor. Thus default keys cannot be used and custom keys are required to encrypt the grades obtained in the individual courses. 1. Key for encrypting Student1's CS 350 grade: KcS1 C350 e NDean NCChair NCF1 NCF2 NS1 2. Key for encrypting Student1's ECE 373 grade: KcS1 E373 e NDean NCChair NEChair NCF1 NEF1 NS1 If key KcS1 C350 is used for encrypting CS 350 grade of Student1, then the encrypted grade can be decrypted by any of the default decryption keys of Student1, CS Faculty1,CS Faculty2,CSChair and the Dean.IfkeyKcS1 E373 is used for encrypting ECE 373 grade of Student1, then the encrypted grade can be decrypted by any of the default decryption keys of Student1, CS Faculty1, ECE Faculty1, CS Chair, ECE Chair and the Dean. Dean CS Chair ECE Chair Student 2 (Sensitive File F) Figure 4: Access Requirements for a Sensitive Project 5.5 Example 3: Sensitive Project Information As a part of the curriculum, the students are required to do a Software Design Project. Some of these projects involve proprietary data whose disclosure should be kept to a minimum level. Thus, there is a need for encrypting the ?les associated with the project. The access requirements are as follows. 1. The faculty members who advise the student on this project have access to the ?les. 2. The student has access to the ?les. 3. No other person is given access to the ?les. Student2 is working on a project with faculty membersCS Faculty2 and ECE Faculty1. File F contains sensitive information which only the student and the project advisors can view. This case is illustrated in ?gure 4. Note that in ?gure 4 the entire organizational hierarchy is not shown ? only the part pertinent to the example is given. The access requirements are as follows: 1. Sensitive ?le F can be viewed by Student2, CS Faculty2 and ECE Faculty1. This is an example where the access requirement does not follow the organizational hierarchy. People higher up in the hierarchy (the Dean, the Department Chairperson) is not given the access to these ?les. 5.5.1 Access Control for Example 3 In this case the organizational structure is that given in ?gure 2. The sensitive ?le F must be protected such that only the faculty advisors (CS Faculty2, ECE Faculty2 and ECE Faculty1)andthe student (Student2) have access to this ?le. No other person can have access. Thus, for protecting the project work the default encryption key is not adequate. Customized encryption key must be used. The student encrypts ?le F using the following customized encryption KcS2. The customized encryption key KcS2 is generated using Algorithm 2. For this example, The encrypted ?le F can be decrypted by the default decryption keys of CS Faculty2, ECE Faculty2, ECE Faculty1 or Student1. 6. SECURITY OF THE PROPOSED MECH- Our scheme is based on the RSA cryptosystem. Its security is based on the dif?culty of factoring large prime numbers. We do need to mention, however, that the low exponent attack on the RSA cryptosystem [10] does not arise in our case. The low exponent attack occurs in the following way: suppose the same message m is encrypted with different keys sharing the same exponent. Let the exponent e 3 and the different keys are K1 e N1 , K2 e N2 , K3 e N3 , etc. By using the Chinese Remainder Theorem [13] an attacker can get me. Now if he can guess e correctly, then by extracting the eth root of me, he can obtain m. To avoid this problem, we choose a large exponent e (e is substantially larger than the number of levels). Since the complexity of raising a number to the eth power grows as loge, choosing a large exponent does not signi?cantly increase the computational com- plexity. Also in our mechanism, the data is appropriately encrypted and stored only in one place. Having multiple copies of same data encrypted with different keys does not arise in our case. 7. ORGANIZATIONAL CHANGES AFFECTING KEY MANAGEMENT As outlined in Section 5, the default encryption keys generated are dependent on the hierarchical structure of the organization. If restructuring takes place in the organization, the Hasse diagram representing the personnel hierarchy will be modi?ed resulting in the change of the default encryption keys. In this section we give algorithms that result in the modi?cation of default encryption keys when the organization structure is changed. Any restructuring can be expressed as modi?cations to the Hasse Diagram. A Hasse diagram can be modi?ed by using combinations of the four primitive operations 1. adding an edge between existing nodes ? this corresponds to the scenario when a new hierarchical relationship is established between two existing persons in the organization. 2. deleting an edge from an existing node ? this corresponds to the scenario when an existing hierarchical relationship is broken. 3. adding a node ? this corresponds to the case when a new person joins the organization. 4. deleting a node ? this corresponds to the case when a person leaves the organization. For each of these operations we give an algorithm stating how the default encryption key must be changed because of the operation. The default decryption key, however, remains the same. ALGORITHM 3. Default enc. key change with edge insertion Input: (i) L - the set of levels, (ii) - the dominance relation, (iii) the new directed edge that is to be inserted from level i to level j, (iv) K - the set of default encryption keys. Output: K - the set containing the default encryption key for each level. Procedure ChangeDefEncKeysEdgeIns(L, , begin for each descendant k of j do insertion results in a new path from i to k then for each Li Ldo begin return K ALGORITHM 4. Default enc. key change with edge deletion Input: (i) L - the set of levels, (ii) - the dominance relation, (iii) the directed edge from level i to level j that will be removed, (iv) K - the set of default encryption keys. Output: K - the set containing the default encryption key for each level. Procedure ChangeDefaultEncKeysEdgeDel(L, , begin foreach descendant k of j do deletion results in no path from i to k then begin eliminate access to parents of Li foreach parent l of i do if there is a path from l to k after deleting i j do for each Li Ldo begin return K ALGORITHM 5. Default key generation with node insertion Input: (i) L - the set of levels, (ii) - the dominance relation, (iii) - the node that will be added, (iv) K - the set of default encryption the set of default decryption keys. Output: (i) K - the set containing the default encryption key for each level, (ii) K 1 - the set containing the default decryption key for each level. Procedure AddDefaultEncKeyNodeIns(L, ,i,K, K 1) begin Choose random Ni that is relatively prime to existing Nk Get the exponent e of any key Kk in K removal with node deletion Input: (i) L -thesetoflevels,(i) - the dominance relation, (iii) - the node that will be added, (iv) K - the set of default encryption the set of default decryption keys. Output: (i) K - the set containing the default encryption key for each level, (ii) K 1 - the set containing the default decryption key for each level. Procedure RemoveDefaultKeysNodeDel(L, ,i,K, K 1) begin if there are no edges incident on node i then begin 8. AN ALTERNATE SOLUTION The problem of access control in a hierarchy can be solved using an alternative solution1 that can potentially have lesser computational requirements. Everybody in the organization has a public-private Suppose an employee wants to share a message m with his n superiors, denoted by S1, S2, ,Sn. For each superior Si, the employee encrypts the message m with the public key of superior Si. Thus, he performs n 1 encryptions (n for his superiors and one for himself). He stores these n 1 encrypted messages. Each superior Si can use his private key to retrieve the message. This alternate solution has a problem: multiple encrypted copies of data must be stored. Storing multiple copies of encrypted data can be a source of inconsistency. For example, suppose the employee decides to change m to m^. After making this change, the employee is supposed to encrypt m^ with the public keys of his n superiors. However, the employee forgets to encrypt m^ for superiors Sj and Sk. In such a case superiors Sj and Sk will be accessing the previous version of the data which is m. This source of inconsistency associated with redundant data does not arise in our case because there is only one copy of the encrypted data. Moreover, keeping multiple encrypted copies of the same data leads to more exposure for the data which may not be desirable. This problem does not arise in our case. Our solution has another advantage, namely, mutual access awareness ? each person having the encrypted data has the knowledge of who else can view this data. For any data object we can have the need-to-know list which speci?es the persons who can access the document. Anyone having this list can verify that only those persons in the list and no one else can decrypt the corresponding data object. This is not possible for the alternate solution. One might, however, argue that our scheme is more computation intensive for large hierarchies. This is because the base N of the 1The authors would like to thank the anonymous reviewer for this alternate solution. default encryption key increases with the number of levels in the [10] hierarchy. However there are techniques (for example using Fast Fourier Transforms) by which the encryption can be done in an [11] ef?cient manner [5]. These techniques are especially useful if the base of the keys are large. Details of computational complexity will be treated in a future work. 9. CONCLUSION AND FUTURE WORK In this paper we have presented a new encryption technique for securing information in a database. The implementation that we propose is completely general; we have shown how the different [13] access control policies in an organization can be implemented by our technique. [14] A lot of work is yet to be done. The ?rst step is to implement the algorithms that have been proposed; this experience will help us in detecting subtle ?aws that we may have overlooked. Performance analysis and scalabilty studies need to be done before our method can be used in real world scenarios. Finally, we wish to show how the discretionary and the mandatory access control policies [22] of [15] an organization can be implemented using the technology that we proposed. The need for hierarchical access control arises in other contexts as well. For example, different kinds of hierarchies, such as, class [16] composition hierarchy and class inheritance hierarchy arises in object-oriented database systems [16, 21]. We need to investigate whether our mechanism can be applied to implement the access control policies [3, 7, 11] (such as, visibility from above and visiblity from [17] below policies) desirable in such hierarchies. 10. --R The Secure Use of RSA. In Database Security III Optimal Algorithm for Assigning Cryptographic Keys to Access Control in a Hierarchy. An Introduction to the Theory of Numbers. Authentication for Hierarchical Multigroup using the Advances in Cryptology: Information Security Policy. Information Security: An Integrated Collection of Essays pages 160? A Model of Authorization for Next-Generation Database Systems Transactions on Database Systems A Method for Communications of the ACM Cryptographic Implementation of a Tree Hierarchy for Access Control. Mathematics: A Discrete Introduction. Brooks/Cole Principles and Practice. Mandatory Security in Trusted Computer System Evaluation Criteria (The Orange New Solutions to the Problem of Access Control in a Hierarchy. Preprint 93-2 of Wollongong http://citeseer. --TR An optimal algorithm for assigning cryptographic keys to control access in a hierarchy Cryptographic implementation of a tree hierarchy for access control Mandatory security in object-oriented database systems A cryptographic key generation scheme for multilevel data security A model of authorization for next-generation database systems Membership authentication for hierarchical multigroups using the extended Fiat-Shamir scheme Flexible access control with master keys Database security Cryptography and network security (2nd ed.) Cryptographic solution to a problem of access control in a hierarchy A method for obtaining digital signatures and public-key cryptosystems Mathematics Fundamentals of Database Systems Access Control in Object-Oriented Database Systems - Some Approaches and Issues --CTR H. Ragab Hassen , A. Bouabdallah , H. Bettahar , Y. Challal, Key management for content access control in a hierarchy, Computer Networks: The International Journal of Computer and Telecommunications Networking, v.51 n.11, p.3197-3219, August, 2007 Jason Crampton, Applying hierarchical and role-based access control to XML documents, Proceedings of the 2004 workshop on Secure web service, p.37-46, October 29-29, 2004, Fairfax, Virginia Alfredo De Santis , Anna Lisa Ferrara , Barbara Masucci, Unconditionally secure key assignment schemes, Discrete Applied Mathematics, v.154 n.2, p.234-252, 1 February 2006 Mikhail J. Atallah , Marina Blanton , Keith B. Frikken, Key management for non-tree access hierarchies, Proceedings of the eleventh ACM symposium on Access control models and technologies, June 07-09, 2006, Lake Tahoe, California, USA Batrice Finance , Sada Medjdoub , Philippe Pucheral, The case for access control on XML relationships, Proceedings of the 14th ACM international conference on Information and knowledge management, October 31-November 05, 2005, Bremen, Germany Luc Bouganim , Franois Dang Ngoc , Philippe Pucheral, Client-based access control management for XML documents, Proceedings of the Thirtieth international conference on Very large data bases, p.84-95, August 31-September 03, 2004, Toronto, Canada Mikhail J. Atallah , Keith B. Frikken , Marina Blanton, Dynamic and efficient key management for access hierarchies, Proceedings of the 12th ACM conference on Computer and communications security, November 07-11, 2005, Alexandria, VA, USA

access control;cryptography;hierarchy

508173

Computational paradigms and protection.

We investigate how protection requirements may be specified and implemented using the imperative, availability and coercion paradigms. Conventional protection mechanisms generally follow the imperative paradigm, requiring explicit and often centralized control over the sequencing and the mediation of security critical operations. This paper illustrates how casting protection in the availability and/or coercion styles provides the basis for more flexible and potentially distributed control over the sequencing and mediation of these operations.

INTRODUCTION The sequencing of operations in a computation may be classified in terms of three fundamental paradigms. In the traditional imperative paradigm, the programmer explicitly determines the sequencing constraints of operations; in the availability paradigm, the sequencing of operations depends only on the availability of operand data; and, in the coercion paradigm, operations are executed when, and only when, their results are needed. These paradigms can be interpreted in the context of pro- tection. Conventional protection mechanisms generally follow the imperative paradigm by enforcing explicit mediation and sequencing on operations. For example, when mediating a purchase order transaction [order; validate; invoice; payment], an imperative protection mechanism might ensure Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. that the operations are done in the correct sequence, and that suitable separation of duties are applied at each stage. A weakness of the imperative approach is that protection is based on the explicit control of sequencing and mediation. Explicit control can become difficult when flexibility over the sequencing of operations is required. For example, it is often desirable to allow an un-validated order to progress through the system with the assumption that validation will be done at some stage, but before payment is made. Sometimes it may even be expedient to make payment without any validation While such requirements can be constructed in terms of explicit sequencing control, it may be more natural to consider the requirements in terms of the data dependencies between the operations. For example, operation pay depends on operations validate and invoice, operation validate depends on operation order, and so forth. These relationships can be expressed in terms of a graph of operations (nodes) linked together by the data that is passed between them. The implicit parallelism of operations in the graph gives rise to two different paradigms for specifying and enforcing protection requirements. In the availability paradigm, sequencing and mediation of operations depend only on the availability of operands. This data-flow like sequencing of operations gives rise to eager execution. For example, once an order is proposed then validation and invoice processing can be done at any stage (before payment). In the coercion paradigm, sequencing and mediation of operations is determined on the basis of when results are needed. This is the functional style of operation sequencing and gives rise to lazy evaluation. For example, only if and when payment is finally required should order validation be sought. In this paper we investigate how protection requirements can be specified and implemented using the imperative, availability and coercion paradigms. This is done using the Condensed Graphs model of computation [8, 11] which provides a single framework that unifies these three paradigms. In addition to providing flexibility in the sequencing and control over security critical operations, these paradigms have facilitated the development of novel distributed protection mechanisms that are also described in this paper. By following the inherently parallel availability and coercion paradigms, the proposed protection mechanisms need not necessarily rely on centralized security state. Section 2 provides a brief outline of the Condensed Graphs model. Using the purchase order transaction as an exam- ple, Section 3 considers various sequencing constraints that one may wish to enforce. This section also serves to illustrate the notation and semantics of the Condensed Graph model. A protection model based on permissions and protection domains is described in Section 4. Specific protection mechanisms for this model are considered in Section 5. 2. CONDENSED GRAPHS Like classical dataflow [1], the Condensed Graphs (CG) model [8, 11] is graph-based and uses the flow of entities on arcs to trigger execution. In contrast, Condensed Graphs are directed acyclic graphs in which every node contains not only operand ports, but also an operator and a destination port. Arcs incident on these respective ports carry other Condensed Graphs representing operands, operators and destinations. Condensed Graphs are so called because their nodes may be condensations, or abstractions, of other Condensed Graphs. (Condensation is a concept used by graph theoreticians for exposing meta-level information from a graph by partitioning its vertex set, defining each subset of the partition to be a node in the condensation, and by connecting those nodes according to a well-defined rule [6].) Condensed Graphs can thus be represented by a single node (called a condensed node) in a graph at a higher level of abstraction. The basis of the CG firing rule is the presence of a Condensed Graph in every port of a node. That is, a Condensed Graph representing an operand is associated with every operand port, an operator Condensed Graph with the operator port and a destination Condensed Graph with the destination port. This way, the three essential ingredients of an instruction are brought together (these ingredients are also present in the dataflow model; only there, the operator and destination are statically part of the graph). Any Condensed Graph may represent an operator. It may be a condensed node, a node whose operator port is associated with a machine primitive (or a sequence of machine primitives) or it may be a multi-node Condensed Graph.5 f Figure 1: Condensed Graphs congregating at a node to form an instruction The present representation of a destination in the CG model is as a node whose own destination port is associated with one or more port identifications. Figure 1 illustrates the congregation of instruction elements at a node and the resultant rewriting that takes place. We decorate connections to distinguish between different kinds of ports, and use numbers to distinguish input ports. Executing a Condensed Graph corresponds to scheduling its fireable nodes to run on ancillary processors, based on the constraints of the graph. The nodes in a graph are represented as triples (operation, operands, destination) and are constructed by the Triple Manager (TM) as the graph exe- cutes. Once a node is ready to fire, the triple manager can schedule it for execution on an ancillary processor. The CG operators can be divided into two categories: those that are 'value-transforming' and those that only move Condensed Graphs from one node to another in a well-defined manner. Value-transforming operators are intimately connected with the ancillary processors and can range from simple arithmetic operations to the invocation of software components that form part of an application system. In contrast, Condensed Graph moving instructions are few in number and are architecture independent. These TM primitives include the condensed node evaporation operator and the ifel node. A number of working prototypes that use Condensed Graphs have been developed, demonstrating it usefulness as a general model of computation. Prototypes include a sequential TM interpreter [8] and a web-based distributed computing engine [10]. WebCom [10] schedules the execution of coarse-grain computations described as Condensed Graphs. Web clients (ancillary processors) connect to a WebCom server (Triple Manager) whereupon they are served appropriate computations (CG operations). By statically constructing a Condensed Graph to contain operators and destinations, the flow of operand Condensed Graphs sequences the computation in a dataflow manner. Similarly, constructing a Condensed Graph to statically contain operands and operators, the flow of destination Condensed Graphs will drive the computation in a demand-driven manner. Finally, by constructing Condensed Graphs to statically contain operands and destinations, the flow of operators will result in a control-driven evaluation. This latter evaluation order, in conjunction with side-effects, is used to implement imperative semantics. The power of the CG model results from being able to exploit all of these evaluation strategies in the same computation, and dynamically move between them, using a single, uniform, formalism. 3. OPERATION SEQUENCING In this section we consider a variety of controls that one may wish to place on the purchase order transaction example discussed in Section 1. The examples also serve to illustrate the notation and semantics of the CG model. Example 1. Condensed node POI specifies the allowable behaviour for order processing (Figure 2). By definition [8], O PO I Figure 2: Imperative Style Defintion of POI the behaviour of a Condensed Node such as POI is constructed as a Condensed Graph with a single entry node single exit node (X ). The other nodes in the graph represent the operations available: propose an order (O), validate the order (V), process the associated invoice (I) and make a payment (P). Arcs represent data paths between the operations which, in this case, may fire (execute) when data arrives at their input ports. For example, when the order has been proposed a value (the details) is output from O and passed to V, which may, in turn, fire, and so forth. Firing a condensed node evaporates it into the graph that defines it, with input available from the E node and final output emanating from its X node. Figure 2 may be regarded as specifying sequential ordering constraints [O;V;I;P] in an imperative style. 4 Example 2. Figure 3 specifies the purchase order trans-action in a simple dataflow or availability manner. Once the order has been proposed, details are available to both V and I, which may fire in either order. Only when both inputs A O I PO Figure 3: Availability Style Definition of POA (outputs from V and I) are available, can payment proceed.Example 3. In Figure 4, orders are validated only when needed, that is, V is executed in a demand or coercion - driven manner. The V node acts as an input value to node O I PO Figure 4: Demand-Driven Validation of Orders POC which results in V becoming fireable only when needed, as illustrated in Figure 5. If id represents a transaction identifier then POC id evaporates on firing to its defining graph with id passed on from entry node E to operation O, which fires (denoted as *) generating, as output, a purchase order po (Figure 5(a)). This acts as input to V and I. Operation I fires (Figure 5(b)) since its input value is present and its output port is bound to a destination. However, while an input value is present for V, its output port is not bound to any destination, and therefore, it may not yet fire. Once operation I has fired, operation P has values at both ports (a simple value inv, and a graph connected to node V), has an output destination, and therefore, is fireable. How- ever, P expects the values on its input ports to be atomic values (such as po), and not an executable graph object. A 'preliminary' firing of P does not execute P, but `grafts' node V to the input port to which it acts as a value (Figure 5(c)). As a result, P is no longer fireable; the output port of V becomes bound and fires (Figure 5(d)). As a result, operation P has atomic input values, fires, and generates a check Figure 5(e)). In this example, V is executed in an availability or demand-driven manner: only when a result (validation) is required, is it scheduled for execution. Execution of I may be regarded as eager, while execution of V regarded as lazy. 4 Example 4. Figure 6 specifies a variation of the purchase order graph whereby the validation requirement may be by-passed for invoices up to a certain limit. The operation lim? O I ifel X PO L1 lim? true Figure inspects the invoice, returning true if its value is below a certain limit; otherwise it returns false. Payment operation P' has the same behaviour as P, except that it takes only one input (from I). The conditional operation ifel is provided by the Triple Manager. It takes a boolean value on its B input port, and if true then it passes the value at its T(hen) port to output, otherwise it passes the value at its E(lse) port to output. Note that the decorations on ports T and E indicate that they are non-strict , that is, their input values will not be grafted if they are not atomic. This is unlike strict ports where non-atomic values are always grafted. Thus, when ifel fires the graph at its T or E port simply passes through it, depending on the value at the B port. Figure 7 illustrates part of the behaviour when lim? returns true. The input port to X is strict and the output of P' will be grafted to the input of X , making P' fireable. Note that in this case P never fires, and consequently, V never fires. If P is selected by the ifel then it becomes grafted to X and fireable (and never fires). If the graph in Figure 6 had, instead, specified a direct grafted connection from outputs of P and P' to the ifel operation then then operation V may eventually fire, regardless of whether it is needed. This is analogous to a kind of speculative validation (may validate regardless), as opposed to the conservative validation (validate only if required) specified in the original graph. 4 Example 5. Non-strictness provides a degree of higher- orderedness to graphs: an operation/graph may be treated as data as it moves around the graph with execution deferred until it arrives at a strict port whereupon it becomes grafted. In Figure 8, a revised invoice-processing operation is non- PO L2 O P" I' Figure 8: Lazy Validation II POL2 strict on its 'validation' port; it checks the invoice against the order and outputs a suitable value (graph) that includes the yet-to-be executed V operation (and its inputs). The payment operation P" also has a non-strict input port; it generates a print-check operation Ck with a dependency on the V operation (see Figure 9(a)). This graph value O po po I O I O I O (b) (a) (c) (d) (e) ok ck id po po id po po ok id po po Figure 5: Firing Sequence for POC . true I ifel ifel lim? true (b) (a) Figure 7: Snapshot of Lazy Order Validation POL1 may be thought of as representing the behaviour "before issuing the check, validation must be done". When this graph arrives at the strict port of X, it is grafted (Figure 9(b)), which in turn, results in the grafting of V (Figure 9(c)), which in turn becomes fireable. Only when validation is done, can the check be printed, and POL2 completed. 4 Condensed Graphs provide an executable notation that allows us to precisely specify how operations should be 'glued' together. The next section proposes a protection framework for this 'glue'. 4. PROTECTION FRAMEWORK A Triple Manager schedules the nodes of a graph to be fired on the ancillary processors that are participating in the computation. These processors could be the components of a parallel machine, a network of workstations or a variety of heterogenous systems, connected over local networks and/or the Internet. From a security perspective, we assume that when a node fires, it does so within some security domain, which reflects the resources that can be accessed by the node. Thus a domain could correspond to a specific host on a network, a subnet, and so forth. However, we are not limited to a network computing model: a domain could represent a traditional protection domain [7]. For example, a node that performs a secret operation could be scheduled to domain secret. Alternatively, an authenticated domain might be represented by the public-key that speaks for it. An operation may be scheduled to a particular security domain only if the security domain holds the correct permission that provides authorization to execute the operation. Each node has a permission attribute that reflects the necessary authorization (required by a domain) to execute it. The Triple Manager provides a primitive operation Perm perm(NonStrict Node n); where, perm(n) returns the permission associated with node n. Non-strictness is required since examining the permission attribute of a node should not result in its execution. Permissions are treated as primitive value nodes within a condensed graph and are assumed to be structured as a lattice means that permission y provides no less authorization than x. A simple example is the powerset lattice of fread; writeg, with ordering defined by subset, and lowest bound (t) defined by union. Thus the lowest upper bound operator may be used to compose permissions. A Triple Manager schedules the nodes of the graph it is executing to fire in security domains that have appropriate permissions. A primitive operation is provided. Perm sdom(NonStrict Node n); Given node A, then sdom(A) returns the permission assigned to the domain that A is scheduled to. If A is not yet ready to fire then sdom may either return the domain planned for A or it may block until it is known and/or ready to fire. Only a single permission need be associated with each security domain since composite permissions may be constructed using t. If the node A is not a primitive operation, that is, it is a condensed node, and if it is to be scheduled to the same do- XP" (b) (a) (c) Ck Ck Ck Figure 9: Strictness and Eventual Validation in POL2 main as that of the current Triple Manager, then the current Triple Manager will manage the scheduling for the graph that A defined. If A is scheduled to fire in a different domain then another Triple Manager running in this domain will schedule the graph that A defines. The primitive TM operation Perm cdom(); returns the permission assigned to the domain of the graph currently executing, that is, the permission assigned to that the Triple Manager executing the graph. Figure 10 illustrates the relationship between the security-related TM prim- itives: a triple manager has scheduled the condensed node A (security attribute a) to be executed by another Triple Manager that is running in a domain with permission x. The graph defined by A is said to run in a security context (x,a). The Triple Manager is regarded as a trusted component in the sense that the triples that it manages may be accessed only by the Triple Manager and that it constructs and schedules triples faithfully and according to the graph it is executing 1 . When a node with permission attribute a fires in a domain with permission x then it is said to have a security context (x; a). Security is defined in terms of whether a graph in one security context may schedule a node to fire in another security context (or possibly the same). A node with permission attribute b that is part of a graph with a security context (x; a) may be scheduled to a domain y if implies is a partial ordering relation between security contexts. This relation, called the scheduling constraint , controls how graphs evaporate We do not prescribe a specific definition for the implies relation. However, one possible definition could be based on the permission orderings. Considering Figure 10, the Triple Manager must have sufficient permissions to execute (the graph defined by) A (a x). This Triple Manager must also have sufficient permission to schedule any node of this graph to another domain (y x). Similarly, B must be authorized to run in this domain (b y), and thus we have implies((x; a); (y; b)). Example 6. A Triple Manager schedules the nodes of a graph to be fired on the ancillary processors participating in the computation. Suppose that the purchase order system is implemented across a network of personal workstations connected to a trusted server. Define the set of permissions as the powerset of fclk; mgrg, with subset as the ordering relation. Operations O and I We believe that assuring the correctness of the Triple Manager should be straightforward; the core of its current implementation stands at a few hundred lines of C code. have permission attribute fclkg; operations V and P have permission attribute fmgrg, and condensed node POI has permission attribute fg. Alice is a manager and is bound to permission fmgrg, while Bob, a clerk, is bound to permission fclkg. Suppose that Alice requests that an instance of POI is to be executed on a trusted server (domain fclk; mgrg). This provides a context (fclk; mgrg; fg) from which the operations O, I, V and P will be scheduled. Operations O and I may be scheduled to Alice's domain (context (fclkg,fg) on her workstation). Similarly, V and P may be scheduled to Bob.5. PROTECTION MECHANISMS The security of a graph (based on the scheduling con- straint) can be defined in an operational or denotational manner. Operationally, a graph is secure if the Triple Manager schedules only those nodes that uphold the scheduling constraint. The disadvantage of this approach is that the security mechanism must be hard-coded as part of the Triple Manager and is implementation dependent. The alternative is to define security in a denotational way, that is, define the enforcement of the scheduling constraint in terms of a Condensed Graph. We take this approach, guaranteeing that our proposed security mechanisms can be implemented, not having to worry at this stage about low-level operational de- tails. Another advantage of defining security in this way is that we can program alternative protection mechanisms. 5.1 Fragile Protection The fragile protection operation F takes a node A as its input operand and if the scheduling constraint is upheld then A may fire, that is, A F evaluates to A. If the scheduling constraint fails then A may not fire and the result of the evaluation is null. Figure 11 defines the operation as a condensed node. For the purposes of this cdom perm sdom F null ifel Figure 11: Definition of Fragile Protection Operator paper we assume that the graph that is defined by F executes Context (y,b) fires/evaporates sdom cdom perm A a x z Graph defined by A scheduled to aTriple Manager running sdom cdom Context (x,a) perm y x in domain with permission x Graph containing condensed node A runs in a domain that holds permission w. Context (w,p) Operation B executes in a domain that holds permission y schedules Figure Condensed Node B Scheduled to Fire executes (is scheduled) in the same protection domain as its parent. This ensures that the value cdom referenced in Figure 11 corresponds to the cdom of its parent, that is, the domain that schedules the node input (A) to F . Since the input port of the fragile protection operation is non-strict, its operand A passes into its graph without grafting/firing. Lazy evaluation within the graph ensures that A passes to the X node only if it is to be scheduled to an appropriate domain, whereupon it becomes grafted to the strict port of X and fires. Example 7. Figure 12 protects the ordering process defined in Figure 8. The protection nodes that protect operations O, I' and P'' are immediately available to fire. Figure 13 illustrates the result of these nodes firing. An alternate firing sequence might fire the protection node of O, followed by O, and so forth. The validation operation is mediated on a lazy basis: The protected V operation (sub-graph passes through the I' and P'' nodes. On becoming an input to the X node the protection operation becomes grafted, and fires, mediating the scheduling of V. 4 In this paper we consider only the security constraints on the scheduling of nodes. How exactly a Triple Manager decides when to schedule fireable nodes must be left to the Triple Manager. It would be straightforward to implement a Triple Manager that tried to ensure that the scheduling constraint was always upheld when scheduling. In practice, we expect that a fragile protection node would be implemented as a TM primitive, rather than as a condensed node. An implementation of the Triple Manager must also decide whether the protection operation should fire as soon as possible or whether it should wait until the node it mediates has all of its input ports bound. Immediately firing a protection node gives rise to the notion of speculative protection, whereby the Triple Manager schedules, in advance, an (au- thorized) domain for an operation before it is ready to fire. Alternatively, deferring the firing of the protection node until the operation it mediates has all of its input ports bound gives rise to conservative protection. Like speculative and conservative computation these can be controlled within the Triple Manager. 5.2 Tenacious Protection The disadvantage of the fragile protection operator is that potential results are lost if the scheduling constraint is not upheld. Rather than failing, it would be preferable to re-schedule the node for later evaluation, or allow it to be scheduled by another Triple Manager that has authority to assign an appropriate domain. This is achieved by the tenacious protection operation defined in Figure 14. Graph A T is defined recursively. If the scheduling constraint is cdom perm sdom ifel Figure 14: Definition of the Tenacious Protection Operator upheld then node A becomes grafted to the input port of X and may be fired. If the scheduling constraint is not upheld then the result is A, lazily protected, that is, A T . Unlike fragile protection, the tenacious protection operator behaves like a security wrapper that can be repeatedly probed, but can only be unwrapped (scheduled) in an authorized domain The tenacious protection operator could be implemented as a TM primitive. One interpretation is that the TM postpones the scheduling of a node until an authorized domain is available. However, more general interpretations are pos- sible. For example, if the current Triple Manager cannot assign an authorized domain then the protected operation can be scheduled to another Triple Manager that can assign an authorized domain. Example 8. Suppose that a network is partitioned in terms of a clerk subnet and a management subnet and each subnet has its own server which routes traffic to other sub- nets. A Triple Manager on the clerk server starts a PO F F F F Figure 12: Protecting the Ordering Process F I' Figure 13: Protecting the Ordering Process transaction graph, and schedules the requested O operation to a clerk's workstation. Since it cannot schedule a management operation, it passes the 'wrapped' operation to the Triple manager on the management server for, scheduling, which in turn schedules it to an appropriate management workstation. 4 Domain scheduling heuristics, such as that discussed in Example 8 should be considered part of the implementation of the Triple Manager. The tenacious protection operator could be thought of as a scout node that can be sent out across the network looking for a suitable Triple Manager to schedule the protected node. Once found, the underlying Triple Manager transparently retrieves the protected node. An alternative and speculative approach would be to multicast the protection operator across the network; as soon as one Triple Manager can schedule the protected node, the node migrates and all other speculative protection nodes are garbage collected. Low-level protocols to support, what is in effect, remote node invocation has been investigated else- where: a Triple Manager scheduling PVM processes [9] and a traditional dataflow system [13]. Investigating suitable domain scheduling heuristics is a topic for future research. Example 9. Condensed Graphs are used to exploit parallelism in a computation and the Triple Manager(s) can schedule the computation across networks of workstations [10]. Figure 15 gives an example of a graph that schedules a distributed brute-force key search given known plain/cipher text. The key space is split into a series of intervals indexed as Primitive operation cr(Int interval); searches a specified interval for the key. If found the key is returned, otherwise 0 is returned. Operation search is defined recursively and has a high degree of parallelism that can be exploited by the Triple Manager which schedules operation cr to be executed on participating processors. Operation search is passed the initial value maxindex. An organization wishes to use this application to find a particular key. For the purposes of security, search operations may be scheduled only to systems within the organization intranet, while the cr operation may be scheduled to any recognized system. Figure 15 illustrates how these requirements are selectively programmed within an appli- cation. Operations search and cr are assigned permission attributes in and out, respectively. A permission may be associated with a node by introducing an additional permission input port to the node and is illustrated by using ifel ifel searchsearch0 -cr out in Figure 15: Programming Protection a solid input arrow-head. Given the permission ordering (out in), then systems (domains) within the intranet are given permission in, and recognized external systems are given permission out. Schedules implies((in,in),(in,in)), im- plies((in,in),(out,out)) and implies((in,in),(in,out)) hold, while does not. 4 5.3 Emergent Protection Many protection policies base access decisions on previous decisions and/or behaviour, for example Chinese Walls [3] and Dynamic Separation of Duties [12]. Condensed Graphs represent distributed computation and it is preferable not to rely upon a centralized-state approaches such as [4] to provide mechanisms that enforce these requirements. The wrapping protection mechanism, specified in Figure 16, can be used to support, in a distributed fashion, a limited form of dynamic separation of duties. The wrapping operator takes as input a node A, and permits it to fire in any domain y that is strictly more authorized, or has an uncomparable authorization, to x. The result R from firing A is then 'wrapped' as W(x t R). Thus, the first parameter of W is used to continue a local state (for this node) by acting as a high-water mark of the permissions of past domains. Figure 17 gives a snapshot of this mechanism in opera- tion. Given W(A;x) and if we have sdom(A) 6! x, then A is explicitly grafted to the second input port of a new W operation. This makes A fireable, but the non-strictness of this port of W will not graft the resulting output R. This resulting output R of A is protected and may fire only in a suitably different domain, and so forth. This wrapping operation is tenacious and is easily extended to enforce the (v=w z) x A R z S R fires in domain w!!=z, outputs S A fires in domain y!!=x, outputs R R z (z=x y) Figure 17: Snapshot of a Wrapping and Unwrapping sdom sdom ifel Figure Wrapping Protection Mechanism scheduling constraint. Example 10. Consider a simplified version of the purchase order transaction (Figure 18). The order operation takes as input an order-id, and (non-strict) a payment op- eration, and outputs the payment operation P appropriately transformed to include order value, and so forth. The fg O W X Figure Dynamic Separation of Duty order operation is mediated as W(O; fg), where fg is the empty permission. A manager (permission fmgrg) may execute the O operation, and the result is the wrapped node Payment P may now fire only in domains with permissions fclkg or fclk,mgrg. 4 While tailored to a specific requirement, the proposed wrapping mechanism illustrates the flexibility in using Condensed Graphs to specify (and implement) protection re- quirements. Rather than maintaining a centralized security state, the operator W can be thought of as providing emergent protection: mediation results in the emergence of a further protection mechanism to mediate a subsequent op- eration. Investigating how this mechanism might be applied in practice and developing emergent mechanisms for general protection policies is a topic for future research. 6. DISCUSSION AND CONCLUSION The Condensed Graphs model provides a single frame-work in which protection requirements can be specified and implemented within the imperative, availability and coercion paradigms. Section 3 illustrated how using these paradigms provide flexibility in the sequencing and control over security critical operations. Sections 4 and 5 draw on these techniques and develop novel protection mechanisms. A Tena- ciously protected node (operation or data) can be repeatedly probed, and passed around, but may only be unwrapped in the appropriate domain. Referential transparency in the Condensed Graphs model means that this tenacity may be further applied to the results generated by an operation which emerge protected by a mechanism created on the fly. Triple Managers transparently schedule graph operations to appropriate security domains. This allows protection requirements to be coded as part of the graph program, independently of the underlying architecture. Graph-based protection operators such as tenacious protection can be viewed as a protection wrapper that may be unwrapped only in an authorized security domain. Scheduling a tenaciously protected node to an authorized domain is completely trans- parent, even though it may have been necessary to migrate the protected node through a number of Triple Managers before it could be successfully scheduled. Secure WebCom [5] provides one possible implementation of the protection model described in this paper. We- bCom [10] Masters schedule Condensed Graph applications over remote WebCom clients (ancillary processors). Web- Com Masters use KeyNote credentials [2] to determine the operations that the client is authorized to execute; Web- Com master credentials are used by clients to determine if the master had the authorization to schedule the (trusted) mobile-computation that the client is about execute. This implementation can be interpreted in terms of the protection mechanisms described in this paper. Client and Master public keys provide security domains, while credentials define their associated permissions. The authorization check is similar to a fragile mediation on every node in the graph. Much work remains to be done investigating how the protection model described in this paper might be used in prac- tice. The protection model might also be used as part of a conventional secure system. A Condensed Graph can be regarded as a sophisticated job-control language used to schedule operations, such as multilevel transactions, to the protection domains of a separation kernel [14]. Acknowledgements Thanks to the anonymous referees and the Workshop audience for their useful comments on this paper. This research was suported in part by Enterprise Ireland National Software Directorate. --R The keynote trust-management system version 2 The Chinese Wall security policy. The specification and implementation of commercial security requirements including dynamic segregation of duties. Exploiting KeyNote in Web- Com: Architecture neutral glue for trust management Structural models: An introduction to the theory of directed graphs. ACM Operating Systems Review 8 Condensed Graphs: Unifying Availability-Driven Facilitating Parallel Programming in PVM using Condensed Graphs. A Condensed Graphs Engine to Drive Metacomputing. Managing and exploiting speculative computations in a flow driven Some conundrums concerning separation of duty. The design and verification of secure systems. --TR The specification and implementation of MYAMPERSANDldquo;commercialMYAMPERSANDrdquo; security requirements including dynamic segregation of duties Facilitating Parallel Programming in PVM Using Condensed Graphs Protection Design and verification of secure systems

protection mechanisms;condensed graphs;functional and dataflow programming;security models;imperative

508810

A fault-tolerant directory service for mobile agents based on forwarding pointers.

A reliable communication layer is an essential component of a mobile agent system. We present a new fault-tolerant directory service for mobile agents, which can be used to route messages to them. The directory service, based on a technique of forwarding pointers, introduces some redundancy in order to ensure resilience to stopping failures of nodes containing forwarding pointers; in addition, it avoids cyclic routing of messages, and it supports a technique to collapse chains of pointers that allows direct communications between agents. We have formalised the algorithm and derived a fully mechanical proof of its correctness using the proof assistant Coq; we report on our experience of designing the algorithm and deriving its proof of correctness. The complete source code of the proof is made available from the WWW.

INTRODUCTION Mobile agents have emerged as a major programming paradigm for structuring distributed applications [3, 5]. For in- stance, the magnitude project [13] investigates the use of mobile agents as intermediary entities capable of negotiating access to information resources on behalf of mobile users. Several important issues remain to be addressed before mobile agents become a mainstream technology for such appli- cations: among them, a communication system and a security infrastructure are needed respectively for facilitating communications between mobile agents and for protecting agents and their hosts. Here, we focus solely on the problem of communications, for which we have adopted a peer-to-peer communication model using a performative-based agent communication language [11], as prescribed by KQML and FIPA. Various authors Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage, and that copies bear this notice and the full citation of the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. SAC 2002, Madrid, Spain ACM 1-58113-445-2/02/03 . $5.00 have previously investigated a communication layer for mobile agents based on forwarding pointers [16, 10]. In such an approach, when mobile agents migrate, they leave forwarding pointers that are used to route messages. A point of concern is to avoid cyclic routing when agents migrate to previously visited sites; additionally, lazy updates and piggy-backing of information on messages can be used to collapse chains of pointers [12]. For structuring and clarity purposes, a communication layer is usually dened in terms of a message router and a directory service; the latter tracks mobile agents' locations, whereas the former forwards messages using the information provided by the latter. Directory services based on forwarding pointers are currently not tolerant to failures: the failure of a node containing a forwarding pointer may prevent nding agents' positions. The purpose of this paper is to present a directory service, fully distributed and resilient to failures exhibited by intermediary nodes, possibly containing forwarding pointers. This algorithm may be used by a fault-tolerant message router (which itself will be the object of another publication). We consider stopping failures according to which processes are allowed to stop during the course or their execution [7]. The essence of our fault-tolerant distributed directory service is to introduce redundancy of forwarding pointers, typically by making N copies of agents' location information. This type of redundancy ensures the resilience of the algorithm to a maximum of N 1 failures of intermediary nodes. We will show that the complexity of the algorithm remains linear in N . Our specic contributions are: 1. A new directory service based on forwarding pointers, fault-tolerant, preventing cyclic routing, and not involving any static location; 2. A full mechanical proof of its correctness, using the proof assistant Coq [1]; the complete source code of the proof (involving some 25000 tactic invocations) may be downloaded from the following URL [9]. We begin this paper by a survey of background work (Sec- tion 2) and follow by a summary of a routing algorithm based on forwarding pointers (Section 3). We present our new directory service and its formalisation as an abstract machine (Section 4). The purpose of Section 5 is to summarise the correctness properties of the algorithm: its safety states that the distributed directory service correctly and uniquely identies agents' positions, whereas the liveness property shows that the algorithm reaches a stable state after a nite number of transitions, once agents stop mi- grating. Then, in Section 6, we report on our experience of designing the algorithm and deriving its proof of correctness, and we suggest possible variants or extensions. 2. BACKGROUND The topic of mobile agent tracking and communication has been researched extensively by the mobile agent commu- nity. Very early on, location-aware communications were proposed: they consist of sending messages to locations where agents are believed to be, but typically result in failure when the receiver agent has migrated [15, 19]. For a number of applications, such a service is not satisfactory because the key property is to get messages reliably delivered to a recipient, wherever its location and whatever the route adopted (for instance, when two mobile agents undertake a negotiation on how to solve a specic prob- lem). Location-transparent communication services were introduced as a means to route and deliver messages automatically to mobile agents, independently of their migra- tion. (Such services have been shown to be implementable on top of a location-aware communication layer [19].) In the category of location-transparent communication lay- ers, there are essentially two approaches, respectively based on home agents and forwarding pointers. In systems based on home agents, such as Aglets [5], each mobile agent is associated with a non-mobile home agent. In order to communicate with a mobile agent, a message has to be sent to its associated home agent, which forwards it to the mobile one; when a mobile agent migrates, it informs its home agent of its new position. Alternatively, in mobile agent systems such as Voyager [16], agents that migrate leave trails of forwarding pointers, which are used to route messages. In situations such as the pervasive computing environment, the mechanism of a home agent may defeat the purpose of using mobile agents by re-introducing centralisation: the home agent approach puts a burden on the infrastructure, which may hamper its scalability, in particular, in massively distributed systems. A typical illustration is two mobile agents with respective home bases in the US and Europe having to communicate at a host in Australia. In such a scenario, routing via home agents is not desirable, and may not be possible when the host is temporarily disconnected from the network. If we introduce a mechanism by which home agents change location dynamically according to the task at hand, we face the problem of how to communicate reliably with a home agent, which is itself mobile. Alter- natively, we could only use the home agent to bootstrap communication, and then shortcut the route, but this approach becomes unreliable once agents migrate. Finally, the home agent also appears as a single point of failure: when it exhibits a failure, it becomes impossible to track the mobile agent or to route messages to it. A naive forwarding pointer implementation causes communications to become more expensive as agents migrate, because chains of pointers increase. Chains of pointers need to be collapsed promptly so that mobile agents become independent of the hosts they previously visited. Once the chain has collapsed direct communications become possible and avoid the awkward scenario discussed above. As far as tolerance to failures is concerned, the crash of an intermediary node with a forwarding pointer prevents upstream nodes to forward messages. Collapsing chains of pointers also has the benet of reducing the system's exposure to failures. Coordination models oer a more asynchronous form of com- munication, typically involving a tuple space [4]. As coordination spaces are non-mobile, they may suer from the same problem as the home agent; solutions such as distributed spaces may be introduced for that purpose but maintaining consistency is a non-trivial problem. An inconvenient of the coordination approach is that it requires coordinated processes to poll tuple spaces, which may be ine-cient in terms of both communication and computation. As a result, tuple spaces generally provide a mechanism by which registered clients can be notied of the arrival of a new tuple: when clients are mobile, we are back to the problem of how to deliver such notications reliably. If the tuple space itself is mobile [17], the problem is then to deliver messages to the tuple space. This discussion shows that reliable delivery of messages to mobile agents without using static locations to route messages is essential, even if peer-to-peer communications are not adopted as the high-level interaction paradigm between agents. Previous work has focused on formalisation [10] and implementation [16] of forwarding pointers, but solutions were not fault-tolerant. We summarise such an approach in Section 3 before extending it with support for failures in Section 4. 3. OF DIRECTORY SERVICE In this section, we summarise the principles of a communication layer based on forwarding pointers [10] without any fault-tolerance. The algorithm comprises two components: a distributed directory service and a message router, which we describe below. Distributed Directory Service. Each mobile agent is associated with a timestamp that is increased every time the agent migrates. When an agent has autonomously decided to migrate to a new location, it requests the communication layer to transport it to its new destination. When the agent arrives at a new location, an acknowledgement message containing both its new position and its newly- incremented timestamp is sent to its previous location. As a result, for each site, one of the following three cases is valid for each agent A: (i) the agent A is local, (ii) the agent A is in transit but has not acknowledged its new position yet, or (iii) the agent A is known to have been at a remote location with a given timestamp. Timestamps are essential to avoid race conditions between acknowledgement messages: by using timestamps, a site can decide which position information is the most recent, and therefore can avoid creating cycles in the graph of forwarding pointers. In order to avoid an increasing cost of communication when the agent mi- grates, a mechanism was specied to propagate information about agent's position, which in turn reduces the length of chains of pointers [10]. Message Router. Sites rely on the information about agents' positions in order to route messages. For any in-coming message aimed at an agent A, the message will be delivered to A if A is known to be local. If A is in transit, the message will be enqueued, until A's location becomes known; otherwise, the message is forwarded to A's known location. Absence of Fault Tolerance. There is no redundancy in the information concerning an agent's location. Indeed, sites only remember the most recent location of an agent, and only the previous agent's location is informed of the new agent's position after a migration. As a result, a site (transitively) pointing at a site exhibiting a failure has lost its route to the agent. 4. FAULT-TOLERANT ALGORITHM The intuition of our solution to the problem of failures is to introduce some redundancy in the information about agents' positions. Two essential elements are used for this purpose. First, agents remember N previous dierent sites that they have visited; once an agent arrives at a new location, it informs its N previous locations of its new position. Second, sites remember up to N dierent positions for an agent, and their associated timestamps. We shall establish that the algorithm is able to determine the agent's position cor- rectly, provided that the number of stopping failures remains smaller or equal to N 1. Remark We aim to design an algorithm which is resilient to failures of intermediary nodes. We are not concerned with reliability of agents themselves. Systems replicating agents and using failure detectors such as [8] may be used for that purpose; they are complementary to our approach. We adopt an existing framework [10] to model the distributed directory service as an abstract machine, whose state space is summarised in Figure 1. For the sake of clarity, we consider a single mobile agent; the formalisation can easily be extended to multiple agents by introducing names by which agents are being referred to. An abstract machine is composed of a set of sites taking part in a computation. Agent timestamps, which we call mobility counters, are dened as natural numbers. A memory is dened as an association list, associating locations with mobility counters; we represent an empty memory by ;. The value N is a parameter of the algorithm. We will show that the agent's memory has a size N and that the algorithm tolerates at most N 1 failures. The set of messages is inductively dened by two construc- tors. These constructors are used to construct messages, which respectively represent an agent in transit and an arrival acknowledgement. The message representing an agent in transit, typically of the form agent(s; l; ~ contains the site s that the agent is leaving, the value l of the mobility counter it had on that site, and the agent's memory ~ the N previous sites it visited and associated mobility coun- ters. The message representing an arrival acknowledgement, ack(s; l), contains the site s (and associated mobility counter l) where the agent is. We assume that the network is fully connected, that communications are reliable, and that the order of messages in between pairs of sites is preserved. These communication hypotheses are formalised in the abstract machine by point-to-point communication links, which we dene as queues using the following notations, where the expression q1xq2 denotes the concatenation of two queues q1 ; q2 , and first(q) the head of a queue q. Each site maintains some information, which we abstract as \tables" in the abstract machine. The location table maps each site to a memory; for a site s, the location table indicates the sites where s believes the agent has migrated to (with their associated mobility counter). The present table is meant to be empty for all sites, except for the site where the agent is currently located, when the agent is not in transit; there, the present table contains the sites previously visited by the agent. The mobility counter table associates each site with the mobility counter the agent had when it last visited the site; the value is zero if the agent has never visited the site. After the agent has reached a new destination, acknowledgement messages have to be sent to the N previous sites it visited. We decouple the agent's arrival from acknowledgement sending, so that transitions that deal with in-coming messages are dierent from those that generate new messages. Consequently, we introduce a further table, the acknowledgement table, indicating which acknowledgements still have to be sent. In our formalisation, we use a variable to indicate whether a machine is up and running. A site's failure state is allowed to change from false to true, which indicates that the site is exhibiting a failure. We are modelling stopping failures [7] since no transition allows a failure state to change from true to false. A complete conguration of the abstract machine is dened as the Cartesian product of all tables and message queues. Our formalisation can be regarded as an asynchronous distributed system [7]. In a real implementation, tables are not shared resources, but their contents can be distributed at each site. The behaviour of the algorithm is represented by transitions, which specify how the state of the abstract machine evolves. Figure 2 contain all the transitions of the distributed directory service. Transitions are assumed to be executed atom- ically. For convenience, we use some notations such as post , receive or table updates, which give an imperative look to the algorithm; their denitions is as follows. Given a cong- uration hloc present present fail T; ki, such that mob T 0 similar notation is used for other tables. Given a conguration, post(s1 ; s2 ; m) denotes hloc T; present A similar notation is used for receive. sns g (Set of Sites) list(S L) (Memory) (Location Tables) Acknowledgement ~ loc present ack Figure 1: State Space In each rule of Figure 2, the conditions that appear to the left-hand side of an arrow are guards that must be satised in order to be able to re a transition. For instance, the rst four rules contain a proposition of the form :fail T (s), which indicates that the rule has to occur for a site s that is up and running. The right-hand side of a rule denotes the conguration that is reached after transition. We assume that guard evaluation and new conguration construction are performed atomically. In order to illustrate our rules, we present graphical representations of congurations; the rst part of Figure 3 illustrates an agent that has successively visited sites s0 respective timestamps t+2. In this example, we assume that the value of N is 3. (Note that s0 is not represented in the gure.) The rst transition of Figure 2 models the actions to be performed, when an agent decides to migrate from s1 to s2 . In the guard, we see that the present table at s1 must be non-empty, which indicates that the agent is present at s1 . After transition, the present table at s1 is cleared, and an agent message is posted between s1 and s2 ; the message contains the agent's origin s1 , its mobility counter mob T (s1 ), and the previous content of the present table at s1 . Note that s2 , the destination of the agent, is only used to specify which communication channel the agent message must be enqueued into. The site s1 does not need to be communicated this information, nor does it have to remember that site. In a real implementation, the agent message would also contain the complete agent state to be restarted by the re- ceiver. The second part of Figure 3 illustrates changes in the system, when an agent has initiated its migration. The second transition is concerned with s2 handling a messageagent(s3 ; l; ~ M) coming from s1 . Tables are updated to re ect that s2 is becoming the new agent's location, with new mobility counter. Our algorithm prescribes the agent to remember N dierent sites it has visited. As s2 may have been visited recently, we remove s2 from ~ before adding the site s3 where it was located before mi- gration. The call add(N; s; l; ~ adds an association (s; l) to the memory ~ , keeping at most N dierent entries with the highest timestamps. (Appendix A contains the com- 1 Note that s3 is not required to be equal to s1 . Indeed, we want the algorithm to be able to support sites that forward incoming agents to other sites. plete denition of add.) In addition, the acknowledgement table of s2 is updated, since acknowledgements have to be sent back to those previously visited sites. At this point, a proper implementation would reinstate the agent state and resume its execution. The third part of Figure 3 illustrates the system as an agent arrives at a new location. According to the third transition, if the acknowledgement table on s1 contains a pair (s2 ; l 2 ), then an acknowledgement message has to be sent from s1 to s2 ; the acknowledgement message indicates that the agent is on s1 with a mobility counter mob T (s1 ). If a site s2 receives an acknowledgement message about site s3 and mobility counter l, its location table has to be up-dated accordingly. Let us note two properties of this rule. First, we do not require the emitter s1 of the acknowledgement message to be equal to s3 ; this property allows us to use the same message for propagating more information about the agent's location. Second, we make sure that up-dating the location table (i) maintains information about dierent locations, (ii) does not overwrite existing location information with older one. This functionality is implemented by the function add , whose specication may be found in appendix A. According to rule inform of Figure 2, any site s1 believing that the agent is located at site s3 , with a mobility counter l, may elect to communicate its belief to another site s2 . Such a belief is also communicated by an ack message. It is important to distinguish the roles of the send ack and inform transitions. The former is mandatory to ensure the correct behaviour of the algorithm, whereas the latter is optional. The purpose of inform is to propagate information about the agent's location in the system, so that the agent may be found in less steps. As opposed to previous rules, the inform rule is non-deterministic in the destination and location information in an acknowledgement message. At this level, our goal is to dene a correct specication of an algorithm: any implementation strategy will be an instance of this specication; some of them are discussed in Section 6. The rst part of Figure 4 illustrates the states of the system after sending acknowledgement messages, whereas the second one shows the eect of such messages. For a conguration hloc present legal transitions are: migrate agent(s1 present T (s1) in present T (s1) := ; receive in loc T (s2) := ; present T (s2) := S 0 ack T (s2) := S 0 g send ack receive loc Figure 2: Fault-Tolerant Directory Service Failure. The rst ve rules of Figure 2 require the site s where the transition takes place to be up and running, i.e. :fail T (s). Our algorithm is designed to be tolerant to stopping failure, according to which processes are allowed to stop somewhere in the middle of their execution [7]. We model a stopping failure by the transition stop failure, changing the failure state of the site that exhibits the failure. Con- sequently, a site that has stopped will be prevented from performing any of the rst ve transitions of Figure 2. As far as distributed system modelling is concerned, it is unrealistic to consider that messages that are in transit on pres(s loc(s pres(s pres(s loc(s pres(s pres(s pres(s loc(s pres(s loc(s pres(s pres(s pres(s loc(s pres(s loc(s pres(s Figure 3: Agent Migration (part 1) a communication link remain present if the destination of the communication link exhibits a failure. Rule msg failure shows how messages in transit to a stopped site may be lost. A similar argument may also hold for messages that were posted (but not sent yet) at a site that stops. We could add an extra rule handling such a case, but we did not do so in order to keep the number of rules limited. As a result, our communication model can be seen as using buered inputs and unbuered outputs. Initial and Legal Congurations. In the initial con- guration, noted c i , we assume that the agent is at a given site origin with a mobility counter set to N + 1. Obviously, at creation time, an agent cannot have visited N sites previ- ously. Instead, the creation process elects a set S i of dierent sites that act as \backup routers" for the agent in the initial Each site is associated with a dierent mobility counter in the interval [1; N ]. Such N sites could be chosen non-deterministically by the system or could be pres(s pres(s loc(s pres(s loc(s pres(s pres(s loc(s pres(s loc(s pres(s pres(s Figure 4: Agent Migration (part 2) congured manually by the user. For each site in S i , the location table points to the origin and to sites of S i with a higher mobility counter; the location table at all other sites contains the origin and the N 1 rst sites of S i . The present table at origin contains the sites in S i . A detailed formalisation of the initial conguration is available from [9]. A conguration c is said to be legal if there is a sequence of transitions t1 ; is reachable from the initial dene 7! as the re exive, transitive closure of 7!. 5. CORRECTNESS The correctness of the distributed directory service is based on two properties: safety and liveness. The safety of the distributed directory service ensures that it correctly tracks the mobile agent's location, in particular in the presence of failures. The liveness guarantees that agent location information eventually gets propagated. We intuitively explain the safety property proof as follows. An acknowledgement message results in the creation of a forwarding pointer that points towards the agent's loca- tion. Forwarding pointers may be modelled by a relationship parent that denes a directed acyclic graph leading to the agent's location. In the presence of failures, we show that the relationship parent contains su-cient redundancy in order to guarantee the existence of a path leading to the agent, without involving any failed site: (i) Sites that belong to the agent's memory have the agent's location as a parent. (ii) Sites that do not belong to the agent's memory have at least N parents. Consequently, if the number of failures is strictly inferior to N , each site has always at least one parent that is closer to the agent's location; by repeating this argument, we can nd the agent's location. We summarise the liveness result similar to the one in [10]. A nite amount of transitions can be performed from any legal conguration (if we exclude migrate agent and inform). Furthermore, we can prove that, if there is a message at the head of a communication channel, there exists a transition of the abstract machine that consumes that message. Con- sequently, if we assume that message delivery and machine transitions are fair, and if the mobile agent is stationary at a location, then location tables will eventually be updated, which proves the liveness of the algorithm. All proofs were mechanically derived using the proof assistant Coq [1]. Coq is a theorem prover whose logical foundation is constructive logic. The crucial dierence between constructive logic and classical logic is that ::p =) p does not hold in constructive logic. The consequence is that the formulation of proofs and properties must make use of constructive and decidable statements. Due to space restriction, we do not include the proofs but they can be downloaded from [9]. The notation adopted here are pretty-printed versions of the mechanically established ones. 6. ALGORITHMANDPROOFDISCUSSION The constructive proof of the initial algorithm without fault-tolerance helped us understand the dierent invariants that needed to be preserved. In particular, the algorithm maintains a directed acyclic graph leading to the agent's position; interestingly, short-cutting chains of pointers by propagating acknowledgement messages ensures that the graph remains connected and acyclic. Using the same mechanism of timestamp in combination with replication preserves a similar invariant in the presence of failures. The resulting algorithm turned out to be simpler because it uses less rules, and its correctness proof was easier to derive. When N is equal to 1, the algorithm has the same observable behaviour as [10]. From a practical point of view, generating the mechanical proof still remained a tedious process, though simpler, because it needed some 25000 tactic invo- cations, of which 5000 for the formalisation of the abstract machine were reused from our initial work. The complexity of the algorithm is linear in N as far as the number of messages (N acknowledgement messages per mi- gration), message length (size of a memory is O(N)), space per site (size of a memory is O(N)), and time per migration are concerned. Our proof established the correctness in the worst-case scenario. Indeed, the algorithm may tolerate more than N failures provided that one parent, at least, remains up and running for each site. For a given application, the designer will have to choose the value of N . If N is chosen to be equal to the number of nodes in the network, the system will be fully realiable but its complexity, even though linear, is too high on an Internet scale. Instead, an engineering decision should be made: in a practical network, from network statistics, one can derive the probability of obtaining simultaneous fail- ures. For each application, and for the quality of service it requires, the designer selects the appropriate failure proba- bility, which determines the number of simultaneous failures the system should be able to tolerate. A remarkable property of the algorithm is that it does not impose any delay upon agents when they initiate a migra- tion. Forwarding pointers are created temporarily until a stable situation is reached and they are removed. This has to be contrasted with the home agent approach, which requires the agent to notify its homebase, before and after each migration. Interestingly, our algorithm does not preclude us also from using other algorithms; we could envision a system where such algorithms are selected at runtime according to the network conditions and the quality of service requirements of the application. Propagating agent location information with rule inform is critical in order to shorten chains of forwarding pointers, because shorter chains reduce the cost of nding an agent's location. The ideal strategy for sending these messages depends on the type of distributed system, and on the applications using the directory service. A range of solutions is possible and two extremes of the spectrum are easily iden- tiable. In an eager strategy, every time a mobile agent mi- grates, its new location is broadcasted to all other sites; such a solution is clearly not acceptable for networks such as the Internet. Alternatively, a lazy strategy could be adopted [12] but it requires cooperation with the message router. The recipient of a message may inform its emitter, when the recipient observes that that the emitter has out-of-date routing information. In such a strategy, tables are only up-dated when user messages are sent. In Section 4, communication channels in the abstract machine are dened as queues. We have established that swapping any two messages in a given channel does not change the behaviour of the algorithm; in other words, messages do not need to be delivered in order. Message Router. This paper studied a distributed directory service, and we can sketch two possible uses for message routing. Simple Routing. The initial message router [10] can be adopted to the new distributed directory service. A site receiving a message for an agent that is not local forwards the message to the site appearing in its location table with the highest mobility counter; if the location table is empty, messages are accumulated until the table is updated. This simple algorithm does not use the redundancy provided by the directory service and is therefore not tolerant to failure. Parallel Flooding. A site must endeavour to forward a message to N sites. If required, it has to keep copies of messages until N acknowledgements have been received. By making use of redundancy, this algorithm would guarantee the delivery of messages. We should note that the algorithm needs a mechanism to clear messages that have been delivered and are still held by intermediate nodes. Further Related Work. Murphy and Picco [14] present a reliable communication mechanism for mobile agents. Their study is not concerned with nodes that exhibit failures, but with the problem of guaranteeing delivery in the presence of runaway agents. Whether their approach could be combined with ours remains an open question. Lazar et al. [6] migrate mobile agents along a logical hierarchy of hosts, and also use that topology to propagate mes- sages. As a result, they are able to give a logarithmic bound on the number of hops involved in communication. Their mechanism does not oer any redundancy: consequently, stopping failures cannot be handled, though they allow reconnections of temporarily disconnected nodes. Baumann and Rothermel [2] introduce the concept of a shadow as a handle on a mobile agent that allows applications to terminate a mobile agent execution by notifying the termination to its associated shadow. Shadows are also allowed to be mobile. Forwarding pointers are used to route messages to mobile agents and mobile shadows. Some fault-tolerance is provided using a mechanism similar to Jini leases, requiring message to be propagated after some timeout. This differs from our approach that relies on information replication to allow messages to be routed through multiple routes. Mobile computing devices share with mobile agents the problem of location tracking. Prakash and Singhal [18] propose a distributed location directory management scheme that can adapt to changes in geographical distribution of mobile hosts population in the network and to changes in mobile host location query rate. Location information about mobile hosts is replicated at O( m) base stations, where m is the total number of base stations in the system. Mobile hosts that are queried more often than others have their location information stored at a greater number of base stations. The proposed algorithm uses replication to oer improved performance during lookups and updates, but not to provide any form of fault tolerance. 7. CONCLUSION In this paper, we have presented a fault-tolerant distributed directory service for mobile agents. Combined with a message router, it provides a reliable communication layer for mobile agents. The correctness of the algorithm is stated in terms of its safety and liveness. Our formalisation is encoded in the mechanical proof assistant Coq, also used for carrying out the proof of correctness. The constructive proof gives us a very good insight on the al- gorithm, which we want to use to specify a reliable message router. This work is part of an eort to dene a mechanically proven correct mobile agent system. Besides message routing, we also intend to investigate and formalise security and authentication methods for mobile agents. 8. ACKNOWLEDGEMENTS Thanks to Nick Gibbins, Dan Michaelides, Victor Tan and the anonymous referees for their comments. This research is funded in part by QinetiQ and EPSRC Magnitude project (reference GR/N35816). 9. --R The shadow detection protocol for mobile agents. Migratory Applications. Reactive Tuple Spaces for Mobile Agent Coordination Program and Deploying Java Mobile Agents with Aglets. A Scalable Location Tracking and Message Delivery Scheme for Mobile Agents. Distributed Algorithms. Providing Fault Tolerance to Mobile Intelligent Agents. A Fault-Tolerant Distributed Directory Service for Mobile Agents: the Constructive Proof in Coq Distributed Directory Service and Message Router for Mobile Agents. SoFAR with DIM Agents: An Agent Framework for Distributed Information Management. Mobile Objects in Java. MAGNITUDE: Mobile AGents Negotiating for ITinerant Users in the Distributed Enterprise. Reliable Communication for Highly Mobile Agents. An RPC mechanism for transportable agents. LIME: Linda meets mobility. A Dynamic Approach to Location Management in Mobile Computing Systems. Nomadic Pict: Language and Infrastructure Design for Mobile Agents. --TR Distributed directory service and message routing for mobile agents Programming and Deploying Java Mobile Agents Aglets The Shadow Approach Reactive Tuple Spaces for Mobile Agent Coordination Migratory Applications Reliable Communication for Highly Mobile Agents Pict An RPC Mechanism for Transportable Agents --CTR Denis Caromel , Fabrice Huet, An adaptative mechanism for communicating with mobile objects, Proceedings of the 1st French-speaking conference on Mobility and ubiquity computing, June 01-03, 2004, Nice, France Luc Moreau , Peter Dickman , Richard Jones, Birrell's distributed reference listing revisited, ACM Transactions on Programming Languages and Systems (TOPLAS), v.27 n.6, p.1344-1395, November 2005

mobile agents;distributed directory service;fault tolerance

508845

Proxy-based security protocols in networked mobile devices.

We describe a resource discovery and communication system designed for security and privacy. All objects in the system, e.g., appliances, wearable gadgets, software agents, and users have associated trusted software proxies that either run on the appliance hardware or on a trusted computer. We describe how security and privacy are enforced using two separate protocols: a protocol for secure device-to-proxy communication, and a protocol for secure proxy-to-proxy communication. Using two separate protocols allows us to run a computationally-inexpensive protocol on impoverished devices, and a sophisticated protocol for resource authentication and communication on more powerful devices.We detail the device-to-proxy protocol for lightweight wireless devices and the proxy-to-proxy protocol which is based on SPKI/SDSI (Simple Public Key Infrastructure / Simple Distributed Security Infrastructure). A prototype system has been constructed, which allows for secure, yet efficient, access to networked, mobile devices. We present a quantitative evaluation of this system using various metrics.

INTRODUCTION Attaining the goals of ubiquitous and pervasive computing [6, 2] is becoming more and more feasible as the number of computing devices in the world increases rapidly. How- ever, there are still signicant hurdles to overcome when This work was funded by Acer Inc., Delta Electronics Inc., Research Center, and Philips Research under the MIT Project Oxygen partnership, and by DARPA through the O-ce of Naval Research under contract number N66001-99-2-891702. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. SAC 2002, Madrid, Spain integrating wearable and embedded devices into a ubiquitous computing environment. These hurdles include designing devices smart enough to collaborate with each other, increasing ease-of-use, and enabling enhanced connectivity between the dierent devices. When connectivity is high, the security of the system is a key factor. Devices must only allow access to authorized users and must also keep the communication secure when transmitting or receiving personal or private information. Implementing typical forms of secure, private communication using a public-key infrastructure on all devices is difcult because the necessary cryptographic algorithms are CPU-intensive. A common public-key cryptographic algorithm such as RSA using 1024-bit keys takes 43ms to sign and 0.6ms to verify on a 200MHz Intel Pentium Pro (a 32- bit processor) [30]. Some devices may have 8-bit micro-controllers running at 1-4 MHz, so public-key cryptography on the device itself may not be an option. Nevertheless, public-key based communication between devices over a net-work is still desirable. This paper presents our approach to addressing these is- sues. We describe the architecture of our resource discovery and communication system in Section 2. The device-to- proxy security protocol is described in Section 3. We review SPKI/SDSI and present the proxy-to-proxy protocol that uses SPKI/SDSI in Section 4. Related work is discussed in Section 5. The system is evaluated in Section 6. 1.1 Our Approach To allow the architecture to use a public-key security model on the network while keeping the devices themselves simple, we create a software proxy for each device. All objects in the system, e.g., appliances, wearable gadgets, software agents, and users have associated trusted software proxies that either run on an embedded processor on the appliance, or on a trusted computer. In the case of the proxy running on an embedded processor on the appliance, we assume that device to proxy communication is inherently secure. 1 If the device has minimal computational power, 2 and communicates to its proxy through a wired or wireless network, we force the communication to adhere to a device- to-proxy protocol (cf. Section 3). Proxies communicate with 1 For example, in a video camera, the software that controls various actuators runs on a powerful processor, and the proxy for the camera can also run on the embedded processor. 2 This is typically the case for lightweight devices, e.g., remote controls, active badges, etc. each other using a secure proxy-to-proxy protocol based on SPKI/SDSI (Simple public Key Infrastructure / Simple Distributed Infrastructure). Having two dierent protocols allows us to run a computationally-inexpensive security protocol on impoverished devices, and a sophisticated protocol for resource authentication and communication on more powerful devices. We describe both protocols in this paper. 1.2 Prototype Automation System Using the ideas described above, we have constructed a prototype automation system which allows for secure, yet e-cient, access to networked, mobile devices. In this system, each user wears a badge called a K21 which identies the user and is location-aware: it \knows" the wearer's location within a building. User identity and location information is securely transmitted to the user's software proxy using the device-to-proxy protocol. Devices themselves may be mobile and may change loca- tions. Attribute search over all controllable devices can be performed to nd the nearest device, or the most appropriate device under some metric. 3 By exploiting SPKI/SDSI, security is not compromised as new users and devices enter the system, or when users and devices leave the system. We believe that the use of two different protocols, and the use of the SPKI/SDSI framework in the proxy-to-proxy protocol has resulted in a secure, scal- able, e-cient, and easy-to-maintain automation system. 2. SYSTEM ARCHITECTURE The system has three primary component types: devices, proxies and servers. A device refers to any type of shared network resource, either hardware or software. It could be a printer, a wireless security camera, a lamp, or a software agent. Since communication protocols and bandwidth between devices can vary widely, each device has a unique proxy to unify its interface with other devices. The servers provide naming and discovery facilities to the various devices We assume a one-to-one correspondence between devices and proxies. We also assume that all users are equipped with K21s, whose proxies run on trusted computers. Thus our system only needs to deal with devices, proxies and the server network. The system we describe is illustrated in Figure 1. 2.1 Devices Each device, hardware or software, has an associated trusted software proxy. In the case of a hardware device, the proxy may run on an embedded processor within the de- vice, or on a trusted computer networked with the device. In the case of a software device, the device can incorporate the proxy software itself. Each device communicates with its own proxy over the appropriate protocol for that particular device. A printer wired into an Ethernet can communicate with its proxy using TCP/IP. A wireless camera uses a wireless protocol for the same purpose. The K21 (a simple device with a lightweight processor) communicates with its proxy using the particular device-to-proxy protocol described in Section 3 For example, a user may wish to print to the nearest printer that he/she has access to. Event Play Tape Play Tape Event Device K21 VCR Proxy K21 Proxy Proxy Farm VCR Proxy Device-to-proxy protocol Server Network Name Resolution Routing (Section Device-to-proxy protocol Proxy-to-proxy (Section protocol Figure 1: System Overview 3. Thus, the device-side portion of the proxy must be customized for each particular device. 2.2 Proxy The proxy is software that runs on a network-visible com- puter. The proxy's primary function is to make access-control decisions on behalf of the device it represents. It may also perform secondary functions such as running scripted actions on behalf of the device and interfacing with a directory service. The proxy provides a very simple API to the device. The sendToProxy() method is called by the device to send messages to the proxy. The sendToDevice() method is a called by the proxy to send messages to the device. When a proxy receives a message from another proxy, depending on the message, the proxy may translate it into a form that can be understood by the proxy's particular device. It then forwards the message to the device. When a proxy receives a message from its device, it may translate the message into a general form understood by all proxies, and then forward the message to other proxies. Any time a proxy receives a mes- sage, before performing a translation and passing the message on to the device, it performs the access control checks described in Section 4. For ease of administration, we group proxies by their ad- ministrators. An administrator's set of proxies is called a proxy farm. This set specically includes the proxy for the administrator's K21, which is considered the root proxy of the proxy farm. When the administrator adds a new device to the system, the device's proxy is automatically given a default ACL, a duplicate of the ACL for the administra- tor's K21 proxy. The administrator can manually change the ACL later, if he desires. A noteworthy advantage of our proxy-based architecture is that it addresses the problem of viruses in pervasive computing environments. Sophisticated virus scanning software can be installed in the proxy, so it can scan any code before it is downloaded onto the device. CommandEvent Used to instruct a device to turn on or o, for example. ErrorEvent Generated and broadcast to all listeners when an error condition occurs. StatusChangeEvent Generated when, for example, a device changes its location. QueryEvent When a server receives a QueryEvent, it performs a DNS (Domain Name Service) or INS lookup on the query, and returns the results of the lookup in a ResponseEvent. ResponseEvent Generated in response to a QueryEvent. Figure 2: Predened Event Types 2.3 Servers and the Server Network This network consists of a distributed collection of independent name servers and routers. In fact, each server acts as both a name server and a router. This is similar to the name resolvers in the Intentional Naming System (INS) [1], which resolve device names to IP addresses, but can also route events. If the destination name for an event matches multiple proxies, the server network will route the event to all matching destinations. When a proxy comes online, it registers the name of the device it represents with one of these servers. When a proxy uses a server to perform a lookup on a name, the server searches its directory for all names that match the given name, and returns their IP addresses. 2.4 Communication via Events We use an event-based communication mechanism in our system. That is, all messages passed between proxies are signals indicating that some event has occurred. For example, a light bulb might generate light-on and light-o events. To receive these messages, proxy x can add itself as an eventlistener to proxy y. Thus, when y generates an event, x will receive a copy. In addition, the system has several pre-dened event categories which receive special treatment at either the proxy or server layer. They are summarized in Figure 2. A developer can dene his own events as well. The server network simply developer-dened events through to their destination. The primary advantage of the event-based mechanism is that it eliminates the need to repeatedly poll a device to determine changes in its status. Instead, when a change oc- curs, the device broadcasts an event to all listeners. Systems like Sun Microsystems' Jini [26] issue \device drivers" (RMI stubs) to all who wish to control a given device. It is then possible to make local calls on the device driver, which are translated into RMI calls on the device itself. 2.5 Resource discovery The mechanism for resource discovery is similar to the resource discovery protocol used by Jini. When a device comes online, it instructs its proxy to repeatedly broadcast a request for a server to the local subnetwork. The request contains the device's name and the IP address and port of its proxy. When a server receives one of these requests, it issues a lease to the proxy. 4 That is, it adds the name/IP address pair to its directory. The proxy must periodically renew 4 Handling the scenario where the device is making false claims about its attributes in the lease request packet is Gateway Gateway Device 1 Device 3 Device 2 Proxy Farm Proxy 3 UDP RF Figure 3: Device-to-Proxy Communication overview its lease by sending the same name/IP address pair to the server, otherwise the server removes it from the directory. In this fashion, if a device silently goes oine, or the IP address changes, the proxy's lease will no longer get renewed and the server will quickly notice and either remove it from the directory or create a new lease with the new IP address. For example, imagine a device with the name [name=foo] which has a proxy running on 10.1.2.3:4011. When the device is turned on, it informs its proxy that it has come online, using a protocol like the device-to-proxy protocol described in Section 3. The proxy begins to broadcast lease-request packets of the form h[name=foo], 10.1.2.3:4011i on the local subnetwork. When (or if) a server receives one of these pack- ets, it checks its directory for [name=foo]. If [name=foo] is not there, the server creates a lease for it by adding the name/IP address pair to the directory. If [name=foo] is in the directory, the server renews the lease. Suppose at some later time the device is turned o. When the device goes down, it brings the proxy oine with it, so the lease request packets no longer get broadcast. That device's lease stops getting renewed. After some short, pre-dened period of time, the server expires the unrenewed lease and removes it from the directory. 3. DEVICE-TO-PROXY PROTOCOL FOR WIRELESS DEVICES 3.1 Overview The device-to-proxy protocol varies for dierent types of devices. In particular, we consider lightweight devices with low-bandwidth wireless network connections and slow CPUs, and heavyweight devices with higher-bandwidth connections and faster CPUs. We assume that heavyweight devices are capable of running proxy software locally (i.e., the proxy for a printer could run on the printer's CPU). With a local proxy, a sophisticated protocol for secure device-to-proxy communication is unnecessary, assuming critical parts of the device are tamper resistant. For lightweight devices, the proxy must run elsewhere. This section gives an overview of a protocol which is low-bandwidth and not CPU-intensive that we use for lightweight device-to-proxy communication. 3.2 Communication Our prototype system layers the security protocol described below over a simple radio frequency (RF) protocol. The the subject of ongoing research. RF communication between a device and its proxy is handled by a gateway that translates packetized RF communication into UDP/IP packets, which are then routed over the network to the proxy. The gateway also works in the opposite direction by converting UDP/IP packets from the proxy into RF packets and transmitting them to the device. An overview of the communication is shown in Figure 3. This gure shows a computer running three proxies; one for each of three separate devices. The gure also shows how multiple gateways can be used; device A is using a dierent gateway from devices B and C. 3.3 Security The proxy and device communicate through a secure channel that encrypts and authenticates all the messages. The HMAC-MD5 [13][20] algorithm is used for authentication and the RC5 [21] algorithm is used for encryption. Both of these algorithms use symmetric keys; the proxy and the device share 128-bit keys. 3.3.1 Authentication HMAC (Hashed Message Authentication Code) produces a MAC (Message Authentication Code) that can validate the authenticity and integrity of a message. HMAC uses secret keys, and thus only someone who knows a particular can create a particular MAC or verify that a particular MAC is correct. 3.3.2 Encryption The data is encrypted using the RC5 encryption algo- rithm. We chose RC5 because of its simplicity and perfor- mance. Our RC5 implementation is based on the OpenSSL [16] code. RC5 is a block cipher; it usually works on eight-byte blocks of data. However, by implementing it using output feedback (OFB) mode, it can be used as a stream cipher. This allows for encryption of an arbitrary number of bytes without having to worry about blocks of data. OFB mode works by generating an encryption pad from an initial vector and a key. The encryption pad is then XOR'ed with the data to produce the ciphertext. Since the ciphertext can be decrypted by producing the same encryption pad and XOR'ing it with the ciphertext. Since this only requires the RC5 encryption routines to generate the encryption pad, separate encrypt and decrypt routines are not required. For our implementation, we use 16 rounds for RC5. We use dierent 128-bit keys for encryption and authentication. 3.4 Location Device location is determined using the Cricket location system[18, 17]. Cricket has several useful features, including user privacy, decentralized control, low cost, and easy deployment. Each device determines its own location. It is up to the device to decide if it wants to let others know where it is. In the Cricket system, beacons are placed on the ceilings of rooms. These beacons periodically broadcast location information (such as \Room 4011") that can be heard by Cricket listeners. At the same time that this information is broadcast in the RF spectrum, the beacon also broadcasts an ultrasound pulse. When a listener receives the RF mes- sage, it measures the time until it receives the ultrasound pulse. The listener determines its distance to the beacon using the time dierence. 4. PROXY TO PROXY PROTOCOL SPKI/SDSI (Simple public Key Infrastructure/Simple Distributed Security Infrastructure) [7, 22] is a security infrastructure that is designed to facilitate the development of scalable, secure, distributed computing systems. SPKI/SDSI provides ne-grained access control using a local name space architecture and a simple, exible, trust policy model. SPKI/SDSI is a public key infrastructure with an egalitarian design. The principals are the public keys and each public key is a certicate authority. Each principal can issue certicates on the same basis as any other principal. There is no hierarchical global infrastructure. SPKI/SDSI communities are built from the bottom-up, in a distributed manner, and do not require a trusted \root." 4.1 SPKI/SDSI Integration We have adopted a client-server architecture for the prox- ies. When a particular principal, acting on behalf of a device or user, makes a request via one proxy to a device represented by another proxy, the rst proxy acts like a client, and the second as a server. Resources on the server are either public or protected by SPKI/SDSI ACLs. A SPKI/SDSI ACL consists of a list of entries. Each entry has a subject (a key or group) and a tag which species the set of operations that that key or group is allowed to perform. To gain access to a resource protected by an ACL, a requester must include, in his request, a chain of certicates demonstrating that he is a member of a group in an entry on the ACL. 5 If a requested resource is protected by an ACL, the princi- pal's request must be accompanied by a \proof of authentic- ity" that shows that it is authentic, and a \proof of autho- rization" that shows the principal is authorized to perform the particular request on the particular resource. The proof of authenticity is typically a signed request, and the proof of authorization is typically a chain of certicates. The principal that signed the request must be the same principal that the chain of certicates authorizes. This system design, and the protocol between the proxies, is very similar to that used in SPKI/SDSI's Project Geron- imo, in which SPKI/SDSI was integrated into Apache and Netscape, and used to provide client access control over the web. Project Geronimo is described in two Master's theses [3, 14]. 4.2 Protocol The protocol implemented by the client and server proxies consists of four messages. This protocol is outlined in Figure 4, and following is its description: 1. The client proxy sends a request, unauthenticated and unauthorized, to the server proxy. 2. If the client requests access to a protected resource, the server responds with the ACL protecting the resource 6 5 For examples of SPKI/SDSI ACLs and certicates, see [7] or [3]. 6 The ACL itself could be a protected resource, protected by another ACL. In this case, the server will return the latter ACL. The client will need to demonstrate that the user's key is on this ACL, either directly or via certicates, before gaining access to the ACL protecting the object to which access was originally requested. and the tag formed from the client's request. A tag is a SPKI/SDSI data structure which represents a set of re- quests. There are examples of tags in the SPKI/SDSI IETF drafts [7]. If there is no ACL protecting the requested resource, the request is immediately honored. 3. (a) The client proxy generates a chain of certicates using the SPKI/SDSI certicate chain discovery algorithm [4, 3]. This certicate chain provides a proof of authorization that the user's key is authorized to perform its request. The certicate chain discovery algorithm takes as input the ACL and tag from the server, the user's public key (principal), the user's set of cer- ticates, and a timestamp. If it exists, the algorithm returns a chain of user certicates which provides proof that the user's public key is authorized to perform the operation(s) specied in the tag, at the time specied in the timestamp. If the algorithm is unable to generate a chain because the user does not have the necessary certi- cates, 7 or if the user's key is directly on the ACL, the algorithm returns an empty certicate chain. The client generates the timestamp using its local clock. (b) The client creates a SPKI/SDSI sequence [7] consisting of the tag and the timestamp. It signs this sequence with the user's private key, and includes a copy of the user's public key in the SPKI/SDSI signature. The client then sends the tag-time- stamp sequence, the signature, and the certicate chain generated in step 3a to the server. 4. The server veries the request by: (a) Checking the timestamp in the tag-timestamp sequence against the time on the server's local clock to ensure that the request was made recently. 8 (b) Recreating the tag from the client's request and checking that it is the same as the tag in the tag- timestamp sequence. (c) Extracting the public key from the signature. (d) Verifying the signature on the tag-timestamp sequence using this key. Validating the certicates in the certicate chain. (f) Verifying that there is a chain of authorization from an entry on the ACL to the key from the signature, via the certicate chain presented. The authorization chain must authorize the client to perform the requested operation. 7 If the user does not have the necessary certicates, the client could immediately return an error. In our design, however, we choose not to return an error at this point; instead, we let the client send an empty certicate chain to the server. This way, when the request does not verify, the client can possibly be sent some error information by the server which lets the user know where he should go to get valid certicates. 8 In our prototype implementation, the server checks that the timestamp in the client's tag-timestamp sequence is within ve minutes of the server's local time. 4. Server verifies request. If the request is verified, it is honored. If the request does not verify, it is denied and an error is returned. 2. Server verification fails. ACL and tag are returned. chain. Client signs request. Client sends signed request with certificates. 3. Client uses ACL and tag to generate certificate Client Proxy Server Proxy ({tag, timestamp} , certificate chain) (requested resource / error) (ACL, tag) 1. Initial unauthenticated, unauthorized request Ku Figure 4: SPKI/SDSI Proxy to Proxy Access Control Protocol If the request veries, it is honored. If it does not verify, it is denied and the server proxy returns an error to the client proxy. This error is returned whenever the client presents an authenticated request that is denied. The protocol can be viewed as a typical challenge-response protocol. The server reply in step 2 of the protocol is a challenge the server issues the client, saying, \You are trying to access a protected le. Prove to me that you have the credentials to perform the operation you are requesting on the resource protected by this ACL." The client uses the ACL to help it produce a certicate chain, using the SPKI/SDSI certicate chain discovery algorithm. It then sends the cer- ticate chain and signed request in a second request to the server proxy. The signed request provides proof of authentic- ity, and the certicate chain provides proof of authorization. The server attempts to verify the second request, and if it succeeds, it honors the request. The timestamp in the tag-timestamp sequence helps to protect against certain types of replay attacks. For example, suppose the server logs requests and suppose that this log is not disposed of properly. If an adversary gains access to the logs, the timestamp prevents him from replaying requests found in the log and gaining access to protected resources. 9 4.2.1 Additional Security Considerations The SPKI/SDSI protocol, as described, addresses the issue of providing client access control. The protocol does not ensure condentiality, authenticate servers, or provide protection against replay attacks from the network. The Secure Sockets Layer (SSL) protocol is the most widely used security protocol today. The Transport Layer Security (TLS) protocol is the successor to SSL. Principal goals of SSL/TLS [19] include providing condentiality and data integrity of tra-c between the client and server, and providing authentication of the server. There is support for client 9 In order to use timestamps, the client's clock and server's clock need to be fairly synchronized; SPKI/SDSI already makes an assumption about fairly synchronized clocks when time periods are specied in certicates. An alternative approach to using timestamps is to use nonces in the protocol. SPKI/SDSI Access Control Protocol Application Protocol Key-Exchange Protocol with Server Authentication TCP/IP Figure 5: Example Layering of Protocols authentication, but client authentication is optional. The SPKI/SDSI Access Control protocol can be layered over a key-exchange protocol like TLS/SSL to provide additional security. TLS/SSL currently uses the X.509 PKI to authenticate servers, but it could just as well use SPKI/SDSI in a similar manner. In addition to the features already stated, SSL/TLS also provides protection against replay attacks from the network, and protection against person-in- the-middle attacks. With these considerations, the layering of the protocols is shown in Figure 5. In the gure, 'Applica- tion Protocol' refers to the standard communication protocol between the client and server proxies, without security. SSL/TLS authenticates the server proxy. However, it does not indicate whether the server proxy is authorized to accept the client's request. For example, it may be the case that the client proxy is requesting to print a 'top secret' docu- ment, say, and only certain printers should be used to print 'top secret' documents. With SSL/TLS and the SPKI/SDSI Client Access Control Protocol we have described so far, the client proxy will know that the public key of the proxy with which it is communicating is bound to a particular address, and the server proxy will know that the client proxy is authorized to print to it. However, the client proxy still will not know if the server proxy is authorized to print 'top se- cret' documents. If it sends the `top secret' document to be printed, the server proxy will accept the document and print it, even though the document should not have been sent to it in the rst place. To approach this problem, we propose extending the SPKI- /SDSI protocol so that the client requests authorization from the server and the server proves to the client that it is authorized to handle the client's request (before the client sends the document o to be printed). To extend the proto- col, the SPKI/SDSI protocol described in Section 4.2 is run from the client proxy to the server proxy, and then run in the reverse direction, from the server proxy to the client proxy. Thus, the client proxy will present a SPKI/SDSI certicate chain proving that it is authorized to perform its request, and the server proxy will present a SPKI/SDSI certicate chain proving that it is authorized to accept and perform the client's request. Again, if additional security is needed, the extended protocol can be layered over SSL/TLS. Note that the SPKI/SDSI Access Control Protocol is an example of the end-to-end argument [23]. The access control decisions are made in the uppermost layer, involving only the client and the server. 5. RELATED WORK 5.1 Device to Proxy Communication The Resurrecting Duckling is a security model for ad-hoc wireless networks [25, 24]. In this model, when devices begin their lives, they must be \imprinted" before they can be used. A master (the mother duck) imprints a device (the duckling) by being the rst one to communicate with it. After imprinting, a device only listens to its master. During the process of imprinting, the master is placed in physical contact with the device and they share a secret key that is then used for symmetric-key authentication and encryption. The master can also delegate the control of a device to other devices so that control is not always limited to just the mas- ter. A device can be \killed" by its master then resurrected by a new one in order for it to swap masters. 5.2 Proxy to Proxy Communication Jini [26] network technology from Sun Microsystems centers around the idea of federation building. Jini avoids the use of proxies by assuming that all devices and services in the system will run the Java Virtual Machine. The project [8] at the Helsinki University of Technology has succeeded in building a framework for integrating Jini and SPKI/SDSI. Their implementation has some latency concerns, however, when new authorizations are granted. UC Berkeley's Ninja project [27] uses the Service Discovery Service [5] to securely perform resource discovery in a wide-area network. Other related projects include Hewlett- Packard's CoolTown [9], IBM's TSpaces [11] and University of Washington's Portolano [29]. 5.3 Other projects using SPKI/SDSI Other projects using SPKI/SDSI include Hewlett-Pack- ard's e-Speak product [10], Intel's CDSA release [12], and Berkeley's OceanStore project [28]. HP's eSpeak uses SPKI- /SDSI certicates for specifying and delegating authoriza- tions. Intel's CDSA release, which is open-source, includes a SPKI/SDSI service provider for building certicates, and a module (AuthCompute) for performing authorization com- putations. OceanStore uses SPKI/SDSI names in their naming architecture. 6. EVALUATION 6.1 Hardware Design Details on the design of a board that can act as the core of a lightweight device, or as a wearable communicator, are given in [15]. 6.2 Device-to-Proxy Protocol In this section we evaluate the device-to-proxy protocol described in Section 3 in terms of its memory and processing requirements. 6.2.1 Memory Requirements Table breaks down the memory requirements for various software components. The code size represents memory used in Flash, and data size represents memory used in RAM. The device functionality component includes the packet and location processing routines. The RF code component includes the RF transmit and receive routines as well as the Cricket listener routines. The miscellaneous component is code that is common to all of the other components. The device code requires approximately 12KB of code space and 1KB of data space. The security algorithms, HMAC-MD5 and RC5, take up most of the code space. Component Code Size Data Size Device Functionality 2.0 191 RF Code 1.1 153 RC5 3.2 256 Miscellaneous 1.0 0 Total 11.9 986 Table 1: Code and data size on the Atmel processor Function Time (ms) Clock Cycles decrypt (n bytes) 0:163n up to 56 bytes 11.48 45,920 Table 2: Performance of encryption and authentication code Both of these algorithms were optimized in assembly, which reduced their code size by more than half. The code could be better optimized, but this gives a general idea of how much memory is required. The code size we have attained is small enough that it can be incorporated into virtually any device. 6.2.2 Processing Requirements The security algorithms put the most demand on the device Table breaks down the approximate time for each algorithm as detailed in [15]. The RC5 processing time varies linearly with the number of bytes being encrypted or decrypted. The HMAC-MD5 routine, on the other hand, takes a constant amount of time up to 56 bytes. This is because HMAC-MD5 is designed to work on blocks of data, so anything less than 56 bytes is padded. Since we limit the RF packet size to 50 bytes, we only analyze the HMAC- MD5 running time for packets of size less than or equal to 50 bytes. 6.3 SPKI/SDSI Evaluation The protocol described in Section 4 is e-cient. The rst two steps of the protocol are a standard request/response no cryptography is required. The signicant steps in the protocol are step 3, in which a certicate chain is formed, and step 4, where the chain is veried. Table 3 shows analyses of these two steps. The paper on Certicate Chain Discovery in SPKI/SDSI [4] should be referred to for a discussion of the timing analyses. The CPU times are approximate times measured on a Sun Microsystems Ultra-1 running SunOS 5.7. 7. CONCLUSIONS We believe that the trends in pervasive computing are increasing the diversity and heterogeneity of networks and their constituent devices. Developing security protocols that can handle diverse, mobile devices networked in various ways represents a major challenge. In this paper, we have taken a rst step toward meeting this challenge by observing the need for multiple security protocols, each with dierent characteristics and computational requirements. While we have described a prototype system with two dierent protocols, other types of protocols could be included if deemed necessary The two protocols we have described have vastly dier- ent characteristics, because they apply to dierent scenar- ios. The device-to-proxy protocol was designed to enable secure communication of data from a lightweight device. The SPKI/SDSI-based proxy-to-proxy protocol was designed to provide exible, ne-grained, access control between prox- ies. The proxy architecture and the use of two dierent protocols has resulted, we believe, in a secure, yet e-cient, resource discovery and communication system. 8. --R The Design and Implementation of an Intentional Naming System. An Application Model for Pervasive Computing. SPKI/SDSI HTTP Server An Architecture for a Secure Service Discovery Service. The Future of Computing. Simple public Key Certi Decentralized Jini Security. See http://cooltown. See http://www. Intelligent Connectionware. An Implementation of a Secure Web Client Using SPKI/SDSI Certi An Architecture and Implementation of Secure Device Communication in Oxygen. Providing Precise Indoor Location Information to Mobile Devices. The Cricket Location-Support System SSL and TLS: Designing and Building Secure Systems. The MD5 Message-Digest Algorithm The RC5 Encryption Algorithm. The Resurrecting Duckling - What next? <Proceedings>In Proc The Resurrecting Duckling: Security Issues for Ad-hoc Wireless Networks Sun Microsystems Inc. The Ninja Project: Enabling Internet-scale Services from Arbitrarily Small Devices The OceanStore Project: Providing Global-Scale Persistent Data University of Washington. Performance Comparison of public-key Cryptosystems --TR An architecture for a secure service discovery service The design and implementation of an intentional naming system The Cricket location-support system Challenges SSL and TLS Certificate chain discovery in SPKI?SDSI The Resurrecting Duckling - What Next? The Resurrecting Duckling --CTR Joerg Abendroth , Christian D. Jensen, A unified security framework for networked applications, Proceedings of the ACM symposium on Applied computing, March 09-12, 2003, Melbourne, Florida Sanjay Raman , Dwaine Clarke , Matt Burnside , Srinivas Devadas , Ronald Rivest, Access-controlled resource discovery for pervasive networks, Proceedings of the ACM symposium on Applied computing, March 09-12, 2003, Melbourne, Florida Taejoon Park , Kang G. Shin, LiSP: A lightweight security protocol for wireless sensor networks, ACM Transactions on Embedded Computing Systems (TECS), v.3 n.3, p.634-660, August 2004 Feng Zhu , Matt Mutka , Lionel Ni, Facilitating secure ad hoc service discovery in public environments, Journal of Systems and Software, v.76 n.1, p.45-54, April 2005 Kui Ren , Wenjing Lou, Privacy-enhanced, attack-resilient access control in pervasive computing environments with optional context authentication capability, Mobile Networks and Applications, v.12 n.1, p.79-92, January 2007 Georgios Kambourakis , Angelos Rouskas , Stefanos Gritzalis, Experimental Analysis of an SSL-Based AKA Mechanism in 3G-and-Beyond Wireless Networks, Wireless Personal Communications: An International Journal, v.29 n.3-4, p.303-321, June 2004 Domenico Cotroneo , Almerindo Graziano , Stefano Russo, Security requirements in service oriented architectures for ubiquitous computing, Proceedings of the 2nd workshop on Middleware for pervasive and ad-hoc computing, p.172-177, October 18-22, 2004, Toronto, Ontario, Canada Arun Kejariwal , Sumit Gupta , Alexandru Nicolau , Nikil Dutt , Rajesh Gupta, Proxy-based task partitioning of watermarking algorithms for reducing energy consumption in mobile devices, Proceedings of the 41st annual conference on Design automation, June 07-11, 2004, San Diego, CA, USA

protocol;mobile device;ubiquitous;authorization;certificate;wireless;certificate chain discovery;certificate chain;proxy;security;pervasive

508854

A comprehensive model for arbitrary result extraction.

Within the realms of workflow management and grid computing, scheduling of distributed services is a central issue. Most schedulers balance time and cost to fit within a client's budget, while accepting explicit data dependencies between services as the best resolution for scheduling. Results are extracted from one service in total, and then simply forwarded to the next service. However, distributed objects and remote services adhere to various standards for data delivery and result extraction. There are multiple means of requesting results and multiple ways of delivering those results. By examining several popular and idiosyncratic methods, we have developed a comprehensive model that combines the functionality of all component models. This model for arbitrary result extraction from distributed objects provides increased flexibility for object users, and an increased audience for module providers. In turn, intelligent schedulers may leverage these result extraction features.

INTRODUCTION 1.1 Traditional RPCs and asynchronous extraction We address the problem of obtaining results from any of multiple computational servers in response to requests made by a client program. The simplest form of result extraction is the traditional synchronous remote procedure call (RPC). Parameters are passed in, the client waits patiently for the results, and finally all results are simultaneously available. Only a single object is returned with certain function calls, but most languages offer procedure calls where you can have more than one OUT-parameter. It is even possible in C++ with the use of pointers, and it is cleanly implemented in CORBA. This has been the dominant form of result extraction in most programming languages and for many distributed systems (e.g. CORBA, where a typical procedure call is defined in the IDL syntax like void mymethod(in TYPE1 param1, out TYPE2 param2, out TYPE3 param3). 1.2 Alternative extraction models There are other models of data extraction that have received little attention in most programming languages and paradigms. These alternative models generally are strictly more functional than the traditional procedure call, and are being used increasingly as a replacement to RPC-style result extraction. They all, in at least some aspect, provide more functionality or flexibility than the RPC. One of the most important enhancements has been the addition of asynchrony. Clients that do not have to stall while results are computed or delivered are strictly more powerful than those that must. Asynchronous result extraction has been used even in many sequential computing languages like Java and Ada, and is enabled in various ways, through message passing, threading, rendezvous, etc. Asynchrony as a tool to delay or reorder result extraction is well understood and widely used [12, 13]. Partial extraction of data results has become almost ubiquitous with the advent of web browsing. There are examples of other extraction models (such as the progressive extraction of results) that have seen limited use in specialized arenas. We can also envision other models encompassing some of the more esoteric extraction models that are potential more flexible than any in wide or limited use today. Perhaps the most important of these model we refer to is "partial extraction." Partial extraction is taking only the desired portions of a result set, thus saving the costs associated with extracting the entire set. For instance, almost all modern web browsers have the ability to download only text, without images, to speed browsing on slower connections. Many browsers also allow users to filter unwanted objects (i.e., embedded audio) out of html documents. The web has brought "partial extraction" to the desktop as the default browsing model. Another model for result extraction, again strictly more powerful than the traditional RPC, is the "progressive extraction" model. In many scientific computations, such as adaptive mesh refinements, answers become "better" (e.g., more precise) over time. It can be very important to extract results as progress is made toward an acceptable solution for steering, early termination, or other reasons. A typical application of this type is the design of an aircraft wing [2, 3, 5, 12]. Certain scientific and mathematical processes, like Newton's method of successive approximation for roots of an equation, can also utilize this type of extraction [4]. The traditional RPC result extraction model does not lend itself well to progressive extractions. Decomposition of the traditional call model has been discussed for some time [10]; we advocate a further breakdown of the extraction phase of this model. We propose an extraction model, developed in the course of research and development of the language CLAM (Composition Language for Autonomous Megamodules) [8] within the CHAIMS (Composing High-level Access Interfaces for Multisite Software) megaprogramming project [1], that encompasses these three important augmentations (asynchronous, partial, progressive) to RPC-style result extraction models. The amalgamation of these three extraction paradigms leads to a general model, more expressive than any of the three taken alone. 1.3 Current system support for partial and progressive extraction For many languages and systems, the generic asynchronous remote procedure call model of result extraction provides enough flexibility. In cases where it does not, custom solutions abound. Occasionally, these custom solutions become widespread, often circumstantially, e.g., because of runtime support rather than language specification. For instance, partial extraction within the available community of web-browsers does not arise from html, per se. It is a consequence of parsing html pages and making only selective http requests. In effect, the web-browser implements the partial extraction of results while the html provides a "schema" of the results available. For example, when web users choose not to download certain types of content, the web-browser implements the filtering and is not affected by the html or the http protocol. We accept this html/http/browsing model as appropriate for partial extractions when a schema for the data is available. Of course, without such a schema, partial extractions would be meaningless. At some level, this constrains the domain for which partial extraction is semantically meaningful or even practical. However, since the model of partial extraction wholly encompasses the traditional RPC method of result extraction, this is not a problem. Progressive extractions are frequently consequences of elaborate development projects or again arise as a consequence of the nature of the data involved. We can look again to web-browsing to find (an extremely limited) form of progressive extraction, one that arises from the data at hand rather than html or http. Within certain result sets, like weather maps, the results change over time. By simply re-requesting the same data, progressive updates to the data may be seen by a client. On the other hand, to see a broader picture of progressive result extraction, we turn to (usually) hand-tooled codes specialized for single applications. As far as we know, there is currently no language with primitives supporting progressive extractions. Additionally, what marginal runtime and protocol support for progressive extractions there is seems to only exist because of specialized data streams (like browsing weather services), and not because of intentional support. We have found examples of hand-coded systems that allow partial result extractions, at predefined process points, to allow early result inspection [3]. Such working codes have been built to allow users to steer executions and to test for convergence on iterative solution generators. However, a language and runtime system to develop arbitrary codes that allow progressive extractions is not to be found. This is especially true of compositions languages geared toward distributed objects. Herein we outline a set of language primitives expressive enough to capture each of these result extraction models simultaneously. We also review an implementation of this general result extraction model, as encompassed within the megaprogramming language CLAM and the supporting components within the CHAIMS system [7]. 2. RESULT EXTRACTION WITHIN 2.1 Definitions and motivation We focus on result extraction. In our studies, interesting "results" come from computational services rather than simple data services like databases, web servers, etc. Computational services are those that add value to input data. Results from computational services are tailored to their inputs, unlike data services where results are often already available before and after the service is used (like static pages from a web server). Various extraction methods have potentially more value in the context of computational services than in the context of data services. We define "partial extraction" as the extraction of a subset of available results. We define "Progressive extraction" as the extractions of the same result parameter with different content/data at different points in a computation. The different points of computation can deliver different accuracy of specific results, as is typical for simulations or complex calculations for images. Or they can signify dependencies of the results on the actual time, as is the case for weather data that changes over time. We went through several versions of CLAM that led to our current infrastructure. Originally limited to partial data extractions, repeated iterations in the design process yielded the general-purpose client-centric result examination and extraction model that seems to be maximally flexible. We say this model is "client-centric" because, unlike an exception or callback-type model, clients initiate all data inspection and collection. In this "data on demand" approach, clients expect that servers do not "push" either data or status, but rather must make requests for both. We specifically contrast this client-centric approach to the CORBA event model seen in 3.4. Finally, we present CLAM as one possible language and implementation of our generic result extraction model. CLAM is a composition language, rather than computationally oriented language [8, 14]. As such, it is aptly suited to the presentation of this extraction model, though not necessarily a programmer's language of choice for a given problem. We advocate this extraction model here, and present it within a particular working framework (CHAIMS). 2.2 Language specification In CLAM, there are two basic primitives essential to our result extraction model: EXAMINE and EXTRACT. We use EXAMINE to inspect the progress of a calculation (or method invocation, procedure, etc.) by requesting data status or, when provided by the server, also information about the invocation progress concerning computation time. . Purpose The EXAMINE primitive is used to determine the state and progress of the invocation referred to by an invocation_handle. EXAMINE has two status fields: state and progress. The state can be any of {DONE, NOT_DONE, PARTIAL, ERROR}. The progress field is an integer, used to indicate progress of results as well as of the invocation. The various pieces of status and progress information are only returned when the client requests them, in line with the client-centric approach. . Syntax (mystatus=status, (mystatus=status, Imagine a megamodule (a module providing a service) that has a method "foo" which returns three distinct <results/data B, and C. If foo is just invoked, and no work has been done, A, B, and C will all be incomplete. A call to will thus return NOT_DONE. When complete, a call to foo_handle.EXAMINE() will return DONE, because the work has been performed. If there are some meaningful results available for extraction before all results are ready, foo_handle.EXAMINE() returns PARTIAL. In case of error with the invocation of foo, ERROR is returned. If the megamodule supports invocation progress information, invocation_handle.EXAMINE() is used instead of (mystatus in order to get progress information about the invocation. This progress information indicates how much progress the computation has made already in terms of time used and estimated time needed to complete. Ideally this progress information is in agreement with pre invocation cost estimation which is also provided by CLAM (see primitive ESTIMATE in [8]). Yet as conditions like server load can change very rapidly, it is essential to be able to track the progress of the computation from its invocation to its completion. Upon receipt of invocation status "PARTIAL," a client knows that some subset of the results are extractable, though not that any particular element of the result data is ready for extraction (nor that the result contains progressive and thus temporary values) or has been finalized. Yet PARTIAL indicates to the client that it would be worth while to inspect individual result parameters to get their particular status information. Once again, imagine foo with return data elements A, B, and C. In this case, A is done, B is partially done with extractable results available, and no progress has been made on C. A call to would return PARTIAL, because some subset of the available data is ready. Subsequently, a client issue of foo_handle.EXAMINE(A) will return DONE, foo_handle.EXAMINE(B) will return {PARTIAL, 50}, and foo_handle.EXAMINE(C) will return NOT_DONE. Interpretations of the results from the examination A and C are obvious. In the case of result B, assuming that the result value is 50% completed, and the server makes this additional information available to the client, a return tuple such as {PARTIAL, 50} would express this. Remember, the second parameter returned by EXAMINE is an integer. CLAM places no restriction on its general use. Servers impart their own meaning on such parameters. However, the recommended usage is for the value to indicate a "percentage (0- 100) value of completion." Such semantic meaning associated with a particular server is available to client builders via a special repository. CLAM couples the EXAMINE primitive with an equally powerful EXTRACT mechanism. . Purpose The EXTRACT call collects the results of an invocation. A subset of all parameters returned by the invocation can be extracted. In fact, parameters which are extracted are the ones explicitly specified at the left hand side of the command. . Syntax Returning to our previous example of the method foo, we look at EXTRACT. The return fields are name/value pairs, and may contain any subset of the available data. For the example method foo, we might do the following: (tmpa=A, tmpb=B) foo_handle.EXTRACT(). This call would return the current values for A and B, leaving C on the server. This extract primitive allows to extract just those results that are ready as well as needed. This in contrast to a much simpler extract where simply all results would be returned with each extract command, independent of their state. 2.3 Analysis of the extraction model Different flavors of extraction are available with different levels of functionality in EXTRACT and EXAMINE. We present here the various achievable extraction modes when only portions of the complete set of CLAM data extraction primitives are available. This analysis shows how EXAMINE and EXTRACT relate and interact to provide each of the result extraction flavors outlined in this paper. Within EXAMINE, simple inspection returns a per invocation status value from {DONE, NOT_DONE, PARTIAL, ERROR}. There are two augmentations made to the EXAMINE primitive. The first augmentation is the addition of a second parameter, the progress indicator. The second augmentation is the ability to inspect individual parameters. These two augmentations provide four distinct possible examination schemes: (1) without parameter inspection, and without progress information (2) without parameter inspection, but with progress information which is invocation specific (3) with parameter inspection, but without progress information (4) with parameter inspection, and with progress information which is parameter specific There are two types of EXTRACT in CLAM: (a) EXTRACT returns all values from an invocation (like an RPC) (b) EXTRACT retrieves specific return values (like requesting specific embedded html objects). The two extraction options, when coupled with the four possible types of EXAMINE, form eight possible models of extraction. Reference source not found. shows the table of possible combinations and outlines basic functionality achievable with each potential scheme. Even with the simplest case as shown in entry 1a, an EXAMINE that works on a per invocation basis only (cannot examine particular parameters) and an EXTRACT that only returns the entire set of results. The extraction model remains more powerful than a typical C++/Fortran-like procedure call because of the viable asynchrony achieved. The model retains its client-centric orientation. The client polls at chosen times (with EXAMINE) and can extract the data whenever desired. The assumption in CLAM is that the client may also extract the same data multiple times. This places some burdens on the support system that we discuss in the next section. Table 1. EXAMINE/EXTRACT relationships single EXTRACT (a) per return element EXTRACT (b) per invocation EXAMINE, One parameter (1) 1a. Like an asynchronous procedure call e.g., Java RMI 1b. Limited partial extraction. Like 1a, with the added ability to extract a subset of return data at a time indicated by PARTIAL (limited to one checkpoint) or after all results are completed. per invocation EXAMINE, Two parameters (2) 2a. Very limited. Progressive extraction becomes possible, with data completion level indirectly indicated by second parameter. Must retrieve entire data set. 2b. Very limited. Progressive extraction still possible, no legitimate potential for partial extractions other than a unique set as in 1b. per result EXAMINE, One parameter (3) 3a. Allows for data retrieval as particular return values are complete, but entire set must be retrieved each time (semantic partial extraction). 3b. True partial extraction becomes possible here. No real progressive extraction. e.g., Web-browsing per result EXAMINE, Two parameters (4) 4a. Progressive extraction becomes possible, with data completion level indicated by second parameters. Must retrieve entire data set. Can determine more detail than 2a. 4b. Partial and progressive extraction are both possible. Single results may be extracted at various stages of completion. e.g., CLAM The mechanism of table entry 1b is mainly used for partial result extraction when all results are completed, yet not all results are needed. The client has the possibility to only extract those results needed right away, and can leave the other results on the server until they are extracted at a later point or time, or the client indicates it is no longer interested in the results. The main advantage of this level of partial extraction is the avoidance of transmitting huge amounts of data when it is not needed. This is particularly the case when one result parameter contains some meta-information that tells the client if it needs the other result parameters at all. This mechanism can be compared to the partial download of web-pages by a web-browser. In a first download the browser might exclude images. At a later point the person using the web-browser might request the download of some specific or all images based on the textual information received in the first download. This usage of partial extraction has become very important as it allows to partially download small amount of information, and only spend the costs (mainly time when using slow connections) for the costly images when really needed. There is a limited capability for process progress inspection even with only one return parameter in table entry 1b. With the set {DONE, NOT_DONE, PARTIAL, ERROR}, there is room for limited process inspection. The return of "PARTIAL" from a server may indicate a unique meaningful checkpoint to a client. It could be used to indicate some arbitrary level of completion or that some arbitrary subset of data was currently meaningful. This single return value is really a special binary case of the second return value from EXAMINE. Together with a per return element EXTRACT, this model can only reasonably be used to extract one specific, pre-defined subset of results before final extraction of the entire return data set. Figure 1 one shows this use of PARTIAL for creating a single binary partition of results this way: PARTIAL indicates in this specific case that the results A and B are ready, yet C is not yet ready. If e.g., A and C were ready yet not yet B, this could not be indicated and the status NOT_DONE would be returned. For clarity, we examine more closely the power associated with each entry in the above table. Entry 1a, when there is only one status parameter per invocation (i.e., for all of foo) and only the ability to extract the entire return data set, is clearly the most limited. It is very much like an asynchronous remote procedure call, with the clear distinction that the client polls for result readiness. Also data are only delivered when the client requests it, as noted previously. In table entry 1b, we see that adding per element extraction capability allows clients to reasonably extract one portion of the data set if it is done earlier than the whole. This capability is still very limited. This model can only reasonably be used to extract one specific, pre-defined subset of results before final extraction of the entire return data set. The return value "PARTIAL" can be used to indicate that a partial set of results are done, whereas the examination return value "DONE" may indicate that all results are ready. This use of PARTIAL can be seen in Figure 1. Again, it is only possible to create a single binary partition of results this way. ready ready not ready RETURNVALUESFROMFOO Figure 1. Overloading PARTIAL with additional semantic information In table entry 2a, we see that very limited progressive extraction of the data is made possible by the addition of the second parameter to EXAMINE. The status of the results can be indirectly derived from the progress of the invocation. This indirect progress indication only applies to the entire result return set, which is really only useful if there is only one result parameter or all the result parameters belong tightly together. This is a superset of the web-browser extraction method, actually. With the web-browser extract method for weather data to be computed, images, or simulations , no meta-information about the data is returned to the client (i.e., status about the data being 20% complete, etc. To add such meta-data to web browsing, the browser must be further extended to actively process return values through another mechanism like Java or JavaScript. In table entry 2b, we see that very limited progressive extraction of the data is again made possible by the addition of the second parameter to EXAMINE. There is really no more power in this examination and extraction model than table entry 2a. However, creative server programmers could take advantage of a scheme similar to that shown in Figure 1. Still, even under such circumstance, programmers are again limited to one predefined subset for extraction and are not allowed the flexibility seen in entries 3b and 4b. In table entry 3a, we see the addition of per return value examination information. If there is only one return value from a method, the functions of entries 3a, 3b, 4a, and 4b are identical to entries 1a, 1b, 2a, and 2b, respectively. We refer to entry 3a as semantic partial extraction because the per result examination allows the user to know exactly when individual return results are ready, but the entire data set must be extracted to get any particular element. This can cause unnecessary delay and overhead, especially with large return data sets. In table entry 3b, we see the first fully functional partial extraction. Whereas in entries 1b and 2b individual return values could be selected from the return set, partial extraction was not meaningful without a per return value EXAMINE primitive as offered here. In table entry 4a, progressive extraction is possible, with progress indicators for each data return value. On the other hand, it still suffers from the same overhead as table entry 3a: all data must be extracted at each stage. This expense cannot be avoided. In table entry 4b, we see the manifestation of the complete CLAM examine/extraction model, one that has all facets of the general result extraction model. Both partial and progressive extractions are possible, including partial-progressive extractions (where arbitrary individual elements may be extracted at various stages of completion). Without exception, this model is strictly more powerful than any of the others. 2.4 Granularity in partial extraction In our discussion of Table 1 we had the special case where it is only possible to extract exactly one subset of parameters (case 1b) and the more general case wherein the partial extraction of an arbitrary set of parameters can be extracted (case 3b). In both cases, the granularity of extraction is given by the individual method parameter. One method parameter is either extracted as a whole, or not at all. If a module supplier wants to provide a very fine granularity for partial extraction, the results of a method have to be split up into as many different parameters as possible. This allows fine granularity for partial examination and extraction, yet has two distinct disadvantages: - if the result parameters are correlated and together form a larger logical structure, this higher level structure gets lost in the specification of the method as well as in the extraction - the client is burdened with reconstructing any higher level data structure. This is an overhead concerning programming as well as concerning information dissemination (the client needs to get the information from somewhere how to reconstruct a higher level structure, information that is not necessarily part of the method specification). add as a figure part of the repository definition in XMLwith the definition for the parameter PersDat - leave out the details about short cuts for simple parameters - from schema spec. (DOROTHEA'STREE) Figure 2. Sample CLAM repository information There is another way to provide a fine granularity for partial extraction of result parameters without burdening the client with the reconstruction of higher level data structures. The module provider makes the substructure of the result parameter public, and allows users to examine and request only part of a result parameter. One possible way to do that is by using XML for parameters [15]. The structure of parameters are defined by DTDs (or XML-Schemas) and made public along with the method interfaces [16]. For CLAM the DTD for the XML- parameters is defined in the CHAIMS repository as shown in Figure 2. If the client wants to extract parameters as a whole, the client uses the CLAM syntax as discussed in section 2.2. If the client wants to examine and extract just part of a parameter, the client adds an XQL query string to each parameter name in the examine or extract command. For the parameter "PersDat" specified in Figure 2 a client could just examine and extract the element "Lastname" with the following CLAM commands: XQL is a very simple query language that allows clietns to search for and extract specific sub-elements of an XML- parameter. In the above example, the whole data structure "PersDat" is searched for an element with the tag "Lastname," which is then returned inclusive all of its sub-elements. XQL would also allow clients to specify the whole path (e.g. "/Persdat/Name/Lastname"), or to search for an element anywhere within another element (e.g., "/Persdat//Lastname") or anywhere within the entire parameter (e.g., "//Lastname"). In our specific example, all of these queries return the same data. XQL also allows more complex queries including conditions and subscripts (for more details see [17]). Using XQL queries for extracting partial results of computational methods should not be confused with using XQL queries to extract data from an XML database, in spite of the apparent similarities. There are several differences: - here we query non-persistent data; the lifetime of the result parameters is linked to the duration of the invocation there exists no overall schema for all the result parameters of one module or even of several modules. The scope of the XML specification (the DTD in the repository) is one single parameter. The relationships between the parameters is not an issue for partial extraction. - Due to the first two differences, there is also no need or use for a join operation, and a simple query language like XQL fulfills all the needs for partial extraction (whereas for XML-databases more complex query languages like XML-QL might be better suited [18 ]). 2.5 Implementation issues 2.5.1 Wrapping Within CHAIMS, all server modules have certain compatibility requirements. Many server modules are actually wrapped legacy code that do not have the necessary components to act as remote servers. For minimal CHAIMS compliance, any legacy module can trivially support an EXAMINE/EXTRACT relationship like that in table entry 1a. This is a single EXTRACT with a per invocation EXAMINE. Simply treat the legacy module like a black box that returns only {DONE, NOT_DONE, ERROR} (without PARTIAL). Also, because return values are collected in the CHAIMS wrapper, the client can freely choose when to request the data, though it must request the data explicitly. The client may also perform multiple requests for the same data without further augmentation of the original code (table entry 1b). connection wrapper legacy code INVOKE "foo" INVOKE "foo" Figure 3. Client-wrapper communication To use the more powerful models of data extraction, significant modification would be required of na-ve modules. We originally classified the two augmentation types (to legacy modules) as either partial or progressive-type augmentations. Partial extraction augmentations are those that make a particular subset of the return data externally available before the completion of the entire invocation. Progressive extraction augmentations are those that post information in multiple stages, i.e., at implementer-defined checkpoints. Native modules designed for partial and progressive extraction must have a way to post interim results that may be extracted by clients. The interim results must be held in some structure so that request for data may be serviced without interrupting the working code. The CHAIMS wrapper is a threaded component that handles messages and provides a means of storing interim results and delivering those results to clients. To implement partial and progressive extractions, two pieces of information are required: status and data. When a module posts an interim result (to be delivered by the wrapper or a native server), both pieces of information must be given about the result value. This status does not need to be provided per method or pre- invocation, however. Such information is extracted from collected knowledge about all partial results. When no partial results are ready, status is NOT_DONE. When all partial results are ready, status is DONE. When some results are ready at any level, per invocation status is simply PARTIAL. When any partial result indicates ERROR, however, the per invocation status should be set to ERROR. This prevents a blind per invocation extraction of all data elements when some may be corrupt. 2.5.2 Result marshalling methods There are two equally appropriate methods for marshalling partial results, depending upon application: passive and active. The marshalling concerns are basically the servers', not the clients'. With a passive approach, whenever a client specifically requests status or partial results, the message handler requests that information from the working code. The appropriate routines for doing so are provided in the runtime layer (CPAM - Composition Protocol for Autonomous Megamodules) in the form of a Java class, or they may be user developed. Of course, native codes written with the intent of posting partial results are easier to work with than wrapped codes and can use any suitable language (Java, or otherwise). Figure 4 better shows the marshalling and examination interaction between a wrapper and legacy code. The status and progress information is held in the wrapper. Also, temporary storage locations for the extractable parameters are located in the wrapper. The active approach to data marshalling is more appropriate for certain problem types. In this method, when a server program reaches a point where it is appropriate to post status and results, it does so, directly to the CPAM objects or wrapper layer. The trade-offs between the approaches should be clear. Active posting is conceptually simpler and easier to code, but requires the overhead posting and storing interim results that may never be examined. Figure 4 shows how an active approach to data marshalling would proceed through time. After the wrapper receives an EXAMINE request, the appropriate routines actively inspect the legacy code to update the status/progress structure in the wrapper. After the EXTRACT is received, the request is passed to the legacy code, and the data structures are then updated, before results are passed back to the client. WRAPPER LEGACY CODE result A result C status progress Figure 4. Wrapper result marshalling 2.5.3 Termination The ability to delay extractions and make repeated extractions implies that a server no longer knows exactly when a client is finished with an invocation. With a traditional RPC, this was not a problem. In that case, when the work was complete, results were returned, and the server (or procedure) had no further obligations to the client. With arbitrary extraction, the server is obligated to hold data for the client. Even without allowing repeated extractions, there are more subtle reasons for which the server must hold data for clients. In the case of a partial extraction from our example method foo, a client may extract result fields A and B, but the server does not know that the client is not also interested in result field C. Since there is no a priori communication of intent from client to server, their relationship must be changed somewhat. The obligation for a server to cache and store results is balanced by a client's obligation to explicitly terminate an invocation. This explicit termination merely signals to a server that a client is no longer interested in further extractions from a particular invocation, but is an integral detail of this model of result extraction. 2.5.4 Repository There should be a repository of method interfaces and the structure of return values available to programmers using this arbitrary extraction model. Of course, when programming "in the small," (i.e., stand-alone programs, in-house projects, etc.), this is not really an issue at all. When making services available for sale/use externally, service providers must provide the appropriate information about the results which can be extracted. For instance, if delivering foo over the net, a provider should indicate to users that fields A, B, and C may be extracted separately. This information in the context of CHAIMS (where we assume "programming in the large") is provided via a repository. 3. COMPARISONS In the following we compare the extraction model as defined in CLAM and mirrored primitive by primitive in the CHAIMS access protocol CPAM to the extraction models found in the following protocols: . web browsing . JointFlow . SWAP . CORBA-DII 3.1 Partial and progressive result extraction in web browsing Web browsing generally falls into the category of services that we refer to as data services. Recall, data services primarily deliver specific data requested by clients, rather than computational services which add value to client supplied data. Clients are usually represented by one of many available web browsers or crawlers while web servers deliver data to those clients. Clients request data using the http protocol. Data extracted (documents delivered) from servers are often written in html, and often have components of varying types, including images, audio, and video. Web browsing occurs in batch and interactive ways. Batch browsing is performed by crawlers for many reasons, such as indexing, archiving, etc. Interactive browsing is performed by humans for numerous reasons, such as information gathering, electronic commerce, etc. A browser of either sort makes a request to a server for a specific document. That document is returned to the client, and serves as a template for further requests. If the document is html, the browser may parse the document and determine that there are other elements that form the complete document (i.e., images tags). The document serves as a schema, describing the other elements that may be extracted from the server. After a main document has been fetched, we can consider the possible partial and progressive extractions that can take place. To extract a sub-element of a web page, an http request is sent to a server, and the data or an error message is returned. In batch browsing, the textual information contained in the page is frequently enough to be meaningful. This is very different from the generalized result extraction model we discuss in that the schema of the results is not meaningful in itself. But, in web- browsing, the page retrieved is often meaningful, not just for the sub-elements it describes. This aside, we consider result extraction in terms of gathering sub-elements from pages In interactive browsing, partial extraction is a simple process, and is at least marginally exploitable in the most widely used interactive browsers (Netscape and Microsoft's Internet Explorer). Both feature an "auto-load" feature that can be toggled (to varying degrees) to automatically load (or not load) different content type such as images, audio, or video. For instance, some users are concerned with images and text, but do not with to be disturbed by audio. Their browser makes the http requests for all sub-elements, save audio. This is partial extraction. In other cases, especially with slower internet connections, images are expensive to download, users may choose to not automatically download images until determining that a particular image or set of images is important enough to invest time in. Partial extraction in web-browsing is a special case of the general partial extraction model in that the first result to be extracted always contains information about the other results to be extracted. Based on this first result, the client not only determines its interest in the other elements of the page, but also gets the information about what other results are available at all. This is in contrast to our general model, where a result parameter may but need not provide information about other result parameters, and where all possible result parameters are specified in a repository beforehand. The most commonly found progressive extraction in web browsing is quite different from progressive extraction in a computational service though progressive extraction of computational services over the web, e.g. improving simulation data, is also feasible. In a computational service, progressive extraction refers to extracting various transformations of input data over the life of a computation. In web browsing, progressive extraction is actually repeated extraction of a changing data stream. Weather services on the web often provide continuous updates and satellite images. Stock tickers provide updated information so users can have current information about their investments. Repeated extractions from the same stream show the stream's progress through time. Sometimes these repeated extractions may be done by manually reloading the source, or they may be pulled from servers by html update commands, JavaScript's, embedded Java-code, etc. Such data is retrievable any time and its progress status is always DONE and 100% accurate, yet we expect the data to contain also information about to which point of time it refers. 3.2 Incremental result extraction and progress monitoring in JointFlow JointFlow is the Joint Workflow Management Facility of CORBA [6]. It is an implementation of the I4 protocol of the workflow reference model of WfMC [11] on top of CORBA. JointFlow adopts an object oriented view of workflow management: processes, activities, requesters, resources, process managers, event audits etc. are distributed objects, collaborating to get the overall job done. Each of these objects can be accessed over an ORB, the JointFlow specification defines their interfaces in IDL. We have chosen JointFlow for comparison as it is a protocol that also support the request of remote computational units which can yet need not to have some degree of autonomy, and the protocol is also based on asynchronous invocation of work and extraction of results, having special primitives for invocation, monitoring, and extraction. Work is started in that a requester asks a process manager to create a new process. The requester then communicates directly with the new process, setting context attributes in the process and invoking the start operation of the process. A process my be a physical device, a wrapper of legacy code, or it may initiate several activity objects which might in turn use resources (e.g. humans) via assignments or itself act as requesters for other processes. Our focus of interest here is the interaction between the requester and the process concerning result extraction and progress monitoring. 3.2.1 Monitoring the Progress of Work Both, processes and activities are in one of the following states: running, not_running.not_started, not_running.suspended, completed (successfully), terminated (unsuccessfully), aborted (unsuccessfully). A requester can query the state of a process, the states of the activities of the process (by querying and navigating the links from processes to activities), and the states of assignments (by querying and navigating the links from activities to assignments). If the requester knows the workflow model with all its different steps implemented by the process, the requester might be able to interpret the state information of all subactivities and assignments and figure out what the progress of the process is. If the model is not known, e.g., due to an autonomy boundary as they are assumed in CHAIMS, the only status information provided by the JointFlow protocol itself is essentially completed or not yet completed. In contrast, CHAIMS supports the notion that certain services may support progress information (e.g. 40% done) that can be monitored. This information is more detailed than just running or complete, and more aggregated and better suited for autonomous services than detailed information about component activities. In contrast to CHAIMS that polls all progress information, in JointFlow a process signals its completion to the requester by an audit event. These audit events could also be used to implement CHAIMS-like progress monitoring on top of JointFlow: a process can have a special result attribute for progress information and the process is free to update that attribute regularly. It then can send an audit event with the old and new value of the progress indicator result to its requester after each update. Yet this result attribute cannot be polled by a requester (in contrast to CPAM and SWAP), because get_result only returns results if all results are available at least as intermediate results. 3.2.2 Extracting Results Incrementally Both, processes and activities have an operation get_result():ProcessData (returning a list of name value pairs). Get_result does not take any input parameter and thus returns all the results. The get_result operation may be used to request intermediate result data, which may or may not be provided depending upon the work being performed. If the results cannot yet be obtained, the operation get_result raises an exception and returns garbage. The results are not final until the whole unit of work is completed, resulting in a state change to the state complete and a notification of the container process or the requester. This kind of extracting intermediate results corresponds to the progressive extraction of all result attributes in CHAIMS. The following features found in CHAIMS are not available in JointFlow: . Partial extraction with get_result: only all or none of the result values can be extracted by get_result, and there is no mechanism to return an exception only for some of the values. . Progressive extraction with get_result of just one result attribute when not yet all other results are ready for intermediate or final extraction . There is no accuracy information for intermediate results, unless it is in a separate result attribute. There is no possibility to find out about the availability or the accuracy of intermediate results unless requesting these results. Though partial and progressive result extraction are not part of the design of JointFlow, they also can be achieved by using audit events and by pushing progressive and partial results onto the requester, instead of letting the requester poll for them. A process or an activity can send out an audit event to its requester or to the containing process whenever one of the result values has been updated. This event would then contain the old as well as the new result value. In case of large data and frequent updates, this messaging mechanism could result in huge amounts of traffic. The mechanism would have to be extended by special context attributes that tell the process or activity in advance which results should be reported in which intervals. Yet this results in a very static and server centric approach, in contrast to the client-centric approach in CHAIMS that is based on data on demand. Also, as partial and progressive result extraction are not mandated by the JointFlow protocol itself, it is questionable how many processes and activities would actually offer it. 3.3 Incremental result extraction and progress monitoring in SWAP SWAP (Simple Workflow Access Protocol) is a proposal for a workflow protocol based on extending http. It mainly implements I4 (to some extend also I2 and I3) of the WfMC reference model. SWAP defines several interfaces for the different components (which are internet resources) of the workflow system that interact via SWAP. The three main components are of type ProcessInstance, ProcessDefinition and Observer. The messages exchanged between these components are extended http-messages with headers defined by SWAP. The data to be exchanged is encoded as text/xml in the body of the message. A process instance (having the interface ProcessInstance) is created and started by sending a CREATEPROCESSINSTANCE message to the appropriate ProcessDefinition resource. This message also contains the context data to be set and the URI of an observer resource that should be notified about completion and other events. The response contains the URI of the newly created process instance. The process is started either automatically by the ProcessDefinition resource if the CREATEINSTANCEMESSAGE contains the startImmediately flag, or by sending a PROPPATCH message to the process instance with the new state running. A process instance resource can delegate work to other resources by acting itself as an observer and ask some ProcessDefinition resources for the creation of other process instances. As in JointFlow and in CHAIMS, the process instance creation, setting of context attributes, start of the process, and the extraction of results are done asynchronously. 3.3.1 Result Extraction and Result Monitoring Results are extracted from a process instance by sending it the message PROPFIND at any time during the execution of a process instance. This message either returns all available results, or if it contained a list of requested result attributes, it only returns the selected ones. Only result attributes are returned that are available. If requested attributes are not yet available, presumably an exception should be returned for these result attributes. SWAP does not specify if the results returned by PROPFIND have to be final or not. Given the possibility to ask for specific result attributes, and to get exceptions for specific result attributes in case they are not available (made possible by having exceptions encoded in XML instead of having just one possible exceptions for one procedure call as in the CORBA based JointFlow protocol), allows some degree of partial and maybe even progressive extraction. A process instance signals the completion of work to an observer with the COMPLETE message. This message also contains the result data: all the name value pairs that represent the final set of data as of the time of completion. After sending the COMPLETE message, the resource does not have to exist any longer, this in contrast to CHAIMS where the no result data is lost until the client (observer) sends a TERMINATE. A process instance can also send NOTIFY messages to an observer resource. These messages transmit state change events, data change events, and role change events, data change events containing the names and values of data items that have changed. 3.3.2 Incremental Result Extraction as Defined in the CHAIMS Model The mechanisms of SWAP allow the following kind of result extraction and progress monitoring: . Partial result extraction: Either pushing results via NOTIFY messages or pulling results via PROPFIND messages is possible. NOTIFY sends all new result data, PROPFIND returns all available result data whether or not they have already been returned by a previous PROPFIND. Notification of result changes without sending also the new values is not possible unless additional result attributes are added. The same is true for getting the status of individual results: asking for the status of results without getting also the results, is not possible unless a state attribute is added for each data attribute to the set of result attributes. . Progressive result extraction: The SWAP specification does not explicitly specify if progressive result updates in a process instance are allowed or not. If not, the result attributes would not be available until their values are final. If yes, then progressive results can be extracted either by pushing results via NOTIFY messages or by pulling results via PROPFIND messages. Accuracy indication is not provided, it would have to be implemented via additional result attributes. 3.3.3 Process Progress Monitoring PROPFIND not only returns all result values available, it also returns the state of the process instance and additional descriptive information about the process. As possible states can be specified by the process itself, PROPFIND also returns the list of all possible state values, yet in most cases it would probably just be not_yet_running, running, suspended, completed, terminated, etc (the basic set of states defined by I4). A process instance can be asked for the URI of all the processes it has delegated work to, and an observer then can directly ask this subprocesses about their statuses. This is analogue to the model found in JointFlow, and thus has the same drawbacks concerning autonomy and concerning amalgated progress information. progress information is not specified by SWAP, but it could be implemented by a special result attribute assuming that result attributes can be changed over time. Such result attributes could be extracted any time by PROPFIND, independent of the availability of other result attributes. Though SWAP does not support incremental result extraction as defined in CHAIMS, it could quite easily either be added to the SWAP protocol itself or done by using the SWAP protocol as defined and applying the simple workarounds mentioned above. As SWAP has very similar goals in accessing remote processes as CHAIMS, and as it is a very open and flexible protocol, its result extraction model is already very close to the one of CHAIMS and could be easily extended to contain all aspects of the CHAIMS extraction model. Yet as SWAP has not been designed with incremental extraction in mind, it does not have the strong duality between extract and monitoring command as found in CHAIMS between EXAMINE and EXTRACT. 3.4 Incremental result extraction and progress monitoring in CORBA 3.4.1 CORBA- DII CORBA offers two modes for interaction between a client and remote servers: the static and the dynamic interface to an ORB. For the static interface an IDL must exist that is compiled into stub code that can be linked with the client. The client then executes remote procedure calls as if the remote methods were local. The dynamic invocation interface (DII) offers dynamic access where no stub code is necessary. The client has to know (or can ask for) the IDL from the remote object, i.e., the names of the methods and the parameters they take. The client then creates a request for a method of that object. In this request the method name appears as a string and the parameters appear as a list of named values, with each named value containing the name of the parameter, the value as type any (or a pointer to the value and a CORBA type code), the length of the parameter, and some flags. Once the request is created, the method can be invoked. This is either done synchronously with invoke or asynchronously with send (in fact, some flags allow more elaborate settings). Invoke returns after the remote computation has completed, and the client can read all OUT parameters in the named value list. In case of a send, the client is not blocked. In order to figure out when the invocation has finished, the client can use get_response, either in a blocking (it waits until invocation is done) or a non-blocking mode. As soon as the return status of get_response indicates that the remote computation is done, the client can read OUT parameters from the named value list. In case of the asynchronous method invocation in CORBA-DII, the progress of an invocation can be monitored and asked for by the client as far as completion is concerned, but no further progress information is available. Progressive extraction of results is not supported by DII. Of course a client is free not to read and use all results after the completion of an invocation, yet while the computation is going on no partial extraction is supported. 3.4.2 CORBA Notification Service In order to mimic the incremental result extraction of CHAIMS, one could use asynchronous method invocation with DII coupled with the event service of CORBA. The client could be implemented as a PullConsumer for a special event channel CHAIMSresults, the servers could push results into that channel as soon as they are available, together with accuracy information. Though event channels could be used for that purpose (we could require that every megamodule uses event channels for this), an integration of incremental result extraction and invocation progress monitoring into the access protocol itself is definitely more adequate when we consider this to be an integral part of the protocol. The same is true for the languages used to program the client: while CLAM directly supports incremental extraction and progress monitoring, this is not the case for any of the languages in used for programming CORBA clients. 4. CONCLUSIONS In the CHAIMS project, we sought to build a composition-only language for remote, autonomous services. To do this, we had to consider many different extraction models used in different domains. This examination led to the realization that a simple asynchronous RPC-style approach was not enough. To build a language and an access protocol to support arbitrary result extractions took careful consideration of the myriad ways extractions were currently being used in widespread as well as hand-crafted systems. Building support and primitives for all of these result extraction methods (autonomously, progressively, and partially) and then binding them within one system has led to the formulation of a comprehensive model for arbitrary result extraction. Our model captures the notions of traditional result extraction, partial extraction and progressive extraction. By combining two simple primitives in CLAM (EXAMINE and EXTRACT), the full power of each of these extraction types can be achieved. This extraction model is appropriate as a template for existing systems, and future languages as well. It is generic to result extraction, and only assumes that the necessary asynchrony can be achieved among components through distributed communication, threading, or other available means. 5. --R "A Language and System for Composing Autonomous, Heterogeneous and Distributed Megamodules," "Opus: A Coordination Language for Multidisciplinary Applications," "Exploiting Parallelism in Multidisciplinary Applications Using Opus," Porgram Design for Engineers ICASE Research Quarterly "CPAM, A Protocol for Software Composition," "CLAM: Composition Language for Autonomous Megamodules," Simple Workflow Access Protocol (SWAP) "Towards Megaprogramming: A Paradigm for Component-Based Programming" The Workflow Reference Model "Pipeline Expansion in Coordinated Applications," Design and Implementation "Composition of Multi-site Services," "Extensible Markup Language (XML) 1.0," "Schema for Object-Oriented XML 2.0," "XQL Tutorial," "XML-QL: A Query Language for XML," --TR Toward megaprogramming Programming languages (3rd ed.) C Program Design for Engineers CPAM, A Protocol for Software Composition A Case for Economy Grid Architecture for Service Oriented Grid Computing Opus: A Coordination Language for Multidisciplinary Applications --CTR Andrea Omicini , Sascha Ossowski, Editorial message: special track on coordination models, languages and applications, Proceedings of the 2002 ACM symposium on Applied computing, March 11-14, 2002, Madrid, Spain

partial;result extraction;scheduling;progressive;CHARMS

509092

Making sparse Gaussian elimination scalable by static pivoting.

We propose several techniques as alternatives to partial pivoting to stabilize sparse Gaussian elimination. From numerical experiments we demonstrate that for a wide range of problems the new method is as stable as partial pivoting. The main advantage of the new method over partial pivoting is that it permits a priori determination of data structures and communication pattern for Gaussian elimination, which makes it more scalable on distributed memory machines. Based on this a priori knowledge, we design highly parallel algorithms for both sparse Gaussian elimination and triangular solve and we show that they are suitable for large-scale distributed memory machines.

Introduction In our earlier work [8, 9, 22], we developed new algorithms to solve unsymmetric sparse linear systems using Gaussian elimination with partial pivoting (GEPP). The new algorithms are highly efficient on workstations with deep memory hierarchies and shared memory parallel machines with a modest number of processors. The portable implementations of these algorithms appear in the software packages SuperLU (serial) and SuperLU MT (multithreaded), which are publically available on Netlib [10]. These are among the fastest available codes for this problem. Our shared memory GEPP algorithm relies on the fine-grained memory access and synchronization that shared memory provides to manage the data structures needed as fill-in is created dynamically, to discover which columns depend on which other columns symbolically, and to use a centralized task queue for scheduling and load balancing. The reason we have to perform all these dynamically is that the computational graph does not unfold until runtime. (This is in contrast to Cholesky, where any pivot order is numerically stable.) However, these techniques are too expensive This research used resources of the National Energy Research Scientific Computing Center, which is supported by the Office of Energy Research of the U.S. Department of Energy under Contract No. DE-AC03-76SF00098. y This research was supported in part by NSF grant ASC-9313958, DOE grant DE-FG03-94ER25219, UT Sub-contract No. ORA4466 from ARPA Contract No. DAAL03-91-C0047, DOE grant DE-FG03-94ER25206, NSF Infrastructure grants CDA-8722788 and CDA-9401156 and DOE grant DE-FC03-98ER25351. (1) Row/column equilibration and row permutation: A / P r \Delta D r \Delta A \Delta D c , where D r and D c are diagonal matrices and P r is a row permutation chosen to make the diagonal large compared to the off-diagonal (2) Find a column permutation P c to preserve sparsity: A / c with control of diagonal magnitude set a ii to p endif using the L and U factors, with the following iterative refinement iterate: multiply Solve A \Delta goto iterate endif Figure 1: The outline of the new GESP algorithm. on distributed memory machines. Instead, for distributed memory machines, we propose to not pivot dynamically, and so enable static data structure optimization, graph manipulation and load balancing (as with Cholesky [20, 25]) and yet remain numerically stable. We will retain numerical stability by a variety of techniques: pre-pivoting large elements to the diagonal, iterative refine- ment, using extra precision when needed, and allowing low rank modifications with corrections at the end. In Section 2 we show the promise of the proposed method from numeric experiments. We call our algorithm GESP for Gaussian elimination with static pivoting. In Section 3, we present an MPI implementation of the distributed algorithms for LU factorization and triangular solve. Both algorithms use an elaborate 2-D (nonuniform) block-cyclic data distribution. Initial results demonstrated good scalability and a factorization rate exceeding 8 Gflops on a 512 node Cray T3E. stability Traditionally, partial pivoting is used to control the element growth during Gaussian elimination, making the algorithm numerically stable in practice 1 . However partial pivoting is not the only way to control element growth; there are a variety of alternative techniques. In this section we present these alternatives, and show by experiments that appropriate combinations of them can effectively stabilize Gaussian elimination. Furthermore, these techniques are usually inexpensive compared to the overall solution cost, especially for large problems. 2.1 The GESP algorithm In Figure 1 we sketch our GESP algorithm that incorporates some of the techniques we considered. To motivate step (1), recall that a diagonally dominant matrix is one where each diagonal entry a ii is larger in magnitude than the sum of magnitudes of the off-diagonal entries in its row ( P exist where even GEPP is unstable, but these are very rare [7, 19]. or column ( P j). It is known that choosing the diagonal pivots ensures stability for such matrices [7, 19]. So we expect that if each diagonal entry can somehow be made larger relative to the off-diagonals in its row or column, then diagonal pivoting will be more stable. The purpose of step (1) is to choose diagonal matrices D r and D c and permutation P r to make each a ii larger in this sense. We have experimented with a number of alternative heuristic algorithms for step (1) [13]. All depend on the following graph representation of an n \Theta n sparse matrix A: it is represented as an undirected weighted bipartite graph with one vertex for each row, one vertex for each column, and an edge with appropriate weight connecting row vertex i to column vertex j for each nonzero entry a ij . Finding a permutation P r that puts large entries on the diagonal can thus be transformed into a weighted bipartite matching problem on this graph. The diagonal scale matrices D r and D r can be chosen independently, to make each row and each column of D r AD c have largest entries equal to 1 in magnitude (using the algorithm in LAPACK subroutine DGEEQU [3]). Then there are algorithms in [13] that choose P r to maximize different properties of the diagonal of P r D r AD c , such as the smallest magnitude of any diagonal entry, or the sum or product of magnitudes. But the best algorithm in practice seems to be the one in [13] that picks P r , D r and D c simultaneously so that each diagonal entry of P r D r AD c is \Sigma1, each off-diagonal entry is bounded by 1 in magnitude, and the product of the diagonal entries is maximized. We will report results for this algorithm only. The worst case serial complexity of this algorithm is O(n \Delta nnz(A) \Delta log n), where nnz(A) is the number of nonzeros in A. In practice it is much faster; actual timings appear later. Step (2) is not new and is needed in both SuperLU and SuperLU MT [10]. The column permutation P c can be obtained from any fill-reducing heuristic. For now, we use the minimum degree ordering algorithm [23] on the structure of A T A. In the future, we will use the approximate minimum degree column ordering algorithm by Davis et. al. [6] which is faster and requires less memory since it does not explicitly form A T A. We can also use nested dissection on A + A T or A T A [17]. Note that we also apply P c to the rows of A to ensure that the large diagonal entries obtained from Step (1) remain on the diagonal. In step (3), we simply set any tiny pivots encountered during elimination to p " is machine precision. This is equivalent to a small (half precision) perturbation to the original problem, and trades off some numerical stability for the ability to keep pivots from getting too small. In step (4), we perform a few steps of iterative refinement if the solution is not accurate enough, which also corrects for the perturbations in step (3). The termination criterion is based on the componentwise backward error berr [7]. The condition berr - " means that the computed solution is the exact solution of a slightly different sparse linear system each nonzero entry a ij has been changed by at most one unit in its last place, and the zero entries are left unchanged; thus one can say that the answer is as accurate as the data deserves. We terminate the iteration when the backward error berr is smaller than machine epsilon, or when it does not decrease by at least a factor of two compared with the previous iteration. The second test is to avoid possible stagnation. (Figure 5 shows that berr is always small.) 2.2 Numerical results In this subsection, we illustrate the numerical stability and runtime of our GESP algorithm on 53 unsymmetric matrices from a wide variety of applications. The application domains of the matrices are given in Table 1. Most of them, except for two (ECL32, WU), can be obtained from the Harwell-Boeing Collection [14] and the collection of Davis [5]. Matrix ECL32 was provided by Jagesh Sanghavi from EECS Department of UC Berkeley. Matrix WU was provided by Yushu Discipline Matrices fluid flow, CFD af23560, bbmat, bramley1, bramley2, ex11, fidapm11, garon2, graham1, lnsp3937, lns 3937, raefsky3, rma10, venkat01, wu fluid mechanics goodwin, rim circuit simulation add32, gre 1107, jpwh 991, memplus, onetone1, onetone2, twotone device simulation wang3, wang4, ecl32 chemical engineering extr1, hydr1, lhr01, radfr1, rdist1, rdist2, rdist3a, west2021 petroleum engineering orsirr 1, orsreg 1, sherman3, sherman4, sherman5 finite element PDE av4408, av11924 stiff ODE fs 541 2 Olmstead flow model olm5000 aeroelasticity tols4000 reservoir modelling pores 2 crystal growth simulation cry10000 power flow modelling gemat11 dielectric waveguide dw8192 (eigenproblem) astrophysics mcfe plasma physics utm5940 economics mahindas, orani678 Table 1: Test matrices and their disciplines. Wu from Earth Sciences Division of Lawrence Berkeley National Laboratory. Figure 2 plots the dimension, nnz(A), and nnz(L i.e. the number of nonzeros in the L and U factors (the fill-in). The matrices are sorted in increasing order of the factorization time. The matrices of most interest for parallelization are the ones that take the most time, i.e. the ones on the right of this graph. From the figure it is clear that the matrices large in dimension and number of nonzeros also require more time to factorize. The timing results reported in this subsection are obtained on an SGI ONYX2 machine running IRIX 6.4. The system has 8 195 MHz MIPS R10000 processors and 5120 Mbytes main memory. We only use a single processor, since we are mainly interested in numerical accuracy. Parallel runtimes are reported in section 3. Detailed performance results from this section in tabular format are available at http://www.nersc.gov/-xiaoye/SC98/. Among the 53 matrices, most would get wrong answers or fail completely (via division by a zero pivot) without any pivoting or other precautions. 22 matrices contain zeros on the diagonal to begin with which remain zero during elimination, and 5 more create zeros on the diagonal during elimination. Therefore, not pivoting at all would fail completely on these 27 matrices. Most of the other 26 matrices would get unacceptably large errors due to pivot growth. For our experiment, the right-hand side vector is generated so that the true solution x true is a vector of all ones. IEEE double precision is used as the working precision, with machine epsilon - 10 \Gamma16 . Figure 3 shows the number of iterations taken in the iterative refinement step. Most matrices terminate the iteration with no more than 3 steps. 5 matrices require 1 step, 31 matrices require 2 steps, 9 matrices require 3 steps, and 8 matrices require more than 3 steps. For each matrix, we present two error metrics, in Figure 4 and Figure 5, to assess the accuracy and stability of GESP. Figure 4 plots the error from GESP versus the error from GEPP (as implemented in SuperLU) for each matrix: A red dot on the green diagonal means the two errors were the same, a red dot below the diagonal means LU Factorization time in seconds dimension - nonzeros in A - nonzeros in L+U Figure 2: Characteristics of the matrices. Condition number Number of iterative refinement steps GESP (Red), GEPP(Blue) Figure 3: Iterative refinement steps in GESP. GESP is more accurate, and a red dot above means GEPP is more accurate. Figure 4 shows that the error of GESP is at most a little larger, and can be smaller (21 out of 53), than the error from GEPP. Figure 5 shows that the componentwise backward error [7] is also small, usually near machine epsilon, and never larger than 10 \Gamma12 . Although the combination of the techniques in steps (1) and (3) in Figure 1 works well for most matrices, we found a few matrices for which other combinations are better. For example, for FIDAPM11, JPWH 991 and ORSIRR 1, the errors are large unless we omit P r from step (1). For EX11 and RADRF1, we cannot replace tiny pivots by p Therefore, in the software, we provide a flexible interface so the user is able to turn on or off any of these options. We now evaluate the cost of each step in GESP Figure 1. This is done with respect to the serial implementation, since we have only parallelized the numerical phases of the algorithm (steps (3) and (4)), which are the most time-consuming. In particular, for large enough matrices, the LU factorization in step (3) dominates all the other steps, so we will measure the times of each step with respect to step (3). Simple equilibration in step (1) (computing D r and D c using the algorithm in DGEEQU from LAPACK) is usually negligible and is easy to parallelize. Both row and column permutation algorithms in steps (1) and (2) (computing P r and P c ) are not easy to parallelize (their parallelization is future work). Fortunately, their memory requirement is just O(nnz(A)) [6, 13], whereas the memory requirement for L and U factors grows superlinearly in nnz(A), so in the meantime we can run them on a single processor. Figure 6 shows the fraction of time spent finding P r in step (1) using the algorithm in [13], as a fraction of the factorization time. The time is significant for small problems, but drops to 1% to 10% for large matrices requiring a long time to factor, the problems of most interest on parallel machines. The time to find a sparsity-preserving ordering P c in step (2) is very much matrix dependent. It is usually cheaper than factorization, although there exist matrices for which the ordering is more expensive. Nevertheless, in applications where we repeatedly solve a system of equations with the same nonzero pattern but different values, the ordering algorithm needs to be run only once, and its cost can be amortized over all the factorizations. We plan to replace this part of the algorithm Error from partial pivoting with Refine from GESP Figure 4: The error jjx true \Gammaxjj 1 Condition number Backward error from GESP Figure 5: The backward error LU factorization (GENP) time in seconds Fraction of GENP time - Permute large diagonal - Triangular solve Figure The times to factorize, solve, permute large diagonal, compute residual and estimate error bound, on a 195 MHz MIPS R10000. Order BBMAT 38744 1771722 .0224 .5398 49.1 4.3 Table 2: Characteristics of the test matrices. NumSym is the fraction of nonzeros matched by equal values in symmetric locations. StrSym is the fraction of nonzeros matched by nonzeros in symmetric locations. with something faster, as outlined in Section 2.1. As can be seen in Figure 6, computing the residual (sparse matrix-vector multiplication cheaper than a triangular solve both take a small fraction of the factorization time. For large matrices the solve time is often less than 5% of the factorization time. Both algorithms have been parallelized (see section 3 for parallel performance data). Finally, our code has the ability to estimate a forward error bound for the true error jjx true \Gammaxjj 1 This is by far the most expensive step after factorization. (For small matrices, it can be more expensive than factorization, since it requires multiple triangular solves.) Therefore, we will do this only when the user asks for it. 3 An implementation with MPI In this section, we describe our design, implementation and the performance of the distributed algorithms for two main steps of the GESP method, sparse LU factorization (step (3)) and sparse triangular solve (used in step (4)). Our implementation uses MPI [26] to communicate data, and so is highly portable. We have tested the code on a number of platforms, such as Cray T3E, IBM SP2, and Berkeley NOW. Here, we only report the results from a 512 node Cray T3E-900 at NERSC. To illustrate scalability of the algorithms, we restrict our attention to eight relatively large matrices selected from our testbed in Table 1. They are representative of different application domains. The characteristics of these matrices are given in Table 2. 3.1 Matrix distribution and distributed data structure We distribute the matrix in a two-dimensional block-cyclic fashion. In this distribution, the P processes (not restricted to be a power of 2) are arranged as a 2-D process grid of shape P r \Theta P c . The matrix is decomposed into blocks of submatrices. Then, these blocks are cyclically mapped onto the process grid, in both row and column dimensions. Although a 1-D decomposition is more natural to sparse matrices and is much easier to implement, a 2-D layout strikes a good balance among locality (by blocking), load balance (by cyclic mapping), and lower communication volume (by 2-D mapping). 2-D layouts were used in scalable implementations of sparse Cholesky factorization [20, 25]. We now describe how we partition a global matrix into blocks. Our partitioning is based on the notion of unsymmetric supernode first introduced in [8]. Let L be the lower triangular matrix in the LU factorization. A supernode is a range of columns of L with the triangular block just below the diagonal being full, and with the same row structure below this block. Because of the identical row structure of a supernode, it can be stored in a dense format in memory. This supernode partition is used as our block partition in both row and column dimensions. If there are N supernodes in an n-by-n matrix, the matrix will be partitioned into N 2 blocks of nonuniform size. The size of each block is matrix dependent. It should be clear that all the diagonal blocks are square and full (we store zeros from U in the upper triangle of the diagonal block), whereas the off-diagonal blocks may be rectangular and may not be full. The matrix in Figure 7 illustrates such a partitioning. By block-cyclic mapping we mean block mapped onto the process at coordinate (I mod P r , J mod P c ) of the process grid. Using this mapping, a block L(I ; J) in the factorization is only needed by the row of processes that own blocks in row I . Similarly, a block U(I ; J) is only needed by the column of processes that own blocks in column J . In this 2-D mapping, each block column of L resides on more than one process, namely, a column of processes. For example in Figure 7, the k-th block column of L resides on the column processes f0, 3g. Process 3 only owns two nonzero blocks, which are not contiguous in the global matrix. The schema on the right of Figure 7 depicts the data structure to store the nonzero blocks on a process. Besides the numerical values stored in a Fortran-style array nzval[] in column major order, we need the information to interpret the location and row subscript of each nonzero. This is stored in an integer array index[], which includes the information for the whole block column and for each individual block in it. Note that many off-diagonal blocks are zero and hence not stored. Neither do we store the zeros in a nonzero block. Both lower and upper triangles of the diagonal block are stored in the L data structure. A process owns dN=P c e block columns of L, so it needs dN=P c e pairs of index/nzval arrays. For matrix U , we use a row oriented storage for the block rows owned by a process, although for the numerical values within each block we still use column major order. Similarly to L, we also use a pair of index/nzval arrays to store a block row of U . Due to asymmetry, each nonzero block in U has the skyline structure as shown in Figure 7 (see [8] for details on the skyline structure). Therefore, the organization of the index[] array is different from that for L, which we omit showing in the figure. Since we do no dynamic pivoting, the nonzero patterns of L and U can be determined during symbolic factorization before numerical factorization begins. Therefore, the block partitioning and the setup of the data structure can all be performed in the symbolic algorithm. This is much cheaper to execute as opposed to partial pivoting where the size of the data structure cannot be forecast and must be determined on the fly as factorization proceeds. 3.2 Sparse LU factorization Figure 8 outlines the parallel sparse LU factorization algorithm. We use Matlab notation for integer ranges and submatrices. There are three steps in the K-th iteration of the loop. In step (1), only a column of processes participate in factoring the block column L(K : N; K). In step (2), only a row of processes participate in the triangular solves to obtain the block row U(K;K The rank-b update by L(K in step (3) represents most of the work and also exhibits more parallelism than the other two steps, where b is the block size of the K-th block column/row. For ease of understanding, the algorithm presented here is simplified. The actual implementation uses a pipelined organization so that processes PROCC (K step (1) of iteration K soon as the rank-b update (step (3)) of iteration K to block column index Storage of block column of L # of blocks nzval block # row subscripts # of full rows LDA in nzval block # row subscripts # of full rows0000000001111111110000000001111111110000000000000000111111111111111111110000000000000001111111111111111111100001111000000000000000000001111111111111111111100000000011111111111100000011111100011111100011111100000011111100110011100110000111111001101100111000111000111 Global Matrix Process Mesh U Figure 7: The 2-D block-cyclic layout and the data structure to store a local block column of L. Let mycol (myrow) be my process column (row) number in the process grid Let PROCC (K) (PROCR (K)) be the column (row) processes that own block column (row) K for block to N do Obtain the block column factor to the processes in my row who need it else Receive need it endif Perform parallel triangular solves to processes in my column who need it else Receive need it endif (3) for to N do for to N do endif end for Figure 8: Distributed sparse LU factorization algorithm. before completing the update to the trailing matrix A(K owned by PROCC (K 1). The pipelining alleviates the lack of parallelism in both steps (1) and (2). On 64 processors of Cray T3E, for instance, we observed speedups between 10% to 40% over the non-pipelined implementation. In each iteration, the major communication steps are send/receive L(K : N; K) across process rows and send/receive U(K;K process columns. Our data structure (see Figure 7) ensures that all the blocks of L(K : N; K) and U(K;K on a process are contiguous in memory, thereby eliminating the need for packing and unpacking in a send-receive operation or sending many more smaller messages. In each send-receive pair, two messages are exchanged, one for index[] and another for nzval[]. To further reduce the amount of communication, we employ the notion of elimination dags (EDAGs) [18]. That is, we send the K-th column of L rowwise to the process owning the J-th column of L only if there exists a path between (super)nodes K and J in the elimination dags. This is done similarly for the columnwise communication of rows of U . Therefore, each block in L may be sent to fewer than P c processes and each block in U may be sent to fewer than P r processes. In other words, our communication takes into account the sparsity of the factors as opposed to "send-to-all" approach in a dense factorization. For example, for AF23560 on (4 \Theta 8) processes, the total number of messages is reduced from 351052 to 302570, or 16% fewer messages. The reduction is even more with more processes or sparser problems. 3.3 Sparse triangular solve The sparse lower and upper triangular solves are also designed around the same distributed data structure. The forward substitution proceeds from the bottom of the elimination tree to the root, whereas the back substitution proceeds from the root to the bottom. Figure 9 outlines the algorithm for sparse lower triangular solve. The algorithm is based on a sequential variant called "inner product" formulation. In this formulation, before the K-th subvector x(K) is solved, the update from the inner product of must be accumulated and subtracted from b(K). The diagonal process, at the coordinate (K mod P r , K mod of the process grid, is responsible for solving x(K). Two counters, frecv and fmod, are used to facilitate the asynchronous execution of different operations. frecv[K] counts the number of process updates to x(K) to be received by the diagonal process owning x(K). This is needed because distributed among the row processes PROCR (K), and due to sparsity, not all processes in PROCR (K) contribute to the update. When frecv(K) becomes zero, all the necessary updates to x(K) are complete and x(K) is solved. fmod(K) counts the number of block modifications to be summed into the local inner product update (stored in lsum(K)) to x(K). When fmod(K) becomes zero, the partial sum lsum(K) is sent to the diagonal process that owns x(K). The execution of the program is message-driven. A process may receive two types of messages, one is the partial sum lsum(K), another is the solution subvector x(K). Appropriate action is taken according to the message type. The asynchronous communication enables large overlapping between communication and computation. This is very important because the communication to computation ratio is much higher in triangular solve than in factorization. The algorithm for the upper triangular solve is similar to that illustrated in Figure 9. However, because of the row oriented storage scheme used for matrix U , there is a slight complication in the actual implementation. Namely, we have to build two vertical linked lists to enable rapid access of the matrix entries in a block column of U . Let mycol (myrow) be my process column (row) number in the process grid Let PROCC (K) be the column processes that own block column K for each block K that I own Send x(K) to the column processes PROCC (K) endif end for while ( I have more work Receive a message (*) if ( message is lsum(K) ) Send x(K) to the column processes PROCC (K) endif else if ( message is x(K) ) for each I ? K, L(I ; K) 6= 0 that I own Send lsum(I) to the diagonal process who owns L(I ; I) endif end for endif while Figure 9: Distributed lower triangular solve L Symbolic Numeric Table 3: LU factorization time in seconds and Megaflop rate on the 512 node T3E-900. 3.4 Parallel performance Recall that we partition the blocks based on supernodes, so the largest block size equals the number of columns of the largest supernode. For large matrices, this can be a few thousand, especially towards the end of matrix L. Such a large granularity would lead to very poor parallelism and load balance. Therefore, when this occurs, we break the large supernode into smaller chunks, so that each chunk does not exceed our preset threshold, the maximum block size. By experimenting, we found that a maximum block size between 20 and 30 is good on the Cray T3E. We used 24 for all the performance results reported in this section. Table 3 shows the performance of the factorization on the Cray T3E-900. The symbolic analysis (steps (1) and (2) in Figure 1) is not yet parallel, so we start with a copy of the entire matrix on each processor, and run steps (1) and (2) independently on each processor. Thus the time is independent of the number of processors. The first column of Table 3 reports the time spent in the symbolic analysis. The memory requirement of the symbolic analysis is small, because we only store and manipulate the supernodal graph of L and the skeleton graph of U , which are much smaller than the graphs of L and U . The subsequent columns in the table show the factorization time with a varying number of processors. For four large matrices (BBMAT, ECL32, FIDAPM11 and WANG4), the factorization time continues decreasing up to 512 processors, demonstrating good scalability. The last column reports the numeric factorization rate in Mflops. More than 8 Gflops is achieved for matrix ECL32. This is the fastest published result we have seen for any implementation of parallel sparse Gaussian elimination. Table 3 starts with processors because some of the examples could not run with fewer processors. As a reference, we compare our distributed memory code to our shared memory SuperLU MT code using small numbers of processors. For example, using 4 processor DEC AlphaServer the factorization times of SuperLU MT for matrices AF23560 and EX11 are 19 and 23 seconds, respectively, comparable to the 4 processor T3E timings. This indicates that our distributed data structure and message passing algorithm do not incur much overhead. Table 4 shows the performance of the lower and upper triangular solves altogether. When the number of processors continues increasing beyond 64, the solve time remains roughly the same. Although triangular solves do not achieve high Megaflop rates, the time is usually much less than that for factorization. The efficiency of a parallel algorithm depends mainly on how the workload is distributed and how much time is spent in communication. One way to measure load balance is as follows. Let Each processor is the same as one T3E processor, except there is a 4 MB tertiary cache. BBMAT 3.69 3.42 2.27 2.23 1.83 56 ECL32 2.95 2.60 1.66 1.57 1.17 128 Table 4: Triangular solves time in seconds and Megaflop rate on the T3E-900. Comm fact sol Table 5: Load balance and communication on 64 processors Cray T3E. f i denote the number of floating-point operations performed on process i. We compute the load balance In other words, B is the average workload divided by the maximum workload. It is clear that better load balance. The parallel runtime is at least the runtime of the slowest process, whose workload is highest. In Table 5 we present the load balance factor B for both factorization and solve phases. As can be seen from the table, the distribution of workload is good for most matrices, except for TWOTONE. In the same table, we also show the fraction of the runtime spent in communication. The numbers were collected from the performance analysis tool called Apprentice on the T3E. The amount of communication is quite excessive. Even for the matrices that scale well, such as BBMAT, ECL32, FIDAPM11 and WANG4, more than 50% of the factorization time is spent in communication. For the solve, which has much smaller amount of computation, communication takes more than 95% of the total time. We expect the percentage of communication will be even higher with more processors, because the total amount of computation is more or less constant. Although TWOTONE is a relatively large matrix, the factorization does not scale as well as for the other large matrices. One reason is that the present submatrix to process mapping results in very poor load distribution. Another reason is due to long time in communication. When we look further into communication time using Apprentice, we found that processes are idle 60% of the time waiting to receive the column block of L sent from a process column on the left (step (1) in Figure 8), and are idle 23% of the time waiting to receive the row block of U sent from a process row from above (step (2) in Figure 8). Clearly, the critical path of the algorithm is in step (1), which must preserve certain precedence relation between iterations. Our pipelining method shortens the critical path to some extent, but we expect the length of the critical path can be further reduced by a more sophisticated DAG (task graph) scheduling. For the solve, we found that processes are idle 73% of the time waiting for a message to arrive (at line (*) in Figure 9). So on each process there is not much work to do but a large amount of communication. These communication bottlenecks also occur for the other matrices, but the problems are not so pronounced as TWOTONE. Another problem with TWOTONE is that supernode size (or block size) is very small, only 2.4 columns on average. This results in poor uniprocessor performance and low Megaflop rate. Concluding remarks and future work We propose a number of techniques in place of partial pivoting to stabilize sparse Gaussian elim- ination. Their effectiveness is demonstrated by numerical experiments. These techniques enable static analysis of the nonzero structure of the factors and the communication pattern. As a result, a more scalable implementation becomes feasible on large-scale distributed memory machines with hundreds of processors. Our preliminary software is being used in a quantum chemistry application at Lawrence Berkeley National Laboratory, where a complex unsymmetric system of order 200,000 has been solved within 2 minutes. 4.1 More techniques for numerical stability Although the current GESP algorithm is successful for a large number of matrices, it fails to solve one finite element matrix, AV41092, because the pivot growth is still too large with any combination of the current techniques. We plan to investigate other complementary techniques to further stabilize the algorithm. For example, we can use a judicious amount of extra precision to store some matrix entries more accurately, and to perform internal computations more accurately. This facility is available for free on Intel architectures, which performs all arithmetic most efficiently in 80-bit registers, and at modest cost on other machines. The extra precision can be used in both factorization and residual computation. We can also mix static and partial pivoting by only pivoting within a diagonal block owned by a single processor (or SMP within a cluster of SMPs). This can further enhance stability. We can use a more aggressive pivot size control strategy in step (4) of the algorithm. That is, instead of setting tiny pivots to " \Delta jjAjj, we may set it to the largest magnitude of the current column. This incurs a non-trivial amount of rank-1 perturbation to the original matrix. In the end, we use Sherman-Morrison-Woodbury formula [7] to recover the inverse of the original matrix, at the cost of a few more steps of inverse iteration. It remains to be seen in what circumstances these ideas should be employed in practice. There are also theoretical questions to be answered. 4.2 High performance issues In order to make the solver entirely scalable, we need to parallelize the symbolic algorithm. In this case, we will start with the matrix initially distributed in some manner. The symbolic algorithm then determines the best layout for the numeric algorithms, and redistributes matrix if necessary. This also requires us to provide a good interface so the user knows how to input the matrix in the distributed manner. For the LU factorization, we will investigate more general functions for matrix-to-process mapping and scheduling of computation and communication by exploiting more knowledge from the EDAGs. This is expected to relax much of the synchrony in the current factorization algorithm, and reduce communication. We also consider switching to a dense factorization, such as the one implemented in ScaLAPACK [4], when the submatrix at the lower right corner becomes sufficiently dense. The uniprocessor performance can also be improved by amalgamating small supernodes into large ones. To speed up the sparse triangular solve, we may apply some graph coloring heuristic to reduce the number of parallel steps [21]. There are also alternative algorithms other than substitutions, such as those based on partitioned inversion [1] or selective inversion [24]. However, these algorithms usually require preprocessing or different matrix distributions than the one used in our factorization. It is unclear whether the preprocessing and redistribution will offset the benefit offered by these algorithms, and will probably depend on the number of right-hand sides. 5 Related work Duff and Koster [13] applied the techniques of permuting large entries to the diagonal in both direct and iterative methods. In their direct method using a multifrontal approach, the numeric factorization first proceeds with diagonal pivots as previously chosen by the analysis on the structure of A+A T . If a diagonal entry is not numerically stable, its elimination will be delayed, and a larger frontal matrix will be passed to the later stage. They showed that using the initial permuta- tion, the number of delayed pivots were greatly reduced in factorization. They experimented with some iterative methods such as GMRES, BiCGSTAB and QMR using ILU preconditioners. The convergence rate is substantially improved in many cases when the initial permutation is employed. Amestoy, Duff and L'Excellent [2] implemented the above multifrontal approach for distributed memory machines. The host performs the fill-reducing ordering, estimates each frontal matrix structure, and statically maps the assembly tree, all based on the symmetric pattern of A and then sends the information to the other processors. During numerical factorization, each frontal matrix is factorized by a master processor and one or more slave processors. Due to possible delayed pivots, the frontal matrix size may be different than predicted by the analysis phase. So the master processor dynamically determines how many slave processors will be actually used for each frontal matrix. They showed good performance on processors IBM SP2. MCSPARSE [16] is a parallel unsymmetric linear system solver. The key component in the solver is the reordering step, which transforms the matrix into a bordered block upper triangular form. Their reordering first uses an unsymmetric ordering to put relatively large entries on the diagonal. The algorithm is a modified version of Duff [11, 12]. After this unsymmetric ordering, they use several symmetric permutations, which preserve the diagonal, to order the matrix into the desired form. With large diagonal entries, there is a better chance of obtaining a stable factorization by pivoting only within the diagonal blocks. The number of pivots from the border is thus reduced. Large and medium grain parallelism is then exploited to factor the diagonal blocks and eliminate the bordered blocks. They implemented the parallel factorization algorithm on a Cedar, an experimental shared memory machine. Fu, Jiao and Yang [15] designed a parallel LU factorization algorithm based on the following static information. The sparsity pattern of the Householder QR factorization of A contains the union of all sparsity patterns of the LU factors of A for all possible pivot selections. This has been used to do both memory allocation and computation conservatively (on possibly zero entries), but it can be arbitrarily conservative, particularly for matrices arising from circuit and device simulations. For several matrices that do not incur much overestimation, they showed good factorization speed on 128 processors Cray T3E. It will be interesting to compare the performance of the different approaches. 6 Acknowledgement We are grateful to Iain Duff for giving us access to the early version of the Harwell subroutine MC64, which permutes large entries to the diagonal. --R Highly parallel sparse triangular solution. Multifrontal parallel distributed symmetric and unsymmetric solvers. ScaLAPACK Users' Guide. University of Florida sparse matrix collection. Approximate minimum degree ordering for unsymmetric matrices. Applied Numerical Linear Algebra. A supernodal approach to sparse partial pivoting. An asynchronous parallel supernodal algorithm for sparse gaussian elimination. Algorithm 575. On algorithms for obtaining a maximum transversal. The design and use of algorithms for permuting large entries to the diagonal of sparse matrices. Users' guide for the Harwell-Boeing sparse matrix collection (release 1) Efficient sparse LU factorization with partial pivoting on distributed memory architectures. Solving large nonsymmetric sparse linear systems using MCSPARSE. Nested dissection of a regular finite element mesh. Elimination structures for unsymmetric sparse LU factors. Matrix Computations. Optimally scalable parallel sparse cholesky factorization. Scalable iterative solution of sparse linear systems. Sparse Gaussian elimination on high performance computers. Modification of the minimum degree algorithm by multiple elimination. Efficient parallel sparse triangular solution with selective inversion. An efficient block-oriented approach to parallel sparse cholesky factorization --TR Elimination structures for unsymmetric sparse <italic>LU</italic> factors An efficient block-oriented approach to parallel sparse Cholesky factorization Scalable iterative solution of sparse linear systems Modification of the minimum-degree algorithm by multiple elimination Solving large nonsymmetric sparse linear systems using MCSPARSE Applied numerical linear algebra user''s guide Efficient Sparse LU Factorization with Partial Pivoting on Distributed Memory Architectures On Algorithms for Obtaining a Maximum Transversal Algorithm 575: Permutations for a Zero-Free Diagonal [F1] Sparse Gaussian Elimination on High Performance Computers An Asynchronous Parallel Supernodal Algorithm for Sparse Gaussian A Supernodal Approach to Sparse Partial Pivoting --CTR Bergen , F. Hulsemann , U. Rude, Is 1.7 x 10^10 Unknowns the Largest Finite Element System that Can Be Solved Today?, Proceedings of the 2005 ACM/IEEE conference on Supercomputing, p.5, November 12-18, 2005 Laura Grigori , Xiaoye S. Li, A new scheduling algorithm for parallel sparse LU factorization with static pivoting, Proceedings of the 2002 ACM/IEEE conference on Supercomputing, p.1-18, November 16, 2002, Baltimore, Maryland Mark Baertschy , Xiaoye Li, Solution of a three-body problem in quantum mechanics using sparse linear algebra on parallel computers, Proceedings of the 2001 ACM/IEEE conference on Supercomputing (CDROM), p.47-47, November 10-16, 2001, Denver, Colorado Olaf Schenk , Klaus Grtner, Two-level dynamic scheduling in PARDISO: improved scalability on shared memory multiprocessing systems, Parallel Computing, v.28 n.2, p.187-197, February 2002 Patrick R. Amestoy , Iain S. Duff , Jean-Yves L'excellent , Xiaoye S. Li, Analysis and comparison of two general sparse solvers for distributed memory computers, ACM Transactions on Mathematical Software (TOMS), v.27 n.4, p.388-421, December 2001 Xiaoye S. Li, An overview of SuperLU: Algorithms, implementation, and user interface, ACM Transactions on Mathematical Software (TOMS), v.31 n.3, p.302-325, September 2005 Xiaoye S. Li , James W. Demmel , David H. Bailey , Greg Henry , Yozo Hida , Jimmy Iskandar , William Kahan , Suh Y. Kang , Anil Kapur , Michael C. Martin , Brandon J. Thompson , Teresa Tung , Daniel J. Yoo, Design, implementation and testing of extended and mixed precision BLAS, ACM Transactions on Mathematical Software (TOMS), v.28 n.2, p.152-205, June 2002 Anshul Gupta, Recent advances in direct methods for solving unsymmetric sparse systems of linear equations, ACM Transactions on Mathematical Software (TOMS), v.28 n.3, p.301-324, September 2002 Xiaoye S. Li , James W. Demmel, SuperLU_DIST: A scalable distributed-memory sparse direct solver for unsymmetric linear systems, ACM Transactions on Mathematical Software (TOMS), v.29 n.2, p.110-140, June

MPI;static pivoting;iterative refinement;2-D matrix decomposition;sparse unsymmetric linear systems

509094

Tuning Strassen''s matrix multiplication for memory efficiency.

Strassen's algorithm for matrix multiplication gains its lower arithmetic complexity at the expense of reduced locality of reference, which makes it challenging to implement the algorithm efficiently on a modern machine with a hierarchical memory system. We report on an implementation of this algorithm that uses several unconventional techniques to make the algorithm memory-friendly. First, the algorithm internally uses a non-standard array layout known as Morton order that is based on a quad-tree decomposition of the matrix. Second, we dynamically select the recursion truncation point to minimize padding without affecting the performance of the algorithm, which we can do by virtue of the cache behavior of the Morton ordering. Each technique is critical for performance, and their combination as done in our code multiplies their effectiveness.Performance comparisons of our implementation with that of competing implementations show that our implementation often outperforms the alternative techniques (up to 25%). However, we also observe wide variability across platforms and across matrix sizes, indicating that at this time, no single implementation is a clear choice for all platforms or matrix sizes. We also note that the time required to convert matrices to/from Morton order is a noticeable amount of execution time (5% to 15%). Eliminating this overhead further reduces our execution time.

Introduction The central role of matrix multiplication as a building block in numerical codes has generated a significant amount of research into techniques for improving the performance of this basic operation. Several of these efforts [3, 6, 12, 13, 14, 19] focus on algorithms whose arithmetic complexity This work supported in part by DARPA Grant DABT63-98-1-0001, NSF Grants CDA-97-2637 and CDA-95- 12356, Duke University, and an equipment donation through Intel Corporation's Technology for Education 2000 Program. Chatterjee is partially supported by NSF CAREER Award CCR-95-01979. Lebeck is partially supported by NSF CAREER Award MIP-97-02547. is O(n log 2 7 ) instead of the conventional O(n 3 ) algorithm. Strassen's algorithm [23] for matrix multiplication and its variants are the most practical of such algorithms, and are classic examples of theoretically high-performance algorithms that are challenging to implement efficiently on modern high-end computers with deep memory hierarchies. Strassen's algorithm achieves its lower complexity using a divide-and-conquer approach. Unfor- tunately, this technique has two potential drawbacks. First, if the division proceeds to the level of single matrix elements, the recursion overhead (measured, for instance, by the recursion depth and additional temporary storage) becomes significant and reduces performance. This overhead is generally limited by stopping the recursion early and performing a conventional matrix multiplication on submatrices that are below the recursion truncation point [13]. Second, the division step must efficiently handle odd-sized matrices. This can be solved by one of several schemes: by embedding the matrix inside a larger one (called static padding), by decomposing into submatrices that overlap by a single row or column (called dynamic overlap), or by performing special case computation for the boundary cases (called dynamic peeling). Previous implementations have addressed these two drawbacks independently. We present a novel solution that simultaneously addresses both issues. Specifically, dynamic peeling was introduced as a method to avoid large amounts of static padding. The large amount of static padding is an artifact of using a fixed recursion truncation point. We can minimize padding by dynamically selecting the recursion truncation point from a range of sizes. However, this scheme can induce significant variations in performance when using a canonical storage scheme (such as column-major) for the matrices. By using a hierarchical matrix storage scheme, we can dynamically select the recursion truncation point within a range of sizes that both ensures high and stable performance of the computations at the leaf of the recursion tree and limits the amount of static padding. We measured execution times of our implementation (MODGEMM) and two alternative im- plementations, DGEFMM uses dynamic peeling [13] and DGEMMW uses dynamic overlap [6], on both a DEC Alpha and a SUN UltraSPARC II. Our results show wide variability in the performance of all three implementations. On the Alpha, our implementation (MODGEMM) ranges from 30% slower to 20% faster than DGEFMM for matrix sizes from 150 to 1024. On the Ultra, MOD- GEMM is generally faster (up to 25%) than DGEFMM for large matrices (500 and larger), while DGEFMM is generally faster for small matrices (up to 25%). We also determine the time to convert matrices to/from Morton order ranges from 5% to 15% of total execution time. When eliminating this conversion time, by assuming matrices are already in Morton order, MODGEMM outperforms DGEFMM for nearly all matrix sizes on both the Alpha and Ultra, with greater benefits on the Ultra. The remainder of this paper is organized as follows. Section 2 reviews the conventional description of Strassen's algorithm. Section 3 discusses the implementation issues that affect memory efficiency, and our solutions to these issues. Section 4 presents performance results for our code. Section 5 cites related work and compares our techniques to them. Section 6 presents conclusions and future work. Background Strassen's original algorithm [23] is usually described in the following divide-and-conquer form. Let A and B be two n \Theta n matrices, where n is an even integer. Partition the two input matrices A and B and the result matrix C into quadrants as follows. A 11 A 12 A 21 A 22 (1) The symbol ffl in equation (1) represents matrix multiplication. We then compute the four quadrants of the result matrix as follows. +A 22 ffl (B 21 In this paper, we discuss and implement Winograd's variant [7] of Strassen's algorithm, which uses seven matrix multiplications and 15 matrix additions. It is well-known that this is the minimum number of multiplications and additions possible for any recursive matrix multiplication algorithm based on division into quadrants. The division of the matrices into quadrants follows equation (1). The computation proceeds as follows. C 22 U 6 Compared to the original algorithm, the noteworthy feature of Winograd's variant is its identification and reuse of common subexpressions. These shared computations are responsible for reducing the number of additions, but also contribute to worse locality of reference unless special attention is given to this aspect of the computation. We do not discuss in this paper numerical issues concerning these fast matrix multiplication algorithms, as they are covered elsewhere [10]. 2.1 Interface In order to stay consistent with previous work in this area and to permit meaningful comparisons, our implementation of Winograd's variant follows the same calling conventions as the dgemm subroutine in the Level 3 BLAS library [5]. Thus, the implementation computes C / ff are scalars, op(A) is an m \Theta k real matrix, op(B) is a k \Theta n real matrix, C is an m \Theta n real matrix, and op(X) is either X or X T . The matrices A, B, and C are stored in column-major order, with leading dimensions ldA, ldB, and ldC respectively. 3 Memory Efficiency: Issues and Techniques To be useful in practice, an implementation of the Strassen-Winograd algorithm must answer three questions: when to truncate the recursion, how to handle arbitrary rectangular matrices, and how to lay out the array data in memory to promote better use of cache memory. These questions are not independent, although past implementations have treated them thus. We now review these implementation issues, identify possible solution strategies, and justify our specific choices. 3.1 Recursion truncation point The seven products can be computed by recursively invoking Strassen's algorithm on smaller sub- problems, and switching to the conventional algorithm at some matrix size T (called the recursion truncation point [13]) at which Strassen's construction is no longer advantageous. If one were to estimate running time by counting arithmetic operations, the recursion truncation point would be around 16. However, the empirically observed value of this parameter is at least an order of magnitude higher. This discrepancy is a direct result of the poor (algorithmic) locality of reference of Strassen's algorithm. All implementations of Strassen's algorithm that we have encountered use an empirically chosen cutoff criterion for determining the matrix size T at which to terminate recursion. 3.2 Handling arbitrary matrices Divide-and-conquer techniques are most effective when the matrices can be evenly partitioned at each recursive invocation of the algorithm. We first note that rectangular matrices present no particular problem for partitioning if all matrix dimensions are even. The trouble arises when we encounter matrices with one or more dimensions of odd size. There are several possible solutions to this problem. ffl The simplest solution, static padding, is to pad the n \Theta n matrices with additional rows and columns containing zeros such that the padded n 0 \Theta n 0 matrix satisfies the "even-dimensions" condition at each level of recursion, i.e., n is the recursion depth. This is Strassen's original solution and the solution most often quoted in algorithms textbooks. However, for a statically predetermined value of T , the overhead of static padding can become quite severe, adding almost three times the number of original matrix elements in the worst case. Furthermore, in a naive implementation of this idea, the arithmetic done on these additional zero elements is pure overhead. Finally, based on the relation between n 0 and interference phenomena can reduce performance. These interference effects can be mitigated using non-standard data layouts, as we discuss further in Section 3.3. ffl A second solution, dynamic peeling, peels off the extra row or column at each level, and separately adds their contributions to the overall solution in a later fix-up computation [13]. This eliminates the need for extra padding, but reduces the portion of the matrix to which Strassen's algorithm applies, thus reducing the potential benefits of the recursive strategy. The fix-up computations are matrix-vector operations rather than matrix-matrix operations, which limits the amount of reuse and reduces performance. ffl A third solution, dynamic overlap, finesses the problem by subdividing the matrix into submatrices that (conceptually) overlap by one row or column, computing the results for the shared row or column in both subproblems, and ignoring one of the copies. This is an interesting solution, but it complicates the control structure and performs some extra computations. Ideally, we would like to avoid both the implementation complexity of dynamic peeling or dynamic overlap and the possibility of excessive static padding. Figure 1: Morton-ordered matrix layout. Each small square is a T \Theta T tile that is laid out contiguously in column-major order. The number in each tile gives its relative position in the sequence of tiles. 3.3 Data layout A significant fraction of the computation of the Strassen-Winograd algorithm occurs in the routine that multiplies submatrices when the recursion truncates. The performance of this matrix product is largely determined by its cache behavior. This issue has not been explicitly considered in previous work on implementing fast recursive matrix multiplication algorithms, where the default column-major layout of array data has been assumed. A primary condition for performance is to choose a tile size T such that the tiles fit into the first-level cache, thus avoiding capacity misses [11]. This is easily achieved using tile sizes in the range shown in Figure 2. Second, to achieve performance stability as T varies, it is also important to have the tile contiguous in memory, thus avoiding self-interference misses [17]. Given the hierarchical nature of the algorithm (the decomposition is by quadrants within quadrants within hierarchical layouts such as Morton ordering [8] naturally suggest themselves for storing the matrices. An operational definition of Morton ordering is as follows. Divide the original matrix into four quadrants, and lay out these quadrants in memory in the order NW, NE, SW, SE. A submatrix larger on the side than T is laid out recursively using the Morton ordering; a T \Theta T tile is laid out using column-major ordering. See Figure 1. A secondary benefit of keeping tiles contiguous in memory is that the matrix addition operations can be performed with a single loop rather than two nested loops, thus reducing loop overheads. 3.4 The connections among the issues We begin by observing that the recursion truncation point T determines the amount of padding since the matrix must evenly divide at all recursive invocations of the algorithm. We also note that at the recursion truncation point, we are multiplying T \Theta T submatrices using the conventional algorithm. By carefully selecting the truncation point, we can minimize the amount of padding required. Figure 2 shows the effects of tile size selection on padding. The four lines correspond to the Size Size Padding with Tilesize Best Case Padding Tile-size for Best Case Padding Matrix Size Figure 2: Effect of tile size on padding. When minimizing padding, tiles are chosen from the range to 64. original matrix size n, the padded matrix size n 0 with the tile size T chosen to minimize padding, the padded matrix size n 0 for fixed tile size and the tile size T that achieves the minimum padding. This figure demonstrates that the ability to select from a range of tile sizes can dramatically reduce the amount of extra storage, making it independent of the matrix size n. In contrast, a fixed tile size can require significant padding, proportional to the matrix size in the worst case. Consider a square matrix size of 513. With a fixed tile size of 32, static padding requires a padded matrix of size 1024, nearly twice the original matrix dimension. In contrast, flexibility in choosing tile size allows us to select a tile size of 33, which requires padding with only 15 (our worst case amount) extra elements in each dimension. The padded matrix size, 528, is recursively divided four times to achieve 33 \Theta 33 tiles. However, we can exploit this flexibility in tile size selection only if we can ensure that the performance of the matrix multiplication algorithm for these tiles is not sensitive to the choice of tile size. This is an important consideration, since a significant portion of the algorithm's computation occurs in the routine that multiplies tiles. Figure 3a and Figure 3b show how the performance of matrix multiplication, C / A \Theta B, varies as a function of the relation between tile size and leading dimension. Each line in the graph corresponds to a different submatrix size 24, 28, 32). The submatrices to operate on are chosen from a base matrix (M) as follows: Non-contiguous submatrices are created by setting the leading dimension of each submatrix to the leading dimension of the base matrix, M, which corresponds to the x-axis. Contiguous submatrices are formed by setting the leading dimension of each submatrix to the tile size, T, which corresponds to each line in the graph. From Figure 3 we see that contiguous submatrices exhibit much more stable performance than non-contiguous submatrices. As expected, the non-contiguous submatrices exhibit a dramatic drop in performance when the base matrix is a power of two (256 in this case), due to self-interference. On the Alpha, the contiguous submatrices clearly outperform the non-contiguous. The performance differential is not as pronounced on the Ultra, however the instability of the non-contiguous Matrix Size50.070.090.0MFlops Contiguous Sub-Matrices Non-Contiguous Sub-Matrices Matrix Size50.070.090.0MFlops Contiguous Sub-Matrices Non-Contiguous Sub-Matrices a.) DEC Miata b.) Sun Ultra Figure 3: Effect of Data Layout on Matrix Multiply Performance submatrices still exists. These results justify the use of Morton ordering within our implementation. Implementation details We can envision three different alternatives for storage layout: the input and output matrices are assumed to be laid out in Morton order at the interface level; the input and output matrices are copied from column-major storage to Morton order at the interface level; and the conversion to Morton order is done incrementally across the levels of recursion. The first alternative is not feasible for a library implementation. Among the two other options, converting the matrices to and from Morton order at the top level was easier to implement, and performed relatively fast in practice (5% to 15% of total execution time, see Section 4.1). We incorporate any necessary matrix transposition operations during the conversion from column-major to Morton order. This is handy, because it requires only a single core routine for the Strassen-Winograd algorithm. The alternative solution requires multiple code versions or indirection using pointers to handle these cases correctly. Choosing tile sizes from a range as described above will in general induce a small constant amount of padding. In our implementation, we explicitly padded out the matrix with zeros and performed redundant computation on the pad. We could afford to do this because the pad was guaranteed to be small. The alternative scheme of avoiding this overhead would have created tiles of uneven sizes and required more control overhead to keep track of tile start offsets and similar pieces of information. We handle rectangular cases by treating the two dimensions separately. Each tile dimension is chosen to minimize padding in that dimension. This method of choosing each tile dimension independently works when the ratio of columns to rows (or rows to columns) is within certain limits. Highly rectangular matrices pose a problem because the two recommended tile dimensions may require different levels of recursion. The following example illustrates the difficulties of this method for such matrices. Consider a highly rectangular matrix of dimensions 1024x256. We choose the tile dimensions independently. First, we consider that the matrix has 1024 rows, and choose the number of rows x x x a.) Lean A and wide B. b.) Wide A and lean B Figure 4: Handling of Highly Rectangular Matrices. in the tile that minimizes row padding (i.e., the number of additional rows). In this case, 32 is chosen, and the recursion is required to unfold to a depth of 5. Next, we consider that the matrix has 256 columns. The number of columns in the tile is chosen to minimize the number of columns that are to be padded. Again, we choose 32, but the recursion must unfold to a depth of only 3. Clearly, naively choosing the two tile dimensions independently does not work for highly rectangular matrices, since we can not unfold the recursion to both a depth of 5 and to a depth of 3. To overcome this limitation, the matrix is divided into submatrices such that all submatrices require the same depth of recursion unfolding for both dimensions. The matrix product is reconstructed in terms of the submatrix products. A given matrix can be ffl wide, meaning its columns-to-rows ratio exceeds the desired ratio, ffl lean, meaning its rows-to-columns ratio exceeds the desired ratio, or ffl well-behaved meaning both its columns-to-rows ratio and rows-to-columns ratio are within the desired ratio. Since there are two input matrices and B), and each can any one of the above forms, there are a total of nine possible combinations. Figure 4a and Figure 4b show two examples of how the input matrices and B) are divided, and how the result (C) is reconstructed from results of submatrix multiplications. Finally, we note that 1 and 0 are common values for the ff and fi parameters. In order to avoid performing extra arithmetic for these parameter values, the core routine for the Strassen-Winograd algorithm computes D / A ffl B, with D being set to C if and to a temporary otherwise. We then post-process to compute C / ff D post-processing is necessary. This section compares the performance of our implementation (MODGEMM) of Strassen's algorithm to a previous implementation that uses dynamic peeling (DGEFMM) [13] (we use the au- thor's original code), and to a previous implementation that uses dynamic overlap (DGEMMW) [6]. We measure the execution time of the various implementations on a 500 MHz DEC Alpha Miata and a 300 MHz Sun Ultra 60. The Alpha machine has a 21164 processor with an 8KB direct-mapped level 1 cache, a 96KB 3-way associative level 2 cache, a 2MB direct-mapped level 3 cache, and 512MB of main memory. The Ultra has two UltraSPARC II processors, each with a level 1 cache, a 2MB level 2 cache, and 512MB of main memory. We use only one processor on the Ultra 60. Matrix Size0.801.001.201.40Normalized Execution Time MODGEMM Matrix Size0.801.001.201.40Normalized Execution Time a.) MODGEMM vs DGEFMM b.) DGEMMW vs DGEFMM Figure 5: Performance of Strassen Winograd Implementations on DEC Miata We timed the execution of each implementation using the UNIX system call getrusage for matrix sizes ranging from 150 to 1024, and For DGEFMM we use the empirically determined recursion truncation point of 64. For matrices less than 500 we compute the average of invocations of the algorithm to overcome limits in clock resolution. Execution times for larger matrices are large enough to overcome these limitations. To further reduce experimental error, we execute the above experiments three times for each matrix size, and use the minimum value for comparison. The programs were compiled with vendor compilers (cc and f77) with the -fast option. The Sun compilers are the Workshop Compilers 4.2, and the DEC compilers are DEC C V5.6-071 and DIGITAL Fortran 77 V5.0-138-3678F. Figure 5 and Figure 6 show our results for the Alpha and UltraSPARC, respectively. We report results in execution time normalized to the dynamic peeling implementation (DGEFMM). On the Alpha we see that DGEFMM generally outperforms dynamic overlap (DGEMMW), see Figure 5b. In contrast, our implementation (MODGEMM) varies from 30% slower to 20% faster than DGEFMM. We also observe that MODGEMM outperforms DGEFMM mostly in the range of matrix sizes from 500 to 800, whereas DGEFMM is faster for smaller and larger matrices. Finally, by comparing Figure 5a and Figure 5b, we see that MODGEMM generally outperforms DGEMMW. The results are quite different on the Ultra (see Figure 6). The most striking difference is the performance of DGEMMW (see Figure 6b), which outperforms both MODGEMM and DGEFMM for most matrix sizes on the Ultra. Another significant difference is that MODGEMM is generally faster than DGEFMM for large matrices (500 and larger), while DGEFMM is generally faster for small matrices. An important observation from the above results is the variability in performance both across platforms and across matrix sizes. Our ongoing research efforts are targeted at understanding these variations. Section 4.2 reports on some of our preliminary findings, and the following section analyzes the penalty of converting to Morton order. Matrix Size0.801.001.201.40Normalized Execution Time MODGEMM Matrix Size0.801.001.201.40Normalized Execution Time a.) MODGEMM vs DGEFMM b.) DGEMMW vs DGEFMM Figure Performance of Strassen Winograd Implementations on Sun Ultra Matrix Size5.015.0 Conversion Cost (%age of Execution Matrix Size5.015.0 Conversion Cost (%age of Execution a.) DEC Miata b.) Sun Ultra Figure 7: Morton Conversion Time as Percentage of Total Execution Time Matrix Size0.801.001.201.40Normalized Execution Time MODGEMM Matrix Size0.801.001.201.40Normalized Execution Time MODGEMM a.) DEC Miata b.) Sun Ultra Figure 8: Performance of MODGEMM without Matrix Conversion 4.1 Morton Conversion Time An important aspect of our implementation is the recursive data layout, which provides stable performance for dynamic tile size selection. The previous performance results include the time to convert the two input matrices from column-major to Morton order, and to convert the output matrix from Morton order back to column-major. Figure 7a and Figure 7b show the cost of this conversion as a percentage of the entire matrix multiply for each of our platforms. From these graphs we see that Morton conversion accounts for up to 15% of the overall execution time for small matrices and approximately 5% for very large matrices. These results show that Morton conversion is a noticeable fraction of the execution time. Eliminating the conversion cost (i.e, assuming the matrices are already in Morton order) produces commensurate improvements in the performance of our implementation. Figure 8a and Figure 8b shows that without conversion costs MODGEMM does indeed execute faster, and increases the number of matrix sizes at which it outperforms DGEFMM. Specifically, on the Alpha (Figure 8a) MODGEMM is superior for most matrix sizes larger than 500, and on the Ultra (Figure 8b) we see that for only a few matrix sizes DGEFMM outperforms MODGEMM. Furthermore, without Morton conversion, MODGEMM is very competitive with DGEMMW, and outperforms it for many matrix sizes. 4.2 Cache Effects Our initial efforts to gain further insight into the performance variability of our implementation begins with analysis of its cache behavior. Here, we present preliminary results. We used ATOM [22] to perform cache simulations of a 16KB direct-mapped cache with blocks of both the DGEFMM, and MODGEMM implementations. Figure 9 shows the miss ratios of each implementation for matrix sizes ranging from 500 to 523. The first observation from this graph is that MODGEMM miss ratios (6% to 2%) are lower than DGEFMM (8%), which matches our expec- tations. The second observation is the unexpected dramatic drop in MODGEMM's miss ratio at a matrix size of 513. Preliminary investigations using CProf [18] reveal that this drop is due to a reduction in conflict misses. Matrix Size0.0400.080Miss Figure 9: Cache Miss Ratios for 16KB Direct-Mapped with To understand this phenomenon, consider that for matrix sizes of 505 to 512 the padded matrix size is 512 and the recursion truncation point is at tile size 32. The conventional algorithm is applied to submatrices that are each 8KB in size (32x32x8), and correspond to the four quadrants of a 64x64 submatrix. With Morton ordering the quadrants are allocated contiguously in memory, and quadrants separated by a multiple of the cache size (16KB) conflict in the cache. For example, since the NW and SW quadrants are separated by the NE quadrant, they map to the same locations in cache (i.e, the first elements of the NW and SW quadrants are separated by 16KB). Therefore, any operations involving these two quadrants will incur a significant number of cache misses. We are currently examining ways to eliminate these conflict misses. 5 Related Work We discuss three separate areas of related work: implementations of Strassen-type algorithms, hierarchical schemes for matrix storage, and compiler technology for improving the cache behavior of loop nests. 5.1 Other implementations Previous implementations of Strassen's algorithm include Bailey's implementation for the CRAY- 2 [3], the DGEMMW implementation by Douglas et al. [6], and the DGEFMM implementation by Huss-Lederman et al.[13]. Bailey, coding in Fortran, used a static two-level unfolding of the recursion by code duplication. Douglas et al. introduced the dynamic overlap method for handling odd-sized dimensions. Huss-Lederman et al. introduced the dynamic peeling method. While all of these implementations are careful to limit the amount of temporary storage, they do not specifically consider the performance of the memory hierarchy on their code. In some cases (such as on the CRAYs), this issue did not arise because the memory system was not cache-based. Section 4 gives extensive performance comparisons of our implementation vs. DGEFMM and DGEMMW. Kreczmar [16] proposes an elegant memory-efficient version of Strassen's algorithm based on overwriting one of the input arguments. His scheme for space savings is not directly applicable for two reasons: we cannot assume that the input matrices can be overwritten, and his scheme requires several copying operations that reduce performance. 5.2 Hierarchical schemes for matrix storage Wise and his coauthors [1, 24] have investigated the algorithmic advantages of quad-tree representations of matrices. Morton ordering has also appeared in the parallel computing literature, where is has been used for load balancing of irregular problems [20]. Most recently, Frens and Wise [8] discuss an implementation of a recursive O(n 3 ) matrix multiplication algorithm using hierarchical matrix layout, in which they sequence the recursive calls in an unusual manner to get better reuse in cache. We do not carry the recursion to the level of single matrix elements as they do, but truncate the recursion when we reach tile sizes that fit in the upper levels of the memory hierarchy. 5.3 Cache behavior Several authors [17, 15, 4, 21] discuss loop transformations such as tiling that attempt to reduce the number of cache misses incurred by a loop nest and thus improve its performance. While these loop transformations are not specific to matrix multiplication, the conventional three-loop algorithm for matrix multiplication falls into the category of codes that they can handle. Lam, Rothberg, and Wolf [17] investigated and modeled the influence of cache interference on the performance of tiled programs. They emphasized the importance of having tiles be contiguous in memory to avoid self-interference misses, and proposed data copying to satisfy this condition. Our top-level conversion between the column-major layout at the interface level and the Morton ordering used internally can be viewed as a logical extension of this proposal. Ghosh et al. [9] present an analytical representation of cache misses for perfect loop nests, which they use to guide selected code optimization problems. Their work, like all of the other work cited above, relies on linear (row- or column-major) storage of arrays, and therefore does not immediately apply to our code. 6 Conclusions and Future Work Matrix multiplication is an important computational kernel, and its performance can dictate the overall performance of many applications. Strassen's algorithm for matrix multiplication achieves lower arithmetic complexity, O(n log 2 7 ), than the conventional algorithm, O(n 3 ), at the cost of worse locality of reference. Furthermore, since Strassen's algorithm is based on divide-and-conquer, an implementation must handle odd-size matrices, and reduce recursion overhead by terminating the recursion before it reaches individual matrix elements. These issues make it difficult to obtain efficient implementations of Strassen's algorithm. In this paper we presented a practical implementation of Strassen's algorithm (Winograd vari- ant) that exploits the ability to dynamically select the recursion truncation point based on matrix size and efficiently handles odd-sized matrices. We achieve this by using a non-standard array layout called Morton order; by converting from standard layouts (e.g., column-major) to internal Morton layout at the interface level; and by exploiting dynamic selection of the recursion truncation point to minimize padding. We compare our implementation to two alternative implementations that use dynamic peeling (DGEFMM) [13] and dynamic overlap (DGEMMW) [6]. Execution time measurements on a DEC Alpha and a SUN UltraSPARC II reveal wide variability in the performance of all three imple- mentations. On the Alpha, our implementation (MODGEMM) ranges from 30% slower to 20% faster than DGEFMM for matrix sizes from 150 to 1024. On the Ultra, MODGEMM is generally faster than DGEFMM for large matrices (500 and larger), while DGEFMM is generally faster for small matrices. When eliminating the time to convert matrices to/from Morton order (5% to 15% of total execution time), MODGEMM outperforms DGEFMM for nearly all matrix sizes on the Ultra, and for most matrices on the Alpha. Our future work includes investigating techniques to further improve the performance and stability of Strassen's algorithm, while minimizing code complexity. We also plan to examine the effects of rectangular input matrices. Our implementation supports the same interface as Level 3 BLAS dgemm routine [2], we plan to examine its performance for a variety of input parameters. --R Experiments with quadtree representation of matrices. Extra high speed matrix multiplication on the Cray-2 Tile size selection using cache organization and data layout. A set of level 3 basic linear algebra subprograms. GEMMW: a portable level 3 BLAS Winograd variant of Strassen's matrix-matrix multiply algorithm Efficient procedures for using matrix algorithms. Cache miss equations: An analytical representation of cache misses. Accuracy and Stability of Numerical Algorithms. Evaluating associativity in CPU caches. A tensor product formulation of Strassen's matrix multiplication algorithm. Implementation of Strassen's algorithm for matrix multiplication. IBM engineering and scientific subroutine library guide and reference On memory requirements of Strassen's algorithms. The cache performance and optimizations of blocked algorithms. Cache profiling and the SPEC benchmarks: A case study. Dynamic partitioning of non-uniform structured workloads with spacefilling curves Data transformations for eliminating conflict misses. Atom a system for building customized program analysis tools. Gaussian elimination is not optimal. Costs of quadtree representation of nondense matrices. --TR Extra high speed matrix multiplication on the Cray-2 Evaluating Associativity in CPU Caches A set of level 3 basic linear algebra subprograms The cache performance and optimizations of blocked algorithms LAPACK''s user''s guide ATOM GEMMW Tile size selection using cache organization and data layout Dynamic Partitioning of Non-Uniform Structured Workloads with Spacefilling Curves multi-level blocking Cache miss equations Auto-blocking matrix-multiplication or tracking BLAS3 performance from source code Data transformations for eliminating conflict misses Implementation of Strassen''s algorithm for matrix multiplication Accuracy and Stability of Numerical Algorithms Cache Profiling and the SPEC Benchmarks Efficient Procedures for Using Matrix Algorithms Experiments with Quadtree Representation of Matrices --CTR Siddhartha Chatterjee , Alvin R. Lebeck , Praveen K. Patnala , Mithuna Thottethodi, Recursive Array Layouts and Fast Matrix Multiplication, IEEE Transactions on Parallel and Distributed Systems, v.13 n.11, p.1105-1123, November 2002 Hossam ElGindy , George Ferizis, On improving the memory access patterns during the execution of Strassen's matrix multiplication algorithm, Proceedings of the 27th Australasian conference on Computer science, p.109-115, January 01, 2004, Dunedin, New Zealand K. Fatahalian , J. Sugerman , P. Hanrahan, Understanding the efficiency of GPU algorithms for matrix-matrix multiplication, Proceedings of the ACM SIGGRAPH/EUROGRAPHICS conference on Graphics hardware, August 29-30, 2004, Grenoble, France Kang Su Gatlin , Larry Carter, Architecture-cognizant divide and conquer algorithms, Proceedings of the 1999 ACM/IEEE conference on Supercomputing (CDROM), p.25-es, November 14-19, 1999, Portland, Oregon, United States John Mellor-Crummey , David Whalley , Ken Kennedy, Improving memory hierarchy performance for irregular applications, Proceedings of the 13th international conference on Supercomputing, p.425-433, June 20-25, 1999, Rhodes, Greece John Mellor-Crummey , David Whalley , Ken Kennedy, Improving Memory Hierarchy Performance for Irregular Applications Using Data and Computation Reorderings, International Journal of Parallel Programming, v.29 n.3, p.217-247, June 2001 Igor Kaporin, The aggregation and cancellation techniques as a practical tool for faster matrix multiplication, Theoretical Computer Science, v.315 n.2-3, p.469-510, 6 May 2004 Sandeep Sen , Siddhartha Chatterjee, Towards a theory of cache-efficient algorithms, Proceedings of the eleventh annual ACM-SIAM symposium on Discrete algorithms, p.829-838, January 09-11, 2000, San Francisco, California, United States Lillian Lee, Fast context-free grammar parsing requires fast boolean matrix multiplication, Journal of the ACM (JACM), v.49 n.1, p.1-15, January 2002 Sandeep Sen , Siddhartha Chatterjee , Neeraj Dumir, Towards a theory of cache-efficient algorithms, Journal of the ACM (JACM), v.49 n.6, p.828-858, November 2002 Paolo D'Alberto , Alexandru Nicolau, Adaptive Strassen's matrix multiplication, Proceedings of the 21st annual international conference on Supercomputing, June 17-21, 2007, Seattle, Washington Nico Galoppo , Naga K. Govindaraju , Michael Henson , Dinesh Manocha, LU-GPU: Efficient Algorithms for Solving Dense Linear Systems on Graphics Hardware, Proceedings of the 2005 ACM/IEEE conference on Supercomputing, p.3, November 12-18, 2005 Siddhartha Chatterjee , Alvin R. Lebeck , Praveen K. Patnala , Mithuna Thottethodi, Recursive array layouts and fast parallel matrix multiplication, Proceedings of the eleventh annual ACM symposium on Parallel algorithms and architectures, p.222-231, June 27-30, 1999, Saint Malo, France Mohamed F. Mokbel , Walid G. Aref , Ibrahim Kamel, Analysis of Multi-Dimensional Space-Filling Curves, Geoinformatica, v.7 n.3, p.179-209, September Chun-Yuan Lin , Yeh-Ching Chung , Jen-Shiuh Liu, Efficient Data Parallel Algorithms for Multidimensional Array Operations Based on the EKMR Scheme for Distributed Memory Multicomputers, IEEE Transactions on Parallel and Distributed Systems, v.14 n.7, p.625-639, July Mohamed F. Mokbel , Walid G. Aref, Irregularity in multi-dimensional space-filling curves with applications in multimedia databases, Proceedings of the tenth international conference on Information and knowledge management, October 05-10, 2001, Atlanta, Georgia, USA Siddhartha Chatterjee , Vibhor V. Jain , Alvin R. Lebeck , Shyam Mundhra , Mithuna Thottethodi, Nonlinear array layouts for hierarchical memory systems, Proceedings of the 13th international conference on Supercomputing, p.444-453, June 20-25, 1999, Rhodes, Greece Richard Vuduc , James W. Demmel , Jeff A. Bilmes, Statistical Models for Empirical Search-Based Performance Tuning, International Journal of High Performance Computing Applications, v.18 n.1, p.65-94, February 2004

matrix multiply;strassen's algorithm;data layout;cache memory

509246

Distributed Memory Parallel Architecture Based on Modular Linear Arrays for 2-D Separable Transforms Computation.

A framework for mapping systematically 2-dimensional (2-D) separable transforms into a parallel architecture consisting of fully pipelined linear array stages is presented. The resulting model architecture is characterized by its generality, high degree of modularity, high throughput, and the exclusive use of distributed memory and control. There is no central shared memory block to facilitate the transposition of intermediate results, as it is commonly the case in row-column image processing architectures. Avoiding shared central memory has positive implications for speed, area, power dissipation and scalability of the architecture. The architecture presented here may be used to realize any separable 2-D transform by only changing the coefficients stored in the processing elements. Pipelined linear arrays for computing the 2-D Discrete Fourier Transform and 2-D separable convolution are presented as examples and their performance is evaluated.

Introduction Separable transforms play a fundamental role in digital signal and image processing. Nearly every problem in DSP is based on the transformation of a time or space domain signal to alternative spaces better suited for efficient storage, transmission, interpretation or estimation. The most commonly employed 2-dimensional (2-D) signal transforms, such as the Discrete Cosine, Fourier, Sine Transforms (DCT, DFT, and DST) and the Discrete Hough and Radon Transforms (DHT, DRT), are known as separable due to a special type of symmetry in their kernel [1]. Separable transforms are computationally less expensive than non-separable ones, having time complexity in O(M) per output data element, as opposed to O(M 2 ) for non-separable forms, when applied to a size M \Theta M input image. The majority of the available VLSI implementations for separable transforms are based on the popular row-column approach (see for example Rao and Yip [2], Guo et al. [3], Y.-P. Lee et al. [4], Bhaskaran and Konstantinides [5], Gertner and Shamash [6], and Chakrabarti and J'aj'a [7]), where the 2-D transform is performed in three steps: (i) 1-D transformation of the input rows, followed by (ii) intermediate result transposition (usually implemented with a transposition memory), and (iii) 1-D transformation of the intermediate result columns, as illustrated in Fig. 1 (a). In this figure, and throughout this paper, we assume that all image inputs become available in raster-scan order, that is, as a continuous sequence of row vectors. Arrays that can accept and produce data in raster-scan avoid expensive host interface memory buffering. Each of the 1-D row and column processing blocks in Fig. 1 (a) may be realized either with highly modular structures, such as standard pipelined arrays of identical Processing Elements (PEs), or by some other type of architecture optimized for the targeted transform. However, one of the main shortcomings of the conventional row-column architecture is that the central memory block severely limits the modularity of the overall structure, despite the possible modularity of the 1- D row and column arrays. In addition, central memory requires moderately complex address generation circuitry, row and column decoders, sense amplifiers depending on its size, and control for coordinating the concurrent access of the two row and column architectures to the shared memory. Large central memories also have negative implications for low-power design. In this paper we present a method for synthesizing fully pipelined modular arrays for 2-D separable transforms. The derived computational structures use only small-size FIFO memories local to the PEs which require neither address generation nor central control and memory-to-array routing. Faster clock speed, smaller area, and low power can be achieved with architectures having a regular set of simple PEs with distributed memory and control. Reducing the size of memory structures has been shown to reduce power consumption [8], largely due to reduced effective bus capacitances of smaller memories, and absence of sense amplifiers. The architectures synthesized are general in the sense that any separable transform can be realized by programming in the PEs memory bank for transposition address row processing (a) row column raster- scan input (b) logic column processing Figure 1: (a) Conventional architecture for raster-scan 2-D separable transforms relying on shared central memory bank for the transposition of intermediate results, and (b) proposed alternative architecture based on two linear arrays where the central memory bank is replaced by FIFO queues distributed to the PEs. the appropriate set of kernel coefficients. The benefits of eliminating memory transposition have been recognized by Lim and Swartzlander [9] and by Wang and Chang [10]. In [9], the authors treat the DFT problem in the context of multi-dimensional systolic arrays. Our method is distinct from [9] primarily in that it is restricted to single input (SI) linear arrays using raster-scan I/O. Fisher and Kung [11] have shown that linear arrays have bounded clock skew regardless of size, whereas higher than 2-D arrays of arbitrary size may not be technologically feasible. (See Lee and Kedem [12] for a discussion on general aspects of mapping algorithms to linear arrays.) Although the DCT arrays presented in [10] avoid transposition by means of local PE storage, the resulting architectures require M 2 multipliers on an M \Theta M input, which in many cases may be prohibitive. The arrays developed in this paper require only multipliers. Another important characteristic of the proposed architecture is that it uses localized communications and its interconnection complexity is in O(M) (for M \Theta M separable 2-D transforms). To the best of our knowledge, parallel architectures that may use less than O(M) PEs also require O(M) input/output ports and may exhibit non-local communications among PEs i.e. their inter-connection complexity may grow faster than O(M ). Furthermore, the reduction in the number of PEs becomes possible by exploiting the specific coefficient symmetries, i.e. these architectures may not compute any desirable 2-D separable transform, something that the architecture developed here is capable of. In addition to the general computational structure, we also derive here as examples arrays for the 2-D DFT and for 2-D separable convolution. Although purely from a computational efficiency point of view the DFT is more expensive than the Fast Fourier Transform (FFT), from a VLSI implementation point of view there are significant reasons why the DFT may be preferable to the FFT (particularly with a small number of coefficients). The FFT's complex data routing limits overall speed and is expensive in terms of chip area (see Swartzlander [13] and Thompson [14]). In this paper we use the 1-D DFT array by Beraldin et al. [15] and derive a modular array structure for computing the 2-D DFT without transposition memory. The 2-D separable convolution has been studied extensively for many years (see for example Dudgeon and Mersereau [16]) and has been used as a basis for constructing general transfer functions. Abramatic et al. [17] has used separable FIR systems for constructing non-separable IIR filters, and Treitel and Shanks [18] have used parallel separable filters for approximating non-separable ones. Despite the limited range of frequency responses of separable filters, they are attractive due to their low computational cost as compared to non-separable systems. In addi- tion, separable convolution plays a central role in the Discrete Wavelet Transform (DWT) [19], as currently most of the basis used for the 2-D DWT are separable. The rest of this paper is organized as follows. In Section 2 we fix notation, define general separable 2-D transforms, and discuss alternative views of conventional row-column 2-D processing that will allow us to formulate equivalent algorithms suitable for raster scan I/O. In Section 3 the derived algorithms are then transformed systematically to fully pipelined linear array structures with distributed memory and control using appropriate linear space-time mapping operators. In Sections 4 and 5 we apply the general method developed in Section 3 to derive modular linear arrays for the 2-D DFT, and for 2-D separable convolution, respectively. Transforms, Definitions and Notation A 2-D transform of an M \Theta M signal x(i; is given by Notice that in a general 2-dimensional (2-D) transform there are M 2 , 2-D distinct orthonormal basis functions spanning the transform space, namely g (where g denotes the complex conjugate of g) [1, 20]. Separable and symmetric transforms have the property that l;m (i; Hence one can express y(l; m) as l (i) In order to formulate the transforms in matrix form, we adopt the following notation: Let A be an M i \Theta M k matrix of elements indexed from zero and interchangeably also denoted as a(i; k) or a i;k . Furthermore, let column vectors M i \Theta 1 denoted by lower-case letters as where the superscript t denotes transposition. We adopt Golub and Van Loan's [21] Matlab-like notation and denote submatrices as A(u; v), where u and v are integer column vectors that pick out the rows and columns of A defining the submatrix. Also, index vectors are specified by colon notation as implying l . For example, is a 2 \Theta 5 submatrix containing rows 2 and 3, and columns 5 through 9 of matrix A. Entire rows or columns can be specified by a single colon as A(:; k), which corresponds to A's kth column, and A(i; :) denoting A's ith row. Index vectors with non-unit increments are specified as denoting a count from p to r in length-q increments. Let X and Y be the transform input and transform coefficient (output) M \Theta M matrices respectively, and M \Theta M matrix G be the separable transformation kernel whose columns are the M , 1-D distinct basis vectors such that G(i; , then matrix With this notation in place, Eqn. (2) can then be rewritten as where m). Consequently, a separable transform can be computed as two consecutive matrix-matrix products. The first matrix product can be written as and can be viewed as operating with matrix G on the rows of input X to produce the rows of the intermediate result Z. The second matrix product, can be viewed as operating with matrix G t on the columns of Z to produce the columns of Y . This interpretation forms the basis for the traditional row-column processing method that necessitates the transposition of Z. In Section 3 we show how Y a be obtained by processing directly the rows of Z in the order in which they become available, thus eliminating the need for transposition. We now give the definition for raster scan ordering. Definition 2.1 Given the set of matrix elements g, we say that element X(i; precedes in raster scan order element X(i 0 only if either i 0 Raster scan induces a total ordering on the input set, and is equivalent to row-major ordering. Let us also define the vectors x and z as x j and z j which are column vector views of matrices X and Z respectively, suitable for raster scan processing. means that the p th row of Z can be derived by multiplying the p th row of X times G, i.e. Z(p; This set of equations can be expressed compactly using Kronecker products [20, 22], by multiplying the 1 \Theta M 2 input vector x t times I M\Omega G to obtain z M\Omega G), or equivalently The matrix-vector product Eqn. (6) is an alternative view of Eqn. (4) that allows us to derive an algorithm of dimension two (i.e. an algorithm having index space of dimension two) for row processing, rather than of dimension three as implied by the matrix-matrix product form of Eqn. (4). Reformulating algorithms into lower dimensional index spaces, as we do here, is an effective technique for simplifying the mapping of algorithms of high dimension into simple and efficient mono-dimensional (linear) array structures, by avoiding the complexities of multiprojection [23]. The intermediate result column vector z in Eqn. (6) can be viewed either as being composed of M consecutive vectors z(i corresponding to the rows of matrix Z, or as M vectors interspersed by a stride factor of M [21], corresponding to the columns of matrix Z, and given by 1g. By adopting the latter view, column 1-D transform Eqn. (5) can be expressed as I M )z; (7) where vector y is defined similarly as x and z. 1 3 Array Synthesis Framework for Separable Transforms In this section we elaborate on the space time mapping for the two components of the general 2-D separable transform array: one for row, and one for column processing. For generality, we treat 1-D transformation as a matrix vector product and do not make use of any additional computational savings that might be possible due to kernel symmetries. Rather than write a three dimensional Single Assignment Form (SAF) [23] 2 algorithm for row processing as suggested by Eqn. (4) (con- ventional M \Theta M matrix-matrix product), we work with Eqn. (6) (matrix-vector product). The nested loop algorithm for Eqn. (6) is only of dimension two, and can be trivially mapped into a linear array for processing inputs in raster-scan order, due to the way we constructed vectors x and z. 1 As a check, recall that an alternate definition for a separable transform is that the basis matrix T of size M 2 \Theta M 2 be written as the Kronecker product of the basis M \Theta M matrix G times itself (see Akansu [1]). Substituting Eqn. (6) into Eqn. (7), we get t\Omega IM )x. Using the property (A\Omega B)(C\Omega (AC\Omega BD) [1, 20], we have that An algorithm in SAF is one in which all variables are assigned a single value throughout the entire execution, so statements of the form x(i) / x(i) b are not allowed. If we define the G as ~ Eqn. (6) is given below Algorithm 3.1 Inputs x(i) and ~ Initially z z k (i) / ~ Outputs z Algorithm 3.1 is defined over 2-D index space I = f(i; g. Using linear space-time mapping methods, under space transformation this algorithm can be transformed into an array with M Processing Elements (PEs), but with efficiency of only 50% due to the block diagonal sparse structure of ~ G. By applying a transformation on index variable i defined cM , we obtain an equivalent algorithm which has a rectangular Constant Bounded Index Space (CBIS) [27] given by I 0 g. This new algorithm can be mapped under into a linear array with M , fully utilized (100% efficient) PEs. Localization of broadcast variables is a complex and rich problem [23, 28, 29]. For maintaining focus we do not address here the details required for writing Algorithm 3.1 in localized form where all variables have a common domain of support, but rather give below as Algorithm 3.2 one possible localized algorithm suitable for raster scan processing. (Note that the variables ~ ~ z(i; are the localized equivalents of x(k), G and z(k) respectively). Algorithm 3.2 Inputs x(k), G(i; for for ~ ~ ~ Outputs z(b k Pictorially, the localized Algorithm 3.2 can be represented by its Dependence Graph (DG) [23] as shown in Fig. 2 (a) for 4, where inputs x(k) appear along the horizontal k-axis and are shown with the corresponding elements in matrix X. The operations modeled by a DG node are shown in Fig. 2 (b). In this figure, arrows represent data dependencies which are seen to be local among DG nodes for variables x and z. Not shown in the figure are the dependence vectors associated with variable G, which are horizontal and of length M ; note that the coefficients of matrix G appear in this DG replicated M times as one term for each DG section. Using the linear space map the DG of Fig. 2 (a) is projected along the i axis giving rise to the linear array in Fig. 2 (c). The linear time map (schedule) used is which is consistent with raster scan processing. Input stream in raster scan is fed to PE 0 , one data point per schedule time period, and the output stream becomes available after an initial delay of three time units at also in raster scan. So for example, at is available at and at at PE 0 again. Note that although outputs do not become available at a single PE, every time instant only one PE produces a z value, and consequently outputs can be collected by means of a bus connecting all PE outputs. 3 The local PE control needed to realize the output of partial results z in Algorithm 3.2, (i.e. recognize the inter-row boundaries at k 2 in the DG of Fig. 2 (a)) can be implemented simply by attaching a bit-tag (called decode bit in Fig. 2 (d)) to each input data token X. This bit is set only for the last element of every row of matrix X, and the PEs accumulate locally the result of the multiply-accumulate operation on z only when the bit of the inputs is not set. When a PE finds that an input in X(:; arrives that has the decode bit set, it places the result on an output link rather than on the internal z register, and then clears the internal z register. A column of transform coefficients G are stored in a bank of M registers, and are fed to the multiplier in sequence with wraparound. Next we derive a SAF algorithm for the transformation which can be viewed either as (i) matrix G t operating on column vectors Z(:; k) to produce column vectors Y (:; or as (ii) matrix Z operating on row vectors G t (i; :) to produce row vectors Y (i; :). Using the latter interpretation, for which in turn is equal to ?From this expression follows the nested-loop SAF algorithm given below which is very similar to Algorithm 3.1 except from the fact that the basic multiply-add operation is of the type "scalar times vector plus vector." 4 Algorithm 3.3 Alternatively, outputs can also be systolically pipelined along with inputs X so that they become available at a single PEM \Gamma1 . See Beraldin et al. [15] for a discussion on output pipelining and systolization for the DFT. 4 So-called saxpy operation in [21]. z x (c) row array 2-bit cntr z x output (d) internal PE structure to column array input (a) row DG (b) DG node Figure 2: (a) Dependence Graph (DG) for transformation on rows of X, with node operations; (c) the resulting row array after applying the linear space-time mapping operators along with the I/O processor-time schedule, and (d) the internal structure of processing element PE 2 . Inputs: matrices Z and G Initially Y Outputs: scalar multiplications, or one multiplication of scalar G(1; 0) by vector Z(1; :) resulting to vector which is then added to vector Y 0 (0; :). The most important aspect of this formulation is that it allows us to operate directly on the rows of intermediate matrix Z thereby exploiting the raster scan order in which Z is becoming available from the row array and, as a result, avoiding any intermediate matrix transposition. Furthermore, although Eqn. (5) is a matrix-matrix product defined over a 3-D index space, we have constructed Algorithm 3.3 as a matrix-vector product over a 2-D index space on variables (i; k) by using vector data objects instead of scalars. Similarly to Algorithm 3.2, Algorithm 3.3 can be localized by propagating variable Z(k; :) along the i-axis and accumulating output Y (i; :) along the k-axis. The DG for localized Algorithm 3.3 is shown in Fig. 3 (a), where basic DG node operation consist of M scalar multiplications and M scalar additions and data dependencies among DG nodes consist of M parallel vectors (depicted as one thick arc). Fig. 3 (b) shows the operation modeled by a DG node (saxpy). We map DG shown in Fig. 3 (a) using the same space transformation applied to the DG of Algorithm 3.2, namely resulting in the column linear array shown in Fig. 3 (c). Scheduling for this DG is accomplished by means of a function mapping a computation within DG node (i; k) to a schedule time period is an index that picks the lth element in row Z(k; :), and is the same schedule vector used for the row array. So for example, if the DG node depicted in Fig. 3 (b) is (i; computation scheduled for execution on PE 1 at at and the first computation of DG node (i; namely is scheduled on PE 2 at Note that this timing function has two resolution levels, one that assigns nodes in the DG to macro time steps, and one that assigns individual multiply-add operations within a DG node to actual time steps (clock cycles). The regularity of this scheduling allows a simple implementation with no additional local PE control. And, very importantly, the width-M data dependencies for Z(k; :) pointing upwards in the column DG of Fig. 3 (a) do not map to M physical links. Rather, only a single physical link suffices to propagate the inputs z since the sequence of z values is produced at a rate of one scalar token per schedule time period by the row array. And in addition, z tokens are consumed by the vector array at the same rate they are produced. (a) column DG (b) DG node Y (i,0)+. Y (i,1)+. Y (i,2)+. Y (i,3)+. 2-bit cntr. clock/4 y z output (d) PE structure (c) column array z from row array dependencies map to single physical link FIFO FIFO Figure 3: (a) Dependence Graph (DG) for transformation DG on columns of Z, with nodes represent scalar-vector multiply and vector-vector add; (c) the resulting column array after applying the linear space-time mapping operators along with the I/O processor-time schedule, and (d) the internal structure of processing element PE 2 . Under array 5 has identical structure as the row array in Fig. 2 (c), and is shown in Fig. 3 (c) along with its I/O schedule. The difference is in the amount of memory required by the column array for holding entire rows of Z; Y rather than single elements. In Fig. 3 (d) we show the internal PE structure, where we see that I/O registers for x and z in the row array have been replaced by length-M FIFO queues for Z and Y in the column array, respectively. 3.1 Performance Characterization We now calculate the latency and pipelining period [23] of the proposed array. Fig. 4 shows the complete I/O schedule for both arrays (row and column) on a transform of size 4, with two consecutive input matrices X 0 and X 1 . In this figure we see that both arrays are perfectly space and time matched: output stream z generated by the row array can be immediately consumed by the column array at the exact same rate of production. As a consequence no intermediate buffering is required. The problem latency (total time) for a single separable transform with input X 0 is calculated as follows: Assuming that the row array starts computation at we know that under schedule the time at which the first intermediate result is produced is i.e. at 4. From that point on, one intermediate result becomes available per schedule time period. Hence, the column array requires 28 time steps for processing the entire stream of z tokens (time between accepting z(0) and producing Y (3; 3)) , or in general MT4 time units. Since the first input to the column array becomes available at , the latency (i.e. the total time required for the parallel architecture to compute the 2-D separable transform for a single image When considering a continuous stream of problems with input matrices fX relevant measure of performance is the block pipelining period i.e. the number of time steps between initiations of two consecutive problems [23]. Referring again to Fig. 4, note first that the last data token of the first input to enter the row array, X 0 does so at at in the figure), and thus the time at which the first element of the next input X 1 (0; may enter the row array is t Next, the time at which the column array may start accepting inputs from intermediate result Z 1 (of problem instance Assuming that the row array does accept X 1 (0; 0) at the earliest possible time then units later at the first data token of the second input stream, becomes available to the column array, which is exactly t column the time at which the column array is ready to accept it. Consequently, the resulting block pipelining period is Loosely speaking we will name this second array column array, or vector array to remind us that it produces the same result Y as the column array implementing the conventional row-column approach in Figure 1 (a). Note however that in our array, Z is provided in raster scan order i.e. as a sequence of rows, not columns. . X(0,3) X(1,0) . X(1,3) X(2,0) . X(2,3) X(3,0) . X(3,3) . Y(0,3) Y(1,0) . Y(1,3) Y(2,0) . Y(2,3) Y(3,0) . Y(3,3) X0 inputs to row array . X(0,3) X(1,0) . X(1,3) X(2,0) . X(2,3) X(3,0) . X(3,3) X1 inputs to row array Intermediate results Z0 Intermediate results Z1 Column array outputs Y0 time last X0 input at 1first intermediate result of Z1 to column array at Figure 4: Overall time schedule for both arrays considering two successive inputs X 0 and X 1 , with as in the example). Since there is a single input line to the row array and thus only one token can be fed per time step the theoretically minimum fi, or maximum throughput, is attained. Efficiency is given by the ratio of the total number of computations to the product of latency times number of processors. Single input efficiency is 50%, and for a fully pipelined array is exactly 100%, implying that one output Y (i; becomes available every schedule time period. Since the performance metrics depend on the particular schedule chosen here, there are a number of available tradeoffs that can be explored. For instance, the non-systolic schedule used for the column array (with the same timing function) leads to a reduced latency of M 2 while the block pipelining period remains minimal, . However, in this case M outputs (a column of Y ) are becoming available simultaneously at every time step and should be collected in parallel (Single Input Parallel Output, or SIPO, architecture). 4 Arrays for the 2-D Discrete Fourier Transform In this Section we derive a parallel architecture for the 2-D Discrete Fourier Transform (DFT) which is a special case of the computational structures derived in Section 3. Therefore the arrays presented here are similar to those shown in Figs. 2 and 3, except that by using a DFT algorithm known as Goertzel's DFT [15] we may reduce kernel coefficient storage and simplify control complexity. 4.1 Goertzel's DFT The DFT of a 1-D, M-sample sequence x(k) is given by the set of M coefficients and the DFT of a 2-D, M \Theta M-sample sequence x(i; k) is given by Clearly, as kernel g lm (i; li M the DFT is separable on 1-D kernel g l li M . The 1-D DFT array by Beraldin et al. [15] based on the so-called Goertzel Algorithm is derived by applying Horner's polynomial multiplication rule. The following recursive equation is equivalent to the 1-D DFT where z(m; k) denotes the m th DFT coefficient after the k th recursive step, and 0 - m; Consider, for example, the computation of DFT coefficient z(2) with which, according to Eqn. (9) is given by 4 x(3). Now, by unraveling the recursion in Eqn. (11) we have that z(2; or just z(2; 4 x(0), which indeed is equal to the DFT coefficient z(2), since W \Gamma6 4 , and W \Gamma8 1. With Goertzel's algorithm, a DFT coefficient is computed using a single twiddle factor W \Gammam M , rather than M twiddle factors as with the direct application of Eqn. (9). 4.2 Modular DFT Array Derivation Let x and z be the input and intermediate result M 2 \Theta 1 vectors representing the corresponding raster scan matrices X and Z, as in Section 2. Examination of Eqn. (11) reveals that, together with the above definitions for x and z, an algorithm very similar to the localized SAF Algorithm 3.2 (see Section can be formulated, where W \Gammai M plays the role of G(i; k), and in the statement for updating z addition takes place before multiplication. Hence, a localized SAF algorithm for Eqn. (11) is given by Algorithm 4.1 Inputs x(k), W \Gammai for 6 In this equation, m is the "frequency" index, and k the "time" index. ~ ~ Outputs z(b k where ~ x(i; k); ~ g(i; z(i; are the localized equivalents of variables x(k), W \Gammai M and z(k) re- spectively. The DG for this algorithm is similar in structure to the one shown in Fig. 2 (a), except that instead of M 2 G coefficients propagated by length-M horizontal dependencies along the k- direction, we have only M complex twiddle factors propagated by length-one dependencies along the k-direction. Also, the order of internal PE operations is reversed. Algorithm 4.1 is mapped under space transformation resulting in the row array shown in Fig. 5 (a). As shown in Fig. 5 (b), the internal PE structure of the DFT row array is simpler than the array for general transforms in that a single complex twiddle factor is always used as a multiplicand. This sort of simplification can be further exploited for designing a specialized multiplier for that particular twiddle factor. For example, depending on the value of W \Gammai M , or using approximations to W \Gammai M which are factors of powers of two, a fast shift-add accumulator can be used rather than an array multiplier. The DG for the DFT column array, which is the vector version of Eqn. (11), is constructed similarly as the one for the general case shown in Fig. 3 (a), with DG node shown in Fig. 3 (b). The vector DG for Goertzel's DFT column algorithm is similar to the general one in Fig. 3 (a), with the exception that there are only M twiddle factors propagating in the k-direction, and the operations in the PEs are exchanged. In order to illustrate a different solution having better latency performance than the arrays of Section 3, we schedule the column DFT DG with function T 0 mapping computation l of DG node (i; k) at time period instead of Mapping the vector DG under (S; T 0 results in the column array shown in Fig. 5 (c), with internal PE structure shown in Fig. 5 (d). In this column array, by using schedule T 0 the need for input FIFO queues holding intermediate results Z is eliminated, resulting in total latency and memory of nearly one half that of the general version in Section 3. Schedule T 0 is also applicable to the general arrays of Section 3 and is specific neither to the DFT nor to Goertzel's formulation. However, to realize the performance gains M parallel output ports are needed in order to collect in parallel one full column of Y results every clock cycle. y z output (d) (c) column array z x output (b) (a) row array FIFO Figure 5: Parallel architecture for 2-D DFT with 4. (a) The row array and processor-time I/O schedule, (b)the internal structure of PE 2 . (c) The column array and processor-time I/O schedule, (d) the internal PE structure. 4.3 Performance Characterization The latency of the overall architecture under schedules for the row and T 0 the column arrays is M 2 +M , which is faster than the that of the general architecture developed in Section 3 (latency of 2M 2 ) by almost a factor of two, due to the broadcasting of intermediate results z(i) to all PEs in the column array and the availability of M parallel output ports. As a result, entire columns of the final result Y become available simultaneously, as shown in Fig.5 (c) at times pipelining period is identical to that of the array in Section 3, i.e. . However, the smaller latency obtained with broadcasting comes at the expense of a potentially larger duration for the each time step, i.e. a reduced maximum rate at which we can run the clock, as M increases. 5 Arrays for 2-D separable Convolution 5.1 2-D Convolution Linear 2-D convolution on M i \Theta M k input x(i; is given by 1. Notice that, in contrast to the general transform case of Eqn. (1), the summation limits in convolution range over a region that depends on the support of both functions g(\Delta; \Delta) and x(\Delta; \Delta). Since we deal here with the case where the kernel g is of smaller support than the input x, the summation limits are defined over the kernel's support region. 7 Separable filters have the property that g(i There is a substantial difference in complexity between non-separable and separable filters, the former requiring L i L k multiplications and L i additions per output sample, while the latter requiring multiplications and additions per sample. Eqn. (13) implies that input x(i; can be first processed in raster-scan order by performing convolutions on its rows, resulting in intermediate result z(i; The final result is then obtained by performing M k , 1-D column-wise convolutions on z(i; k), as k). For simplicity of exposition, but without loss of generality, from now on we assume 7 The method is not restricted to small convolution kernels, but we focus here on VLSI arrays with a relatively small number of PEs. Figure 6 (a) shows an example of the DG for row-wise convolution on raster-scan input x(i; k), 3. In this figure we have omitted the data dependence vector arrowheads indicating the data flow direction for broadcast variables x and g, and have included only those pointing upwards for the localized accumulation variables z. In contrast to the general transform case in Section 3 (see Fig. 2), in convolution there is a slight "spill-over" of outputs at row interface boundaries. In our proposed solution, we let input rows be separated by slots as they enter the array, and, as a result, there is no need to introduce any control mechanism for row interface management. Typically, the inefficiency incurred by leaving empty computations between rows in the schedule is small, in the order of 1%. 8 Let the linear space-time mapping operators tA , be given by and indexing processors and t is denoting the schedule time periods. The resulting array, along with its I/O schedule is shown in Fig. 6 (b), where every input token is made simultaneously available to all PEs. The arrows in the I/O schedule indicate the progression of partial results, that lead to the computation of output z(0; 2). No row interface control is needed as long as are padded at the end of each input row. For example, at time instants (following input x(0; 5)), two zeros are inserted into the input stream clearing the filter for the subsequent insertion of row x(1; k), Using the notation introduced in Section 2, we can write as which is convolution based on saxpy row operations. At the vector level, the DG in Fig. 7 (a) accepts row inputs produces row outputs Y (i; :) (filter coefficients g(i 0 are still scalar values). The internal DG node structure shown in Fig. 7 (b) has M k each one accepting as arguments an element from row :) and a filter coefficient g(i 0 subsequently adding the product to one y partial result that is propagating from bottom to top, leading to the computation of the corresponding element of Y (i; :) In Fig. 7 (c) we show the complete 2-D convolution pipelined architecture consisting of two linear array stages: the first, shown on the left and called the row array, for processing the DG of Fig. 6 (a), followed by another linear array stage, called the column array, for processing the vector DG of Fig. 7 (a). Each PE is 8 An alternative to spacing rows by slots is to introduce control tags carried by the last and first elements of every row, resulting in perfect row pipelining. However, the efficiency improvement of perfect row pipelining over the solution presented here is negligible. row 0 row 1 row 2 (a) (b) Figure (a) Row-wise 1-D convolution on x(i; resulting in z(i; k), with M 3. Note that rows are spaced by slots at the input. (b) array for row-wise convolution and processor-time snapshots of array operation. y(i,0)+. y(i,1)+. y(i,7)+. (a) (b) FIFO y (d) FIFO Figure 7: (a) Column 1-D convolution DG in vector form. Dark nodes represent scalar computations and thick edges represent data dependencies among vector variables with (b) The internal structure of DG node (i; i 0 (c) The parallel architecture for computing 2-D separable convolutions: the first stage linear array performs the row and the second the column processing respectively.(d) The internal row and column array PE structure. identified by two indices (p; q), where q indicates whether a PE is for row processing or for column processing The space transformation for the column array is and the timing function mapping a computation to a schedule time period is given by The choice of this timing function implies that PE i 0 ;1 can serialize the execution of the M k scalar operations modeled by DG node (i; i 0 ) in the following 2)g. The two arrays are perfectly space and time matched because this order corresponds exactly to the order in which outputs are produced from PE 0;0 of the row array (see Fig. 6). As a result, each intermediate output is broadcasted to the column array PEs via a bus as soon as it is produced and there is no need for introducing any additional control or memory to interface the two pipelined array stages. Figure 7 (d) shows the internal structure of the PEs of both arrays. During every schedule time period, PE p;q performs the following three operations: (a) multiplication of the value on the input bus by the local filter coefficient; (b) addition of the product to the incoming partial result from (c) propagation of the result to PE p\Gamma1;q . With regards to partial result accumulation in the column array, since every DG node produces outputs this is also the number of registers in the FIFO queues required between every pair of PEs (see Fig. 7 (c)). Note, however, that buses for broadcasting as well as links for moving partial results y(i; into and out of the FIFOs carry scalar quantities, so their bit-width should be only one word (and does not depend on M k or L). 5.2 Performance Characterization: Latency and Memory Total computation time for a single input problem instance x(i; k) is as follows. The row array computation time is T 9 The column array computation time is Since the communication between both structures is perfectly pipelined, there is only a delay of one time unit in producing the first output y(0; 0). As a result, total computation time is T vector + 1. As an example, consider typical numbers for image processing, where M which result in 96.6% efficiency (loss due to imperfect row pipelining), while with efficiency is 98.3%. 11 This relatively small efficiency loss is the result of having arrays with virtually no control circuitry. Alternatively, this efficiency loss rows, each requiring processing time of Mk +Lk \Gamma 1. The insertion of zeros between rows is the cause of the steps. the total number of computations is and since there are 2L processors the efficiency is given by could be reduced by introducing control at the PE-level for managing the interfaces between rows in the row array (as opposed to insertion of zeros). The block pipelining period on an input sequence fX implying that a complete output y(i; k) is available every fi time steps, substantially improving efficiency. Efficiency for the case when M 1024 and Next we prove a Proposition on the lower bound on memory of any architecture that processes inputs under raster scan order, and show that the 2-D convolution array proposed here nearly attains this lower bound. This Proposition shows that this array does not require larger memory storage than the conventional transposition memory solutions, a result holding also for the structures presented in Sections 3 and 4. In the following I/O model, we draw a boundary around the circuit under consideration, and assume that the input stream x(i; k) crosses this boundary in raster scan, or row-major order. 12 Once an input has entered the circuit, it is held in local storage until used in some computation for the last time, and then it is discarded. It is further assumed that once input x(i; k) has appeared in the input stream, the circuit cannot request x(i; k) from the host again, and has to manage it throughout its lifetime. (This analysis is similar to that in [30].) Proposition 5.1 The lower bound on memory for 2-D convolution by separable L-tap and by non-separable filters with row-scan input is Proof: Consider output y(i; k). By Eqn. (12) (non-separable) or Eqn. (13) (separable) we know that computation of y(i; requires the set of L 2 input elements L+1)g. Under the row-scan order imposed by Def. 2.1, the last element in this set to enter the circuit boundary is x(i; k). The minimum time x(i; spends inside the boundary can be calculated by looking for the last output element that requires x(i; its computation, which can be shown to be y(i 1). Making the assumption that the circuit is operating under perfect pipelining, and hence that one output element is available every single time period, the minimum time needs to spend inside the boundary is are rows, each with Finally, during time t, at least (M k new input elements entered the boundary, and therefore minimum memory is lower bounded by t storage elements. 2 Total memory requirements for the proposed convolution array of Fig. 7 (c) are dominated by the column FIFOs, accounting for a total of (M k memory registers, and meeting exactly the lower memory bound in Proposition 5.1. In addition, there is a total of 2L registers for locally storing filter coefficients (which could be reduced to L by having corresponding PEs in 12 The derivation also holds if we substitute column-major for row-major order, as long as we are consistent. each array share their coefficient), and L registers for the accumulation of z(i; results in the row array. The array proposed in Section 4 also meets this lower memory bound. In the case of a global transform (i.e. DFT) with equal support regions for kernel and input, Proposition 5.1 reduces to the trivial case where storage for an entire M \Theta M frame is required. 6 Conclusions In this paper we have shown that the basis for constructing modular architectures for 2-D separable transforms is an array for the 1-D transform, and that the resulting overall architecture can reflect the modularity of that array. The only restriction required by the 1-D transform array is that it accepts input and generates output in raster scan order. Then a regular architecture consisting of two fully pipelined linear array stages with no shared memory can be derived, based on the key idea that the DG for the first linear array (performing the "row" processing) is structurally equivalent to the DG of the second linear array (for the "column" processing), but with the difference that the latter DG operates at a coarser level of granularity (vector instead of scalar computations). This idea can be exploited even for separable transforms having a complex computation structure, since, in principle, avoiding transposition and shared memory does not depend on the 1-D transform structural details. Rather, it depends only on (i) there being a raster scan order for the 1-D transform and (ii) a change in abstraction level from scalar to vector operations. The 2-D DFT and the 2-D convolution have been used as case studies to demonstrate the generality and effectiveness of the proposed architecture. We further conjecture that the same method can be used to derive pipelined array structures for multi-dimensional separable transforms of any dimension. We have also applied this methodology to the Discrete Wavelet Transform, a time-scale transform that represents finite energy signals with a set of basis functions that are shifts and translates of a single prototype mother wavelet. Modular arrays for computing the 1-D DWT have been proposed in the literature [24, 25, 26], but 2-D solutions usually rely on transposition, or some form of complex data routing mechanism that destroys modularity. Since most of the current applications and research on wavelets is based on separable FIR wavelet filters, we have been able to apply the results from Section 5 to the 1-D arrays presented in [24, 25]. All Dependence Graphs and space-time transformation pairs derived here, can be supplied as input to the rapid prototyping tool DG2VHDL developed in our group [31, 32] that can generate automatically high quality Hardware Description Language models for the individual PEs and the overall processor array architectures. The generated VHDL models can be used to perform behavioral simulation of the array and then synthesized (using industrial strength behavioral compilers, such as Synopsys [35]) to rapidly generate hardware implementations of the algorithm targeting a variety of devices, including the increasingly popular Field Programmable Gate Arrays [33, 34], or the Application Specific Integrated Circuits (ASICS). Using the ideas presented here, modular linear array cores for the 2-D Discrete Cosine Transform (DCT) which supporting a variety of different I/O patterns have been synthesized and compared in [36]. --R "Multiresolution Signal Decomposition" "Discrete Cosine Transform: Algorithms, Advantages, and Applica- tions" "An Novel CORDIC-based Array Architecture for the Multidimensional Discrete Hartley Transform" "A Cost-Effective Architecture for 8x8 Two-Dimensional DCT/IDCT Using Direct Method" "Image and Video Compression Standars" "VLSI Architectures for Multidimensional Fourier Transform Processing" "VLSI Architectures for Multidimensional Transforms" "Low-Power Video Encoder/Decoder Using Wavelet/TSVQ With Conditional Replenishment" "Multidimensional Systolic Arrays for Multidimensional DFTs" "Highly Parallel VLSI Architectures for the 2-D DCT and IDCT Computations" "Synchronizing Large Systolic Arrays" "Synthesizing Linear Array Algorithms from Nested For Loop Algo- rithms" "VLSI Signal Processing Systems" "Fourier Transforms in VLSI" "Efficient One-Dimensional Systolic Array Realization of the Discrete Fourier Transform" "Multidimensional Digital Signal Processing" "Design of 2-D Recursive Filters with Separable Denominator Transfer Functions" "The Design of Multistage Separable Planar Filters" "Parallel Algorithms and Architectures for Discrete Wavelet Transforms" "Fundamentals of Digital Image Processing" "Matrix Computations" "Algorithms for Discrete Fourier Transform and Convolution" VLSI Array Processors. "On the Synthesis of Regular VLSI Architectures for the 1-D Discrete Wavelet Transform" "Distributed Memory and Control VLSI Architectures for the 1-D Discrete Wavelet Transform" "1-D Discrete Wavelet Transform: Data Dependence Analysis and Synthesis of Distributed Memory and Control Array Architectures" "On Time Mapping of Uniform Dependence Algorithms into Lower Dimensional Processor Arrays" "Systolic Array Implemenation of Nested Loop Pro- grams" "Synthesizing Systolic Arrays with Control Signals from Recurrence Equa- tions" "Computational Aspects of VLSI" "DG2VHDL: A tool to facilitate the high level synthesis of parallel processing array architectures" "DG2VHDL: a tool to facilitate the synthesis of parallel VLSI architectures" "Using DG2VHDL to Synthesize an FPGA Implementation of the 1-D Discrete Wavelet Transform" Synthesis of Array Architectures of Block Matching Motion Estimation: Design Exploration using the tool DG2VHDL. Behavioral Synthesis. Design and Synthesis of maximum throughput parallel array architectures for real-time image transforms --TR VLSI array processors VLSI Architectures for multidimensional fourier transform processing Synthesizing Linear Array Algorithms from Nested FOR Loop Algorithms Fundamentals of digital image processing Discrete cosine transform: algorithms, advantages, applications VLSI Architectures for Multidimensional Transforms Behavioral synthesis VLSI Signal Processing Systems Image and Video Compression Standards Multiresolution Signal Decomposition Multidimensional Digital Signal Processing On Time Mapping of Uniform Dependence Algorithms into Lower Dimensional Processor Arrays Synchronizing large VLSI processor arrays Parallel architectures and algorithms for discrete wavelet transforms

VLSI architectures;parallel processing;2-D separable transforms

509251

A High Speed VLSI Architecture for Handwriting Recognition.

This article presents PAPRICA-3, a VLSI-oriented architecture for real-time processing of images and its implementation on HACRE, a high-speed, cascadable, 32-processors VLSI slice. The architecture is based on an array of programmable processing elements with the instruction set tailored to image processing, mathematical morphology, and neural networks emulation. Dedicated hardware features allow simultaneous image acquisition, processing, neural network emulation, and a straightforward interface with a hosting PC.HACRE has been fabricated and successfully tested at a clock frequency of 50 MHz. A board hosting up to four chips and providing a 33 MHz PCI interface has been manufactured and used to build BEATR IX, a system for the recognition of handwritten check amounts, by integrating image processing and neural network algorithms (on the board) with context analysis techniques (on the hosting PC).

Introduction Handwriting recognition [1, 2, 3] is a major issue in a wide range of application areas, including mailing address interpretation, document analysis, signature verification and, in particular, bank check processing. Handwritten text recognition has to deal with many problems such as the apparent similarity of some characters with each other, the unlimited variety of writing styles and habits of different writers, and also the high variability of character shapes issued by the same writer over time. Furthermore, the relatively low quality of the text image, the unavoidable This work has been partially supported by the EEC MEPI initiative DIM-103 Handwritten Character Recognition for Banking Documents. F. Gregoretti et al.: A High Speed VLSI Architecture for Handwriting Recognition presence of background noise and various kinds of distortions (for instance, poorly written, degraded, or overlapping characters) can make the recognition process even more difficult. The amount of computations required for a reliable recognition of handwritten text is therefore very high, and real-time constraints can only be satisfied either by using very powerful and expensive processors, or by developing ad-hoc VLSI devices. To cope with the tight cost and size constraints we had, we decided to develop HACRE, a massively parallel VLSI image processor with the instruction set dedicated to execute both traditional image processing (such as filtering and image enhancement) and mathematical morphology [5] (such as opening, closing, skeletonization) and several types of neuro-fuzzy networks [6] (such as perceptrons, self-organizing maps, cellular networks, fuzzy systems). We have tailored the recognition algorithms to the architecture capabilities. HACRE is based on PAPRICA-3, a dedicated architecture deriving from enhancement of previous works [8, 9, 10]. It is designed to be used both in a minimum-size configuration (consisting of 1 chip, 1 external RAM, plus some glue logic for microprocessor interfacing) and in larger-size configurations (consisting of as many cascaded chips as required, as many external RAM's, some additional global interconnection logic and a host interface, typically a PCI). A configurable PCI board hosting up to four HACRE chips, together with the required RAM chips (for image and program memories), camera and monitor interfaces, controllers, PCI interface and all the required logic, has been designed, manufactured and successfully tested. One such board, populated with two HACRE chips, plugged into a hosting PC, has been used to build and test BEATRIX, a complete system for high-speed recognition of the amount on banking checks, which mixes image processing algorithms, neural networks and a context analysis subsystem [3]. The VLSI device and the overall system (board and PC interface) have been designed in parallel with the development of the algorithms (BEATRIX), so that the hardware and software designs have influenced each other very much. That resulted into an efficient though flexible implementation for a wide class of applications. Section 2 describes the driving application, while Sections 3 and 4 describe the PAPRICA- 3 architecture and the HACRE chip, respectively. Section 5 describes the PCI board, while submitted to Journal of VLSI Signal Processing 3 Section 6 briefly describes the BEATRIX check recognizer. At the end, Section 7 gives the measured performance of the system and compares them with those of a commercial PC. Driving Application The aim of our work was to recognize real-world checks, where handwriting is assumed to be unboxed and usually unsegmented, so that characters in a word may touch or even overlap. The amount is written twice: the legal amount (namely, the literal one), and the courtesy amount (namely, the numerical one). The two fields are placed in well-known areas of the check, and an approximate localization of these two areas can be obtained from the information contained in the code-line printed at the bottom of the check. 2.1 Application Requirements Our aim was to cope with the tight cost and performance requirements which might make our system commercially relevant. From a preliminary system analysis we pointed out the following requirements: average processing speed of at least 3 checks per second (5 per second, peak), with a rejection rate significantly lower than 5% and an error rate approaching zero. This corresponds to a sustained recognition speed of about 500 characters per second (both digits and alphanumeric) and a character accuracy rate in excess of 99%. Cost requirements imposed a higher bound of 50 US$ per chip with a power dissipation of at most 1 W per chip. In addition, one particular application in a low-speed but low-cost system was claiming at a single-chip embedded solution. As far as the host processor is concerned, we selected a commercial PC with a standard PCI bus. The operating system could either be Microsoft Windows or Linux. The choice of such a general-purpose hybrid architecture (a PAPRICA-3 system tightly interconnected to a host PC) was driven by the observation that in many image processing systems the first processing steps (at the bitmap level) mostly require simple operations on a large amount of data, whereas, as processing proceeds, less and less data requires more and more 4 F. Gregoretti et al.: A High Speed VLSI Architecture for Handwriting Recognition complex operations. In our approach, PAPRICA-3 is tailored to a fast and repetitive processing of elementary pixels of an image while the PC is best used for more complex and more symbolic- based operations on a much reduced amount of data. We feel that such a cooperation between the two systems can provide the best cost-performance ratio. 2.2 Chip Requirements The HACRE chip has been designed bearing in mind the following application-dependent con- straints/requirements: ffl cascadability: HACRE implements a slice of Processing Elements (PE's) (all that could fit in 100 mm 2 ), but additional chips can be cascaded and connected to as many RAM chips; ffl neural networks mapping: one PE per each neuron, to have the highest efficiency; ffl a simple 1-bit Processor Element; operations with higher resolution can be computed by means of a bit-serial approach; this provided the best complexity/flexibility/performance trade-off among all the architectures which were analyzed; ffl provisions for both image processing, mathematical morphology [5], neural networks [6] emulations, image acquisition, etc.; ffl easy interface with a host processor (possibly a PC), to improve overall ffl the highest degree of programmability, while keeping the hardware complexity at a low level; this has been achieved by letting the host PC (instead of HACRE) perform a number of operations which normally occur at a lower rate; ffl both "local" and "global" instructions; the former are used to implement a sort of pixel neighborhood, while the latter are available to compute global operators, such as summa- tions, maxima, minima, winner-takes-all, etc.; ffl provisions for simple handling of external look-up tables (external RAM's or ROM's); submitted to Journal of VLSI Signal Processing 5 Internal Registers from camera 8-bit serial input serial output to monitor k lines of binary image plane 0 WCS UNIT CONTROL Processing Elements GLOBAL COMMUNICATION PEQ Figure 1: General architecture of the Processor Array. ffl a set of external "status registers" where HACRE can accumulate neuron outputs and which can be read (or written into) in parallel by the host PC; ffl a set of direct binary I/O channels (6+6) through which HACRE can either interrupt the host PC or activate stepper motors, CCD cameras, etc. 3 Architecture Description PAPRICA-3 is the latest component of a family of massively parallel processor architectures [8, 9, 10] designed at the Politecnico di Torino. As shown in Figure 1, the kernel of PAPRICA-3 is composed of a linear array of identical Processing Elements (PEs) connected to a memory via a bidirectional bus. The memory stores the image and the intermediate results of the computation and is organized in words (each one associated to an address) whose length matches that of the array of processors. Each word in the memory contains information relative to one binary pixel plane (also called layer) of one line of an image. Because the width of the bus that connects the array and the memory is the same as the number of processing elements (and therefore it is the same as the word length), a single memory cycle is required to load or store an entire line of the image 6 F. Gregoretti et al.: A High Speed VLSI Architecture for Handwriting Recognition to/from the PE's internal registers. A Control Unit executes the program stored in the instruction memory (called the Writable Control Store) and generates the signals that control the operations of the processing elements, the image memory, and of the other architectural elements that will be described below. The typical flow of operation consists of first transferring one line of the image from the memory to the array, then processing the data, and finally storing back the results into memory according to a LOAD - EXECUTE - STORE processing paradigm typical of RISC processors. The same cycle is repeated for each line until the entire image has been processed. When executing a program, the correspondence between the line number, the pixel plane of a given image, and the absolute memory address is computed by means of data structures called Image Descriptors. An Image Descriptor is a memory pointer that consists of two parts: a base address that usually represents the first line of the image, and two line counters which can be reset or increased by a specified amount during the execution of the program, which are used as indices to scan different portions of the image. The instruction set includes several ways to modify the sequential flow of control. Branches can be taken unconditionally or on the basis of conditions drawn over the control unit internal registers. In addition, any instruction can be prefixed by an enabling condition. One register in the control unit is dedicated to the implementation of hardware loops: given the iterative nature of the algorithms employed in image processing, this features greatly enhances the performance of the architecture. Two additional conventional counters can be used as indices in the outer loops; instructions are provided to preset, increase and test their value. 3.1 Processing Elements Each PE is composed of a Register File and a 1-bit Execution Unit, and processes one pixel of each line. The core of the instruction set is based on morphological operators [5]: the result of an operation depends, for each processor, on the value assumed by the pixels of a given neighborhood, which in the case of PAPRICA-3 is a reduced 5 \Theta 5 box, as sketched by the grey squares in Figure 1. The morphological function can be selected by changing the value of a template which encodes for each pixel of the neighborhood the kind of required boolean submitted to Journal of VLSI Signal Processing 7 combination. The instruction set includes also logical and algebraic operations (AND, OR, NOT, EXOR, etc.), which can be used either to match input patterns against predefined templates, or to compute algebraic operations such as sums, differences, multiplications, etc. As PE's are 1-bit computing elements, all algebraic operations have to be computed using a bit-serial approach. For each Processor Element, the value of the pixels located in the East and West directions (either 1 or 2 pixels away) is obtained by a direct connection to the neighboring PE's, while the value of the pixels in the North and South directions corresponds to that of previously or to be processed lines. To obtain the outlined neighborhood in the chip implementation, a number of internal registers per each PE, at present), called Morphological Registers (MOR), have a structure which is more complex than that of a simple memory cell, and are actually composed of five 1-bit cells with a S!N shift register connection. When a load operation from memory is performed, all data are shifted northwards by one position and the south-most position is taken by the new line from memory. In this way, data from a 5 \Theta 5 neighborhood are available inside the array for each PE, at the expense of a two-line latency. A second set of registers (48 per each PE, at present), called Logical Registers (LOR), is only 1-bit wide and is used for logical and algebraic operations only. 3.2 Video Interface An important characteristic of the system is the integration of a serial-to-parallel I/O device, called InterFace (VIF), which can be connected to a linear CCD array for direct image input (and optionally to a monitor, for direct image output). The interface is composed of two 8-bit shift registers which serially and asynchronously load/store a new line of the input/output image during the processing of the previous/next line. Two instructions activate the bidirectional transfer between the PE's internal registers and the VIF, ensuring also proper synchronization with the CCD and the monitor. 8 F. Gregoretti et al.: A High Speed VLSI Architecture for Handwriting Recognition R R R R R R R R Count Ones RESET Status Word Count R R R R a) Processor Figure 2: Inter-processor Communication Mechanisms: a) Status Evaluation Network b) Inter-processor Communication Network 3.3 Inter-processor communication Two inter-processor communication mechanisms are available to exchange information among PE's which are not directly connected. The first mechanism consists of a network (called Status Evaluation Network), shown in Figure 2a which spans the extent of the array; each processor sends the 1-bit content of one of its registers and the global network provides a Status Word which summarizes the status of the array. The word is divided into two fields: the first field is composed of two global flags, named SET and RESET, which are true when the contents of the specified registers are all '1's or all '0's, respectively; the second field is the COUNT field which is set equal to the number of processing elements in which the content of the specified register is '1'. This inter-processor communication mechanism can be used to compute global functions such as maxima, minima (e.g., for emulation of fuzzy systems), logical OR and AND of boolean neurons, neighborhood communications, neuron summations in perceptrons, external look-up tables, winner-takes-all, seed propagation algorithms, and many others. This global information may also be accumulated and stored in an external Status Register File and used for further processing, or to conditionally modify program flow using the mechanism of prefix conditions. Status Registers can also be read by the host processor. For instance, Status Registers have been used in the example of Section 2 to implement a neural network by computing the degree of matching between an image and a set of templates (weight and center submitted to Journal of VLSI Signal Processing 9 matrices). The second communication mechanism is an Inter-processor Communication Network, shown in Figure 2b, which allows global and multiple communications among clusters of PE's. The topology of the communication network may be varied at run-time: each PE controls a switch that enables or disables the connection with one of its adjacent processors. The PE's may thus be dynamically grouped into clusters, and each PE can broadcast a register value to the whole cluster with a single instruction. This feature can be very useful in algorithms involving seed- propagation techniques, in the emulation of pyramidal (hierarchical) processing and for cellular neural networks or for local communication (short range neighborhood). 3.4 Host interface A Host Interface allows the host processor to access the WCS and a few internal configuration registers when HACRE is in STOP mode. The access is through a conventional 32-bit data bus with associated address and control lines. The same lines are controlled by HACRE in RUN mode and used to access the private external Image Memory. Some additional control and status bits are used to exchange information with the host processor: these include a START input line and a RUNNING output line, plus another six input and six output lines called Host Communication Channels (HCC's). HCC input lines can be tested during program execution to modify program flow, while HCC output lines can be used as flags to signal certain conditions to the host processor (for instance, interrupts). The kernel of the hardware implementation of the PAPRICA-3 system is the HACRE integrated circuit whose block diagram is shown in Figure 3. The main components are : ffl The Processor Array (PA) which is composed of processing elements; the internal architecture and the main features of the PA and of the PE's are described in detail in Section 4.1. F. Gregoretti et al.: A High Speed VLSI Architecture for Handwriting Recognition WCS REGISTER TO EXTERNAL external memory to the host and CONTROL UNIT ICN Data Address TO EXTERNAL IMAGE MEMORY Descriptors Image Figure 3: Block diagram of the chip. The PA has a direct and fast parallel communication channel towards the Image Memory (IM) which in the current implementation is external to the chip. The decision to keep the IM outside of the chip has been a very critical issue. The direct and fast access to a large internal memory would have allowed to execute at each clock cycle a LOAD or STORE operation with the same timing as other instructions and with a high processing throughput but, on the other hand, the cost of implementing a large memory with a standard CMOS technology would have been too high. In fact architectures such as the IMAP system [15], which have large on chip data memories, employ dedicated, memory oriented fabrication processes. Therefore we decided to implement the IM external to the chip and to reduce the processing overhead by a heavy pipeline of internal operations which may allow the partial overlap of a memory access with the execution of subsequent instructions. ffl The Control Unit (CU) which executes the instructions by dispatching to the PA the appropriate control signals. The choice of implementing the CU on the same chip as the PA is a key characteristic of the PAPRICA-3 architecture with respect to other mono- and bi-dimensional SIMD machines [11, 12, 14, 15, 16] in which the controller is a unit which is physically distinct form the array. In this case the maximum clock speed is limited to a few tens of MHz by the propagation delay of the control signal from the controller to the different processing units. This may be a non critical limit in systems with a large submitted to Journal of VLSI Signal Processing 11 number of PE's, but, since our application was aimed at real time embedded systems, we preferred to push to the limit the performances even of single-chip systems by integrating the CU with the array. This means that with multiple-chip systems the CU is replicated in each chip with an obvious loss in silicon area, but in our case it has been considered a little price to pay, with respect to the possible increase in performance since our design goal was an operating frequency of 100 MHz. Section 4.2 will analyze in detail the design choices of the CU and its internal architecture. ffl The Writable Control Store (WCS, 1K words \Theta 32 bits), in which a block of instructions is stored and from which the CU fetches instructions. The choice of a WCS is a compromise between different constraints. First the goal of executing one instruction per cycle required the Instruction Memory to reside on the same integrated circuit as the CU. Since the amount of memory which may be placed on a chip with a standard CMOS technology is limited the optimal solution would have been a fast associative cache. A preliminary feasibility study showed that the cache approach would have been too expensive in terms of silicon area and too slow to match the target 10 ns cycle time. Considering that most image processing algorithms consist of sequences of low level steps, such as filters, convolutions, etc., to be performed line by line over the whole image this means that the same block of instructions has to be repeated many times. Hence we chose to pre-load each block of instructions into the WCS, a fast static memory, and to fetch instructions from there. After the instructions of one block have been executed for all the lines of an image, a new block of instructions is loaded in the WCS. If the ratio between the loading time and the processing time is small then the performance of a fast cache with a hit ratio close to 1 can be obtained, at a fraction of the cost and complexity. ffl The ICN and SEN communication networks. The former is fully distributed in the PE's; the latter is composed of two parts: the first part is distributed and is the collection of EVAL units (see Section 4.1) integrated in the PE's, while the second one is centralized and is composed of one Count Ones unit and two 32-AND units for the evaluation of the SET and RESET flags. 12 F. Gregoretti et al.: A High Speed VLSI Architecture for Handwriting Recognition Figure 4: Microphotograph of the chip ffl The Host interface which allows a host processor to access the WCS, and a few internal configuration registers. The access is through a conventional 32-bit data bus with associated address and control lines. In addition the interface handles the external protocol for the communication between the PA and the external Image Memory. A microphotograph of the complete chip is shown in Figure 4. The chip has been implemented in a 0:8 -m, single poly, dual metal CMOS technology and has a total area of 99 mm 2 . Multiple chips may be cascaded to build systems with a larger number of processing elements, as explained in detail in Section 5. submitted to Journal of VLSI Signal Processing 13 EXECUTION UNIT UNIT MATCH UNIT COMM UNIT UNIT REGISTER from previous the load from camera store video to the next to MOP/LOP result from Image Memory match result operands to ICN from and to from neighbours to neighbours Figure 5: Block diagram of the processing element. 4.1 Processing Array The Processing Array has been implemented using a full custom design style in order to take advantage of its regular structure and optimize its area. In fact each Processing Element is a layout slice and, since all PE's operate according to a SIMD paradigm, the control signals are common to all of them and may be broadcast by simple abutment of PE cells. Unlike the block diagram of Figure 3 where the array is shown as a single entity, in the implementation the array is divided into two 16-PE sub-blocks which are clearly visible at the top and at the bottom of the photograph of Figure 4. In this way the capacitive load and the peak current on the control lines is reduced and the delay is optimized; in addition, the 16 PE's block has a better aspect ratio and is more easily routed by automatic tools. The block diagram of a PE is shown in Figure 5. Its main components are the Register File (RF) and the Execution Unit (EU). The RF, introduced in Section 3.1, is composed of two sections, corresponding to the 48 LOR registers and to the 16 MOR registers. Address decoding is centralized for each block in order to optimize its area and the decoded selection lines are run over the whole block. LOR's are implemented as 3-port static RAM cells, allowing the simultaneous execution of 3 operations (2 read and 1 write) on the RF at any given time. 14 F. Gregoretti et al.: A High Speed VLSI Architecture for Handwriting Recognition Memory or from Image Cell Central N NN Cell Cell Cell Cell from Address Decoder To EXECUTION UNIT Figure Structure of a MOR register Each MOR register is composed of five 1-bit RAM cells, as shown in Figure 6. The central cell is similar to a LOR cell with 3 ports, while the other four are static RAM cells with a single read port. In addition all cells have an input port (SH) which allows data to be shifted from each cell to its right neighbor in a single clock cycle. When executing a LOAD operation between the Image Memory or the VIF and a LOR register, the value of the 1-bit pixel is simply transferred in the register through one of the ports. When the same operation is performed on a MOR, the value from the memory (or the VIF) is loaded by the SH port into the leftmost cell and the contents of all the others is shifted one position to the right. In this way the central cell contains the current pixel of the image, the right cells the north neighbors with distance 1 and 2 and the left cells the south neighbors. When a read operation is executed on port 1 of a MOR, the values of all cells are sent to the EU for the execution of neighborhood based operations. The execution time of a read/write operation from the RF is in the worst case lower than 10ns and the data is latched at the output by the global system clock. The register file of each PE integrates also one stage of the VIF. From a global point of view the VIF is a shift register with a number of stages equal to the number of PE's and with a 16 bit word width. It is divided in an Input and an Output section, each 8-bit wide, which are connected respectively to a pixel serial input and output device and may be clocked independently. The operations of the VIF and of the Processor Array are independent and may proceed in parallel, overlapping I/O and processing. However they synchronize with each other when one of the two following events takes place: submitted to Journal of VLSI Signal Processing 15 ffl A full line of input data has been shifted in the Input section and the CU has fetched a Load From Camera instruction. In this case in each PE the 8 bits of the Input section of the local VIF stage are transferred to the southmost position of the first 8 MOR registers. ffl A full line of data has been shifted out of the output section of the VIF and a Store To Video instruction has been fetched by the CU. When it happens the value of the first 8 LOR registers are transferred in parallel into the output section of the local stage of the VIF. In order to minimize the transfer time and the interconnection area required, the Input and Output sections of the local VIF stage have been implemented directly inside the RF, close to the corresponding LOR and MOR registers. The EU performs the operations corresponding to the instructions of the program on the data from the RF, under the control of the CU. It is composed of 4 blocks: ffl The LOP unit which is responsible for logical and arithmetic operations. It has been synthesized using a standard cell approach and the final layout has been routed manually to optimize the area. ffl The MATCH unit which is responsible for MATCH operations, the base of all morphological and low level image processing operation. Its basic function is a comparison between a given neighborhood of the pixel and a vector of 3-valued templates (0, 1 and don't care) provided by the CU. The neighborhood of the pixel is obtained from the contents of a MOR register of the PE and from the four neighboring PE's. In order to reduce the silicon area and execute the operation in a single clock cycle the unit has been implemented with custom dynamic logic. ffl The COMM unit which implements in standard cell one section of the ICN communication network depicted in Figure 2b. ffl The EVAL unit which, when the corresponding instruction is executed, takes the contents of one register of the PE, masks it with the contents of another register and sends the F. Gregoretti et al.: A High Speed VLSI Architecture for Handwriting Recognition Figure 7: Layout of a processing element. results to the centralized part of the SEN for the evaluation of the different fields of the Status Word. As explained in the next section, the Control Unit is able to concurrently execute more than one instruction which may in turn activate different functional units in the PE's. Hence, in order to obtain a correct execution, a data pipeline that reflects that of the CU had to be put in place to separate the units in the data path. Figure 7 shows the layout of one PE which occupies approximately 1 mm 2 of silicon area. As clearly visible, all functional units have the same vertical dimension (horizontal in fig. 4). Control signals run in the vertical direction across the PE which may be connected by abutment to its neighbors. The layout also shows how tightly the register file and the VIF I/O structure are integrated in order to obtain a high throughput in I/O operations. 4.2 The Control Unit As already mentioned, PAPRICA-3 exploits both spatial parallelism due to its massively parallel architecture, and instruction level parallelism due to the pipelined design of the control unit. Because of the very different nature of the instructions executed by the array and those executed directly by the control unit, the pipeline had to be designed with particular care both in its architecture and its implementation in order to obtain the best trade-off between complexity and performance. In HACRE the Control Unit is located on each chip of a multi-chip implementation together with the WCS, thus decentralizing the issue of control signals. The WCS is loaded only once every algorithm while each instruction is usually executed many times (often thousands of times) submitted to Journal of VLSI Signal Processing 17 per image. In most cases the overhead is thus reduced by orders of magnitude. The drawback of a similar implementation is the waste of chip area required for the duplicated logic: while this is still significant in the technology used for the design (0:8-m), we anticipate a lower impact in deep sub-micron technologies where the interconnections, and not the logic, play the major role. The main drive in the design of the pipeline is to obtain the highest performance for those sequences of instructions that are most used in image processing algorithms; these include, among others, sequences of morphological operations, loops on the entire image, and bit-serial compu- tations. Although the relative frequency of global and scalar instructions is far less than that of array instructions, their sometimes long latency due to their global nature may impose severe constraints to the execution of other instruction, negatively impacting the overall performance; neglecting their execution would therefore certainly lead to a sub-optimal design. For these reasons the pipeline has been logically divided into two sections which have been partially collapsed in the implementation to better exploit the hardware resources. The first section deals with scalar and global operations and has been designed as a conventional mono- functional queue; the second section controls the execution of array operations and employs a more complex multi-functional pipeline (which increases instruction level parallelism) and a more sophisticated conflict resolution mechanism. Despite the implementation as a single continuous structure, the two sections are designed to operate independently and they synchronize by means of mutual exclusion on shared resources during the execution of instructions that involve both array and global or scalar operations. 4.2.1 Pipeline implementation Figure 8 shows the first section of the pipeline together with the supporting registers. Dotted lines represent data flow, while block lines represent the flow of the instruction codes. The first stage of the pipeline (IF) is responsible for loading the instructions from the program memory. In order to achieve a high clock rate in spite of a relatively slow memory, instructions are fetched two at a time (technique known as double fetch). The entire mechanism is handled by this stage, so that to the rest of the pipeline it appears as if the memory were able F. Gregoretti et al.: A High Speed VLSI Architecture for Handwriting Recognition M/N Decod. IF IM Status Word ID Stop Register Sum From Status Registers To host system From ext. connection Image Desc. To the CD stage, second section Progr. Mem. Figure 8: First section of the pipeline. to deliver one instruction per clock cycle. The drawback of this technique is an increased penalty due to control conflicts that require to invalidate the initial stages of the pipeline, for example when a branch is taken (this problem is typical of superpipelined architectures where the queue is very deep). If the loop control in bit-serial and morphological computations were subject to this problem, the benefits obtained by the higher clock rate achievable by the double fetch technique would certainly be offset by the increased penalty. For this reason, the stage J of the pipeline is dedicated to the support of a hardware loop mechanism that controls the program counter and minimizes the negative effects. When the number of instructions in the loop is even (or if there is only one instruction), the pipeline is able to deliver one instruction per clock cycle with no control overhead. In all other cases the penalty is just one clock cycle per iteration. In the ID stage the scalar instructions get executed and leave the pipeline, while the array instruction are dispatched to the AR queue that computes the effective addresses for the internal registers and the image memory. Figure 9 shows the second section of the pipeline. The stage CD is dedicated to the conflict resolution as will be explained later. The rest of the pipeline is a multi-functional queue, where each branch of the queue is dedicated to a particular feature of the array. In particular SE controls the Status Evaluation; CO the inter-processor communication mechanism; the submitted to Journal of VLSI Signal Processing 19 CD OW CO MU SM SS From AR3 acc. Figure 9: Second section of the pipeline. sequence OR-BP-EX-OW executes the array operations by first reading the internal registers, then propagating the data to the neighborhood, then computing the result of the operation and finally writing back the result in the internal registers; MU handles the access to the image memory; and finally SL-LV and SS-SV is the path followed by the instructions that control the VIF. Because of the multi-functional nature of the pipeline and the presence of different execution stages (including ID for the scalar instructions), it is possible that instructions be executed out of order (although the architecture still issues no more than one instruction per clock cycle). This feature enhances temporal parallelism and therefore performance. In addition, as mentioned earlier, the first and the second sections of the pipeline are decoupled, so that a stall in the second doesn't halt the execution of scalar instructions in the first. By doing this, the impact of the sequential part of the algorithm is minimized. 4.2.2 Conflict Resolution The concurrent execution of instructions gives rise to conflicts in the pipeline that must be resolved in the most efficient way. In this section we will present the mechanism employed to resolve conflicts involving the internal registers. In a multi-functional queue there are essentially two ways to detect a conflict condition: F. Gregoretti et al.: A High Speed VLSI Architecture for Handwriting Recognition Scoreboard method: In the scoreboard method, all conflicts are resolved in a single pipeline stage. A table, called the scoreboard, contains an entry for each resource in the architecture (e.g. registers). Every instruction, before entering the execution stages, verifies the availability of the source and destination operands, and once the permission is obtained, it gets possession of them by registering with the scoreboard. The resources are released only after the execution is complete. Tomasulo algorithm: In the Tomasulo algorithm the conflict resolution is distributed in the pipeline. For example, the stage responsible for reading the operands checks their avail- ability, as does the stage that writes them. Since the resolution is distributed, resources are reserved only when they are actually needed, so that an instruction is allowed to proceed even if only part of the required operands are available. The advantage of the scoreboard method is a much simpler implementation, which is traded-off with the higher throughput achievable with the Tomasulo algorithm. HACRE adopts a combination of the two approaches. In a traditional implementation the scoreboard is basically a map where a set of flags signals the availability of a particular register for reading or for writing. Given the number of registers in our architecture, the distributed access to this map that would be used in the Tomasulo algorithm results in a big and slow implementation. Our solution is to build a map that does not represent the registers, but rather the interaction among the instructions in the different stages of the pipeline. Once in the CD stage, each instruction compares the effective addresses of its operands to those of the instructions in the following stages of the pipeline (in the picture, those to the right); in doing so, the instruction builds a map where each stage of the pipeline is marked with a flag: an active flag in the map means that there is a potential conflict between the instruction which owns the map and the instruction in the stage of the pipeline corresponding to the flag. The conflict is only potential because the instruction that builds the map has yet to be routed to the stages where the operands are needed: at that later time, the instruction that gives rise to the conflict might have already left the pipeline. submitted to Journal of VLSI Signal Processing 21 When our instruction leaves the CD stage, it brings the map with it; in addition, at each cycle, the flags in the map are updated to reflect the change in state of the pipeline, i.e. each flag is routed in the map to the stage where the corresponding instruction ends up. When an instruction corresponding to a flag leaves the pipeline, the flag is cleared in the maps. Note that each instruction has its very own different copy of the map, reflecting the fact that the conflicts are the expression of a precedence relation which involves two instructions at a time. The data collected by each instruction in the map can then be used to establish whether a certain operand is available at a certain time, thus implementing the Tomasulo technique. Note however that by optimizing this approach not only is the collective size of the maps much smaller than that of a register map (approximately one third in our case), but the update of the flags is much simpler to implement than a multiple indexed access to the scoreboard in all stages. Assuming a 2 clock cycle access time to the image memory, the pipeline has been measured to perform at about 2 clock cycles per instruction on real applications. Under these conditions, the speed-up obtained with respect to a conventional, non-pipelined controller running at a similar clock speed is close to 6. The use of more aggressive optimization techniques for the software may further improve these numbers. 5 The PCI Board We designed a PCI board hosting up to four HACRE chips, a large image memory, a program memory, a system controller, the Status Register File and the PCI interface. A piggy-back connector provides a direct interface to an input imaging device, such as a linear scanner or a video camera and to an output device, such as a video monitor. Some bus drivers and the clock generation and synchronization logic are also part of the board. A block diagram of the board is depicted in Figure 10. The Image Memory is a fast static RAM (15 ns access time), 256K \Theta 32 \Theta n bits, where n is the number of HACRE chips installed. One 32-bit memory module is associated to each of the chips. The memory is of exclusive use of the array during program execution, while it is 22 F. Gregoretti et al.: A High Speed VLSI Architecture for Handwriting Recognition Figure 10: System block diagram. memory mapped on the PC through the PCI interface when it is in STOP mode. Isolation is by means of bidirectional drivers. The External Program Memory is a static memory module analogous to those used for the Image Memory and is used to store the complete HACRE program. It is mapped on the PC address space while in STOP mode. The System Controller autonomously transfers to the internal Writable Control Store the correct instruction block during program execution, so that a continuously repeated program, even if very long, can be downloaded from the PC only once. The array outputs a completion word to the System Controller every time it completes the execution of the active program block. The completion word contains a code that directs the system controller to load a new block at a specific address or to signal program completion or an anomaly to the PC. The System Controller provides several functions. It contains a set of registers, memory mapped on the PC address space, always accessible, that enable system management. It is also responsible of the control of the piggy-back board dedicated to image acquisition and display. When debugging an application it is sometimes desirable to test an algorithm on a well-defined set of images. To be able to do so without having to modify the HACRE program, the System Controller provides means of excluding the piggy-back I/O board and give the PC an access to the VIF. The PC writes data to a register and the System Controller shifts them into the VIF. The VIF output can also be redirected to a PC memory location for test purposes. submitted to Journal of VLSI Signal Processing 23 Different types of Video I/O Boards can be connected through the piggy-back video con- nector. This connector makes available to the interface board all video signals from the VIF structure, the Host Communication Channels and several control lines from the System Con- troller. We defined a simple interface both for digital video input from a camera or a scanner and digital video output to a monitor. Input data are shifted into HACRE asynchronously, using a pixel clock provided by the video source. At each end of line a handshake protocol insures that no data are missed or overwritten. Another handshake protocol allows the interface to read from the array the output video lines as soon as they are available. A frame buffer memory, a DAC and a simple scan logic are needed to provide an analog video signal for an output monitor. The Status Register File (SRF) is used to accumulate and broadcast the global information collected by the Status Evaluation Network described in Section 3. When executing the EVAL instruction, every HACRE chip, using the Status Evaluation Network, calculates its Status Word, asserts a dedicated output line and suspends program execution. The EVAL line of the first chip is connected to the System Controller. The System Controller, using a Bus, reads the Status Word from every chip. The COUNT field is accumulated, while the SET and RESET fields are ANDed to detect all-ones and all-zeros condition. The address field of the EVAL instruction is then retrieved from the first chip and used to store the result in the corresponding Status Register. The result is also propagated to the chips, which resume program execution and can test the result in conditional instructions. The same basic mechanism is used to reset, accumulate and normalize or read the contents of a specific register. A special SRF operation interrupts the PC in correspondence of a Status Register Write, thus allowing for non-linear or custom functions, implemented in software on the PC, to be used when accumulating values in a Status Register. The SRF is very useful in object classification algorithms. Every register stores the degree of matching of the image with a different template. By analyzing the SRF, the PC can interpret the image contents accordingly. We designed the PCI board to cope with both the available chips, limited in frequency to 45-50MHz, and the future ones running at 100MHz. The clock is obtained from the system PCI clock. This was done to avoid synchronization problems between the PCI interface and the rest of the system, which have to interact. The array clock is derived from the PCI clock by a F. Gregoretti et al.: A High Speed VLSI Architecture for Handwriting Recognition Figure 11: The PCI board. PLL multiplier. The PLL can generate 33, 50, 66 or 99MHz frequencies (jumper configurable) from the 33MHz PCI clock, with known phase relationship. This high frequency signal feeds the different HACRE chips. To be able to compensate for the different PCB line lengths of the clock signals and to comply with the stringent requirements on clock loading and delay on PCI specifications, we found very useful the adoption of skew buffers to control the clock signal feeding the different devices. These are PLL devices with four outputs at the same frequency as the input one, but with delays presettable via jumpers. In this way it is possible to compensate for PCB transmission delays. All clock signals are one-to-one connections to avoid multiple reflections. All lines longer than few centimeters are series terminated at the source to match PCB line impedance. All high frequency signals run on internal layers of the PCB, shielded by ground or power supply planes, to limit EMI. The PCI interface has PCI target capabilities and conforms to PCI 2.1 specifications. All of the board registers and memories are mapped on the PCI memory address space, to form a contiguous 8 MB memory block. The board can interrupt the PC to signal error conditions, end of run and to permit non-standard SRF operations. The PCI interface, the System Controller and the SRF were all implemented using a single FPGA (Altera 10K30), providing internal memory blocks and the availability of a large number of gates. This solution was adopted for submitted to Journal of VLSI Signal Processing 25 both cost reasons and flexibility. A photograph of the PCI board is shown in Figure 11. 6 Application Implementation This section describes BEATRIX, that is an implementation of the proposed handwriting recognizer on the PCI board populated with: two HACRE chips (namely, 64 PE 0 s); 2 high-speed chips (for a total of 2 MB of image memory); 256 KWord of program memory; 256 Status Registers; 6+6 direct I/O channels; a direct interface to an image scanner; PCI interface to a hosting 100 MHz Pentium PC. The system has been tested and Section 7 shows its performance. See also [3] for additional details on the complete recognizer algorithm and performance. 6.1 System Description BEATRIX integrates four logical subsystems which are in cascade [3]: a mechanical and optical scanner, to acquire a bit-map image of the check; an image preprocessor for preliminary image filtering, scaling, and thresholding; a neural subsystem, based on an ensemble of morphological feature extractors and neuro-fuzzy networks, which detects character centers and provide hypotheses of recognition for each detected character; a context analysis subsystem based on a lexical and syntactic analyzer. The neural subsystem carries out a pre-recognition of the individual characters, based on an integrated segmentation and recognition technique [7]. Legal and courtesy amounts are preprocessed and recognized independently (at the character level) and then the two streams of information are sent to the common context analysis subsystem, which exploits all the mutual redundancy. The context analysis subsystem combines the candidate characters and, guided by the mutual redundancy present in the legal and courtesy amounts, produces hypotheses about the amount so as to correct errors made by the neural subsystem alone. The image preprocessor and the neural subsystem are executed on the PAPRICA-3 system (namely, the PCI board), while the context analysis subsystem is executed by an external Pen- 26 F. Gregoretti et al.: A High Speed VLSI Architecture for Handwriting Recognition tium processor, which can implement these types of algorithms more efficiently. 6.2 Image Preprocessor The first preprocessing subsystem is the image preprocessor which consists of the blocks described below. ffl A WINDOW EXTRACTOR acquires the input image from the SCANNER, at a resolution of approximately 200 dpi, 16 gray levels. The scanner is an 876-pixel CCD line camera scanned mechanically over the image, from right to left (due to practical reasons), at a speed of 2m/s (which is equivalent to about 700 characters/s); Fig. 12.a. Image acquisition is performed by the VIF, in parallel with processing, and a whole image line is acquired in just one clock cycle. ffl A FILTER block computes a simple low-pass filter with a 3 \Theta 3 pixel kernel (Fig. 12.b), while a BRIGHTNESS block compensates for the non-uniform detector sensitivity and paper color (Fig. 12.c). ffl A THRESHOLD block converts the gray-scale image into a B/W image by a comparison with an adaptive threshold (Fig. 12.d). ffl A THINNING block reduces the width of all the strokes to 1 pixel (Fig. 12.f). Thinning is a morphological operator [5] which reduces the width of lines, while preserving stroke connectivity. ffl A BASELINE block detects the baseline of the handwritten text, which is a horizontal stripe intersecting the text in a known position (Fig. 12.f). ffl A FEATURES block detects and extracts from the image a set of 12 stroke features, which are helpful for further character recognition. As shown in Fig. 12.h (crosses), this block detects the four left, right, top and bottom concavities, and the terminal strokes in the eight main directions. ffl A FEATURE REDUCTION, a ZOOM and a COMPRESS blocks reduce, respectively, the number of features (by removing both redundant and useless ones), the vertical size of the manuscript submitted to Journal of VLSI Signal Processing 27 a) c) d) f) Figure 12: Preprocessing steps of handwritten images (an "easy" example): a) original image, 200 dpi, 16 gray levels; b) low-pass filtered image; c) compensated for brightness; d) thresholded; spot noise removal; f) thinned, after 6 steps; g) finding baseline (at the left side of the image); features detection (features are tagged by small crosses); i) compressed. (to approximately 25-30 pixels), and the overall size of the manuscript (by a linear factor of 2), by means of ad-hoc topological transformations which do not preserve image shape, although they do preserve its connectivity. (Fig. 12.i) After all the preprocessing steps, the B/W image is ready for the following neural recognition steps (see Section 6.3). The image is reduced both in size (down to 14 \Theta (=252) pixels for the courtesy and the legal amounts, respectively), in number of gray levels (2), and in stroke thickness (1 pixel), and noise is removed. Table 1 lists execution times of individual blocks. 6.3 Neural subsystem The neural subsystem is made of two cascaded subsystems, namely a CENTERING DETECTOR and a CHARACTER RECOGNIZER. See [3] for further details on the algorithms and the implementation 28 F. Gregoretti et al.: A High Speed VLSI Architecture for Handwriting Recognition of each block. Table 1 lists execution times of individual blocks. ffl The CENTERING DETECTOR scans the preprocessed and compressed image from right to left (for mechanical reasons) and extracts a sliding window of fixed size. It then tries to locate the characters, by detecting the position of their center, based on the type, quantity and mutual positions of the detected features. Note that windows without strokes are immediately skipped, as they contain no useful information. Lines can be skipped in as low as one clock cycle. ffl The CHARACTER RECOGNIZER recognizes each individual pseudo-character, using a hybrid approach, which mixes feature-based [2] and neural [1] recognizers. First of all, features extracted by the FEATURES block are used to identify all easy-to-recognize characters. For instance, most "0", "6", "9" digits (but not only these) are written well enough that a straightforward and fast analysis of the main features and strokes is sufficient to recognize those characters with high accuracy. Other characters are more difficult to recognize using only features; for instance, digits "4", "7" and some types of "1" can be recognized more easily using neural techniques. All characters which have not been recognized using features are isolated and passed to a neural network (an ad-hoc 2-layers WRBF [4]) trained by an appropriate training set. Therefore the CHARACTER RECOGNIZER is made of two blocks, namely a feature-based recognizer and a neural recognizer, each one optimized to recognize a particular subset of the whole alphabet. The CHARACTER RECOGNIZER is "triggered" for each pseudo-character center detected by the CENTERING DETECTOR. As shown in Table 1, the CHARACTER RECOGNIZER is the slowest piece of code. Fortunately it is run at a relatively low rate, namely every 15 lines, in the average, therefore its effects on computing time are limited. submitted to Journal of VLSI Signal Processing 29 Performance Table 1 lists the execution times of the various processing blocks for the example presented in Section 2; figures are given for a system with 64 PE's (namely, 2 chips), running at 33 MHz. All the programs were also tested on both a Pentium at 100 MHz and a Sparc Station 10, using the same algorithms based on mathematical morphology, which are well suited to the specific problems of bitmap processing and character recognition. Some programs (FILTER, BRIGHTNESS, THRESHOLD, ZOOM, CENTERING DETECTOR, CHARACTER RECOGNIZER) could be implemented more efficiently on a sequential computer using more traditional methods (ad-hoc programs). These were also implemented on the Pentium and their performance listed in Table 1 for comparison. It can be seen that the performance of PAPRICA-3 is 100 to 1000 times faster than that of Pentium 100 MHz and Sparc Station, for nearly all the programs considered. This improvement factor reduces by at most five folds when a Pentium 500 MHz is used. For further comparison, Table 2 lists the execution times of a single-chip PAPRICA-3 system running at 100 MHz, for other well-known neural algorithms such as Perceptrons (MLP) and Kohonen maps [6]. As all mathematical operations are implemented in a bit-serial fashion, system performance depends heavily on input and weight resolution. Furthermore, the best performance can be obtained when the number of either neurons or inputs matches the number of PE's. The overall recognition accuracy of BEATRIX (not quoted here; see [3]) are comparable with other existing handwriting recognition systems, but our system achieves that at a much lower price. F. Gregoretti et al.: A High Speed VLSI Architecture for Handwriting Recognition worst case morphol. ad-hoc morphol. Image preprocessor -s/line ms/check -s/line -s/line -s/line BRIGHTNESS 5.90 11.8 1,970 320 1,660 FEATURES 6.10 12.2 9,490 - 10,430 ZOOM 4.48 8.96 820 160 760 COMPRESS y 61.6 123.2 21,350 - 24,950 PAPRICA-3 Pentium 100 MHz worst case morphol. ad-hoc Neural subsystem ms/psd-char ms/check ms/check ms/check RECOGNIZER RECOGNIZER (NEURAL) yy 3.36 161.4 - 13,440 TOTAL RECOGNIZER 11.4 545.4 27,840 Table 1: Average execution times of the various processing steps, while processing the courtesy amount. y COMPRESS acts on an image zoomed by an average factor 3.2, therefore processing times are scaled accordingly. yy CHARACTER RECOGNIZER acts a few times per each pseudo-character, namely once every 15 lines on average. submitted to Journal of VLSI Signal Processing 31 Neural Network Paradigm internal weights external weights MCPS MCUPS MCPS MCUPS adaptive learning rate neighborhood Kohonen, neurons, 8 inputs, 8 bits/input, 5 \Theta 5 neighborhood 90 Table 2: Performance of the BEATRIX system, in single-chip configuration (namely, running at 100 MHz, with either internal weights (max. 60 bits/neuron) or external weights (no size limitation). MCPS and MCUPS stand for, respectively, mega connections per second and mega connection updates per second. 8 Conclusion From the hardware implementation point of view not all the original goals have been reached. In particular all the main full custom blocks (memory, PA) have been designed and verified by simulation to operate within a ns clock cycle in the worst case, but the whole chip is fully functional up to a maximum frequency of 50 MHz. This is due to strict deadlines on the availability of funds for chip fabrication which have reduced the time available for the optimization of the CU layout on the basis of backannotated simulation. Moreover, as clearly visible in the microphotograph of figure 4, a large area of approximately wasted. This is mainly due to the limitations of the tools employed in the placing and routing phase of the CU which has been synthesized into a standard cell library form a HDL description, This has led to a large increase of the ratio between the CU area and the PA area and Preliminary tests with new tools have shown that the current layout size could be reduced by approximately 15 %. This would allow to place on board of a single chip system an integrated version of the Status Register File [17] which has been designed in order to minimize the components of a single chip system. With the current technological evolution of VLSI circuits preliminary evaluations made with F. Gregoretti et al.: A High Speed VLSI Architecture for Handwriting Recognition a 0:35 -m technology have shown that a single chip system could integrate 64 PE's, the SRF and 64 Kbit of image memory, making it possible a really fully integrated system for handwritten character recognition. For what system performance concerns, the performance of the BEATRIX system have shown that the proposed PAPRICA-3 architecture, even in a medium-size configuration, outperforms the Pentium processor by much more than a factor ten. In addition, the recognition accuracy of BEATRIX is comparable with other much more expensive systems. In addition the development environment and the image processing language (not described here), which have been developed explicitly for PAPRICA-3 allow a straightforward design of new mathematical morphology and image processing algorithms, reducing the design time of new algorithms. --R "Off-line cursive script word recognition" "Invariant handwriting features useful in cursive-script recognition" "High-Speed Recognition of Handwritten Amounts on Italian Checks" " Weighted Radial Basis Functions for Improved Pattern Recognition and Signal Processing" "Image Analysis and Mathematical Morphology" "Neural Networks: A Comprehensive Foundation" "A Chip Set Implementation of a Parallel Cellular Architecture" "Design and Implementation of the PAPRICA Parallel Architecture" Implementation of a slim array processor. A linear array parallel image processor: Slim-ii The PAPRICA SIMD array: Critical reviews and perspectives. Efficient image processing algorithms on the scan line array processor. A 3.84 GIPS integrated memory array processor with 64 processing elements and a 2-MB SRAM An array processor for general purpose digital image compression. ''An intelligent Register File for Neural Processing applications''. --TR Off-Line Cursive Script Word Recognition Efficient Image Processing Algorithms on the Scan Line Array Processor The evolution of the PAPRICA system Design and Implementation of the PAPRICA Parallel Architecture High-speed recognition of handwritten amounts on Italian checks Neural Networks Implementation of a SliM Array Processor Recognition-Based Segmentation of On-Line Hand-Printed Words A Linear Array Parallel Image Processor

parallel architectures;image processing;handwriting recognition;artificial neural networks;VLSI implementations

509594

The effects of communication parameters on end performance of shared virtual memory clusters.

Recently there has been a lot of effort in providing cost-effective Shared Memory systems by employing software only solutions on clusters of high-end workstations coupled with high-bandwidth, low-latency commodity networks. Much of the work so far has focused on improving protocols, and there has been some work on restructuring applications to perform better on SVM systems. The result of this progress has been the promise for good performance on a range of applications at least in the 16-32 processor range. New system area networks and network interfaces provide significantly lower overhead, lower latency and higher bandwidth communication in clusters, inexpensive SMPs have become common as the nodes of these clusters, and SVM protocols are now quite mature. With this progress, it is now useful to examine what are the important system bottlenecks that stand in the way of effective parallel performance; in particular, which parameters of the communication architecture are most important to improve further relative to processor speed, which ones are already adequate on modern systems for most applications, and how will this change with technology in the future. Such information can assist system designers in determining where to focus their energies in improving performance, and users in determining what system characteristics are appropriate for their applications.We find that the most important system cost to improve is the overhead of generating and delivering interrupts. Improving network interface (and I/O bus) bandwidth relative to processor speed helps some bandwidth-bound applications, but currently available ratios of bandwidth to processor speed are already adequate for many others. Surprisingly, neither the processor overhead for handling messages nor the occupancy of the communication interface in preparing and pushing packets through the network appear to require much improvement.

Introduction With the success of hardware cache-coherent distributed shared memory (DSM), a lot of effort has been made to support the programming model of a coherent shared address space using commodity- oriented communication architectures in addition to commodity nodes. The techniques for the communication architecture range from using less customized and integrated controllers [17, 2] to supporting shared virtual memory (SVM) at page level through the operating system [12, 8]. While these techniques reduce cost, unfortunately they usually lower performance as well. A great deal of research effort has been made to improve these systems for large classes of applications. Our focus in this paper is on SVM systems. In the last few years there has been much improvement of SVM protocols and systems, and several applications have been restructured to improve performance [12, 8, 25, 14]. With this progress, it is now interesting to examine what are the important system bottlenecks that stand in the way of effective parallel performance; in particular, which parameters of the communication architecture are most important to improve further relative to processor speed, which are already adequate on modern systems for most applications, and how will this change with technology in the future. Such studies can hopefully assist system designers in determining where to focus their energies in improving performance, and users in determining what system characteristics are appropriate for their applications. This paper examines these questions through detailed architectural simulation using applications with widely different behavior. We simulate a cluster architecture with SMP nodes and a fast system area interconnect with a programmable network interface (i.e. Myrinet). We use a home-based SVM protocol that has been demonstrated to have comparable or better performance than other families of SVM protocols. The base case of the protocol, called home-based lazy release consistency (HLRC) does not require any additional hardware support. We later examine a variant, called automatic update release consistency (AURC) that uses automatic (hardware) propagation of writes to remote nodes to perform updates to shared data, and also extend our analysis to the use of uniprocessor nodes where this is useful. The major performance parameters we consider are the host processor overhead to send a message, the network interface occupancy to prepare and transfer a packet, the node-to-network bandwidth (often limited by I/O bus bandwidth), and the interrupt cost. We do not consider network link latency, since it is a small and usually constant part of the end-to-end latency, in system area networks (SAN). After dealing with performance parameters, we also briefly examine the impact of key granularity parameters of the communication architecture. These are the page size, which is the granularity of coherence, and the number of processors per node. Assuming a realistic system that can be quite easily implemented today, and a range of applications that are well optimized for SVM systems [10], we see (Figure 1) that, for most applications, protocol and communication overheads are substantial. The speedups obtained in the realistic implementation are much lower than in the ideal case, where all communication costs are zero. This motivates the current research, whose goal is twofold. First, we want to understand how performance changes as FFT LU Ocean (contiguous) Water (nsquared) Water (spatial) Radix Volrend Raytrace Barnes (rebuild) Barnes (space)26101418Speedup Figure 1. Ideal and realistic speedups for each application. The ideal speedup is computed as the ratio of the uniprocessor execution time divided by the sum of the compute and the local cache stall time in the parallel execution, i.e. ignoring all communication and synchronization costs. The realistic speedup corresponds to a realistic set of values (see Section communication architecture parameters today, in a configuration with four processors per node. the parameters of the communication architecture are varied relative to processor speed, both to see where we should invest systems energy and to understand the likely evolution of system performance as technology evolves. For this, we use a wide range of values for each parameter. Second, we focus on three specific points in the parameter space. The first is the point for which the system generally achieves its best performance within the ranges of parameter values we examine. The performance on an application at this point is called its best performance. The second point is an aggressive set of values that the communication parameters can have in current or near-future systems, especially if certain operating system features are well optimized. The performance at this point in the space is called the achievable performance. The third point is the ideal point, which represents a hypothetical system that incurs no communication or synchronization overheads, taking into consideration only compute time and stall time on local data accesses. Our goal is to understand the gaps between the achievable, best and ideal performance by identifying which parameters contribute most to the performance differences. This leads us to the primary communication parameters that need to be improved to close the gaps. We find, somewhat surprisingly, that host overhead to send messages and per-packet network interface occupancy are not critical to application performance. In most cases, interrupt cost is by far the dominant performance bottleneck, even though our protocol is designed to be very aggressive in reducing the occurrence of interrupts. Node-to-network bandwidth, typically limited by the I/O bus, is also significant for a few applications, but interrupt cost is important for all the applications we study. These results suggest that system designers should focus on reducing interrupt costs to support SVM well, and SVM protocol designers should try to avoid interrupts as possible, perhaps by using polling or by using a programmable communication assist to run part of the protocol avoiding the need to interrupt the main processor. Our results show the relative effect of each of the parameters, i.e. relative to the processor speed and F O F F O I Processor M e r y s Processor I/O s First level Cache Buffer Write Second Level Cache Network Interface e r y Snooping Device Figure 2. Simulated node architecture. to one another. While the absolute parameter values that are used for the achievable set match what we consider achievable with aggressive current or near-future systems, viewing the parameters relative to processor speed allows us to understand what the behavior will be as technology trends evolve. For instance, if the ratio of bandwidth to processor speed changes, we can use these results to reason about system performance. Section 2 presents the architectural simulator that we use in this work. In Section 3 we discuss the parameters we use and the methodology of the study. Section 4 presents the applications suite. Sections 5 and 6 present our results for the communication performance parameters. In Section 5 we examine the effects of each parameter on system performance and in Section 6 we discuss for each application the parameters that limit its performance. Section 7 presents the effects of page size and degree of clustering on system performance. We discuss related work in Section 8. Finally we discuss future work directions and conclusions in Sections 9 and 10 respectively. Simulation Environment The simulation environment we use is built on top of augmint [18], an execution driven simulator using the x86 instruction set, and runs on x86 systems. In this section we present the architectural parameters that we do not vary. The simulated architecture (Figure 2) assumes a cluster of c-processor SMPs connected with a commodity interconnect like Myrinet [3]. Contention is modeled at all levels except in the network links and switches themselves. The processor has a P6-like instruction set, and is assumed to be a 1 IPC pro- cessor. The data cache hierarchy consists of a 8 KBytes first-level direct mapped write-through cache and a 512 KBytes second-level two-way set associative cache, each with a line size of 32 Bytes. The buffer [19] has 26 entries, 1 cache line wide each, and a retire-at-4 policy. Write buffer stalls are simulated. The read hit cost is one cycle if satisfied in the write buffer and first level cache, and 10 cycles if satisfied in the second-level cache. The memory subsystem is fully pipelined. Each network interface (NI) has two 1 MByte memory queues, to hold incoming and outgoing packets. The size of the queues is such that they do not constitute a bottleneck in the communication subsystem. If the network queues fill, the NI interrupts the main processor and delays it to allow queues to drain. Network links operate at processor speed and are 16 bits wide. We assume a fast messaging system [5, 16, 4] as the basic communication library. The memory bus is split-transaction, 64 bits wide, with a clock cycle four times slower than the processor clock. Arbitration takes one bus cycle, and the priorities are, in decreasing order: outgoing network path of the NI, second level cache, write buffer, memory, incoming path of the NI. The I/O bus is bits wide. The relative bus bandwidth and processor speed match those on modern systems. If we assume that the processor has a 200 MHz clock, the memory bus is 400 MBytes/s. Protocol handlers themselves cost a variable number of cycles. While the code for the protocol handlers can not be simulated since the simulator itself is not multi-threaded, we use for each handler an estimate of the cost of its code sequence. The cost to access the TLB from a handler running in the kernel is 50 processor cycles. The cost of creating and applying a diff is 10 cycles for every word that needs to be compared and 10 additional cycles for each word actually included in the diff. The protocols we use are two versions of a home-based protocol, HLRC and AURC [8, 25]. These protocols either use hardware support for automatic write propagation (AURC) or traditional software diffs (HLRC) to propagate updates to the home node of each page at a release point. The necessary pages are invalidated only at acquire points according to lazy release consistency (LRC). At a subsequent page fault, the whole page is fetched from the home, where it is guaranteed to be up to date according to the lazy release consistency [8]. The protocol for SMP nodes attempts to utilize the hardware sharing and synchronization within an SMP as much as possible, reducing software involvement [1]. The optimizations used include the use of hierarchical barriers and the avoidance of interrupts as much as possible. Interrupts are used only when remote requests for pages and locks arrive at a node. Requests are synchronous (RPC like), to avoid interrupts when replies arrive at the requesting node. Barriers are implemented with synchronous messages and no interrupts. Interrupts are delivered to processor 0 in each node. More complicated schemes (i.e. round robin, random assignment) that result in better load balance in interrupt handling can be used if the operating system provides the necessary support. These schemes however, may increase the cost of delivering interrupts. In this paper we also examine a round robin scheme. As mentioned earlier, we focus on the following performance parameters of the communication archi- tecture: host overhead, I/O bus bandwidth, network interface occupancy, and interrupt cost. We do not examine network link latency, since it is a small and usually constant part of the end-to-end latency, in system area networks (SAN). These parameters describe the basic features of the communication sub- system. The rest of the parameters in the system, for example cache and memory configuration, total number of processors, etc. remain constant. When a message is exchanged between two hosts, it is put in a post queue at the network interface. In an asynchronous send operation, which we assume, the sender is free to continue with useful work. The network interface processes the request, prepares packets, and queues them in an outgoing network queue, incurring an occupancy per packet. After transmission, each packet enters an incoming network queue at the receiver, where it is processed by the network interface and then deposited directly in host memory without causing an interrupt [2, 4]. Thus, the interrupt cost is an overhead related not so much to data transfer but to processing requests. While we examine a range of values for each parameter, in varying a parameter we usually keep the others fixed at the set of achievable values. Recall that these are the values we might consider achievable currently, on systems that provide optimized operating system support for interrupts. We choose relatively aggressive fixed values so that the effects of the parameter being varied are observed. In more detail: Host Overhead is the time the host processor itself is busy sending a message. The range of this parameter is from a few cycles to post a send in systems that support asynchronous sends, up to the time needed to transfer the message data from the host memory to the network interface when synchronous sends are used. If asynchronous sends are available, an achievable value for the host overhead is a few hundred processor cycles. Recall that there is no processor overhead for a data transfer at the destination end. The range of values we consider is between 0 (or almost cycles and 10000 processor cycles (about 50-s with a 5ns processor clock). Systems that support asynchronous sends will probably be closer to the smaller values and systems with synchronous sends will be closer to the higher values depending on the message size. The achievable value we use is an overhead of 600 processor cycles per message. ffl The I/O Bus Bandwidth determines the host to network bandwidth (relative to processor speed). In contemporary systems this is the limiting hardware component for the available node-to-network network links and memory buses tend to be much faster. The range of values for the I/O bus bandwidth is from 0.25 MBytes per processor clock MHz up to 2 MBytes per processor clock MHz (or 50 MBytes/s to 400 MBytes/s assuming a 200 MHz processor clock). The achievable value is 0.5 MBytes/MHz, or 100 MBytes/s assuming a 200 MHz processor clock. ffl Network Interface Occupancy is the time spent on the network interface preparing each packet. Network interfaces employ either custom state machines or network processors (general purpose or custom designs) to perform this processing. Thus, processing costs on the network interface vary widely. We vary the occupancy of the network interface from almost 0 to 10000 processor cycles (about 50-s with a 5ns processor clock) per packet. The achievable value we use is 1000 main processor cycles, or about 5-s assuming a 200 MHz processor clock. This value is realistic for the currently available programmable NIs, given that the programmable communication assist on the NI is usually much slower than the main processor. ffl Interrupt cost is the cost to issue an interrupt between two processors in the same SMP node, or the cost to interrupt a processor from the network interface. It includes the cost of context switches and operating system processing. Although the interrupt cost is not a parameter of the communication subsystem, it is an important aspect of SVM systems. Interrupt cost depends on the operating system used; it can vary greatly from system to system, affecting the performance portability of SVM across different platforms. We therefore vary the interrupt cost from free interrupts (0 processor cycles) to 50000 processor cycles for both issuing and delivering an interrupt (total 100000 processor cycles or 500 -s with a 5ns processor clock). The achievable value we use is 500 processor cycles, which results in a cost of 1000 cycles for a null interrupt. This choice is significantly more aggressive than what current operating systems provide. However it is achievable with fast interrupt technology [21]. We use it as the achievable value when varying other parameters to ensure that interrupt cost does not swamp out the effects of varying those parameters. To capture the effects of each parameter separately, we keep the other parameters fixed at their achievable values. Where necessary, we also perform additional guided simulations to further clarify the results In addition to the results obtained by varying parameters and the results obtained for the achievable parameter values, an interesting result is the speedup obtained by using the best value in our range for each parameter. This limits the performance that can be obtained by improving the communication architecture within our range of parameters. The parameter values for the best configuration are: host overhead 0 processor cycles, I/O bus bandwidth equal to the memory bus bandwidth, network interface occupancy per packet 200 processor cycles and total interrupt cost 0 processor cycles. In this best con- figuration, contention is still modeled since the values for the other system parameters are still nonzero. Table 1 summarizes the values of each parameter. With a 200 MHz processor, the achievable set of values discussed above assumes the parameter values: host overhead 600 processor cycles, memory bus bandwidth 400 MBytes/s, I/O bus bandwidth 100 MBytes/s, network interface occupancy per packet 1000 processor cycles and total interrupt cost 1000 processor cycles. Parameter Range Achievable Best Host Overhead (cycles) 0-10000 600 -0 I/O Bus Bandwidth (Mbytes/MHz) 0.25-2 0.5 2 NI Occupancy (cycles) 0-10000 1000 200 Table 1. Ranges and achievable and best values of the communication parameters under consideration. Applications We use the SPLASH-2 [22] application suite. This section briefly describes the basic characteristics of each application relevant to this study. A more detailed classification and description of the application behavior for SVM systems with uniprocessor nodes is provided in the context of AURC and LRC in [9]. The applications can be divided in two groups, regular and irregular. 4.1 Regular Applications The applications in this category are FFT, LU and Ocean. Their common characteristic is that they are optimized to be single-writer applications; a given word of data is written only by the processor to which it is assigned. Given appropriate data structures they are single-writer at page granularity as well, and pages can be allocated among nodes such that writes to shared data are almost all local. In HLRC we do not need to compute diffs, and in AURC we do not need to use a write through cache policy. Protocol action is required only to fetch pages. The applications have different inherent and induced communication patterns [22, 9], which affect their performance and the impact on SMP nodes. Application Page Faults Page Fetches Local Lock Acquires Remote Lock Acquires Barriers Water(nsquared) (512) 69.19 22.06 8.04 68.26 19.01 7.29 Water(spatial) (512) 97.86 21.42 9.23 93.81 17.73 6.04 0.01 1.83 2.60 3.94 2.16 1.39 4.19 Volrend (head) 105.09 44.06 34.49 104.78 29.35 6.53 0.00 29.34 43.80 44.34 17.64 3.97 1.61 Raytrace (car) 89.80 25.64 6.83 89.79 25.57 6.76 0.03 2.21 3.96 4.89 3.26 1.34 0.10 Barnes(rebuild) Barnes(space) Table 2. Number of page faults, page fetches, local and remote lock acquires and barriers per processor per 10 7 cycles for each application for 1,4 and 8 processors per node. FFT: The all-to-all, read-based communication in FFT is essentially a transposition of a matrix of complex numbers. We use two problem sizes, 256K(512x512) and 1M(1024x1024) elements. FFT has a high inherent communication to computation ratio. LU: We use the contiguous version of LU, which allocates on each page data assigned to only one processor. LU exhibits a very small communication to computation ratio but is inherently imbalanced. We used a 512x512 matrix. Ocean: The communication pattern in the Ocean application is largely nearest-neighbor and iterative on a regular grid. We run the contiguous (4-d array) version of Ocean on a 514x514 grid with an error tolerance of 0.001. 4.2 Irregular Applications The irregular applications in our suite are Barnes, Radix, Raytrace, Volrend and Water. Barnes: We ran experiments for different data set sizes, but present results for 8K particles. Access patterns in Barnes are irregular and fine-grained. We use two versions of Barnes, which differ in the manner they build the shared tree at each time step. In the first version (Barnes-rebuild, which is the one in SPLASH-2) processors load the particles that were assigned to them for force calculation directly into the shared tree, locking (frequently) as necessary. The second version, Barnes-space [10], is optimized for SVM, and it avoids locking as much as possible. It uses a different tree-building algorithm, in which disjoint subspaces that match tree cells are assigned to different processors. These subspaces include particles which are not the same as the particles that are assigned to the processors for force calculation. Each processor builds each own partial tree, and all partial trees are merged to the global tree without locking. Radix: Radix sorts a series of integer keys. It is a very irregular application with highly scattered writes to remotely allocated data and a high inherent communication to computation ratio. We use the unmodified SPLASH-2 version. FFT LU Ocean (contiguous) Water (nsquared) Water (spatial) Radix Volrend Raytrace Barnes (rebuild) Barnes (space)200600100014001800Normalized Number of Messages Sent Figure 3. Number of messages sent per processor per 10 7 compute cycles for each application for 1,4 and 8 processors per node. Raytrace: Raytrace renders complex scenes in computer graphics. The version we use is modified from the SPLASH-2 version to run more efficiently on SVM systems. A global lock that was not necessary was removed, and task queues are implemented better for SVM and SMP [10]. Inherent communication is small. Volrend: The version we use is slightly modified from the SPLASH-2 version, to provide a better initial assignment of tasks to processes before stealing [10]. This improves SVM performance greatly. Inherent communication volume is small. Water: We use both versions of Water from SPLASH-2, Water-nsquared and Water-spatial. Water- nsquared can be categorized as a regular application, but we put it here to ease the comparison with Water-spatial. In both versions, updates to water molecules positions and velocities are first accumulated locally by processors and then performed to the shared data once at the end of each iteration. The inherent communication to computation ratio is small. We use a data set size of 512 molecules. Table 2 and Figures 3 and 4 can be used to characterize the applications. Table 2 presents counts of protocol events for each application, for 1, 4 and 8 processors per node (16 processors total in all cases). Figures 3 and 4 show the numbers of messages and MBytes of data (both application and protocol) that are sent by each processor in the system. These characteristics are measured per 10 7 cycles of application compute time per processor, and are averaged over all processors in the system. We can use them to categorize the applications in terms of the communication they exhibit. Both the number of messages and MBytes of data exchanged are important to performance; if we use the geometric mean of these properties, which captures their multiplicative effect, as a metric then we can divide the applications in FFT LU Ocean (contiguous) Water (nsquared) Water (spatial) Radix Volrend Raytrace Barnes (rebuild) Barnes Normalized MBytes Sent Figure 4. Number of MBytes sent per processor per 10 7 compute cycles for each application for 1,4 and 8 processors per node. three groups. In the first group are Barnes-rebuild, FFT and Radix that exhibit a lot of communication. In the second group belong Water-nsquared and Volrend that exhibit less communication and in the third group the rest of the applications, LU, Ocean, Water-spatial, Raytrace and Barnes-space that exhibit very little communication. It is important to note that this categorization holds for the 4-processor per node configuration. Changing the number of processors in the node, can dramatically change the behavior of some applications, and the picture can be very different. For instance Ocean exhibits very high communication with 1 processor per node. 5 Effects of Communication Parameters In this section we present the effects of each parameter on the performance of an all-software HLRC protocol for a range of values. Table 3 presents the maximum slowdowns for each application for the parameters under consideration. The maximum slowdown is computed from the speedups for the smallest and biggest values considered for each parameter, keeping all other parameters at their achievable val- ues. Negative numbers indicate speedups. The rest of this section discusses the parameters one by one. For each parameter we also identify the application characteristics that most closely predict the effect of that parameter. The next section will take a different cut, looking at the bottlenecks on a per-application rather than per-parameter basis. At the end of this section we also present results for AURC. Host Overhead: Figure 5 shows that the slowdown due to the host overhead is generally low, especially for realistic values of asynchronous message overheads. However, it varies among applications from less than 10% for Barnes-space, Ocean-contiguous and Raytrace to more than 35% for Volrend, Radix and Barnes-rebuild across the entire range of values. In general, applications that send more messages exhibit a higher dependency on the host overhead. This can be seen in Figure 6, which shows two Application Host Overhead NI Occupancy I/O Bus Bandwidth Interrupt Cost Page Size Procs/Node FFT 22.6% 11.9% 40.8% 86.6% 72.6% 13.8% LU(contiguous) 17.9% 7.5% 15.9% 70.8% 34.4% -35.3% Ocean(contiguous) 4.5% 2.8% 6.5% 35.2% 19.6% 63.2% Water(nsquared) 32.4% 16.6% 10.8% 83.2% 62.2% -87.1% Water(spatial) 23.7% 8.5% 8.9% 67.9% 51.0% -87.5% Radix 35.8% -31.8% 77.6% 58.7% -368.2% -699.4% Volrend 34.7% 12.8% 15.7% 91.3% 63.9% -68.1% Raytrace 8.2% 2.9% 8.9% 52.3% 9.1% -16.1% Barnes(rebuild) 40.7% 21.8% 44.8% 80.3% 71.5% -383.4% Barnes(space) 4.4% -0.6% 27.5% 59.0% -109.6% -49.4% Table 3. Maximum Slowdowns with respect to the various communication parameters for the range of values with which we experiment. Negative numbers indicate speedups. curves. One is the slowdown of each application between the smallest and highest host overheads that we simulate, normalized to the biggest of these slowdowns. The second curve is the number of messages sent by each processor per 10 6 compute cycles, normalized to the biggest of these numbers of messages. Note that with asynchronous messages, host overheads will be on the low side of our range, so we can conclude that host overhead for sending messages is not a major performance factor for coarse grain SVM systems and is unlikely to become so in the near future. Network Interface Occupancy: Figure 7 shows that network interface occupancy has even a smaller effect than host overhead on performance, for realistic occupancies. Most applications are insensitive to it, with the exception of a couple of applications that send a large number of messages. For these applications, slowdowns of up to 22% are observed at the highest occupancy values. The speedup observed for Radix is in reality caused by timing issues (contention is the bottleneck in Radix). I/O Bus Bandwidth: Figure 8 shows the effect of I/O bandwidth on application performance. Reducing the bandwidth results in slowdowns of up to 82%, with 4 out of 11 applications exhibiting slowdowns of more than 40%. However, many other applications are not so dependent on bandwidth, and only FFT, Radix, and Barnes-rebuild benefit much from increasing the I/O bus bandwidth beyond the achievable relationships to processor speed today. Of course, this does not mean that it is not important to worry about improving bandwidth. As processor speed increases, if bandwidth trends do not keep up, we will quickly find ourselves at the relationship reflected by the lower bandwidth case we examine (or even worse). What it does mean is that if bandwidth keeps up with processor speed, it is not likely to be the major limitation on SVM systems for applications. Figure 9 shows the dependency between bandwidth and the number of bytes sent per processor for each application. As before, units are normalized to the maximum of the numbers presented for each curve. Applications that exchange a lot of data, not necessarily a lot of messages, need higher bandwidth. shows that interrupt cost is a very important parameter in the system. Unlike bandwidth, it affects the performance of all applications dramatically, and in many cases a relatively FFT LU-contiguous Ocean-contiguous Water-nsquared Water-spatial Radix Volrend Raytrace Barnes-space Barnes-rebuild Figure 5. Effects of host overhead on application performance. The data points for each application correspond to a host overhead of 0, 600, 1000, 5000, and 10000 processor cycles. small increase in interrupt cost leads to a big performance degradation. For most applications, interrupt costs of up to about 2000 processor cycles for each of initiation and delivery do not seem to hurt much. However, commercial systems typically have much higher interrupt costs. Increasing the interrupt cost beyond this point begins to hurt sharply. All applications have a slowdown of more than 50% when the interrupt cost varies from 0 to 50000 processor cycles (except Ocean-contiguous that exhibits an anomaly since the way pages are distributed among processors changes with interrupt cost). This suggests that architectures and operating systems should work harder to improving interrupt costs if they are to support SVM well, and SVM protocols should try to avoid interrupts as much as possible, Figure 11 shows that the slowdown due to the interrupt cost is closely related to the number of protocol events that cause interrupts-page fetches and remote lock acquires. With SMP nodes there are many options for how interrupts may be handled within a node. Our protocol uses one particular method. Systems with uniprocessor nodes have less options, so we experimented with such configurations as well. We found that interrupt cost is important in that case as well. The only difference is that the system seems to be a little less sensitive to interrupt costs of between 2500 and 5000 cycles. After this range, performance degrades quickly as in the SMP configuration. We also experimented with round robin interrupt delivery and the results look similar to the case where all interrupts are delivered to a fixed processor in each SMP. Overall performance seems to increase slightly, compared to the static interrupt scheme, but as in the static scheme it degrades quickly as cost increases. Moreover implementing such a scheme in a real system may be complicated and may incur additional costs. LU(contiguous) Ocean(contiguous) Radix Raytrace Volrend Water(nsquared) Water(spatial)0.10.30.50.70.9Normalized units Slowdown due to Host Overhead (normalized to the largest slowdown) Number of Messages sent/Processor/1M Compute Cycles (normalized to the largest) Figure 6. Relation between slowdown due to Host Overhead and Number of Messages sent. AURC: As mentioned in the introduction, besides HLRC, we also used AURC to study the effect of the communication parameters when using hardware support for automatic write propagation instead of software diffs. The results look very similar to HLRC, with the exception that network interface occupancy is much more important in AURC. The automatic update mechanism may generate more traffic through the network interface because new values for the same data may be sent multiple times to the home node before a release. More importantly, the number of packets may increase significantly since updates are sent at a much finer granularity, so if they are apart in space or time they may not be coalesced well into packets. Figure 12 shows how performance changes as the NI overhead increases for both regular and irregular applications. 6 Limitations on Application Performance In this section we examine the difference in performance between the best configuration and an ideal system (where the speedup is computed only from the compute and local stall times, ignoring communication and synchronization costs), and the difference in performance between the achievable and the best configuration on a per application basis. Recall that best stands for the configuration where all communication parameters assume their best value, and achievable stands for the configuration where the communication parameters assume their achievable values. The goal is to identify the application properties and architectural parameters that are responsible for the difference between the best and the FFT LU-contiguous Ocean-contiguous Water-nsquared Water-spatial Radix Volrend Raytrace Barnes-space Barnes-rebuild Figure 7. Effects of network interface occupancy on application performance. The data points for each application correspond to a network occupancy of 50, 250, 500, 1000, 2000, and 10000 processor cycles. ideal performance, and the parameters that are responsible for the difference between the achievable and the best performance. The speedups for each configuration will be called ideal, best and achievable respectively. Table 4 shows these speedups for all applications. In many cases, the achievable speedup is close to the best speedup. However, in some cases (FFT, Radix, Barnes) there remains a gap. The performance with the best configuration is often quite far from the ideal speedup. To understand these effects, let us examine each application separately. FFT: The best speedup for FFT is about 13.5. The difference from the ideal speedup of 16.2 comes from data wait time at page faults, which have a cost even for the best configuration, despite the very high bandwidth and the zero-cost interrupts. The achievable speedup is about 7.7. There are two major parameters responsible for this drop in performance: the cost of interrupts and the bandwidth of the I/O bus. Making the interrupt cost 0 results in a speedup of 11, while increasing the I/O bus bandwidth to the memory bus bandwidth gives a speedup of 10. Modifying both parameters at the same time gives a speedup almost the same as the best speedup. LU: The best speedup is 13.7. The difference from the ideal speedup is due to load imbalances in communication and due to barrier cost. The achievable speedup for LU is about the same as the best speedup, since this application has very low communication to computation ratio, so communication is not the problem. Ocean: The best speedup for Ocean is 10.55. The reason for this is that when the interrupt cost is 0 an anomaly is observed in first touch page allocation and the speedup is very low due to a large number of MBytes/MHz13579111315Speedup FFT LU-contiguous Ocean-contiguous Water-nsquared Water-spatial Radix Volrend Raytrace Barnes-space Barnes-rebuild Figure 8. Effects of I/O bandwidth on application performance. The data points for each application correspond to an I/O bandwidth of 2, 1, 0.5 and 0.25 MBytes per processor clock MHz, or 400, 200, 100, and 50 MBytes/s assuming a 200 MHz processor. page faults. The achievable speedup is 13.0, with the main cost being that of barrier synchronization. It is worth noting that speedups in Ocean are artificially high because of local cache effects: a processor's working set does not fit in cache on a uniprocessor, but does fit in the cache with processors. Thus the sequential version performs poorly due to the high cache stall time. Barnes-rebuild: The best speedup for Barnes-rebuild is 5.90. The difference from the ideal is because of page faults in the large number of critical sections (locks). The achievable speedup is 3.9. The difference between the best and achievable speedups in the presence of page faults is because synchronization wait time is even higher due to the increased protocol costs. These increased costs are mostly because of the host overhead (a loss of about 1 in the speedup) and the NI occupancy (about 0.8). To verify all these we disabled remote page fetches in the simulator so that all page faults appear to be local. The speedup becomes 14.64 in the best and 10.62 in the achievable cases respectively. The gap between the best and the achievable speedups is again due to host and NI overheads. Barnes-space: The second version of Barnes we run is an improved version with minimal locking [10]. The best speedup is 14.5, close to the ideal. The achievable speedup is 12.5. The difference between these two is mainly because of the lower available I/O bandwidth in the achievable case. This increases the data wait time in an imbalanced way. Water-Nsquared: The best speedup for Water-Nsquared is 9.9 and the achievable speedup is about 9. The reason for the not very high best speedup is page faults that occur in contended critical sections, LU(contiguous) Ocean(contiguous) Radix Raytrace Volrend Water(nsquared) Water(spatial)0.10.30.50.70.9Normalized units Slowdown due to I/O Bus Bandwidth (normalized to the largest slowdown) Number of Bytes sent/Processor/1M Compute Cycles (normalized to the largest) Figure 9. Relation between slowdown due to I/O Bus Bandwidth and Number of Bytes transferred. greatly increasing serialization at locks. If we artificially disable remote page faults the best speedup increases from 9.9 to 14.1. The cost for locks in this artificial case is very small and the non-ideal speedup is due to imbalances in the computation itself. Water-Spatial: The best speedup is 13.75. The difference from ideal is mainly due to small imbalances in the computation and lock wait time. Data wait time is very small. The achievable speedup is about 13.3. Radix: The best speedup for Radix is 7. The difference from the ideal speedup of 16.1 is due to data wait time, which is exaggerated by contention even at the best parameter values, and the resulting imbalances among processors which lead to high synchronization time. The imbalances are observed to be due to contention in the network interface. The achievable speedup is only 3. The difference from the best speedup is due to the same factors: data wait time is much higher and much more imbalanced due to much greater contention effects. The main parameter responsible for this is I/O bus bandwidth. For instance, if we quadruple I/O bus bandwidth the achievable speedup for Radix becomes 7, just like the best speedup. Raytrace: Raytrace performs very well. The best speedup is 15.64 and the achievable speedup 14.80. FFT LU-contiguous Ocean-contiguous Water-nsquared Water-spatial Radix Volrend Raytrace Barnes-space Barnes-rebuild Figure 10. Effects of interrupt cost on application performance. The six bars for each application correspond to an interrupt cost of 0, 500, 1000, 2500, 5000, 10000, and 50000 processor cycles. Volrend: The best speedup is 10.95. The reason for this low number is imbalances in the computation itself due to the cost of task stealing, and large lock wait times due to page faults in critical sections. If we artificially eliminate all remote page faults, then computation is perfectly balanced and synchronization costs are negligible (speedup is 14.9 in this fictional case). The achievable speedup is 9.40, close to the best speedup. We see that the difference between ideal and best performance is due to page faults that occur in critical sections, I/O bandwidth limitations and imbalances in the communication and computation, and the difference between best and achievable performance is primarily due to the interrupt cost and I/O bandwidth limitations and less due to the host overhead. Overall, application performance on SVM systems today appears to be limited primarily by interrupt cost, and next by I/O bus bandwidth. Host overhead and NI occupancy per packet are substantially less significant, and in that order. 7 Page Size and Degree of Clustering In addition to the performance parameters of the communication architecture discussed above, the granularities of coherence and data transfer-i.e. the page size-and the number of processors per node are two other important parameters that affect the behavior of the system. They play an important role in determining the amount of communication that takes place in the system, the cost of which is then determined by the performance parameters. Page Size: The page size in the system is important for many reasons. It defines the size of the transfers, since in all software protocols data fetches are performed at page sizes. It also affects the LU(contiguous) Ocean(contiguous) Radix Raytrace Volrend Water(nsquared) Water(spatial)0.10.30.50.70.9Normalized units Slowdown due to Interrupt cost (normalized to the largest slowdown) Number of Page Fetches and Remote Lock Acquires (normalized to the largest) Figure 11. Relation between slowdown due to Interrupt cost and Number of Page Fetches and Remote Lock Acquires. amount of false sharing in the system, which is very important for SVM. These two aspects of the page size conflict with each other: bigger pages reduce the number of messages in the system if spatial locality is well exploited in communication, but they increase the amount of false sharing, and vice versa. Moreover, different page sizes lead to different amounts of fragmentation in memory, which may result in wasted resources. Figure 13 shows that the effects of page size on applications vary a lot. Most applications seem to favor smaller page sizes, with the exception of Radix that benefits a lot from bigger pages. We vary the page size between 2 KBytes and 32 KBytes pages. Most systems today support either 4 KBytes or 8 KBytes pages. We should note two caveats in our study with respect to page size. First, we did not tune the applications specifically to the different page sizes. Second, the effects of the page size are often related to the problem sizes that are used. For applications in which the amount of false sharing and fragmentation (i.e. the granularity of access interleaving in memory from different processors) changes with problem size, larger problems that run on real systems may benefit from larger pages (i.e. FFT). Size: The degree of clustering is the number of processors per node. Figure 14 shows that for most applications greater clustering helps even if the memory configuration and bandwidths are kept the same 1 . We use cluster sizes of 1, 4, 8 and 16 processors, always keeping the total number of processors This assumption, of keeping the memory subsystem the same and increasing the number of processors per node is not very realistic, since systems with higher degrees of clustering usually have a more aggressive memory subsystem as well, FFT LU-contiguous Ocean-contiguous Water-spatial Volrend Raytrace Barnes-rebuild Figure 12. Effects of network interface occupancy on application performance for AURC. The data points for each application correspond to a network occupancy of 50, 250, 500, 1000, 2000, and 10000 processor cycles. in the system at 16. These configurations cover the range from a uniprocessor node configuration to a cache-coherent, bus-based multiprocessor. Typical SVM systems today use either uniprocessor or 4-way nodes. A couple of interesting points emerge. First, unlike most applications, for Ocean-contiguous the optimal clustering is four processors per node. The reason is that Ocean-contiguous generates a lot of local traffic on the memory bus due to capacity and conflict misses, and more processors on the bus exacerbate this problem. On the other hand, Ocean-contiguous benefits a lot from clustering because of the communication pattern. Thus when four processors per node are used, the performance improvement over one processor per node comes from sharing. When the system has more than four processors per node, the memory bus is saturated and although the system benefits from sharing, performance is degrading because of memory bus contention. Radix and FFT also put greatly increased pressure on the shared bus. The cross-node SVM communication however, is very high and the reduction in it via increased spatial locality at page grain due to clustering outweighs this problem. The second important point is that the applications that perform very poorly under SVM do very well on a shared bus system at this scale. The reason is that these applications either exhibit a lot of synchronization or make fine grain accesses, both of which are much cheaper on a hardware-coherent shared bus architecture. For example, applications where the problem in SVM is page faults within critical sections (i.e. Barnes- rebuild) perform much better on this architecture. These results show that bus bandwidth is not the most significant problem for these applications at this scale, and the use of hardware coherence and synchronization outweighs the problems of sharing a bus. and are likely to provide greater node-to-network bandwidth. Application Best Achievable Ideal FFT 13.5 7.7 16.2 Ocean 10.5 13.0 16.0 Water(nsquared) 9.9 9.0 15.8 Water(spatial) 13.7 13.3 15.8 Radix 7.0 3.0 16.1 Volrend 10.9 9.40 15.4 Raytrace 15.6 14.8 16.4 Barnes(rebuild) 5.9 3.9 15.4 Barnes(space) 14.5 12.5 15.6 Table 4. Best and Achievable Speedups for each application 8 Related Work Our work is similar in spirit to some earlier studies, conducted in [15, 7], but in different context. In [15], the authors examine the impact of communication parameters on end performance of a network of workstations with the applications being written in Split-C on top of Generic Active Messages. They find that application performance demonstrates a linear dependence on host overhead and on the gap between transmissions of fine grain messages. For SVM, we find these parameters to not be so important since their cost is usually amortized over page granularity. Applications were found to be quite tolerant to latency and bulk transfer bandwidth in the split-C study as well. In [7], Holt et al. find that the occupancy of the communication controller is critical to good performance in DSM machines that provide communication and coherence at cache line granularity. Overhead is not so significant there (unlike in [15]) since it is very small. In [11], Karlsson et al. find that the latency and bandwidth of an ATM switch is acceptable in a clustered SVM architecture. In [13] a Lazy Release Consistency protocol for hardware cache-coherence is presented. In a very different context, they find that applications are more sensitive to the bandwidth than the latency component of communication. Several studies have also examined the performance of different SVM systems across multiprocessor nodes and compared it with the performance of configurations with uniprocessor nodes. Erlichson et al. [6] find that clustering helps shared memory applications. Yeung et al. in [23] find this to be true for SVM systems in which each node is a hardware coherent DSM machine. In [1], they find that the same is true in general for all software SVM systems, and for SVM systems with support for automatic write propagation. 9 Discussion and Future Work This work shows that there is room for improving SVM cluster performance in various directions: ffl Interrupts. Since reducing the cost of interrupts in the system can improve performance signifi- cantly, an important direction for future work is to design SVM systems that reduce the frequency FFT LU-contiguous Ocean-contiguous Water-nsquared Water-spatial Radix Volrend Raytrace Barnes-space Barnes-rebuild Figure 13. Effects of page size on application performance. The data points for each application correspond to a page size of 2 KBytes, 4 KBytes, 8 KBytes, 16 KBytes, and KBytes. and/or the cost of interrupts. Polling, better operating system support, or support for remote fetches that do not involve the remote processor are mechanisms that can help in this direction. Operating system and architectural support for inexpensive interrupts would improve system performance. Unfortunately this is not always achieved, especially in commercial systems. In these cases, protocol modifications (using non-interrupting remote fetch operations) or implementation optimizations (using polling instead of interrupts) can improve system performance and lead to more predictable and portable performance across different architectures and operating systems. Polling can be done either by instrumenting the applications or (in SMP systems) by reserving one processor for protocol processing. Recent results for interrupts versus polling in SVM systems vary. One study finds that polling may add a significant overhead, leading to inferior performance than interrupts for page grain SVM systems [24]. On the other hand, Stets et al. find that polling gives generally better results than interrupts [20]. We believe more research is needed on modern systems to understand the role of polling. Another interesting direction that we are exploring is moving some of the protocol processing itself to the network processor found in programmable network interfaces like such as Myrinet, thus reducing the need for interrupting the main processor. ffl System bandwidth. Providing high bandwidth is also important, to keep up with increasing processor speeds. Although fast system interconnects are available, software performance is, in practice, rarely close to what the hardware provides. Low level communication libraries fail to deliver close to raw hardware performance in many cases. Further work on low level communication interfaces may also be helpful in providing low-cost, high-performance SVM systems. Multiple network interfaces per node is another approach that can increase the available bandwidth. In this case protocol changes may be necessary to ensure proper event ordering. FFT LU-contiguous Ocean-contiguous Water-nsquared Water-spatial Radix Volrend Raytrace Barnes-space Barnes-rebuild Figure 14. Effects of cluster size on application performance. The data points for each application correspond to a cluster size of 1, 4, 8, and 16 processors per node. ffl Clustering. Up to the scale we examined, adding more processors per node helps in almost all cases. In applications where performance does not increase quickly with the cluster size, scaling of other system parameters, such as memory bus and I/O bandwidth, can have the desirable effects. ffl Applications. In doing this work we found that restructuring applications is an area that can make a big difference. Understanding how an application behaves and restructuring it properly can dramatically improve performance far beyond the improvement in system parameters or protocols [10]. This however, is not always easy, and, unfortunately, not many tools are available in parallel systems to help easily discover the cause of bottlenecks and obtain insight about application restructuring needs, especially when contention is a major problem as it often is in commodity-based communication architectures. Architectural simulators are one of the few tools that can currently be used to understand how an application behaves in detail. We should point out that this work is limited to a certain family of home-based SVM protocols. Other systems-for instance fine grain SVM systems-may exhibit different behavior and dependencies on communication parameters. Similar studies for other protocols and architectures can help us understand better the differences and similarities among SVM systems. This work was based on a 16 processor system. To address the question of what happens in bigger systems we run some experiments with a processor configuration and compared the number of protocol events between the two configurations. Table 5 shows the ratios of protocol events and communication traffic between a 32 and a 16 processor configuration. In most cases the event counts scale proportionally with the size of the system which leads us to believe that the results presented so far will hold for bigger configurations as well (at least up to 32 processors). Moreover, with larger problem sizes the problems related to the communication architecture are usually alleviated. However, more sophisticated Application Page Faults Page Fetches Remote Lock Acquires Local Lock Acquires Barriers MBytes Sent Messages Sent LU 1.94 2.53 1.86 2.00 1.90 12.90 2 3.66 Ocean 0.75 0.53 2.77 1.57 1.99 2.50 1.95 Water-nsquared 2.89 2.63 1.40 2.50 1.99 2.80 2.37 Water-spatial 1.85 2.05 1.68 2.26 1.98 2.00 2.08 Radix 1.83 2.43 2.70 4.10 1.99 2.19 2.38 Volrend Raytrace 2.08 2.08 1.33 2.40 2.00 2.08 1.83 Table 5. Ratios of protocol events for a 32 and a 16 processor configuration (4 processor per node). scaling models, that take into account the problem size, may be necessary for more detailed and accurate predictions. Another important question is how are these communication parameters going to scale with time. It seems that the parameters that closely follow hardware performance (host overhead, network interface occupancy, bandwidth) have more potential for getting better (relative to processor speeds) than interrupt cost which depends on the operating system and on special architectural support. Conclusions We have examined the effects of communication parameters to a family of SVM protocols. Through detailed architectural simulations of a cluster of SMPs and a variety of applications, we find that most applications are very sensitive to interrupt cost, and a few would benefit from improvements in band-width relative to processor speed as well. Unbalanced systems with relatively high interrupt costs and low I/O bandwidth can result in substantial losses in application performance. In these cases we observe slowdowns of more than 90% (a factor of 10 longer execution time). However, most applications are not sensitive to host overhead and network interface occupancy. Most regular applications can achieve very good SVM performance under the best configuration of parameters. For irregular applications, though, even this best performance can be low. This is mainly due to serialization effects in critical sections, i.e. due to page faults incurred inside critical sections, which dilate the critical sections and increase serialization. For example by reducing the amount of locking by using a different algorithm for parallel tree building, the performance of Barnes improves by a factor of 2-3. Overall, the achievable application performance today is limited primarily by interrupt cost and then by node to network bandwidth. Host overhead and NI occupancy appear less important to improve relative to processor speed. If interrupts are free and bandwidth high relative to the processor speed, then the achievable performance approaches the best performance in most cases. Acknowledgments We thank Hongzhang Shan for making available to us the improved version of Barnes, and the anonymous reviewers for their comments and feedback. --R Comparison of shared virtual memory across uniprocessor and SMP nodes. A virtual memory mapped network interface for the shrimp multicomputer. A gigabit-per-second local area network Design and implementation of virtual memory-mapped communication on myrinet Active messages: A mechanism for integrated communication and computation. The benefits of clustering in shared address space multiprocessors: An applications-driven investigation The effects of latency Improving release-consistent shared virtual memory using automatic update Understanding application performance on shared virtual memory. Application restructuring and performance portability on shared virtual memory and hardware-coherent multiprocessors Performance evaluation of cluster-based multiprocessor built from atm switches and bus-based multiprocessor servers Distributed shared memory on standard workstations and operating systems. Lazy release consistency for hardware-coherent multi- processors Effect of communication latency The Fast Messages (FM) 2.0 streaming interface. Tempest and typhoon: User-level shared memory Augmint: a multiprocessor simulation environment for intel x86 architectures. Design issues and tradeoffs for write buffers. Fast interrupt priority management in operating system kernels. Methodological considerations and characterization of the SPLASH-2 parallel application suite MGS: a multigrain shared memory system. Relaxed consistency and coherence granularity in DSM systems: A performance evaluation. Performance evaluation of two home-based lazy release consistency protocols for shared virtual memory systems --TR Active messages Virtual memory mapped network interface for the SHRIMP multicomputer Tempest and typhoon The benefits of clustering in shared address space multiprocessors Lazy release consistency for hardware-coherent multiprocessors MGS Understanding application performance on shared virtual memory systems Performance evaluation of two home-based lazy release consistency protocols for shared virtual memory systems Application restructuring and performance portability on shared virtual memory and hardware-coherent multiprocessors VM-based shared memory on low-latency, remote-memory-access networks Myrinet Design and Implementation of Virtual Memory-Mapped Communication on Myrinet Fast Interrupt Priority Management in Operating System Kernels Improving Release-Consistent Shared Virtual Memory using Automatic Update Performance Evaluation of a Cluster-Based Multiprocessor Built from ATM Switches and Bus-Based Multiprocessor Servers Design Issues and Tradeoffs for Write Buffers The Effects of Latency, Occupancy, and Bandwidth in Distributed Shared Memory Multiprocessors Effect of Communication Latency, Overhead, and Bandwidth on a Cluster --CTR Mainak Chaudhuri , Mark Heinrich , Chris Holt , Jaswinder Pal Singh , Edward Rothberg , John Hennessy, Latency, Occupancy, and Bandwidth in DSM Multiprocessors: A Performance Evaluation, IEEE Transactions on Computers, v.52 n.7, p.862-880, July Cheng Liao , Dongming Jiang , Liviu Iftode , Margaret Martonosi , Douglas W. Clark, Monitoring shared virtual memory performance on a Myrinet-based PC cluster, Proceedings of the 12th international conference on Supercomputing, p.251-258, July 1998, Melbourne, Australia Soichiro Araki , Angelos Bilas , Cezary Dubnicki , Jan Edler , Koichi Konishi , James Philbin, User-space communication: a quantitative study, Proceedings of the 1998 ACM/IEEE conference on Supercomputing (CDROM), p.1-16, November 07-13, 1998, San Jose, CA Angelos Bilas , Courtney R. Gibson , Reza Azimi , Rosalia Christodoulopoulou , Peter Jamieson, Using System Emulation to Model Next-Generation Shared Virtual Memory Clusters, Cluster Computing, v.6 n.4, p.325-338, October Angelos Bilas , Liviu Iftode , Jaswinder Pal Singh, Evaluation of hardware write propagation support for next-generation shared virtual memory clusters, Proceedings of the 12th international conference on Supercomputing, p.274-281, July 1998, Melbourne, Australia Angelos Bilas , Dongming Jiang , Jaswinder Pal Singh, Accelerating shared virtual memory via general-purpose network interface support, ACM Transactions on Computer Systems (TOCS), v.19 n.1, p.1-35, Feb. 2001 Sanjeev Kumar , Yitzhak Mandelbaum , Xiang Yu , Kai Li, ESP: a language for programmable devices, ACM SIGPLAN Notices, v.36 n.5, p.309-320, May 2001 Zoran Radovi , Erik Hagersten, Removing the overhead from software-based shared memory, Proceedings of the 2001 ACM/IEEE conference on Supercomputing (CDROM), p.56-56, November 10-16, 2001, Denver, Colorado Angelos Bilas , Cheng Liao , Jaswinder Pal Singh, Using network interface support to avoid asynchronous protocol processing in shared virtual memory systems, ACM SIGARCH Computer Architecture News, v.27 n.2, p.282-293, May 1999 Salvador Petit , Julio Sahuquillo , Ana Pont , David Kaeli, Addressing a workload characterization study to the design of consistency protocols, The Journal of Supercomputing, v.38 n.1, p.49-72, October 2006

distributed memory;bandwidth;latency;network occupancy;shared memory;host overhead;communication parameters;interrupt cost;clustering

509603

Compiling parallel code for sparse matrix applications.

We have developed a framework based on relational algebra for compiling efficient sparse matrix code from dense DO-ANY loops and a specification of the representation of the sparse matrix. In this paper, we show how this framework can be used to generate parallel code, and present experimental data that demonstrates that the code generated by our Bernoulli compiler achieves performance competitive with that of hand-written codes for important computational kernels.

Introduction Sparse matrix computations are ubiquitous in computational science. However, the development of high-performance software for sparse matrix computations is a tedious and error-prone task, for two reasons. First, there is no standard way of storing sparse matri- ces, since a variety of formats are used to avoid storing zeros, and the best choice for the format is dependent on the problem and the architecture. Second, for most algorithms, it takes a lot of code reorganization to produce an efficient sparse program that is tuned to a particular format. We illustrate these points by describing two formats - a classical format called Compressed Column Storage (CCS) [10] and a modern one used in the BlockSolve library [11] - which will serve as running examples in this abstract. CCS format is illustrated in Fig. 1. The matrix is compressed along the columns and is stored using three arrays: COLP, VALS and ROWIND. The values of the non-zero elements of each column j are stored in the array section 1)). The row indices for the non-zero elements of the column j are stored in 1)). This is illustrated in Fig. 1(b). If a matrix has many zero columns, then the zero columns are not stored, which results in what is called Compressed Compressed Column Storage format (CCCS). In this case, another level of indirection is added (the COLIND array) to compress the column dimension, as well (Fig. 1(c)). COLP VALS ROWIND COLP VALS ROWIND 3 6 8(a) An example matrix (b) CCS format (c) CCCS format Figure 1: Illustration of Compressed Column Storage format This is a very general and simple format. However, it does not exploit any application specific structure in the matrix. The format used in the BlockSolve library exploits structure present in sparse matrices that arise in the solution of PDE's with multiple degrees of freedom. Figure 2(a) (adapted from [11]) illustrates a grid that would arise from 2-D, linear, multi-component finite-element model with three degrees of freedom at each discretization point. The degrees of freedom are illustrated by the three dots at each discretization point. The stiffness matrix for such model would have groups of rows with identical column structure called i-nodes ("identical nodes"). Non-zero values for each i-node can be gathered into a dense matrix as shown in Fig. 2(c). Such matrices are also rich in cliques (a partition into cliques is shown in Fig. 2(a) using dashed rectangles). The library colors the contracted graph induced by the cliques and reorders the matrix as shown in Fig. 2(b). For symmetric matrices, only the lower half is stored together with the diagonal. Black triangles along the diagonal correspond to dense matrices induced by the cliques. Gray off-diagonal blocks correspond to sparse blocks of the matrix (stored using i-nodes). Notice that the matrix is stored as a collection of smaller dense matrices. This fact helps reduce sparse storage overhead and improve performance of matrix-vector products. For parallel execution, each color is divided among the processors. Therefore each processor receives several blocks of contiguous rows. On each processor, the off-diagonal blocks are actually stored by column (in column i-nodes). When performing a matrix-vector product, this storage organization makes the processing of messages containing non-local values of the vector more efficient. In addition, this allows the overlap of computation and communication by separating matrix-vector product into a portion which accesses only local data and one that deals with non-local data in incoming messages. The main algorithm we will consider in this paper is matrix-vector product which is the core computation in iterative solvers for linear systems. Consider the performance (in Mflops) of sparse matrix-vector product on a single processor of an IBM SP-2 for a variety of matrices and storage formats, shown in Table 1 (descriptions of the matrices and the formats can be found in Appendix A). Boxed numbers indicate the highest performance for a given matrix. It is clear from this set of experiments that there is no single format that is appropriate for all kinds of problems. This demonstrates the difficulty of developing a "sparse BLAS" for sparse matrix computations. Even if we limit ourselves to the formats in Table 1, one still has to provide at least 6 versions of sparse matrix-matrix product Color #1 Color #0 Color #2 (a) A subgraph generated by 2D linear finite element model (b) Color/clique reordering in the Block- Solve library a b f e a c d e f (c) I-node storage Figure 2: Illustration of the BlockSolve format (assuming that the result is stored in a single format)! The lack of extensibility in the sparse BLAS approach has been addressed by object-oriented solver libraries, like the PETSc library from Argonne [4]. Such libraries provide templates for a certain class of solvers (for example, Krylov space iterative solvers) and allow a user to add new formats by providing hooks for the implementations of some algebraic operations (such as matrix-vector product). However, in many cases the implementations of matrix-vector products themselves are quite tedious (as is the case in the BlockSolve library). Also, these libraries are not very useful in developing new algorithms. A radically different solution is to generate sparse matrix programs by using restructuring compiler technology. The compiler is given a dense matrix program with declarations about which matrices are actually sparse, and it is responsible for choosing appropriate storage formats and for generating sparse matrix programs. This idea has been explored by Bik and [6, 7], but their approach is limited to simple sparse matrix formats that are not representative of those used in high-performance codes. Intuitively, they trade the ability to handle a variety of formats for the ability to compile arbitrary loop nests. We have taken a different approach. Previously, we have shown how efficient sparse sequential code can be generated for a variety of storage formats for DOALL loops and loops with reductions [13, 14]. Our approach is based on viewing arrays as relations, and the execution of loop nests as evaluation of relational queries. We have demonstrated that our method of describing storage formats through access methods is general enough to specify Name Diagonal Coordinate CRS ITPACK JDiag BS95 small 21.972 8.595 16.000 7.446 21.818 2.921 685 bus 1.133 5.379 20.421 4.869 31.406 2.475 gr memplus 0.268 4.648 15.299 0.250 12.111 4.138 sherman1 Table 1: Performance of sparse matrix-vector product a variety of formats yet specific enough to allow important optimizations. Since the class of "DOANY" loops covers not only matrix-vector and matrix-matrix products, but also important kernels within high-performance implementations of direct solvers and incomplete preconditioners, this allows us to address the needs of a number of important applications. One can think of our sparse code generator as providing an extensible set of sparse BLAS codes, which can be used to implement a variety of applications, just like dense BLAS routines. For parallel execution, one need to specify how data and computation are partitioned. Such information (we call it distribution relation) can come in a variety of formats. Just as is the case with sparse matrix formats, the distribution relation formats are also application dependent. In the case of regular block/cyclic distributions the distribution relations can be specified by a closed-form formula. This allows ownership information to be computed at compile-time. However, regular distributions might not provide adequate load-balance in many irregularly structured applications. The HPF-2 standard [9] provides for two kinds of irregular distributions: generalized block and indirect. In generalized block distribution, each processor receives a single block of continuous rows. It is suggested in the standard that each processor should hold the block sizes for all processors - that is the distribution relation should be replicated. This permits ownership to be determined without communication. Indirect distributions are the most general: the user provides an array MAP such that the element MAP(i) gives the processor to which the ith row is assigned. The MAP array itself can be distributed a variety of ways. However, this can require communication to determine ownership of non-local data. The Chaos library [15] allows the user to specify partitioning information by providing the list of row indices assigned to each processor. The list of indices are transferred into distributed translation table which is equivalent to having a MAP array partitioned block- wise. This scheme is as general as the indirect scheme used in HPF-2 and it also requires communication to determine ownership and to build the translation table. As we have already discussed, the partitioning scheme used in the BlockSolve library is somewhat different. It is more general than the generalized block distribution provided by HPF-2, yet it has more structure than the indirect distribution. Furthermore, the distribution relation in the BlockSolve library is replicated, since each processor usually receives only a small number of contiguous rows. Our goal is to provide a parallel code generation strategy with the following properties: ffl The strategy should not depend on having a fixed set of sparse matrix formats. ffl It should not depend on having a fixed set of distributions. ffl The system should be extensible. That is it should be possible to add new formats without changing the overall code generation mechanism. ffl At the same time, the generality should not come at the expense of performance. The compiler must exploit structure available in sparse matrix and partitioning formats. To solve this problem we extend our relational approach to the generation of parallel sparse code starting from dense code, a specification of sparse matrix formats and data partitioning information. We view arrays as distributed relations and parallel loop execution as distributed query evaluation. In addition, different ways of representing partitioning information (regular and irregular) are unified by viewing distribution maps themselves as relations. Here is the outline of the rest of the paper. In Section 2, we outline our relational approach to sequential sparse code generation. In Section 3, we describe our sparse parallel code generation algorithm. In Section 4, we present experimental evidence of the advantages of our approach. Section 5 presents a comparison with previous work. Section 6 presents conclusions and ongoing work. Relational Model of Sparse Code Generation Consider the matrix-vector product Suppose that the matrix A and the vector x are sparse, and that the vector y is dense. To execute this code efficiently, it is necessary to perform only those iterations (i,j) for which and X(j) are not zero. This set of iterations can be described by the following set of constraints:! a (1) The first row represents the loop bounds. The constraints in the second row associate values with array indices: for example, the predicate A(i; j; a) constraints a to be the value of j). Finally, the constraints in the third row specify which iterations update Y with non-zero values. Our problem is to compute an efficient enumeration of the set of iterations specified by the constraints (1). For these iterations, we need efficient access to the corresponding entries in the matrices and vectors. Since the constraints are not linear and the sets being computed are not convex, we cannot use methods based on polyhedral algebra, such as Fourier-Motzkin elimination [2], to enumerate these sets. Our approach is based on relational algebra, and it models A, X and Y as relations (tables) that hold tuples of array indices and values. Conceptually, the relation corresponding to a sparse matrix contains both zero and non-zero values. We view the iteration space of the loop as a relation I of hi; ji tuples. Then we can write the first two rows of constraints from (1) as the following relational query (relational algebra notation is summarized in Appendix To test if elements of sparse arrays A and X are non-zero, we use predicates NZ(A(i; j)) and NZ(X(j)). Notice that because Y is dense, NZ(Y (i)) evaluates to true for all array . Therefore, the constraints in the third row of (1) can be now rewritten as: The predicate P is called the sparsity predicate. We use the algorithm of Bik and Wijshoff [6, 7] to compute the sparsity predicate in general. Using the definition of the sparsity predicate, we can finally write down the query which defines the indices and values in the sparse computation: (oe is the relational algebra selection operator.) We have now reduced the problem of efficiently enumerating the iteration points that satisfy the system of constraints (1) to the problem of efficiently computing a relational query involving selections and joins. This problem in turn is solved by determining an efficient order in which the joins in (4) should be performed and determining how each of the joins should be implemented. These decisions depend on the storage formats used for the sparse arrays. 2.1 Describing Storage Formats Following ideas from relational database literature [16, 20], each sparse storage format is described in terms of its access methods and their properties. Unlike database relations, which are usually stored as "flat" collections of tuples, most sparse storage formats have hierarchical structure, which must be exploited for efficiency. For example, the CCS format does not provide a way of enumerating row indices without first accessing a particular column. We use the following notation to describe such hierarchical structure of array indices: which means that for a given column index j we can access a set of hi; vi tuples of row indices and values of the matrix. The - operator is used to denote the hierarchy of array indices. For each term in the hierarchy (J and (I; V ) in the example), the programmer must provide methods to search and enumerate the indices at that level, and must specify the properties of these methods such as the cost of the search or whether the enumeration produces sorted output. These methods and their properties are used to determine good join orders and join implementations for each relational query extracted from the program, as described in [14]. This way of describing storage formats to the compiler through access methods and properties solves the extensibility problem: a variety of storage formats can be described to the compiler, and the compilation strategy does not depend on a fixed set of formats. For the details on how the formats are specified to the compiler, see [13]. indices. Permutations and other kinds of index translations can be easily incorporated into our framework. Suppose we have a permutation P which is stored using two integer arrays: PERM and IPERM - which represent the permutation and its inverse. We can view P as a relation of tuples hi; i 0 i, where i is the original index and i 0 is the permuted index. Now suppose that rows of the matrix in our example have been permuted using P . Then we can view A as relation of hi tuples and the query for sparse matrix-vector product is: where the sparsity predicate is: 2.3 Summary Here are the highlights of our ffl Arrays (sparse and dense) are relations ffl Access methods define the relation as a view of the data structures that implement a particular format ffl We view loop execution as relational query evaluation ffl The query optimization algorithm only needs to know the high-level structure of the relations as provided by the access methods and not the actual implementation (e.g. the role of the COLP and ROWIND arrays in the CCS storage). ffl Permutations also can be handled by our compiler ffl The compilation algorithms are independent of any particular set of storage formats and new storage formats can be added to the compiler. Generating parallel code Ancourt et al. [1] have described how the problem of generating SPMD code for dense HPF programs can be reduced to the computation of expressions in polyhedral algebra. We now describe how the problem of generating sparse SPMD code for a loop nest can be reduced to the problem of evaluating relational algebra queries over distributed relations. Section 3.1 describes how distributed arrays are represented. Section 3.2 describes how a distributed query is translated into a sequence of local queries and communication statements. In Section 3.3 we discuss how our code generation algorithm is used in the context of the BlockSolve data structures. 3.1 Representing distributed arrays In the uniprocessor case, relations are high-level views of the underlying data structures. In the parallel case, each relation is a view of the partitions (or fragments) stored on each processor. The formats for the fragments are defined using access methods as outlined in Sec. 2.1. The problem we must address is that of describing distributed relations from the fragments. Let's start with the following simple example: ffl The matrix A is partitioned by row. Each processor p gets a fragment matrix A (p) . ffl Let i and j be the row and column indices of an array element in the original matrix, and let i 0 and j 0 be the corresponding indices in a fragment A (p) . Because the partition is by row, the the column indices are the same (j is the global row index, whereas i 0 can be thought of the local row offset. To translate between i and i 0 , each processor keeps an integer array IND (p) such that IND (p) (i 0 each processor keeps the list of global row indices assigned to it. How do we represent this partition? Notice that on each processor p the array IND (p) can be viewed as a relation IND (p) (i; i 0 ). The local fragment of the matrix can also be viewed as a relation: A (p) (i a). We can define the global matrix as follows: (The projection operator - is defined in the Appendix B.) In this case, each processor p carries the information that translates its own fragment A (p) into the contribution to the global relation. But there are other situations, when a processors other than p might own the translation information for the fragment stored on p. A good example is the distributed translation table used in the Chaos library [15]. Suppose that the global indices fall into the range be the number of processors. Let e. Then for a given global index i the index of the owner processor p and the the local offset i 0 are stored on processor Each processor q holds the array of hp; indexed by We need a general way of representing such index translation schemes. The key is to view the index translation relation itself as a distributed relation. Then, in the first case this global relation is defined as: In the example from the Chaos library, the relation is defined by: where IND (q) (h; is the view of the above mentioned array of hp; i 0 i tuples and the relation BLOCK(i; q; h) is the shorthand for the constraints in (8) and (9). Once we have defined the index translation relation IND(i; where IND can be defined by, for example, (10) or (11). Similarly, we can define the global relations X and Y for the vectors in the matrix-vector product (assuming they are distributed the same way as the rows of A): In general, distributed relations are described by: - a IND(a; where R is the distributed relation, R (p) is the fragment on processor p and IND is the global-to-local index translation relation. The index translation relation can be different for different arrays, but we assume that it always specifies a 1-1 mapping between the global index a and the pair hp; a 0 i. Notice that our example partitioning of the IND relation in (10) and (11) themselves satisfy definition (15). We call (15) the fragmentation equation. How do we specify the distribution of computation? Recall that the iteration set of the loop is also represented as a relation: I(i; j), in our matrix-vector product example. We A (p) Access methods Fragmentation COLP Global relations VALS Low-level data structures A ROWIND A (p) Bernoulli Compiler HPF A Distributed arrays Alignment/Distribution +Compiler Figure 3: Flow of information in HPF and Bernoulli Compiler could require the user to supply the full fragmentation equation for I. But this would be too burdensome: the user would have to provide the local iteration set I (p) , but this set should really be determined by the compiler using some policy (such as the owner-computes rule). In addition, because the relation I is not stored, there is no need to allow multiple storage formats for it. Our mechanisms are independent of the policy used to determine the distribution relation for iterations; given any distribution relation IND, we can define the local iteration set by: I (p) (i This simple definition allows us to treat the iteration set relation I uniformly together with other relations in question. Notice that the fragmentation equation (15) is more explicit than the alignment-distri- bution scheme used in HPF. In the Bernoulli compiler global relations are described through a hierarchy of views: first local fragments are defined through access methods as the views of the low-level data structures. Then the global relations are defined as views of the local fragments through the fragmentation equation. In HPF, alignment and distribution provide the mapping from global indices to proces- sors, but not the full global-to-local index translation. Local storage layout (and the full index translation) is derived by the compiler. This removes from the user the responsibility for (and flexibility in) defining local storage formats. The difference in the flow of information between HPF and Bernoulli Compiler is illustrated in Fig. 3. By mistake, the user may specify inconsistent distribution relations IND. These incon- sistencies, in general, can only be detected at runtime. For example, it can only be verified at run-time if a user specified distribution relation IND in fact provides a 1-1 and onto map. This problem is not unique to our framework - HPF with value-based distributions [21] has a similar problem. Basically, if a function is specified by its values at run-time, its properties can only be checked at run-time. It is possible to generate a "debugging" version of the code, that will check the consistency of the distributions, but this is beyond the scope of this paper. 3.2 Translating distributed queries Let us return to the query for sparse matrix-vector product: The relations A, X and Y are defined by (12), (13) and (14). We translate the distributed query (17) into a sequence of local queries and communication statements by expanding the definitions of the distributed relations and doing some algebraic simplification, as follows. 3.2.1 General strategy In the distributed query literature the optimization problem is: find the sites that will evaluate parts of the query (17). In the context of, say, a banking database spread across branches of the bank, the partitioning of the relations is fixed, and may not be optimal for each query submitted to the system. This is why the choice of sites might be non-trivial in such ap- plications. See [20] for a detailed discussion of the general distributed query optimization problem. In our case, we expect that the placement of the relations is correlated with the query itself and is given to us by the user. In particular, the placement of the iteration space relation I tells us where the query should be processed. That is the query to be evaluated on each processor p is: I (p) (i; where I (p) is the set of iterations assigned to processor p. We resolve the references to the global relations A, X and Y by, first, exploiting the fact that the join between some of them (in this case A and Y ) do not require any communication at all and can be directly translated into the join between the local fragments. Then, we resolve the remaining references by computing communication sets (and performing the actual communication) for other relations (X in our example). We now outline the major steps. 3.2.2 Exploiting collocation In order to expose the fact that the join between A and Y can be done without communi- cation, we expand the join using the definitions of the relations: Because we have assumed that the index translation relation IND provides a 1-1 mapping between global index and processor numbers, we can deduce that q. This is nothing more than the statement of the fact that A and Y are aligned [3, 5]. So the join between A and Y can be translated into: A (p) (i Notice that the join on the global index i has been translated into the join on the local offsets . The sparsity predicate P originally refers to the distributed relations: In the translated query, we replace the references to the global relations with the references to the local relations. 3.2.3 Generating communication The query: Used (p) A (p) (i computes the set of global indices j of X that are referenced by each processor. The join of this set with the index translation relation will tell us where to get each element: RecvInd (p) Used (p) This tells us which elements of X must be communicated to processor p from processor q. If the IND relation is distributed (as is the case in the Chaos library), then evaluation of the query (22) might itself require communication. This communication can also be expressed and computed in our framework by applying the parallel code generation algorithm recursively. 3.2.4 Summary Here is the summary of our We represent distributed arrays as distributed relations. We represent global-to-local index translation relations as distributed relations. We represent parallel DOANY loop execution as distributed query evaluation. ffl For compiling dense HPF programs, Ancourt et al. [1] describe how the computation sets, communication sets etc. can be described by expressions in polyhedral algebra. We derive similar results for sparse programs, using relational algebra. 3.3 Compiling for the BlockSolve formats. As was discussed in the introduction, the BlockSolve library splits the matrix into two disjoint data structures: the collection of dense matrices along the diagonal, shown using black triangles in Figure 2(b), and the off-diagonal sparse portions of the matrix stored using i-node format (Figure 2(c)). In the computation of a matrix-vector product x the dense matrices along the diagonal refer only to the local portions of the vector x. Also the off-diagonal sparse blocks are stored in a way that makes it easy to enumerate separately over those elements of the matrix that refer only to the local elements of x and over those that require communication. Altogether, we can view a matrix A stored in the BlockSolve library format as a sum AD +A SL +A SNL , where: ffl AD represents the dense blocks along the diagonal ffl A SL represents the portions of the sparse blocks that refer to local elements of x ffl A SNL represents the portions of the sparse blocks that refer to non-local elements of x AD , A SL and A SNL are all partitioned by row. The distribution in the library assigns a small number of continuous rows to each processor. The distribution relation is also replicated, thus reducing the cost of computing the ownership information. The hand-written library code does not have to compute any communication sets or index translations for the products involving AD and A SL - these portions of the matrix access directly the local elements of x. How can we use our code generation technology to produce code competitive with the hand-written code? The straight-forward approach is to start from the sequential dense matrix data-parallel program for matrix-vector product. Since the matrix is represented as three fragments (AD , A SL and A SNL ), our approach essentially computes three matrix vector products: The performance of this code is discussed in the next section. Careful comparison of this code with the handwritten code reveals that the performance of our code suffers from the fact that even though the products involving AD and A SL do not require any communication, they still require global-to-local index translation for the elements of x that are used in the computation. If we view AD and A SL as global relations that stored global row and column indices, then we hide the fact that the local indices of x can be determined directly from the data structures for AD and A SL . This redundant index translation introduces extra level of indirection in the accesses to x and degrades node program performance. At this point, we have no automatic approach to handling this problem. We can however circumvent the problem at the cost of increasing the complexity of the input program by specifying the code for the products with AD and A SL at the node program level. The code for the product with A SNL is still specified at the global (data-parallel) level: local: y local: y where y (p) , etc are the local portions of the arrays and y, A SNL and x are the global views. The compiler then generates the necessary communication and index translations for the product with A SNL . This mixed specification (both data-parallel and node level programs) is not unique to our approach. For example, HPF allows the programmer to "escape" to the node program level by using extrinsics [9]. In general, sophisticated composite sparse formats, such as the one used in the BlockSolve library, might require algorithm specification at a different level than just a dense loop. We are currently exploring ways of specifying storage formats so that we can get good sequential performance without having to drop down to node level programs for some parts of the application. 4 Experiments In this section, we present preliminary performance measurements on the IBM SP-2. The algorithm we studied is a parallel Conjugate Gradient [18] solver with diagonal preconditioning (CG), which solves large sparse systems of linear equations iteratively. Following the terminology from Chaos project, the parallel implementation of the algorithm can be divided into the inspector phase and the executor phase [15]. The inspector determines the the set of values to be communicated and performs some other preprocessing. The executor performs the actual computation and communication. In iterative applications the cost of the inspector can usually be amortized over several iterations of the executor. In order to verify the quality of the compiler-generated code and to demonstrate the benefit of using the mixed local/global specification (24) of the algorithm in this application we have measured the performance of the inspector and the executor in the following implementations of the CG algorithm: ffl BlockSolve is the hand-written code from the BlockSolve library. ffl Bernoulli-Mixed is the code generated by the compiler starting from the mixed lo- cal/global specification in (24). ffl Bernoulli is the "naive" code generated by the compiler starting from fully data-parallel specification (23). We ran the different implementations of the solver on a set of synthetic three-dimensional grid problems. The connectivity of the resulting sparse matrix corresponds to a 7-point stencil with 5 degrees of freedom at each discretization point. Then, we ran the solver on 2, 4, 8, 16, 32 and 64 processors of the IBM SP-2 at Cornell Theory Center. During each run we kept the problem size per processor constant at \Theta 30. This places 135 \Theta 10 3 rows with about 4:5 \Theta 10 6 non-zeroes total on each processor. We limited the number of solver iterations to 10. Tab. 2 shows the times (in seconds) for the numerical solution phase (the executor). Tab. 3 shows the overhead of the inspector phase as the ratio of the time taken by the inspector to the time taken by a single iteration of the executor. The comparative performance of the Bernoulli-Mixed and BlockSolve versions verifies the quality of the compiler generated code. The 2-4% difference is due to aggressive BlockSolve Bernoulli-Mixed Bernoulli sec sec diff. sec diff. Table 2: Numerical computation times (10 iterations) BlockSolve Bernoulli-Mixed Bernoulli Indirect-Mixed Indirect Table 3: Inspector overhead overlapping of communication and computation done in the hand-written code. Currently, the Bernoulli compiler generates simpler code, which first exchanges the non-local values of x and then does the computation. While the inspector in Bernoulli-Mixed code is about twice as expensive as that in the BlockSolve code, its cost is still quite negligible (2:7% of the executor with 10 iterations). The comparison of the Bernoulli and Bernoulli-Mixed code illustrates the importance of using the mixed local/global specification (24). The Bernoulli code has to perform redundant work in order to discover that most of the references to x are in fact local and do not require communication. The amount of this work is proportional to the problem size (the number of unknowns) and is much larger than the number of elements of x that are actually communicated. As the result, the inspector in the Bernoulli code is an order of magnitude more expensive than the one in the BlockSolve or Bernoulli-Mixed implementations. The performance of the executor also suffers because of the redundant global-to-local translation, which introduces an extra level of indirection in the final code even for the local references to x. As the result, the executor in Bernoulli code is about 10% slower than in the Bernoulli-Mixed code. To demonstrate the benefit of exposing structure in distribution relations, we have measured the inspector overhead for using the indirect distribution format from the HPF-2 standard [9]. We have implemented two versions of the inspectors using the support for the indirect distribution in the Chaos library [15]: ffl Indirect-Mixed is the inspector for the mixed local/global specification of (24). ffl Indirect is the inspector for the fully data parallel specification. Tab. 3 shows the ratio of the time taken by the Indirect-* inspectors to the time taken by the single iteration of the Bernoulli-* executors - the executor code is exactly the same in both cases and we have only measured the executors for the Bernoulli-* implementations. The order of magnitude difference between the performance of Indirect-Mixed and Bernoulli-Mixed inspectors is due to the fact that the Indirect-Mixed inspector has to perform asymptotically more work and requires expensive communication. Setting up the distributed translation table in the Indirect-Mixed inspector, which is necessary to resolve non-local references, requires the round of all-to-all communication with the volume proportional to the problem size (i.e. the number of unknowns). Additionally, querying the translation table (in order to determine the ownership information) again requires all-to-all communication: for each global index j the processor queried for the ownership information - even though the communication pattern for our problems has limited "nearest-neighbor" connectivity. The difference between Indirect and Bernoulli inspectors is not as pronounced - the number of references that has to be translated is proportional to the problem size. Still, the Indirect inspector has to perform all-to-all communication to determine the ownership of the non-local data. The relative effect of the inspector performance on the overall solver performance de- pends, of course, on the number of iterations taken by the solver, which, in turn, depends on the condition number of the input matrix. To get a better idea of the relative performance of the Bernoulli-Mixed and Indirect-Mixed implementation for a range of problems we have plotted in Fig. 4 the ratios of the time that the Indirect-Mixed implementation would take to the time that the Bernoulli-Mixed implementation would take on 8 and 64 processors for a range of iteration counts 5 - k - 100. The lines in Fig. 4 plot the values of the ratio: is the inspector overhead for the Bernoulli-Mixed version, r I is inspector overhead for the Indirect-Mixed version and k is the iteration count. A simple calculation shows that it would take 77 iterations of an Indirect-Mixed solver on 64 processors to get within 10% of the performance of the Bernoulli-Mixed. On 8 processors the number is 43 iterations. To get within 20% it would take 21 and 39 iteration on 8 and 64 processors, respectively. These data demonstrate that, while the inspector cost is somewhat amortized in an iterative solver, it is still important to exploit the structure in distribution relations - it can lead to order of magnitude savings in the inspector cost and improves the overall performance of the solver. It should also be noted that the Indirect-Mixed version is not only slower than the two Bernoulli versions but also requires more programming effort. Our compiler starts with the specification at the level of dense loops both in (23) and (24), whereas an HPF compiler needs sequential sparse code as input. For our target class of problems - sparse DOANY loops - our approach results in better quality of parallel code while reducing programming effort. Indirect-mixed Bernoulli-mixed Number of iterations Figure 4: Effect of problem conditioning on the relative performance 5 Previous work The closest alternative to our work is a combination of Bik's sparse compiler [6, 7] and the work on specifying and compiling sparse codes in HPF Fortran [19, 21, 22]). One could use the sparse compiler to translate dense sequential loops into sparse loops. Then, the Fortran D or Vienna Fortran compiler can be used to compile these sparse loops. However, both Bik's work and the work done by Ujaldon et al. on reducing inspector overheads in sparse codes limits a user to the fixed set of sparse matrix storage and distribution formats, this reducing possibilities for exploiting problem-specific structure. 6 Conclusions We have presented an approach for compiling parallel sparse codes for user-defined data struc- tures, starting from DOANY loops. Our approach is based on viewing parallel DOANY loop execution as relational query evaluation and sparse matrices and distribution information as distributed relations. This relational approach is general enough to represent a variety of storage formats. However, this generality does not come at the expense of performance. We are able to exploit both the properties of the distribution relation in order to produce inexpensive inspectors, as well as produce quality numerical code for the executors. Our experimental evidence shows that both are important for achieving performance competitive with hand-written library codes. So far, we have focused our efforts on the versions of iterative solvers, such as the Conjugate Gradient algorithm, which do not use incomplete factorization preconditioners. The core operation in such solvers is the sparse matrix-vector product or the product of a sparse matrix and a skinny dense matrix. We are currently investigating how our techniques can be used in the automatic generation of high-performance codes for such operations as matrix factorizations (full and incomplete) and triangular linear system solution. --R A linear algebra framework for static hpf code distribution. Scanning polyhedra with do loops. Global optimizations for parallelism and locality on scalable parallel machines. Solving alignment using elementary linear algebra. Advanced compiler optimizations for sparse computations. Automatic data structure selection and transformation for sparse matrix computations. The Quality of Numerical Software: Assessment and Enhancement High Performance Fortran Forum. Computer Solution of Large Sparse Positive Definite Systems. users manual: Scalable library software for the parallel solution of sparse linear systems. Algorithm 586 ITPACK 2C: A FORTRAN package for solving large sparse linear systems by adaptive accelerated iterative methods. Compiling parallel sparse code for user-defined data structures A relational approach to sparse matrix compilation. Database Management Systems. Solving Elliptic Problems Using ELLPACK. Kyrlov subspace methods on supercomputers. New data-parallel language features for sparse matrix computations Principles of Database and Knowledge-Base Systems Distributed memory compiler design for sparse problems. --TR Solving elliptic problems using ELLPACK Principles of database and knowledge-base systems, Vol. I Krylov subspace methods on supercomputers Scanning polyhedra with DO loops Global optimizations for parallelism and locality on scalable parallel machines Runtime compilation techniques for data partitioning and communication schedule reuse Advanced compiler optimizations for sparse computations Automatic Data Structure Selection and Transformation for Sparse Matrix Computations Database management systems Matrix market Algorithm 586: ITPACK 2C: A FORTRAN Package for Solving Large Sparse Linear Systems by Adaptive Accelerated Iterative Methods Computer Solution of Large Sparse Positive Definite Distributed Memory Compiler Design For Sparse Problems Solving Alignment Using Elementary Linear Algebra --CTR Chun-Yuan Lin , Yeh-Ching Chung , Jen-Shiuh Liu, Efficient Data Compression Methods for Multidimensional Sparse Array Operations Based on the EKMR Scheme, IEEE Transactions on Computers, v.52 n.12, p.1640-1646, December Yuan Lin , David Padua, On the automatic parallelization of sparse and irregular Fortran programs[1]This work is supported in part by Army contract DABT63-95-C-0097; Army contract N66001-97-C-8532; NSF contract MIP-9619351; and a Partnership Award from IBM. This work is not necessarily representative of the positions or policies of the Army or Government., Scientific Programming, v.7 n.3-4, p.231-246, August 1999 Roxane Adle , Marc Aiguier , Franck Delaplace, Toward an automatic parallelization of sparse matrix computations, Journal of Parallel and Distributed Computing, v.65 n.3, p.313-330, March 2005 Chun-Yuan Lin , Yeh-Ching Chung, Data distribution schemes of sparse arrays on distributed memory multicomputers, The Journal of Supercomputing, v.41 n.1, p.63-87, July 2007 Eun-Jin Im , Katherine Yelick , Richard Vuduc, Sparsity: Optimization Framework for Sparse Matrix Kernels, International Journal of High Performance Computing Applications, v.18 n.1, p.135-158, February 2004 Chun-Yuan Lin , Jen-Shiuh Liu , Yeh-Ching Chung, Efficient Representation Scheme for Multidimensional Array Operations, IEEE Transactions on Computers, v.51 n.3, p.327-345, March 2002 Chun-Yuan Lin , Yeh-Ching Chung , Jen-Shiuh Liu, Efficient Data Parallel Algorithms for Multidimensional Array Operations Based on the EKMR Scheme for Distributed Memory Multicomputers, IEEE Transactions on Parallel and Distributed Systems, v.14 n.7, p.625-639, July

sparse matrix computations;parallelizing compilers

509605

Compiling stencils in high performance Fortran.

For many Fortran90 and HPF programs performing dense matrix computations, the main computational portion of the program belongs to a class of kernels known as stencils. Stencil computations are commonly used in solving partial differential equations, image processing, and geometric modeling. The efficient handling of such stencils is critical for achieving high performance on distributed-memory machines. Compiling stencils into efficient code is viewed as so important that some companies have built special-purpose compilers for handling them and others have added stencil-recognizers to existing compilers.In this paper we present a general compilation strategy for stencils written using Fortran90 array constructs. Our strategy is capable of optimizing single or multi-statement stencils and is applicable to stencils specified with shift intrinsics or with array-syntax all equally well. The strategy eliminates the need for pattern-recognition algorithms by orchestrating a set of optimizations that address the overhead of both intraprocessor and interprocessor data movement that results from the translation of Fortran90 array constructs. Our experimental results show that code produced by this strategy beats or matches the best code produced by the special-purpose compilers or pattern-recognition schemes that are known to us. In addition, our strategy produces highly optimized code in situations where the others fail, producing several orders of magnitude performance improvement, and thus provides a stencil compilation strategy that is more robust than its predecessors.

Introduction High-Performance Fortran (HPF)[14], an extension of Fortran90, has attracted considerable attention as a promising language for writing portable parallel programs. HPF offers a simple programming model shielding programmers from the intricacies of concurrent programming and managing distributed data. Programmers express data parallelism using Fortran90 array operations and use data layout directives to direct partitioning of the data and computation among the processors of a parallel machine. In many programs performing dense matrix computations, the main computational portion of the program belongs to a class of kernels known as stencils. For HPF to gain acceptance as a vehicle for parallel scientific programming, it must achieve high performance on this important class of problems. Compiling stencils into efficient code is viewed as so important that some companies have built special-purpose compilers for handling them [4, 5, 6] and others have added stencil-recognizers to existing HPF compilers [1, 2]. Each of these previous approaches to stencil compilation had significant limitations that restricted the types of stencils that they could handle. In this paper, we focus on the problem of optimizing stencil computations, no matter how they are instantiated by the programmer, for execution on distributed-memory archi- tectures. Our strategy orchestrates a set of optimizations that address the overhead of both intraprocessor and interprocessor data movement that results from the translation of For- array constructs. Additional optimizations address the issues of scalarizing array assignment statements, loop fusion, and data locality. In the next section we briefly discuss stencil computations and their execution cost on distributed-memory machines. In Section 3 we give an overview of our compilation strategy, and then discuss the individual optimizations. In Section 4 we present an extended example to show how our strategy handles a difficult case. Experimental results are given in Section 5, and in Section 6 we compare this strategy with other known efforts. Computations In this section we introduce stencil computations and give an overview of their execution cost on distributed-memory machines. We also introduce the normalized intermediate form which our compiler uses for all stencils. 2.1 Stencils A stencil is a stylized matrix computation in which a group of neighboring data elements are combined to calculate a new value. They are typically combined in the form of a sum of products. This type of computation is common in solving partial differential equations, image processing, and geometric modeling. The Fortran90 array assignment statement in Figure 1 is commonly referred to as a 5-point stencil. In this statement src and dst are arrays, and C1-C5 are either scalars or arrays. Each interior element of the result array dst is computed from the corresponding element of the source array src and the neighboring Figure 1: 5-point stencil computation. Figure 2: 9-point stencil computation. Figure 3: Problem 9 from the Purdue Set. elements of src on the North, West, South, and East. A 9-point stencil that computes all grid elements by exploiting the cshift intrinsic might be specified as shown in Figure 2. In the previous two examples the stencils were specified as a single array assignment statement, but this need not always be the case. Consider again the 9-point stencil above. If the programmer attempted to optimize the program by hand, or if the stencil was pre-processed by other optimization phases of the compiler, we might be presented with the code shown in Figure 3 1 . A goal of our work is to generate the same, highly-optimized code for all stencil computa- tions, regardless of how they have been written in HPF. For this reason, we have designed our optimizer to target the most general, normalized input form. All stencil and stencil-like computations can be translated into this normal form by factoring expressions and introducing temporary arrays. In fact, this is the intermediate form used by several distributed-memory compilers [18, 23, 3]. The normal form has several distinguishing characteristics: ffl cshift intrinsics and temporary arrays have been inserted to perform data movement needed for operations on array sections that have different processor mappings. This example was taken from Problem 9 of the Purdue Set [21] as adapted for Fortran D benchmarking by Thomas Haupt of NPAC. Figure 4: Intermediate form of 5-point stencil computation. ffl Each cshift intrinsic occurs as a singleton operation on the right-hand side of an array assignment statement and is only applied to whole arrays. ffl The expression that actually computes the stencil operates on operands that are perfectly aligned, and thus no communication operations are required. For example, given the 5-point stencil computation presented in Figure 1, the CM Fortran compiler would translate it into the sequence of statements shown in Figure 4. For the rest of this paper we assume that all stencil computations have been normalized into this form, and that all arrays are distributed in a block fashion. And although we concentrate on stencils expressed using the cshift intrinsic, the techniques presented can be generalized to handle the eoshift intrinsic as well. 2.2 Stencil Execution The execution of a stencil computation on a distributed-memory machine has two major components: the data movement associated with a set of cshift operations and the calculation of the sum of products. In the first phase of a stencil computation, all data movement associated with cshift operations is performed. We illustrate the data movement for a single cshift using an example. Figure 5 shows the effects of a cshift by -1 along the second dimension of a two-dimensional block-distributed array. When a cshift operation is performed on a distributed array, two major actions take place: 1. Data elements that must be shifted across processing element (PE) boundaries are sent to the appropriate neighboring PE. This is the interprocessor component of the shift. In Figure 5, the dashed lines represent this type of data movement, in this case the transfer of a column of data between neighboring processors. 2. Data elements shifted within a PE are copied to the appropriate locations in the destination array. This is the intraprocessor component of the shift. The solid lines in Figure 5 represent this data movement. Following data movement, the second phase of a stencil computation is the execution of a loop nest to calculate a sum of products. The loop nest for a stencil computation is Figure 5: constructed during compilation in two steps. First the compiler applies scalarization [24] to replace Fortran 90 array operations with a serial loop nest that operates on individual data elements. Next, the compiler transforms this loop nest into SPMD code [8]. The SPMD code is synthesized by reducing the loop bounds so that each PE computes values only for the data it owns. A copy of this transformed loop nest, known as the subgrid loop nest, executes on each PE of the parallel machine. Due to the nature of stencils which make many distinct array references, these subgrid loops can easily become memory bound. In such loops, the CPU must often sit idle while it waits for the array elements to be fetched from memory. 3 Compilation Strategy In this section we start with an overview of our compilation strategy, and then present the individual component optimizations. Given a stencil computation in normal form (as described in Section 2.1), we optimize it by applying a sequence of four optimizations. The first addresses the intraprocessor data movement associated with the cshift operations, eliminating it when possible. The second rearranges the statements into separate blocks of computation operations and communication operations. This optimizes the stencil by promoting loop fusion for the computation operations and it prepares the communication operations for further optimization by the following phase. Next, the interprocessor data movement of the cshift operations is optimized by eliminating redundant and partially-redundant communication. Finally, loop-level transformations are applied to optimize the computation. 3.1 Optimizing Intraprocessor Data Movement Intraprocessor data movement associated with shift intrinsics is completely eliminated when possible. This is accomplished by an optimization we call offset arrays [15]. This optimization determines when the source array (src) and the destination array (dst) of the cshift can share the same memory locations. If this is the case only the interprocessor data movement needs to occur. We exploit overlap areas [11] to receive the data that is copied between processors. After this has been accomplished, appropriate references to the destination array can be rewritten to refer to the source array with indices offset by the shift amount. The principal challenge then is to determine when the source and destination arrays can share storage. We have established a set of criteria to determine when it is safe and profitable to create an offset array. These criteria, and an algorithm used to verify them are described in detail elsewhere [15, 22]. In general, our approach allows the source and destination arrays of a shift operation to share storage between destructive updates to either array when the shift offset is a small constant. Once we have determined that the destination array of an assignment statement CSHIFT(SRC,SHIFT,DIM) may be an offset array, we perform the following transformations on the code. These transformations take advantage of the data that may be shared between the source array src and destination array dst and move only the required data between the PEs. First we replace the shift operation with a call to a routine that moves the off-processor data of SRC into an overlap area: CALL OVERLAP SHIFT(SRC,SHIFT,DIM). We then replace all uses of the array dst, that are reached from this definition, with a use of the array src. The newly created references to src carry along special annotations representing the values of shift and dim. Finally, when creating subgrid loops during the scalarization phase, we alter the subscript indices used for the offset arrays. The array subscript used for the offset reference to src is identical to the subscript that would have been generated for dst with the exception that the dim-th dimension has been incremented by the shift amount. The algorithm that we have devised for verifying the criteria and for performing the above transformations is based upon the static single assignment (SSA) intermediate representation [9]. The algorithm, after validating the use of an offset array at a shift operation, transforms the program and propagates that information in an optimistic manner. The propagation continues until there are no more references to transform or one of the criteria has been violated. When a criterion has been violated, it may be necessary to insert an array copy statement into the program to maintain its original semantics. The inserted copy statement performs the intraprocessor data movement that was avoided with the overlap shift. Due to the offset array algorithm's optimistic nature, it is able to eliminate intraprocessor data movement associated with shift operations in many difficult situations. In particular, it can determine when offset arrays can be exploited even when their definition and uses are separated by program control flow. This allows our stencil compilation strategy to eliminate intraprocessor data movement in situations where other strategies fail. 3.2 Statement Reordering We follow the offset array optimization with our context partitioning optimization [17]. This optimization partitions a set of Fortran90 statements into groups of congruent array statements 2 , scalar expressions, and communication operations. This assists the compilation of stencils in the following two ways: 1. First, by grouping congruent array statements together, we ensure that as subgrid loops are generated, via scalarization and loop fusion, as much computation as possible is placed within each loop without causing the loops to be over-fused [22]. Loops are over-fused when the code produced for the resulting parallel loops exhibits worse performance than the code for the separate parallel loops. Also, the structure of the subgrid loops produced is very regular. These characteristics increase the chances that loop transformations performed later are successful in exploiting data reuse and data locality. 2. Second, by grouping together communication operations, we simplify the task of reducing the amount of interprocessor data movement, which we discuss in the next subsection. To accomplish context partitioning, we use an algorithm proposed by Kennedy and M c Kinley [16]. While this algorithm was developed to partition parallel and serial loops into fusible groups, we use it to partition Fortran90 statements into congruence classes. The algorithm works on the data dependence graph (ddg)which must be acyclic. Since we apply it to a set of statements within a basic block, our dependence graph contains only loop-independent dependences and thus is acyclic. A complete description of our context partitioning algorithm is available elsewhere [17, 22], along with a discussion of its advantages for both SIMD and MIMD machines. Context partitioning is key to our ability to optimize multi-statement stencils as fully as single-statement stencils. No other stencil compilation strategy has this capability. 3.3 Minimizing Interprocessor Data Movement Once intraprocessor data movement has been eliminated and we have partitioned the statements into groups of congruent operations, we focus our attention on the interprocessor data movement that occurs during the calls to cshift. Due to the nature of offset arrays, we are presented with many opportunities to eliminate redundant and partially redundant data movement. We call this optimization communication unioning [22], since it combines a set of communication operations to produce a smaller set of operations. There are two key observations that allow us to find and eliminate redundant inter-processor data movement. First, shift operations, including overlap shift, are commutative Array statements are congruent if they operate on arrays with identical distributions and cover the same iteration space. Thus, for arrays that are shifted more than once, we can order the shift operations in any manner we like without affecting the result. Second, since all overlap shifts move data into the overlap areas of the subgrids, a shift of a large amount in a given direction and dimension may subsume all shifts of smaller amounts in the same direction and dimension. More formally, an overlap shift of amount i in dimension k is redundant if there exists an overlap shift of amount j in dimension k such that jjj - jij and Since we have already applied our context partitioning optimization to the program, we can restrict our focus to the individual groups of calls to overlap shift. To eliminate redundant data movement using communication unioning, we first use the commutative property to rewrite all the shifts for multi-offset arrays such that the overlap shifts for the lower dimensions occur first and are used as input to the overlap shifts for higher dimensions. We then reorder all the calls to overlap shift, sorting them by the shifted dimension, lowest to highest. We now scan the overlap shifts for the lowest dimension and keep only the largest shift amount in each direction. All others can be eliminated as redundant. Communication unioning then proceeds to process the overlap shifts for each higher dimension in ascending order by performing the following three actions: 1. We scan the overlap shifts for the given dimension to determine the largest shift amount in each direction. 2. We look for source arrays that are already offset arrays, indicating a multi-offset array. For these, we use the annotations associated with the source array to create an RSD to be used as an optional fourth argument in the call to overlap shift. The argument indicates those data elements from the adjacent overlap areas that should also be moved during the shift operation. Mapping the annotations to the RSD is simply a matter of adding the annotations to the corresponding RSD dimension; the annotation is added to the lower bound of the RSD if the shift amount is negative, otherwise it is added to the upper bound. As with shift amounts, larger RSDs subsume smaller RSDs. 3. We generate a single overlap shift in each direction, using the largest shift amount and including the RSD as needed - all other overlap shifts for that dimension can be eliminated. This procedure eliminates all redundant offset-shift communication, including partially redundant data movement associated with accessing "corner elements" of stencils. This algorithm is unique in that it is based upon the understanding and analysis of the shift intrinsics, rather than being based upon pattern-matching as is done in many stencil compilers. This optimization eliminates all communication for a shifted array, except for a single message in each direction of each dimension. The number of messages for the stencil is thus minimized. As an example, consider again the 9-point stencil computation that we presented in Figure 2. The original stencil specification required twelve cshift intrinsics. After applying communication unioning, only the four calls to overlap shift shown in Figure 6 are required. Figures 7-10 display the data movement that results from these calls. The figures contain a 5 \Theta 5 subgrid (solid lines) surrounded by its overlap area (dashed lines). Portions of the Figure of communication unioning for 9-point stencil. Figure 7: First half of 9-point stencil com- munication Figure 8: Result of communication operation adjacent subgrids are also shown. Figure 7 depicts the data movement specified by the first two calls. The result of that data movement is shown in Figure 8, where the overlap areas have been properly filled in. The data movement of the last two calls is shown in Figure 9. Notice how the last two calls pick up data from the overlap areas that were filled in by the first two calls, and thus they populate all overlap area elements needed for the subsequent computation, as shown in Figure 10. Figure 9: Second half of 9-point stencil communication Figure 10: Result of communication operation 3.4 Optimizing the Computation Finally, after scalarization has produced a subgrid loop nest, we can optimize it by applying a set of loop-level transformations designed to improve the performance of memory-bound programs. These transformations include unroll-and-jam, which addresses memory refer- ences, and loop permutation, which addresses cache references. Each of these optimize the program by exploiting reuse of data values. These optimizations are described in detail elsewhere [7, 19] and are not addressed in this paper. 4 An Extended Example In this section, we trace our compilation strategy through an extended example. This detailed examination shows how our strategy is able to produce code that matches or beats hand-optimized code. It also demonstrates how we are able to handle stencil computations that cause other methods to fail. For this exercise, we have chosen to use Problem 9 of the Purdue Set [21], as adapted for Fortran D benchmarking by Thomas Haupt of NPAC [20, 13]. The program kernel is shown in Figure 3. The arrays T, U, RIP, and RIN are all two-dimensional and have been distributed in a (block,block) fashion. This kernel computes a standard 9-point stencil, identical to that computed by the single-statement stencil shown in Figure 2. The reason it has been written in this fashion is to reduce memory requirements. Given the single-statement 9-point stencil, most Fortran90 compilers will generate 12 temporary arrays, one for each cshift. This greatly restricts the size of the problem that can be solved on a given machine. In contrast, the Problem 9 specification can be computed with only 3 temporary arrays since the live-ranges of the last 6 cshifts do not overlap. This reduces the temporary storage requirements by a factor of four! Additionally, the assignments of the cshifts into RIP and RIN perform a common subexpression elimination, removing four duplicate cshifts from the original specification of the stencil. Figure 11 shows a comparison of execution times for the single-statement cshift stencil in Figure 2 and the multi-statement Problem 9 stencil in Figure 3. The programs were compiled with IBM's xlhpf compiler and executed on a 4-processor SP-2 for varying problem sizes. As can be seen, the single-statement stencil specification exhausted the available memory for the larger problem sizes, even though each PE had 256Mbytes of real RAM. 4.1 Program Normalization We now step through the compilation of the stencil code in Figure 3 using the strategy presented in this paper. Figure 12 shows the stencil code after normalization. The six cshifts that are subexpressions in the assignment statements to array T are hoisted from the statements and assigned to compiler-generated temporary arrays. Since the live ranges of the temporary arrays do not overlap, a single temporary can be shared among all the statements. Alternatively, each cshift could receive its own temporary array - that would not affect the results of our stencil compilation strategy. Execution time (ms) Subgrid size (squared) exceeded memory CSHIFT specification Purdue Problem 9 Figure 11: Comparison of two 9-point stencil specifications. Figure Problem 9 after normalization. 4.2 Offset Array Optimization Once all shift operations have been identified and hoisted into their own assignment state- ments, we apply our offset array optimization. For this example, our algorithm determines that all the shifted arrays can be made into offset arrays. As can be seen in Figure 13, all the cshift operations have been changed into overlap shift operations, and references to the assigned arrays have been replaced with offset references to the source array U. All intraprocessor data movement has thus been eliminated. In addition, notice how the temporary arrays, both the compiler-generated TMP array Figure 13: Problem 9 after offset array optimization. Figure 14: Problem 9 after context partitioning optimization. and the user-defined RIP and RIN, are no longer needed to compute the stencil. If there are no other uses of these arrays in the routine, they need not be allocated. This reduction in storage requirements allows for larger problems to be solved on a given machine. 4.3 Context Partitioning Optimization After offset array optimization, we apply our context partitioning algorithm. This algorithm begins by determining the congruence classes present in the section of code. In this example there are only two congruence classes: the array statements, which are all congruent, and the communication statements. The dependence graph is computed next. There are only two types of dependences that exist in the code: true dependences from the overlap shift operations to the expressions that use the offset arrays, and the true and anti-dependences that exist between the multiple occurrences of the array T. Since all the Figure 15: Problem 9 after communication unioning optimization. dependences between the two classes are from statements in the communication class to statements in the congruent array class, the context partitioning algorithm is able to partition the statements perfectly into two groups. The result is shown in Figure 14. Since the array statements are now adjacent, scalarization will be able to fuse them into a single loop nest. Similarly, the communication statements are adjacent and communication unioning will be successful at its task. 4.4 Communication Unioning Optimization We now turn our attention to the interprocessor data movement specified in the overlap shift operations. As described in Section 3.3, we first exploit the commutativity of overlap shift operations and rewrite multi-dimensional overlap shifts so that the lower dimensions are shifted first. No rewriting is necessary for this example since all the dimension 1 shifts occur first, as can be seen in Figure 14. Next we look at the shifts across the first dimension. Since there is only a single shift of distance one in each direction, there is no redundant communication to eliminate. In the second dimension we again find only shifts of distance one. However, we discover four multi-offset arrays. Examining the annotations of the offset arrays, we create RSD's that summarize the overlap areas that are necessary. We generate the two calls to overlap shift that include the RSD's and then eliminate all other overlap shift calls for the second dimension. The resulting code is shown in Figure 15. Communication unioning has reduced the amount of communication to a minimum: a single communication operation for each dimension in each direction. 4.5 Scalarization and Memory Optimizations Figure shows the code after scalarization. The code now contains only 4 interprocessor communication operations, and no intraproceesor data movement is performed. Final transformations refine the loop bounds to generate a node program that only accesses the subgrids local to each PE. Our strategy has generated a single loop nest which, due to the nature of stencil computations, is ripe with opportunities for memory hierarchy optimization. We hand the final code to an optimizing node compiler that performs loop-level transformations such as scalar replacement and unroll-and-jam. ENDDO ENDDO Figure Problem 9 after scalarization. It is important to note that our strategy also produces the exact same code when given the single-statement 9-point stencil from Figure 2. This example shows how our stencil compilation algorithm is capable of fully optimizing stencils, no matter how they are instantiated by the programmer. 5 Experimental Results To measure the performance boost supplied by each step of our stencil compilation strategy, we ran a set of tests on a 4-processor IBM SP-2. We started by generating a naive translation of the Problem 9 test case into Fortran77+MPI. This is considered our "original" version. We then successively applied the transformations as outlined in the preceding section and measured the execution time. The results are shown in Figure 17. Before analyzing the results in Figure 17, it is worthwhile to compare them to the results shown in Figure 11 for the Problem 9 code. The performance of our "original" MPI version of the code for this example is already an order of magnitude faster than the code produced by IBM's xlhpf compiler: 0.475 seconds versus 4.77 seconds for the largest problem size. After applying our offset array optimization to the Fortran77+MPI test case as shown in Figure 13, execution time improves by 45%, equivalent to a speedup of 1.80. Next, after applying context partitioning, as shown in Figure 14, scalarization was able to merge all of the computation into a single loop nest, improving execution time an additional 31%. At this point, we have reduced the execution time of the original program by 62%, a speedup of 2.64. As shown in Figure 15, our communication unioning optimization eliminates four communication operations, which reduces the execution time by 41% when compared to the context-optimized version. Applying memory optimizations such as scalar replacement and unroll-and-jam further reduce the execution time another 14%. The execution time of the original program has been trimmed by 81%, equivalent to a speedup of 5.19. Comparing our code to the code produced by IBM's xlhpf compiler shows a speedup by a factor of 52! Lest someone think that we have chosen IBM's xlhpf compiler as a straw man, we have collected some additional performance numbers. We generated a third version of a 9-point Execution time (ms) Subgrid size (squared) Original program Offset arrays Context Partitioning Communication Unioning Memory Optimizations Figure 17: Step-wise results from stencil compilation strategy on Problem 9 when executed on an SP-2. stencil computation, this one using array syntax similar to the 5-point stencil shown in Figure 1. This 9-point stencil computation only computes the interior elements of the matrix; that is, elements 2:N-1 in each dimension. A graph comparing its execution time to the other two 9-point stencil specifications is given in Figure 18. The IBM xlhpf compiler was used in all cases. It is interesting to note that for the array syntax stencil the xlhpf compiler produced performance numbers that tracked our best performance numbers for all problem sizes except the largest, where we had a 10% advantage. It is important to note that the stencil compilation strategy that we have presented handles all three specifications of the 9-point stencil equally well. That is because our algorithm is based upon the analysis and optimization of the base constructs upon which stencils are built. Our algorithm is designed to handle the lowest common denominator - a form into which our compiler can transform all stencil computations. 6 Related Work One of the first major efforts to specifically address the compilation of stencil computations for a distributed-memory machine was the stencil compiler for the CM-2, also known as the convolution compiler [4, 5, 6]. The compiler eliminated intraprocessor data movement and optimized the interprocessor data movement by exploiting the CM-2's polyshift communication [10]. The final computation was performed by hand-optimized library microcode that took advantage of several loop transformations and a specialized register allocation scheme. Our general compilation methodology produces the same style code as this specialized compiler. We both eliminate intraprocessor data movement and minimize interprocessor Execution time (ms) Subgrid size (squared) exceeded memory CSHIFT specification Purdue Problem 9 Array syntax Figure 18: Comparison of three 9-point stencil specifications. data movement. Finally, our use of a loop-level optimizer to perform the unroll-and-jam optimization accomplish the same data reuse as the stencil compiler's ``multi-stencil swath''. The CM-2 stencil compiler had many limitations however. It could only handle single- statement stencils. The stencil had to be specified using the cshift intrinsic; no array-syntax stencils would be accepted. Since the compiler relied upon pattern matching, the stencil had to be in a very specific form: a sum of terms, each of which is a coefficient multiplying a shift expression. No variations were possible. And finally, the programmer had to recognize the stencil computation, extract it from the program and place it in its own subroutine to be compiled by the stencil compiler. Our compilation scheme handles a strict superset of patterns handled by the CM-2 stencil compiler. In their own words, they "avoid the general problem by restricting the domain of applicability." [6] We have placed no such restrictions upon our work. Our strategy optimizes single-statement stencils, multi-statement stencils, cshift intrinsic stencils, and array-syntax stencils all equally well. And since our optimizations were designed to be incorporated into an HPF compiler, they benefit those computations that only slightly resemble stencils. There are also some other commercially available compilers that can handle certain styl- ized, single-statement stencils. The MasPar Fortran compiler avoids intraprocessor data movement for single-statement stencils written using array notation. This is accomplished by scalarizing the Fortran90 expression (avoiding the generation of cshifts) and then using dependence analysis to find loop-carried dependences that indicate interprocessor data movement. Only the interprocessor data is moved, and no local copying is required. How- ever, the compiler still performs all the data movement for single-statement stencils written using shift intrinsics. This strategy is shared by many Fortran90/HPF compilers that focus on handling scalarized code. As with the CM-2 stencil compiler, our methodology is a strict superset of this strategy. Gupta, et al. [12], in describing IBM's xlhpf compiler, state that they are able to reduce the number of messages for multi-dimensional shifts by exploiting methods similar to ours. However, they do not describe their algorithm for accomplishing this, and it is unknown whether they would be able to eliminate the redundant communication that arises from shifts over the same dimension and direction but of different distances. The Portland Group's pghpf compiler, as described by Bozkus, et al. [1, 2], performs stencil recognition and optimizes the computation by using overlap shift communication. They also perform a subset of our communication unioning optimization. However, they are limited to single-statement expressions in both cases. In general, there have been several different methods for handling specific forms of stencil computations. Our strategy handles a more general form of stencil computations than these earlier methods. 7 Conclusion In this paper, we presented a general compilation scheme for compiling HPF stencil computations for distributed-memory architectures. The strategy optimizes such computations by orchestrating a unique set of optimizations. These optimizations eliminate unnecessary intraprocessor data movement resulting from cshift intrinsics, rearrange the array statements to promote profitable loop-fusion, eliminate redundant interprocessor data movement, and optimize memory accesses via loop-level transformations. The optimizations are general enough to be included in a general-purpose HPF/Fortran90 compiler as they will benefit many computations, not just those that fit a stencil pattern. The strength of these optimizations is that they operate on a normal form into which all stencil computations can readily be translated. This enables us to optimize all stencil computations regardless of whether they are written using array syntax or explicit shift in- trinsics, or whether the stencil is computed by a single statement or multiple statements. This approach is significantly more general than stencil compilation approaches in previous compilers. Even though we focused on the compilation of stencils for distributed-memory machines in this paper, the techniques presented are equally applicable to optimizing stencil computations on shared-memory and scalar machines (with the exception of reducing interprocessor movement). Acknowledgments This work has been supported in part by the IBM Corporation, the Center for Research on Parallel Computation (an NSF Science and Technology Center), and DARPA Contract DABT63-92-C-0038. This work was also supported in part by the Defense Advanced Re-search Projects Agency and Rome Laboratory, Air Force Materiel Command, USAF, under agreement number F30602-96-1-0159. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. The views and conclusions contained herein are those of the authors and should not be interpreted as representing the official policies or endorsements, either expressed or implied, of the Defense Advanced Research Projects Agency and Rome Laboratory or the U.S. Government. --R Techniques for compiling and executing HPF programs on shared-memory and distributed-memory parallel systems Compiling data parallel programs to message passing programs for massively parallel MIMD systems. A stencil compiler for the Connection Machine models CM-2/200 A stencil compiler for the Connection Machine model CM-5 Compiling Fortran 77D and 90D for MIMD distributed-memory machines Efficiently computing static single assignment form and the control dependence graph. communications software for the Connection Machine systems CM-2 and CM-200 Updating distributed variables in local computations. An HPF compiler for the IBM SP2. Low level HPF compiler benchmark suite. High Performance Fortran Forum. Optimizing Fortran 90 shift operations on distributed-memory multicomputers Context optimization for SIMD execution. Optimization techniques for SIMD Fortran compilers. Improving data locality with loop transfor- mations Applications benchmark set for Fortran-D and High Performance For- tran Problems to test parallel and vector languages. Optimizing Fortran90D/HPF for Distributed-Memory Computers A compiler for a massively parallel distributed memory MIMD computer. Optimizing Supercompilers for Supercomputers. --TR Updating distributed variables in local computations Fortran at ten gigaflops Efficiently computing static single assignment form and the control dependence graph Compiler optimizations for improving data locality POLYSHIFT communications software for the connection machine system CM-200 An HPF compiler for the IBM SP2 Improving data locality with loop transformations PGHPFMYAMPERSANDmdash;an optimizing High Performance Fortran compiler for distributed memory machines Optimizing Fortran90D/HPF for distributed-memory computers Optimizing Supercompilers for Supercomputers Optimizing Fortran 90 Shift Operations on Distributed-Memory Multicomputers --CTR David Wonnacott, Achieving Scalable Locality with Time Skewing, International Journal of Parallel Programming, v.30 n.3, p.181-221, June 2002 M. Kandemir, 2D data locality: definition, abstraction, and application, Proceedings of the 2005 IEEE/ACM International conference on Computer-aided design, p.275-278, November 06-10, 2005, San Jose, CA Hitoshi Sakagami , Hitoshi Murai , Yoshiki Seo , Mitsuo Yokokawa, 14.9 TFLOPS three-dimensional fluid simulation for fusion science with HPF on the Earth Simulator, Proceedings of the 2002 ACM/IEEE conference on Supercomputing, p.1-14, November 16, 2002, Baltimore, Maryland G. Chen , M. Kandemir, Optimizing inter-processor data locality on embedded chip multiprocessors, Proceedings of the 5th ACM international conference on Embedded software, September 18-22, 2005, Jersey City, NJ, USA Armando Solar-Lezama , Gilad Arnold , Liviu Tancau , Rastislav Bodik , Vijay Saraswat , Sanjit Seshia, Sketching stencils, ACM SIGPLAN Notices, v.42 n.6, June 2007 Steven J. Deitz , Bradford L. Chamberlain , Lawrence Snyder, Eliminating redundancies in sum-of-product array computations, Proceedings of the 15th international conference on Supercomputing, p.65-77, June 2001, Sorrento, Italy Gerald Roth , Ken Kennedy, Loop fusion in high performance Fortran, Proceedings of the 12th international conference on Supercomputing, p.125-132, July 1998, Melbourne, Australia A. Ya. Kalinov , A. L. Lastovetsky , I. N. Ledovskikh , M. A. Posypkin, Compilation of Vector Statements of C[] Language for Architectures with Multilevel Memory Hierarchy, Programming and Computing Software, v.27 n.3, p.111-122, May-June 2001 Mary Jane Irwin, Compiler-directed proactive power management for networks, Proceedings of the 2005 international conference on Compilers, architectures and synthesis for embedded systems, September 24-27, 2005, San Francisco, California, USA Zhang , Zhengqian Kuang , Baiming Feng , Jichang Kang, Auto-CFD-NOW: A pre-compiler for effectively parallelizing CFD applications on networks of workstations, The Journal of Supercomputing, v.38 n.2, p.189-217, November 2006 Daniel J. Rosenkrantz , Lenore R. Mullin , Harry B. Hunt III, On minimizing materializations of array-valued temporaries, ACM Transactions on Programming Languages and Systems (TOPLAS), v.28 n.6, p.1145-1177, November 2006

stencil compilation;high performance Fortran;communication unioning;statement partitioning;shift optimization

509611

Portable performance of data parallel languages.

A portable program executes on different platforms and yields consistent performance. With the focus on portability, this paper presents an in-depth study of the performance of three NAS benchmarks (EP, MG, FT) compiled with three commercial HPF compilers (APR, PGI, IBM) on the IBM SP2. Each benchmark is evaluated in two versions: using DO loops and using F90 constructs and/or HPF's Forall statement. Base-line comparison is provided by versions of the benchmarks written in Fortran/MPI and ZPL, a data parallel language developed at the University of Washington.While some F90/Forall programs achieve scalable performance with some compilers, the results indicate a considerable portability problem in HPF programs. Two sources for the problem are identified. First, Fortran's semantics require extensive analysis and optimization to arrive at a parallel program; therefore relying on the compiler's capability alone leads to unpredictable performance. Second, the wide differences in the parallelization strategies used by each compiler may require an HPF program to be customized for the particular compiler. While improving compiler optimizations may help to reduce some performance variations, the results suggest that the foremost criteria for portability is a concise performance model that the compiler must adhere to and that the users can rely on.

Introduction Portability is defined as the ability to use the same program on different platforms and to achieve consistent performance. Developing a parallel program that is both portable and scalable is well recognized as a challenging endeavor. However, the difficulty is not necessarily an intrinsic property of parallel computing. This assertion is especially clear in the case of data parallel algorithms which provide abundant parallelism and tend to involve computation that is very regular. The data parallel model is not adequate for general parallel programming. However, its simplicity coupled with the prevalence of data parallel problems in scientific applications has This research was supported by the IBM Resident Study Program and DARPA Grant N00014-92-J-4041 and motivated the development of many data parallel languages, all with the goal of simplifying programming while achieving scalable and portable performance. Of these languages, High Performance Fortran [10] constitutes the most widespread effort, involving a large consortium of companies and universities. One of HPF's distinctions is that it is the first parallel language with a recognized standard - indeed, HPF can be regarded as the integration of several similar data parallel languages including Fortran D, Vienna Fortran and CM Fortran [6, 14, 22]. The attractions of HPF are manifold. First, its use of Fortran as a base language promises quick user acceptance since the language is well established in the target community. Second, the use of directives to parallelize sequential programs implies ease of programming since the directives can be added incrementally without affecting the program's correctness. In particular cases, the compiler may even be able to parallelize the program without user assistance. On the other hand, HPF also has potential disadvantages. First, as an extension of a sequential language, it is likely to inherit language features that are either incompatible with parallelization or difficult for a compiler to analyze. Second, the portability of a program must not be affected by differences in the technology of the compilers or the machines since the principal purpose for creating a standard is to ensure that programs are portable. HPF's design presents some potential conflicts with the goal of portability. For instance, hiding most aspects of communication from the programmer is convenient, but it forces the user to rely completely on the compiler for generating efficient communication. Differences between compilers will always be present. However, to maintain the program portability in the language, the differences must not force the users to make program modifications to accommodate a specific compiler. In other words, the user should be able to use any compiler to develop a program that scales, then have the option of migrating to a different machine or compiler for better scalar performance. This requires a tight coupling between the language specification and the compiler in the sense that the compiler implementations must provide a consistent behavior for the abstractions provided in the language. To this end, the language specification must serve as a consistent contract between the compiler and the programmer. We call this contract the performance model of the language [18]. A robust performance model has a dual effect: the program performance is (1) predictable to the user and (2) portable across different platforms. With the focus on the portability issue, we study in-depth the performance of three NAS benchmarks compiled with three commercial HPF compilers on the IBM SP2. The benchmarks are: Embarrassingly Parallel (EP), Multigrid (MG), and Fourier Transform (FT). The HPF compilers include Applied Parallel Research, Portland Group, and IBM. To evaluate the effect of data dependences on compiler analysis, we consider two versions of each benchmark: one programmed using DO loops, and the second using F90 constructs and/or HPF's Forall statement. For the comparison, we also consider the performance of each benchmark written in MPI and ZPL [16], a data parallel language developed at the University of Washington. Since message passing programs yield scalable performance but are not convenient, the MPI results represent a level of performance that the HPF programs should use as a point of reference. The motivation for including the ZPL results is as follows. ZPL is a data parallel language developed from first principles. The lack of a parent language allows ZPL to introduce new language constructs and incorporate a robust performance model, creating a concrete delineation between parallel and sequential execution. Consequently, the programming model presented to the user is clear, and the compiler is relatively unhindered by artificial dependencies and complex interactions between language features. One may expect that it is both easier to develop a ZPL compiler and to write a ZPL program that scales well. Naturally, the downside of designing a new language without a legacy is the challenge of gaining user acceptance. For this study, the ZPL measurement gives an indication as to whether consistent and scalable performance can be achieved when the compiler is not hampered by language features unrelated to parallel computation. Our results show that programs that scale well using a particular HPF compiler may not perform similarly with a different compiler, indicating a lack of portability. Some F90/Forall programs achieve scalable performance, but the results are not uniform. For the other programs, the results suggest that Fortran's sequential nature leads to considerable difficulties in the compil- er's analysis and optimization of the communication. By analyzing in detail the implementations by the HPF compilers, we find that the wide differences in the parallelization strategies and their varying degrees of success contribute to the portability problem of HPF programs. While improving compiler optimizations may help to reduce some performance variations, it is clear that a robust solution will require more than a mature compiler technology. The results suggest that the foremost criteria for portability is a concise performance model that the compiler must adhere to and that the users can rely on. This performance model will serve as an effective contract between the users and the compiler. In related work, APR published the performance of its HPF compiler for a suite of HPF pro- grams, along with detailed descriptions of their program restructuring process using the APR FORGE tool to improve the codes [3, 11]. The programs are well tuned to the APR compiler and in many cases rely on the use of APR-specific directives rather than standard HPF directives. Although the approach that APR advocates (program development followed by profiler-based program restructuring) is successful for these instances, the resulting programs may not be portable with respect to performance, particularly in cases that employ APR directives. There- fore, we believe that the suite of APR benchmarks is not well suited for evaluating HPF compilers in general. Similarly, papers by vendors describing their individual HPF compilers typically show some performance numbers; however it remains difficult to make comparisons across compilers [8, 12, 13]. Lin et al. used the APR benchmark suite to compare the performance of ZPL versions of the programs against the corresponding HPF performance published by APR and found that ZPL generally outperforms HPF [17]. However, without access to the APR compiler at the time, detailed analysis was not possible, limiting the comparison to the aggregate timings. This paper makes the following contributions: 1. An in-depth comparison and analysis of the performance of HPF programs with three current HPF compilers and alternative approaches (MPI, ZPL). 2. A comparison of the DO loop with the F90 array syntax and the Forall construct. 3. An assessment of the parallel programming model presented by HPF. The remainder of the paper is organized as follows: Section 2 describes the methodology for the study, including a description of the algorithms and the benchmark implementations. In Section 3, we examine and analyze the benchmarks' performance, detailing the communication generated in each implementation and quantifying the effects of data dependences in the HPF programs. Section 4 provides our observations and our conclusions. 2.1 ZPL Overview ZPL is an array language designed at the University of Washington expressly for parallel ex- ecution. In the context of this paper, it serves two purposes. First, it sets a bound on the performance that can be expected from a high level data parallel language that is not an extension of an existing sequential language. Second, it illustrates the importance of the performance model in a parallel language. ZPL is implicitly parallel - i.e. there are no directives. The concurrency is derived entirely from the semantics of the array operations. Array decompositions, specified at run time, partition arrays into either 1D or 2D blocks. Processors perform the computations for the values they own. Scalars are replicated on all processors and kept coherent by redundantly computing scalar computations. ZPL introduces a new abstraction called region which is used to allocate distributed arrays and to specify distributed computation. ZPL provides a full complement of operations to define regions relative to each other (e.g. [east of R]), to refer to adjacent elements (e.g. A@west), to perform full and partial prefix operations (e.g. big := max !! A), to express strided computations, to establish boundary conditions (e.g. wrap A), and to accomplish other powerful operations (e.g. flooding). ZPL also contains standard operators, data types and control structures, using a syntax similar to Modula-2. 2.2 Benchmark selection To emphasize the portability issue, we establish these criteria: 1. The benchmarks should be derived from an independent source to insure objectivity. 2. A message passing version should be included in the study to establish the target performance 3. For HPF, there should be separate versions that employ F77 DO loop and F90/Forall because there is a significant difference between the two types of constructs. 4. The algorithm should be parallel and there should be no algorithmic differences between versions of the same benchmark. 5. Tuning must adhere to the language specification rather than any specific compiler capability 6. Because support for HPF features is not uniform, the benchmarks should not require any feature that is not supported by all HPF compilers. Following these criteria proves to be challenging given the disparate benchmark availabil- ity. The NAS benchmark version 1.0 (NPB1) was implemented by computer vendors and was intended to measure the best performance possible on a parallel machine without regard to portability. The available sources for NPB1 are generally sequential implementations. Although they are valid HPF programs, the sequential nature of the algorithms may be too difficult for the compilers to parallelize and may not reflect a natural approach to parallel programming. For instance, a programmer may simply choose a specific parallel algorithm to implement in HPF. The NAS version 2.1 (NPB2.1), intended to measure portable performance, is a better choice since the programs implement inherently parallel algorithms and they use the same MPI interface as the compilers in the study. NPB2.1 contains 7 benchmarks 1 , all of which should ideally be included in the study. Unfortunately, a portable HPF version of these benchmarks is not available, severely limiting an independent comparison. While APR and other HPF vendors publish the benchmarks used to acquire their performance measurements, these benchmarks are generally tuned to the specific compiler and are not portable. This limitation forces us to carefully derive HPF versions from the NPB2.1 sources with the focus on portability while avoiding external effects such as algorithmic differences. Among the benchmarks, CG is not available in NPB2.1. SP, BT and LU are not included in this study because they require block cyclic and 3-D data distributions that are not supported by all HPF compilers. These limitations do not prevent these benchmarks from being implemented in HPF and ZPL; however, the implementations will have an algorithmic difference that cannot be factored from the performance. This leaves only FT and MG as potential candidates. Fortunately, EP is by definition highly parallel; therefore its sequential implementation can be trivially parallelized. 2.3 Benchmark implementation The HPF implementations are derived by reverse engineering the MPI programs: communication calls are removed and the local loop bounds are replaced with global loop bounds. To parallelize the programs, HPF directives are then added to recreate the data partitioning of the MPI versions. The HPF compilers are thus presented with a program that is fully data parallel (not sequential) and ready to be parallelized. Conceptually, the task for the compilers is to repartition the problem as specified by the HPF directives and to regenerate the communication. The benchmarks we chose require only the basic BLOCK distribution which is supported by all three compilers and use only the basic HPF intrinsics. Therefore they can stress the compilers without exceeding their capability. The implementations in the ZPL language are derived from the same source as the HPF implementations, but in the following manner: the sequential computation is translated directly from Fortran to the corresponding ZPL syntax, while the parallel execution is expressed using ZPL's parallel constructs. One was added recently. 2.3.1 EP The benchmark EP generates N pairs of pseudo-random floating point values in the interval (0,1) according to the specified algorithm, then redistributes each value x j and y j onto the range (-1,1) by scaling them as 2x Each pair is tested for the condition: If true, the independent Gaussian deviates are computed: Then the new pair (X; Y ) is tested to see if it falls within one of the 10 square annuli and a total count is tabulated for each annulus. l - max(jXj; jY The pseudo-random numbers are generated according to the following linear congruential recursion The values in a pair are consecutive values of the recursion. To scale to the (0,1) range, the value x k is divided by 2 46 . The computation for each pair of Gaussian deviates can proceed independently. Each processor would maintain its own counts of the Gaussian deviates and communicate at the end to obtain the global sum. The random number generation, however, presents a challenge. There are two ways to compute a random value 1. x k can be computed quickly from the preceding value x using only one multiplication and one mod operation, leading to a complexity of O(n). However, the major drawback is the true data dependence on the value x 2. x k can be computed independently using k and the defined values of a and x 0 . This will result in an overall complexity of O(n 2 ). Fortunately, the property of the mod operation allows x k to be computed in O(log steps by using a binary exponentiation algorithm [4]. The goal then is to balance between method (1) and (2) to achieve parallelism while maintaining the O(n) cost. Because EP is not available in the NPB2.1 suite, we use the implementation provided by APR as the DO loop version. This version is structured to achieve the balance between (1) and (2) by batching: the random values are generated in one sequential batch at a time and saved; the seed of the batch is computed using the more expensive method (2), and the remaining values are computed using the less expensive method (1). A DO loop then iterates to compute the number of batches required, and this constitutes the opportunity for parallel execution. The F90/Forall version is derived from the DO loop version with the following modifications: ffl All variables in the main DO loop that cause an output dependence are expanded into arrays of the size of the loop iteration. In other words, the output dependence is eliminated by essentially renaming the variables so that the computation can be expressed in a fully data parallel manner. Since the iteration count is just the number of sequential batches, the expansion is not excessive. ffl Directives are added to partition the arrays onto a 1-D processor grid. ffl The DO loop for the final summation is also recoded using the HPF reduction intrinsic. A complication arises involving the subroutine call within the Forall loop, which must be free of side effects in order for the loop to be distributed. Some slight code rearrangement was done to remove a side effect in the original subroutine, then the PURE directives were added to assert freedom from side effects. The ZPL version is translated in a straightforward manner from the DO loop version. The only notable difference is the use of the ZPL region construct to express the independent batch computations. 2.3.2 MG Multigrid is interesting for several reasons. First, it illustrates the need for data parallel languages such as HPF or ZPL. The NPB2.1 implementation contains over 700 lines of code for the communication - about 30% of the program - which are eliminated when the program is written in a data parallel language. Second, since the main computation is a 27-points stencil, the reference pattern that requires communication is simply a shift by a constant, which results in a simple neighbor exchange in the processor grid. All compilers (ZPL and HPF) recognize this pattern well and employ optimizations such as message vectorization and storage preallocation for the nonlocal data [3, 8, 9, 12]. Therefore, although the benchmark is rather complex, the initial indication is that both HPF and ZPL should be able to produce efficient parallel programs. The benchmark is a V-cycle multigrid algorithm for computing an approximate solution to the discrete Poisson problem: where r 2 is the Laplacian operator r 2 The algorithm consists of 4 iterations of the following three steps: residual correction apply correction where A is the trilinear finite element discretization of the Laplace operator r 2 , M k is the V-cycle multigrid operator as defined in the NPB1 benchmark specification [1, 4]. The algorithm implemented in the NPB2.1 version consists of three phases: the first phase computes the residual, the second phase is a set of steps that applies the M k operator to compute the correction while the last phase applies the correction. The HPF DO loop version is derived from the NPB2.1 implementation as follows: ffl The MPI calls are removed. ffl The local loop bounds are replaced with the global bounds. ffl The use of a COMMON block of storage to hold the set of hierarchical arrays (different sizes) is incompatible with HPF; therefore the arrays are renamed and declared explicitly. ffl HPF directives are added to partition the arrays onto a 3-D processor grid. The array distribution is maintained across subroutine calls by using the transcriptive directives to prevent unnecessary redistribution. The HPF F90/Forall version requires the additional step of rewriting all data parallel loops in F90 syntax. The ZPL version has a similar structure to the HPF F90/Forall version, the notable difference being the use of strided region to express the hierarchy of 3-D grids. A strided region is a sparse index set over which data can be declared and computation can be specified. 2.3.3 FT Consider the partial differential equation for a point x in 3-D space: ffit The FT benchmark solves the PDE by (1) computing the forward 3-D Fourier Transform of u(x; 0), (2) multiplying the result by a set of exponential values, and (3) computing the inverse 3-D Fourier Transform. The problem statement requires 6 solutions, therefore the benchmark consists of 1 forward FFT and 6 pairs of dot products and inverse FFTs. The NPB2.1 implementation follows a standard parallelization scheme [5, 2]. The 3-D FFT computation consists of traversing and applying the 1-D FFT along each of the three dimensions. The 3-D array is partitioned along the third dimension to allow each processor to independently carry out the 1-D FFT along the first and second dimension. Then the array is transposed to enable the traversal of the third dimension. The transpose operation constitutes most of the communication in the program. The program requires moving the third dimension to the first dimension in the transpose so that the memory stride is favorable for the 1-D FFT; therefore the HPF REDISTRIBUTE function alone is not sufficient 2 . The HPF DO loop implementation is derived with the following modifications: 1. HPF directives are added to distribute the arrays along the appropriate dimension. Tran- scriptive directives are used at subroutine boundaries to prevent unnecessary redistribution. 2. The communication for the transpose step is replaced with a global assignment statement. 3. A scratch array that is recast into arrays of different ranks and sizes between subroutines is replaced with multiple arrays of constant rank and size. Although passing an array section in a formal argument is legitimate in HPF, some HPF compilers have difficulty managing array sections. data distribution specifies the partition to processor mapping, not the memory layout. The HPF F90/Forall version requires the additional step of rewriting all data parallel loops in F90 syntax. The ZPL implementation allocates the 3-D arrays as regions of 2-D arrays; the transpose operation is realized with the ZPL permute operator. 2.4 Platform Program portability should be evaluated across multiple platforms of different compilers and machines. In this study, we elect to factor out the difference due to the machine architecture and focus on the compiler difference. Although the machine architecture sets the upper bound on the possible performance, a language and its compiler determine the achievable performance. The targeted parallel platform is the IBM SP2 at the Cornell Theory Center. The HPF compilers used in the study include: ffl Portland Group pghpf version 2.1 ffl IBM xlhpf version 1.0 Applied Parallel Research xhpf version 2.0 All compilers generate MPI calls for the communication and use the same MPI library, ensuring that the communication fabric is identical for all measurements. The PGI and APR HPF compilers generate intermediate Fortran code that is then processed by the standard IBM Fortran compiler (xlf). The IBM HPF compiler generates machine code directly but otherwise is based on the same xlf compiler. The ZPL compiler generates intermediate C code. All measurements use the same compiler options and system environment that NAS and APR specified in their publications (using SP2 wide nodes only), and spotchecks confirmed that the published NAS and APR performances are reproduced. 3 Parallel Performance In this section we examine the performance of the programs. Figure 1 shows the aggregate timing for all versions (MPI, HPF, ZPL) and for the small and large problem size (class S, class A). Notes: ffl Because the execution time may be excessive depending on the success of the compilers, we first examine the small problem size (class S), then only the programs with a reasonable performance and speedup with the large problem size (class A). ffl The time axis uses a log scale because of the wide performance spread. ffl For EP, an MPI version is not used; for class S, the F90/Forall version is not shown. ffl For MG class A, the APR F90/Forall version does not scale and is not included. ffl For FT class A, only one data point is available for the IBM F90/Forall version. Processors time(secs) (b) EP class A time(secs) (a) EP class S Processors time(secs) (c) MG class S Processors time(secs) (d) MG class A time(secs) Processors time(secs) (f) FT class A MPI APR_do IBM_do PGI_do ZPL APR_F90 IBM_F90 PGI_F90 Figure 1: Performance for EP, MG and FT (see notes in Section 3.1 NAS EP benchmark In Figure 1(a), the first surprising observation is that the IBM and PGI compilers achieve no speedup with the HPF DO loop version although the APR compiler produces a program that scales well (recall that the EP DO loop version is from the APR suite). Inspecting the code reveals that no distribution directives were specified for the arrays, resulting in a default data distribution. Although the default distribution is implementation dependent, the conventional choice is to replicate the array. The IBM and PGI compilers distribute the computation strictly by the owner-computes rule; therefore, in order for the program to be parallelized, some data structures must be distributed. Since the arrays in EP are replicated by default, no computation is partitioned among the processors: each processor executes the full program and achieves no speedup. By contrast, the APR parallelization strategy does not strictly adhere to the owner-computes rule. This allows the main loop to be partitioned despite the fact that none of the arrays within the loop are distributed. Note that the HPF language specification does not specify the default distribution for the data nor the partitioning scheme for the computation. The omission was likely intended to maximize the opportunity for the compiler to optimize; however the observation for EP suggests that the different schemes adopted by the compilers may result in a portability problem with HPF programs. When directives were inserted to distribute the arrays, it was found that the main array in EP is intended to hold pseudo-random values generated sequentially, therefore there exists a true dependence in the loop computing the values. If the array is distributed, the compiler will adjust the loop bounds to the local partition, but the computation will be serialized. The HPF F90/Forall version corrects this problem by explicitly distributing the arrays and the IBM and PGI compilers were able to parallelize. The class A performance in Figure 1(b) shows that all compilers achieve the expected linear speedup. However, expanding the arrays to express the computation into a more data parallel form seems to introduce overhead and degrade the scalar performance. It is possible for advanced compiler optimizations such as loop fusion and array contraction to remove this overhead, but these optimizations were either not available or not successful in this case. The ZPL version scales linearly as expected and the scalar performance is slightly better than the APR version. 3.2 NAS MG benchmark Compared to EP, MG allows a more rigorous test of the languages and compilers. We first discuss the performance for the class S in Figure 1(c). The p=1 column shows considerable variation in the scalar performance with all versions showing overhead of 1 to 2 orders of magnitude over the MPI performance. For the base cases, both the original MPI program and the ZPL version scale well. The ZPL compiler partitions the problem in a straightforward manner according to the region and strided region semantics, and the communication is vectorized with little effort. The scalar performance does however show over a 6x overhead compared to the MPI version. The HPF DO loop version clearly does not scale with any HPF compiler, as explained below. The PGI compiler performs poorly in vectorizing the communication when the computation is expressed with DO loops: the communication calls tend to remain in the innermost loop, resulting in a very large number of small messages being generated. In addition, the program uses guards within the loop instead of adjusting the loop bound. The APR compiler only supports a 1-D processor grid, therefore the 3-D distribution specified in the HPF directives is collapsed by default to a 1-D distribution. This limitation affects the asymptotic speedup but does not necessarily limit the parallelization of the 27-point stencil computation. For one subroutine, the compiler detects through interprocedural analysis an alias between two formal arguments, which constitutes an inhibitor for the loop parallelization within the subroutine. However, the analysis does not go further to detect from the index expressions of the array references that no dependence actually exists. For most of the major loops in the program, the APR compiler correctly partitions the computation along the distributed array dimension, but generates very conservative communication before the loop to obtain the latest value for the RHS and after the loop to update the LHS. As a result, the performance degrades with the number of processors. The IBM compiler does not parallelize because it detects an output dependence on a number of variables although the arrays are replicated. In this case, the compiler appears to be overly conversative in maintaining the consistency of the replicated variables. Other loops do not parallelize because they contain an IF statement. The INDEPENDENT directive is treated differently by the compilers. The PGI compiler interprets the directive literally and parallelizes the loop as directed. The IBM compiler on the other hand ensures correctness by performing a more rigorous dependence check nevertheless and because of detected dependence, it does not parallelize the loop. For the HPF F90/Forall version, the IBM and PGI compilers are more successful: the IBM compiler performance and scalability approach ZPL's, while the PGI compiler now experiences little problem in vectorizing the communication; indeed its scalar performance exceeds IBM's. The APR compiler does not result in slowdown but does not achieve any speedup either. It partitions the computation in the F90/Forall version similarly to the DO loop version, but is able to reduce the amount of communication. It continues to be limited by its 1-D distribution as well as an alias problem with one subroutine. Note that the version of MG from the APR suite employs APR's directives to suppress unnecessary communication. These directives are not used in our study because they are not a part of HPF, but it is worth noting that it is possible to use APR's tools to analyze the program and manually insert APR's directives to improve the speedup with the APR compiler. Given that the DO loop version fails to scale with any compiler, one may conjecture whether the program may be written differently to aid the compilers. The specific causes for each compiler described above suggest that the APR compiler would be more successful if APR's directives are used, that the PGI compiler may benefit from the HPF INDEPENDENT directive, and that the IBM compiler would require actual removal of some data dependences. Therefore, it does not appear that any single solution is portable across the compilers. 3.3 NAS FT benchmark FT presents a different challenge to the HPF compilers. In terms of the reference pattern, FT consists of a dot product and the FFT butterfly pattern. The former requires no communication and is readily parallelized by all compilers. For the latter, the index expression is far too complex to optimize the communication, but fortunately the index variable is limited to one dimension at a time; therefore the task for the compiler is to partition the computation along the appropriate dimensions. The intended data distribution is 1-D and is thus within the capability of the APR compiler. Figure 1(e) shows the full set of performance results for the small problem size. As with MG, the MPI and ZPL versions scale well and the scalar performance of all data parallel implementations shows an overhead of 1 to 2 orders of magnitude over the MPI implementation. For the HPF DO loop version, the APR compiler exhibits the same problem as with MG: it generates very conservative communication before and after many loops. In addition, the APR compiler does not choose the correct loop to parallelize. The discrepancy arises because APR's strategy is to choose the partitioning based on the array references within the loop. In this case the main computation and thus the array references are packaged in a subroutine called from the loop, the intention being that the loop is parallelized and the subroutine operates on the local data. When the compiler proceeds to analyze the loops in this subroutine (1-D FFT), it finds that the loops are not parallelizable. The PGI compiler also generates suboptimal communication, although its principal limitation is in vectorizing the messages. The IBM compiler does not parallelize because of assignments to replicated variables. The HPF F90/Forall version requires considerable experimentation and code restructuring to arrive at a version that is accepted by all compilers, partly because of differences in supported features among the compilers and partly because of the nested subroutines structure of the original program. All HPF compilers achieve speedup to varying degrees. APR is particularly successful since the principal parallel loop has been moved to the innermost subroutine. Its scalar performance approaches MPI performance, although communication overhead limits the speedup. PGI shows good speedup while IBM's speedup is more limited. 3.4 Communication The communication generated by the compilers is a useful indication of the effectiveness of the parallelized programs. For the versions of the benchmarks that scale, table 1 shows the total number of MPI message passing calls and the differences in the communication scheme employed by each compiler. The APR and PGI compilers only use the generic send and receive while the IBM compiler also uses the the nonblocking calls and the collective communication; this may have ramifications in the portability of the compiler to other platforms. The ZPL compiler uses nonblocking MPI calls to overlap computation with communication as well as MPI collective communication. Benchmark version point-to-point collective type of MPI calls EP (class Allreduce, Barrier IBM F90 70 120 Send, Recv, Bcast MG (class S) MPI 2736 40 Send, Irecv, Allreduce, Barrier ZPL 9504 56 Isend, Recv, Barrier APR F90 126775 8 Send, Recv, Barrier Bcast FT (class S) MPI 0 104 Alltoall, Reduce APR F90 58877 8 Send, Recv, Barrier IBM F90 728 258048 Send, Irecv, Bcast Table 1: Dynamic communication statistics for EP class A, MG class S and FT class S: p=8 3.5 Data Dependences HPF compilers derive parallelism from the data distribution and the loops that operate on the data. Loops with no dependences are readily parallelized by adjusting the loop bounds to the local bounds. Loops with dependences may still be parallelizable but will require analysis; for instance, the IBM compiler can recognize the dependence in a loop that performs a reduction and generate the appropriate HPF reduction intrinsic. In other instances, loop distribution may isolate the portion containing the dependence to allow the remainder of the original loop to be parallelized. To approximately quantify the degree of difficulty that a program presents to the parallelizing compiler in terms of dependence analysis, we use the following simple metric: count of all loops with dependences count of all loops A value of 0 would indicate that all loops can be trivially parallelized, while a value of 1 would indicate that any parallelizable loops depend on the analysis capability of the compiler. Using the KAPF tool, we collect the loop statistics from the benchmarks for the major subroutines; they are listed in Table 2. This metric is not complete since it does not account for the data distribution; for instance, for 3 nested loops and a 1-D distribution, only 1 loop needs to be partitioned to parallelize the program and 2 loops may contain dependences with no ill effects. In addition, this metric is static and may not correlate directly with the dynamic characteristics of the programs. However, the metric gives a coarse indication for the demand on the compiler. The loop dependence statistics show clear trends that correlate directly with the performance data. We observe the expected reduction in dependences from the DO loop version to the F90/Forall version. The reduction greatly aids the compilers in parallelizing the F90/Forall programs, but also highlights the difficulty with parallelizing programs with DO loops. subroutine DO F90/Forall EP embar 3/5 1/31 get start seed - 1/1 FT fftpde 2/16 2/16 cfft3 0/6 0/6 cfftz 3/5 1/4 subroutine DO F90/Forall MG hmg 1/2 1/27 psinv 4/4 0/6 resid 4/4 0/6 rprj3 4/4 0/6 norm2u3 3/3 0/0 Table 2: Statistics on dependences for EP, MG and FT: m is the count of loops with data dependences or subroutine calls, and n is the total loop count; for F90/Forall, the counts are obtained after the array statements have been scalarized For MG, the difference is significant; the array syntax eliminates the dependence in most cases. Some HPF compilers implement optimizations for array references that are affine functions of the DO loop indices, particularly for functions with constants. These optimizations should have been effective for the MG DO loop version, however it does not appear that they were successful. Note that the loops in the subroutine norm2u3 are replaced altogether with the HPF reduction intrinsics. For FT, the low number of dependences in fftpde comes from the dot-products which are easily parallelized. The top-down order of the subroutines listed also represents the nesting level of the subroutines. The increasing dependences in the inner subroutine reflect the need to achieve parallelism at the higher level. As explained earlier, this proves to be a challenge to the APR compiler which focuses on analyzing individual loops to partition the work. For data parallel applications, the recent progress in languages and compilers allows us to experimentally evaluate an important issue: program portability. We recognize that many factors affect the development and success of a parallel language and our study only focuses on the portability factor. Three NAS benchmarks were studied across three current HPF compilers. We examined different styles of expressing the computation in HPF and we also consider the same benchmarks written in MPI and ZPL to understand the interaction between performance, portability and convenient programming. The HPF compilers show a general difficulty in detecting parallelism from DO loops. The compilers are more successful with the F90 array syntax and Forall construct, although even in this case the success in parallelization is not uniform. Significant variation in the scalar performance also exists between the compilers. While the HPF directives and constructs provide information on the data and computation partitioning, the sequential semantics of Fortran leave many potential dependences in the pro- gram. An HPF compiler must analyze these dependences, and when unable to do so, it must make a conservative assumption. This analysis capability differentiates the various vendor imple- mentations. However, because it is difficult for the compilers to parallelize reliably, a user cannot consistently estimate the parallel behavior and thus the speedup of the program. In addition, because the parallelization strategy by the compilers varies widely, different ways to express the same computation can lead to drastically different performance. The unpredictable variations reflect a shortcoming in the performance model of HPF; as a result, a user needs to continually experiment with each compiler to learn the actual behavior. In doing so, the user is effectively supplementing the performance model provided by the language with empirical information. Yet, such an enhanced model tends to be platform specific and not portable. ZPL programs show consistent scalable performance, illustrating that it is possible to incorporate a robust performance model in a high level language. The language design ensures that the language abstractions behave in a predictable manner with respect to parallel performance. Although ZPL is supported on multiple parallel systems, the results in this study do not directly show ZPL's portability across platforms because multiple independent compiler implementations for ZPL are not available. However, the existence of the performance model, evident in the predictable performance behavior, ensures that ZPL programs will be portable across independently developed platforms. The results also show that significant overhead in the scalar performance remains in all implementations compared to the MPI programs. One source for the overhead is the large number of temporary arrays generated by the compiler across subroutine calls and parallelized loops. They require dynamic allocation/deallocation and copying, and generally degrade the cache perfor- mance. The index computation also contributes significantly to the overhead. It is clear that to become a viable alternative to explicit message passing, compilers for data parallel languages must achieve a much lower scalar overhead. --R David Klepacki Rick Lawrence An efficient parallel algorithm for the 3-D FFT NAS parallel benchmark Applied Parallel Research. The NAS parallel benchmarks. Rob van der Wijngaart Vienna Fortran 90. Compiling Fortran 90D/HPF for distributed memory mimd computers. Compiling High Performance Fortran. Factor-Join: A unique approach to compiling array languages for parallel machines. High Performance Fortran Forum. Fortran parallelization hand- book An HPF compiler for the IBM SP2. Compiling High Performance Fortran for distributed-memory systems Evaluating compiler optimizations for Fortran D. ZPL language reference manual. ZPL: An array sublanguage. ZPL vs. HPF: A comparison of performance and programming style. The Role of Performance Models in Parallel Programming and Languages. On the influence of programming models on shared memory computer performance. NAS parallel benchmark 2.1 results: 8/96. A ZPL programming guide. CM Fortran Programming Guide --TR Compiling Fortran 90D/HPF for distributed memory MIMD computers Evaluating compiler optimizations for Fortran D High-performance parallel implementations of the NAS kernel benchmarks on the IBM SP2 Compiling high performance Fortran for distributed-memory systems The role of performance models in parallel programming and languages Factor-Join ZPL --CTR Bradford L. Chamberlain , Sung-Eun Choi , E Christopher Lewis , Lawrence Snyder , W. Derrick Weathersby , Calvin Lin, The Case for High-Level Parallel Programming in ZPL, IEEE Computational Science & Engineering, v.5 n.3, p.76-86, July 1998 Bradford L. Chamberlain , Steven J. Deitz , Lawrence Snyder, A comparative study of the NAS MG benchmark across parallel languages and architectures, Proceedings of the 2000 ACM/IEEE conference on Supercomputing (CDROM), p.46-es, November 04-10, 2000, Dallas, Texas, United States M. Govett , L. Hart , T. Henderson , J. Middlecoff , D. Schaffer, The scalable modeling system: directive-based code parallelization for distributed and shared memory computers, Parallel Computing, v.29 n.8, p.995-1020, 1 August Bradford L. Chamberlain , Sung-Eun Choi , E. Christopher Lewis , Calvin Lin , Lawrence Snyder , W. Derrick Weathersby, ZPL: A Machine Independent Programming Language for Parallel Computers, IEEE Transactions on Software Engineering, v.26 n.3, p.197-211, March 2000

MPI;performance model;NAS;HPF;ZPL;data paraller language

509627

Massively parallel simulations of diffusion in dense polymeric structures.

An original computational technique to generate close-to-equilibrium dense polymeric structures is proposed. Diffusion of small gases are studied on the equilibrated structures using massively parallel molecular dynamics simulations running on the Intel Teraflops (9216 Pentium Pro processors) and Intel Paragon (1840 processors). Compared to the current state-of-the-art equilibration methods this new technique appears to be faster by some orders of magnitude. The main advantage of the technique is that one can circumvent the bottlenecks in configuration space that inhibit relaxation in molecular dynamics simulations. The technique is based on the fact that tetravalent atoms (such as carbon and silicon) fit in the center of a regular tetrahedron and that regular tetrahedrons can be used to mesh the three-dimensional space. Thus, the problem of polymer equilibration described by continuous equations in molecular dynamics is reduced to a discrete problem where solutions are approximated by simple algorithms. Practical modeling applications include the construction of butyl rubber and ethylene-propylene-dimer-monomer (EPDM) models for oxygen and water diffusion calculations. Butyl and EPDM are used in O-ring systems and serve as sealing joints in many manufactured objects. Diffusion coefficients of small gases have been measured experimentally on both polymeric systems, and in general the diffusion coefficients in EPDM are an order of magnitude larger than in butyl. In order to better understand the diffusion phenomena, 10,000 atoms models were generated and equilibrated for butyl and EPDM. The models were submitted to a massively parallel molecular dynamics simulation to monitor the trajectories of the diffusing species. The massively parallel molecular dynamics code used in this paper achieves parallelism by a spatial-decomposition of the workload which enables it to run large problems in a scalable way where both memory cost and per-timestep execution speed scale linearly with the number of atoms being simulated. It runs efficiently on several parallel platforms, including the Intel Teraflops at Sandia. There are several diffusion modes observed depending if the diffusion is probed at short time scale (anomalous mode) or long time scale (normal mode). Ultimately, the diffusion coefficient that need to be compared with experimental data corresponds to the normal mode. The dynamics trajectories obtained with butyl and EPDM demonstrated that the normal mode was reached for diffusion within one nanosecond of simulation. In agreement with experimental evidences, the oxygen and water diffusion coefficients were found larger for EPDM than butyl.

Introduction Many technological processes depend on the design of polymers with desired permeation characteristics of small molecules. Examples include gas separation with polymeric membranes, food packaging, and encapsulant of electronic components in polymers that act as barriers to atmospheric gases and moisture. To solve the polymer design problem successfully, one needs to relate the chemical composition of the polymer to the diffusivities of the the penetrants molecules within it. One way of establishing structure-property relationships is the development of empirical relationships (namely QSPRs - quantitative structure property relationships). A more sophisticated approach is the application of simulation techniques that rely directly upon fundamental molecular science. While QSPRs can generally be established from the the monomers configurations, simulations involve the generation of configurations representing the whole polymer. From these configurations, structural, thermodynamics, and transport properties are estimated. In principle, simulations can provide exact results for a given model representation of the polymer/penetrant system. In practice, computer time consideration necessitates the introduction of approximation, and particularly for polymeric systems, the use of high performance computing. In this paper, we propose an original technique to generate close-to-equilibrium polymeric structures. We are presenting data that suggest our technique is some orders of magnitude faster than the current state-of-the-art methods used to prepare and equilibrate dense polymeric systems. The technique is used to generate initial model structures for polymers present in O-ring systems. Oxygen and water diffusions though the O-ring are then probed using massively parallel molecular dynamics. Results are compared with similar simulations and experimental data. Methodology As pointed out in the introduction, the ability to represent the molecular level structure and mobility of polymeric structures is a prerequisite for simulating diffusion in them. One fundamental parameter characterizing molecular structures is the potential energy. Polymer chains in a equilibrium melt or amorphous glass remains essentially at minimum energy configurations. The potential energy is the sum of bond and bond angle distortion terms, tortional potentials, as well as intermolecular and intramolecular repulsive and attractive interactions (i.e., van der Waals), and electrostatic (i.e., Coulombic) interactions. These energy terms are expressed as a function of the atoms coordinates constituting the polymer and a set of parameters computed from experimental data or quantum mechanics calculations. The functional forms of the energy terms and their associated parameters is called a forcefield. In the present study, we are making use of the commercially available CHARMm forcefield [1], as well as forcefield parameters taken from Muller-Plathe et al. [2] Molecular mechanics and molecular dynamics. Molecular mechanics is the procedure by which one locates local minima of energy. Molecular mechanics simply consists of a minimization routine (conjugate gradient, steepest descent, etc.) that finds the first energy minimum from a starting configuration. In the present paper, a parallel implementation of conjugate gradient is utilized [3]. The main disadvantage of molecular mechanics is that thermal fluctuations are not explicitly taken into account, and therefore cannot be used for diffusion calculations. Moreover, the procedure followed in generating minimum energy configurations does not correspond to any physical process of polymer formation. Nonetheless, static minimum energy structures provide satisfactory starting configurations for molecular dynamics simulations. Molecular dynamics (MD) follows the temporal evolution of a microscopic model system through numerical integration of the equations of motions for all the degrees of freedom. MD simulations can be performed in the microcanonical ensemble (constant number of molecules, volume, and total energy, or NVE), the canonical ensemble (constant number of molecules, volume, and temperature, or NVT), as well as the isothermal-isobaric ensemble (constant number of molecules, pressure, and temperature, or NPT). The major advantage of MD is that it provides detailed information of short-time molecular motions. Its limitation resides in computer time consideration. Hundreds of CPU hours on a vector supercomputer are required to simulate a nanosecond of actual atomistic motions. However, computer time can be decreased by making use of massively parallel processing. In the present work we are using a large-scale atomic/molecular massively parallel simulator (LAMMPS) [4]. LAMMPS is a new parallel MD code suitable for modeling large molecular systems. LAMMPS has been written as part of a CRADA (Cooperative Research and Development Agreement) between two DOE labs: Sandia and Lawrence Livermore, and three industrial partners: Bristol-Myers Squibb, Cray Research, and Dupont. LAMMPS is capable of modeling a variety of molecular systems such as bio-membranes, polymers, liquid-crystals, and zeolites. The code computes two kinds of forces: (1) short-range forces such as those due to van der Waals interactions and molecular bond stretching, bending, and torsions, and (2) long-range forces due to Coulombic effects. In the latter case, LAMMPS uses either Ewald or particle particle/particle-mesh (PPPM) techniques to speed the calculation [5]. LAMMPS achieves parallelism by a spatial-decomposition of the workload which enables it to run large problems in a scalable way where both memory cost and per-timestep execution speed scale linearly with the number of atoms being simulated [6]. It runs efficiently on several parallel platforms, including the Intel Teraflops at Sandia (9216 processors), the large Intel Paragon at Sandia (1840 processors) and large Cray T3D machines at Cray Research. Generate initial structure. Creating an atomic-level model of a completely equilibrated dense polymer melt is a challenging task, given current computational capabilities. van der Vegt et al. [7] discuss the various approaches to this problem. The common techniques used in simulations of liquids which consists of melting an idealized structure generally takes tens of picosecond of equilibration using MD. Unfortunately, the equilibration time for dense polymers is many orders of magnitude larger than are feasible with MD [8]. Consequently, one needs other efficient ways of preparing initial polymer structures, that resemble the equilibration structure. One of the most common methods has been to pack chains into the simulation box at the experimental density, either by random placement followed by energy minimization or with Monte Carlo chain growth techniques. However, these methods have tended to produce structures which are rather inhomogeneous and which lead to an overestimation of solubility values for small permeants. In the present paper, we are proposing and comparing two alternative construction methods: the compression box technique and the lattice construction technique. Compression box technique. van der Vegt et al. [7] suggest that a more efficient way of producing near-equilibrium structures is by starting with a dilute model polymer and compressing it slowly until the target experimental density is reached. A set of chains is built with the correct dihedral angle distributions, at about 1/8 of the experimental density. Using an NPT MD technique with a pressure ramp, the model system is compressed to the desired density over ~ 500 picoseconds. During this compression, only the repulsive part of the nonbond (van der Waals) interaction is used, to avoid "clustering" of the polymer. Then the structure is further equilibrated with the full nonbond interactions at the new density, for ~ 1000 picoseconds. van der Vegt et al. [7] observed that a model of poly(dimethylsiloxane) built with this "slow-compression" technique had significantly less stress in the dihedral angles than a model built with a "random packing" method. Furthermore, the slowly compressed model yielded small-permeant solubility results which were in much better agreement with experiment. Lattice construction technique. Many natural and synthetic polymers including all structures considered in this paper are formed with hydrocarbon chains. More precisely, these polymers are composed of linear chains of tetravalent carbon atoms to which are attached molecular groups (methyl, phenyl,. The lattice technique is based on the fact that any tetravalent atom (such as carbon and silicon) fits in the center of a regular tetrahedron where its four bonds are perpendicular to the faces of the tetrahedron (cf. Figure 1a). It is well known that regular tetrahedrons can be used to mesh the three-dimensional space. Figure 1. Tetrahedron meshing. a) Carbon atom in center of tetrahedron. b) Two dimensional projection of meshed space. Using the previous observation, the experimental polymer volume is meshed and the polymer chains are constructed by generating random walks following adjacent tetrahedrons (cf. Figure 1.b). One may notice from Figure 1 that the distance between two adjacent tetrahedrons is equal to a carbon-carbon bond length, and the angle between three adjacent tetrahedrons is equal to a carbon-carbon-carbon angle. Therefore, bond and angle energy terms are directly minimized by the lattice construction procedure. Excluded volume interactions are treated by keeping track of the occupancy of each tetrahedron. As the chain construction progresses the next atom position is always chosen in a non occupied tetrahedron. In the unprobable event where all adjacent tetrahedrons are already occupied the construction routine back-tracks to the previous position. In order to keep homogeneous density the first atom of the chain is chosen at random in the least occupied region of the volume box. The goal of above excluded volume procedure is to keep the intermolecular and intramolecular repulsive forces to a minimum value. The lattice construction technique can be used with cubic cells instead of tetrahedrons. In such a case, each cell represents a monomer, and the cell length, width, and height are those of the boundary box of the monomer. The advantage of the cubic cells lattice is a reduction of the computational complexity since the objects manipulated are monomers instead of atoms (for instance, for butyl, there are 12 atoms per monomer). The disadvantage of the cubic lattice is that bond lengths and bond angles between monomers are no longer necessarily valid. Therefore, structures generated with cubic lattices must be energy minimized prior using MD simulations in order to restore the correct bond lengths and bond angles. An important parameter characterizing equilibrium melts or amorphous glasses is the mean square end-to-end distance of the polymer chains. According to Flory "random coil hypothesis", for which there is ample experimental support [9], at equilibrium the mean square end-to-end distance of any polymer chain in the bulk is related to the number n of skeletal bond lengths l as , where is the characteristic ratio of the polymer. It is also known that the correct end-to-end distance starting from any random initial configuration can be reached by MD using a simulation time proportional to O(n 3 ) [10]. Our lattice construction technique makes use of Flory's result while avoiding the computational complexity of MD simulations. For each chain to be constructed the initial atom position is chosen at random (in the least occupied region) and the final atom position is chosen within a distance equal to the value given by the Flory equation. Then, a path is constructed between the two chosen positions. At each step of the construction the non-occupied adjacent tetrahedrons are ranked using their respective distance to the final position. The tetrahedron or cubic cell corresponding to the shortest distance is chosen as the next position. When the last position is reached, if the path length is greater than n, then the chain is deleted and two new positions are chosen. Note that this situation is unlikely to occur . However, in the likely event where the path length is smaller than n, additional atoms are added to the chain until the correct length is reached. As illustrated in Figure 2, the addition of new atoms is carried out by deleting at random a bond along the chain, thus creating two non-bonded atoms from which the chain is extended. Thus, the chain extension is grown using the same path construction as above with the exception that the initial and final positions are not chosen at random but are the location of two non-bonded atoms. Figure 2. Increase of polymer chain length. a) Bond i-j is deleted. b) Chain extension is carried out by growing a new chain from position i and j. Note that the lattice construction procedure allows one to construct cross-linked polymers since the basic operation is to generate a chain between two given positions. Hence, cross-linking chains are constructed by generating chains between pairs of branch-points. Once the polymeric structures have been constructed, minimization followed by NVT MD simulations are used to further equilibrate the structures. While minimization is carried out using the aforementioned conjugate gradient algorithm, MD is performed using the massively parallel LAMMPS code. Diffusion calculations. Once polymeric structures have been generated and equilibrated using the compression box or lattice construction technique, diffusion calculations can be carried out. Microscopically, the diffusion coefficient can be calculated from the motion of the diffusing particles, provided they have been traced long enough so that they perform random walks. There are several formulations to derive coefficients of diffusion. With MD simulations the most often used expression is the so called Einstein relation where D is the diffusion coefficient, is the number of spatial dimension, t is the time, and r(t) is the position vector of a particle at time t. The angle brackets denote the ensemble average, which in MD simulations is realized by averaging over all particles of the diffusing species. The calculation of diffusion coefficients rests on the fact that for sufficiently long times the mean-square displacement of a diffusing particle increases linearly with time. There are however, cases in which the mean-square displacement is not linearly proportional to time, but obeys a different power law where n is lower than 1 (normal diffusion). This process is called anomalous diffusion. It is caused by some mechanism which forces the particles onto a path that is not a random walk. These can be obstacles in the way of the diffusant and thus inhibit random motion, such as to force the diffusant to remain inside a cavity. Anomalous diffusion persists only on short time scales. At long enough time scales (and hence length scales) the trajectory of the diffusant becomes randomized and a change to normal diffusion occurs. Results and discussion The main goal of our study is to measure oxygen and water diffusion coefficient in two polymeric materials, butyl rubber and ethylene-propylene-dimer-monomer (EPDM). Butyl and EPDM are used in O-ring systems and serve as sealing joints in many manufactured objects. Butyl rubber is a copolymer of isobutene and isoprene (Figure 3). EPDM is a terpolymer of ethylene, propylene, and 1,4 hexadiene (Figure 3). The experimental density for both butyl and EPDM is 0.91 g/cm 3 . Small gases diffusion coefficients have been measured experimentally on both polymeric systems, and in general the diffusion coefficients in EPDM are an order of magnitude larger than in butyl. Presently this decrepency is poorly understood. Butyl rubber: EPDM: Figure 3. Butyl rubber and EPDM structural formulas. The results outlined in this section were obtained using the lattice construction code and the energy minimization program both running on a SGI R10000 platform. MD simulations were carried out using the MD LAMMPS code running on both the 9216 processors Teraflops Intel and 1840 processors Intel Paragon. All butyl MD runs were carried out on the Paragon using 216 nodes, EPDM and polyisobutylene MD runs were performed on the Teraflops using 125 nodes. We observed a speed up factor of 7 between the Teraflops and the Paragon for the same number of nodes. Polymer construction. In prior calculations of diffusion coefficients we probed the performances of the lattice construction code for dense polymeric systems. It has been suggested by van der Vegt et al. [7] that polymer building techniques which create models directly at the experimental density may introduce significant nonequilibrium strain in the angle bending and dihedral torsion degrees of freedom. Furthermore, the time required to anneal this strain may be beyond MD time scales, thus prohibiting the generation of true equilibrium polymer structures. van der Vegt et al. [7] showed that model polymers built by slow compression from a low density initial state exhibit less strain in the angle and dihedral degrees of freedom, and they suggest that these structures may be more representative of an equilibrated polymer. Since the lattice construction technique we introduce here is novel, and it creates models directly at experimental densities, we thought it appropriate to compare it to the "slow-compression" building method. The polymer used in this study was poly(isobutylene), or PIB, this polymer is enssentially an non cross-linked butyl rubber. The forcefield parameters were taken from Muller-Plathe et al. [2] The Lorenz-Berthelot mixing rules were used for nonbond interactions between different atomic species, and all studies were performed at 300K. Model A was built with the slow-compression method described in the Methodology section. First, a set of polymer chains was built at a density of 0.11 g/cm 3 , which is approximately 1/8 of the experimental density. There were 21 chains with lengths distributed randomly between 80 and 120 monomer units. The full-atom model consisted of 25104 atoms, 25076 bonds, 50040 angles, and 74594 dihedrals. To anneal out nonbond (van der Waals) overlaps in the initial configuration, an MD simulation with a small timestep was run for a few picoseconds (ps). We then performed two MD stages as suggested by van der Vegt et al. [7] The first stage was a compression using only the repulsive part of the nonbond potential (i.e., the Lennard-Jones interaction between species i and j was truncated at ij ) but full bonded interactions (bonds, angles, dihedrals). The model was compressed from its initial density to the experimental density, 0.91 g/cc, over 525 picoseconds, using constant-temperature MD with a pressure ramp. The Nose-Hoover method was used to control both temperature and pressure, with time constants of 0.1 and 0.5 ps, respectively. The second stage was initiated from the last configuration of the first stage, with the attractive nonbond interactions turned on by increasing the Lennard-Jones cutoffs to 74.5 nm. A constant pressure MD simulation was performed such that the experimental density (0.91 g/cm 3 ) was maintained. This run lasted for about 120 picoseconds. About 20 picoseconds into the second stage, the polymer seemed to be equilibrated; changes in the various components of potential energy were not significant or systematic. Model B was built at the experimental density of 0.91 g/cm 3 using the lattice construction technique described earlier. It was created to be close in size to Sample A, although it ended up 2.8% smaller. It contained 21 chains with lengths distributed randomly between 80 and 120 monomer units. It was a full-atom model consisting of 24408 atoms, 24387 bonds, 48690 angles, and 72658 dihedrals. Prior to running dynamic simulations, this model was minimized to reduce nonbond overlaps using the procedure described in the Methodology section. A MD run at constant volume and temperature was then performed for 380 picoseconds. The Nose-Hoover method with a time constant of 0.1 ps was used to maintain temperature at 300K. Equilibration was reached within the first 10 picoseconds. Our first comparison of the two polymer models (i.e, PIB A and B) is given in Table I, which shows the various contributions to potential energy in the systems. The averages and standard deviations were taken over the last 10 ps of the dynamics for Model A and for the last 80 picoseconds of the dynamics for Model B. All values represent total amounts for each contribution; since the two models were approximately the same size, we did not normalize the values. Interestingly, we observe differences between the two models in the van der Waals, angle, and dihedral contributions; in each case the model created with the lattice contruction technique (Model B) has a higher energy. The total potential energy difference between the models is 8650 kJ/mol, or 0.14 k B T per atom. The largest differences are found in the angle and dihedral energies, accounting for about 80% of the total difference. Both of these observations are consistent with those of van der Vegt et al. [7] for poly(dimethylsiloxane) polymer models. They observed a total difference of per atom between models made by direct packing and slow compression methods, with about 70% of that due to angle, dihedral, 1-4 nonbond, and 1-5 nonbond contributions. Table I. Potential energy contributions in PIB models (kcal/mol) PIB Model van der Waals bonds angles dihedrals Since the concern of this paper is the calculation of diffusion coefficients of small molecules in polymers, we are interested in how the differences between Models A and B will affect such calculations. We studied the diffusion of helium (He) through each of the polymer models, using the Lennard-Jones parameters given by Muller-Plathe et al. [2] for He. In the case of Sample A, a polymer configuration with a density of exactly 0.91 g/cm 3 was taken from the end of the second-stage runs. Then 100 He atoms were added to the configuration. In the case of sample B, 100 He atoms were added to the final configuration. The overlaps of the new He atoms with the polymer were relaxed using constant volume and temperature MD with a small time step. Diffusion MD runs were then performed with constant volume and temperature (300K) for 360 picoseconds for model A and 770 picoseconds for model B. Figure 4 visualizes the mean-square displacement of He molecules in samples A and B. The straight lines are drawn with slopes of exactly unity and one can see that at long times eq. 2 is valid for 1. The normal mode is reached for both systems within 1000 picosecond. The diffusion coefficient for model A is model B. Both diffusion coefficients are close to the value found by Muller-Plathe et al. this is due to the fact we are using the same forcefield. Most importantly the diffusion coefficients differences between systems A and B are not significant. Hence, the lattice construction technique appears to generate equilibrated polymer structures that have the same behavior that those created by the slow compression method in so far as diffusion calculations are concerned. Figure 4. Trajectories of helium particles versus time in PIB models A & B (cf. text). The straight line represents the normal diffusion regime. We further probed the performances of the lattice construction technique with PIB systems of increasing sizes. As shown in Figure 5, the computational complexity appears to scale linearly with the number of atoms. This is a substantial gain compare to other MD techniques in which the number of steps of the simulation must be at least n 3 , where n is the number of atoms of the polymer chains (cf. Methodology section). As a consequence, we were able to generate PIB structures up to 1,380,000 atoms. To the best of our knowledge, these models are the largest non-crystalline bonded atomic systems ever generated and in general are several orders of magnitude larger than the current models used in polymer science. It is also important to note that the structures generated by the lattice construction program are equilibrated using, at most, ten picoseconds MD simulations, rather than several hundred picoseconds with other techniques. Figure 5. Computational running time of the lattice construction code versus number of atoms and number of monomers. Oxygen and water diffusion calculations. The lattice construction code was used to generate initial structures for butyl and EPDM. Both structures contained approximately 10,000 atoms and were generated in a cubic box of 4.5 nm size, which lead to a density of 0.91 g/cm 3 . In both systems, oxygen molecules were added up to 3% of the total weight, while water molecules were added up to 9% weight. Although 3% weight is the experimental value for both oxygen and water in butyl and EPDM, a higher value for water was chosen to obtain a better statistical average when calculating diffusion coefficients. Once the diffusing species were added, the two resulting structures were energy minimized and equilibrated with MD simulations. The simulation time used for equilibration did not exceed 10 picoseconds. In order to compute diffusion coefficients using eq. 1, MD was run up to a 1000 picoseconds simulation. This large simulation time was chosen with the hope of being able to reach the normal diffusion mode (cf. eq. 2 and discussion below). Figures 6 and 7 visualize the mean-square displacement of oxygen and water molecules in butyl and EPDM as a function of time. With the exception of water in EPDM the normal mode was reached in all cases. The diffusion coefficients D were evaluated from the positions of these lines. Figure 6a. Trajectories of oxygen particles versus time in butyl rubber. 6b. Trajectories of oxygen particles versus time in EPDM. The straight line represents the normal diffusion regime. Figure 7a. Trajectories of water particles versus time in butyl rubber. 7b. Trajectories of water particles versus time in EPDM. The non convergence for diffusion calculations of water in EPDM may be attributed to the fact that several diffusing molecules can eventually fit into EPDM cavities. Indeed, when visualizing the solvated EPDM model it was observed that some cavities contained several water molecules. Because of the electrostatic attractions between water molecules, one can hypothesize that when several penetrants are present in one cavity, it is energetically favorable for the penetrants to remain inside the cavity rather than jumping to neighboring cavities. Since 9% weight for water is overestimated compared to the corresponding experimental value, a new EPDM model was constructed containing only 3% water. With this new model, the normal mode was reached in less than 1000 picoseconds. The diffusion coefficient values are listed in Table II. Table II. Diffusion coefficients for O 2 and water (10 -6 cm 2 /s) Model system Diffusion coef. (this work) Diffusion coef. (experiment) [11] O 2 in rubber 0.285 0.081 O in rubber 0.159 not reported O 2 in EPDM 0.781 0.177 O in EPDM 1.921 not reported The diffusion coefficients for EPDM are larger than for butyl as expected from experiments. A tentative explanation was provided when visualizing the volumes of the equilibrated polymers. Free volumes are void spaces in model structures that are accessible by penetrant molecules. The free volumes were computed using a program developed by one of the authors of this paper. [12] The determination of the free volume within a given model system is conducted as follows. The radii of all the polymer atoms are augmented by the length equal to the penetrant radius, and the unoccupied volume of the resulting model system is then calculated. For both butyl and EPDM it was observed that the free volumes did not percolate for penetrant having a diameter greater 0.30 nm (diameter of atomic oxygen is 0.35 nm). In other words, all cavities having an entrance size greater than 0.30 nm were not connected. This observation is consistent with the hoping mechanism picture proposed by some authors. [13][14][15] Penetrant molecules spend relatively long times in cavities before performing infrequent jumps between adjacent cavities. The same authors have performed visual inspection of polymer models in the vicinity of penetrants, and it appeared that jumping events occur after channels between neighboring cavities are formed. Once the channels are formed, the penetrants slip through it without much effort. If such a picture is true, since EPDM has higher diffusion coefficients than butyl, EPDM should have a greater number of channels than butyl. We probed the free volume distribution for butyl and EPDM versus the cavity entrance size. As shown in Figure 8, butyl and EPDM have a different free volume distribution. EPDM contains significantly more free volumes having an entrance size smaller than 0.30 nm than butyl. It is important to note that these small free volumes provide links between larger cavities since the free volume network percolates for entrance sizes smaller than 0.30 nm. Consequently, the curves presented in Figure 8 are consistent with the hoping mechanism picture, and according to this picture, EPDM has higher diffusion coefficients because it comprises more channels between cavities. Figure 8. Free volume distribution in butyl rubber and EPDM. Conclusion We have proposed a new technique to generate close-to-equilibrium dense polymeric structures. This technique leads to diffusion results that are similar to results obtained by previously published methods, while being several order of magnitude faster. We have shown the new technique to scale linearly with the number of atoms, and have used it to generate dense polymer models comprising up to 1,380,000 atoms. To the best of our knowledge these models are the largest non-crystalline bonded atomic systems ever generated and are many times larger than the current models used in polymer science. We have used the new technique to construct 10,000 atoms models of butyl rubber and EPDM for the purpose of simulating the diffusion of oxygen and water molecules. The simulations were carried out using LAMMPS molecular dynamics code, and were run on Sandia's Intel Teraflop and Sandia's Intel Paragon. In agreement with experimental results the diffusion coefficients in EPDM were found an order of magnitude larger than in butyl. A tentative explanation of the differences between the coefficients was advanced when comparing the free volume distributions of the two polymer models: diffusion is facilitated in EPDM due to a larger number of channels between cavities than in butyl. Acknowledgments We are pleased to acknowledge the funding provided by the Accelerated Strategic Computing Initiative (ASCI) of the U.S. Department of Energy, Sandia National Laboratories under contract DE-AC04-76DP00789. --R "CHARMm: A program for macromolecular energy, minimization and dynamics calculations" Personal communication "Particle-Mesh Ewald and rRESPA for Parallel Molecular Dynamics Simulations" Principles of Polymer Chemistry Scaling Concepts in Polymer Physics in Polymer Handbook --TR

gas diffusion;teraflop;molecular builder;molecular dynamics;polymer

509641

A scalable mark-sweep garbage collector on large-scale shared-memory machines.

This work describes implementation of a mark-sweep garbage collector (GC) for shared-memory machines and reports its performance. It is a simple ''parallel'' collector in which all processors cooperatively traverse objects in the global shared heap. The collector stops the application program during a collection and assumes a uniform access cost to all locations in the shared heap. Implementation is based on the Boehm-Demers-Weiser conservative GC (Boehm GC). Experiments have been done on Ultra Enterprise 10000 (Ultra Sparc processor 250 MHz, 64 processors). We wrote two applications, BH (an N-body problem solver) and CKY (a context free grammar parser) in a parallel extension to C++.Through the experiments, We observe that load balancing is the key to achieving scalability. A naive collector without load redistribution hardly exhibits speed-up (at most fourfold speed-up on 64 processors). Performance can be improved by dynamic load balancing, which exchanges objects to be scanned between processors, but we still observe that straightforward implementation severely limits performance. First, large objects become a source of significant load imbalance, because the unit of load redistribution is a single object. Performance is improved by splitting a large object into small pieces before pushing it onto the mark stack. Next, processors spend a significant amount of time uselessly because of serializing method for termination detection using a shared counter. This problem suddenly appeared on more than processors. By implementing non-serializing method for termination detection, the idle time is eliminated and performance is improved. With all these careful implementation, we achieved average speed-up of 28.0 in BH and 28.6 in CKY on 64 processors.

Introduction Shared-memory architecture is attractive platform for implementation of general-purpose parallel programming languages that support irregular, pointer-based data structures [4, 20]. The recent progress in scalable shared-memory technologies is also making these architectures attractive for high-performance, massively parallel computing one of the important issues not yet addressed in the implementation of general-purpose parallel programming languages is scalable garbage collection (GC) technique for shared-heaps. Most previous work on GC for shared-memory machines is concurrent GC [6, 10, 17], by which we mean that the collector on a dedicated processor runs concurrently with application programs, but does not perform collection itself in parallel. The focus has been on shortening pause time of applications by overlapping the collection and the applications on different processors. Having a large number of processors, however, such collectors may not be able to catch up allocation speed of applications. To achieve scalability, we should parallelize collection itself. This paper describes the implementation of a parallel mark-sweep GC on a large-scale (up to 64 processors), multiprogrammed shared-memory multiprocessor and presents the results of empirical studies of its performance. The algorithm is, at least conceptually, very simple; when an allocation requests a collection, the application program is stopped and all the processors are dedicated to collection. Despite its simplicity, achieving scalability turned out to be a very challenging task. In the empirical study, we found a number of factors that severely limit the scalability, some of which appear only when the number of processors becomes large. We show how to eliminate these factors and demonstrate the speed-up of the collection. We implemented the collector by extending the Boehm-Demers-Weiser conservative garbage collection library (Boehm GC [2, 3]) on two systems: a 64-processor Ultra Enterprise 10000 and a 16-processor Origin 2000. The heart of the extension is dynamic task redistribution through exchanging contents of the mark stack (i.e., data that are live but yet to be examined by the collector). At present, we achieved 14-28-fold speed-up on Ultra Enterprise 10000, and 3.7 to 6.3-fold speed-up on Origin 2000. The rest of the paper is organized as follows. Chapter 2 compares our approach with previous work. Chapter 3 briefly summarizes Boehm GC, on which our collector is based. Chapter 4 describes our parallel marking algorithm and solutions for performance limiting factors. Chapter 5 describes the experimental conditions. Chapter 6 shows experimental results, and we conclude in Chapter 7. Chapter 2 Previous Work Most previous published work on GCs for shared-memory machines has dealt with concurrent GC [6, 10, 17], in which only one processor performs a collection at a time. The focus of such work is not on the scalability on large-scale or medium-scale shared-memory machines but on shortening pause time by overlapping GC and the application by utilizing multiprocessors. When GC itself is not parallelized, the collector may fail to finish a single collection cycle before the application exhausts the heap (Figure 2.1). This will occur on large-scale machines, where the amount of live data will be large and the (cumulative) speed of allocation will be correspondingly high. We are therefore much more interested in "parallel" garbage collectors, in which a single collection is performed cooperatively by all the processors. Several systems use this type of collectors [7, 16] and we believe there are many unpublished work too, but there are relatively few published performance results. To our knowledge, the present paper is the first published work that examines the scalability of parallel collectors on real, large-scale, and multiprogrammed shared-memory machines. Most of previous publications have reported only preliminary measurements. Uzuhara constructed a parallel mark sweep collector on symmetric multiprocessors [22]. When the amount of free space in the shared heap becomes smaller than a threshold, some processors start a collection, while other processors continue application execution. The collector processors cooperatively mark all reachable objects with dynamic load balancing by using global task pool. Then they sweep the heap and The time when memory region can be reused GC Time Time Application PEs Concurrent GC Parallel GC (Our approach) Figure 2.1: Difference between concurrent GC and our approach. If only one dedicated processor performs GC, a collection cycle becomes longer in proportion to the number of processors. join the application workers. This approach has the same advantage as concurrent GC, and it can prevent a single collection cycle from becoming longer on large-scale machines. Ichiyoshi and Morita proposed a parallel copying GC for a shared heap [11]. It assumes that the heap is divided into several local heaps and a single shared heap. Each processor collects its local heap individually. Collection on the shared-heap is done cooperatively but asynchronously. During a collection, live data in the shared- heap (called 'from-space' of the collection) are copied to another space called `to-space'. Each processor, on its own initiative, copies data that is reachable from its local heap to to-space. Once a processor has copied data reachable from its local heap, it can resume application on that processor, which works in the new shared-heap (i.e., to-space). Our collector is much simpler than both of Uzuhara's collector and Ichiyoshi and Morita's collector; it simply synchronizes all the processors at a collection and all the processors are dedicated to the collection until all reachable objects are marked. Although Ichiyoshi and Morita have not mentioned explicitly, we believe that a potential advantage of their method over ours is its lower susceptibility to load imbalance of a collection. That is, the idle time that would appear in our collector is effectively filled by the application. The performance measurement in Chapter 6 shows a good speed-up up to our maximum configuration, 64 processors, and indicates that there is no urgent need to consider using the application to fill the idle time. We prefer our method because it does not interfere with SPMD-style applications, in which global synchronizations are frequent. 1 Both of Uzuhara's method and Ichiyoshi and Morita's method may interact badly with such applications because it exhibits a very long marking cycle, during which the applications cannot utilize all the processors. Taura also reached a similar conclusion on distributed-memory machines [21]. Our collector algorithm is most similar to Imai and Tick's parallel copying collector [12]. In their study, all processors perform copying tasks cooperatively and any memory object in one shared heap can be copied by any processor. Dynamic load balancing 1 A global synchronization occurs even if the programming language does not provide explicit barrier synchronization primitives. It implicitly occurs in many places, such as reduction and termination detection. is achieved by exchanging memory pages to be scanned in the to-space among proces- sors. Speed-up is calculated by a simulation that assumes processors become idle only because of load imbalance-the simulation overlooks other sources of performance degrading factors such as spin-time for lock acquisition. As we will show in Chapter 6, such factors become quite significant, especially in large-scale and multiprogrammed environments. Chapter 3 Boehm-Demers-Weiser Conservative GC Library The Boehm-Demers-Weiser conservative GC library (Boehm GC) is a mark-sweep GC library for C and C++. The interface to applications is very simple; it simply replaces calls to malloc with calls to GC MALLOC. The collector automatically reclaims memory no longer used by the application. Because of the lack of precise knowledge about types of words in memory, a conservative GC is necessarily a mark-sweep collector, which does not move data. Boehm GC supports parallel programs using Solaris threads. The current focus seems to support parallel programs with minimum implementation efforts; it serializes all allocation requests and GC is not parallelized. 3.1 Sequential Mark-Sweep Algorithm The mark-sweep collector's work is to find all garbage objects, which are unreachable from the root set (machine registers, stacks and global variables) via any pointer paths, and to free those objects. To tell whether an object is live (reachable) or garbage, each object has its mark bit, which shows 0(= 'unmarked') before a collection cycle. We mention how Boehm GC maintains the mark bits in section 3.2. A collection cycle consist of two phases; in the mark phase, the collector traverse objects that are reachable from root set recursively, and sets (marks) their mark bits at 1(= 'marked'). To mark objects recursively, Boehm GC uses a data structure called mark stack as shown in section 3.3. In the sweep phase, the collector scans all mark bits and frees objects whose mark bits are still 'unmarked'. The sweeping method heavily depends on how the free objects are managed. We describe aspects relevant to the sweep phase in section 3.4. 3.2 Heap Blocks and Mark Bitmaps Boehm GC manages a heap in units of 4-KB blocks, called heap blocks. Objects in a single heap block must have the same size and be word-aligned. For each block separate header record (heap block header) is allocated that contains information about the block, such as the size of the objects in it. Also kept in the header is a mark bitmap for the objects in the block. A single bit is allocated for each word (32 bits in our experimental environments); thus, a mark bitmap is 128-byte length. The j th bit of the th byte in the mark bitmap describes the state of an object that begins at (BlockAddr BlockAddr is the start address of the corresponding heap block. Put differently, each word in a mark bitmap describes the states of consecutive words in the corresponding heap block, which may contain multiple small objects. Therefore, in parallel GC algorithms, visiting and marking an object must explicitly be done atomically. Otherwise, if two processors simultaneously mark objects that share a common word in a mark bitmap, either of them may not be marked properly. 3.3 Mark Stack Boehm GC maintains marking tasks to be performed with a vector called mark stack . It keeps track of objects that have been marked but may directly point to an unmarked object. Each entry is represented by two words: ffl the beginning address of an object, and ffl the size of the object. Figure 3.1 shows the marking process in pseudo code; each iteration pops an entry from the mark stack and scans the specified object, 1 possibly pushing new entries onto precisely, when the specified object is very large (? 4 KB), the collector scans only the first 4 KB and keeps the rest in the stack. push all roots (registers, stack, global variables) onto mark stack. while (mark stack is not empty) f for size of o; i++) f if (o[i] is not a pointer) do nothing else if (mark bit of o[i] == 'marked') do nothing else f mark bit of push(o[i], mark stack) Figure 3.1: The marking process with a mark stack. the mark stack. A mark phase finishes when the mark stack becomes empty. 3.4 Sweep In the sweep phase, Boehm GC does not free each garbage object actually. Instead, it distinguish empty heap blocks from other heap blocks. Boehm GC examines the mark bitmaps of all heap blocks in the heap. A heap block that contains any marked object is linked to a list called reclaim list, to prepare for future allocation requests 2 . Heap blocks that are found empty are linked to a list called list, in which heap blocks are sorted by their addresses, and adjacent ones are coalesced to form a large contiguous block. Heap block free list is examined when an allocation cannot be served from a reclaim list. 2 The system does not free garbage objects on nonempty heap blocks, until the program requests objects of the proper size (lazy sweeping). In order to find a garbage objects from those heap blocks, mark bitmaps are preserved until next collection Chapter 4 Parallel GC Algorithm Our collector supports parallel programs that consist of several UNIX processes. We assume that all processes are forked at the initialization of a program and are not added to the application dynamically. Interface to the application program is the same as that of the original Boehm GC; it provides GC MALLOC, which now returns a pointer to shared memory (acquired by a mmap system call). We could alternatively support Solaris threads. The choice is arbitrary and somewhat historical; we simply thought having private global variables makes implementation simpler. We do not claim one is better than the other. 4.1 Basic Algorithm 4.1.1 Parallel Marking Each processor has its own local mark stack. When GC is invoked, all application processes are suspended by sending signals to them. When all the signals have been delivered, every processor starts marking from its local root, pushing objects onto its local mark stack. When an object is marked, the corresponding word in a mark bitmap is locked before the mark bit is read. The purpose of the lock is twofold. one is to ensure that a live object is marked exactly once, and the other is to atomically set the appropriate mark bit of the word. When all reachable objects are marked, the mark phase is finished. This naive parallel marking hardly results in any recognizable speed-up because of Objects marked by PE2 Objects marked by PE1 PE1's root PE2's root heap Figure 4.1: In the simple algorithm, all nodes of a shared tree are marked by one processor. the imbalance of marking tasks among processors. Load imbalance is significant when a large data structure is shared among processors through a small number of externally visible objects. For example, a significant imbalance is observed when a large tree is shared among processors only through a root object. In this case, once the root node of the tree is marked by one processor, so are all the internal nodes (Figure 4.1). To improve marking performance, our collector performs dynamic load balancing by exchanging entries stored in mark stacks. 4.1.2 Dynamic Load Balancing of Marking Besides a local mark stack, each processor maintains an additional data structure named stealable mark queue, through which "tasks" (entries in mark stacks) are exchanged 4.2). During marking, each processor periodically checks its stealable mark queue. If it is empty, the processor moves all the entries in the local mark stack Tasks Mark stack Lock Stealable mark queue Figure 4.2: Dynamic load balancing method: tasks are exchanged through stealable mark queues. (except entries that point to the local root, which can be processed only by the local processor) to the stealable mark queue. When a processor becomes idle (i.e., when its mark stack becomes empty), it tries to obtain tasks from stealable mark queues. The processor examines its own stealable mark queue first, and then those of other processors, until it finds a non-empty queue. Once it finds one, it steals half of the entries 1 in the queue and stores them into its mark stack. Because several processors may become idle simultaneously, this test-and-steal operation must acquire a lock on a queue. The mark phase is terminated when all the mark stacks and stealable mark queues become empty. The termination is detected by using a global counter to maintain the number of empty stacks and empty queues. The counter is updated whenever a processor becomes idle or obtains tasks. 1 If the queue has n entries and n is an odd number, (n + 1)=2 entries are stolen. 4.1.3 Parallel Sweeping In the parallel algorithm, all processors share a single heap block free list, while each processor maintains a local reclaim list. In the sweep phase, each processor examines a part of the heap and links empty heap blocks to the heap block free list and non-empty ones to its local reclaim list. Since each processor has a local reclaim list, inserting blocks to a reclaim list is straightforward. Inserting blocks to the heap block free list is, however, far more difficult, because the heap block free list is shared, blocks must sorted by their addresses, and adjacent blocks must be coalesced. To reduce the contention and the overhead on the shared list, we make the unit of work distribution in the sweep phase larger than a single heap block and perform tasks as locally as possible; each processor acquires a large number of (64 in the current implementation) contiguous heap blocks at a time and processes them locally. Empty blocks are locally sorted and coalesced within the blocks acquired at a time and accumulated in a local list called partial heap block free list. Each processor repeats this process until all the blocks have been examined. Finally, the lists of empty blocks accumulated in partial heap block free lists are chained together to form the global heap block free list, possibly coalescing blocks at joints. When this sweep phase is finished, we restart the application. 4.2 Performance Limiting Factors and Solutions The basic marking algorithm described in previous section exhibits acceptable speed-up on small-scale systems (e.g., approximately fourfold speed-up on eight processors). As we will see in Chapter 6, however, several factors severely limit speed-up and this basic form never yields more than a 12-fold speed-up. Below we list these factors and describe how did we address them in turn. Load imbalance by large objects: We often found that a large object became a source of significant load imbalance. Recall that the smallest unit of task distribution is a single entry in a stealable mark queue, which represents a single object in memory. This is still too large! We often found that only some processors were busy scanning large objects, while other processors were idle. This behavior was most prominent when applications used many stacks or large arrays. In one of our parallel applications, the input data, which is a single 800-KB array caused significant load imbalance. In the basic algorithm, it was not unusual for some processors to be idle during the entire second half of a mark phase. We address this problem by splitting large objects (objects larger than 512 bytes) into small (512-byte) pieces before it is pushed onto the mark stack. In the experiments described later, we refer to this optimization as SLO (Split Large Object). Delay in testing mark bitmap: We observed cases where processors consumed a significant amount of time acquiring locks on mark bits. A simple way to guarantee that a single object is marked only once is to lock the corresponding mark bit (more precisely, the word that contains the mark bit) before reading it. How- ever, this may unnecessarily delay processors that read the mark bit of an object to just know the object is already marked. To improve the sequence, we replaced this "lock-and-test" operation with optimistic synchronization. We tests a mark bit first and quit if the bit is already set. Otherwise, we calculate the new bitmap for the word and write the new bitmap in the original location, if the location is the same as the originally read bitmap. This operation is done atomically by compare&swap instruction in SPARC architecture or load-link and store-conditional instructions in MIPS architecture. We retry if the location has been overwritten by another processor. These operations eliminate useless lock acquisitions on mark bits that are already set. We refer to this optimization as MOS (Marking with Optimistic Synchronization) in the experiments below. Another advantage of this algorithm is that it is a non-blocking algorithm [8, 18, 19], and hence does not suffer from untimely preemption. A major problem with the basic algorithm is, however, that locking a word in a bitmap every time we check if an object is marked causes contention (even in the absence of preemption). We confirmed that a "test-and-lock-and-test" sequence that checks the mark bit before locking works equally well, though it is a blocking algorithm. Serialization in termination detection: When the number of processors becomes large, we found that the GC speed suddenly dropped. It revealed that processors spent a significant amount of time to acquire a lock on the global counter that maintains the number of empty mark stacks and empty stealable mark queues. We updated this counter each time a stack (queue) became empty or tasks were thrown into an empty stack (queue). This serialized update operation on the counter introduced a long critical path in the collector. We implemented another termination detection method in which two flags are maintained by each processor; one tells whether the mark stack of the processor is currently empty and the other tells whether the stealable mark queue of the processor is currently empty. Since each processor maintains its own flags on locations different from those of the flags of other processors, setting flags and clearing flags are done without locking. Termination is detected by scanning through all the flags in turn. To guarantee the atomicity of the detecting process, we maintain an additional global flag detection-interrupted , which is set when a collector recovers from its idle state. A detecting processor clears the detection-interrupted flag, scans through all the flags until it finds any non-empty queue, and finally checks the detection- interrupted flag again if all queues are empty. It retries if the process has been interrupted by any processor. We must take care of the order of updating flags lest termination be detected by mistake. For example, when processor A steals all tasks of processor B, we need to change flags in the following order: (1) stack- empty flag of A is cleared, (2) detection-interrupted flag is set, and (3) queue- empty flag of B is set. We refer to this optimization as NSB (Non-Serializing Barrier). Chapter 5 Experimental Conditions We have implemented the collector on two systems: the Ultra Enterprise 10000 and the Origin 2000. The former has a uniform memory access (UMA) architecture and the latter has a nonuniform memory access (NUMA) architecture. The implementation is based on the source code of Boehm GC version 4.10. We used four applications written in C++: BH (an N-body problem solver), CKY (a context free (a life game simulator) and RNA (a program to predict RNA secondary structure). 5.1 Ultra Enterprise 10000 Ultra Enterprise 10000 is a symmetric multiprocessor with sixty-four 250 MHz Ultra processors. All processors and memories are connected through a crossbar interconnect whose bandwidth is 10.7 GB/s. The L2 cache block size is 64 bytes. 5.2 Origin 2000 Origin 2000 is a distributed shared memory machine. The machine we used in the experiment has sixteen 195 MHz R10000 processors. That system consists of eight modules, each of which has two processors and the memory module. The modules are connected through a hypercube interconnect whose bandwidth is 2.5 GB/s. The memory bandwidth of each module is 0.78 GB/s and the L2 cache block size is 128 bytes. In the default configuration, each memory page (whose size is 16 KB) is placed on the same node as the processor that accessed the page first. Therefore processors can have desired pages on local by touching the pages at the initializing phase of the program. We used two physical memory allocation policies in the experiment: Local to allocator (LA) Each heap block and corresponding mark bitmap are local to the processor that allocate the heap block first. Round-robin (RR) The home node of a heap block is determined by its address rather than the allocator of the block. The heap block is local to processor P such that #Processors. The home of a mark bitmap is determined in the same rule. 5.3 Applications We used following four applications written in C++. BH and CKY are parallel ap- plications. We wrote Enterprise version of those applications in a parallel extension to C++ [14]. This extension allows programmer to create user level threads dynamically. The runtime implicitly uses fork system call at the beginning of the program. Since this extension does not work on Origin now, we wrote those applications on Origin by using fork system call explicitly. Life and RNA are sequential applications; even those sequential applications can utilize our parallel collection facility. BH simulates the motion of N particles by using the Barnes-Hut algorithm [1]. At each time step, BH makes a tree whose leaves correspond to particles and calculates the acceleration, speed, and location of the particles by using the tree. In the experiment, we simulate 10000 particles for 50 time steps. CKY takes sentences written in natural language and the syntax rules of that language as input, and outputs all possible parse trees for each sentence. CKY calculates all nonterminal symbol for each substring of the input sentence in bottom-up. In the experiment, each of the given 256 sentences consists of 10 to words. Life solves "Conway's Game of Life''. It simulates the cells on square board. Each cell have either of two states, ON and OFF. The state of a cell is determined by states of adjacent cells at the previous time step. The program takes a list that contains ON cells in an initial state. The number of initial ON cells is 5685 in our experiment. We simulate them for 150 time steps. RNA predicts the secondary structure of an RNA sequence. The input data is a set of stack regions and each stack region has its position and energy. A set of stack regions is called feasible if any pair of its elements fulfills a certain condition. The problem is to find all feasible subsets of given stack regions whose total energy is no smaller than a threshold. The size of input stack regions is 119 in our experiment. 5.4 Evaluation Framework Ideally, the speed-up of the collector should be measured by using various numbers of processors and applying the algorithm to the same snapshot of heap. It is difficult, however, to reproduce the same snapshot multiple times because of the indeterminacy of application programs. The amount of data is so large that we cannot simply dump the entire image of the heap. Even if such dumping were feasible, it would still be difficult to continue from a dumped image with a different number of processors. Thus the only feasible approach is to formulate the amount of work needed to finish a collection for a given heap snapshot and then calculate how fast the work is processed at each occurrence of a collection. A generally accepted estimation of the workload of marking for a given heap configuration is the amount of live objects, or equivalently, the number of words that are scanned by the collector. This, however, ignores the fact that the load on each word differs depending on whether it is a pointer, and the density of pointers in a live data may differ from one collection to another. Given a word in heap, Boehm GC first performs a simple test that rules out most non-pointers and then examines the word more elaborately. To measure the speed-up more accurately, we define the workload W of a collection as a 4 x 4 a 5 x 5 is the number of marked objects, x 2 the number of times to scan already marked objects, x 3 the number of times to scan non-pointers, x 4 the number of empty heap blocks, and x 5 the number of non-empty blocks 1 . Each x n is totaled over all processors. The GC speed S is defined as is the elapsed time of the collection. And the GC speed-up on N processors is the ratio of S on N processors to S on a single processor. When we measure S on a single processor, we eliminate overhead for parallelization. The constants an were determined through a preliminary experiment. To determine a 3 , for example, we created a 1000-word object that contained only non-pointers and we measured the time to scan the object. We ran this measurement several times and used the shortest time. It took 20 us to scan a 1000-word object on Enterprise 10000; that means 0.020 us per word. From this result, we let a 0:020. The other constants were determined similarly. The intention of this preliminary experiment is to measure the time for the workload without any cache misses. In the experiment, the constants were set at a 1 0:16, a 3 a 4 2:0, and a 5 = 1:3 on Enterprise 10000, and a 1 0:13, a 3 a 4 2:0, and a 5 = 1:3 on Origin 2000. 1 The marking workload is derived from x1 ; x2 ; x3 and the sweeping workload is from x4 ; x5 . Chapter 6 Experimental Results 6.1 Speed-up of GC Figures 6.1-6.16 show performance of GC using the four applications on two systems. We measured several versions of collectors. "Sequential" refers to the original Boehm GC and "Simple" refers to the algorithm in which each processor simply marks objects that are reachable from the root of that processor without any further task distribution. "Basic" refers to the basic algorithm described in Section 4.1, and the following three versions refer to ones that implement all but one of the optimizations described in Section 4.2. "No-XXX" stands for a version that implements all the optimizations but XXX. "Full" is the fully optimized version. We measured an additional version on Origin 2000, "Full-LA". This is the same as "Full" but takes different physical memory allocation policy. "Full-LA" takes "Local to Allocator" policy, while all other versions do "Round-robin" policy. The applications were executed four times in each configuration and invoked collections more than 40 times. The table shows the average performance of the invocations. When we used all or almost all the processors on the machine, we occasionally observed invocations that performed distinguishably worse than the usual ones. They were typically times worse than the usual ones. The frequency of such unusually bad invocations was about once in every five invocations when we used all processors. We have not yet determined the reason for these invocations. It might be the effect of other processes. For the purpose of this study, we exclude these cases. Figure 6.1-6.4 and 6.8-6.11 compare three versions, namely, Simple, Basic, and Full. The graphs show that Simple does not exhibit any recognizable speed-up in any application. As Figure 6.1-6.4 show, Basic on Enterprise 10000 performs reasonably until a certain point, but it does not scale any more beyond it. The exception is RNA, where we do not see the difference between Basic and Full. The saturation point of Basic depends on the application; Basic of CKY reaches the peak on 32 processors, while that of BH reaches the saturation point on 8 processors. The peak of Life is 48 processors. Full achieved about 28-fold speed-up in BH and in CKY, and about 14-fold speed-up in Life and RNA on 64 processors. On 16-processor Origin 2000, the difference between Basic and Full is little, except in BH. The some performance problems in Basic, however, appear only when the number of processors becomes large as we have observed on Enterprise 10000; thus, Full would be more significant on larger system. Full achieved 3.7-6.3-fold speed-up on processors. 6.2 Effect of Each Optimization Figure 6.5-6.7 show how each optimization affects scalability on Enterprise 10000. Especially in BH and in CKY, removing any particular optimization yields a sizable degradation in performance when we have a large number of processors. Without the improved termination detection by the non-serializing barrier (NSB), neither BH nor CKY achieves more than a 17-fold speed-up. Without NSB, Life does not scale on more than 48 processors, too. Sensitivity to optimizations differs among the applications; Splitting large objects (SLO) and marking with optimistic synchronization (MOS) have significant impacts in BH, while they do not in other applications. SLO is important when we have a large object in the application. In BH, we use a single array named particles to hold all particles data, whose size is 800 KB in our experiments. This large array became a bottleneck when we omitted SLO optimization. This phenomenon was noted on Origin 2000, as Figure 6.12 indicates. Generally, MOS have significant effects when we have objects with big reference counts, because these objects cause many contentions between collectors that try to visit them. The experiment revealed that the array particles was the source of problem again; in one collection cycle, we observed that we had about 70,000 pointers to this array. That caused significant contentions. This big reference count was produced by the stack of user threads. Because our BH implementation computes forces to the particles in parallel, each thread has references to its responsible particles. Although those references are directed to distinct addresses (for example, ith thread has references to particles[i]), all of them are regarded as pointers to a single object particles. MOS optimization effectively alleviate the contentions in such case and improve the performance. We observe significant impact of NSB optimization on GC speed in three applica- tions, but we do not see that in RNA even on 64 processors. Although the reason for this difference is not understood well, in general, NSB is important when collectors tend to become idle frequently. This is because collectors often update the idle coun- ters, when we do not implement NSB. In RNA, the frequency of the task shortage may be low. We will investigate whether this hypothesis is the case in the future. 6.3 Effect of Physical Memory Allocation Policy Figure 6.13-6.16 compare two memory placement policies: 'Local to allocator (LA)' and 'Round-robin (RR)' on Origin 2000, described in Section 5.2. Full adopts RR policy and Full-LA LA. As we can see easily, collection speed with RR is significantly faster than that with LA in three applications, BH, CKY and RNA. When we adopt LA policy, GC speed does not improve on more than eight processors. While we have not fully analyzed this, we conjecture that this is mainly due to the imbalance in the amount of allocated physical pages among nodes. With LA policy, the access to objects and mark bitmaps in the mark phase contend at nodes that have allocated many pages. Actually, BH has significant memory imbalance because only one processor construct a tree of particles. And all objects in RNA are naturally allocated by one processor because our RNA is sequential program 1 . We will investigate why this is not the case in Life, which is also sequential program. 6.4 Discussion on Optimized Performance As we have seen in Section 6.1, the GC speed of fully optimized version always get faster as the number of processors increases in any applications. But they considerably differ in GC speed; for instance, it is 28-fold speed-up in BH and CKY, while 14-fold in Life and RNA on Enterprise 10000. In order to try to find the cause of this difference, we examined how processors spend time during the mark phase 2 . Figure 6.17-6.20 show the breakdowns. From these figures, we can say that the biggest problem in Life is load imbalance, because processors spend a significant amount of time in idle. The performance improvement may be possible by refining the load balancing method. On the other hand, we currently can not specify the reasons of the relatively bad performance in RNA, where processors are busy during 90% of the mark phase. In most collection cycles, the sweep phase is five to ten times shorter than mark phase. We therefore focus on the mark phase. Sequential Sequential code without overhead for parallelization. Simple Parallelized but no load balancing is done. Basic Only load balancing is done. All optimizations but SLO (splitting large object) are done. All optimizations but MOS (marking with optimistic synchronization) are done. All optimizations but NSB (non-serializing barrier) are done. All optimizations but SLO (splitting large object) are done. Full All optimizations are done. Full-LA Origin 2000 only. Same as Full, but the physical memory allocation policy is "Local to Allocator". Table 6.1: Description of labels in following graphs. Except Full-LA, the physical memory allocation policy on Origin 2000 is "Round-robin". number of processors speed-up Full Basic Simple Linear Figure 6.1: Average GC speed-up in BH on Enterprise 10000. number of processors speed-up Full Basic Simple Linear Figure 6.2: Average GC speed-up in CKY on Enterprise 10000. number of processors speed-up Full Basic Simple linear Figure 6.3: Average GC speed-up in Life on Enterprise 10000. number of processors speed-up Full Basic Simple linear Figure 6.4: Average GC speed-up in RNA on Enterprise 10000. number of processors speed-up Full Linear Figure 6.5: Effect of each optimization in BH on Enterprise 10000. number of processors speed-up Full Linear Figure 6.6: Effect of each optimization in CKY on Enterprise 10000. number of processors speed-up Full linear Figure 6.7: Effect of each optimization in Life on Enterprise 10000. BH (Origin 2000)2610 number of processors speed-up Full Basic Simple linear Figure 6.8: Average GC speed-up in BH on Origin 2000. CKY (Origin 2000)2610 number of processors speed-up Full Basic Simple linear Figure 6.9: Average GC speed-up in CKY on Origin 2000. Life (Origin 2000)2610 number of processors speed-up Full Basic Simple linear Figure 6.10: Average GC speed-up in Life on Origin 2000. RNA (Origin 2000)2610 number of processors speed-up Full Basic Simple linear Figure 6.11: Average GC speed-up in RNA on Origin 2000. BH (Origin 2000)2610 number of processors speed-up Full linear Figure 6.12: Effect of each optimization in BH on Origin 2000. BH (Origin 2000)2610 number of processors speed-up Full Full-LA linear Figure 6.13: Effect of physical memory allocation policy in BH on Origin 2000. CKY (Origin 2000)2610 number of processors speed-up Full Full-LA linear Figure 6.14: Effect of physical memory allocation policy in CKY on Origin 2000. Life (Origin 2000)2610 number of processors speed-up Full Full-LA linear Figure 6.15: Effect of physical memory allocation policy in Life on Origin 2000. RNA (Origin 2000)2610 number of processors speed-up Full Full-LA linear Figure 6.16: Effect of physical memory allocation policy in RNA on Origin 2000. BH (Enterprise 10000) 0% 20% 40% 80% 100% number of processors Busy Lock Balance Idle Figure 6.17: Breakdown of the mark phase in BH on Enterprise 10000. This shows busy, waiting for lock, moving tasks, and idle. CKY (Enterprise 10000) 0% 20% 40% 80% 100% number of processors Busy Lock Balance Idle Figure 6.18: Breakdown of the mark phase in CKY on Enterprise 10000. Life (Enterprise 10000) 0% 20% 40% 80% 100% number of processors Busy Lock Balance Idle Figure 6.19: Breakdown of the mark phase in Life on Enterprise 10000. RNA (Enterprise 10000) 0% 20% 40% 80% 100% number of processors Busy Lock Balance Idle Figure 6.20: Breakdown of the mark phase in RNA on Enterprise 10000. Chapter 7 Conclusion We constructed a highly scalable parallel mark-sweep garbage collector for shared-memory machines. Implementation and evaluation are done on two systems: Ultra Enterprise 10000, a symmetric shared-memory machine with 64 processors and Origin 2000, a distributed shared-memory machine with 16 processors. This collector performs dynamic load balancing by exchanging objects in mark stacks. Through the experiments on the large-scale machine, we found a number of factors that severely limit the scalability, and presented the following solutions: (1) Because the unit of load balancing was a single object, a large object that cannot be divided degraded the utilization of processors. Splitting large objects into small parts when they are pushed onto the mark stack enabled a better load balancing. (2) We observed that processors spent a significant time for lock acquisitions on mark bits in BH. The useless lock acquisitions were eliminated by using a optimistic synchronization instead of a "lock-and-test" operation. (3) Especially on 32 or more processors, processors wasted a significant amount of time because of the serializing operation used in the termination detection with a global counter. We implemented non-serializing method using local flags without locking, and the long critical path was eliminated. On Origin 2000, we must pay attention to physical page placement. With the default policy that places a physical page to the node that first touches it, the GC speed was not scalable. We improved performance by distributing physical pages in a round robin fashion. We conjecture that this is because the default policy causes imbalance of access traffic between nodes; since some nodes have much more physical pages allocated than other nodes, accesses to these highly-loaded nodes tend to contend, hence the latency of such remote accesses accordingly increases. For now, we do not have enough tools to conclude. When using all these solutions, we achieved 14 to 28-fold speed-up on 64-processor Enterprise 10000, and 3.7 to 6.3-fold speed-up on 16-processor Origin 2000. Chapter 8 Future Work We would like to improve the GC performance further. In Section 6.4, we have seen that the collectors in some applications still spend a significant amout of time in idle. We will investigate how we can improve the load balancing method. Instead of using buffers for communication (the stealable mark queues), stealing tasks from victim's mark stack directly may enable faster load distributing. We have also noticed that we cannot explain the relatively bad performance of RNA by the load imbalance alone. This may be due to the number of cache misses, which are included by 'Busy' in Figure 6.20. We can capture the number of cache misses by using performance counters, which recent processors are equipped with. We can use the R10000's counters through /proc file system on Origin 2000. And we have constructed a simple tool to use the Ultra SPARC's counters on Enterprise 10000. With these tools, we are planning to examine how often processors meets cache misses. In Section 6.3, we mentioned that we can obtain better performance with the RR (Round-robin) physical memory allocation policy than with the LA (Local to allocator) policy. So far, the focus of our discussion is the speed of GC alone. But the matter will be complicated when we take account into the locality of application programs. LA policy may be advantageous when a memory region tends to be accessed by the allocator. The ideal situation would be that most accesses by application are local, while collection task is balanced well. --R A hirarchical O(N log N) force-calculation algorithm Space efficient conservative garbage collection. Garbage collection in an uncooperative environment. Concurrent garbage collection for C A concurrent generational garbage collector for a multithreaded implementation of ML. A language for concurrent symbolic computa- tion A methodology for implementing highly concurrent data ob- jects A concurrent copying garbage collector for languages that distinguish (Im)mutable data. A shared-memory parallel extension of KLIC and its garbage collection Evaluation of parallel copying garbage collection on a shared-memory multiprocessor Garbage Collection The SGI Origin: A ccNUMA highly scalable server. Garbage collection in MultiScheme (preliminary version). Concurrent replicating garbage collection. Remarks on a methodology for implementing highly concurrent data objects. "a methodology for implementing highly concurrent data objects" An effective garbage collection strategy for parallel programming languages on large scale distributed-memory machines Parallel garbage collection on shared-memory multiprocessors Uniprocessor garbage collection techniques. --TR MULTILISP: a language for concurrent symbolic computation Garbage collection in an uncooperative environment Garbage collection in MultiScheme Space efficient conservative garbage collection A concurrent copying garbage collector for languages that distinguish (im)mutable data A concurrent, generational garbage collector for a multithreaded implementation of ML A methodology for implementing highly concurrent data objects Concurrent replicating garbage collection Remarks on A methodology for implementing highly concurrent data Notes on MYAMPERSANDldquo;A methodology for implementing highly concurrent data objectsMYAMPERSANDrdquo; Garbage collection An effective garbage collection strategy for parallel programming languages on large scale distributed-memory machines Lock-Free Garbage Collection for Multiprocessors Evaluation of Parallel Copying Garbage Collection on a Shared-Memory Multiprocessor Uniprocessor Garbage Collection Techniques ICC++-AC++ Dialect for High Performance Parallel Computing --CTR Guy E. Blelloch , Perry Cheng, On bounding time and space for multiprocessor garbage collection, ACM SIGPLAN Notices, v.34 n.5, p.104-117, May 1999 Guy E. Blelloch , Perry Cheng, On bounding time and space for multiprocessor garbage collection, ACM SIGPLAN Notices, v.39 n.4, April 2004 David Siegwart , Martin Hirzel, Improving locality with parallel hierarchical copying GC, Proceedings of the 2006 international symposium on Memory management, June 10-11, 2006, Ottawa, Ontario, Canada Toshio Endo , Kenjiro Taura, Reducing pause time of conservative collectors, ACM SIGPLAN Notices, v.38 n.2 supplement, February H. Gao , J. F. Groote , W. H. Hesselink, Lock-free parallel and concurrent garbage collection by mark&sweep, Science of Computer Programming, v.64 n.3, p.341-374, February, 2007 David Detlefs , Christine Flood , Steve Heller , Tony Printezis, Garbage-first garbage collection, Proceedings of the 4th international symposium on Memory management, October 24-25, 2004, Vancouver, BC, Canada Yoav Ossia , Ori Ben-Yitzhak , Irit Goft , Elliot K. Kolodner , Victor Leikehman , Avi Owshanko, A parallel, incremental and concurrent GC for servers, ACM SIGPLAN Notices, v.37 n.5, May 2002 Katherine Barabash , Ori Ben-Yitzhak , Irit Goft , Elliot K. Kolodner , Victor Leikehman , Yoav Ossia , Avi Owshanko , Erez Petrank, A parallel, incremental, mostly concurrent garbage collector for servers, ACM Transactions on Programming Languages and Systems (TOPLAS), v.27 n.6, p.1097-1146, November 2005 Yossi Levanoni , Erez Petrank, An on-the-fly reference-counting garbage collector for java, ACM Transactions on Programming Languages and Systems (TOPLAS), v.28 n.1, p.1-69, January 2006 Yossi Levanoni , Erez Petrank, An on-the-fly reference counting garbage collector for Java, ACM SIGPLAN Notices, v.36 n.11, p.367-380, 11/01/2001 Guy E. Blelloch , Perry Cheng, On bounding time and space for multiprocessor garbage collection, ACM SIGPLAN Notices, v.39 n.4, April 2004

scalability;shared-memory machine;parallel algorithm;dynamic load balancing;garbage collection

509644

Loop re-ordering and pre-fetching at run-time.

The order in which loop iterations are executed can have a large impact on the number of cache misses that an applications takes. A new loop order that preserves the semantics of the old order but has a better cache data re-use, improves the performance of that application. Several compiler techniques exist to transform loops such that the order of iterations reduces cache misses. This paper introduces a run-time method to determine the order based on a dependence-driven execution. In a dependence-driven execution, an execution traverses the iteration space by following the dependence arcs between the iterations.

Introduction Despite rapid increases in CPU performance, the primary obstacles to achieving higher performance in current processor organizations remain control and data hazards. An estimate [5] shows that the performance of single-chip microprocessors are improving at a rate of 80% annually, while DRAM speeds are improving at a rate of only 5-10% in that same amount time [5] [8]. The growing inability of the memory systems to keep up with the processors increases the importance of cache data re-use to reduce traffic to main memory and pre-fetching mechanisms to hide memory access latencies. These technological trends pose a challenge to interesting scientific and engineering applications whose data requirements are much larger than the processor's cache. Because scientific and engineering applications spend most of their execution time in loops, most of the effort in locality optimizations has focused on restructuring loops. Changing the iteration order by restructuring loops can significantly improve the performance of an application. Re-ordering iterations of a loop are conventionally done at compile time by applying transformations such as loop interchange, skewing and reversal. Unfortunately, compile-time transformations do not apply to certain types of loops because these transformations must be provably correct without knowing the values of the variables, forcing the compiler to make conservative assumptions. In this paper, we present a hybrid compile-time/run-time method to re-order loop iterations using a dependence-driven execution model that is loosely based on the concept of systolic arrays [9,11] and coarse grain dataflow [3]. In a dependence-driven execution, the system enables a block of iterations when the dependence constraints on those iterations are satisfied. The immediate execution of newly enabled iterations produces a depth-first traversal of the iteration space which improves data re-use. We maintain symbolic data-dependence information based on array subscript expression found in the body of the loop which is evaluated at run-time. This meta-computation on symbolic dependences allows us to avoid an early commitment to any specific order, giving the system a greater flexibility, which in turn increases the class of loops that can be optimized by re-ordering iterations. Furthermore, by maintaining dependence information during run-time, the run-time system can pre-fetch the sinks of dependences to hide the latency of memory accesses. Conventional wisdom suggests that determining the iteration order dynamically would add too much computational overhead. However, there are other overheads in addition to computational ones, such as those overheads caused by control and data hazards. On previous generations of computers with a more balanced memory system, the cost may indeed have been unjustified. On contemporary processors, the CPU cycles are relative cheap in comparison to memory cycles which can be up two orders of magnitude more expensive. This imbalance suggests that computational overhead for logic to avoid cache misses may not be significant if this logic can reduce traffic to memory, or hide the latency of memory operations. Elsewhere [16], we discussed the parallelism and scalability of a dependence-driven execution on a multiprocessor. In this paper, we evaluate the efficacy of run-time loop ordering to improve temporal locality in a contemporary uniprocessor. Background and Related Work Many important numerical applications in science and engineering consist of composite functions of the form where the f i 's are functions that are not necessarily distinct, D is some large data set, much greater than the processor's cache size, and n denotes the number of times the function sequence is to be applied. Implemented in imperative languages, the composite function would appear as a nested loop with each f i being expressed as a simple loop iterating over the data space D. The semantics of loops in these languages orders the iterations lexicographically with respect to the induction variables forcing the computation to traverse the data space in a strict function-at-a-time order. This execution order leads to a poor re-use of cache data: Because D does not fit entirely in the cache, the cache contains only the last c bytes of D upon completion of some function f i , forcing f i+1 to re-load every byte of D on to the cache. It may be possible and desirable to execute the iterations in a different order to improve locality. How do we determine which are the desirable and legal orders? How do we specify or express these orders for efficient execution? Much of the work on locality optimization relies on compile-time transformations to re-order the iterations of loops. Unfortunately, these transformations apply only to loops that are perfectly nested (that is, loops in which all assignment statements occur only in the innermost loop) or loops that can be transformed into perfectly nested loops. Put differently, compile-time transformations are applicable to the case where all the f i 's are the same in equation 1 above. In this section, we briefly review compile time loop transformations. In the following section, we describe how loop re-ordering can be extended to composite functions where not all f i 's are the same. To discuss compile transformation, it is useful to define the notion of an iteration space. An iteration space is an n-dimensional space of integers that models nested loops of depth n. A loop iteration is a point in the iteration space described by a vector I, (i 1 ,i 2 ,.i n ) where each i p is an index variable delimited by the iteration range for the corresponding loop at depth p. A linear loop transformation is a transformation from the original iteration space to another iteration space with some desired properties such as better locality or parallelism. A sequence of loop transformations can be modeled as a product of a non-singular matrices, each matrix making a transformation, such as skewing, loop interchange, reversal, etc. Thus finding the possible and desirable iteration order can be formulated as a search for a non-singular matrix with some objective function satisfying some set of constraints. These transformations are also called unimodular transformations because they preserve the volume of the iteration space of integers. A loop transformation is legal if the transformation preserves the dependence relations. If there is a dependence between two points I and J in the iteration space, then the difference between vector J and vector I, J-I is called the dependence distance vector. The set of distance vectors makes up the data dependences that determine the allowable re-ordering transformation. Based on these dependences, optimizing compilers may make the following transformations to improve locality: Loop Interchange: Loop Interchange [19,1] swaps an inner loop with an outer loop. Optimizing compilers will apply this transformation to improve memory locality if the interchange reduces the array access stride. Blocking (or Tiling): Blocking [20,12,17] takes advantage of applications with spatial locality by traversing a rectangle of iteration space at a time. If most of the memory accesses in the application are limited to addresses within the rectangle and this rectangle of data fits wholly in the cache, the processor will access the cache line multiple times before it leaves the cache. Skewing: Blocking may not be legal on some iteration spaces if the distance vectors contain negative distances. In some cases, skewing [19] can be applied to enable blocking transformation. Skewing traverses the iteration space diagonally in waves. Figure 1: Skewing and Tiling Transformations on Hyperbolic 1D PDE. From the transformed iteration space, the compiler generates code in the form of new loops. As an example, consider the hyperbolic 1D PDE. Figure 1 shows the dependences and the iteration space prior, to and after, the skewing and blocking transformation with a block size of two. While the generated code is more complex than the original, the new code has better locality and parallelism. A survey of compiler transformations for high performance computing can be found in [4,18]. Dependence-Driven Execution Given the transformed iteration space, compilers must generate code that describes the traversal over the entire iteration space. This early commitment to a specific order limits flexibility. More specifically, compile time transformations have the following limitations: Unimodular transformation do not apply to composite functions with multiple distinct functions, that is, they do not apply to a large class of imperfectly nested loops. Some dependences in loop iterations involve unknown user variables in the subscript expressions. For example, consider the following loop: for (I=0; I < N1; I++) { for (J=1; J < N2; J++) { The compiler cannot apply unimodular transformations without knowing the values of K and L, or at the very least knowing whether the values are negative or positive. Because compilers must give a static specification of iteration order, the code generated for the transformed iteration space can become complex, as in the example of skewing and blocking transformation in figure 1. The complex code with many levels of nesting and conditionals causes control hazards which reduces instruction level parallelism on contemporary processors. Furthermore, it is difficult for compilers to apply other optimizations on complex code. For example, none of the compilers on various architectures with which we experimented were able to apply loop unrolling to the code generated in figure 1. Furthermore, compile time linear loop transformations do not give us a general technique to automate pre-fetching data to hide memory access latencies. In this section, we describe the DUDE (Def-Use-Descriptor Environment) run-time system. DUDE is meant to be used either as a target for optimizing compilers or as a set of library calls that programmers can use directly 1 to optimize their code. The basic model is loosely based on the underlying concept of systolic arrays. Like systolic arrays, computation in DUDE consists of a large number of processing elements (cells) which are of the same type. However, for efficient computation on commercial processors, the granularity of computation in DUDE is much coarser. In our implementation, these cells are actually C++ objects that consist of an operation and a descriptor describing a region of data to which the operation is applied. We term these objects Iterates and an array of these Iterates make up an IterateCollection. The procedures in the cells of systolic arrays may consist of several alternative options. Similarly, operators in the Iterates of a IterateCollection may be overloaded. Like the cells of systolic arrays, Iterates are interconnected through links. But unlike the interconnection between cells of systolic arrays which are physical, hardwired links, the links in DUDE are symbolic expression of indices from the index space of IterateCollections. The expression for symbolic links, called dependence rule, is derived from array access patterns in the statements of the original loop, and therefore it summarizes dependences in the iteration space. The symbolic meta-computation on the dependence rule determines the path of execution through the iteration space. Also like the computation in systolic arrays, data is processed and transfered from one element to another by pipelining. Since there is only one physical processing element on a uniprocessor, there is no computational speedup due to pipelining. However, because of the temporal locality that this model offers, we can expect a performance improvement even on an uniprocessor. Unlike the computation in systolic arrays, the computations in DUDE are not synchronized by a global clock (and in that sense, our model is closer to wavefront arrays [10]). The asynchronous computation, together with the pipelining of the function applications, allows the system to apply multiple functions to the same block of data before that data block leaves the processor's cache. Figure 2: Dependence-driven Execution Model Describing Loops in DUDE A goal of the run-time system is to be able to optimize complex loops of the form shown in equation 1. To achieve these goals, we have taken an object-oriented loops and blocks of iterations are extensible first class objects which can be put together to describe complex loops. By putting together and specializing these objects, the user specializes the system to create a "software systolic array" for the application at hand. This object-oriented model is based on AWESIME [7] and the Chores [6] run-time systems. The following is a list of objects in DUDE: Data Descriptor: Data Descriptors describe a subsection of the data space. For example, a matrix can be divided into sub-matrices with each sub-matrix being defined by a data descriptor. The methods on this object, SX(), EX(), SY(), EY(), etc., retrieve the corners of the sub-matrix. Iterate: An Iterate is a tuple <data descriptor, operator>. The user specializes an Iterate by overloading the default operator with an application specific operator consisting of statements found in the body of a simple loop. The system applies the virtual operator to the data described by the descriptor. IterateCollection: An IterateCollection, as the name implies, is an array of Iterates. An IterateCollection represents a simple loop in a nested loop (or a simple function in a composite function) that performs an operation on the entire data space. The dimensionality of an IterateCollection is normally the same as that of the data array on which it operates. LOOP: LOOP is a template structure used to describe a composite function by putting together one or more IterateCollections. The user relies on the following methods provided by the LOOP object to glue together different IterateCollections and begin the computation: makes IterateCollection the nth simple function in the composite function. defines the symbolic link dep, from IterateCollection IC1 to IC2. This symbolic link is expressed in terms of a dependence rule with variables that range over the index space of the IterateCollections. Execute() executes the entire loop nest described by the loop descriptor. Because loops are objects, they can be created and defined at run-time, giving the system the flexibility to describe and compute complex loop nests. Figure 2 shows the basic model that the run-time system uses. Initially, the system pushes only the unconstrained Iterates onto to the system LIFO Queue. This allows the scheduler to pop off an Iterate to perform the operation with which that Iterate is associated. The completion of an Iterate can potentially enable other Iterates based on the data dependences specified in SetDependence() and based on what other Iterates have completed. This creates a cycle shown in figure 2 which the system repeats until the entire loop nest is completed. Example: Red/Black SOR We now describe a dependence-driven execution with respect to the example of Red/Black SOR which has the form , (Red (D) where N is the number of time steps. Note that since the loops for the Red and Blk operations are not nested within each other, they are not perfectly nested and hence unimodular transformation does not apply. Figure 3: Multi-Loop Dependences on Red/Black Figure 4: Red/Black SOR on DUDE Figure 3 shows the original code and the inter-loop dependences. The Red and the Blk operations in the loop body simply take the average of an element's neighboring point creating the dependence shown in the figure. Figure 4 shows the application as it would appear when written for DUDE. For this application, there are two Iterates, the RED and the BLK with corresponding BLK::main and RED::main methods that overload the operator to specialize the Iterate for this application. These Iterates compose the RedColl and BlkColl collections. Finally, these collections themselves are combined in the loop descriptor to create a composite function that iterates up to 10 iterations. The Execute() function of loop in figure 4 starts the system by pushing all of the initially unconstrained Iterates of the RedColl collection onto the system LIFO queue. In a dependence-driven execution, the memory locality of the entire execution of the nested loop is sensitive to the order in which the initially unconstrained Iterates are loaded. For applications which have block memory access patterns such as this one, the system loads the Iterates in Morton order [15]. Now the computation begins. After the initial Iterates have been loaded, the system scheduler pops off a (Red) Iterate from the system LIFO queue and applies the main operator to the data described by the descriptor for that Iterate. When completed, the system determines the list of sinks of the dependences arcs for that Iterate based on the dependence rule. For each sink, the system decrements the counter in the destination Iterates which at this point in the execution, are Blk Iterates. If the count is zero, the Iterate becomes unconstrained or enabled. The dependence satisfaction engine pushes these enabled Iterates onto the system LIFO queue. Because of the LIFO order, the next time the scheduler pops off an Iterate from the LIFO queue, it would be a Blk Iterate. Continuing, the completion of a Blk Iterate can further enable a Red Iterate from second time step, and so forth. This describes a depth-first traversal of the iteration space since a Blk operation can begin before all of the Red operations are completed. Note that using a FIFO queue would enforce a breadth-first traversal order of the iteration space as would the order produced by the original source loop. Figure 5: Iteration order for Hyperbolic 1D pde using DUDE Having described the run-time system, we are now in a position to compare it with the compile-time transformations. Because compile-time optimization cannot transform the loop structure for Red/Black SOR, we now revert back to the example of Hyperbolic 1D PDE to compare the iteration order of a loop nest as it would run on DUDE with a compile-time loop re-ordering. Figure 5 shows the snapshot of the iteration order of the Hyperbolic 1D PDE as it runs on DUDE. The code shown below the diagram is the body of the operator for the PDE Iterate. Note that this code is much simpler than the code required for skewing/tiling shown in figure 1. Simpler code with less conditionals runs more efficiently on contemporary processors with deep pipelines. It also enables the possibility of further optimizations. As shown in figure 1, there are really two orderings to consider in a dependence-driven execution: intra-Iterate order (indicated by numbers) and inter-Iterate order (enforced by arrows). While the inter-Iterate order is determined by the dependence rule, the order within an Iterate is exactly the same as that in the original source loop. Note that the given dependence rule causes a diagonal traversal of the Iterates, much like the skewing transformation. Support for Automated Pre-fetches So far we have discussed methods to reduce the number of cache misses. How do we hide memory access latency when cache misses are unavoidable? In this section, we discuss how DUDE inserts pre-fetch instructions to mask the memory accesses latencies with useful computation for cases when a cache miss is unavoidable. Figure 2 shows where the pre-fetch logic fits into the dependence-driven model. Just before the system executes the currently ready Iterate C the pre-fetch logic tries to predict what the next Iterate N to execute would be, based on C and the dependence rule. This prediction simply simulates what the dependence satisfaction engine does to enable new Iterates with the following exception: instead of pushing the newly generated Iterates N onto the LIFO queue, the pre-fetch logic invokes the pre-fetch command on the region of data that the Iterate N is associated with. This causes the data needed by N to be delivered to processor cache while the system is executing Iterate C with the intention that when the system finally executes N, the data for N would be available in the processor's cache. Experimental Results and Analysis We measured the performance of six applications: Red/Black SOR, Odd-Even sort, Multi-grid, Levialdi's Component labeling algorithm, Hyperbolic 1D PDE, and Vector chain addition. One would not ordinarily run some of these algorithms on a uniprocessor since more efficient scalar algorithms exist. Our aim is to ultimately run these algorithm on a multiprocessor but by conducting these experiments on a uniprocessor, we isolate the benefits of increased parallelism in a dependence-driven execution from the benefits of improved temporal locality. All experiments were conducted on a single processor DEC 21164 running at 290 MHz with 16KB of first level cache and 4 MB of second level cache. The cache penalties for the two caches are 15 cycles and 60 cycles respectively. All programs were compiled with DEC C++ (cxx) compiler with -O5 option. To determine where the cycles (cache miss, branch mis-predict, useful computation, etc.) were spent, we used the Digital Continuous Profiling Infrastructure (DCPI) [2] available on the Alpha platforms. DCPI runs on the background with low overhead (slowdown of 1-3%) and unobtrusively generates profile data for the applications running on the machines by sampling hardware performance counters available on the Alphas. The cycles are attributed to the following categories: Computation: This is the cycles spent in doing useful computation. Static Stalls: These are stalls due to static resource conflicts among instructions, dependencies to previous instructions, and conflicts in the functional units. D-Cache Miss Stalls: These are dynamic stalls caused by D-cache misses. Other Dynamic Stalls: Other sources of dynamic stalls are missed branch predictions, I-cache misses, ITB and DTB misses. Figure Cycles Breakdown for Various Applications Figure 6 shows a breakdown of where the cycles were spent for the six applications. Because some of the methods are not relevant to certain applications, some graphs compares fewer methods than others. All measurements are averages of 15 runs with negligible standard deviations. Red/Black SOR Figure 7: Analysis of SOR (2048x2048) Running on DUDE This application has good spatial locality since each element only accesses its neighboring elements. It also has the potential for temporal locality if we pipeline the iterations from different time steps. We compared using a dependence-driven execution to three other methods: unoptimized, tiling by row and tiling by block. We chose a matrix size of 2048x2048 (64-bit floats) to insure that the entire matrix did not entirely fit into the processor's cache. For each method, we used the optimal block size for that method which were 256x2048 for tiling by row, 256x256 for block and 32x32 for DUDE. Since the time steps in Red/Black SOR are normally controlled by a while loops, we use the following while (!done) { for (i=0; i<10; i++) Figure 6 shows that the Red/Black SOR using the unoptimized method spends as much as 76% of time on D-cache stalls. This is not surprising given how expensive memory accesses are relative to cpu cycles. Since tiling by row creates the same iteration order as the unoptimized case, there is little benefit in using tiling by row.Due to its access patterns (access of north, south, west, and east neighbors), tiling by block does a little better because of spatial locality. As shown in the figure, DUDE incurs the greatest overhead in terms of number of instructions executed. The run-time dependence satisfaction is partly responsible for these overheads. Another source of overheads are caused by the use of smaller block sizes (32x32) which increases the number of loops and hence the number of instructions. Comparing these overheads with the cycles spent in D-cache stalls, it is clear that these overheads are relatively insignificant. Neverthless, there is a tension between the overheads caused by smaller block sizes and the benefits of greater temporal locality; a smaller block allows the algorithm to explore deeper into time steps improving temporal locality, but it also increases the total overhead. The right hand side of figure 7 shows the effect of grain sizes on this application. To further analyze the cache behavior of SOR using various method, we used ATOM [14] to instrument this application. The left-hand side of the figure shows the total number of references, L1 cache, misses and L2 cache misses that we derived from instrumenting the executables. As expected, there are more memory references required for DUDE, but it also suffers the least from L2 cache misses. The working set for this application is too large for the dependence-driven execution to entirely fit in L1 causing a DUDE slight increase in L1 cache misses over the tiling by block method. Hyperbolic 1D PDE To compare the run-time method with skewing and blocking compiler transformation, we also measured the breakdown in cycles of the Hyperbolic 1D PDE, which is a wavefront computation with a perfectly nested loop. Figure 6 shows the performance of the three methods. Both the static and run-time re-ordering optimization significantly improve locality. The static skewing transformation has the best locality, but introduces more overhead than the dependence-driven execution. The control hazards introduced in the compiler generated code (right side of figure 1) increase the static stalls as shown in the figure. Further analysis of the skewing code using DCPI also revealed that this method suffered from resource conflict stalls due to all the integer arithmetic in the subscript expression required by the compiler-transformed code (see right side of figure 1). Component Labeling Levialdi's algorithm [13] for component labeling is used in image processing to detect connected components of a picture. It involves a series of phases, each phase consisting of changing a 1-pixel to a 0-pixel if its upper, left, and upper-left neighbors are 0-pixels. A 0-pixel is changed to a 1-pixel if both its upper and left neighbors are 1-pixels. Comparison of the performances of various methods are shown in figure 6. Again, we see that the higher overhead for determining the iteration order dynamically is small compared to the benefits of avoiding stalls. Odd-Even Sort Because the overheads in a dependence-driven execution are proportional to the number of dependence arcs emanating from an Iterate, and because this application has only two arcs per Iterate, DUDE dramatically outperforms the unoptimized method. We do not include the performance of tiling methods because tiling would not change the iteration order in this one-dimensional problem. Compile-time skewing transformation does not apply here because there are two functions, the odd and the even operations. Multi-Grid Multi-grid is another iterative PDE solver but it is an interesting one because it has five distinct operators (smooth even elements, smooth odd elements, restrict, form, and prolong) and different dependence relations between these operations. The dimension of the data space also changes depending on what level of the multi-grid pyramid you are in. Vector Chain Additions To measure the performance of DUDE when a purely static method can determine the same iteration order, we analyzed the performance of simply adding 14 vectors each of length 1048576 double floats. In the unoptimized version, two vectors were added in their entirety before adding the next vector. In the tiled versions, vectors were broken into chunks containing elements which gave the best performance. Elements 0.31 of all the 14 vectors were added before adding elements and so forth. Finally, in DUDE, we used a chunk size consisting of 512 elements with the Dependence Rule set to . While DUDE incurs slightly more overhead for determining the iteration order, it also benefits from data pre-fetches. Effect of Pre-fetching To study how well the system pre-fetches data, we 1) inspected the footprint of the execution, and 2) compared the cycles breakdown of the applications with or without enabling pre-fetch instruction. There are two pre-fetch instruction on the Alphas, fetch(M) instruction which is documented but not implemented on the 21164 and the instuction to load (ldt or ldl) to the zero register which is implemented but not documented. Since we were running our experiments on the 21164, we used the load to the zero register to generate our pre-fetches. By looking at the the footprint of an execution of what the system was pre-fetching and what the system was executing, we were able to verify that the system was indeed pre-fetching the next Iterate while executing the current one. However, the performance result that we observed showed that there was little benefit from pre-fetching for some of the applications at which we looked. The left sided of 7 gives us a clue as to why pre-fetching was not very effective on some applications. This figure shows that there were hardly any L2 misses, but a lot of L1 misses. This implies that the working set for a dependence-driven execution was too large to fit in L1. Pre-fetching the next iterate may help reduce L2 cache misses (which is not really the source of most stalls), but it does not help misses to L1 because of the conflicts with the working set of the current Iterate. We can reduce the size of the working set by reducing granularity, but this would increase the total overhead. Figure 8: Effect of Pre-fetch on Odd-Even Sort and Vector Chain Addition For applications with small working sets, we were able to get about a 7% performance improvement by pre-fetching as shown in figure 8. The working set for Odd-Even sort consists of only the left and right neighbors on an one-dimensional array while the working set of vector chain addition consist of only the vector containing the total sum and the current vector that is being added. We expect that the efficacy of pre-fetching would be greater if L1 was sufficiently big enough to fit the working set. Pre-fetching at the granularity of dependence-driven execution is more suited for distributed shared memory systems with larger caches and larger latencies. Nevertheless, our preliminary experiment indicates that pre-fetching can be automated in a dependence-driven execution. Conclusion Compile-time optimizations have the advantage that compilers can apply the optimizations with little overhead ( in most cases) to the total execution time. Run-time optimization have the advantage that there is more information available (assuming information from compile-time is kept until run-time). This information includes the state of computation such as values of the variables, dependences that are satisfied, and dependences that are yet to be satisfied. Having this information, the run-time can be more flexible. This flexibility comes at the cost of more instructions executed and more cpu cycles. However, on modern commercial machines, processor cycles are cheap. If the additional cycles can be used to use the memory system more effectively, then we can reduce overall execution time of an application. We have described a run-time system that uses symbolic dependence information to determine loop-order during run-time. We have also shown that a computation that is driven by dependences significantly reduces the number of cache miss stalls while adding relatively insignificant overhead. Furthermore, when cache misses cannot be avoided, we have shown that the system can also be used to pre-fetch data to avoid memory access latencies. As memory access costs become more expensive; cpu speeds continue to get faster; the size of caches continue to get larger; and as the mechanism for pre-fetching becomes better supported - the overhead for a dependence-driven execution becomes more justified. --R Automatic loop interchange. Continuous profiling: Where have all the cycles gone? Parallel processing with large-grain data flow techniques Compiler transformations for high-performance computing Keynote address. Enhanced run-time support for shared memory parallel computing A users guide to AWESIME: An object oriented parallel programming and simulation system. Computer Architecture: a Quantitative Approach. Systolic arrays (for VLSI). Wavefront array processors. VLSI Array Processors. The cache performance and optimization of blocked algorithm. On shrinking binary picture patterns. The Design and Analysis of Spatial Data Structures. Improving locality and parallelism in nested loops. High Performance Compilers for Parallel Computing. Optimizing supercompilers for supercomputers. More iteration space tiling. --TR VLSI array processors More iteration space tiling Computer architecture: a quantitative approach The design and analysis of spatial data structures The cache performance and optimizations of blocked algorithms Chores: enhanced run-time support for shared-memory parallel computing Link-time optimization of address calculation on a 64-bit architecture Continuous profiling On shrinking binary picture patterns Automatic loop interchange Compiler Transformations for High-Performance Computing Optimizing supercompilers for supercomputers --CTR Suvas Vajracharya , Dirk Grunwald, Dependence driven execution for multiprogrammed multiprocessor, Proceedings of the 12th international conference on Supercomputing, p.329-336, July 1998, Melbourne, Australia Suvas Vajracharya , Steve Karmesin , Peter Beckman , James Crotinger , Allen Malony , Sameer Shende , Rod Oldehoeft , Stephen Smith, SMARTS: exploiting temporal locality and parallelism through vertical execution, Proceedings of the 13th international conference on Supercomputing, p.302-310, June 20-25, 1999, Rhodes, Greece

systolic arrays;coarse-grain dataflow;temporal locality;dependence-driven;data locality;run-time systems;loop transformations

509647

Performance characteristics of gang scheduling in multiprogrammed environments.

Gang scheduling provides both space-slicing and time-slicing of computer resources for parallel programs. Each thread of execution from a parallel job is concurrently scheduled on an independent processor in order to achieve an optimal level of program performance. Time-slicing of parallel jobs provides for better overall system responsiveness and utilization than otherwise possible. Lawrence Livermore National Laboratory has deployed three generations of its gang scheduler on a variety of computing platforms. Results indicate the potential benefits of this technology to parallel processing are no less significant than time-sharing was in the 1960's.

Introduction Interest in parallel computers has been propelled by both the economics of commodity priced microprocessors and a growth rate in computational requirements exceeding processor speed increases. The symmetric multiprocessor (SMP) and massively parallel processor (MPP) architectures have proven quite popular, yet both suffer significant shortcomings when applied to large scale problems in multiprogrammed environments. The problems must first be recast in a form supporting a high levels of parallelism. Then to achieve the inherent parallelism and performance, it is necessary to concurrently schedule CPU resources to all threads and processes associated with each program. System throughput is frequently used as a metric of success; however, ease of use, good interactivity, and "fair" distribution of resources are of substantial importance to customers in a multiprogrammed environment. Efficiently harnessing the power of a multitude processors while satisfying customer requirements is a difficult proposition for schedulers. Most MPP computers provide concurrent scheduling through space-slicing schemes. A program is allocated a collection of processors and retains those processors until completion of the program. Scheduling is critical, yet each decision has an unknown impact upon the future: should a job be scheduled at the risk of blocking larger jobs later or should processors be left idle in anticipation of future arrivals? The lack of a time-slicing mechanism precludes good interactivity at high levels of utilization. Gang scheduling solves this dilemma by combining concurrent resource scheduling, space-slicing, and time-slicing. The impact of each scheduling decision is limited to a time-slice rather than the job's entire lifetime. Empirical evidence from gang scheduling on a Cray T3D installed at Lawrence Livermore National Laboratory (LLNL) demonstrates this additional flexibility can improve overall system utilization and responsiveness. Most SMP computers provide both space-sharing and time-slicing, but schedule each process independently. While good parallelism may be achieved this way on a lightly loaded system, that is a luxury rarely available. The purpose of gang scheduling in this environment is to improve the throughput of parallel jobs by concurrent scheduling, without degrading either overall system throughput or responsiveness. Moreover, this scheme can be extended to scheduling of parallel jobs across a cluster of computers in order to address larger problems. Gang scheduling of DEC Alpha computers at LLNL is explored in both stand-alone and cluster environments, and shown to fulfill these expectations. Overview of Gang Scheduling The term "gang scheduling" refers to all of a program's threads of execution being grouped into a gang and concurrently scheduled on distinct processors. Furthermore, time-slicing is supported through the concurrent preemption and later rescheduling of the gang [4]. These threads of execution are not necessarily POSIX threads, but components of a program which can execute simultaneously. The threads may span multiple computers and/or UNIX processes. Communications between threads may be performed through shared memory, message passing, and/or other means. Concurrent scheduling of a job's threads has been shown to improve the efficiency of both the individual parallel jobs and the system [3, 13]. The job's perspective is similar to that of a dedicated machine during the time-slices of its execution. Some reduction in I/O bandwidth may be experienced due to interference from other jobs, but CPU and memory resources should be dedicated. Job efficiency improvements results from a reduction in communications latency, promoting fine-grained parallelism. System efficiency can be improved by reductions in context switching, virtual memory paging, and cache refreshing. The advantages of gang scheduling are similar to those of time-sharing in uniprocessor systems. The ability to preempt jobs permits the scheduler to more efficiently utilize the system in several ways: Long running jobs and those with high resource requirements can be executed without monopolizing resources Interactive and other high priority jobs can be provided with near real-time response, even jobs with high resource requirements during periods of high system utilization Jobs with high processor requirements can be initiated in a timely fashion, without waiting for processors to be made available in a piecemeal fashion as other jobs terminate Low priority jobs can be executed, provided with otherwise unused resources, and preempted when higher priority jobs become available High system utilization can be sustained under a wide range of workloads Job preemption does incur some additional overhead. The CPU resources sacrificed for context switching is slight and explored later in the paper. The increase in storage requirements is possibly substantial. All running or preempted jobs must have their storage requirements satisfied simultaneously. Preempted jobs must also vacate memory for other jobs, with the memory contents written to disk. The acquisition of additional disks in order to increase utilization of CPU and memory resources is likely to be cost-effective, but this need must be considered. Other Approaches to Parallel Scheduling Most MPP computers use the variable partitioning paradigm. In variable partitioning, the job specifies its processor count requirement at submission time. Processors allocated to a job are retained until its termination. The inability to preempt a job can prevent the timely allocation of resources to interactive or other high priority jobs. A shortage of jobs with small processor counts can result in the incomplete allocation of processors. Conversely, the execution a job with high processor requirements can be delayed by fragmentation, which necessitates the accumulation of processors in a piecemeal fashion as multiple jobs with smaller processor counts terminate. Variable partitioning can result in poor resource utilization due to resource fragmentation [10, 18], processors left idle in anticipation of high priority job arrival [12], or processors left idle in order to accommodate jobs with substantial resource requirements. Another option is dynamic partitioning, in which the operating system determines the number of processors to be allocated to each job. While dynamic partitioning does possess the allure of high utilization and interactivity [11, 15, 16], it can make program development significantly more difficult. The program is required to operate efficiently without knowledge of processor count until execution time. The variable processor count also causes execution time variability, which may be unacceptable for workloads containing very long running jobs or even moderate size jobs with high priority. MPP Scheduling The variable partition paradigm prevalent on MPP architectures and its lack of job preemption makes responsiveness particularly difficult to provide. In order to ensure responsive service, some MPP computers divide resources into "pools" reserved for interactive jobs, batch jobs, or available to any job type. These pools can also be used to reserve portions of the computer for specific customers on some systems. Since jobs can not normally span multiple pools, partitioning the computer in this fashion reduces the maximum problem size which can be addressed while fragmentation reduces scheduling flexibility and system utilization. The optimal configuration places all resources into a single pool available to any job type, assuming that support for resource allocation and interactivity can be provided by other means. Scheduling Space-sharing and time-sharing are the norm on SMP computers, providing both good interactivity and utilization. Most computers schedule each process independent, which works well for a workload consisting of many independent processes. However, the solution of large problems is dependent upon the use of parallel jobs, which suffer significant inefficiencies without the benefit of concurrent scheduling [3, 13]. Parallel job development efforts at the National Energy Research Supercomputer Center (NERSC) illustrates difficulties in parallel job scheduling [2]. In order to encourage parallel job development, NERSC provided dedicated time to parallel jobs on a Cray C90 computer. Several of the parallel jobs running in dedicated mode were able to achieve a parallel performance (CPU time/wall time) over 15.5 on this 16 CPU machine and approach 10 GFlops per second of the 16 GFlops per second peak speed. While all batch jobs were suspended during the dedicated period, interactive jobs were initially permitted to execute concurrently with the parallel job. Several instances were observed of a single compute-bound interactive program reducing the parallel job's throughput and system utilization by almost 50 percent. Suspension of interactive jobs was found to be necessary in order to achieve reasonable parallel job performance. The benefits of SMP computers can be scaled for larger problems by clustering. Efficient execution of parallel jobs in this environment requires coordination of scheduling across the entire cluster. If communications between the threads consisted exclusively of message passing, it is possible to base scheduling upon this message traffic and achieve a high level of parallelism [13, 14]. The LLNL workload makes extensive use of shared memory for communications, preventing us from pursuing this strategy. LLNL Workload Characterization LLNL customers have long relied upon interactive supercomputing for program development and rapid throughput of jobs with short to moderate execution times. While some of this work can be performed on smaller computers, most customers use the target platform for reasons of problem size, hardware compatibility, and software compatibility. LLNL began its transition to parallel computing with the acquisition of a BBN TC2000 in 1989. Additional parallel computers at LLNL have included the Cray C90, Cray T3D, Meiko CS-2, and IBM SP2. Many of our problems are of substantial size and have been well parallelized. The LLNL Cray T3D workload is typical of that on our other parallel computers. The model at LLNL has 256 processors, each with 64 megabytes of DRAM. All processors are configured into a single pool available to any job. The LLNL Cray T3D is configured to permit interactive execution of jobs up to 64 processors and 2 hours execution time. Interactive jobs account for 67 percent of all jobs executed and consume 13 percent of all CPU resources delivered. The interactive workload typically ranges from 64 to 256 processors (25 to 100 percent of the computer's processors) during working hours and drops to a negligible level in the late night hours. Peak interactive workloads reach 320 processors for a 25 percent oversubscription rate. Timely execution of interactive jobs is dependent upon the preemption of batch jobs and, in extreme cases, time-sharing processors among interactive jobs. Large jobs account for a high percentage of resources utilized on the Cray T3D as shown in Table 1. Memory requirements are quite variable, but most jobs use between 42 and 56 megabytes per processor. Due to Cray T3D hardware constraints, a contiguous block of processors with a specific shape must be made available to execute a job [1], making fragmentation a particularly onerous problem. Gang scheduling permits us to preempt jobs as needed to execute large jobs in a timely fashion with minimal overhead. Job Size (CPU CPU Utilization (%): Table 1: CPU Utilization by Job Size on Cray T3D at LLNL Most programs use either PVM (Parallel Virtual Machine) or MPI (Message Passing Interface) libraries for communications between the threads. Although the Cray T3D has a distributed memory architecture, it will support a shared memory programming model permitting a job to read and write memory local of another processor assigned to that job. This shared memory programming model has considerably lower overhead than message passing and is widely used. A small number of programs utilize a combination of both paradigms: shared memory within the Cray T3D and message passing to communicate with threads or serial program components executing on a Cray YMP front-end. This multiple paradigm model is expected to be common on our DEC Alpha cluster and IBM SP2 containing SMP nodes. LLNL Gang Scheduler Design Strategy The LLNL customers expect rapid response, even on a heavily utilized multiprogrammed computer; however the need for rapid response is not uniform across the entire workload. In order to provide interactivity where required and minimize the overhead of context switching, we divide our workload into six different classes. Each job class has significantly different scheduling characteristics as described below: Express jobs have been deemed by management to be mission critical and are given rapid response and optimal throughput. Interactive jobs require rapid response time and very good throughput during extended working hours. The jobs' response time and throughput can be reduced at other hours for the sake of improved system utilization and throughput of production jobs. Debug jobs require rapid response time during extended working hours. The jobs' response time can be reduced at other hours for the sake of improved system utilization. Debug jobs can not be preempted on the Cray T3D. Production jobs do not require rapid response time, but should receive very good throughput at night and on weekends. Benchmark jobs do not require rapid response time, but can not be preempted. Standby jobs have low priority and are suitable for absorbing otherwise idle compute resources. The default classes are production for batch jobs, debug for totalview debugger initiated jobs, and interactive for other jobs directly initiated from a user terminal. Jobs can be placed in the express class only by the system administrator. Benchmark and standby classes can be specified by the user at job initiation time. Each job class has a number of scheduling parameters, including relative priority and processor limit. There are also several system-wide scheduling parameters such as aggregate processor limit for all gang scheduled jobs. The scheduling parameters can be altered in real-time, which permits periodic reconfiguration. LLNL emphasizes interactity during extended work hours (7:30 AM to 10:00 PM) and throughput at other times. The gang scheduler daemon itself can also be updated at any time. Upon receipt of the appropriate signal and completion of any critical tasks, the daemon writes its state to a file and initiates a new daemon program. Sockets are utilized for user communications, most of which occur at job initiation. Sockets are also used for job state change requests and scheduling parameter changes, which are rare. Job and processor state information is written periodically to a data file, which is globally readable. This file is read by the "gangster" application, which is the users' window into gang scheduling status. We have delegated the issue of "fair" resource distribution to other systems, including the Centralized User Bank (CUB) [8] and Distributed Production Control System (DPCS) [17]. These systems provide resource allocation and accounting capabilities to manage the entire computing environment with a single tool. Both systems exercise coarse grained control by the scheduling of batch jobs and fine grained control by adjusting nice values of both interactive and batch processes. Our gang scheduler is designed to recognize changes in nice value and schedule accordingly. Jobs of classes which can be preempted will automatically be changed to standby class when at high nice value and returned to their requested class upon reduction in nice value. This mechanism has proven to function very well in managing resource distribution while minimizing interdependence of the systems. Significant differences exist in the three LLNL gang scheduler implementations. These differences were largely the result of architectural differences, but some were based upon experiences with previous implementations. Each implementation is described below with results. BBN TC2000 In order to pioneer parallel computing at LLNL, a BBN TC2000 with 126 processors was acquired in 1989. The TC2000 has a shared memory architecture and originally supported space-sharing only. The gang scheduler for the TC2000 reserves all resources at system startup and controls all resource scheduling from that time [6, 7]. User programs require no code changes to communicate with the gang scheduler. However, the program must load with a modified version of the mandatory parallel program initiation library. Rather than securing resources directly from the operating system, this library secures resources from the gang scheduler daemon. The program may also increase or decrease its processor count during execution by explicitly notifying the gang scheduler daemon. Otherwise, a job must continue execution on the specific processors originally assigned to it. Time-sharing is performed by the use of sleep and wake signals, issued concurrently to all threads of a job. This mechanism can provide for very rapid context switches, on the order of milliseconds. This implementation did have "fair share" mechanism with the objective of providing each user with access to comparable resources. The gang scheduler would determine resource allocation to be made for each ten second time-slice with the objective of insuring interactivity, equitable distribution of resources, and high utilization. While the gang scheduler implemented for the TC2000 was able to provide good responsiveness and throughput, some shortcomings should be noted. Despite the shared memory architecture to the TC2000, it was not possible to relocate threads in order to better balance the continuously changing workload. The scheduler could react to workload changes only at time-slice boundaries, which limited responsiveness. The user based fair share mechanism was abandoned in later implementations due to the availability of an independent and more comprehensive resource allocation system. Cray T3D The Cray T3D is a massively parallel computer incorporating DEC alpha 21064 microprocessors capable of 150 MFLOPS peak performance. Each processor has its own local memory. The system is configured into nodes, consisting of two processors with their local memory and a network interconnect. The nodes are connected by a bidirectional three-dimensional torus communications network. There are also four synchronization circuits (barrier wires) connected to all processors with a tree shaped interconnect [1]. getting into great detail, the T3D severely constrains processor and barrier wire assignments. Jobs must be allocated a processor count which is a power of two, with a minimum of two processors (one node). A job's can be built to run with any valid processor count, but its processor count can not change after execution begins. The processors allocated to a job must have a specific shape with specific dimensions for a given problem size. For example, an allocation of processors must be made with a contiguous block of eight processors in the X direction, two processors in the Y direction, and two processors in the Z direction. Furthermore, the possible locations of the processor assignments are restricted. These very specific shapes and locations for processor assignment are the result of the barrier wire structure. Jobs must be allocated one of the four barrier wires when initiated. The barrier wire assignment to a job can not change if the job is relocated and, under some circumstances, two jobs sharing a single barrier wire may not be located adjacent to each other. Prior to the installation of a gang scheduler on our Cray T3D, we were forced to make several significant sacrifices in order to satisfy our customers' need for interactivity [5, 9]. The execution time of batch jobs was restricted to insure "reasonable" responsiveness to jobs with large processor requirements and interactive jobs, although one may argue that delays on the order of hours may not be reasonable. Most batch jobs were limited to four hours. One batch queue permitted execution times up to 19 hours, but this queue was enabled during only brief periods on weekends and holidays. Since jobs could not be preempted, our batch workload was dramatically reduced at 4:00 AM. As batch jobs completed, their released resources might remain unused by any interactive job for many hours. At times of heavy interactive use, the initiation of an interactive job had to wait for other jobs to terminate and release resources. The processor allocation restrictions also made for severe fragmentation problems. While interactivity was generally good, the processor utilization rate of 33 percent was considered unacceptable. The Cray T3D gang scheduler implementation has a several significant differences from that of the BBN TC2000. Since the initiation of all parallel jobs on the T3D is conducted by a single application program, we placed a wrapper around this to communicate with the gang scheduler daemon. No changes to the user application or scripts were required. The fixed time-slice period was replaced with an event driven scheduler with the ability to react instantly to changes in workload. The allocation of specific processors and barrier wires can dramatically effect performance on the T3D, so substantial effort was placed in developing software to optimize these selections. Processor selection criterion includes best-fit packing, placement of jobs with similar scheduling characteristics in close proximity, and minimizing contention of the barrier wire circuits. Context switching logic also required substantial modification. The BBN TC2000 supported memory paging, while the Cray T3D lacks support for paging. Cray T3D job preemption results in the entire memory image of the preempted job being written to disk before the processors can be reassigned, which requires about one second per preempted processor. In the case of a 64 processor jobs being context switched, about one minute is required to store the preempted job's context and another minute to load the state of another job of similar size. The T3D gang scheduler calculates a value for each job including its processor count, job type, location (disk or memory) to make job preemption and initiation decisions. Additional information associated with each job class and used in this calculation include: maximum wait time, processor limit, and do-not-disturb time multiplier. The minimum time-slice for a job is the product of the job's processor count and its class' do-not-disturb time multiplier. While the minimum time-slice mechanism does reduce responsiveness, it prevents the costly thrashing of jobs between memory and disk and is critical for maintaining a high utilization level. The Cray T3D gang scheduler has been in operation since March 1996. We were able to dramatically modify the batch environment to fully subscribe the machine during the day and oversubscribe it at night by as much as 100 percent. The normal mode of operation in the daytime is for interactive class jobs to preempt production class jobs shortly after initiation. The interactive jobs then usually continue execution until completion without preemption. At night, the processors are oversubscribed temporarily and only for the express purpose of executing larger jobs (128 or 256 processor). This strategy results in only about eight percent of all jobs ever being preempted. The batch queue for long running jobs has been substantially reconfigured: its time limit has been increased from 19 to 40 hours, its maximum processor allocation (for all running jobs in the queue) increased from 64 to 128 processors, and is enabled at all times. System utilization increased substantially and weekly CPU utilization rates over 96 percent have been sustained. Figure 1 shows monthly CPU utilization for a period of 15 months. Three different schedulers were utilized over this period. UNICOS MAX is the native Cray T3D operating system from Cray Research. DJM or Distributed Job Manager is a parallel job scheduler originally developed by the Minnesota Supercomputer Center and substantially modified by Cray Research. The LLNL developed gang scheduler is also shown. The CPU utilization reported is that during which a CPU is actually assigned to a program which is memory resident. CPU utilization is reduced by three things: 1. Context switch time, a CPU is unavailable while a program's image is being transferred between disk and memory 2. Sets of processors on which no job can fit, a packing problem 3. Insufficient work, particularly on weekends Figure 1: Cray T3D CPU Utilization Benchmarks run against both UNICOS MAX and an early prototype of the LLNL gang scheduler showed a 21 percent improvement in interactive job throughput with no reduction in aggregate throughput. The cost of moving jobs' state between memory and disk to provide responsiveness was fully compensated for by more efficient packing of the processor torus. The current implementation has additional performance enhancements and should show an aggregate throughput improvement of a few percent. Figure 2 shows a gangster display of typical Cray T3D daytime utilization. Note that only 12 of the 256 processors (six of 128 nodes) are not assigned to some job. A total of ten interactive and debug class jobs are running and using 148 of the processors. Over half of the batch workload is currently paged out for these interactive jobs. Left side of display shows the contents of each node (two processors). Each letter indicates the job and a period indicates an unused node. The right side of the display describes each job. The W field shows the barrier wire used. The MM:SS field shows the total execution time accumulated. The ST field shows the job's state: i=swapping in, N=new job, not yet assigned nodes or barrier wire, o=swapping out, O=swapped out, R=running, W=awaiting commencement of execution, assigned nodes and barrier wire. The gang scheduler state information is written to disk at 15 second intervals. The gangster display is updated at the same rate. gangster - 13:25 - 42377 pet_test 4 700 0 R 5:24 a a . h a a . f Dbug a - susan 81280 aaaa8123 a a d . a a d g Prod s - mahdi 83818 ge 64 420 0 R1462:43 a a . h Prod u - eduardo 79879 new 64 -1 a a . w Prod m - delarub 79890 new 64 -1 a a d . Prod n - tomas 80331 mdterrep Figure 2: Typical daytime gangster display on Cray T3D While the utilization rate is quite satisfactory, responsiveness is also of great importance. Responsiveness can be quantified by slowdown, the ratio of total time spent in the system to the run time. During the three week period of July 23 through August 12, 2659 interactive jobs were executed using a total of 44.5 million CPU-seconds and 1328 batch jobs were executed using a total of 289.5 million CPU-seconds. The slowdown of the aggregate interactive workload was 18%, which is viewed quite favorably. Further investigation shows a great deal of variation in slowdown. Most longer running interactive jobs enjoy slowdowns of only a few percent. Interactive jobs executed during the daytime typically begin execution within seconds and are not preempted. Interactive jobs executed during late night and early morning hours experienced slowdowns as high as 1371 (a one second job delayed for about 23 minutes). However the computer is configured for high utilization and batch job execution during these hours, so high slowdowns are not unexpected. DEC Alpha The Digital Unix 4.0D operating system includes a "class scheduler", which is essentially a very fine grained fair share scheduler. The class scheduler permits processes, process groups, or sessions to be grouped in a class. The class can then be allocated some share of CPU resources. For example, the eight threads of a job could be placed into a single class and allocated 80 percent of the CPU resources on a ten CPU computer. While the class scheduler does not explicitly reserve eight CPUs for the eight threads of this parallel job, the net effect is very close. The operating system also recognizes advantage in keeping a process on a particular CPU to avoid refreshing its cache. We have found the class scheduler actually delivers over percent of the CPU resources as desired for gang scheduling with minimal thread migration between processors. For the job which fails to sustain its target number of runable threads, the class scheduler will allocate the CPU resources to other jobs in order to sustain high overall system utilization. Since parallel jobs are not normally registered with the Digital UNIX operating system, the gang scheduler relies upon explicit registration through the use of a library. While it is highly desirable that user code modification for gang scheduler be avoided, that is impossible to achieve at this time. Imbedding a few gang scheduler remote procedure calls directly in MPI and PVM libraries would free many users from having to modify their programs. Presently, the gang scheduled application must be modified with a few simple library calls, including: Register job with gang scheduler: A global gang scheduler job ID is returned Register resource requirements: CPU, memory, and disk space requirements are specified for the job on each computer to be gang scheduled Register the processes: Associate a specific process, process group, or session with the gang scheduler job ID Since it is necessary to coordinate activities across multiple computers, the DEC Alpha gang scheduler returned to the fixed time-slice model used on the BBN TC2000. In order to manage these time-slices across multiple computers, the concept of "tickets" was introduced. These tickets represent specific resource allocations at specific times and are issued only for jobs spanning multiple computers. Jobs which execute only on a single computer are not pre-issued tickets, but are managed by that computer's gang scheduler daemon which makes scheduling decisions at the start of each time-slice. This design permits each computer to operate as independently as possible, while permitting the gang scheduling of jobs across the cluster as needed. The tickets are managed by the gang scheduler daemon on each computer in the cluster and are associated with a job for its lifetime. A job may be given additional tickets or have some revoked depending upon changes in overall system load. The job is also permitted to alter its resource requirements during execution. A change in the number of CPU required for a job may result in revoked or additional tickets. The gangster display program was re-written in Tcl/Tk in order to provide a richer environment. Detailed information on system and job status now includes a history of CPU, real memory, and virtual memory use which is updated at time-slice boundaries. This information can be helpful for system administrators and programmers tuning their systems. For example, if a program's CPU allocation and CPU use are substantially different, the desired level of parallelism is not being achieved and the matter should be investigated. A dramatic changes in real memory use for a program during its execution may indicate substantial paging, which also warrants investigation. Figures 3 and 4 show displays of system and program status. C machine status Figure 4: Gangster display of DEC job status In order to insure that good responsiveness is preserved for all jobs, this gang scheduler permits processors to be reserved for non-gang scheduled jobs. This prevents gang scheduled jobs from reserving all processors on a computer for the entire time-slice, which could prevent logins or other interactions for a significant period. The gang scheduler daemon also reacts to changes in the overall workload, insuring resources are fairly distributed to all active jobs. Gangster also permits several system-wide operations to a gang scheduled job including: suspend, resume, kill, and change job class. For example, the kill job operation will send kill signals to all processes registered to a gang scheduled job on all computers running that job. Conclusions Gang scheduling has been shown to provide an attractive multiprogrammed environment for multiprocessor systems in several ways: Parallel jobs can be provided with access to all required resources simultaneously, providing the illusion of a dedicated environment Interactive and other high priority jobs can be provided with rapid response and excellent throughput Jobs with large resource requirements can be initiated rapidly, without having to wait for multiple jobs to terminate and release resources A high level of utilization can be maintained under a wide range of workloads Experience on the Cray T3D has been overwhelmingly positive. It can now sustain processor utilization rates over 95 percent through the course of a week while providing the aggregate interactive workload with a slowdown of less than 20 percent. Such performance give this distributed memory MPP a range of performance spanning large-scale parallel applications to general purpose interactive computing. Experience on the DEC Alpha environment has also been very favorable during our testing period. A very rich and interactive environment is available on these fast SMPs, while true supercomputing class problems can be addressed by harnessing the power of a cluster for parallel jobs. Our plans call for bringing two 80 CPU DEC Alpha 8400 clusters under the control of gang scheduling in the Fall of 1997. --R Cray Research Inc. Dedicated Computing on a YMP/C916. Effective Distributed Scheduling of Parallel Workloads. A Survey of Scheduling in Multiprogrammed Parallel Systems. Improved Utilization and Responsiveness with Gang Scheduling. Timesharing massively parallel machines. Centralized User Banking and User Administration on UNICOS. Gang Scheduler - Timesharing the Cray T3D" A dynamic processor allocation policy for multiprogrammed shared-memory multiprocessors Analysis of non-work-conserving processor partitioning policies Dynamic Coscheduling on Workstation Clusters. Comparing Gang Scheduling with Dynamic Space Sharing on Symmetric Multiprocessors Using Automatic Self-Allocating Threads (ASAT) Process control and scheduling issues for multiprogrammed shared-memory multiprocessors Distributed Production Control System. A new graph approach to minimizing processor fragmentation in hypercube multiprocessors. --TR Process control and scheduling issues for multiprogrammed shared-memory multiprocessors A two-dimensional buddy systems for dynamic resource allocation in a partitionable mesh connected system A dynamic processor allocation policy for multiprogrammed shared-memory multiprocessors Effective distributed scheduling of parallel workloads A New Graph Approach to Minimizing Processor Fragmentation in Hypercube Multiprocessors Comparing Gang Scheduling with Dynamic Space Sharing on Symmetric Multiprocessors Using Automatic Self-Allocating Threads (ASAT) Analysis of Non-Work-Conserving Processor Partitioning Policies Demand-Based Coscheduling of Parallel Jobs on Multiprogrammed Multiprocessors Improved Utilization and Responsiveness with Gang Scheduling --CTR Adrian T. Wong , Leonid Oliker , William T. C. Kramer , Teresa L. Kaltz , David H. Bailey, ESP: a system utilization benchmark, Proceedings of the 2000 ACM/IEEE conference on Supercomputing (CDROM), p.15-es, November 04-10, 2000, Dallas, Texas, United States Jung-Lok Yu , Jin-Soo Kim , Seung-Ryoul Maeng, A runtime resolution scheme for priority boost conflict in implicit coscheduling, The Journal of Supercomputing, v.40 n.1, p.1-28, April 2007 Atsushi Hori , Hiroshi Tezuka , Yutaka Ishikawa, Highly efficient gang scheduling implementation, Proceedings of the 1998 ACM/IEEE conference on Supercomputing (CDROM), p.1-14, November 07-13, 1998, San Jose, CA Shahaan Ayyub , David Abramson, GridRod: a dynamic runtime scheduler for grid workflows, Proceedings of the 21st annual international conference on Supercomputing, June 17-21, 2007, Seattle, Washington Gyu Sang Choi , Jin-Ha Kim , Deniz Ersoz , Andy B. Yoo , Chita R. Das, A comprehensive performance and energy consumption analysis of scheduling alternatives in clusters, The Journal of Supercomputing, v.40 n.2, p.159-184, May 2007 Gyu Sang Choi , Jin-Ha Kim , Deniz Ersoz , Andy B. Yoo , Chita R. Das, Coscheduling in Clusters: Is It a Viable Alternative?, Proceedings of the 2004 ACM/IEEE conference on Supercomputing, p.16, November 06-12, 2004 Bin Lin , Peter A. Dinda, VSched: Mixing Batch And Interactive Virtual Machines Using Periodic Real-time Scheduling, Proceedings of the 2005 ACM/IEEE conference on Supercomputing, p.8, November 12-18, 2005

multiprogramming;gang scheduling;time-slicing;parallel system;space-slicing;scheduling

509649

A common data management infrastructure for adaptive algorithms for PDE solutions.

This paper presents the design, development and application of a computational infrastructure to support the implementation of parallel adaptive algorithms for the solution of sets of partial differential equations. The infrastructure is separated into multiple layers of abstraction. This paper is primarily concerned with the two lowest layersof this infrastructure: a layer which defines and implements dynamic distributed arrays (DDA), and a layer in which several dynamic data and programming abstractions are implemented in terms of the DDAs. The currently implemented abstractions are those needed for formulation of hierarchical adaptive finite difference methods, hp-adaptive finite element methods, and fast multipole method for solution of linear systems. Implementation of sample applications based on each of these methods are described and implementation issues and performance measurements are presented.

Introduction This paper describes the design and implementation of a common computational infrastructure to support parallel adaptive solutions of partial differential equations. The motivations for this research are: 1. Adaptive methods will be utilized for the solution of almost all very large-scale scientific and engineering models. These adaptive methods will be executed on large-scale heterogeneous parallel execution environments. 2. Effective application of these complex methods on scalable parallel architectures will be possible only through the use of programming abstractions which lower the complexity of application structures to a tractable level. 3. A common infrastructure for this family of algorithms will result in both, enormous savings in coding effort and a more effective infrastructure due to pooling and focusing of effort. The goal for this research is to reduce the intrinsic complexity of coding parallel adaptive algorithms by providing an appropriate set of data structures and programming abstractions. This infrastructure has been developed as a result of collaborative research among computer scientists, computational scientists and application domain specialists working on three different projects: An DARPA project for hp-adaptive computational fluid dynamics and two NSF sponsored Grand Challenge projects, one on numerical relativity and the other on composite materials. 1.1 Conceptual Framework Figure 1.1: Hierarchical Problem Solving Environment for Parallel Adaptive Algorithms for the Solution of PDEs Figure 1 is a schematic of our perception of the structure of a problem solving environment (PSE) for parallel adaptive techniques for the solution of partial differential equations. This paper is primarily concerned with the lowest two layers of this hierarchy and how these layers can support implementation of higher levels of abstraction. The bottom layer of the hierarchical PSE is a data-management layer. The layer implements a Distributed Dynamic Array (DDA) which provides array access semantics to distributed and dynamic data. The next layer is a programming abstractions layer which adds application semantics to DDA objects. This layer implements data abstractions such as grids, meshes and trees which underlie different solution methods. The design of the PSE is based on a separation of concerns and the definition of hierarchical abstractions based on the separation. Such a clean separation of concerns [1] is critical to the success of an infrastructure that can provide a foundation for several different solution methods. In particular the PSE presented in this paper supports finite difference methods based on adaptive mesh refinement, hp-adaptive finite element methods, and adaptive fast multipole methods. 1.2 Overview This paper defines the common requirements of parallel adaptive finite difference and finite element methods for solution of PDEs and fast multipole solution of linear systems, and demonstrates that one data management system based on Distributed Dynamic Arrays (DDA) can efficiently meet these common requirements. The paper then describes the design concepts underlying DDAs and sketches implementations of parallel adaptive finite difference, finite element, multipole solutions using DDAs based on a common conceptual basis as the common data management system. Performance evaluations for each method are also presented. The primary distinctions between the DDA-based data management infrastructures and other packages supporting adaptive methods are (1) the separation of data management and solution method semantics and (2) the separation of addressing and storage semantics in the DDA design. This separation of concerns enables the preservation of application locality in multi-dimensional space when it is mapped to the distributed one-dimensional space of the computer memory and the efficient implementation of dynamic behavior. The data structure which has traditionally been used for implementation of these problems has been the multi-dimensional array. Informally an array consists of: (1) an index set (a lattice of points in an n-dimensional discrete space), (2) a mapping from the n-dimensional index set to one dimensional storage, and (3) a mechanism for accessing storage associated with indices. A DDA is a generalization of the traditional array which targets the requirements of adaptive algorithms. In contrast to regular arrays, the DDA utilizes a recursively defined hierarchical index space where each index in the index space may be an index space. The storage scheme then associates contiguous storage with spans of this index space. The relationship between the application and the array is defined by deriving the index space directly from the n-dimensional physical domain of the application. Problem Description Adaptive algorithms require definition of operators on complex dynamic data structures. Two problems arise: (1) the volume and complexity of the bookkeeping code required to construct and maintain these data structures overwhelms the actual computations and (2) maintaining access locality under dynamic expansion and contraction of the data requires complex copying operations if standard storage layouts are used. Implementation on parallel and distributed execution environments adds the additional complexities of partitioning, distribution and communication. Application domain scientists and engineers are forced to create complex data management capabilities which are far removed from the application domain. Further, standard parallel programming languages do not provide explicit support for dynamic distributed data structures. Data management requirements for the three different adaptive algorithms for PDEs are described below. 2.1 Adaptive Finite Difference Data Management Requirements Finite difference methods approximate the solution of the PDE on a discretized grid overlaid on the n-dimensional physical application domain. Adaptation increases the resolution of the discretization in required regions by refining segments of the grid into finer grids. The computational operations on the grid may include local stencil based operations at all levels of resolution, transfer operations between levels of resolution and global linear solves. Thus the requirement for the data-management system for adaptive finite difference methods is seamless support of these operations across distribution and refining and coarsening of the grid. Storage of the dynamic grid as an array where each point of the grid is mapped to a point in a hierarchical index space is natural. Then the refinements and coarsening of the grid become traversal of the hierarchy of the hierarchical index space. 2.2 Adaptive Finite Element Data Management Requirement The finite element method requires storage of the geometric information defining the mesh of elements which spans the application domain. Elements of the linear system arising from finite element solutions are generally computed on the fly so that they need not be stored. HP-adaptive finite element methods adapt by partitioning elements into smaller elements (h refinement) or by increasing the order of the polynomial approximating the solution on the element (p refinement). Partitioning adds new elements and changes relationships among elements. Changing the approximation function enlarges the descriptions of the elements. The data-management requirements for hp-adaptive finite elements thus include storage of dynamic numbers of elements of dynamic sizes. These requirements can be met by mapping each element of the mesh to a position in a hierarchical index space which is associated with a dynamic span of the one-dimensional storage space. Partitioning an element replaces a position in the index space by a local index space representing the expanded mesh. 2.3 Adaptive Fast Multipole Data Management Requirements Fast multipole methods partition physical space into subdomains. The stored representation of the subdomain includes a charge configuration of the subdomain and various other descriptive data. Adaptation consists of selectively partitioning subdomains. The elements are often generated on the fly so that storage of the values of the forces and potential need not be stored. The requirement for data management capability is therefore similar to that of adaptive finite element methods. A natural mapping is to associate each subdomain with a point in a hierarchical index space. 2.4 Requirements Summary It should be clear that an extended definition of an array where each element can itself be an array and where the entity associated with each index position of the array can be an object of arbitrary and variable size provides one natural representation of the data management requirements for all of adaptive finite element, adaptive finite difference and adaptive fast multipole solvers. The challenge is now to demonstrate an efficient implementation of such an array and to define implementations of each method in terms of this storage abstraction. The further and more difficult challenge is an efficient parallel/distributed implementation of such an array. 3 Distributed Dynamic Data-Management The distributed dynamic data-management layer of the PSE implements Distributed Dynamic Arrays (DDAs). This layer provides pure array access semantics to dynamically structured and physically distributed data. DDA objects encapsulate distribution, dynamic load-balancing, communications, and consistency management, and have been extended with visualization and analysis capabilities. There are currently two different implementations of the data management layer, both built on the same DDA conceptual framework: (1) the Hierarchical Dynamic Distributed Array (HDDA) and (2) the Scalable Dynamic Distributed Array (SDDA). The HDDA is a hierarchical array in that each element of the array can recursively be an array; it is a dynamic array in that each array (at each level of the hierarchy) can expand and contract at run-time. Instead of hierarchical arrays the SDDA implements a distributed dynamic array of objects of arbitrary and heterogeneous types. This differentiation results from the differences in data management requirements between structured and unstructured meshes. For unstructured meshes it is more convenient to incorporate hierarchy information into the programming abstraction layer which implements the unstructured mesh. These two implementations derived from a common base implementation and can and will be re-integrated into a single implementation in the near future. The arrays of objects defined in the lowest layer are specialized by the higher layers of the PSE to implement application objects such as grids, meshes, and tree. A key feature of a DDA based on the conceptual framework described following is its ability to extract out the data locality requirements from the application domain and maintain this locality despite its distribution and dynamic structure. This is achieved through the application of the principle of separation of concerns [1] to the DDA design. An overview of this design is shown in Figure 3.1. Distributed dynamic arrays are defined in the following subsection. Figure 3.1: DDA Design - Separation of Concerns -> Hierarchical Abstractions 3.1 Distributed Dynamic Array Abstraction The SDDA and HDDA are implementations of a distributed dynamic array. The distributed dynamic array abstraction, presented in detail in [2], is summarized as follows. In general, an array is defined by a data set D, an index space I, and an injective An array is dynamic if elements may be dynamically inserted into or removed from its data set D. An array is distributed if the elements of its data set D are distributed. A data set D is simply a finite set of data objects. An index space I is a countable set of indices with a well-defined linear ordering relation, for example the set of natural numbers. Two critical points of this array abstraction are (1) that the cardinality of the data set D is necessarily equal to or less than the cardinality of the index space I and (2) that each element of the data set uniquely maps to an element of the index space. 3.2 Hierarchical Index Space and Space-Filling Curves An application partitions its N-dimensional problem domain into a finite number of points and/or regions. Each region can be associated with a unique coordinate in the problem domain. Thus the "natural" indexing scheme for such an application is a discretization of these coordinates. Such an index space can defined as: I 1 -I 2 -I N where each I j corresponds to the discretization of a coordinate axis. An index space may be hierarchical if the level of discretization is allowed to vary, perhaps in correspondence with a hierarchical partitioning of the problem domain. The I j components of hierarchical index space could be defined as (d,(i 1 ,i 2 ,.,i d )), where d is the depth of the index. Recall that an index space requires a well-defined linear ordering relation. Thus the "natural" N dimensional hierarchical index space must effectively be mapped to a linear, or one-dimensional, index space. An efficient family of such maps are defined by space-filling curves (SFC) [3]. One such mapping is defined by the Hilbert space-filling curve, illustrated in Figure 3.2. In this mapping a bounded domain is hierarchically partitioned into regions where the regions are given a particular ordering. In theory the partitioning "depth" may be infinite and so any finite set of points in the domain may be fully ordered. Thus an SFC mapping defines a hierarchical index space, of theoretically infinite depth, for any application domain which can be mapped into the SFC "bounding box". Figure 3.2: Hilbert Space-Filling Curve The "natural" index space of the SFC map efficiently defines a linear ordering for all points and/or subregions of the application's problem domain. Given such a linear ordering the these subregions can be distributed among processors by simply partitioning the index space. This partitioning is easily and efficiently obtained from a partitioning the linearly ordered index space such that the computational load of each partition is roughly equal. Figure 3.3, Space-Filling Curve Partitioning, illustrates this process for an irregularly partitioned two dimensional problem domain. Note in Figure 3.3 that the locality preserving property of the Hilbert SFC map generates "well-connected" subdomains. Figure 3.3: Space-Filling Curve Partitioning 3.3 DDA Implementation An array implementation provides storage for objects of the data set D and access to these objects through a converse function F D. The quality and efficiency of an array implementation are determined by the correlation between storage locality and index locality and by the expense of the converse function F -1 . For example a conventional one-dimensional FORTRAN array has both maximal quality and efficiency. A DDA implements distributed dynamic arrays where storage for objects in the data set D is distributed and dynamic, and the converse function F -1 provides global access to these objects. A DDA's storage structure and converse function consists of two components: (1) local object storage and access and (2) object distribution. The HDDA and SDDA implementations of a DDA use extendible hashing [4] & [5] and red-black balanced binary trees [6] respectively for local object storage and access. An application instructs the DDA as to how to distribute data by defining a partitioning of index space I among processors. Each index in the index space is uniquely assigned to a particular processor i -> P. The storage location of a particular data object is now determined by its associated index d -> i -> P. Thus the storage location of any dynamically created data object is well-defined. Each DDA provides global access to distributed objects by transparently caching objects between processors. For example, when an application applies a DDA's converse function d) if the data object d is not present on the local processor the data object is transparently copied from its owning processor into a cache on the local processor. 3.4 Locality, Locality, Locality! A DDA's object distribution preserves locality between object storage and the object's global indices. Given the "natural" index space of the space-filling curve map and the corresponding domain partitioning, a DDA's storage locality is well-correlated with geometric locality, as illustrated in Figure 3.4. Figure 3.4: Locality, Locality, Locality! In Figure 3.4 an application assigns SFC indices to the data objects associated with each subregion of the domain. A DDA stores objects within a span of indices in the local memory of a specified processor. Thus geometrically local subregions have their associated data objects stored on the same processor. Application Programming Interface (API) The DDA application programming interface consists a small set of simple methods which hide the complexity of the storage structure and make transparent any required interprocessor communication. These methods include: GET d <- F -1 (i) INSERT D <- D + d new REMOVE D <- D - d old REPARTITION Forcing a redistribution of objects PUT/LOCK/UNLOCK Cache coherency controls Implementation of these methods varies between the SDDA and HDDA; however, the abstractions for these methods are common to both versions of the DDA. 3.6 DDA Performance The DDA provides object management services for an application's distributed dynamic data structures. This functionality introduces an additional overhead cost when accessing local objects. For example, local object access could be accomplished directly via 'C' pointers instead of DDA indices; however, note that the pointer to a data object may be invalidated under data object redistribution. This overhead cost is measured for the SDDA version of the DDA, which uses a red-black balanced binary tree algorithm for local object management. The overhead cost of the local get, local insert, and local remove method is measured on an IBM RS6000. The local get method retrieves a local data object associated with an input index value, i.e. the converse function. The local insert method inserts a new data object associated with its specified index (d new ,F(d new )) into the local storage structure. The local remove method removes a given data object from the local storage structure. The computational time of each method is measured given the size of the existing data structure, as presented in Figure 3.5, SDDA Overhead for Local Methods. Figure 3.5: SDDA Overhead for Local Methods The overhead cost of local get method, as denoted by the middle line in Figure 3.5, increases logarithmically from 2 microseconds for an empty SDDA to 8 microseconds for an SDDA containing one million local data objects. This slow logarithmic growth is as expected for a search into the SDDA's balanced binary tree. The local insert method performs two operations: (1) search for the proper point in the data structure to insert the object and (2) modification of the data structure for the new object. The cost of the insert method is a uniform "delta" cost over the get operation. Thus the overhead cost of modifying the data structure for the new data object independent of the size of the SDDA. Note that the overhead cost of the local remove method, as denoted in Figure 3.5 by the lowest line, is also independent of the size of the SDDA. 4 Method Specific Data & Programming Abstractions The next level of the PSE specializes DDA objects with method specific semantics to create high-level programming abstractions which can be directly used to implement parallel adaptive algorithms. The design of such abstractions for three different classes of adaptive solution techniques for PDEs, hierarchical, dynamically adaptive grids, hp-adaptive finite elements and dynamic tree (dynamic trees are the data abstractions upon which the fast multipole methods are implemented) are descried below. 4.1 Hierarchical Adaptive Mesh-Refinement Problem Description Figure 4.1: Adaptive Grid Hierarchy - 2D (Berger-Oliger AMR Scheme) Dynamically adaptive numerical techniques for solving differential equations provide a means for concentrating computational effort to appropriate regions in the computational domain. In the case of hierarchical adaptive mesh refinement (AMR) methods, this is achieved by tracking regions in the domain that require additional resolution and dynamically overlaying finer grids over these regions. AMR-based techniques start with a base coarse grid with minimum acceptable resolution that covers the entire computational domain. As the solution progresses, regions in the domain requiring additional resolution are tagged and finer grids are overlayed on the tagged regions of the coarse grid. Refinement proceeds recursively so that regions on the finer grid requiring more resolution are similarly tagged and even finer grids are overlayed on these regions. The resulting grid structure is a dynamic adaptive grid hierarchy. The adaptive grid hierarchy corresponding to the AMR formulation by Berger & Oliger [7] is shown in Figure 4.1. Distributed Data-Structures for Hierarchical AMR Two basic distributed data-structures have been developed, using the fundamental abstractions provided by the HDDA, to support adaptive finite-difference techniques based on hierarchical AMR: (1) A Scalable Distributed Dynamic Grid (SDDG) which is a distributed and dynamic array, and is used to implement a single component grid in the adaptive grid hierarchy; and (2) A Distributed Adaptive Grid Hierarchy (DAGH) which is defined as a dynamic collection of SDDGs and implements the entire adaptive grid hierarchy. The SDDG/DAGH data-structure design is based on a linear representation of the hierarchical, multi-dimensional grid structure. This representation is generated using space-filling curves described in Section 3 and exploits the self-similar or recursive nature of these mappings to represent a hierarchical DAGH structure and to maintain locality across different levels of the hierarchy. Space-filling mapping functions are also used to encode information about the original multi-dimensional space into each space-filling index. Given an index, it is possible to obtain its position in the original multi-dimensional space, the shape of the region in the multi-dimensional space associated with the index, and the space-filling indices that are adjacent to it. A detailed description of the design of these data-structures can be found in [8]. Figure 4.2: SDDG Representation - Figure 4.3: DAGH Composite Representation SDDG Representation: A multi-dimensional SDDG is represented as a one dimensional ordered list of SDDG blocks. The list is obtained by first blocking the SDDG to achieve the required granularity, and then ordering the SDDG blocks based on the selected space-filling curve. The granularity of SDDG blocks is system dependent and attempts to balance the computation-communication ratio for each block. Each block in the list is assigned a cost corresponding to its computational load. Figure 4.2 illustrates this representation for a 2-dimensional SDDG. Partitioning a SDDG across processing elements using this representation consists of appropriately partitioning the SDDG block list so as to balance the total cost at each processor. Since space-filling curve mappings preserve spatial locality, the resulting distribution is comparable to traditional block distributions in terms of communication overheads. DAGH Representation: The DAGH representation starts with a simple SDDG list corresponding to the base grid of the grid hierarchy, and appropriately incorporates newly created SDDGs within this list as the base grid gets refined. The resulting structure is a composite list of the entire adaptive grid hierarchy. Incorporation of refined component grids into the base SDDG list is achieved by exploiting the recursive nature of space-filling mappings: For each refined region, the SDDG sub-list corresponding to the refined region is replaced by the child grid's SDDG list. The costs associated with blocks of the new list are updated to reflect combined computational loads of the parent and child. The DAGH representation therefore is a composite ordered list of DAGH blocks where each DAGH block represents a block of the entire grid hierarchy and may contain more than one grid level; i.e. inter-level locality is maintained within each DAGH block. Figure 4.3 illustrates the composite representation for a two dimensional grid hierarchy. The AMR grid hierarchy can be partitioned across processors by appropriately partitioning the linear DAGH representation. In particular, partitioning the composite list to balance the cost associated to each processor results in a composite decomposition of the hierarchy. The key feature of this decomposition is that it minimizes potentially expensive inter-grid communications by maintaining inter-level locality in each partition. Figure 4.4: SDDG/DAGH Storage Data-structure storage is maintained by the HDDA described in Section 3. The overall storage scheme is shown in Figure 4.4. Programming Abstractions for Hierarchical AMR Figure 4.5: Programming Abstraction for Parallel Adaptive Mesh-Refinement We have developed three fundamental programming abstractions using the data-structures described above that can be used to express parallel adaptive computations based on adaptive mesh refinement (AMR) and multigrid techniques (see Figure 4.5). Our objectives are twofold: first, to provide application developers with a set of primitives that are intuitive for expressing the application, and second, to separate data-management issues and implementations from application specific operations. Grid Geometry Abstractions: The purpose of the grid geometry abstractions is to provide an intuitive means for identifying and addressing regions in the computational domain. These abstractions can be used to direct computations to a particular region in the domain, to mask regions that should not be included in a given operation, or to specify region that need more resolution or refinement. The grid geometry abstractions represent coordinates, bounding boxes and doubly linked lists of bounding boxes. Coordinates: The coordinate abstraction represents a point in the computational domain. Operations defined on this class include indexing and arithmetic/logical manipulations. These operations are independent of the dimensionality of the domain. Bounding Boxes: Bounding boxes represent regions in the computation domain and is comprised of a triplet: a pair of Coords defining the lower and upper bounds of the box and a step array that defines the granularity of the discretization in each dimension. In addition to regular indexing and arithmetic operations, scaling, translations, unions and intersections are also defined on bounding boxes. Bounding boxes are the primary means for specification of operations and storage of internal information (such as dependency and communication information) within DAGH. Bounding Boxes Lists: Lists of bounding boxes represent a collection of regions in the computational domain. Such a list is typically used to specify regions that need refinement during the regriding phase of an adaptive application. In addition to linked-list addition, deletion and stepping operation, reduction operations such as intersection and union are also defined on a BBoxList. Grid Hierarchy Abstraction: The grid hierarchy abstraction represents the distributed dynamic adaptive grid hierarchy that underlie parallel adaptive applications based on adaptive mesh-refinement. This abstraction enables a user to define, maintain and operate a grid hierarchy as a first-class object. Grid hierarchy attributes include the geometry specifications of the domain such as the structure of the base grid, its extents, boundary information, coordinate information, and refinement information such as information about the nature of refinement and the refinement factor to be used. When used in a parallel/distributed environment, the grid hierarchy is partitioned and distributed across the processors and serves as a template for all application variables or grid functions. The locality preserving composite distribution [9] based on recursive Space-filling Curves [3] is used to partition the dynamic grid hierarchy. Operations defined on the grid hierarchy include indexing of individual component grid in the hierarchy, refinement, coarsening, recomposition of the hierarchy after regriding, and querying of the structure of the hierarchy at any instant. During regriding, the re-partitioning of the new grid structure, dynamic load-balancing, and the required data-movement to initialize newly created grids, are performed automatically and transparently. Grid Function Abstraction: Grid Functions represent application variables defined on the grid hierarchy. Each grid function is associated with a grid hierarchy and uses the hierarchy as a template to define its structure and distribution. Attributes of a grid function include type information, and dependency information in terms of space and time stencil radii. In addition the user can assign special (FORTRAN) routines to a grid function to handle operations such as inter-grid transfers (prolongation and restriction), initialization, boundary updates, and input/output. These function are then called internally when operating on the distributed grid function. In addition to standard arithmetic and logical manipulations, a number of reduction operations such as Min/Max, Sum/Product, and Norms are also defined on grid functions. GridFunction objects can be locally operated on as regular FORTRAN 90/77 arrays. 4.2 Definition of hp-Adaptive Finite Element Mesh The hp-adaptive finite element mesh data structure consists of two layers of abstractions, as illustrated in Figure 4.6. The first layer consists of the Domain and Node abstractions. The second layer consists of mesh specific abstractions such as Vertex, Edge, and Surface, which are specializations of the Node abstraction. Figure 4.6: Layering of Mesh Abstraction A mesh Domain is the finite element application's specialization of the SDDA. The Domain uses the SDDA to store and distribute a dynamic set of mesh Nodes among processors. The Domain provides the mapping from the N-dimensional finite element domain to the one-dimensional index space required by a DDA. A finite element mesh Node associates a set of finite element basis functions with a particular location in the problem domain. Nodes also support inter-Node relationships, which typically capture properties of inter-Node locality. Specializations of the finite element mesh Node for a two-dimensional problem are summarized in the following table and illustrated Figure 4.7. Mesh Object Reference Location Relationships Vertex vertex point Edge midside point Vertex endpoints Element "owners" Irregular edge constraints Element centroid point Edge boundaries Element refinement heirarchy Figure 4.7: Mesh Object Relationships Extended relationships between mesh Nodes are obtained through the minimul set of relationships given above. For example: Extended Relationship Relationship "Path" Element ->Vertex Element <-> Edge -> Vertex Element <-> Element Element <-> Edge <-> Element (normal) Element <-> Element Element <-> Edge <-> Edge <-> Element (constrained) Finite element h-adaptation consists of splitting elements into smaller elements, or merging previously split elements into a single larger elements. Finite element p-adaptation involves increasing or decreasing the number of basis functions associated with the elements. An application performs these hp-adaptations dynamically in response to an error analysis of a finite element solution. HP-adaptation results in the creation of new mesh Nodes and specification of new relationships. Following an hp-adaptation the mesh partitioning may lead to load imbalance, as such the application may repartition the problem. A DDA significantly simplifies such dynamic data structure update and repartitioning operations while insuring data structure consistency throughout these operations. 4.3 Adaptive Trees An adaptive distributed tree requires two main pieces of information. First it needs a tree data structure with methods for gets, puts, and pruning nodes of the tree. This infrastructure requires pointers between nodes. Second an adaptive tree needs an estimation of the cost associated with each node of the tree in order to determine if any refinement will take place at that node. With these two abstractions, an algorithm can utilize an adaptive tree in a computation. At this point, we are developing a distributed fast multipole method based on balanced trees, with the goal of creating a mildly adaptive tree in the near future. Adaptive trees could be defined in either of the DDAs. The implementation described here is done using the SDDA. All references in the tree are made through a generalization of the pointer concept. These pointers are implemented as indices into the SDDA, and access is controlled by accessing the SDDA data object with the appropriate action and index. This control provides a uniform interface into a distributed data structure for each processor. Thus, distributed adaptive trees are supported on the SDDA. The actual contents of a node includes a list of items. 1. An index of each node derived from the geometric location of the node. 2. Pointers to a parent and to children nodes. 3. An array of coefficients used by the computation. 4. A list of pointers to other nodes with which the given node interacts. All of this information is stored in a node, called a subdomain. The expected work for each subdomain is derived from the amount of computation to be performed as specified by the data. Adaptivity can be determined on the basis of the expected work of a given node, relative to some threshold. In addition, since each node is registered in the SDDA, we can also compute the total expected work per processor. By collecting the total expected work per processor with the expected work per subdomain, a simple load balance can be implemented by repartitioning the index space. 5 Application Codes There follow sketches of applications expressed in terms of each of the parallel adaptive mesh refinement method, the parallel hp-adaptive method and the parallel many-body problem each built on programming abstractions built upon a DDA. 5.1 Numerical Relativity using Hierarchical AMR A distributed and adaptive version of H3expresso 3-D numerical relativity application has been implemented using the the data-management infrastructure presented in this paper. The H3expresso 3-D numerical relativity application is developed at the National Center for Supercomputing Applications (NCSA), University of Illinois at Urbana, has H3expresso (developed at National Center for Supercomputing Applications (NCSA), University of Illinois at Urbana) is a "concentrated" version of the full H version 3.3 code that solves the general relativistic Einstein's Equations in a variety of physical scenarios [10]. The original H3expresso code is non-adaptive and is implemented in FORTRAN 90. Representation Overheads Figure 5.1: DAGH Overhead Evaluation The overheads of the proposed DAGH/SDDG representation are evaluated by comparing the performance of a hand-coded, unigrid, Fortran 90+MPI implementation of the H3expresso application with a version built using the data-management infrastructure. The hand-coded implementation was optimized to overlap the computations in the interior of each grid partition with the communications on its boundary by storing the boundary in separate arrays. Figure 5.1 plots the execution time for the two codes. The DAGH implementation is faster for all number of processors. Composite Partitioning Evaluation The results presented below were obtained for a 3-D base grid of dimension 8 X 8 X 8 and 6 levels of refinement with a refinement factor of 2. Figure 5.2: DAGH Distribution: Snap-shot I - Figure 5.3: DAGH Distribution: Snap-shot II Figure 5.4: DAGH Distribution: Snap-shot III - Figure 5.5: DAGH Distribution: Snap-shot IV Load Balance: To evaluate the load distribution generated by the composite partitioning scheme we consider snap-shots of the distributed grid hierarchy at arbitrary times during integration. Normalized computational load at each processor for the different snap-shots are plotted in Figures 5.2-15.5. Normalization is performed by dividing the computational load actually assigned to a processor by the computational load that would have been assigned to the processor to achieve a perfect load-balance. The latter value is computed as the total computational load of the entire DAGH divided by the number of processors. Any residual load imbalance in the partitions generated can be tuned by varying the granularity of the SDDG/DAGH blocks. Smaller blocks can increase the regriding time but will result in smaller load imbalance. Since AMR methods require re-distribution at regular intervals, it is usually more critical to be able to perform the re-distribution quickly than to optimize each distribution. Communications: Both prolongation and restriction inter-grid operations were performed locally on each processor without any communication or synchronization. Partitioning Overheads Table 5.1: Dynamic Partitioning Overhead Partitioning is performed initially on the base grid, and on the entire grid hierarchy after every regrid. Regriding any level l comprises of refining at level l and all level finer than generating and distributing the new grid hierarchy; and performing data transfers required to initialize the new hierarchy. Table 5.1 compares the total time required for regriding, i.e. for refinement, dynamic re-partitioning and load balancing, and data-movement, to the time required for grid updates. The values listed are cumulative times for 8 base grid time-steps with 7 regrid operations. 5.2 HP-Adaptive Finite Element Code An parallel hp-adaptive finite element code for computational fluid dynamics is in development. This hp-adaptive finite element computational fluid dynamics application has two existing implementations: (1) a sequential FORTRAN code and (2) a parallel FORTRAN code with fully a duplicated data structure. The hp-adaptive finite element data structure has a complexity so great that is was not tractable to distribute the FORTRAN data structure. As such the parallel FORTRAN implementation is not scalable due to the memory consumed on each processor by duplicating the data structure. To make tractable the development of a fully distributed hp-adaptive finite element data structure is was necessary to achieve a separation of concerns between the complexities of the hp-adaptive finite element data structure and complexities of distributed dynamic data structures in general. This separation of concerns in the development of a fully distributed hp-adaptive finite element data structure provided the initial motivation for developing the SDDA. The organization of the new finite element application is illustrated in Figure 5.6. At the core of the application architecture is the index space. The index space provides a very compact and succinct specification of how to partition the application's problem among processors. This same specification is used to both distribute the mesh structure through the SDDA and to define a compatible distribution for the vectors and matrices formed by the finite element method. Figure Finite Element Application Architecture The finite element application uses a second parallel infrastructure which supports distributed vectors and matrices, as denoted in the lower right corner of Figure 5.6. The current release of this infrastructure is documented in [11]. Both DDA and linear algebra infrastructures are based upon the common abstraction of an index space. This commonality provides the finite element application with uniform abstraction for specifying data distribution. 5.3 N-Body Problems General Description Figure 5.7: Data flow in the fast multipole algorithm The N-body particle problems arising in various scientific disciplines appear to require an computational method. However, once a threshold in the number of particles is surpassed, approximating the interaction of particles with interactions between sufficiently separated particle clusters allows the computational effort to be substantially reduced. The best known of these fast summation approaches is the fast multipole method [12], which, under certain assumptions, gives a method of O(N) The fast multipole method is a typical divide and conquer algorithm. A cubic computational domain is recursively subdivided into octants. At the finest level, the influence of the particles within a cell onto sufficiently separated cells is subsumed into a multipole series expansion. These multipole expansions are combined in the upper levels, until the root of the oct-tree contains a multipole expansion. Then local series expansions of the influence of sufficiently separated multipole expansions on cells are formed. Finally, in a reverse traversal of the oct-tree the contributions of cells are distributed to their children (cf. Figure 5.7). The algorithm relies on a scaling property, which allows cells on the scale of children to be sufficiently separated when they were too close on the scale of the current parents. At the finest level the influence of these sufficiently separated cells is taken into account together with the interaction of the particles in the remaining closeby cells. For a more mathematically oriented description of the shared memory implementation with or without periodic boundary conditions we refer to [13][14] and the references therein. Figure 5.7 shows that the fast multipole algorithm is readily decomposed into three principal stages: 1. populating the tree bottom-up with multipole expansions 2. converting the multipole expansions to local expansions 3. distribute local expansions top-down These three stages have to be performed in sequential order, but it is easily possible to parallelize each of the stages individually. For the second stage the interaction set of a cell is defined as the set of cells which are sufficiently separated from the cell, but their parents are not sufficiently separated from the cells parent cell. The local expansion about the center of a cell is found by adding up all the influences from the cells of the interaction set. For each cell these operations are found to be completely independent of each other. But notice that the majority of communication requirements between different processors is incurred during this stage. Hence, an optimization of the communication patterns during the second stage can account for large performance gains. A more detailed analysis of the (non-adaptive) algorithm reveals that optimal performance should be attained when each leaf-cell contains an optimal number of particles, thereby balancing the work between the direct calculations and the conversion of the multipole expansions to the local expansions in the second stage. Distributed Fast Multipole Results Figure 5.8: Total Execution Time on SP2 vs. Problem Size for Multiple Levels of Approximation Next we describe preliminary performance results for the distributed fast multipole method implemented on the HDDA. We verified the fast multipole method by computing an exact answer for a fixed resolution of the computational domain. Test problems were constructed by storing one to a few particles chosen randomly per leaf cell. This comparison was repeated for several small problems until we were certain that the algorithm behaved as expected. Performance measurements were taken from a 16 node IBM SP2 parallel computer running AIX Version 4.1. The total run times using 3, 4, and 5 levels of approximation on a variety of problem sizes for 1, 2, 4, and 8 processors are presented in Figure 5.8. We also plot the expected O(N) run time aligned with the 8 processor results. Each curve represents the total execution time for a fixed level of spatial resolution while increasing the number of particles in the computation. Two major components of each run time are an approximation time and a direct calculation for local particles. The approximation time is a function of the number of levels of resolution and is fixed for each curve. The direct calculation time grows as O(N 2 ) within a given curve. The problem size for which these two times are equal represents the optimal number of particles stored per subdomain. This optimal problem size appears as a knee in the total execution time curves. The curves for 8 processors align relatively well with the O(N) run time for all problem sizes. Thus, the algorithm appears scalable for the problems considered. Furthermore, these results show that 1 processor has the longest time and 8 processors have the shortest time, which indicates that some speedup is being attained. Ideally, one would expect that processors achieve a speedup of 8. We have presented results for a fast multipole method which demonstrates the O(N) computational complexity. These results exhibit both scalability and speedup for a small number of processors. Our preliminary results focused on validity and accuracy. The next step is to performance tune the algorithm. 6 Conclusion and Future Work The significant conclusions demonstrated herein are: 1. That there is a common underlying computational infrastructure for a wide family of parallel adaptive computation algorithms. 2. That a substantial decrease in effort in implementation for these important algorithms can be attained without sacrifice of performance through use of this computational infrastructure. There is much further research needed to complete development of a robust and supportable computational infrastructure for adaptive algorithms. The existing versions of DDA require extension and engineering. The programming abstractions for each solution method require enriching. It is hoped to extend the programming abstraction layer to support other adaptive methods such as wavelet methods. There is a need to do many more applications to define the requirements for the programming abstraction layer interfaces. Acknowledgements This research has been jointly sponsored by the Argonne National Laboratory Enrico Fermi Scholarship awarded to Manish Parashar, by the Binary Black-Hole NSF Grand Challenge (NSF ACS/PHY 9318152), by ARPA under contract DABT 63-92-C-0042, and by the NSF National Grand Challenges program grant ECS-9422707. The authors would also like to acknowledge the contributions of J-rgen Singer, Paul Walker and Joan Masso to this work. --R System Engineering for High Performance Computing Software: The HDDA/DAGH Infrastructure for Implementation of Parallel Structured Adaptive Mesh Refinement A Parallel Infrastructure for Scalable Adaptive Finite Element Methods and its Application to Least Squares C-infinity Collocation Database System Concepts Linear Hashing: a New Tool for File and Table Addressing Adaptive Mesh-Refinement for Hyperbolic Partial Differential Equations Distributed Dynamic Data-Structures for Parallel Adaptive Mesh-Refinement On Partitioning Dynamic Adaptive Grid Hierarchies Hyperbolic System for Numerical Relativity Robert van de Geijn The rapid evaluation of potential fields in particle systems The Parallel Fast Multipole Method in Molecular Dynamics Parallel Implementation of the Fast Multipole Method with Periodic Boundary Conditions --TR Algorithms The parallel fast multipole method in molecular dynamics Database System Concepts On Partitioning Dynamic Adaptive Grid Hierarchies --CTR Yun He , Chris H. Q. Ding, Coupling Multicomponent Models with MPH on Distributed Memory Computer Architectures, International Journal of High Performance Computing Applications, v.19 n.3, p.329-340, August 2005 Faith E. Sevilgen , Srinivas Aluru, A unifying data structure for hierarchical methods, Proceedings of the 1999 ACM/IEEE conference on Supercomputing (CDROM), p.24-es, November 14-19, 1999, Portland, Oregon, United States Sumir Chandra , Manish Parashar, Towards autonomic application-sensitive partitioning for SAMR applications, Journal of Parallel and Distributed Computing, v.65 n.4, p.519-531, April 2005 S. Chandra , X. Li , M. Parashar, Engineering an autonomic partitioning framework for Grid-based SAMR applications, High performance scientific and engineering computing: hardware/software support, Kluwer Academic Publishers, Norwell, MA, 2004 Karen Devine , Bruce Hendrickson , Erik Boman , Matthew St. John , Courtenay Vaughan, Design of dynamic load-balancing tools for parallel applications, Proceedings of the 14th international conference on Supercomputing, p.110-118, May 08-11, 2000, Santa Fe, New Mexico, United States Andrew M. Wissink , Richard D. Hornung , Scott R. Kohn , Steve S. Smith , Noah Elliott, Large scale parallel structured AMR calculations using the SAMRAI framework, Proceedings of the 2001 ACM/IEEE conference on Supercomputing (CDROM), p.6-6, November 10-16, 2001, Denver, Colorado J. M. Malard , R. D. Stewart, Distributed dynamic hash tables using IBM LAPI, Proceedings of the 2002 ACM/IEEE conference on Supercomputing, p.1-11, November 16, 2002, Baltimore, Maryland Valerio Pascucci , Randall J. Frank, Global static indexing for real-time exploration of very large regular grids, Proceedings of the 2001 ACM/IEEE conference on Supercomputing (CDROM), p.2-2, November 10-16, 2001, Denver, Colorado James C. Browne , Madulika Yalamanchi , Kevin Kane , Karthikeyan Sankaralingam, General parallel computations on desktop grid and P2P systems, Proceedings of the 7th workshop on Workshop on languages, compilers, and run-time support for scalable systems, p.1-8, October 22-23, 2004, Houston, Texas Johan Steensland , Sumir Chandra , Manish Parashar, An Application-Centric Characterization of Domain-Based SFC Partitioners for Parallel SAMR, IEEE Transactions on Parallel and Distributed Systems, v.13 n.12, p.1275-1289, December 2002

fast multipole methods;problem solving environment;adaptive mesh-refinement;HP-adaptive finite elements;parallel adaptive algorithm;distributed dynamic data structures

509912

New results on monotone dualization and generating hypergraph transversals.

This paper considers the problem of dualizing a monotone CNF (equivalently, computing all minimal transversals of a hypergraph), whose associated decision problem is a prominent open problem in NP-completeness. We present a number of new polynomial time resp. output-polynomial time results for significant cases, which largely advance the tractability frontier and improve on previous results. Furthermore, we show that duality of two monotone CNFs can be disproved with limited nondeterminism (more precisely, in polynomial time with $O(\log^2 n)$ suitably guessed bits). This result sheds new light on the complexity of this important problem.

INTRODUCTION # Part of the work carried out while visiting TU Wien. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for pro-t or commercial advantage and that copies bear this notice and the full citation on the -rst page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior speci-c permission and/or a fee. STOC'02, May 19-21, 2002, Montreal, Quebec, Canada. Recall that the prime CNF of a monotone Boolean function f is the unique formula # c#S c in conjunctive normal form where S is the set of all prime implicates of f , i.e., minimal clauses c which are logical consequences of f . In this paper, we consider the following problem: Problem Dualization Input: The prime CNF # of a monotone Boolean Output: The prime CNF # of its dual It is well known that problem Dualization is equivalent to the Transversal Computation problem, which requests to compute the set of all minimal transversals (i.e., minimal hitting sets) of a given hypergraph H, in other words, the transversal hypergraph Tr(H) of H. Actually, these problems can be viewed as the same problem, if the clauses in a monotone CNF # are identified with the sets of variables they contain. Dualization is a search problem; the associated decision problem Dual is to decide whether two given monotone prime CNFs # and # represent a pair (f, g) of dual Boolean functions. Analogously, the decision problem Trans-Hyp associated with Transversal Computation is deciding, given hypergraphs H and G, whether Tr(H). Dualization and several problems which are like transversal computation known to be computationally equivalent to Dualization (see [13]) are of interest in various areas such as database theory (e.g., [34, 43]), machine learning and data mining (e.g., [4, 5, 10, 18]), game theory (e.g., [22, 38, 39]), artificial intelligence (e.g., [17, 24, 25, 40]), mathematical programming (e.g., [3]), and distributed systems (e.g., [16, 23]) to mention a few. While the output CNF # can be exponential in the size of #, it is currently not known whether # can be computed in output-polynomial (or polynomial total) time, i.e., in time polynomial in the combined size of # and #. Any such algorithm for Dualization (or Transversal Computation) would significantly advance the state of the art of many problems in the application areas. Similarly, the complexity of Dual and Trans-Hyp is open since more than 20 years now (cf. [2, 13, 26, 27, 29]). Note that Dualization is solvable in polynomial total time on a class C of hypergraphs i# Dual is in PTIME for all pairs Dual is known to be in co-NP and the best currently known upper time-bound is n o(log n) [15]. Determining the complexities of Dualization and Dual, and of equivalent problems such as the transversal problems, is a prominent open problem. This is witnessed by the fact that these problems are cited in a rapidly growing body of literature and have been referenced in various survey papers and complexity theory retrospectives, e.g. [26, 30, 36]. Given the importance of monotone dualization and equivalent problems for many application areas, and given the long standing failure to settle the complexity of these prob- lems, emphasis was put on finding tractable cases of Dual and corresponding polynomial total-time cases of Dualiza- tion. In fact, several relevant tractable classes were found by various authors; see e.g. [6, 7, 8, 10, 12, 13, 31, 32, 35, 37] and references therein. Moreover, classes of formulas were identified on which Dualization is not just polynomial total-time, but where the conjuncts of the dual formula can be enumerated with incremental polynomial delay, i.e., with delay polynomial in the size of the input plus the size of all conjuncts so far computed, or even with polynomial delay, i.e., with delay polynomial in the input size only. Main Goal. The main goal of this paper is to present important new polynomial total time cases of Dualization and, correspondingly, PTIME solvable subclasses of Dual which significantly improve previously considered classes. Towards this aim, we first present a new algorithm Dualize and prove its correctness. Dualize can be regarded as a generalization of a related algorithm proposed by Johnson, Yannakakis, and Papadimitriou [27]. As other dualization algorithms, Dualize reduces the original problem by self- reduction to smaller instances. However, the subdivision into subproblems proceeds according to a particular order which is induced by an arbitrary fixed ordering of the vari- ables. This, in turn, allows us to derive some bounds on intermediate computation steps which imply that Dualize, when applied to a variety of input classes, outputs the conjuncts of with polynomial delay or incremental polynomial delay. In particular, we show positive results for the following input classes: . Degenerate CNFs. We generalize the notion of k-degen- erate graphs [44] to hypergraphs and define k-degener- ate monotone CNFs resp. hypergraphs. We prove that for any constant k, Dualize works with polynomial delay on k-degenerate inputs. Moreover, it works in output-polynomial time on O(log n)-degenerate CNFs. . Read-k CNFs. A CNF is read-k, if each variable appears at most k times in it. We show that for read-k CNFs, problem Dualization is solvable with polynomial delay, if k is constant, and in total polynomial time, if O(log(#). Our result for constant k significantly improves upon the previous best known algorithm [10], which has a higher complexity bound, is not polynomial delay, and outputs the clauses of in no specific order. The result for O(log #) is a non-trivial generalization of the result in [10], which was posed as an open problem [9]. . Acyclic CNFs. There are several notions of hyper-graph resp. monotone CNF acyclicity [14], where the most general and well-known is #-acyclicity. As shown in [13], Dualization is polynomial total time for #- acyclic CNFs; #-acyclicity is the hereditary version of #-acyclicity and far less general. A similar result for #- acyclic prime CNFs was left open. We give a positive answer and show that for #-acyclic prime #, Dualization is solvable with polynomial delay. . Formulas of Bounded Treewidth. The treewidth [41] of a graph expresses its degree of cyclicity. Treewidth is an extremely general notion, and bounded treewidth generalizes almost all other notions of near-acyclicity. Following [11], we define the treewidth of a hyper-graph resp. monotone CNF # as the treewidth of its associated (bipartite) variable-clause incidence graph. We show that Dualization is solvable with polynomial delay (exponential in k) if the treewidth of # is bounded by a constant k, and in polynomial total time if the treewidth is O(log log #). . Recursive Applications of Dualize and k-CNFs. We show that if Dualize is applied recursively and the recursion depth is bounded by a constant, then Dualization is solved in polynomial total time. We apply this to provide a simpler proof of the known result [6, 13] that monotone k-CNFs (where each conjunct contains at most k variables) can be dualized in output-polynomial time. After deriving the above results, we turn our attention (in Section 5) to the fundamental computational nature of problems Dual and Trans-Hyp in terms of complexity theory. Complexity: Limited nondeterminism. In a landmark paper, Fredman and Khachiyan [15] proved that problem Dual can be solved in quasi-polynomial time. More pre- cisely, they first gave an algorithm A solving the problem in n) time, and then a more complicated algorithm B whose runtime is bounded by n 4#(n) where #(n) is defined by #(n) As noted in [15], #(n) # log n/ log log o(log n); therefore, duality checking is feasible in n o(log n) time. This is the best upper bound for problem Dual so far, and shows that the problem is most likely not NP-complete. A natural question is whether Dual lies in some lower complexity class based on other resources than just runtime. In the present paper, we advance the complexity status of this problem by showing that its complement is feasible with limited nondeterminism, i.e, by a nondeterministic polynomial-time algorithm that makes only a poly-logarithmic number of guesses. For a survey on complexity classes with limited nondeterminism, and for several references, see [19]. We first show by a simple and self-contained proof that testing non-duality is feasible in polynomial time with O(log 3 n) nondeterministic steps. We then observe that this can be improved to O(log 2 n) nondeterministic steps. This result is surprising, because most researchers dealing with the complexity of Dual and Trans-Hyp believed so far that these problems are completely unrelated to limited nondeterminism We believe that the results presented in this paper are signif- icant, and we are confident they will prove useful in various contexts. First, we hope that the various polynomial/output- polynomial cases of the problems which we identify will lead to better and more general methods in various application areas (as we show, e.g. in learning and data mining [10]), and that based on the algorithm Dualize or some future modifi- cations, further relevant tractable classes will be identified. Second, we hope that our discovery on limited nondeterminism provides a new momentum to complexity research on Dual and Trans-Hyp, and will push it towards settling these longstanding open problems. 2. PRELIMINARIES AND NOTATION A Boolean function (in short, function) is a mapping f : {0, 1} n # {0, 1}, where v # {0, 1} n is called a Boolean vector (in short, vector). As usual, we write g # f if f and g satisfy A function f is monotone (or positive), if for all i) implies f(v) # f(w) for all v, w # {0, 1} n . Boolean variables x1 , x2 , . , xn and their complements - x1 , - x2 , . , - xn are called literals. A clause (resp., term) is a disjunction (resp., conjunction) of literals containing at most one of x i and - x i for each variable. A clause c (resp., term t) is an implicate (resp., implicant) of a function f , if f # c (resp., moreover, it is prime, if there is no implicate c # < c no implicant t # > t) of f , and monotone, if it consists of positive literals only. We denote by P I(f) the set of all prime implicants of f . A conjunctive normal form (CNF) (resp., disjunctive normal form, DNF) is a conjunction of clauses (resp., disjunction of terms); it is prime (resp. monotone), if all its members are prime (resp. monotone). For any CNF (resp., DNF) #, we denote by |#| the number of clauses (resp., terms) in it. Furthermore, for any formula #, we denote by V (#) the set of variables that occur in #, and by # its length, i.e., the number of literals in it. As well-known, a function f is monotone i# it has a monotone CNF. Furthermore, all prime implicants and prime implicates of a monotone f are monotone, and it has a unique prime CNF, given by the conjunction of all its prime impli- cates. For example, the monotone f such that (0111), (1111)} has the unique prime Recall that the dual of a function f , denoted f d , is defined by f d x is the complement of f and respectively. By definition, we have (f d From De Morgan's law, we obtain a formula for f d from any one of f by exchanging # and # as well as the constants 0 and 1. For example, if f is given by represented by For a monotone be the prime CNF of f d . Then by De Morgan's law, f has the (unique) prime DNF Thus, we will regard Dualization also as the problem of computing the prime DNF of f from the prime CNF of f . 3. ORDERED GENERATION OF TRANSVERSAL In what follows, let f be a monotone function and # its prime CNF, where we assume w.l.o.g. that all variables x j n) appear in #. Let # i n) be the CNF obtained from # by fixing variables x with 1. By definition, we have Example 3.1. Consider Similarly, for the prime DNF of f , we denote by # i the DNF obtained from # by fixing variables x Clearly, we have denoted by f i . Proposition 3.1. Let # and # be any CNF and DNF for f , respectively. Then, (a) # i # and |# i | #|, and Denote by # i n) the CNF consisting of all the clauses in # i but not in # i-1 . Example 3.2. For the above example, we have Note that # have denote the CNF consisting of all the clauses c such that c contains no literal in t i-1 and appears in # i . For example, if It follows from (2) that for all Lemma 3.2. For any term t #PI (f i-1 ), let g i,t be the function represented by # i [t]. Then |PI (g i,t )|# i | #|. Proof. Let Then by (3), t # s is an implicant of # i . Hence, some t s exists such that t s # t#s. Note that V (t)#V #, and hence we have V otherwise there exists a clause c in # i [t] such that V (c) # For any s # PI (g i,t ) such that s #= s # , let t s , t s # such that t s # t#s and t s # respectively. By the above discussion, we have t s This completes the proof. We now describe our algorithm Dualize for generating the set PI (f ). It is inspired by a similar graph algorithm of Johnson, Yannakakis, and Papadimitriou [27], and can be regarded as a generalization. Here, we say that term s is smaller than term t if i.e., as vector, s is lexicographically smaller than t. Algorithm Dualize Input: The prime CNF # of a monotone function f . Output: The prime DNF of f , i.e. all prime implicants of function f . Compute the smallest prime implicant t min of f and set Q := { t min }; Step 2: while Q # do begin Remove the smallest t from Q and output t; for each i with x do begin Compute the prime DNF # (t,i) of the function represented by # i [t]; for each term t # in # (t,i) do begin if t i-1 # t # is a prime implicant of f i then begin Compute the smallest prime implicant t # of f such that t # Theorem 3.3. Algorithm Dualize correctly outputs all increasing order. Proof. (Sketch) First note that the term t # inserted in when t is output is larger than t. Indeed, t #= 1) and t i-1 are disjoint and V (t # {x1 ,. , x i-1}. Hence, every term in Q is larger than all terms already output, and the output sequence is increasing. We show by induction that, if t is the smallest prime implicant of f that was not output yet, then t is already in Q. This clearly proves the result. Clearly, the above statement is true if . Assume now that t #= t min is the smallest among the prime implicants not output yet. Let i be the largest index such that t i is not a prime implicant of f i . This i is well-defined, since otherwise must hold, a contradiction. Now we have (1) i < n and (2) is a prime implicant of fn (= f) and (2) follows from the maximality of i. Let s # PI (f i ) such that V let (s). Then K # holds, and since x i+1 / V (t), the term t # x j #K x j is a prime implicant of # i+1 [s]. There exists s # PI (f) such that s # since s#x i+1 # PI (f i+1 ). Note that # i+1 [s] #= 0. Moreover, since s # is smaller than t, by induction s # has already been output. Therefore, t been considered in the inner for-loop of the algorithm. Since s # is a prime implicant of f i+1 , the algorithm has added the smallest prime implicant t # of f such that t # We finally claim that t t. Otherwise, let k be the first index in which t # and t di#er. Then k contradicting the maximality of i. Let us consider the time complexity of algorithm Dualize. We store Q as a binary tree, where each leaf represents a term t and the left (resp., right) son of a node at depth the root has depth 0, encodes x In Step 1, we can compute t min in O(#) time and initialize Q in O(n) time. As for Step 2, let T (t,i) be the time required to compute the prime DNF # (t,i) from # i [t]. By analyzing its substeps, we can see that each iteration of Step 2 requires # x i #V (t) (T (t,i) time; note that t # is the smallest prime implicant of the function obtained from f by fixing x and 0 if x Thus, we have Theorem 3.4. The output delay of Algorithm Dualize is bounded by t#PI (f) time, and Dualize needs in total time t#PI (f) If the T (t,i) are bounded by a polynomial in the input length, then Dualize becomes a polynomial delay algorithm, since holds for all t # PI (f) and x the other hand, if they are bounded by a polynomial in the combined input and output length, then Dualize is a polynomial total time algorithm, where |# (t,i) | #| holds from Lemma 3.2. Using results from [2], we can construct from Dualize an incremental polynomial time algorithm for Dualization, which however might not output PI (f) in increasing order. Summarizing, we have the following corollary. Corollary 3.5. Let is bounded by (i) a polynomial in n and #, then algorithm Dualize is an O(n#T ) polynomial delay algorithm; (ii) a polynomial in n, #, and #, then algorithm Dualize is an O(n| |(T + | |#)) polynomial total time algorithm; moreover, Dualization is solvable in incremental polynomial time. In the next section, we identify su#cient conditions for the boundedness of T and fruitfully apply them to solve open problems and improve previous results. 4. POLYNOMIAL CLASSES 4.1 Degenerate CNFs We first consider the case of small # i [t]. Generalizing a notion for graphs (i.e., monotone 2-CNFs) [44], we call a monotone CNF # k-degenerate, if there exists a variable ordering x1 , . , xn in which |# i | # k for all We call a variable ordering x1 , . , xn smallest last as in [44], if x i is chosen in the order | is smallest for all variables that were not chosen. Clearly, a smallest last ordering gives the least k such that # is k- degenerate. Therefore, we can check for every integer k # 1 whether # is k-degenerate in O(#) time. If this holds, then we have |# (t,i) | # n k and T apply the distributive law to # i [t] and remove terms t where some x j # V (t) has no }). Thus Theorem 3.4 implies the following. Theorem 4.1. For k-degenerate CNFs #, Dualization is solvable with O(#n k+1 ) polynomial delay if k # 1 is constant. Applying the result of [33] that any monotone CNF which has O(log n) many clauses is dualizable in incremental polynomial time, we obtain a polynomiality result also for non-constant degeneracy: Theorem 4.2. For O(log #)-degenerate CNFs #, problem Dualization is polynomial total time. In the following, we discuss several natural subclasses of degenerate CNFs. 4.1.1 Read-bounded CNFs A monotone CNF # is called read-k, if each variable appears in # at most k times. Clearly, read-k CNFs are k-degenerate, and in fact # is read-k i# it is k-degenerate under every variable ordering. By applying Theorems 4.1 and 4.2, we obtain the following result. Corollary 4.3. For read-k CNFs #, problem Dualization is solvable (i) with O(#n k+1 ) polynomial delay, if k is constant; (ii) in polynomial total time, if Note that Corollary 4.3 (i) trivially implies that Dualization is solvable in O(|#|n k+2 ) time for constant k, since # kn. This improves upon the previous best known algorithm [10], which is only O(|#|n k+3 ) time, not polynomial delay, and outputs PI (f) in no specific order. Corollary 4.3 (ii) is a non-trivial generalization of the result in [10], which was posed as an open problem [9]. 4.1.2 Acyclic CNFs Like in graphs, acyclicity is appealing in hypergraphs resp. monotone CNFs from a theoretical as well as a practical point of view. However, there are many notions of acyclicity for hypergraphs (cf. [14]), since di#erent generalizations from graphs are possible. We refer to #-, and Berge- acyclicity as stated in [14], for which the following proper inclusion hierarchy is known: Berge-acyclic #-acyclic #-acyclic #-acyclic. The notion of #-acyclicity came up in relational database theory. A monotone CNF # is #-acyclic reducible by the GYO-reduction [21, 45], i.e., repeated application of one of the two rules: (1) If variable x i occurs in only one clause c, remove x i from clause c. (2) If distinct clauses c and c # satisfy clause c from #. to 0 (i.e., the empty clause). Note that #-acyclicity of a monotone CNF # can be checked, and a suitable GYO- reduction output, in O(#) time [42]. A monotone CNF # is #-acyclic i# every CNF consisting of clauses in # is #- acyclic. As shown in [13], the prime implicants of a monotone f represented by a #-acyclic CNF # can be enumerated (and thus Dualization solved) in p(#| time, where p is a polynomial in #. However, the time complexity of Dualization for the more general #-acyclic prime CNFs was left as an open problem. We now show that it is solvable with polynomial delay. Let #= 1 be a prime CNF. Let a = a1 , a2 , . , aq be a GYO- reduction for #, where a the #-th operation removes x i from c, and a removes c from #. Consider the unique variable ordering b1 , b2 , . , bn such b i occurs after b j in a, for all i < j. Example 4.1. Let x3 since it has the GYO-reduction . From this sequence, we obtain the variable ordering As easily checked, this ordering shows that # is 1-degenerate. Under this ordering, we have # That # is 1-degenerate in this example is not accidental. Lemma 4.4. Every #-acyclic prime CNF is 1-degenerate. Note that the converse is not true. Lemma 4.4 and Theorem 4.1 imply the following result. Corollary 4.5. For #-acyclic CNFs #, problem Dualization is solvable with O(#n 2 ) delay. Observe that for a prime #-acyclic #, we have |# n. Thus, if we slightly modify algorithm Dualize to check # in advance (which can be done in linear time in a preprocessing phase) such that such # i need not be considered in step 2, then the resulting algorithm has O(n|#) delay. Observe that the algorithm in [13] solves, minorly adapted for enumerative output, Dualization for #-acyclic CNFs with O(n|#) delay. Thus, the above modification of Dualize is of the same order. 4.1.3 CNFs with bounded treewidth A tree decomposition (of type I) of a monotone CNF # is a tree T =(W,E) where each node w#W is labeled with a set X(w)#V (#) under the following conditions: 1. # w#W 2. for every clause c in #, there exists some w # W such that V (c) # X(w); and 3. for any variable x X(w)} induce a (connected) subtree of T . The width of T is maxw#W |X(w)| - 1, and the treewidth of #, denoted by Tw 1 (#), is the minimum width over all its tree decompositions. Note that the usual definition of treewidth for a graph [41] results in the case where # is a 2-CNF. Similarly to acyclic- ity, there are several notions of treewidth for hypergraphs resp. monotone CNFs. For example, tree decomposition of type II of CNF # c#C c is defined as type-I tree decomposition of its incident 2-CNF (i.e., graph) G(#) [11, 20]. That is, for each clause c #, we introduce a new variable yc and construct denote the type-II treewidth of #. Proposition 4.6. For every monotone CNF #, it holds that Tw 2 Proof. Let be any tree decomposition of # having width Tw 1 (#). Introduce for all c # new variables y c , and add y c to every X(w) such that Clearly, the result is a type-I tree decomposition of G(#), and thus a type-II tree decomposition of #. Since at most 2 |X(w)| many yc are added to X(w) and for every w # W , the result follows. This means that if Tw 1 (#) is bounded by some constant, then so is Tw 2 (#). Moreover, Tw 1 implies that # is a k-CNF; we discuss k-CNFs in Section 4.2 and only consider Tw 2 (#) here. We note that, as shown in the full paper, there is a family of prime CNFs # which have Tw 2 (#) bounded by constant k but are not k-CNF for any k < n not read-k for any k < n - 1), and a family of prime CNFs which are k-CNFs for constant k (resp., #-acyclic) but Tw 2 (#) is not bounded by any constant. As we show now, bounded-treewidth implies bounded degeneracy Lemma 4.7. Let # be any monotone CNF with Tw 2 k. Then # is 2 k -degenerate. Proof. (Sketch) Let E) with show From this, we reversely construct a variable ordering , an on (#) such that |# i for all i. Choose any leaf w # of T , and let p(w # ) be a node in W adjacent to w # . If X(w # ) \ X(p(w # {yc | c #}, then remove w # from T . On the other hand, if only for We complete a by repeating this process, and claim it shows that |# i for all i. Let w # be chosen during this process, and assume that a i # X(w # ) \ X(p(w # )). Then, for each clause c # i we must have either yc # X(w # ) or V (c) # X(w # ). Let | # Corollary 4.8. For CNFs # with Tw 2 (# k, Dualization is solvable (i) with O(#n 2 k +1 ) polynomial delay, if k is constant; (ii) in polynomial total time, if 4.2 Recursive application of algorithm Dual- ize Algorithm Dualize computes in step 2 the prime DNF # (t,i) of the function represented by # i [t]. Since #[t] is the prime CNF of some monotone function, we can recursively apply Dualize to # i [t] for computing # (t,i) . Let us call this variant R-Dualize. Then we have the following result. Theorem 4.9. If its recursion depth is d, R-Dualize solves Dualization in O(n d-1 |#| d-1 #) time. Proof. If min and every # 1. This means that PI (f)={tmin} and # is a 1-CNF (i.e., each clause in # contains exactly one variable). Thus in this case, R-Dualize needs O(n) time. Recall that algorithm Dualize needs, by (5), time |# (t,i) |O(#)). If Therefore, R-Dualize needs time O(n|#). For d # 3, Corollary 3.5.(ii) implies that algorithm R-Dualize needs time O(n d-1 |#| d-1 #). Recall that a CNF # is called k-CNF if each clause in # has at most k literals. Clearly, if we apply algorithm R-Dualize to a monotone k-CNF #, the recursion depth of R-Dualize is at most k. Thus we obtain the following result; it re- establishes, with di#erent means, the main positive result of [6, 13]. Corollary 4.10. Algorithm R-Dualize solves Dualization in time O(n k-1 |#| k-1 #), i.e., in polynomial total time for monotone k-CNFs # where k is constant. 5. LIMITED NONDETERMINISM In the previous section, we have discussed polynomial cases of monotone dualization. In this section, we now turn to the issue of the precise complexity of this problem. For this purpose, we consider the decision problem Dual instead of the search problem Dualization. It appears that problem Dual can be solved with limited nondeterminism, i.e., with poly-log many guessed bits by a polynomial-time non-deterministic Turing machine. This result might bring new insight towards settling the complexity of the problem. We adopt Kintala and Fischer's terminology [28] and write g(n)-P for the class of sets accepted by a nondeterministic Turing machine in polynomial time making at most g(n) nondeterministic steps on every input of length n. For every n)-P. The #P Hierarchy consists of the classes and lies between P and NP. The #kP classes appear to be rather robust; they are closed under polynomial time and logspace many-one reductions and have complete problems (cf. [19]). The complement class of #kP is denoted by co-#kP. We start by recalling algorithm A of [15], reformulated for CNFs. In what follows, we view CNFs # also as sets of clauses, and clauses as sets of literals. Algorithm A. (reformulated for CNFs) Input: Monotone CNFs #, representing monotone f , g s.t. V (c)#V (c #, for all c#, c # . Output: yes if vector w of form Delete all redundant (i.e., non-minimal) implicates from # and #. Step 2: Check that V If any of these conditions fails, f #= g d and a witness w is found in polynomial time (cf. [15]). Step 3: If |# 1, test duality in O(1) time. Step 4: If |# 2, find a variable x i that occurs in # or # (w.l.o.g. in #) with frequency # 1/ log(|#|). Let Call algorithm A on the two pairs of forms: If both calls return yes, then return yes (as otherwise we obtain w such that f(w) #= g d (w) in polynomial time (cf. [15]). be the original input for A. For any pair (#) of CNFs, define its volume by |. As shown in [15], step 4 of algorithm A divides the current (sub)problem of volume self-reduction into subproblems (A.1) and (A.2) of respective volumes (assuming that x i frequently occurs in #): (#) be the recursion tree generated by A on input (#), i.e., the root is labeled with (#). Any node a labeled with (#) is a leaf, if A stops on input (#) during steps 1-3; otherwise, a has a left child a l and a right child ar corresponding to (A.1) and (A.2), i.e., labeled (#1 , #0 # #1 ) and (#1 , #0 #1 ) respectively. That is, a l is the "high frequency move" by the splitting variable. We observe that every node a in T is determined by a unique path from the root to a in T and thus by a unique sequence seq(a) of right or left moves starting from the root of T and ending at a. The following key lemma bounds the number of moves of each type for certain inputs. Lemma 5.1. Suppose |# i |. Then for any node a in T , seq(a) contains # v right and # log 2 v left moves, where |. Proof. By (6) and (7), each move decreases the volume v of a node label. Thus, the length of seq(a), and in particular the number of right moves, is bounded by v. To obtain the better bound for the left moves, we will use the following well-known inequality: # 1/e, for m # 1. (8) In fact, the sequence (1 -1/x i monotonically converges to 1/e from below. By inequality (6), the volume va of the label of any node a such that seq(a) contains log 2 v left moves is bounded as follows: log n) log 2 v . Because | - |# i and because of (8) it follows that: log v Thus, a must be a leaf in T . Hence for every a in T , seq(a) contains at most log 2 v left moves. Theorem 5.2. Problem Dual is in co-#3P. Proof. (Sketch) Instances such that either c # c # for some c # i and c # i , the sequence seq(a) is empty, or |# i | > |# i | are easily solved in deterministic polynomial time. In the remaining cases, if f #= g d , then there exists a leaf a in T labeled by a non-dual pair (# ). If seq(a) is known, we can compute, by simulating A on the branch described by seq(a), the entire path from the root to a with all labels check that non-dual in steps 2 and 3 of A in polynomial time. We observe that, as noted in [15], the binary length of any standard encoding of the input # i , # i is polynomially related to |# i | if algorithm A reaches step 3. Thus, to prove the theorem, it is su#cient to show that seq(a) is obtainable in polynomial time from O(log 3 v) suitably guessed bits, where |. To see this, let us represent every seq(a) as a sequence seq # #0 is the number of leading right moves and # i is the number of consecutive right moves after the i-th left move in seq(a), for then seq # (a) = [2, 3, 0]. By Lemma 5.1, seq # (a) has length at most log 2 v + 1. Thus, seq # (a) occupies only O(log 3 v) bits in binary; moreover, seq(a) is trivially computed from seq # (a) in polynomial time. Remark 5.1. It also follows that if f #= g d , a witness w can be found in polynomial time within O(log 3 n) nondeterministic steps. In fact, the sequence seq(a) to a "failing labeled (# ) describes a choice of values for all variables in V (#) \ V (# ). By completing it with values for show non-duality of (# ), we obtain in polynomial time a vector w such that f(w) #= g d (w). The aim of the above proof was to show with very simple means that duality can be polynomially checked with limited nondeterminism. With a more involved proof, applied to the algorithm B of [15] (which runs in n 4#(n)+O(1) and thus time), we can prove the following sharper result. Theorem 5.3. Deciding if monotone CNFs # and # are non-dual is feasible in polynomial time with O(#(n) log n) nondeterministic steps. Thus, problem Dual is in co-#2P. While our independently developed methods are di#erent from those in [1], the previous result may also be obtained from Beigel and Fu's Theorem 11 in [1]. They show how to convert certain recursive algorithms that use disjunctive self-reductions and have runtime bounded by f(n) into polynomial algorithms using log f(n) nondeterministic steps (cf. [1, Chapter 5]). However, this yields a somewhat more complicated nondeterministic algorithm. In the full paper, we also prove that algorithm B qualifies for this. 6. ACKNOWLEDGMENTS This work was supported in part by the Austrian Science Fund project Z29-INF, by TU Wien through a scientific collaboration grant, and by the Scientific Grant in Aid of the Ministry of Education, Science, Sports and Culture of Japan. We would like to thank the reviewers for their constructive comments on this paper. 7. --R Molecular computing Complexity of identification and dualization of positive Boolean functions. On generating all minimal integer solutions for a monotone system of linear inequalities. On the complexity of generating maximal frequent and minimal infrequent sets. Dual subimplicants of positive Boolean functions. time recognition of 2-monotonic positive Boolean functions given by an oracle Dualization of regular Boolean functions. Private communication. Conjunctive query containment revisited. Exact transversal hypergraphs and application to Boolean Identifying the minimal transversals of a hypergraph and related problems. Degrees of acyclicity for hypergraphs and relational database schemes. On the complexity of dualization of monotone disjunctive normal forms. How to assign votes in a distributed system. Incremental recompilation of knowledge. Data mining Limited nondeterminism. Hypertree decompositions and tractable queries. On the universal relation. A theory of coteries: Mutual exclusion in distributed systems. Translating between Horn representations and their characteristic models. "G. Stampacchia" On generating all maximal independent sets. Refining nondeterminism in relativized polynomial-time bounded computations Generating all maximal independent sets: NP-hardness and polynomial-time algorithms Combinatorial optimization: Some problems and trends. The maximum latency and identification of positive Boolean functions. A fast and simple algorithm for identifying 2-monotonic positive Boolean functions Generating all maximal independent sets of bounded-degree hypergraphs A retrospective An O(nm)-time algorithm for computing the dual of a regular Boolean function Coherent Structures and Simple Games. Every one a winner A theory of diagnosis from first principles. Graph minors II: Algorithmic aspects of tree-width Simple linear time algorithms to test chordality of graphs Minimal keys and antikeys. Colouring, stable sets and perfect graphs. An algorithm for tree-query membership of a distributed query --TR Simple linear-time algorithms to test chordality of graphs, test acyclicity of hypergraphs, and selectively reduce acyclic hypergraphs How to assign votes in a distributed system Design by exmple: An application of Armstrong relations The minimal keys and antikeys A theory of diagnosis from first principles Dualization of regular Boolean functions On generating all maximal independent sets An O(<italic>nm</italic>)-time algorithm for computing the dual of a regular Boolean function Exact transversal hypergraphs and application to Boolean MYAMPERSANDmgr;-functions Identifying the Minimal Transversals of a Hypergraph and Related Problems Complexity of identification and dualization of positive Boolean functions Colouring, stable sets and perfect graphs Limited nondeterminism On the complexity of dualization of monotone disjunctive normal forms Polynomial-Time Recognition of 2-Monotonic Positive Boolean Functions Given by an Oracle Data mining, hypergraph transversals, and machine learning (extended abstract) The Maximum Latency and Identification of Positive Boolean Functions Generating all maximal independent sets of bounded-degree hypergraphs A fast and simple algorithm for identifying 2-monotonic positive Boolean functions Hypertree decompositions and tractable queries Degrees of acyclicity for hypergraphs and relational database schemes Efficient Read-Restricted Monotone CNF/DNF Dualization by Learning with Membership Queries Dual-Bounded Generating Problems A Theory of Coteries Conjunctive Query Containment Revisited On Generating All Minimal Integer Solutions for a Monotone System of Linear Inequalities NP-Completeness On Horn Envelopes and Hypergraph Transversals On the Complexity of Generating Maximal Frequent and Minimal Infrequent Sets --CTR Dimitris J. Kavvadias , Elias C. Stavropoulos, Monotone boolean dualization is in co-NP[log2n], Information Processing Letters, v.85 n.1, p.1-6, January Thomas Eiter , Kazuhisa Makino, On computing all abductive explanations, Eighteenth national conference on Artificial intelligence, p.62-67, July 28-August 01, 2002, Edmonton, Alberta, Canada Georg Gottlob , Reinhard Pichler , Fang Wei, Tractable database design through bounded treewidth, Proceedings of the twenty-fifth ACM SIGMOD-SIGACT-SIGART symposium on Principles of database systems, June 26-28, 2006, Chicago, IL, USA Leonid Khachiyan , Endre Boros , Khaled Elbassioni , Vladimir Gurvich, A global parallel algorithm for the hypergraph transversal problem, Information Processing Letters, v.101 n.4, p.148-155, February, 2007 Peter L. Hammer , Alexander Kogan , Bruno Simeone , Sndor Szedmk, Pareto-optimal patterns in logical analysis of data, Discrete Applied Mathematics, v.144 n.1-2, p.79-102, November 2004

hypergraph acyclicity;dualization;treewidth;limited nondeterminism;output-polynomial algorithms;transversal computation;combinatorial enumeration

510729

Multicast Video-on-Demand services.

The server's storage I/O and network I/O bandwidths are the main bottleneck of VoD service. Multicast offers an efficient means of distributing a video program to multiple clients, thus greatly improving the VoD performance. However, there are many problems to overcome before development of multicast VoD systems. This paper critically evaluates and discusses the recent progress in developing multicast VoD systems. We first present the concept and architecture of multicast VoD, and then introduce the techniques used in multicast VoD systems. We also analyze and evaluate problems related to multicast VoD service. Finally, we present open issues on multicast VoD as possible future research directions.

INTRODUCTION A typical Video-on-Demand (VoD) service allows remote users to play back any one of a large collection of videos at any time. Typically, these video les are stored in a set of central video servers, and distributed through high-speed communication networks to geographically-dispersed clients. Upon receiving a client's service request, a server delivers the video to the client as an isochronous video stream. Each video stream can be viewed as a concatenation of a storage-I/O \pipe" and a network pipe. Thus, su-cient storage-I/O bandwidth must be available for continuous transfer of data from the storage system to the network interface card (NIC), which must, in turn, have enough bandwidth to forward data to clients. Thus, a video server has to reserve su-cient I/O and network bandwidths before accepting a This work was partly done during Huadong Ma's visit to RTCL at the University of Michigan. The work was supported in part by the USA NSF under Grant EIA-9806280 and the Natural Science Foundation of China under Grant 69873006. client's request. We dene a server channel as the server resource required to deliver a video stream while guaranteeing a client's continuous playback. This type of VoD service has a wide spectrum of appli- cations, such as home entertainment, digital video library, movie-on-demand, distance learning, tele-shopping, news- on-demand, and medical information service. In general, the VoD service can be characterized as follows. Long-lived session: a VoD system should support long-lived sessions; for example, a typical movie-on-demand service usually lasts 90{120 minutes. High bandwidth requirements: for example, server storage I/O and network bandwidth requirements are 1.5 Mbps (3-10 Mbps) for a MPEG-1 (MPEG-2) stream. Support for VCR-like interactivity: a client requires the VoD system to oer VCR-like interactivity, such as the ability to play, forward, reverse and pause. Other advanced interactive features include the ability to skip or select advertisements, investigate additional details behind a news event (by hypermedia link), save the program for a later reference, and browse, select and purchase goods. QoS-sensitive service: the QoS that VoD consumers and service providers might care includes service latency, defection rate, interactivity, playback eects of videos, etc. A conventional TVoD system uses one dedicated channel for each service request, oering the client the best TVoD service. However, such a system incurs very high costs, especially in terms of storage-I/O and network bandwidths. Moreover, such a VoD service has poor scalability and low performance/cost e-ciency. Although the conventional approach simplies the implementation, not sharing channels for client requests will quickly exhaust the network and the server I/O bandwidth. In fact, the network-I/O bottleneck has been observed in many earlier systems, such as Time Warner Cable's Full Service Network Project in Orlando [68], and Microsoft's Tiger Video Fileserver [12]. In order to support a large population of clients, we therefore need new solutions that e-ciently utilize the server and network resources. Clearly, the popularity or access pattern of video objects plays an important role in determining the eectiveness of a video delivery technique. Because dierent videos are requested at dierent rates and at dierent times, videos are usually divided into hot (popular) and cold (less popular), and requests for the top 10{20 videos are known to constitute 60{80% of the total demand. So, it is crucial to improve the service e-ciency of hot videos. Thus, requests by multiple clients for the same video arriving within a short time interval can be batched together and serviced using a single stream. This is referred to as batching . The multicast facility of modern communication networks [24, 25, 59] oers an e-cient means of one-to-many 1 data transmission. The basic idea is to avoid transmitting the same packet more than once on each link of the net-work by having branch routers duplicate and then send the packet over multiple downstream branches. Multicast can signicantly improve the VoD performance, because it reduces the required network bandwidth greatly, thereby decreasing the overall network load; alleviates the workload of the VoD server and improves the system throughput by batching requests; oers excellent scalability which, in turn, enables servicing a large number of clients; and provides excellent cost/performance benets. In spite of these advantages, multicast VoD (MVoD) introduces new and di-cult challenges, as listed below, that may make the system more complex, and may even degrade a particular customer's QoS. It is di-cult to support VCR-like interactivity with multicast VoD service while improving service e-ciency. Batching makes the clients arriving at dierent times share a multicast stream, which may incur a long service latency (or waiting time) causing some clients to renege. A single VoD stream from one server cannot support clients' heterogeneity due mainly to diverse customer premise equipments (CPEs). A multicast session makes it di-cult to manage the system protocol and diverse clients. Multicast VoD introduces the complex legal issue of copyright protection. Amulticast VoD system must therefore overcome the above drawbacks without losing its advantages. This paper critically reviews the recent progress in multicast VoD (including general VoD techniques) and discusses open issues in multi-cast VoD. The remainder of this paper is organized as follows. Section 2 introduces the concepts and architectures of a multi-cast VoD system, and analyzes the problems in developing it. Section 3 reviews the implementations of multicast VoD. Section 4 discusses the issues related to multicast VoD ser- vice. Finally, Section 5 summarizes the paper and discusses open issues in implementing multicast VoD. 2. OVERVIEW OF MVOD SERVICE Multicast also covers multipoint-to-multipoint communica- tion, but for our purpose in this paper, it su-ces to consider only one-to-many communication. Classication Features similar to broadcast TV, in which the No-VoD user is a passive participant and has no control over the session. in which the user signs up and pays PPV for specic programming, similar to existing CATV PPV services. in which users are grouped based on a QVoD threshold of interest. Users can perform rudimentary temporal control activities by switching to a dierent group. functions like forward and reverse are NVoD simulated by transitions in discrete time intervals. This capability can be provided by multiple channels with the same programming skewed in time. the user has complete control over the TVoD session presentation. The user has full-function VCR capabilities, including forward and reverse play, freeze, and random positioning. Table 1: Classication of VoD system 2.1 The taxonomy of VoD systems A true VoD system supports a user to view any video, at any time and in any interactive mode. Based on the amount of interactivity and the ability of controlling videos, VoD systems are classied as Broadcast (No-VoD), Pay-Per-View (PPV), Quasi Video-On-Demand (QVoD), Near Video-On- Demand (NVoD), True Video-On-Demand (TVoD) [55] which are listed and compared in Table 1. Obviously, TVoD is the most ideal service. For TVoD service, the simplest scheme of scheduling server channels is to dedicate a channel to each client, but it will require too many channels to be aordable. Since a client may be willing to pay more for TVoD service than for non-TVoD service, sharing a channel among clients is a reasonable way to improve the VoD performance and lower clients' cost. In fact, multicast can support all types of VoD services while consuming much less resources. 2.2 VCR interactivity of VoD Interactivity is an essential feature of VoD service. After their admission, customers can have the following types of interactions: Play/Resume, Stop/Pause/Abort, Fast For- ward/Rewind, Fast Search/Reverse Search, Slow Motion as identied in [54]. A TVoD service may also provide the support for other interactions such as Reverse and Slow Reverse, which correspond to a presentation in the reverse direction, at normal or slow speed. Usually, we don't consider them as part of the usual interactive behavior of a customer. We classify interactive operations into two types: (1) forward interactions, such as Fast Forward and Fast Search; (2) backward interactions, such as Rewind, Reverse Search, Slow Motion, and Stop/Pause. This classication depends on whether the playback rate after interactive operations is faster than the normal playback or not. In order to understand the limited support provided by default in multicast VoD systems, one can identify two types of interactivity: continuous or discontinuous interaction [7]. Continuous in- Delivery System CPS-SPS Delivery System Principal service Interface Application Service Interface Service Interface Network Service Interface Physical Interface Figure 1: DAVIC reference model teractive functions allow a customer to fully control the duration of all actions to support TVoD service, whereas discontinuous interactive functions allow actions to be specied only for durations that are integer multiples of predetermined time increment to support NVoD service. Note that the size of discontinuity is a measure of the QoS experienced by the customers from NVoD service. From the implementation's perspective, we also categorize interactions as interactions with picture or interaction without picture. Fast/Reverse Search and Slow Motion are typical interactions with picture, whereas Fast Forward and Rewind are typical interactions without picture. In gen- eral, it is easier to implement interactions without picture because it requires less system resource. 2.3 The architecture of multicast VoD systems 2.3.1 The reference model of VoD systems The Digital Audio-Visual Council (DAVIC) founded in 1994 is a non-prot organization which has charged itself with the task of promoting broadband digital services by the timely availability of internationally-agreed specications of open interfaces and protocols that maximize interoperability across countries and applications or services. According to the DAVIC reference model shown in Figure 1 [27], a VoD system generally consists of the following entities: Content Provider System (CPS) that owns and sells video programs to the service provider; Service Provider System (SPS) is a collection of system functions that accept, process and present information for delivery to a service consumer system; Service Consumer System (SCS) is responsible for the primary functions that allow a consumer to interact with the SPS and implemented in a customer premise equipment (CPE); CPS-SPS and SPS-SCS network provider. A consumer generates a request for service to the provider, who will obtain the necessary material from the program (content) provider and deliver it to the consumer using the network provider's facilities. The SPS acts as an agent for consumers and can access the various types of CPS. The network, CPS, and SPS can be the same organization, but they are generally dierent. DAVIC-based VoD systems have been developed, such as the one in [71], ARMIDATM [52], the NIST VoD system [45], KYDONIA [19], and the Broadband Interactive VoD system at Beijing Telecommunications Media Server Storage Serivce Provider Request Response Network Figure 2: A multicast VoD system The reference model is also suitable for specifying the architecture of MVoD systems. Consider a typical MVoD delivery system shown in Figure 2 [8, 35]. Consumers make program requests to the manager server (Service Provider). A request is received and queued by the manager server until the scheduler is ready to allocate a logical channel to deliver video streams from a video object storage to a group of consumers (multicast group) across a high-speed network. The manager server organizes the media server and network resources to deliver a video stream into a channel. A channel can be either a unicast or multicast channel. The media server receives consumer requests for video objects via the manager server, processes them, and determines when and which channels to deliver requested video objects to the consumers Each consumer accesses the system by a CPE which includes a set-top box (STB), a disk and a display monitor. A consumer is connected to the network via a STB, which selects one or more network channels to receive requested video objects according to the server's instructions. The received video objects are either sent to the display monitor for immediate playback, or temporarily stored on the disk which will later be retrieved and played back. 2.3.2 Hierarchical VoD systems Large-scale VoD systems require the servers to be arranged as a distributed system in order to support a large number of concurrent streams. If the system is hierarchical, an end-node server handles the requests from a particular area, the next server in the hierarchy takes the requests over for end-node servers if they cannot handle them. This architecture provides the cost e-ciency, reliability and scalability of servers. Generally, servers are either tree-shaped [60] or graph-structured [76, 77] in Figure 3. The graph-structured system often oer good QoS for handling the requests, but the managements of requests, videos and streams are complicated in the system. The tree-shaped system can easily manage requests, videos and streams, but it oers poorer QoS than the former. In order to evaluate the eectiveness of distribution strategies in such a hierarchy, the authors of [39] investigated how to reduce storage and network costs while taking the customers' behaviors into account. Although some of hierarchical architectures are originally designed for unicast VoD services, they can also be used for multicast VoD to further improve the e-ciency of service. 2.4 Problems with multicast VoD Given below are the desired properties of a multicast VoD Client Client Client Client Server Client Server Server Server Server Server Server Server Client Server Client Figure 3: Hierarchical architecture of a VoD system system. E-ciency: The system should impose a minimal additional burden on the server and the network, and should sufciently utilize critical resources on the server and the network. Real-Time: The system should respond to the consumer requests and transmit the requested videos in real time. Scalability: The system should scale well with the number of clients. Interactivity: The system should provide the clients full control of the requested video by using VCR-like interactive functions. Reliability: The system should be robust to failures in the server and the network, and easy to recover from fail- ures. The transmission of messages and video streams should also be reliable. Security: The system should provide e-cient support for copyright protection in transmitting video streams to multiple clients. Ability to deal with heterogeneity: The system should deal with heterogeneous networks and CPEs. Fairness: The system should provide \fair" scheduling of videos with dierent popularities so as to treat all customers \fairly." In order to meet the above requirements, we must solve the following key problems. The rst problem is how to deal with the coupling between system throughput and the batching interval. Increasing the batching interval can save server and network resources signicantly at the expense of increasing the chance of cus- tomers' reneging: consumers are likely to renege if they are forced to wait too long, whereas shortening their waiting time will diminish the benets of multicast VoD. In order to make this tradeo, we must shorten all requests' waiting time while enabling each multicast session to serve as many consumers as possible. The second problem is how to support scalability and in- teractivity. Support for full interactivity requires an \in- dividualized" service for each customer by dedicating an interaction-(or I-) channel per consumer, which limits the scalability of multicast VoD. We need a fully-interactive on-demand service in multicast VoD systems without compromising system scalability and economic viability. The third problem is how to guarantee customers' QoS with limited bandwidths. In multicast VoD, customers' QoS can be expressed in terms of the waiting time before receiving service (or service latency), the customers' defection rate due to long waits, and the VCR action blocking probability and playback eect. However, since system resources are limited, we must strive to maximize their utilization. Moreover, the multicast VoD service generally favors popular videos, but how to serve the requests for unpopular videos in a multicast VoD framework is also of importance to the fairness of service. 3. IMPLEMENTATION OF MVOD 3.1 Storage organization There are two types of servers: manager and media servers. The manager server (Service Provider) is responsible for billing and connection management, while the media server, the focus of this section, handles real-time retrieval and delivery of video streams. The main challenge in the design of video server is how to utilize storage e-ciently. When designing a cost-eective video storage, one must consider issues, such as placement of data on disks, disk bandwidth and disk-access QoS. We consider the following main storage requirements. The VoD server requires a large storage capacity. A 100-minute MPEG-2 video with a transfer rate of 4 Mbps requires approximately 3 GBytes of storage space. Video objects are di-cult to handle due to their large volume/size and stringent requirement of real-time continuous playback. Most existing studies consider the use of multiple disks organized in the form of disk-farm or disk-array. A video server typically uses disk-array for large video data. When designing such a disk-array based VoD server, we must deal with several constraints on resource allocation to provide scala- bility, versatility, and load-balancing. Scalability is dened as the ability to absorb signicant workload uctuations and overloads without aecting admission latency, while versatility is dened as the ability to recongure the VoD server with a minimal disturbance to service availability. High-level versatility is also desirable for expandability, to ensure that new devices can be added easily. Each video can be stored on a single disk or stripped over multiple disks. There are two basic types of storage organization. The rst type completely partitions the storage among dier- ent movie titles. Such a storage system is said to have a completely-partitioned (CP) organization, and may be found in small-scale VoD servers which store One Movie title Per Disk (OMPD). The second type completely shares the storage among dierent movie titles, which is said to have a completely-shared (CS) organization. VoD servers store movie titles using ne-grained striping (FGS) or coarse-grained striping (CGS) [65] of videos across disks in order to eec- tively utilize disk bandwidth. In FGS (similar to RAID-3), the stripe unit is relatively small and every retrieval involves all n disks that behave like a single logical disk with band-width nB (B is the bandwidth of one disk). In CGS, each retrieval block consists of a large stripe unit which is read from only a single disk, but dierent disks can simultaneously serve independent requests. CGS with parity information maintained on one or several dedicated disks corresponds to RAID-5 [11, 66]. organizations typically trade availability | disks can fail or be brought o-line for update without aecting the entire service | for increased latency and costly, ine-cient use Batching policy Features Comparison Maximum Queue Length requests for the video with the largest maximizing the server throughput (MQLF) [21] number of pending requests to serve rst. but unfairness to unpopular videos. First-Come-First-Served the oldest request (with the longest fairness but a lower system -First waiting time) to serve next. throughput. Maximum Factored Queue the pending batch with the largest size a throughput close to that of MQLF Length First (MFQLF)[5] weighted by the factor, (the associated without compromising fairness. access frequency) 1=2 ,to serve next. Look-Ahead-Maximize a channel is allocated to a queue if and maximizing the number of admitted -Batch (LAMB) [33] only if a head-of-the-line user is about users in a certain time window to leave the system without being served but unfairness to some requests. Group-Guaranteed Server server capacity is pre-assigned to meeting a given performance objective Capacity (GGSC) [81] groups of objects for the specic group. Table 2: Multicast batching policies of storage capacity. CS organizations ensure a very low latency and high storage utilization, but recongurations risk the availability of the entire VoD server. Studies in [3, 18, have shown that video striping improves disk utilization and load-balancing, and hence increases the number of concurrent streams. [3, 63] considered both CGS and FGS, and concluded that the former can support more concurrent video streams than the latter. This is because a disk has a relatively high latency for data access (10{20 ms), and a su-cient amount of video data must be transferred in each disk access in order to improve the utilization of the eective disk transfer bandwidth. 3.2 User-centered scheduling strategies A conventional VoD system assumes the user-centered scheduling scheme [4, 83] in which a user eventually acquires some dedicated bandwidth. It can be achieved by providing (1) a su-cient bandwidth equal to an object consumption rate multiplied by the number of users, or (2) less bandwidth, for which the users compete by negotiating with a sched- uler. The consumption rate of a video object is equal to the amount of bandwidth necessary to view it continuously. When a client makes a request to the server, the server sends the requested object to the client via a dedicated channel. This scheme incurs high system costs, especially in terms of server storage-I/O and network bandwidths. To maximally utilize these channels, researchers have proposed e-cient scheduling techniques [20, 32, 43, 51, 61, 62, 64, 84]. These techniques are said to be \user-centered," because channels are allocated to users, not data or objects. These simplify the implementation, but dedicating a stream to each viewer will quickly exhaust the network-I/O bandwidth. 3.3 Data-centered scheduling strategies To address the network-I/O bottleneck faced by the user-centered scheduling, one can use the data-centered scheduling which dedicates channels to video objects, instead of users. It allows users to share a server stream by batching their requests. That is, requests by multiple clients for the same video arriving within a short time interval can be batched together and served by using a single stream. The data-centered scheme has the potential for dramatically reducing the network and server bandwidth require- ments. The data-centered multicast VoD service can be either client-initiated or server-initiated [35]. In the client-initiated service, channels are allocated among the users and the service is initiated by clients, so it is also known as a scheduled or client-pull service. In the server-initiated ser- vice, the server channels are dedicated to individual video objects, so it is also called a periodic broadcast or server-push service. Popular videos are broadcast periodically in this scheme, and a new request dynamically joins, with a small delay, the stream that is being broadcast. In practice, it is e-cient to use hybrid batching that combines the above two schemes. 3.3.1 Client-initiated multicast schemes Using a client-initiated multicast, when a server channel becomes available, the server selects a batch to multicast according to the scheduling policies in Table 2. The equally-spaced batching mechanism has a xed maximum service latency and supports NVoD interactivity, but its usually-large service latency may cause some clients to renege. In order to reduce the service latency, dynamic multicast has been proposed, where the multicast tree is expanded dynamically to accommodate new requests. For example, Adaptive Piggybacking [38] allows clients arriving at dierent times to share a data stream by altering the playback rates of in-progress requests (for the same object), for the purpose of merging their respective video streams into a single stream that can serve the entire group of merged requests. This approach can lower the service latency as compared to simple batching. But it is restrictive in that the variation of the playback rate must be within, say 5%, of the normal playback rate, or it will result in a perceivable deterioration of QoS. This limits the number of streams that can be merged. Chaining [77] is also a generalized dynamic multicast technique to reduce the demand on the network-I/O bandwidth by caching data in the client's local storage to facilitate future multicasts. Thus, data are actually pipelined through the client stations residing at the nodes of the respective chaining tree, and the server serves a \chain" of client stations using only a single data stream. The advantage of chaining is that not every request has to receive its data directly from the sever. A large amount of video also becomes available from clients located throughout the network. This scheme scales well because each client station using the service also contributes its resources to the community. Hence, the larger the chaining trees, the more eective the application can utilize the aggregate bandwidth. The authors of [16] present stream tapping that allows a client to greedily \tap" data from any stream on the VoD server containing video data s/he can use. This is accomplished through the use of a small buer on the CPE and requires less than 20% of the disk bandwidth used by conventional systems for popular videos. To eliminate the service latency, patching was introduced in [41]. The objective of patching is to substantially improve the number of requests each channel can serve per time unit, thereby su-ciently reducing the per-customer system cost. In the patching scheme, channels are often used to patch the missing portion of a service or deliver a patching stream, rather than multicasting the video in its entirety. Given that there is an existing multicast video, when to schedule another multicast for the same video is crucial. The time period after a multicast, during which patching must be used, is called the patching window [14]. Two simple approaches to setting the patching window are discussed in [41]. The rst one uses the length of the video as the patching win- dow. That is, no multicast is initiated as long as there is an in-progress multicast session for the video. This approach is called the greedy patching because it tries to exploit an in-progress multicast as much as possible. However, an over- greed can actually reduce data sharing [41]. The second approach, called the grace patching , uses a patching stream for the new client only if it has enough buer space to absorb the skew. Hence, under grace patching, the patching window is determined by the client buer size. Considering such factors as video length, client buer size, and request rate, the authors of [15] generalized patching by determining the optimal patching window for each video. An improved form of patching, called as the transition patching [15], uses either a patching stream or a transition stream and improves performance without requiring any extra download band-width at the client site. Other optimal patching schemes were described in [30, 74]. In patching, a client might have to download data on both regular multicast and patching channels simultaneously. To implement patching, a client station needs three threads: two data loaders to download data from the two channels, and a video player to play back the video. The controlled CIWP (Client-Initiated-With-Prefetching) [35] is another multicast technique similar to patching and tapping for near instantaneous VoD service. The novelty of the controlled CIWP is that it uses a threshold to control the frequency of multicasting a complete video stream. It uses simple FCFS channel scheduling so that a client can be informed immediately of when its request will begin service. 3.3.2 Server-initiated batching In server-initiated batching, the bandwidth is dedicated to video objects rather than to users. Videos are decomposed into segments which are then broadcast periodically via dedicated channels, and hence, it is also called periodic broadcast . Although the worst-case service latency experienced by any subscriber is guaranteed to be less than the interval of broadcasting the leading segment and is independent of the current number of pending requests, this strategy is more e-cient for popular videos than for unpopular ones due to the xed cost of channels. One of earlier periodic broadcast schemes was the Equally-spaced interval Broadcasting (EB) [21]. Since it broadcasts a given video at equally-spaced intervals, the service latency can only be improved linearly with the increase of the server bandwidth. The author of [10] also proposed the staggered VoD which broadcasts multiple copies of the same video at staggered times. To signicantly reduce the service latency, Pyramid Broadcasting (PB) was introduced in [82]. In PB, each video le is partitioned into the segments of geometrically-increasing sizes, and the server capacity is evenly divided into K logical channels. The i-th channel is used to broadcast the i-th segments of all videos sequentially. Since the rst segments are very small, they can be broadcast more frequently through the rst channel. This ensures a smaller waiting time for every video. A drawback of this scheme is that a large buer | which usually corresponds to more than 70% of the video | must be used at the receiving end, requiring disks for buering. Furthermore, since a very high transmission rate is used for each video segment, an extremely high bandwidth is required to write data to the disk as quickly as it receives the video. To address these issues, the authors of [4] proposed a technique called Permutation-based Pyramid Broadcasting (PPB). PPB is similar to PB except that each channel multiplexes its own segments (in- stead of transmitting them sequentially), and a new stream is started once every short period. This strategy allows PPB to reduce both disk space and I/O bandwidth requirements at the receivers. However, the required disk size is still large due to the exponential nature of the data fragmentation scheme. The sizes of successive segments increase exponen- tially, thus causing the size of the last segment to be very large (typically more than 50% of the video). Since the buer sizes are determined by the largest segment, using the same data fragmentation scheme proposed for PB limits the savings achievable by PPB. In PPB, a client needs to tune in dierent logical subchannels to collect its data for a given data fragment if the maximum savings in disk space is desirable. To reduce the disk costs in the client side, the authors of [42] introduced Skyscraper Broadcasting (SB) which uses a new data fragmentation technique and proposes a dier- ent broadcasting strategy. In SB, K channels are assigned to each of the N most popular objects. Each of these K channels transports a specic segment of the video at the playback rate. The progression of relative segments size on the channel, f1,2,2,5,5,12,12,25,25,52,52,105,105,: : :g, is bounded by the width parameter W , in order to limit the storage capacity required at the client end. SB allows for simple and e-cient implementation, and can achieve a low service latency while using only 20% of the buer space required by PPB. The authors of [34] provided a framework for broadcasting schemes, and designed a family of schemes for broadcasting popular videos, called the Greedy Disk- conserving Broadcasting (GDB). They systematically analyze the resource requirements, i.e., the number of server broadcast channels, the client storage space, and the client I/O bandwidth required by GDB. GDB exhibits a trade- between any two of the three resources, and outperforms SB in the sense of reducing resource requirements. The Dynamic Skyscraper Broadcasting (DSB) in [28] dynamically schedules the objects that are broadcast on the skyscraper channels to provide all clients with a precise time at which their requested objects will be broadcast, or an upper bound on that time if the delay is small and reaps the cost/performance benets of the skyscraper broadcasting. The above broadcasting schemes generally assume that the client I/O bandwidths are limited to download data from Scheduling strategy Features Typical methods Client-initiated the channels are allocated among the users, Adaptive Piggybacking [38], Patching [41] (scheduled, the multicast tree can be expanded dynamically Chaining [77], Tapping [16] and client-pull) to accommodate new requests so that the Controlled CIWP [35], etc. service latency is minimized (ideally zero). Server-initiated the channels are dedicated to video objects, Only two download channels: (periodic broadcast, the videos are divided into segments which EB [21], PB [82], PPB [4], server-push) are then broadcast periodically via dedicated SB [42], GDB [34], DSB [28],etc. channels, the worst-case service latency More than two download channels: experienced by any client is less than the Harmonic broadcasting [47],Staircase interval of broadcasting the leading segment scheme [48] and FB [46, 80], etc. Hybrid scheduling the overall performance is improved by combining Controlled multicast [35], Catching client-initiated and server-initiated strategies. and Selective catching [36], etc. Table 3: The summary of existing data-centered approaches only two channels. If the client can download data from more than two channels, there are methods available that can e-ciently reduce the service latency with less broadcasting channels. For example, a broadcasting scheme based on the concept of harmonic series is proposed in [47, 49]; the scheme doesn't require the bandwidth assigned to a video equal to a multiple of a channel's bandwidth. For a movie of length of D minutes, if we want to reduce the viewer's waiting time to D=N minutes, we only need to allocate H(N) video channels to broadcast the movie peri- odically, where H(N) is the harmonic number of N , i.e., N . The staircase scheme in [48] can reduce the storage and disk transfer-rate requirement at the client end. However, both the staircase and harmonic schemes cannot serve buerless users. In [46, 50], a scheme called Fast Broadcasting (FB) is proposed, which can further reduce the waiting time and the buer requirement. Using FB, if a STB does not have any buer, its user can still view a movie insofar as a longer waiting time is ac- ceptable. The authors of [80] proposed two enhancements to FB, showing how to dynamically change the number of channels assigned to the video and seamlessly perform this transition, and presenting a greedy scheme to assign a set of channels to a set of videos such that the average viewers' waiting time is minimal. 3.3.3 Hybrid multicast scheduling All practical scheduling policies are guided by three primary objectives: minimize the reneging probability, minimize average waiting time, and be fair. It was shown in [21, 22, 35, 36] that a hybrid of the above two techniques oered the best performance. For example, the Catching proposed in [36] is a combination of periodic broadcast and client-initiated prex retrieval of popular videos. There are many hybrid schemes to improve the overall performance of multicast VoD. The selective catching in [36] further improves the overall performance by combining catching and controlled multicast to account for diverse user access pat- terns. Because most demands are on a few very popular movies, more channels are assigned to popular videos. However, it is necessary (and important) to support unpopular videos. We assume that scheduled multicasts are used to handle less popular videos, while the server-initiated scheme is used for popular videos. In this approach, a fraction of server channels are reserved and pre-allocated for periodically-broadcasting popular videos. The remaining channels are used to serve the rest of the videos using some scheduled multicasts. This hybrid of server-initiated and client-initiated schemes achieves better overall performance. The existing data-centered approaches are summarized in Table 3. 3.4 Multicast routing and protocols There has been extensive research into multicast routing algorithms and protocol [17, 72, 85]. Multicast can be implemented on both LANs and WANs. Nodes connected to a LAN often communicate via a broadcast network, while nodes connected to a WAN communicate via switched net- works. In a broadcast LAN, transmission from any one node is received by all the other nodes on the network, so it is easy to implement multicast on a broadcast LAN. On the other hand, it is challenging | due mainly to the problem of scalability | to implement multicast on a switched network. Today's WANs are designed to mainly support unicast com- munication, but in future, as multicast applications become more popular and widespread, there will be a pressing need to provide e-cient multicast support on WANs. In fact, the multicast backbone (MBone) of the Internet is an attempt toward this goal. For multicast video transmissions, one of the key issues is QoS routing which selects routes with su-cient resources to provide the requested QoS. For instance, the multicast VoD service requires its data throughput to be guaranteed at or above a certain rate. The goal of QoS routing is twofold: (1) meet the QoS requirements for every admitted connection, and (2) achieve global e-ciency in resource utilization. In most cases, the problems of QoS routing are proven to be NP-complete [86]. Routing strategies can be classied as source routing, distributed or hierarchical routing. Some heuristic QoS routing algorithms have been proposed (see [17, 85] for an excellent survey of existing multicast QoS routing schemes). In an eort to provide QoS for video transmissions, a number of services have been dened in the Internet. A Resource Reservation Protocol (RSVP) has been developed to provide receiver-initiated xed/shared resource reservation for unicast/multicast data ows [13] after nding a feasible path/tree to satisfy the QoS requirements. Furthermore, a protocol framework for supporting continuous media has been developed: RTP (Real-Time Protocol) [73] provides support for timing information, packet sequence numbers and option specication, without imposing any additional error control or sequencing mechanisms. Its companion control protocol, RTCP (Real-Time Control Protocol), can be used for gathering feedback from the receivers, again according to the application's need. 3.5 The client-end system Customer premise equipments (CPEs) include set-top boxes (STBs), disks, and display monitors, where a disk or a RAM is used as a buer. As an example, the disk space of 100 MB can cache about 10 minutes of MPEG-1 video. Such a disk space costs less than $10 today. The high cost of a VoD system is due mostly to the network costs. For instance, the cost of networking contributes more than 90% of the hardware cost of the Time Warner's Full Service Network project. The client's STB, from software perspectives, generally contains a main control thread, video stream receiver threads and a video player thread. A client is connected to the network via a STB. The main control thread processes the client's service request by sending a message indicating his desired video to the server. It then forks the video stream receiver threads to select one or more network channels to receive and decompress video data according to the server's instructions. The received video data are either stored on the disk or sent to the display monitor for immediate play- back. The display monitor can either retrieve stored data from the disk or receive data directly from a channel. The CPE buer plays important roles as follows. Supporting the VCR interactions of a customer [2, 7, 70]. The interaction protocols for multicast VoD are designed by using the CPE buer. Providing instant access to the stored video program so as to minimize the service latency [36, 69]. Preloading and caching, based on the video stored in the CPE buer, can reduce the service latency. Reducing the bandwidth required to transmit the stored video [36, 69]. Because some video data reside in the CPE buer, the overall bandwidth requirement of transmitting videos is reduced. Eliminating the additional bandwidth required to guarantee jitter-free delivery of the compressed video stream. These functions are discussed in detail in the following sub-sections 3.6 Support for interactive functions One of the important requirements is to oer VCR interac- tivity. In order to support customers' interactive behavior in multicast VoD service, there have been e-cient techniques proposed by a combination of tuning and merging as well as using the CPE buer and I-channels (see Table 4). The authors of [10] introduced tuning in staggered VoD which broadcasts multiple copies of the same video at staggered times. Intelligently tuning to dierent broadcast channels is used to perform a user interaction. However, not all interactions can be achieved by jumping to dierent streams. Moreover, even if the system can emulate some interactions, it cannot guarantee the exact eect the user wants. Other solutions to VCR interactivity are proposed in [7] and [23], especially for handling a pause/resume request. Support for continuous service of pause operations was simulated for a NVoD server, but merge operations from I-channels to batching- (or B-) channels were either ignored [23] or did not guarantee continuity in video playout [7]. [7] proposed the use of the CPE buer to provide limited interactive func- tions. In order to implement the interactivity of multicast VoD services, more e-cient schemes have been proposed. For example, the SAM protocol [53] oers an e-cient way for TVoD interactions, and all those introduced in Section 2 are provided by allocating the I-channels as soon as a VCR action request is issued. When playout resumes, the VoD server attempts to merge the users back into a B-channel by using a dedicated synch buer located at access nodes and partially shared by all the users. Should this attempt fail, a request for a new B-channel is then initiated. The drawback of the SAM protocol requires an excessive number of I-channels, thus causing a high blocking rate of VCR interactions. The authors of [2] improved the SAM protocol by using the CPE buer and active buer manage- ment, and hence, more interactions can be supported without I-channel allocation. The BEP (Best-Eort Patching) scheme proposed in [56] presents an e-cient approach to the implementation of continuous TVoD interactions. Compared to the other methods, BEP aims to oer zero-delay (or continuous) service for both request admission and VCR interaction, whereas the SAM protocol just supports continuous VCR interactions without considering service admis- sion. Moreover, BEP uses a dynamic technique to merge interaction streams with a regular multicast stream. This technique signicantly improves the e-ciency of multicast TVoD for popular videos. The authors of [70] proposed another scheme called the Single-Rate Multicast Double-Rate Unicast (SRMDRU) to minimize the system resources required for supporting full VCR functionality in a multicast VoD system. This scheme also supports TVoD service, so customers can be served as soon as their requests are received by the system. It forces customers in unicast streams (on the I-channel) to be served by multicast streams again after they resume from VCR operations. The idea is to double the transmission rate of the unicast stream so that the customer of normal playback can receive the frame synchronized with the transmitting frame of a multicast group. 4. ISSUES RELATED TO MVOD SERVICE 4.1 QoS of multicast VoD The eectiveness of a video delivery technique must be evaluated in terms of both the server and network resources required for delivering a video object and the expected service latency experienced by the clients. Reducing the service latency is an important goal in designing eective scheduling strategies. The existing dynamic multicast and periodic broadcast schemes reviewed in Section 3.3 are shown to achieve good performance. Besides the dynamic scheduling schemes, other schemes, such as caching and preloading , have been proposed to reduce the service latency. In [29, 75], proxy servers are used to cache the initial segments of popular videos to improve the service latency. Because nearly all broadcast protocols assume that the CPE buer is large enough to store up to 40 or 50 % of each video (about 50 minutes of a typical movie), the partial preloading proposed in [69] uses this storage to preload anticipated customers' requests, say, the rst 3 minutes of top 16 to 20 videos. It will provide instantaneous access to these videos and also reduce the bandwidth required Level of interaction Features Typical methods NVoD the interactive functions are simulated tuning [10] by transitions in discrete time interval. Limited TVoD the continuous interaction times are that supported by CPE buer [7] limited by the available resource. TVoD full control the durations of SAM [53], Improved SAM [2] all continuous interactions. BEP [56], SRMDRU [70] Table 4: The summary of interaction schemes for multicast VoD Solution Features Comparison Active transcoding data streams are individually transformed according an administrative burden, to the specication of each requesting receiver di-cult to deploy proxies Layered multicast encodes source data as a series of layers, sends complex adaptation scheme dierent layers for dierent multicast groups. in client-end Table 5: The solutions of handling client heterogeneity to broadcast them as well as the extra bandwidth required to guarantee jitter-free delivery of the compressed video signal. It diers from proxy caching in that the preloaded portions of each video will reside inside the CPE buer rather than at a proxy server. The customers' defection rate is closely related to the service latency, and is inversely proportional to the server throughput, or an average number of service requests granted per program. The shorter the service latency, the lower the defection rate becomes and the higher the server through-put is. Another important QoS parameter is the VCR action blocking probability. All the existing multicast TVoD protocols covered in Section 3.6 aim to reduce the blocking probability or discontinuity of VCR interactions. 4.2 Client heterogeneity As the multicast network expands, there will be various types of end devices ranging from simple palm-top personal digital assistants (PDAs) to powerful desktop PCs or HDTV receivers of multicast VoD. Since there will be multiple VoD transmission rates or paths, the sender alone cannot meet the possibly con icting demands of dierent receivers. Distributing a uniform representation of the video to all receivers could cause low-capacity regions of the network to suer from congestion, and some receivers' QoS cannot be met even when there are su-cient network resources to provide better QoS for these receivers. In the context of VoD, scalability also applies to the server's ability to support the data requirements of multiple terminal types. One way to solve this problem is proxy-based transcoding, where data streams are individually transformed according to the specication of each requesting receiver [31]. However, it typically imposes an administrative burden because it is not transparent to end users. Proxies are also di-cult to deploy because a user behind a constrained network link might not have access to the optimal location of a proxy. There was a proposal to use active networks to solve this problem by oering a common platform for such services as part of the basic network service model [79], but there remain many issues to be addressed before such an infrastructure can be deployed. Furthermore, transcoding proxies must be highly reliable and scalable which can be very costly [31]. Another e-cient solution to heterogeneity is the use of layered media formats. This scheme encodes source data as a series of layers, the lowest layer being called the base layer and higher layers being called the enhancement layers. Layered encoding can be eectively combined with multicast transmission by sending dierent layers for dierent multi-cast groups. Consequently, a receiver using only the basic multicast service (i.e., joining and leaving multicast groups) can individually tailor its service to match its capabilities, independently of other receivers. This basic framework was later rened in a protocol architecture called Receiver-driven Layered Multicast (RLM) [58]. In RLM, a receiver searches for the optimal number of layers by experimentally joining and leaving multicast groups much in the same way as a TCP source searches for the bottleneck transmission rate with the slow-start congestion avoidance algorithm [44]. The receiver adds layers until congestion occurs and backs o to an operating point below this bottleneck. The two solutions to heterogeneity are summarized in Table 5. 4.3 Fairness of multicast VoD service Fairness is one of the performance metrics in VoD service, meaning that every client request should be fairly treated regardless whether it is for a hot video or not. In [41], the unfairness of a multicast VoD system is expressed as a function of the defection rate, that is, d) , where d i denotes the defection rate for video i, d is the mean defection rate, and N is the number of videos. Alternatively, this property can also be measured by video service latencies. The fairness is mainly related to scheduling and resource allocation. When selecting a scheduling strategy, we make it as fair as possible. The fairness of certain batching schemes are surveyed in Section 3.3. However, scheduling strategies like various periodic broadcasts, are only for popular videos, and the fairness depends on the scheduling scheme used for cold videos and the bandwidth allocation among hot and cold videos. Unfortunately, there are only a very few attempts to analyze the fairness of existing scheduling schemes [41]. How to assure the fairness of practical scheduling schemes, particularly for hybrid schemes, and make the optimal bandwidth resource allocation are open issues. 4.4 Customer behavior Understanding customer behaviors is necessary to eciently design a multicast VoD system and take dierent strategies to dierent videos at dierent times. Modeling customers' behaviors includes video selection distribution, variations of video popularity, and the user interaction model. 4.4.1 Video selection distribution For short-term considerations, most researchers assume that the popularity of videos follows the Zipf distribution [87], that is, the probability of choosing the i-th video isi 1 z P N , where N is the total number of videos in the system, and z is called the skew factor . Typically, researchers assume that the skew factor be set to 0.271 [5, 21]. This number is obtained from the analysis of a user access pattern from a video rental store [5]. That is, most of the demand (80%) is for a few (10 to 20) very popular videos. 4.4.2 Time-variation of video popularity In a real VoD system, request arrivals are usually nonsta- tionary. Variations in the request rate can be observed on a daily basis, between \prime time" (e.g., 7 p.m.{10 p.m.) and \o-hours" (e.g., early morning). On a larger time scale (e.g., one week or month), movie popularities may change due to new releases or loss of customers' interest in current titles. At the same time, the dierent types of customers (e.g., children and adult) have dierent prime times. In [1], a time distribution model is expressed as sinusoidal: where 0 is the daily average arrival rate, A(> 0) is the amplitude, being a 24-hour period), and pm the popularity of movie title m. More general models of nonstationarity have been proposed in [8, 39] for the long-term popularity of movie. We call these time-dependent changes of movie popularity the life-cycle of the movie. The authors of [39] observed that the long-term behavior of a movie follows an exponential curve plus a random eect. [8] also assumed that variations in workload are exponential functions with dierent average inter-arrival times. 4.4.3 Interaction model Some interaction models have been proposed in [2, 26, 54]. In [26], the behavior of each user is modeled by a two-state Markov process with states PB (playback) and FF/Rew. The times spent by the user in PB and FF/Rew modes are exponentially distributed. The two-state model in [54] assumes that the user activity is in either the normal or the interaction state. But these two models are too simple to represent the details of user interactions. To be realistic, a model should capture three specic parameters for their potentially signicant impacts on the performance of the VoD server: (1) the frequency of requests for VCR actions; (2) the duration of VCR actions; (3) the bias of interaction behavior. Considering these parameters, the authors of [2] proposed a VCR interaction model. In this model, a set of states corresponding to the dierent VCR actions are designed durations and probabilities of transitioning to neighboring states. The initial state is Play, and the inter-action system randomly transits to other interactive states or stays in the Play state according to the behavior distri- Slow Play/ Fast Fast Reverse1 Forward Search Search Resume Rewind Motion d6 d2 d3 d7 d8 Abort Pause d5 Stop d4 Figure 4: VCR interactive model bution. The user resides at each interaction state for an exponentially-distributed period of time. As shown in Figure 4, transition probabilities P are assigned to a set of states corresponding to dierent VCR actions. It is important to notice that the above-mentioned parameters are captured by this representation of viewers' activity. Finding representative values for P still an open issue. For tractability, customers are classied to be Very Interactive (VI) or Not Very Interactive (NVI). Their interactions can be simulated by taking dierent parameter values [2]. 4.5 Evaluation of multicast VoD systems By batching multiple requests and serving them with a single multicast, the system capability for handling a large number of requests can be greatly improved at the expense of increased admission delay. For zero-delay admission and VCR interactions, more channels are required. It is important to evaluate multicast VoD in terms of throughput, resource requirement, and e-ciency. These evaluation results will in uence pricing, system management as well as resource sharing. 4.5.1 The service throughput For NVoD service, we need to evaluate the throughput of multicast VoD. Common assumptions include the Poisson arrival of clients' requests, and customers' willingness to wait for a xed amount of time, say, 5 minutes, before withdrawing their requests. The authors of [1] modeled the customers' patience with an exponential distribution, i.e., a customer agrees to wait for units of time or more with probability , where is the average time customers agreed to wait. In general, the patience rate can be assumed independently of the videos requested. Based on this assumption, the authors of [1] converted the problem of calculating the number of customers waiting between two consecutive services into that of making a transient analysis of the M=M=1 \self-service" queueing system with arrival rate , and self-service with a negative exponential distribution with rate . They then derived the server throughput and the average loss rate for each movie. In [78], the user wait-tolerant behavior in some batching schemes, such as FCFS (rst-come-rst-serve), MQL (the maximum queue length), Max-Batch and Min-Idle, have also been investi- gated, the problem of maximizing the system throughput is formally discussed and the functional equation dening the optimal scheme is derived. For TVoD service in multicast VoD, there is, to our best knowledge, no literature on evaluating the service through- put. It is still an open issue. 4.5.2 Bandwidth requirements of TVoD service Some multicast VoD protocols can support TVoD services in zero-delay admission if enough channels are used. How to evaluate the channel requirements depends on the underlying multicast scheduling scheme. For example, [14] presents the optimal patching performance, and [15] addresses the bandwidth requirement of transition patching. [57] proposes a method for analyzing the user interactivity and evaluating the number of channels required for multicast TVoD service. The results determine the relationships among the clients' behaviors, the system resources, and the TVoD service pro- tocol. However, rigorously analyzing the requirements of channels for TVoD protocols supporting zero-delay admission and full VCR interactions is still an open issue. 4.5.3 Bandwidth cost vs. QoS Although most server-initiated multicast VoD schemes are not for providing TVoD service, they strive to make a better tradeo between the channel cost and the service latency. For example, most periodic broadcasting schemes try to reduce the service latency and improve the throughput at low channel costs. The performance of related periodic broadcasting schemes have been discussed in [4, 34, 42, 83]. 4.6 Other issues Besides the mentioned issues, there also are other important issues in multicast VoD service, such as: Copyright protection: A typical solution to video copyright protection is an encrypted point-to-point delivery to ensure that only paying customers receive the service, but it is ine-cient for multicast VoD. In [40] a simple scheme is proposed to provide similar protection for the content, which can be used e-ciently with multi-cast and caching. In that scheme, the major part of the video is intentionally corrupted and can be distributed via multicast connections, while the part necessary for reconstructing the original is delivered to each receiver individually. However, the e-cient copyright protection for multicast VoD service is still an open issue. Video replacement: The maintenance of storage is a main server management task in general VoD systems. Due to the limited storage capacity, it is necessary to replace old, unpopular videos by new popular videos. How to select videos to replace depends mainly on the hit ratio of videos. Popularity-based assignment and BSR(Bandwidth-to-Space Ratio)-based assignment are considered as the important policies. Moreover, the replacement operation for one video shouldn't aect the service of other videos residing in the same server. This problem is related to storage organization, and [3] discussed the eect of video partitioning strategies. 5. CONCLUSIONANDFUTUREDIRECTIONS As VoD service becomes popular and widely-deployed, consumers will likely suer from network and server over- loads. Network and server-I/O bandwidths have been recognized as a serious bottleneck in today's VoD systems. Multicast is a good remedy for this problem; it alleviates server bottlenecks and reduces network tra-c, thereby improving system throughput and server response time. We discussed the state-of-art designs and solutions to the problems associated with multicast VoD. We also illustrated the benets of multicast VoD, and feasible approaches to implementing it. Although multicast VoD can make signicant performance improvements, there are still several open issues that warrant further research, including: Eective scheduling and routing schemes, and VCR- type interactions for multicast TVoD. Ad hoc schemes may achieve a better trade-o between QoS and system costs while preserving scalability and interactivity. E-cient active CPE buer management. It is highly dependent on customers' behavior, and utilizing the CPE buer will signicantly improve QoS. Fairness of VoD service. There is a tradeo between fairness and throughput for practical scheduling schemes, particularly for hybrid schemes. We need to achieve the optimal bandwidth resource allocation without loss of fairness. Knowledge of realistic customers' behavior is essential to the design of multicast VoD protocols and resource allocation. Other issues related to customers' behavior are throughput, scaling, and scheduling. An e-cient theoretical framework for evaluating the performance of multicast VoD service, especially for modeling and analyzing multicast TVoD service to support both zero-latency admission and full VCR inter-activity Developing standard protocols for multicast VoD for practical applications. The existing protocols for multicast routing, scheduling and VCR interactions can be viewed as a basis to achieve this goal. The DAVIC protocol also provides a general reference framework, but protocols for multicast TVoD still need to be developed further. 6. --R Human Behaviour and the Principle of Least E --TR Congestion avoidance and control Multicast routing in datagram internetworks and extended LANs Scheduling policies for an on-demand video server with batching Providing VCR capabilities in large-scale video servers Evaluating video layout strategies for a high-performance storage server Choosing the best storage system for video service Channel allocation under batching and VCR control in video-on-demand systems Reducing I/O demand in video-on-demand storage servers The SPIFFI scalable video-on-demand system Dynamic batching policies for an on-demand video server Adaptive piggybacking On optimal piggyback merging policies for video-on-demand systems Fault-tolerant architectures for continuous media servers Metropolitan area video-on-demand service using pyramid broadcasting Adapting to network and client variability via on-demand dynamic distillation Receiver-driven layered multicast Group-guaranteed channel capacity in multimedia storage servers Skyscraper broadcasting Long-term movie popularity models in video-on-demand systems Scheduling video programs in near video-on-demand systems The multimedia multicasting problem Protecting VoD the easier way <italic>Patching</italic> Exploring wait tolerance in effective batching for video-on-demand scheduling Zero-delay broadcasting protocols for video-on-demand Optimal and efficient merging schedules for video-on-demand servers Catching and selective catching An efficient bandwidth-sharing technique for true video on demand systems ARMIDA TM A VoD Application Implemented in Java Prospects for Interactive Video-on-Demand The Split and Merge Protocol for Interactive Video-on-Demand Design of Multimedia Storage Systems for On-Demand Playback A Redundant Hierachical Structure for a Distributed Continuous Media Server A Low-Cost Storage Server for Movie on Demand Databases The KYDONIA Multimedia Information Server A New Scheduling Scheme for Multicast True VoD Service Dynamic Skyscraper Broadcasts for Video-on-Demand Fast broadcasting for hot video access Supplying Instantaneous Video-on-Demand Services Using Controlled Multicast Earthworm Long Term Resource Allocation in Video Delivery Systems Chaining Video-on-Demand Server Efficiency through Stream Tapping Providing Unrestricted VCR Functions in Multicast Video-on-Demand Servers Demand Paging for Video-on-Demand Servers --CTR Jian-Guang Lou , Hua Cai , Jiang Li, Interactive multiview video delivery based on IP multicast, Advances in Multimedia, v.2007 n.1, p.13-13, January 2007 Leonardo Bidese de Pinho , Claudio Luis de Amorim, Assessing the efficiency of stream reuse techniques in P2P video-on-demand systems, Journal of Network and Computer Applications, v.29 n.1, p.25-45, January 2006 Azzedine Boukerche , Richard W. Nelem Pazzi, Scheduling and buffering mechanisms for remote rendering streaming in virtual walkthrough class of applications, Proceedings of the 2nd ACM international workshop on Wireless multimedia networking and performance modeling, October 06-06, 2006, Terromolinos, Spain Ma Huadong , Kang G. Shin, Hybrid broadcast for the video-on-demand service, Journal of Computer Science and Technology, v.17 n.4, p.397-410, July 2002 Huadong Ma , G. Kang Shin , Weibiao Wu, Best-Effort Patching for Multicast True VoD Service, Multimedia Tools and Applications, v.26 n.1, p.101-122, May 2005 Bharadwaj Veeravalli , Long Chen , Hun Yen Kwoon , Goh Kar Whee , See Ying Lai , Lim Peng Hian , Ho Chin Chow, Design, analysis, and implementation of an agent driven pull-based distributed video-on-demand system, Multimedia Tools and Applications, v.28 n.1, p.89-118, January 2006 Carlo K. da S. Rodrigues , Rosa M. M. Leo, Bandwidth usage distribution of multimedia servers using Patching, Computer Networks: The International Journal of Computer and Telecommunications Networking, v.51 n.3, p.569-587, February, 2007

Video-on-Demand VoD;VCR-like interactivity;scheduling;multicast;Quality-of-Service QoS

510775

On securely scheduling a meeting.

When people want to schedule a meeting, their agendas must be compared to find a time suitable for all participants. At the same time, people want to keep their agendas private. This paper presents several approaches which intend to solve this contradiction. A custom (---) made protocol for secure meeting scheduling and a protocol based on secure distributed computing are discussed. The security properties and complexity of these protocols are compared. A trade-off between trust and bandwidth requirements is shown to be possible by implementing the protocols using mobile agents.

INTRODUCTION When negotiating meetings, the participants look up, communicate and process information about each other's agendas trying to nd a moment when they are all free to attend the meeting. Due to the private nature of a person's schedule, as little as possible should be revealed to any other party during that negotiation. Ideally, only the result of the negotiation should be known to the participants (and to the participants only), and any other information about the users' agendas should remain secret. An easy solution for scheduling a meeting is to broadcast the schedules to all participants, but this totally neglects the privacy of the partici- pants' agendas. Another solution is to send all schedules to a trusted third party, but nding one such single third party trusted by every participant, will be very di-cult in practice. Some existing meeting scheduling applications, like for example \Ya- hoo! Calendar", dene access levels for viewing and modifying agenda entries, and dene user groups to which these access levels are assigned. This is only necessary because the comparison between schedules must be done by the users themselves. Our approaches eliminate the need for managing access control, as they are not based on users directly accessing each other's agenda. This paper presents more secure solutions. Their goal is for participants to be able to negotiate a meeting whereby parties have no direct access to each other's agenda, whereby parties do not rely on another party for telling the nal result, and whereby no information about the agendas is revealed, but the nal result, i.e., the particular time the meeting can be scheduled, or the fact that the meeting cannot be scheduled This paper builds on the work done in [6] and [3] and shows the trade- os that can be made in security, level of trust, and e-ciency, when choosing a particular negotiation protocol and a specic implementation approach. The paper is organized as follows. Section 2 presents a custom-made negotiation protocol. Section 3 presents an alternative approach based on secure distributed computing. Both approaches are analyzed from a security and complexity point of view. Section 4 discusses the use of mobile agents for secure meeting scheduling, and presents the \agenTa" prototype implementation. We conclude in Sect. 5. 2. USING A CUSTOM-MADE NEGOTIATION PROTOCOL 2.1. There exists a representation which reduces the problem of deciding if the meeting can be scheduled at a certain moment to a logical AND operation. As shown in Fig. 1, an agenda will be represented as a bit string in the following way: for each time slot in the schedule, there is one bit indicating whether the negotiator can (1) or cannot (0) attend a size of meeting schedule granularity possible grid tation Figure 1 Conversion from agenda to representation meeting of the specied length which would start at that time. The ner the granularity and the longer the negotiation window, the more bits there will be in the representation. 2.2. SCHEDULING MODEL In our model, a meeting scheduling starts with an invitation phase. The initiator broadcasts to the invitees a set of negotiation parameters such as meeting length, negotiation window (limited time span in which to attempt the meeting scheduling) and a complete list of invitees. Each invitee broadcasts to all others a reply indicating whether it will accept or decline the negotiation invitation. Because broadcasts are used, no invitee can be mislead as to the set of negotiators it will encounter in the second phase. In the second phase, called negotiation, the negotiators try the time slots one by one and attempt to schedule the meeting. For each time slot the negotiation takes place according to the protocol outlined be- low. If the meeting was successfully scheduled the negotiators move on to the third phase, otherwise the next time slot is tried. After independently arriving to a result concerning a certain time slot, each participant broadcasts the result to the others and checks whether all results coin- cide. This allows for detection of partial failures and attacks which try to mislead a subset of the negotiators. In the third phase either the common result is presented to the users, or the users are informed that no meeting can take place. If there is a common result, users might conrm their commitment to the scheduled time on a separate channel (e-mail, telephone), independently of the scheduling process. 2.3. SCHEDULING A MEETING For the purpose of this subsection we will refer to the representation of an agenda according to the description in the previous subsection as \schedule." Instead of comparing schedules, the negotiation should be based on comparing protected forms of the schedules. The schedules are protected in a way which still allows scheduling to be performed by broadcasting the protected forms to all negotiators and letting them process the data without fear of the unprotected form to be revealed. The binary XOR operation between the schedule and a mask is a transformation which still allows scheduling to be performed in the sense that the (in)equality of two or more bits is preserved when they are all XORed with the same mask. If all negotiators know the mask, they are able to retrieve the original schedules easily, by unmasking the broadcasted data. The solution is to let the mask be a shared secret, that is, all negotiators will contribute when building it, but it will not be revealed to any of them. The negotiation protocol then goes as follows: In step one of the negotiation protocol, each negotiator chooses a random mask, and XORs it with its schedule. This random mask is actually a partial mask. The shared secret will be the XOR of all partial masks, and is called global mask. Even if only one negotiator keeps its partial mask secret, the others cannot nd the global mask solely using their partial masks. In step two of the protocol, all schedules visit all negotiators exactly one time. At each visit they are masked with the partial mask of that particular negotiator. In the end, all original schedules are thus masked with the global mask, without the need for the negotiators to disclose their partial mask. Since the schedule is rst masked with its owner's partial mask it remains secret during its visits. A negotiator must be unable to identify a protected schedule as representing its own schedule: otherwise performing XOR between the original and the protected schedule reveals the global mask, allowing the negotiator to retrieve all original schedules. Therefore during the trip to all negotiators, the schedule must be forwarded randomly between the negotiators in order to make it impossible to trace. The schedule must have attached a list of negotiators it hasn't visited yet, decremented at each forwarding, in order to prevent multiple maskings with the same partial mask. Note that for countering attempts to trace a schedule by attackers who have a global view on the network, all communications should be encrypted. 3 In step three, all protected schedules are broadcasted. Each negotiator looks independently for a time slot when all protected schedules have the same value. That implies that the original schedules are identical, too, for that time slot but does not provide any clue whether the negotiators are free or busy for that time slot. The clue is provided by each negotiator's schedule for that time slot. If the negotiator is free then, it means all negotiators are free then and the meeting can be scheduled. For time slots when some are busy and some are free, it is not possible to gure out who are the busy ones and who are the free ones. Note that our scheduling protocol does not specify any form of negotiator authentication. This is however needed for linking the protocol messages to their originators. Depending on the meeting application, the desired form of authentication can be added to the protocol. Figure 2 shows the negotiation protocol as performed by three parties. For easy understanding of the protocol the schedules in the simulation are following the same route and the maskings appear to be performed simultaneously by the three negotiators. In reality the process is asynchronous (some negotiators may be idle while others are masking) and routing is random (in the end some negotiators may have nothing to broadcast while others may broadcast several protected schedules). Another dierence is that in reality only one bit is processed at a time (otherwise an attack is possible, see following section). If the meeting can be scheduled in the corresponding time slot the protocol stops, otherwise the next time slot is processed. 2.4. SECURITY ANALYSIS Our custom-made protocol does not require one single entity to be trusted. It however does not completely protect the privacy of the par- ticipants' agenda, as attacks by both passive and active adversaries are possible. Bad slots. There may be time slots for which all users are busy and therefore all protected slots will be equal. By checking against the original schedule each negotiator will avoid scheduling a meeting in that slot but it will also know everybody else's schedule for that slot (i.e., everybody is busy). Because they constitute an infringement on all users' privacy we call these slots bad slots. A step one step two step three Figure Simulation for three negotiators Entropy attack. The reason for performing the negotiation one slot at a time is to prevent the following attack. If the negotiation is done on sequences of slots, when all the broadcasted masked schedules are received, it still is possible for a party to recognize its original schedule. It can be done by testing all the masks which transform the original schedule into one of the protected forms. The correct global mask can be recognized by the fact that by unmasking the other protected schedules with it, bit strings are obtained which have the entropy expected from a schedule. Negotiating one bit at a time, with fresh partial masks for each bit and stopping when a meeting is scheduled counters this attack because each mask bit and schedule bit have maximal entropy. Number of parties. When only two parties are negotiating, each can deduce the schedule of the other based on their own schedule and the comparison between the protected forms of the schedules. Besides that, the global mask is straightforward to nd because the original schedule can be linked to its protected form. Also when only three or four parties are negotiating it is sometimes possible to nd out the global mask by tracing back schedules. For ve or more participants the ability to trace a schedule along its route decreases as the number of participants increases. Dummy negotiators could be introduced to articially increase the number of parties, and thus to alleviate this problem. In a broadcasting communication environment encrypted dummy messages could also be sent to make the real schedules untraceable. Rogue negotiators. Active adversaries could attack the protocol in various ways. A simple denial of service attack can be mounted by negotiating based on a fully busy schedule instead of declining the invitation. Since the protocol relies on the negotiators consistently using their partial mask, the protocol has unpredictable outcomes if a negotiator randomly changes its partial mask during the negotiation of a time slot. Goal-oriented misbehavior is also possible. A negotiator can wait to be the last to broadcast the protected schedule(s) it has. This way it is able to detect rst when a meeting could take place. In that case it can broadcast a false protected form, preventing the meeting from being scheduled. It knows everybody else's schedule for that time slot, while the others do not. 2.5. COMPLEXITY For analyzing the complexity of the scheduling we count the messages that are sent between the negotiators. In a distributed environment it is expected that sending messages will be much more resource consuming than masking or a comparison between bits. Since much of the processing is done in parallel, bandwidth is more important. Remember that the negotiation protocol is performed bit by bit. Note that for n negotiators a broadcast is of complexity n 1. When an all-to-all broadcast is needed it has complexity n(n 1). The scheduling starts with a simple broadcast of the invitation. C n 1. All negotiators (except for the initiator) must announce their position towards the invitation. These broadcasts adds complexity (n 1)(n 1). For getting masked, one bit must visit all negotiators and then be broadcasted: 2(n 1). This happens to each negotiator's bit in a round: C 1). If the number of bits in a schedule is l, after at most l rounds the protocol will end. In the check phase of the scheduling, all negotiators broadcast their result or the fact that no meeting could be scheduled to all others: C 1). Note that only positive results (i.e., a meeting is possible) are broadcasted. If the result is negative, the agents automatically go to the next bit. If the result is still negative after the last bit, it was not possible to schedule a meeting. Therefore at most are sent. For example, for a scheduling window of 3 eight-hour working days, granularity 1 hour (l = 24) and this amounts to at most 1000 messages; for participants in the same conditions, there will be up to 4500 messages sent. 3. USING SECURE DISTRIBUTED COMPUTING 3.1. THE PROBLEM OF SECURE DISTRIBUTED COMPUTING Usually, the problem of Secure Distributed Computing (SDC) is stated as follows. Let f be a publicly known function taking n inputs, and suppose there are n dierent parties, each holding their own private input n). The n parties want to compute the value f(x without leaking any information about their private inputs to the other parties (except of course the information about x i that is implicitly present in the function result). In descriptions of solutions to the Secure Distributed Computing problem, the function f is usually encoded as a boolean circuit, and therefore Secure Distributed Computing is also often referred to as secure circuit evaluation. Over the past two decades, a fairly large variety of solutions (other than the trivial one using a trusted third party) to the problem has been proposed. An overview is given by Franklin [4] and more recently by Cramer [2]. 3.2. HOW TO PERFORM GENERAL The core problem of SDC is that we want to perform computations on hidden data (using encryption, secret sharing or other techniques) without revealing the data. One class of techniques to compute with encrypted data is based on homomorphic probabilistic encryption. An encryption technique is probabilistic if the same cleartext can encrypt to many dierent ciphertexts under the same encryption key. To work with encrypted bits, probabilistic encryption is essential, otherwise only two ciphertexts (the encryption of a zero and the encryption of a one) would be possible, and cryptanalysis would be fairly simple. An encryption technique is homomorphic if it satises at least one equation of the form E(x op operations op and op 0 . A homomorphic encryption scheme allows operations to be performed on encrypted data, and hence is suitable for secure circuit evaluation. In [5], Franklin and Haber present a protocol that evaluates a boolean circuit on data encrypted with such a homomorphic probabilistic encryption scheme. In order to support any number of participants, they use a group oriented encryption scheme, i.e., an encryption scheme that allows anyone to encrypt, but that needs the cooperation of all participants to decrypt. In the group oriented encryption scheme used by Franklin and Haber, a bit b is encrypted for a group of participants S ng as r mod r mod N and q are two primes such that p q mod 4, and r 2R ZN . The public key is given by [N; while K i is the private key of the ith participant. This scheme has some additional properties that are used in the protocol: XOR-Homomorphic. Anyone can compute a joint encryption of the XOR of two jointly encrypted bits. Indeed, if and Blindable. Given an encrypted bit, anyone can create a random ciphertext that decrypts to the same bit. Indeed, if and r 2R ZN , then r mod N; r mod N is a joint encryption of the same bit. Witnessable. Any participant can withdraw from a joint encryption by providing the other participants with a single value. In- it is easy to compute D i First of all, the participants must agree on a value for N and g, choose a secret key K i and broadcast g K i mod N to form the public key. To start the actual protocol, each participant broadcasts a joint encryption of his own input bits. To evaluate an XOR-gate, everyone simply applies the XOR-homomorphism. The encrypted output of a NOT-gate can be found by applying the XOR-homomorphism with a default encryption of a one, e.g. [1; 1 mod N ]. The encryption scheme is not AND-homomorphic, so the evaluation of an AND-gate will be more troublesome. Suppose the encrypted input bits for the AND-gate are ^ E(v). To compute a joint encryption proceed as follows: Each participant i chooses random bits b i and c i and broadcasts Each participant repeatedly applies the XOR-homomorphism to Each participant broadcasts decryption witnesses 3 Everyone can now decrypt ^ repeatedly applying the fact that (a can prove that w (b Each participant is able to compute a joint encryption of w knows b i and c i (he chose them himself) and he received encryptions from the other participants, so he can compute E(b as follows: if b default encryption for a zero will do, e.g. [1; 1]. Otherwise, if b c j is a valid substitution for E(b can be computed in an analogous way. He uses the XOR-homomorphism to combine all these terms, blinds the result and broadcasts this as ^ 4 Each participant combines ^ using the XOR-homomorphism, to form ^ E(w). When all gates in the circuit have been evaluated, every participant has a joint encryption of the output bits. Finally, the participants broadcast decryption witnesses for the output bits to reveal them. 3.3. SECURE MEETING SCHEDULING USING SDC We already showed how to reduce the problem of scheduling a meeting for n secret agendas to a series of logical AND operations on n secret bits. For every time slot in the schedule, each negotiator has one secret input bit: a one if he is available to start the meeting at that time, a zero if he isn't. Because the Secure Distributed Computing protocol we just discussed can only handle binary gates, we implement the n-ary AND operation as a log 2 (n)-depth tree of binary AND-gates. The output bit of the circuit indicates if this slot is an appropriate starting time for the meeting (1) or not (0). 3.4. SECURITY ANALYSIS Franklin and Haber show that their protocol is provably secure against passive adversaries (i.e., adversaries who follow the rules of the protocol, but who try to learn as much information from the communication as possible), given that ElGamal encryption with a composite modulus is secure. This means that under the assumption of passive adversaries, complete privacy of all agendas is guaranteed (except of course for the fact that everybody is available at the time the meeting is scheduled). However, the proof Franklin and Haber give uses a more complicated encryption scheme and they mention the one we used here as an alter- native. To the best of our knowledge, the security of this encryption scheme is still an open problem. The protocol is not provably secure against active adversaries (who can deviate from the protocol). For example, a malicious participant can ip the output of an AND gate by XORing his ^ with the encryption of a one. For this particular application however, the most obvious attacks don't seem to give rise to substantial information leaks. The SDC protocol presented by Chaum, Damgard and van de Graaf in [1] provides provable security against active adversaries at the cost of higher bandwidth requirements. 3.5. COMPLEXITY Let's have a closer look at the message complexity of this protocol. The same public and private keys can be used for every evaluation. This means that the initiator's invitation message can contain N and g (C while g K i can be wrapped together with the message that announces each participant's position towards the invitation (C (n 1)(n 1) messages). The evaluation of a single AND gate consists of four phases, of which the rst three need an all-to-all broadcast (consuming n(n 1) messages while the last one doesn't need any communication. Since the AND gates within one level of the tree can be evaluated in parallel, the evaluation of the entire circuit takes C sages. The broadcast of the encrypted input bits of the circuit and the broadcast of decryption witnesses for the output bit both take another messages. If l slot evaluations are needed before a suitable meeting time is found, the total message complexity is given by C 1 If we consider the same example as we did in the previous section (l = 24), this amounts to 5300 messages for 5 participants and 30330 messages for 10 participants. Before comparing this result to that of the custom-made protocol in the previous section, we should notice that only the number of messages is taken into account, not their size. As jN j should be about 1024 bits to be secure, the messages in the SDC protocol will be larger than the messages in the custom-made protocol. However, since the maximum message length for 10 participants is only 2:5 KB (which easily ts into a single IP packet), we considered the number of transmitted messages more relevant than the number of bits that are strictly needed. It should also be noted that we do not take into account computation or memory overhead for the protocols. The amount of computation and storage needed for the SDC protocol is considerably higher than for the custom-made protocol. 4. USING MOBILE AGENTS In this section, it will be shown how mobile agents can be used to reduce the communication overhead of the two solutions for the agenda scheduling problem. The basic idea is to use mobility to bring agents of the participants closer together. Of course, a mobile agent needs to trust his execution platform but we will show that the trust requirements are less strong than for a classical trusted third party (TTP) solution for the meeting scheduling problem. To compare the trust requirements of the dierent approaches, we use the following simple trust model. We say a participant trusts an execution site if it believes that: (1) the execution site will correctly execute any code sent to it by the participant; (2) the execution site will correctly (i.e., as expected by the participant) handle any data sent to it by the participant. It also implies that the execution site will maintain the privacy of the data or the code if this is expected by the participant. If p trusts E, we denote this as shown in Fig. 3. Figure 3 Notation for \p trusts E" To compare bandwidth requirements (for communication overhead), we make the following simple distinction. High bandwidth is required to execute one of the discussed protocols. Low bandwidth su-ces to transmit data or agent code. Also intermittent connections (e.g. for devices that are sometimes disconnected from the network) are considered low bandwidth. We assume low bandwidth communication is available between any two parties. If high bandwidth communication is possible we denote this as shown in Fig. 4. Figure 4 Notation for high bandwidth connection between E i and Based on these simple models of communication and trust, we compare three options for implementing secure meeting scheduling. 4.1. A TRUSTED THIRD PARTY The rst, perhaps most straightforward option, is to use a globally trusted third party. Every participant sends its agenda to the TTP who will compute an appropriate meeting time and disseminate the result to the participants. Of course, data must be sent to the TTP, through an authenticated and safe channel. This can be accomplished via conventional cryptographic techniques. It is clear that this approach has a very low communication overhead: the data is only sent once to the TTP; later, every participant receives the result of the computation. However, every participant should unconditionally trust the TTP. For the case of 4 participants, the situation is as shown in Fig. 5. Figure 5 Situation with 4 participants and a TTP. It is not clear whether n distrustful participants will easily agree on one single trustworthy third party. This requirement of one single globally trusted execution site is the main disadvantage of this approach. 4.2. CRYPTOGRAPHIC SECURE MEETING The second option is the use of cryptographic techniques (as discussed in previous sections) that make the use of a TTP super uous. The trust requirements are really minimal: every participant only trusts its own execution site. Although this option is very attractive, it should be clear from the previous sections that the communication overhead might be too high to be practically useful in a general networked environment. High bandwidth is required between all of the participants. For the case of 4 participants, the situation can be summarized as shown in Fig. 6. Figure 6 Situation with 4 participants without a TTP. 4.3. USING MOBILE AGENTS Finally, a third solution tries to combine the two previous options: the communication overhead is remedied by introducing semi-trusted execution sites and mobile agents. In this approach, every participant p i sends its representative, agent a i , to a trusted execution site E j . The agent contains a copy of the agenda and is capable of running a secure meeting scheduling protocol. It is allowed that dierent participants send their agents to dierent sites. The only restriction being that the sites should be located closely to each other, i.e., should have high bandwidth communication between them. The amount of long distance communication is moderate: every participant sends its agent to a remote site, and receives the result from its agent. The agents use a cryptographic protocol, which unfortunately involves a high communication overhead. However, since the agents are executing on sites that are near each other, the overhead of the protocol is acceptable. For a situation with 4 participants, we could have the situation as depicted in Fig. 7. Figure 7 Situation with 4 participants using mobile agents communication between the participants is nec- essary, and there is no longer a need for one single trusted execution site. 4.4. CURRENT IMPLEMENTATION WITH \agenTa" is the name of our prototype implementation of a secure meeting scheduling system. Currently it uses the custom-made protocol described in this paper. We have used the Aglets SDK 1.1 beta 3 [7], a mobile agents system development kit which was released to the open source community by its creator, IBM. The SDK contains an agent server, the API needed to write agents in Java (called aglets), examples and documentation. The prototype implementation of agenTa has around 3500 lines of Java code. For the inter-agent communication KQML (Knowledge Query and Manipulation Language) was chosen. KQML was developed at the University of Baltimore Maryland County, and enhanced with security capabilities in [8]. In our implementation, each user's scheduling application is modular, the user interface, the agenda management and the negotiation being performed by distinct intercommunicating aglets. Only the negotiator aglets of all users take advantage of their mobility to gather on a host where they carry out the negotiation protocol by local communication. There are no language limitations for implementing the custom-made protocol. Communication relies on transmitting character strings. There- fore, agents implemented with other agent platforms and in other programming languages can take part in the negotiation, provided the platforms can interoperate. 5. CONCLUSION This paper has shown that there exist several techniques for secure meeting scheduling. Moreover, a trade-o can be made between the level of security that can be obtained, the degree of trust that is required, and the amount of overhead that is caused by the protocol. When a TTP is used, a meeting can be scheduled very e-ciently. The custom-made protocol has more overhead, but does not require trust in a third party. An SDC protocol is more secure than our custom-made protocol, but it is also much less e-cient. Using mobile agents when implementing any protocol can improve the e-ciency, while still avoiding the need for one single trusted entity. Acknowledgements This work was supported in part by the FWO-Vlaanderen project G.0358.99 and by the Concerted Research Action (GOA) Mesto-666. Joris Claessens is funded by a research grant of the Flemish Institute for the Promotion of Industrial Scientic and Technological Research Gregory Neven is an FWO-Vlaanderen Aspirant, and Frank Piessens is an FWO-Vlaanderen Postdoctoral fellow. --R James May --TR Complexity and security of distributed protocols Semi-trusted Hosts and Mobile Agents Multiparty Computations Ensuring Privacy of Each Party''s Input and Correctness of the Result Introduction to Secure Computation Secure Meeting Scheduling with --CTR Marius Calin Silaghi, Meeting Scheduling Guaranteeing n/2-Privacy and Resistant to Statistical Analysis (Applicable to any DisCSP), Proceedings of the 2004 IEEE/WIC/ACM International Conference on Web Intelligence, p.711-715, September 20-24, 2004

meeting scheduling;secure distributed computation;mobile agents

510800

Extended description techniques for security engineering.

There is a strong demand for techniques to aid development and modelling of security critical systems. Based on general security evaluation criteria, we show how to extend the system structure diagrams of the CASE tool AutoFocus (which are related to UML-RT collaboration diagrams) to allow modelling of security critical systems, in particular concerning components and channels. Both high-level and low-level models of systems are supported, and the notion of security patterns is introduced to provide generic solutions for security requirements. We explain our approach on the example of an electronic purse card system.

INTRODUCTION In developing distributed systems-in particular applications that communicate over open networks like the Internet-security is an extremely important issue. Many customers are reluctant to take part in electronic business, confirmed by recent attacks on well-known portal sites or cases of credit card fraud via the Internet. To overcome their reluctance and make E-Commerce and mobile systems a success, these systems need to become considerably more trustworthy. To solve this problem, on the one hand there are highly sophisticated collections of evaluation criteria that security critical systems have to meet, like the ITSEC security evaluation criteria (ITSEC, 1990) or their recent successor, the Common Criteria (CC) (Common Criteria, 1999). The Common Criteria describe security related functionality to be included into a system, like authen- tification, secrecy or auditing, and evaluation assurance levels (EALs) for its This work was supported by the German Ministry of Economics within the FairPay project development. The strictest level is EAL7 (Common Criteria, 1999, part 3, p. 66), where a formal representation of the high-level design is required. On the other hand, research has produced many formal methods to describe and verify properties of security critical systems, ranging from protocol modelling and verification (Burrows et al., 1989; Lowe, 1996; Paulson, 1998; Thayer et al., 1998) to models for access control, like the Bell-LaPadula model (Bell and LaPadula, 1973) or the notion of non-interference (Goguen and Meseguer, 1998). Such formal methods however are rarely used in practice, because they require expert knowledge and are costly and time-consuming. Therefore, an integrated software development process for security critical systems is needed, supported by CASE tools and using graphical description techniques. This reduces cost, as security problems are discovered early in the development process when it is still inexpensive to deal with them, and proof of meeting evaluation criteria is a byproduct of software development. In addition, systems developed along a certain integrated "security engineering process" will be much more trustworthy. In this paper, we describe a first step towards using extended description techniques for security modelling. As a basis for our work, we use the AutoFocus description techniques. The AutoFocus (Huber et al., 1998b; Slo- tosch, 1998; Broy and Slotosch, 1999) system structure diagrams are related to UML-RT collaboration diagrams and describe a system as a set of communicating components. The corresponding CASE tool developed at Munich University of Technology supports user-friendly graphical system design and incorporates simulation, code and test case generation and formal verification of correctness. The main advantage of the use of AutoFocus over a more general description technique as UML is its simplicity. Besides, there exists a clear semantics for the description techniques (for general UML description techniques, defining a formal semantics is still subject of ongoing research). As a start, in our work we focus on security properties of communication channels. We show how certain important security properties of communication channels, such as secrecy and authenticity, can be modelled at different abstraction levels of the system design and explain our ideas on the transition between these levels, using generic security patterns. We give definitions of the meanings of our extended description techniques, based on the AutoFocus semantics. See also (J-rjens, 2001) for first work on integrating access control models into UML description techniques, and (Lotz, 2000) for formal definitions of security properties using the Focus method. This paper is structured as follows. In Section 2, we give a short introduction to AutoFocus. In Section 3, we present the extensions of AutoFocus system structure diagrams to model security properties of channels. The usage of these techniques is demonstrated in Section 4, with the help of an example Extended Description Techniques for Security Engineering 3 model of an electronic purse system. We conclude in Section 5 with a summary and indicate further work. 2. AutoFocus AutoFocus/Quest (Huber et al., 1998a; Slotosch, 1998; Philipps and Slo- tosch, 1999) is a CASE tool recently developed at Munich University of Technology with the goal to combine user-friendly graphical system design and support of simulation, code generation and formal verification of correctness. AutoFocus supports system specification in a hierarchical, view-oriented way, an approach that is well established and facilitates its use in an industrial environment. However, it is also based on the well-founded formal background Focus (Broy et al., 1992), and a fairly elementary underlying concept: communicating extended Mealy machines. System specifications in AutoFocus make use of the following views: System Structure Diagrams (SSDs) are similar to collaboration diagrams and describe structure and interfaces of a system. In the SSD view, the system consists of a number of communicating components, which have input and output ports to allow for sending and receiving messages of a particular data type. The ports can be connected via chan- nels, making it possible for the components to exchange data. SSDs can be hierarchical, i.e. a component belonging to an SSD can have a sub-structure that is defined by an SSD itself. Besides, the components in an SSD can be associated with local variables. Data Type Definitions (DTDs) define the data types used in the model, with the functional language Quest (Philipps and Slotosch, 1999). In addition to basic types as integer, user-defined hierarchical data types are offered that are very similar to those used in functional programming languages like Gofer (Jones, 1993) or Haskell (Thompson, 1999). State Transition Diagrams (STDs) represent extended finite automata and are used to describe the behaviour of a component in an SSD. Transitions consist of input patterns on the channels, preconditions, output patterns, and actions setting local variables when the transition is exe- cuted. As the main focus of this paper is extending SSDs, we will not describe STDs in detail at this place. Extended Event Traces (EETs) finally make it possible to describe exemplary system runs, similar to MSCs (ITU, 1996). The Quest extensions (Slotosch, 1998) to AutoFocus offer various connections of AutoFocus to programming languages and formal verification tools, such as Java code generation, model checking using SMV, bounded model checking and test case generation (Wimmel et al., 2000). 3. EXTENDING SYSTEM STRUCTURE DIAGRAMS The main difference between security critical systems and traditional systems is the consideration of attacks. A potential third party could try to overhear and manipulate security critical data. To decrease the risk of attacks to a minimum, special security functionalities are used. For example, encryption is a common principle to inhibit overhearing of the communication between two agents. It is state of the art to use graphical description techniques for system spec- ification. For the specification of security critical systems, we need special description techniques to deal with the particularities of those systems. We extend the AutoFocus description techniques to fulfill the needs of security engineering. In this paper the extensions of the AutoFocus SSDs, mentioned in Section 2, are described. The extensions of the structure diagrams allow the definition of security requirements. Furthermore it is possible to specify the usage of security functionality to fulfill the defined requirements. We use special tags for the security extensions to the SSDs. These security tags are assigned to particular diagram parts and have a defined semantics. The following sections describe the different security extensions made to the SSDs. 3.1. SECURITY CRITICAL SYSTEM PARTS To model and evaluate security aspects of distributed systems, it is always necessary to define its security critical parts. The identification of security critical parts should be done very early within the system development process. This task is typically part of the analysis phase. In the Common Criteria, the security critical parts of a system together form the Target Of Evaluation (TOE). The following definition will make our notion of security criticality more precise. critical parts of a distributed system we mean parts that deal with data or information that has to be protected against unauthorized operations (e.g. disclosure, manipulation, prevention of access etc. In particular, security critical system parts are connected with security requirements such as secrecy, authentication or auditing, and according to required strictness of evaluation might be subject to formal modelling. We want to make the distinction visible within the graphical system descrip- tion. Therefore we annotate security critical system parts with the security tag -critical-. Both components and channels can be tagged. To mark non critical system parts, -noncritical- is used. A system part without a -critical- or Extended Description Techniques for Security Engineering 5 A data Figure Critical System Parts -noncritical- tag is non critical by default. Figure 1 shows an SSD. It consists of two security critical components and one security critical channel. 3.2. PUBLIC CHANNELS The special aspect of security critical systems is the possibility of attacks. Hostile subjects can manipulate the system by overhearing and manipulating the communication between the subsystems. To distinguish private communication channels of a system from public channels, we use the two security tags -private- and -public-. A -private- channel is a channel a hostile party has no access to. The hostile subject can neither overhear nor manipulate the communication that takes place via the channel. Vice versa a hostile party can overhear and manipulate all communications of a -public- channel. If neither -private- nor -public- are used, we assume by default that the communication channel is not publicly accessible. channel has the same semantics as a normal channel within AutoFocus, i.e. it is a dedicated connection from one component to another. Channel). The semantics of a -public- channel without secrecy and authenticity introduced in Section 3.6, is defined by the SSDs shown in Figure 2. Using a -public- channel (Figure 2(a)) is an abbreviation for having an intruder included in the model that has access to the channel Figure 2(b)). The behaviour of the intruder is defined by the threat model-for example, the intruder usually can overhear, intercept or redirect messages. It is possible to model this behaviour in a flexible way using using AutoFocus State Transition Diagrams (STDs) (Wimmel and Wisspeintner, 2000). The identification of private and public channels should be done during the analysis phase of the system development process, right after the identification of the security critical parts. Every security critical channel must be analyzed with regard to accessibility-for the other channels, this is optional. The result of this analysis is documented by using the tags for private and public channels. data (a) -public- Channel A Intruder data B data (b) Defined as an Intruder between the Two Agents Figure 2 Semantics of a Public Channel 3.3. REPLACEABLE COMPONENTS Conventional system structure diagrams always show a system in normal operation, e.g. an ordinary electronic purse card component with the specified functionality communicating with the point-of-sale component. In addition to manipulation of the communication link (man-in-the-middle attack), another attack scenario is imaginable: the attacker could try to communicate with the rest of the system in place of the purse card. In this case, there is no ordinary purse card in the system, but a faked one (that in particular usually does not contain private keys of ordinary cards, except if they leaked). We mark components that can be replaced by faked ones by the attacker with the -replace- tag, and components that can not be replaced with the -nonre- place- tag. If neither -replace- nor -nonreplace- is used for a component, the component is non replaceable by default. Definition 4 (Replaceable Component). Figure 3 shows the semantics of a -replace- component. Using a replaceable component (Figure 3(a)) is an abbreviation for specifying two different system scenarios. The first scenario describes the structure of the system with the specified component A (Figure 3(b)). In the second scenario (Figure 3(c)) the attacker exchanges the component by another component A'. A' has the same component signature like A but has an arbitrary behaviour that can be defined by the threat model. In the development process, replaceable components should be identified during the analysis phase together with the identification of private and public Extended Description Techniques for Security Engineering 7 A data (a) -replace- Component A data Component not Exchanged A' data (c) Scenario 2: Component Exchanged Figure 3 Semantics of a Replaceable Component channels. It is only necessary to analyze security critical components with regard to replaceability. 3.4. ENCAPSULATED COMPONENTS An encapsulated component is a component that only consists of not publicly accessible subcomponents. In this way an attacker has no possibility to manipulate or exchange the subsystems of this component. Furthermore, the communication within the component cannot be overheard. The security tag -node- is used to mark a component as an encapsulated one. The identification of encapsulated components is done together with the identification of private and public channels and replaceable components. (Consistency Condition of -node-). A -node- component only consists of -private- channels and -nonreplace- components. One example for an encapsulated component is an automated teller machine (ATM). An ATM is encapsulated in a way that unauthorized persons are not able to manipulate system parts. Overhearing the internal communication is also not possible. 3.5. ACTOR COMPONENTS Most systems interact with their system environments. It is often desired to illustrate the system environment in the graphical system design. Components that are not part of the system are called actors. We point out actors by using the tag -actor-. A typical example for an actor is a system user. The system user interacts with the system without being part of it. Actor components can never be marked with the -critical- tag. An actor is not part of the system and therefore there is no need to analyze the actor itself with respect to security aspects. But an actor can interact with our system in a way that affects our security requirements. To visualize these critical interactions, channels between actors and the system components can be annotated with the -critical- tag. 3.6. SECRECY AND AUTHENTICITY The most important security properties of communication channels-in addition to integrity, availability and non-repudiation-are authenticity and secrecy. For this purpose, we will introduce tags -secret- and -auth- for channels in the SSDs. The security properties of channels are identified in the high-level design phase, taking place after the activities of the analysis phase. It is only necessary to specify security properties for security critical, public channels. There are many possible definitions for authenticity and secrecy in the security literature (see (Gollmann, 1996)). In the following, we give a definition based on our model. During the high-level design phase, we assume that the defined requirements of secrecy and authenticity are fulfilled automatically, if the corresponding tags appear on the channels. Consequently these requirements restrict the possibilities of an attacker. In the low-level design, the validity of these requirements has to be ensured by proper mechanisms. Definition 6 (Secret Channel in High-Level Design). A message sent on a -secret- channel can only be read by its specified destination component. Therefore, we can assume in high-level design that a -secret- and -public- channel can not be read by the intruder. But the intruder could write something on it. Figure 4 shows the semantics of a secret and public channel. Definition 7 (Authentic Channel in High-Level Design). A message received on an -auth- channel can only come from its specified source component. Therefore, an -auth- and -public- channel can not be written Extended Description Techniques for Security Engineering 9 data (a) -secret- and -public- Channel A Intruder data data (b) Intruder can Write Data The "S" circle represents a switch component that distributes all incoming data to all outgoing channels. Figure 4 Semantics of a Secret and Public Channel by the intruder. But the intruder could possibly read data from it. Figure 5 illustrates the semantics of an authentic channel. There are some relations between our notion of security critical and secrecy and authenticity. A security critical channel references to data that should be protected against attackers (see Definition 1). Secrecy and authenticity are security properties. These security properties defines concrete requirements concerning the protection of data against attackers. Definition 8 (Consistency Condition of -secret- and -auth-). If a channel is marked to be security critical and the communication is visible for an at- tacker, the data sent via the channel must be protected in a suitable manner. In this case, during the high-level design phase the protection of data must be ensured by a security property. A -critical- and -public- channel must be -secret- or -auth- or both. 3.7. INTEGRITY We assume that -secret- or -auth- channels also provide message integrity, i.e. a message received on an -auth- channel is guaranteed not to have been modified. In future, the integrity property could also be modelled separately. data (a) -auth- and -public- Channel A Intruder data data (b) Intruder can Read Data Figure 5 Semantics of an Authentic and Public Channel 3.8. CHANNEL REQUIREMENTS IN LOW-LEVEL In the low-level design phase, taking place after the high-level design phase, the system specification is refined. In this phase, security functionalities can be used to ensure the security properties and requirements. We can define the usage of symmetric encryption, asymmetric encryption and corresponding signatures to realize the requirements of secrecy and au- thentification. The security tag -sym- marks a channel. It defines that the realization of the channel must use a symmetric encryption algorithm to ensure the secrecy requirements. Furthermore the -asym- tag is used to define the usage of an asymmetric encryption algorithm. It is also possible to specify the encryption algorithm that should be used to guarantee secrecy. The security tag -encalg [Parameters]- can be used together with a channel to specify the usage of a specific encryption algorithm. We can specify additional parameters of the algorithm. For example it is possible to define the key length of the encryption keys. Authenticity can be realized by using specific authentification protocols. The tag -authprotocol [Parameters]- defines a specific authentification protocol to be used. Extended Description Techniques for Security Engineering 11 By choosing a specific encryption algorithm for realizing secrecy, we perform a refinement step. Special encryption drivers are introduced to perform the encryption and decryption tasks. The data that is sent between the two encryption drivers is encrypted. To realize a specific authentification protocol, special protocol drivers are introduced. Furthermore, additional channels between the protocol drivers are needed to allow bidirectional communication. After these refinements, the statements on access of the intruder to the channel in Definition 6 and Definition 7 change for -public- channels: the intruder now does have read and write access to the channel. The encryption mechanism on the channel must ensure that the intruder cannot manipulate the channel in an improper way, so that the security requirements stated by -secret- and -auth- are still fulfilled. Definition 9 (Secret and Authentic Channels in Low-Level Design). In low-level design, -public- channels implementing secret or authentic communication have to be modelled by including appropriate security functionality. A convenient way to do this is using security patterns. Security patterns are generic solutions for common security problems. Figure 6 shows such a pattern for the simple case of encryption (guaranteeing secrecy). The communicating components now include protocol drivers for encryption and decryption of the messages, so the original channel is replaced by an encrypted one. This encryption pattern could be extended by including protocol drivers for key agreement, a public key server or a whole public key infrastructure. Other security patterns, for instance providing auditing functionality, are possible. 4. EXAMPLE-ELECTRONIC PURSE TRANSACTION SYSTEM In the previous section, we have introduced extensions for SSDs to deal with security requirements. Now, we use the extended SSDs to model an example system from the area of E-Commerce, an electronic purse card system. The given specification conforms with the Common Electronic Purse Specification (CEPS), an international standard for an electronic purse environment (CEP- SCO, 2000). Figure 7 shows the complete system structure of the electronic purse system consisting of several sub-components. The system user possesses an electronic purse card. The card has some amount of money stored on it and the user can pay with the card at a dealer using a point-of-sale (POS) terminal. The electronic purse card is involved in several security requirements. For example, together with other components the electronic purse card must ensure that an attacker can not steal any money from the whole transaction system. Therefore the card is a security critical component (-critical-). Furthermore, the purse card should be realized by a single A <<encalg [Parameter]>> data (a) -public- Channel using a Specific Encryption Algorithm A' enc(data) Encrypt Decrypt data data Intruder enc(data) (b) Intruder can Read and Write Encrypted Data Figure 6 The Encryption Security Pattern integrated chip. It is an encapsulated component (-node-) and consequently we assume an attacker has no possibility to overhear or manipulate the communication within the electronic purse card. An attacker could try to produce a faked card and he could replace the valid card by the faked one. The security tag -replace- of the component ElectronicPurse visualizes this possibility. The electronic purse card can communicate with the other components over its interface. It offers ports to read the balance, decrease the balance and set the balance to a specific value. The getBal operation of the card is used by the POS, a card loading terminal at the bank of the card owner (IssuerBank) and by the CardReader component. The operation of reading the balance is not security critical because everybody who possesses an electronic purse is allowed to read the stored balance of the card. Consequently the channels getBal and returnBal do not have the -critical- tag. But an intruder could build a special adapter to overhear all communication between the electronic purse card and the other components. To express this possibility the communication channels getBal, returnBal, decreaseBal and setBal are marked with the -public- tag. If the user wants to load money onto his card, he must go to his bank and use a special card terminal. The channel setBal is used to transfer money from his bank account to the card. This channel is security critical, because it affects the security requirement about an attacker stealing Extended Description Techniques for Security Engineering 13 CardReader ElectronicPurse POS decreaseBal(int) returnBal(int) User pay IssuerBank AcquirerBank getBal loadCard(int) getBal getBal getBal returnBal(int) returnBal(int) returnBal(int) returnBal(int) getBal Figure 7 Electronic Purse Card System money. We use the property of authentification (-auth-) to ensure that the card and the card terminal are valid. The card reader is a simple device for the user to check the amount of money that is stored on the card. An attacker can exchange the card reader by another component (-replace-). On the other hand the card reader component is encapsulated (-node-). No security requirements are defined for this component and therefore the component is not security critical. The POS component is a card terminal located at a dealer. The user can insert the electronic purse card into the terminal in order to pay for goods. The POS is directly involved in the money transaction process. Therefore this component is security critical. The POS is encapsulated to prevent manipulations within the unit. But a malicious dealer has the possibility to exchange the POS terminal by a faked unit. The POS component can instruct the electronic purse card to decrease its balance by a given amount. The operation decreaseBal is part of the money transaction process and the communication to perform the operation is security critical. To comply with our security requirement that an intruder cannot steal any money, we must ensure that only a valid POS is able to decrease the amount of money on a money card. The POS must authenticate itself in a way that the card is sure to communicate with a valid POS. This fact is visualized in the SSD by the -auth- tag annotated to the decreaseBal channel. The dealer can submit the earned money to his bank (AcquirerBank) using the cashCredit channel. This channel is security critical. The communication medium is a standard telephone line and the potential attacker has possibilities to overhear and manipulate the communication. This fact is expressed using the -public- tag on this channel. To ensure that an attacker can not overhear or manipulate the transferred data, the -secret- and -auth- tags are used. The channel cashCredit between the two banks is used to transfer money from one bank to the other. The communication takes place via an open network (-public- channel). We must ensure secrecy and authenticity for this channel to protect the transaction data. Both bank components are security critical, but we do not see a risk that a potential attacker can act as a faked bank component. Thus the -replace- tag is not used for both banks. Finally let us have a look at the User. The user is not part of our system. Therefore the -actor- tag is used. The user can initiate the paying process at a POS. The initiation of the paying process is not security critical (it just corresponds to inserting the card, whereas the amount of money to be withdrawn is negotiated outside of the system). The critical part of the paying process takes place between the money card and the POS. Furthermore, the user can load the card with some amount of money. During this action, an amount of money is transferred from his bank account onto the card. This operation is security critical because an attacker could try to transfer money from a foreign account Extended Description Techniques for Security Engineering 15 to his own electronic purse card. Thus we need some kind of authentification to perform this operation, e.g. the user must enter a PIN code before the money transaction is performed. 5. CONCLUSIONS AND FURTHER WORK This work is only the beginning of an effort to extend graphical description techniques for distributed systems with security aspects to support methodical development of security critical systems. We used the CASE tool AutoFo- cus, the description techniques of which are related to UML-RT, for its simplicity and clear semantics and the possibility to give our security extensions a straightforward and unambiguous meaning. We showed how to extend AutoFocus system structure diagrams by security tags, both for high-level and low-level design. The transition from high-level to low-level design is aided by the possibility to use security patterns. The description techniques were illustrated with the help of an example from the field of E-Commerce, an electronic purse card system. We focused on the consideration of channels and system structure. In fu- ture, additional security properties such as integrity and availability are to be included. The specification of channels and components in low-level design needs to be detailed, using classifications as pointed out in (Eckert, 1998). Be- sides, it seems very promising to further examine security patterns providing generic architectures for specific security functionality and evaluate their use within the development process. The refinement of security requirements and security functionalities together with its influence on correctness verification is also part of our research activities. Also, state transition diagrams (STDs) specifying the behaviour of a component can be extended in a similar way with security properties. For this purpose, it suggests itself to classify the data received and sent on the ports and to use models such as Bell-LaPadula or non-interference-similar as it is done in (J-rjens, 2001) for Statechart diagrams. When the behaviour of components is specified, formal proofs can be carried out (by hand or automatically via model checking) that the specified security properties are fulfilled. EETs (extended event traces) can also be enriched by cryptographic primitives and security properties, and thus be used to specify and verify security functionality of a component. Examining software development of security critical systems with the help of AutoFocus EETs (using protocols from the CEPS purse card system as a case study) is subject of ongoing work. Acknowledgments Helpful comments and encouragement from Oscar Slotosch are gratefully acknowledged. --R Secure computer systems: Mathematical foundations and model. The Design of Distributed Systems-An Introduction to FOCUS Enriching the Software Development Process by Formal Methods. A logic of authentication. Common criteria for information technology security evaluation version 2.1. Security Policy and Security Models. What do We Mean by Entity Authentication? Towards Development of Secure Systems using UML. Formally Defining Security Properties with Relations on Streams. Breaking and fixing the Needham-Schroeder Public-Key Protocol using FDR The inductive approach to verifying cryptographic pro- tocols The Quest for Correct Systems: Model Checking of Diagramms and Datatypes. Quest: Overview over the Project. Strand Spaces: Why is a security protocol correct? Haskell: TheCraft of Functional Programming. Specification Based Test Sequence Generation with Propositional Logic. The Needham-Schroeder Protocol- an AutoFocus Case Study --TR The inductive approach to verifying cryptographic protocols The Haskell Towards Development of Secure Systems Using UMLsec Breaking and Fixing the Needham-Schroeder Public-Key Protocol Using FDR Enriching the Software Development Process by Formal Methods QUEST The Quest for Correct Systems Traffic Lights - An AutoFocus Case Study Tool Supported Specification and Simulation of Distributed Systems What do we mean by entity authentication? --CTR Monika Vetterling , Guido Wimmel , Alexander Wisspeintner, Secure systems development based on the common criteria: the PalME project, ACM SIGSOFT Software Engineering Notes, v.27 n.6, November 2002 Monika Vetterling , Guido Wimmel , Alexander Wisspeintner, Secure systems development based on the common criteria: the PalME project, Proceedings of the 10th ACM SIGSOFT symposium on Foundations of software engineering, November 18-22, 2002, Charleston, South Carolina, USA Jan Jrjens, Modelling audit security for Smart-Card payment schemes with UML-SEC, Proceedings of the 16th international conference on Information security: Trusted information: the new decade challenge, June 11-13, 2001, Paris, France Martin Deubler , Johannes Grnbauer , Jan Jrjens , Guido Wimmel, Sound development of secure service-based systems, Proceedings of the 2nd international conference on Service oriented computing, November 15-19, 2004, New York, NY, USA Folker den Braber , Theo Dimitrakos , Bjrn Axel Gran , Mass Soldal Lund , Ketil Stlen , Jan yvind Aagedal, The CORAS methodology: model-based risk assessment using UML and UP, UML and the unified process, Idea Group Publishing, Hershey, PA,

security patterns;UML-RT;security properties;formal methods;security engineering;autofocus;requirements engineering;graphical description techniques;design patterns;CASE;software engineering

510866

Write barrier removal by static analysis.

We present a set of static analyses for removing write barriers in programs that use generational garbage collection. To our knowledge, these are the first analyses for this purpose. Our Intraprocedural analysis uses a flow-sensitive pointer analysis to locate variables that must point to the most recently allocated object, then eliminates write barriers on stores to objects accessed via one of these variables. The Callee Type Extension incorporates information about the types of objects allocated in invoked methods, while the Caller Context Extension incorporates information about the most recently allocated object at call sites that invoke the currently analyzed method. Results from our implemented system show that our Full Interprocedural analysis, which incorporates both extensions, can eliminate the majority of the write barriers in most of the programs in our benchmark set, producing modest performance improvements of up to 7% of the overall execution time. Moreover, by dynamically instrumenting the executable, we are able to show that for all but two of our nine benchmark programs, our analysis is close to optimal in the sense that it eliminates the write barriers for almost all store instructions observed not to create a reference from an older object to a younger object.

INTRODUCTION Generational garbage collectors have become the memory management alternative of choice for many safe languages. The basic idea behind generational collection is to segregate objects into dierent generations based on their age. Gen- This research was supported in part by an NSF Fellowship, DARPA Contract F33615-00-C-1692, NSF Grant CCR00- 86154, and NSF Grant CCR00-63513. erations containing recently allocated objects are typically collected more frequently than older generations; as young objects age by surviving collections, the collector promotes them into older generations. Generational collectors therefore work well for programs that allocate many short-lived objects and some long-lived objects | promoting long-lived objects into older generations enables the garbage collector to quickly scan the objects in younger generations. Before it scans a generation, the collector must locate all references into that generation from older generations. Write barriers are the standard way to locate these references | at every instruction that stores a heap reference into an object, the compiler inserts code that updates an intergenerational reference data structure. This data structure enables the garbage collector to nd all references from objects in older generations to objects in younger generations and use these references as roots during the collections of younger gen- erations. The write barrier overhead has traditionally been accepted as part of the cost of using a generational collector. This paper presents a set of new program analyses that enables the compiler to statically eliminate write barriers for instructions that never create a reference from an object in an older generation to an object in a younger generation. The basic idea is to use pointer analysis to locate store instructions that always write the most recently allocated ob- ject. Because this object is the youngest object, such a store instruction will never create a reference from an older object to a younger object. The write barrier for this instruction is therefore super uous and the transformation eliminates it. 1 We have implemented several analyses that use this basic approach to write barrier elimination: Intraprocedural Analysis: This analysis analyzes each method separately from all other methods. It uses a ow-sensitive, intraprocedural pointer analysis to nd variables that must refer to the most recently allocated object. At method entry, the analysis conservatively assumes that no variable points to the most recently allocated object. After each method invoca- This analysis assumes the most recently allocated object is always allocated in the youngest generation. In some cases it may be desirable to allocate large objects in older gener- ations. A straightforward extension of our analysis would statically identify objects that might be allocated in older generations and suppress write barrier elimination for stores that write these objects. tion site, the analysis also conservatively assumes that no variable refers to the most recently allocated object. Callee Type Extension: This extension augments the Intraprocedural analysis with information from invoked methods. It nds variables that refer to the object most recently allocated within the currently analyzed method (the method-youngest object). It also tracks the types of objects allocated by each invoked method. For each program point, it extracts a pair is the set of variables that refer to the method-youngest object and T is a set of the types of objects potentially allocated by methods invoked since the method-youngest object was allocated. If a store instruction writes a reference to an object into the method-youngest object, and C is not a super-type of any type in T, the transformation can eliminate the write barrier | the method-youngest object is younger than the object o. Caller Context Extension: This extension augments the Intraprocedural analysis with information about the points-to information at call sites that may invoke the currently analyzed method. If the receiver object of the currently analyzed method is the most recently allocated object at all possible call sites, the algorithm can assume that the this variable refers to the most recently allocated object at the entry point of the currently analyzed method. Full Interprocedural This analysis combines the Callee Type Extension and the Caller Context Extension to obtain an analysis that uses both type information from callees and points-to information from callers. Our experimental results show that, for our set of benchmark programs, the Full Interprocedural analysis is often able to eliminate a substantial number of write barriers, producing modest overall performance improvements of up to a 7% reduction in the total execution time. Moreover, by instrumenting the benchmarks to dynamically observe the age of the source and target objects at each store instruction, we are able to show that in all but two of our nine bench- marks, the analysis is able to eliminate the write barriers at virtually all of the store instructions that do not create a reference from an older object to a younger object during the execution on the default input from the benchmark suite. In other words, the analysis is basically optimal for these benchmarks. Finally, this optimality requires information from both the calling context and the called methods. Neither the Callee Type Extension nor the Caller Context Extension by itself is able to eliminate a signicant number of write barriers. This paper provides the following contributions: Write Barrier Removal: It identies write barrier removal as an eective means of improving the performance of programs that use generational garbage collection. Analysis Algorithms: It presents several new static analysis algorithms that enable the compiler to automatically remove unnecessary write barriers. To the class TreeNode { TreeNode left; TreeNode right; Integer depth; static public void main(String[] arg) { void linkDepth(int d) { new void linkTree(TreeNode l, TreeNode r, int d) { 1: linkDepth(d); 2: right = static TreeNode buildTree(int d) { if (d <= new TreeNode(); return t; Figure 1: Binary Tree Example best of our knowledge, these are the rst algorithms to use program analysis to eliminate write barriers. Experimental Results: It presents a complete set of experimental results that characterize eectiveness of the analyses on a set of benchmark programs. These results show that the Full Interprocedural analysis is able to remove the majority of the write barriers for most of the programs in our benchmark suite, producing modest performance benets of up to a 7% reduction in the total execution time. The remainder of this paper is structured as follows. Section presents an example that illustrates how the algorithm works and how it can be used to remove unnecessary write barriers. Section 3 presents the analysis algorithms. We discuss experimental results in Section 4, related work in Section 5, and conclude in Section 6. 2. AN EXAMPLE Figure presents a binary tree construction example. In addition to the left and right elds, which implement the tree structure, each tree node also has a depth eld that refers to an Integer object containing the depth of the subtree rooted at that node. In this example, the main method invokes the buildTree method, which calls itself recursively to create the left and right subtrees before creating the root TreeNode. The linkTree method links the left and right subtrees into the the current node, and invokes the linkDepth method to allocate the Integer object that holds the depth and link this new object into the tree. We focus on the two store instructions generated from lines 1 and 2 in Figure these store instructions link the left and right subtrees into the receiver of the linkTree method. In the absence of any information about the relative ages of the three objects involved (the left tree node, the right tree node, and the receiver), the implementation must conservatively generate write barriers at each store operation. But in this particular program, these write barriers are super- uous: the receiver object is always younger than the left and right tree nodes. This program is an example of a common pattern in many object-oriented programs in which the program allocates a new object, then immediately invokes a method to initialize the object. Write barriers are often unnecessary for these assignments because the object being initialized is often the most recently allocated object. 2 In our example, the analysis allows the compiler to omit the unnecessary write barriers as follows. The analysis rst determines that, at all call sites that invoke the linkTree method, the receiver object of linkTree is the most recently allocated object. It then analyzes the linkTree method with this information. Since no allocations occur between the entry point of the linkTree method and store instruction at line 1, the receiver object remains the most recently allocated object, so the write barrier at this store instruction can be safely removed. In between lines 1 and 2, the linkTree method invokes the linkDepth method, which allocates a new Integer object to hold the depth. After the call to linkDepth, the receiver object is no longer the most recently allocated object. But during the analysis of the linkTree method, the algorithm tracks the types of the objects that each invoked method may create. At line 2, the analysis records the fact that the receiver referred to the most recently allocated object when the linkTree method was invoked, that the linkTree method itself has allocated no new objects so far, and that the linkDepth method called by the linkTree method allocates only Integer objects. The store instruction from line creates a reference from the receiver object to a TreeNode object. Because TreeNode is not a superclass of Integer, the referred TreeNode object must have existed when the linkTree method started its execution. Because the receiver was the most recently allocated object at that point, the store instruction at line 2 creates a reference to an object that is at least as old as the receiver. The write barrier at line 2 is therefore super uous and can be safely removed. 3. THE ANALYSIS Our analysis has the following structure: it consists of a purely intraprocedural framework, and two interprocedural extensions. The rst extension, which we call the Callee Type Extension, incorporates information about called meth- ods. The second extension, which we call the Caller Context Extension, incorporates information about the calling context. With these two extensions, which can be applied separately or in combination, we have a set of four analyses, which are given in Table 2. 2 Note that even for the common case of constructors that initialize a recently allocated object, the receiver of the constructor may not be the most recently allocated object | object allocation and initialization are separate operations in Java bytecode, and other object allocations may occur between when an object is allocated and when it is initialized With Callee With Caller Type Extension Context Extension Callee Only Yes No Caller Only No Full Interprocedural Yes Yes Figure 2: The Four Analyses The remainder of this section is structured as follows. We present the analysis features in Section 3.1 and the program representation in Section 3.2. In Section 3.3 we present the Intraprocedural analysis. We present the Callee Only analysis in Section 3.4, and the Caller Only analysis in Section 3.5. In Section 3.6, we present the Full Interprocedural analysis. Finally, in Section 3.7, we describe how the analysis results are used to remove unnecessary write barriers. 3.1 Analysis features Our analyses are ow-sensitive, forward data ow analyses that compute must points-to information at each progam point. The precise nature of the computed data ow facts depends on the analysis. In general, the analyses work with a set of variables V that must point to the object most recently allocated by the current method, and optionally a set of types T of objects allocated by invoked methods. 3.2 Program Representation In the rest of this paper, we use v, v0 , v1 to denote local variables, m, m0 , m1 to denote methods, and C, C0 , to denote types. The statements that are relevant to our analyses are as follows: the object allocation statement the move statement and the call statement In the given form, the rst parameter to the call, v1 , points to the receiver object if the method m is an instance method. 3 We assume that a preceding stage of the compiler has constructed a control ow graph for each method and a call graph for the entire program. We use entry m to denote the entry point of the method m. For each statement st in the program, pred(st) is the set of predecessors of st in the control ow graph. We use st to denote the program point immediately before st, and st to denote the program point immediately after st. For each such program point p (of the form st or st), we denote A(p) to be the information computed by the analysis for that program point. We use Callers(m) to denote the set of call sites that may invoke the method m. 3.3 The Intraprocedural Analysis The simplest of our set of analyses is the Intraprocedural analysis. It is a ow-sensitive, forward data ow analysis that generates, for each program point, the set of variables that must point to the most recently allocated object, known as the m-object. We call a variable that points to the m-object an m-variable. The property lattice is P(Var) (the powerset of the set of 3 In Java, an instance method is the same as a non-static method. any other assignment to v V n fvg other statements V Figure 3: Transfer Functions for the Intraprocedural Analysis variables Var) with normal set inclusion as the ordering re- lation, where Var is the set of all program variables. The operator used to combine data ow facts at control- ow merge points is the usual set intersection operator: u \. Figure 3 presents the transfer functions for the analysis. In the case of an allocation statement C," the new object clearly becomes the most recently allocated object. Since v is the only variable pointing to this newly-allocated object, the transfer function returns the singleton fvg. For a call statement the transfer function returns ;, since in the absence of any interprocedural information, the analysis must conservatively assume that the called method may allocate any number or type of objects. For a move the source of the move, v2 , is an m-variable, the destination of the move, v1 , becomes an m-variable. The transfer function therefore returns the union of the current set of m-variables with the singleton fvg. For a move statement where the source of the move is not an m-variable, or for any other type of assignment (i.e., a load from a eld or a static eld), the destination of the move may not be an m-variable after the move. The transfer function therefore returns the current set of m-variables less the destination variable. Other statements leave the set of m-variables unchanged. The analysis result satises the following equations: st entry m ufA(st st The rst equation states that the analysis result at the program point immediately before st is ; if st is the entry point of the method; otherwise, the result is the meet of the analysis results for the program points immediately after the predecessors of st. As we want to compute the set of variables that denitely point to the most recently allocated object, we use the meet operator (set intersection). The second equation states that the analysis result at the program point immediately after st is obtained from applying the transfer function for st to the analysis result at the program point immediately before st. The analysis starts with the set of m-variables initialized to the empty set for the entry point of method and to the full set of variables Var (the top element of our property lattice) for all the other program points, and uses an iterative algorithm to compute the greatest xed point of the aforementioned equations under subset inclusion. 3.4 The Callee Only Analysis The Callee Type Extension builds upon the framework of the Intraprocedural analysis, and extends it by using information about the types of objects allocated by invoked methods. This extension stems from the following observation. The Intraprocedural analysis loses all information at call sites because it must conservatively assume that the invoked method may allocate any number or type of objects. The Callee Type Extension allows us to retain information across a call by computing summary information about the types of the objects that the invoked methods may allocate. To do so, the Callee Type Extension relaxes the notion of the m-object. In the Intraprocedural analysis, the m-object is simply the most recently allocated object. In the Callee Type Extension, the m-object is the object most recently allocated by any statement in the currently analyzed method. The analysis then computes, for each program point, a tuple containing a variable set V and a type set T. The variable set V contains the variables that point to the m-object (the m-variables), and the type set T contains the types of objects that may have been allocated by methods invoked since the allocation of the m-object. The property lattice is now where Var is the set of all program variables and Types is the set of all types used by the program. The ordering relation on this lattice is and the corresponding meet operator is The top element is This lattice is in fact the cartesian product of the lattices and These two lattices have dierent ordering relations because their elements have different meanings: must information, while may information. Figure 4 presents the transfer functions for the Callee Only analysis. Except for call statements, the transfer functions treat the variable set component of the tuple in the same way as in the Intraprocedural analysis. For call statements of unanalyzable methods (for example, native methods), the transfer function produces the (very) conservative approximation h;; ;i. For other call statements, the transfer function returns the variable set unchanged, but adds to the type set the types of objects that may be allocated during the call. Due to dynamic dispatch, the method invoked at st may be one of a set of methods, which we obtain from the call graph using the auxiliary function Callees(st). To determine the types of objects allocated by any particular method, we use another auxiliary function Allocated Types. The set of types that may be allocated during the call at st is simply the union of the result of the Allocated Types function applied to each component of the set Callees(st). The only other transfer function that modies the type set is the st Allocated Types(m)) any other assignment to v hV n fvg; other statements hV; Figure 4: Transfer Functions for the Callee Only Analysis allocation statement, which returns ; as the second component of the tuple. The Callees function can be obtained directly from the program call graph, while the Allocated Types function can be e-ciently computed using a simple ow-insensitive analysis that determines the least xed point for the equation given in Figure 5. The analysis solves the data ow equations in Figure 4 using a standard work list algorithm. It starts with the entry point of the method initialized to h;; ;i and all other program points initialized to the top element hVar; ;i. It computes the greatest xed point of the equations as the solution. 3.5 The Caller Only Analysis The Caller Context Extension stems from the observation that the Intraprocedural analysis has no information about the m-object at the entry point of the method. The Caller Context Extension augments this analysis to determine if the m-object is always the receiver of the currently analyzed method. If so, it analyzes the method with the this variable as an element of the set of variables V that must point to the m-object at the entry point of the method. With the Caller Context Extension, the property lattice, associated ordering relation, and meet operator are the same as for the Intraprocedural analysis. Figure 6 presents the additional data ow equation that denes the data ow result at the entry point of each method. The equation basically states that if the receiver object of the method is the m- object at all call sites that may invoke the method, then the this variable refers to the m-object at the start of the method. Note that because class (static) methods have no receiver, V is always ; at the start of these methods. It is straightforward to extend this treatment to handle call sites in which an m-object is passed as a parameter other than the receiver. Within strongly-connected components of the call graph, the analysis uses a xed point algorithm to compute the greatest xed point of the combined interprocedural and intraprocedural equations. It initializes the analysis with fthisg at each method entry point, Var at all other program points within the strongly-connected component, then iterates to a xed point. Between strongly-connected components, the algorithm simply propagates the caller context information in a top-down fashion, with each strongly-connected component analyzed before any of the components that contain methods that it may invoke. 3.6 The Full Interprocedural Analysis The Full Interprocedural analysis combines the Callee Type Extension and Caller Context Extension. The transfer functions are the same as for the Callee Only analysis, given in Table 4. Likewise, the property lattice, associated ordering relation and meet operator are the same as for the Callee Only analysis. The analysis result at the entry point of the method, however, is subject to the equation given in Figure 7. With this extension, the analysis will recognize that it can use hfthisg; ;i as the analysis result at the entry point entry m of a method m if, at all call sites that may invoke m, the receiver object of the method is the m-object and the type set is ;. Note that if we expand our denition of the safe method, we can additionally propagate type set information from the calling context into the called method. Like the algorithm from the Caller Only analysis, the algorithm for the Full Interprocedural analysis uses a xed point algorithm within strongly-connected components and propagates caller context information in a top-down fashion between components. It initializes the analysis algorithm to compute the greatest xed point of the data ow equations. 3.7 How to Use the Analysis Results It is easy to see how the results of the Intraprocedural analysis can be used to remove unnecessary write barriers. Since an m-variable must point to the most recently allocated ob- ject, the write barrier can be removed for any store to an object pointed to by an m-variable, since the reference created must point from a younger object to an older one. The results of the Caller Only analysis are used in the same way. It is less obvious how the analysis results are used when the Callee Type Extension is applied, since the results now include a type set in addition to the variable set. Consider a store of the form \v1 and the analysis result computed for the program point immediately before the store. If v1 2 V, then v1 must point to the m-object. Any object allocated more recently than the m-object must have type C such that C 2 T. If the actual (i.e., dynamic) type of the object pointed to by v2 is not included in T, then the object that v2 points to must be older than the object that v1 points to. The write barrier associated with Allocated st st i is a CALL@ [ Allocated Types(m j )AC C C C C A Figure 5: Equation for the Allocated Types Function A(entry m is an instance method and st 2 Callers(m); v1 2 V st is of the form Figure Equation for the Entry Point of a Method m for the Caller Only Analysis the store can therefore be removed if v1 2 V, and if the type of v2 is not an ancestor of any type in T. Note that v2 62 T is not a su-cient condition since the static type of v2 may be dierent from its dynamic type. The analysis results are used in this way whenever the Callee Type Extension is applied (i.e., for both the Callee Only and the Full Interprocedural analyses). 4. EXPERIMENTAL RESULTS We next present experimental results that characterize the eectiveness of our optimization. In general, the Full Interprocedural analysis is able to remove the majority of the write barriers for most of our applications. For applications that execute many write barriers per second, this optimization can deliver modest performance benets of up to 7% of the overall execution time. There is synergistic interaction between the Callee Type Extension and the Caller Context Extension; in general, the analysis must use both extensions to remove a signicant number of write barriers. 4.1 Methodology We implemented all four of our write barrier elimination analyses in the MIT Flex compiler system, an ahead-of-time compiler for Java programs written in Java. This system, including our implemented analyses, is available under the GNU GPL at www:flexc:lcs:mit:edu. The Flex runtime uses a copying generational collector with two generations, the nursery and the tenured generation. It uses remembered sets to track pointers from the tenured generation into the nursery [18, 1]. Our remembered set implementation uses a statically allocated array to store the addresses of the created references. Each write barrier therefore executes a store into the next free element of the array and increments the pointer to that element. By manually tuning the size of the array to the characteristics of our applications, we are able to eliminate the array over ow check that would otherwise be necessary for this implementation. 4 We present results for our analysis running on the Java ver- 4 Our write barriers are therefore somewhat more e-cient than they would be in a general system designed to execute arbitrary programs with no a-priori information about the behavior of the program. sion of the Olden Benchmarks [6, 5]. This benchmark set contains the following applications: An implementation of the Barnes-Hut N-body solver [2]. bisort: An implementation of bitonic sort [4]. em3d: Models the propagation of electromagnetic waves through objects in three dimensions [8]. Simulates the health-care system in Colombia [15]. mst: Computes the minimum spanning tree of a graph using Bentley's algorithm [3]. perimeter: Computes the total perimeter of a region in a binary image represented by a quadtree [17]. power: Maximizes the economic e-ciency of a community of power consumers [16]. treeadd: Sums the values of the nodes in a binary tree using a recursive depth-rst traversal. tsp: Solves the traveling salesman problem [14]. voronoi: Computes a Voronoi diagram for a random set of points [9]. We do not include results for tsp because it uses a nonde- terministic, probabilistic algorithm, causing the number of barriers executed to be vastly dierent in each run of the same executable. In addition, for three of the benchmarks (bh, power, and treeadd) we modied the benchmarks to construct the MathVector, Leaf, and TreeNode data structures, respectively, in a bottom-up instead of a top-down manner. We present results for the following compiler options: Baseline: No optimization, all writes to the heap have associated write barriers. A(entry m is an instance method and st 2 Callers(m); v1 2 V; where hV; st is of the form Figure 7: Equation for the Entry Point of a Method m for the Full Interprocedural Analysis Intraprocedural: The Intraprocedural analysis described in Section 3.3. Callee Only: The analysis described in Section 3.4, which uses information about the types of objects allocated in invoked methods. Caller Only: The analysis described in Section 3.5, which uses information about the contexts in which the method is invoked. Specically, the analysis determines if the receiver of the analyzed method is always the most recently allocated object and, if so, exploits this fact in the analysis of the method. Full Interprocedural: The analysis described in Section 3.6, which uses both information about the types of objects allocated in invoked methods and the contexts in which the analyzed method is invoked. The Caller Only and Full Interprocedural analyses view dynamically dispatched calls as :Analyzable. The transfer functions for these call sites conservatively set the analysis information to h;; ;i. As explained below in Section 4.4, including the allocation information from these call sites sig- nicantly increases the analysis times but provides no corresponding increase in the number of eliminated write barriers. For each application and each of the analyses, we used the MIT Flex compiler to generate two executables: an instrumented executable that counts the number of executed write barriers, and an uninstrumented executable without these counts. For all versions except the Baseline version, the compiler uses the analysis results to eliminate unnecessary write barriers. We then ran these executables on a 900MHz Intel Pentium-III CPU with 512MB of memory running RedHat Linux 6.2. We used the default input parameters for the Java version of the Olden benchmark set for each application (given in Table 13). 4.2 Eliminated Write Barriers Figure 8 presents the percentage of write barriers that the dierent analyses eliminated. There is a bar for each version of each application; this bar plots (1 W=WB ) 100% where W is the number of write barriers dynamically executed in the corresponding version of program and WB is the number of write barriers executed in the Baseline version of the program. For bh, health, perimeter, and treeadd, the Full Interprocedural analysis eliminated over 80% of the write barriers. It eliminated less than 20% only for bisort and em3d. Note the synergistic interaction that occurs when exploiting information from both the called methods and the calling context. For all applications except health, the Caller Only and Callee Only versions of the analysis are able to eliminate very few write barriers. But when combined, as in the Full Interprocedural analysis, in many cases the analysis is able to eliminate the vast majority of the write barriers. bh bisort em3d health mst perimeter power treeadd voronoi Full Interprocedural Caller Only Callee Only Intraprocedural 0% 20% 40% 80% 100% Percentage Decrease in Write Barriers Executed Figure 8: Percentage Decrease in Write Barriers Execute To evaluate the optimality of our analysis, we used the MIT Flex compiler system to produce a version of each application in which each write instruction is instrumented to determine if, during the current execution of the program, that write instruction ever creates a reference from an older object to a younger object. If the instruction ever creates such a reference, the write barrier is denitely necessary, and cannot be removed by any age-based algorithm whose goal is to eliminate write barriers associated with instructions that always create references from younger objects to older objects. There are two possibilities if the store instruction never creates a reference from an older object to a younger object: 1) Regardless of the input, the store instruction will never create a reference from an older object to a younger object. In this case, the write barrier can be statically re- moved. Even though the store instruction did not create a reference from an older object to a younger object in the current execution, it may do so in other executions for other inputs. In this case, the write barrier cannot be statically removed. Figure 9 presents the results of these experiments. We present one bar for each application and divide each bar into three categories: Unremovable Write Barriers: The percentage of executed write barriers from instructions that create a reference from an older object to a younger object. Removed Write Barriers: The percentage of executed write barriers that the Full Interprocedural anal-0.10.30.50.70.9bh bisort em3d health mst perimeter power treeadd voronoi Proportion of DynamicWrite Barriers Unremovable Removed Potentially Removable Figure 9: Write Barrier Characterization ysis eliminates. Potentially Removable: The rest of the write barri- ers, i.e., the percentage of executed write barriers that the Full Interprocedural analysis failed to eliminate, but are from instructions that never create a reference from an older object to a younger object when run on our input set. These results show that for all but two of our applications, our analysis is almost optimal in the sense that it managed to eliminate almost all of the write barriers that can be eliminated by any age-based write barrier elimination scheme. 4.3 Execution Times We ran each version of each application (without instrumen- tation) four times, measuring the execution time of each run. The times were reproducible; see Figure 15 for the raw execution time data and the standard deviations. Figure presents the mean execution time for each version of each application, with this execution time normalized to the mean execution time of the Baseline version. In general, the benets are rather modest, with the optimization producing overall performance improvements of up to 7%. Six of the applications obtain no signicant benet from the optimiza- tion, even though the analysis managed to remove the vast majority of the write barriers in some of these applications. Figure 11 presents the write barrier densities for the dier- ent versions of the dierent applications. The write barrier density is simply the number of write barriers executed per second, i.e., the number of executed write barriers divided by the execution time of the program. These numbers clearly show that to obtain signicant benets from write barrier elimination, two things must occur: 1) The Baseline version of the application must have a high write barrier density, and 2) The analysis must eliminate most of the write barriers. 4.4 Analysis Times Figure 12 presents the analysis times for the dierent applications and analyses. We include the Full Dynamic Interprocedural analysis in this table | this version of the analysis includes callee allocated type information for call0.910.930.950.970.99bh bisort em3d health mst perimeter power treeadd voronoi Normalized Execution Time Intraprocedural Callee Only Caller Only Full Interprocedural Figure 10: Normalized Execution Times for Benchmark Programs Benchmark Write Barrier Density (write barriers/s) bisort 4769518 em3d 773375 health 624960 mst 1031059 perimeter 2053484 power 3286 treeadd 955755 Figure 11: Write Barrier Densities of the Baseline Version of the Benchmark Programs sites that (because of dynamic dispatch) have multiple potentially invoked methods. As the times indicate, including the dynamically dispatched call sites signicantly increases the analysis times. Including these sites does not signi- cantly improve the ability of the compiler to eliminate write barriers, however, since the Full Interprocedural analysis is already nearly optimal for seven out of nine of our benchmark programs. 4.5 Discussion The experimental results show that, for many of our benchmark programs, our analysis is able to remove a substantial number of the write barriers. The performance improvement from removing these write barriers depends on the inherent barrier density of the application | the larger the barrier density, the larger the performance improve- ment. While the performance impact of the optimization will clearly vary based on the performance characteristics of the particular execution platform, the optimization produces modest performance increases on our platform. By instrumenting the application to nd store instructions Analysis Time Full Full Dynamic Benchmark Intraprocedural Callee Only Caller Only Interprocedural Interprocedural bh bisort em3d health mst perimeter power treeadd voronoi Figure 12: Analysis Times for Dierent Analysis Versions that create a reference from an older object to a younger object, we are able to obtain a conservative upper bound for the number of write barriers that any age-based write barrier elimination algorithm would be able to eliminate. Our results show that in all but two cases, our algorithm achieves this upper bound. We anticipate that future analyses and transformations will focus on changing the object allocation order to expose additional opportunities to eliminate write barriers. In general, this may be a non-trivial task to automate, since it may involve hoisting allocations up several levels in the call graph and even restructuring the application to change the allocation strategy for an entire data structure. 5. RELATED WORK There is a vast body of literature on dierent approaches to write barriers for generational garbage collection. Comparisons of some of these techniques can be found in [19, 12, 13]. Several researchers have investigated implementation techniques for e-cient write barriers [7, 10, 11]; the goal is to reduce the write barrier overhead. We view our techniques as orthogonal and complementary: the goal of our analyses is not to reduce the time required to execute a write barrier, but to nd super uous write barriers and simply remove them from the program. To the best of our knowledge, our algorithms are the rst to use program analysis to remove these unnecessary write barriers. 6. CONCLUSION Write barrier overhead has traditionally been an unavoidable price that one pays to use generational garbage collec- tion. But as the results in this paper show, it is possible to develop a relatively simple interprocedural algorithm that can, in many cases, eliminate most of the write barriers in the program. The key ideas are to use an intraprocedural must points-to analysis to nd variables that point to the most recently allocated object, then extend the analysis with information about the types of objects allocated in invoked methods and information about the must points-to relationships in calling contexts. Incorporating these two kinds of information produces an algorithm that can often eectively eliminate virtually all of the unnecessary write barriers. Benchmark Input Parameters Used bh bisort 250000 numbers em3d 2000 nodes, out-degree 100 health 5 levels, 500 time steps mst 1024 vertices perimeter power 10000 customers treeadd 20 levels voronoi 20000 points Figure 13: Input Parameters Used on the Java Version of the Olden Benchmarks 7. ACKNOWLEDGEMENTS C. Scott Ananian implemented the Flex compiler infrastructure on which the analysis was implemented. Many thanks to Alexandru Salcianu for his help in formalizing the analyses 8. --R Simple generational collection and fast allocation. A hierarchical O(N log N) force calculation algorithm. A parallel algorithm for constructing minimum spanning trees. Adaptive bitonic sorting: An optimal parallel algorithm for shared-memory machines Data ow analysis for software prefetching linked data structures in Java. Software caching and computation migration in Olden. The Design and Implementation of the Self Compiler Parallel programming in Split-C Remembered sets can also play cards. Garbage Collection Algorithms for Automatic Dynamic Memory Management. Probabilistic analysis of partitioning algorithms for the traveling-salesman problem in the plane A performance study of Time Warp. Decentralized optimal power pricing: the development of a parallel program. Computing perimeters of regions in images represented by quadtrees. Generational scavenging: A non-disruptive high performance storage reclamation algorithm Barrier methods for garbage collection. --TR Adaptive bitonic sorting: an optimal parallel algorithm for shared-memory machines Simple generational garbage collection and fast allocation A comparative performance evaluation of write barrier implementation The design and implementation of the self compiler, an optimizing compiler for object-oriented programming languages Decentralized optimal power pricing Parallel programming in Split-C Software caching and computation migration in Olden Garbage collection Data Flow Analysis for Software Prefetching Linked Data Structures in Java Generation Scavenging

program analysis;generational garbage collection;write barriers;pointer analysis

510867

Efficient global register allocation for minimizing energy consumption.

Data referencing during program execution can be a significant source of energy consumption especially for data-intensive programs. In this paper, we propose an approach to minimize such energy consumption by allocating data to proper registers and memory. Through careful analysis of boundary conditions between consecutive blocks, our approach efficiently handles various control structures including branches, merges and loops, and achieves the allocation results benefiting the whole program. The computational cost for solving the energy minimization allocation problem is rather low comparing with known approaches while the quality of the results are very encouraging.

Introduction Today's high demand of portable electronic products makes low energy consumption as important as high speed and small area in computer system design. Even for non-portable high performance systems, lower power consumption design helps to decrease packing and cooling cost and increase the reliability of the systems [25]. A lot of research has been done in improving the power consumptions of various components in computer systems [8, 13, 19, 20, 25]. In particular, it is well recognized that access to dierent levels of storage components, such as register, cache, main memory and magnetic disk, diers dramatically in speed and power consumption [8, 17]. Hence allocation of variables in a program to dierent storage elements plays an important role toward achieving high performance and low energy consumption [13]. With today's IC technology and computer architecture, memory read/write operations take much more time and consume much more power than register read/write operations [3, 8]. For some data-intensive applications, energy consumption due to memory access can be more than 50%[16]. In this paper, we focus on the problem of allocating variables in a program between registers and main memory to achieve both short execution time and low energy consumption. Register allocation for achieving the optimal execution time of a program on a given system is a well known problem and has been studied extensively. In general, existing results can be divided into two categories, local register allocation [5, 15, 18] and global register allocation [2, 4, 11, 14, 22, 23, 27, 28, 29]. A local register allocator considers the code in one basic block (a sequence of code without any branch) at a time and nds the best allocation of the variables in that basic block to achieve the minimum execution time. A global allocator deals with code containing branches, loops, and procedure calls. The global register allocation problems are in general NP-hard [22] and can be modeled as a graph-coloring problem [1, 4, 6, 7, 11]. Integer linear programming approaches and heuristics have been proposed to solve the global register allocation problems [22, 14]. However, optimal-time register allocations do not necessarily consume the least amount of energy [8, 19, 20]. Some researchers have recognized this challenge and have proposed interesting approaches to deal with the optimal-energy register allocation problem [9, 13, 25]. Chang and Pedram [9] gave an approach to solve the register assignment problem in the behavioral synthesis process. They formulated the register assignment problem for minimum energy consumption as a minimum cost clique covering problem and used max-cost ow algorithm to solve it. But memory allocation is not considered in [9]. Gebotys [13] modeled the problem of simultaneous memory partitioning and register allocation as a network ow problem and an ecient network ow algorithm can be applied to nd the best allocation of variables in a basic block to registers and memory. However, the work in [13] is only applicable to basic blocks. In this paper, we extend the work in [13] by investigating the problem of register and memory allocation for low energy beyond basic blocks. We consider programs that may contain any combination of branches, merges, and loops. The basis of our approach is to perform the allocation block by block. By carefully analyzing boundary conditions between consecutive blocks, we are able to nd the best allocation for each block based on the allocation results from previous blocks. In this way, we maintain the global information and make allocation decision to benet the whole program, not just each basic block. Our approach can be applied to dierent energy and variable access models. The time complexity of our algorithm is O(b l(n; k)), where b is the total number of blocks and l(n; k) is the time complexity for register allocation in a single block that has n variables and k available registers. The function l(n; k) is either O(nlog n), O(knlog n), or O(kn 2 ), depending on the dierent energy and access models used. The rest of the paper is organized as follows. Section 2 reviews the energy models and known approaches for local register allocation for low energy. We also give an overview of how our allocation approach works. Section 3 gives our approach to do the allocation for the branch control structure. Section 4 presents our algorithm for low energy register allocation for the merge structure in a control ow graph. In section 5, we discuss how to treat loops to get better register allocation results. Section 6 summarizes the experimental results and concludes with some discussions. Overview Allocating variables to dierent levels of storage components is usually done by a compiler. Many factors can aect the allocation results. There are a number of techniques that can be applied to reduce the number of cycles needed to execute a program, including instructions reordering and loop unrolling. In this paper, we assume that such optimization techniques have been applied to the program and a satisfactory schedule of the program is already given. 2.1 Energy models and known approaches be the set of variables in the program under consideration. Given a schedule, the lifetime of a variable v in a basic block is well dened and is denoted by its starting time, t s (v), and nishing time, t f (v). We use VR (resp., VM ) to represent the set of variables assigned to the k available registers (resp., memory). Furthermore, let e M wv ) be the energy consumed by reading (resp., writing) variable v from (resp., to) memory and e R wv ) be the energy consumed by reading (resp., writing) variable v from (resp., to) registers. If v is not specied, e M r represents the energy consumed by one read (resp., write) access to the memory and e R r represents the energy consumed by one read (resp., write) access to registers. Denote the total energy consumed by a given program as E. Then the objective of the register allocation problem is to map each variable v 2 V to register les or memory locations in order to achieve the least energy consumption due to data accesses, that is, to minimize v2VR Dierent energy models and dierent data access models have been proposed in the literatures [5, 8, 9, 13, 20]. Depending on the particular models used, the calculation of energy consumption varies. One of the energy model is the static energy model SE, which assumes that references to the same storage component consume the same amount of energy [8]. This model is like the reference-time model, where accesses to the same storage component take the same amount of time, no matter of what is the value of the referenced data. In [9, 20], the authors proposed a more sophisticated energy model, called activity based energy model AE, capturing the actual data conguration in a storage. Under this model, the energy consumed by a program is related to the switched capacitance of successive accesses of data variables which share the same locations. The switched capacitance varies for dierent access sequences. In particular, the product of the Hamming distance, the number of bits two data items dier [10], or other measurements [9] and the total capacitance of a storage is used to represent the switched capacitance for two consecutive references. Whether a data variable is read only once (SR) or multiple times (MR) after it is dened (written) also has an eect on the overall energy estimation. Carlisle and Lloyd [5] presented a greedy algorithm for putting the maximum number of variables into registers to minimize the execution time of a basic block under the signal read model. Their approach also gives an optimal energy consumption for the static energy model. It takes O(nlog n) time and is quite easy to implement. For the multiple-read case, the algorithm proposed in [13] uses a graph which can have O(n 2 ) edges in the worst case. A better graph model is given in [5], which has only O(n) edges. Applying the minimum cost network ow algorithm, the optimal-time solution, which is also the optimal- energy solution, can be obtained in O(knlog n) time [5] rather than O(kn 2 ) as in [13]. As one can see, dierent models result in dierent computational overhead in computing energy consumption. For example, the AE-MR model is the most comprehensive and also most computationally expensive one. Depending on the availability of the models, the architecture features of a system, and the computational overhead that one is willing to incur, any of the above models may be used in energy consumption estimation. Therefore, we consider all the two energy models and the two variable access models in our paper. For the activity based energy model, we will only consider it for register le access to simplify the problem formulation. (Adopting the activity based model for memory is a simple extension to the algorithm discussed later in this paper.) Under this model, the total energy consumption of a program can be calculated as is the Hamming distance between v i and v j , designates that v j is accessed immediately after v i and that v i and v j share the same register, C R rw is the average switched capacitance for register le reference and VR is the operational voltage of the register le. The objective for optimal energy allocation is to minimize the objective function (2). For both the single read and multiple read models, Gebotys [13] formulated the problem as a minimum cost network ow problem with a solution of O(kn 2 complexity. The cost for each edge in the network ow problem for dierent read models is also dierent. In some designs, it is desirable to allow a variable to be assigned to registers in one time interval and switched it to memory in another time interval, in order to minimize the total number of accesses to memory. This is called split lifetime [11]. If the split lifetime model is used, the graph model will need to accommodate all cases where a split occurs. Hence, O(n 2 ) graph edges will be needed, and the algorithm in [5] has an O(kn 2 complexity, the same as that in [13]. The results we discussed above only apply to variables in a basic block, i.e., a piece of straight-line code. But real application programs may contain some control structures, such as branches, merges and loops. A merge is the case where one block has more than one parent blocks. Such type of control ows occurs when an instruction has more than one entry point (due to branching, looping, or subroutine calls). To consider register allocation in a program with these control structures, one must deal with those variables whose lifetimes extend beyond one basic block. A straightforward way of applying the existing approaches to such programs would be to enumerate all possible execution paths and treat each path as a piece of straight line code. Two problems arise with this approach. First, the number of paths increases exponentially in the worst case as the number of basic blocks increases. Secondly, each variable may have dierent lifetimes in dierent paths. Hence such an approach would be computationally costly and be dicult to obtain an optimal allocation. On the other hand, nding an optimal allocation for all the variables in a general program has been known to be computational intractable (an NP-hard problem). 2.2 Overview of our approach In the following, we outline our heuristic approach which can nd superior solutions for the minimum energy register allocation problem. To avoid the diculties encountered by the known approaches discussed above, we use an iterative approach to handle programs containing non- straight-line code. Such program is modeled by a directed graph G in which each node represents a basic block. Given the nature of a program, the corresponding graph always has a node whose in-degree is zero. Furthermore, for each loop, the edges that branch backwards can be identied. A topological order for the nodes in G can thus be established after removing those edges branching backwards. To ease our explanation, we generalize the concept of parent and child used in trees. A parent (resp., child) of a block B is a block whose corresponding node in G is an immediate ancestor (resp., descendent) of the node corresponding to B. Hence, a block may have multiple parents blocks (in the case of a merge) and it may be a parent block to itself (in the case of a loop). We solve the register allocation problem for each basic block in the topological order. The allocation of variables within a block is based on the results from its parent blocks. Any assignment of this variable in the current block must consider the eect of its initial assignment. Furthermore, we do not allow the allocation result in a block to aect the result in its parent blocks. By allowing the allocation results to propagate \downwards" without back-tracking, we eliminate the excessive computational cost yet obtain superior allocations. A main challenge is how to handle those variables whose lifetimes extend beyond a basic block such that the overall energy consumption of the program is minimized. We have made several observations to help deal with such variables. The key idea is to set up appropriate boundary conditions between parent and child basic blocks and use these conditions to guide the register allocation procedure within the child basic blocks. We rst describe the graph model used in [13], and then introduce extensions to handle variables beyond basic blocks. Consider a basic block B in G. A directed graph which is a generalization of an interval graph, is associated with B. For each data variable whose lifetime overlaps with the execution duration of B, two nodes, n s (v) and n f (v) are introduced, which correspond to the starting time t s (v) (when v is rst written) and the nishing time t f (v) (when v is last read), respectively. Note that v can be either a variable referenced in B or a variable which will be referenced by B's descendents. F a b c d e f ns(a) ns(b) ns(c) ns(d) nf(a) nf(d) Figure 1: The graph G for a basic block without considering other blocks in the program Several dierent types of arcs are used in G. First, each pair of n s (v) and n f (v) is connected by an arc a(n s (v); n f (v)). To dene the rest of the arcs, we introduce the concept of critical set. A critical set, C i , is a set of variables with overlapping lifetimes to one another such that the number of variables in the set is greater than k, the number of available registers. For each pair of C i and C i+1 (the index ordering is based on scanning the variables from the beginning to the end of block B), let D i be the set of variables whose lifetimes are in between the minimum nishing time of all variables in C i and the maximum starting time of all variables in C i+1 . Note that D 0 contains the variables whose lifetimes are in between the starting time of B and the maximum starting time of all variables in C 1 , and that D g is the set of variables whose lifetimes are in between the minimum nishing time of all variables in C g (the last critical set in B) and the nishing time of B. Now, an arc is introduced from n f (u) for each u 2 (C i [D i ) to n s (v) for each Intuitively, these arcs represent allowable register sharing among subsequent data references. Similar to [13], a source node, S, and a nish node, F , are introduced at the beginning and end of B, respectively. Arcs are used to connect S to n s (v) for each v 2 (D 0 [C 1 ) and to connect for each u 2 (C g [D g ) to F . The nodes S and F can be considered as the registers available at the beginning and end of block B. An example graph is shown in Figure 1, where the solid lines correspond to the arcs associated with variables in B and the dash lines are the arcs introduced based on C i 's and D i 's. In this graph, variable b, c and d form the rst critical set, C 1 , while variable a and b form the set D 0 . To handle the allocation beyond a block boundary, the graph needs to be modied. Substantial modications to the existing approaches as well as new ideas are proposed here in order to eciently handle the branch, merge, and loop structures. Bc Figure 2: Four dierent allocation results for a cross-boundary variable 3 Register Allocation for the Branch Structure In this section, we discuss in detail our algorithms of register allocation for the branch structure for the two dierent energy models. 3.1 Beyond boundary allocation for the static energy model We rst consider the case in which only single read is allowed, and then extend the result to the multiple read case. For a variable, v, if it is dened in a block B 0 and is also read in another block B, the lifetime of v crosses the boundary of the two blocks. The block B 0 is the parent block, B p , of the block B. We say v is in both B p and B. When our algorithm computes the allocation for B p , it performs the computation as if v did not cross the boundary. When we come to the block B, we will make the best allocation decision for v in B based on the allocation result of v in B p . There are totally four dierent combinations of allocation results for v in the two consecutive blocks: (1) v in register in both B p and B (R!R), (2) v in register in B p but in memory in B (R!M), (3) v in memory in B p but in register in B (M!R), and (4) v in memory both in B p and B (M!M), as shown in Figure 2. In Figure 2, solid line segments mean the variable is in register, while dashed segments mean the variable is in memory. (We will use this convention throughout the paper unless stated otherwise explicitly.) We will analyze the energy consumption for all the four cases, R!R, R!M, M!R, and M!M. For the R!R case, the energy consumed by v in block B p is e R w when it is dened, in block B is e R r when it is read from the same register, and totally the energy consumed by v in the two consecutive blocks, r . By applying the same analysis to other three cases, we obtain the amount of energy consumed by a crossing-boundary variable in dierent blocks, as shown in the column R!R, R!M, M!R, and M!M in Table 1. For a local variable of B, which is written and read only in block B, the energy consumed by Table 1: Energy consumption in dierent blocks under dierent allocation results Block R! R R! M M! R M! M Local R Local w e R r e R r e R r r e R r e R r it in B is e R r or e M r if it is assigned to a register (Local R) or a memory (Local M) location. It consumes no energy in B p since it does not exist in B p . The energy consumed by a local variable in B is shown in the last two columns (Local R and Local M) in Table 1. From Table 1, it is easy to see that, based on the allocation result for B p , the best choice for a global variable in B should be: is in memory in B p , it should stay in memory in B to achieve the lowest energy. Comparing the two columns corresponding to the situation when v is in memory in B p , M!R and M!M, it is clear that if v is in memory in B p , assigned it to a register will cost more energy than assign it to memory in B. R!Local : If v is in register in B p , it should be treated as a brand new local variable in block B to achieve the optimal energy. The energy data for block B and B in the columns for R!R and Local R is same, while the data diers only by e R r in the columns for R!M and Local M in which the total amount of energy is much larger than the dierence. So a simple way is to just treat this kind of global variable as a brand new local variable whose allocation is to be determined in block B. If it turns out to be assigned to register in B too, it should stay in the same register as it uses in B p . Treating v, which extends beyond boundary and is assigned to a register in B p , as a brand new variable in B is clearly an approximation. Another way is to assign the actual energy saving, the energy saved by assigning a variable to register comparing with assigning it to memory, as the weight of the variable and construct a network ow and then use minimum-cost network ow algorithms to solve the problem as described in [5] for the weighted case. A variable, v, can also be dened in B, not read in the child of B, but read in the grandchild or even grand-grandchild of B. The analysis for this kind of variable is as same as the above analysis for the variable that dened in B p and used in B, and same conclusion is drawn for the allocation in the children blocks of B. With these rules, we can execute the local register allocation algorithm on each of the blocks in a program. For the simple case, the time complexity of the allocation for the whole program is O(bnlog n), where n is the largest number of local variables in a basic block. For the more complex case, the time complexity is O(bknlog n). The algorithm is shown in Algorithm 1. As we discussed earlier in Section 2, a variable in a program may be read more than once after it is written. In this case, we can use the same graph model as in [5] but modify the weights Algorithm 1 Algorithm for low energy register allocation in static energy model Input: A scheduled program with blocks, P g. Output: The register assignment for every variable v in program P . Denitions: allocation result for variable v in block B i . B i (v) is 1 if v is in register, 0 if v is in memory. weight(v): the weight of variable v. apply algorithm for the unweighted case in [5] to B 1 to m do if using the simple way for approximation then for all variables, v, live in B i p and B i do else v is treated the same as other local variables end for apply algorithm for the unweighted case in [5] to B i else for all variables, v, live in B i p and B i and is read in B i do else r end for for all variables, v, live in B i p and B i but not read in B i do else end for for all variables written and read in B i do end for for all variables written but not read in B i do w e R end for apply algorithm for the weighted case in [5] to B i end for Table 2: Weight for global variables and local variables in B Block Global v in Register in B p Global v in Memory in B p Local associated with certain graph edges. Specically, we assign the energy saving by assigning a variable to a register as the weight for this variable. For the current block, B, the weight of a variable dened in B is h v e M r +e R is the total number of read accesses to v as dened in Section 2. If a variable is live in the parent block, B p , and is assigned to a register in B p , the weight is for it in the current block B is h v e M r ). Otherwise, if it is allocated to memory in B p , the weight is h v e M r ). Note that the weight is the energy saving when v is allocated to a register in the current block B. Since the total energy of accessing all the variabls from memory is xed for a basic block, the more energy saved by allocating variables to registers, the less the energy the block consumes. The weight assignment is summaried in Table 2. By applying the network ow approach in [5] for weighted intervals on each basic block one after another, we can get a low energy allocation solution for the whole program in O(bknlog n) time. If splitting is allowed, the graph model in [5], which only allows the register transfered from longer sucient. In this case, we can use the graph model (to be described in the next section) and the network ow approach presented in [13]. That is, we assign weights as described in Table 2 and use a minimum-cost network ow algorithm to solve the problem in O(bkn 2 ) time. 3.2 Register allocation for branches for the activity-based energy model To handle the allocation beyond a block boundary, our approach is to decide the register allocation for block B based on the allocation result from B's parent block, B p . Depending on which variable is assigned to which register prior to entering B, the amount of energy consumption by each variable in B can be dierent. Therefore, simply using a single source S to represent the registers available at the beginning of B is no longer sucient. We generalize the construction of graph G for B, where B has both a parent and at least one child, as follows. (The graphs for the root and leave blocks in T are simply special cases of the graph we discuss.) For those variables that are referenced only in B, we introduce arcs and nodes associated with them in the same way as we discussed above. The start and nish nodes, S and F , for B are also maintained. Let the starting and nishing times of block B be t s (B) and t f (B), respectively. For each variable v in B whose starting (resp. nishing) time is earlier (resp. later) than the starting use two end nodes n s (v) and n f (v) in G for the associated arc a(v) (which means that v is considered by all e f c np(a) ns(c) nf(c) F Figure 3: The graph G for block B based on the allocation of its parent block B p . the graphs corresponding to the blocks with which the lifetime of v overlaps). The arcs between these nodes and S or F are dened in the same ways as discussed in the previous paragraphs. Furthermore, we introduce a register set, VRB , which contains the variables residing in registers at the completion of the allocation process for block B. Note that VRB becomes uniquely dened after the allocation process of B is completed, and that the size of VRB , jV RB j, is always less than or equal to k, the number of available registers. We expand G by adding a node n p (v) for each is the parent block of B. It is not dicult to see that the variables in VRBp are the only variables which have a chance to pass on their register locations to variables in B directly. Now, we insert an arc from each n p (u) to n s (v) for each v 2 (D are as dened in the previous paragraphs). Comparing our generalized graph with the original graph, one can see that at most k additional nodes and k jD additional arcs are used in the generalized graph. Figure 3 shows an example graph G for the current block B assuming that there are three available registers. Sometimes, a program may read a variable, v, more than once, In this case, we introduce an additional node n r i (v) for each read of v except the last read. Additional arcs are also introduced to model possible register sharing between variables. Due to the page limit, we omit the discussion on this part. Given the directed graph G for B, we are ready to construct the network ow problem associated with G. Let x(n f (u); n s (v)) and c(n f (u); n s (v)) be the amount of ow and the cost of one unit of ow on arc a(n f (u); n s (v)), respectively. Denote the total amount of energy consumed by B as E. The objective function of our network ow problem can be written as: v2Bjts (v)<ts (B) a(p;q)2A where A is the set of arcs in G. In (3), the rst three terms are the amount of energy consumed by B if all the variables are assigned to memory, and the last term represents the energy saved by allocating certain variables to registers. The values of x(p; q) are unknown and to be determined. If corresponds to a variable v, then v will be assigned to a register. The values of c(p; q) are dependent on the types of arcs associated with them, and can be categorized into the following cases. 1) For an arc from a node of type n f to another node of type n s , i.e. a(n f (u); n s (v)), the cost associated to the arc, c(n f (u); n s (v)), is computed by r H(u; v)C R where N is the set of nodes in G. This is the amount of energy saved by reading u from a register and writing v to the same register. For an arc from a node of type n p to another node of type n s , i.e. a(n p (u); n s (v)), the cost associated to the arc, c(n p (u); n s (v)), is dened dierently. There are a total of 7 cases to be considered. 2.1) If u is not in B(i.e., u's lifetime does not overlap with that of B), and v is written in B, the cost c(n p (u); n s (v)) is computed by 2.2) If u is not in B, and v has been assigned to a register during the allocation process of B p , If u is not in B, and v has been assigned to memory during the allocation process of B p , r H(u; v)C R If u is in B, and v is written in B, the cost c(n p (u); n s (v)) is the same as dened in (6) for Case 2.2. 2.5) If u is in B, and v has been assigned to a register during the allocation process of B p , If u is in B, and v has been assigned to memory during the allocation process of B p , r H(u; v)C R 2.7) If u and v represent the same variable, the cost c(n p (u); n s (v)) is simply assigned to zero. For an arc from start node S to another node of type n s , i.e. need to have three dierent cost functions. 3.1) If v is written in B, where H(0; v) is the average Hamming distance between 0 and a variable v, and is normally assumed to be 0.5. If v has been assigned to a register during the allocation process of B p , If v has been assigned to memory during the allocation process of B p , For an arc from a node of type n f to the nish node, F , i.e. a(n f (v); F ) for v 2 D g [ C g , we need to have two dierent cost functions. If v is read in B, c(n f (v); F r 4.2) If v is not read in B, the cost c(n f (v); F ) is simply assigned to zero. 5) For an arc from a node of type n s to another node of type n f , which is the arc corresponding to the same variable, the cost associated to the arc is assigned to zero. Using the above equations, the objective function for the network ow problem will be uniquely dened. The constraints for the network ow problem is dened based on the number of registers available to the arc. They are summarized as following: Applying a network ow algorithm such as the one in [21] to our network ow problem instance, we can obtain the value of each x(p; q) in O(kn 2 ) time for the block B, where k is the number of available registers and n is the number of variables whose lifetimes overlap with the lifetime of B. If the resulted x value of the arc associated with a variable in B is one, the variable is assigned to the appropriate register based on the ow information. The above formulation can then be applied to each basic block in the program tree in the depth-rst search order as we discussed at the beginning of this section. Here we should point out that our method can also be used if one wishes to explore program execution paths to determine the register allocation. For example, we can associate each execution path of a program with a priority. One way to assign priorities is based on the execution frequency of each path (obtained from proling). Another way is based on the timing requirements. To solve the register allocation problem, one can start with the highest priority path, P 1 , and proceed to the lower priority ones sequentially. For the rst path, form a super block for the path and use the local register allocation algorithm in [13] to nd the best allocation for this single path. For the next path, remove the basic blocks on this path whose register assignments have been determined. Then form a super block B 0 for the rest of the basic blocks in this path. To construct the graph for this super block B 0 , we need to consider all those B p 's that have a child in B 0 and introduce the necessary n p nodes to complete the graph. The network ow problem for the super block B 0 can use the same formulations as we discussed above. The process is repeated for all subsequent paths. By applying our algorithm, the variable allocation decisions made for the higher priority paths will not be changed by the allocation processes of the lower priority paths. Rather, the results from the higher priority paths are used to insure that good allocations are found for the lower priority paths based on these results. Hence, we have eectively eliminated the con icting assignments that can arise when solving the allocation problem for each path separately. 4 Register Allocation for Blocks with Merging Parent Blocks In this section we present our approach to handle the merge case. The fact that the allocation results from dierent parent blocks may be dierent makes the merge case more dicult to handle than the branch case. If a variable, v, is alive in B but is not dened (written) in B, v must be alive when leaving every B's parents. Since the allocation results for v in dierent parent blocks may be dierent, we cannot simply do the allocation for v in B based on the allocation result from one particular parent block. Hence our register allocation approach for programs with branches, where each basic block has only one parent block, is not adequate for a block when more than one blocks merging into it. To handle the control structure of merge blocks, we devise a method to capture the non-unique assignments of variables to registers. Based on this approach, we formulate the problem of register allocation for both the static energy model and the activity based energy model as an instance of the network ow problem. Let the parent blocks of B be B assuming m is the total number of blocks merge into B and m > 1. Each parent block, B p i is associated with a probability, P (B indicating how frequently the block is executed. (Such probabilities can be obtained from proling information.) After the allocation for a parent block is nished, the allocation decision for a variable in B p i which goes beyond the boundary of B p i and B is xed. We dene the probability of v allocated to a register (resp., memory) in B p i as P (v; R; B let P (v; R be the probability of v being assigned to the register R j in block B p i . We also dene P (v; R; as the total probability of v allocated to a register (resp., memory) in all the parent blocks, P (v; R as the total probability of v allocated to the register R j in all the parent blocks. The following relations hold for the variables dened above, where k is the total number of registers. We modify the ow graph to be used in the merge case as follows. The source node, S, and the nish node, F , are kept as the same. The nodes used to represent a variable, v, are still the same as n s (v) for the starting node, and n f (v) for the nishing node. Because the register assignments in dierent parent blocks may be dierent, a cross boundary variable may have dierent allocation assignment in dierent parent blocks. To re ect this fact, a critical set, C i , is dened as a set of variables with overlapping lifetimes to one another such that the number of variables written in B in the set is greater than the number k of available registers. Furthermore, we introduce a new set of register nodes, n r (i), which correspond to the registers. In addition to all the arcs dened in Section 2, we add new arcs from source node, S, to each register node, n r (i). A complete bipartite graph is formed between the register nodes and the starting nodes for all variables in the rst critical set C 1 or D 0 . An example ow graph for a block B with multiple parent blocks is shown in Figure 4, assuming that there are two registers, R 1 and R 2 , available in the system. In minimizing energy consumption by a single block with multiple blocks merging into it, one should minimize the overall energy consumption under all execution scenarios for this block. Thus, the objective function for the ow problem needs to take into account the probabilities of executing dierent parent blocks. For dierent energy models, static energy model and activity based energy model, the objective function and the cost associated with arcs are slightly dierent. To reduce repetition, we will integrate the presentation of the formulations for both models and point out the dierences wherever necessary. The objective function can be written as: F a b c d e f ns(a) nf(a) ns(c) ns(d) nf(c) nf(d) Figure 4: The graph G for a basic block with more than one parent blocks v2Bjts (v)<ts (B) a(p;q)2A where 1 if static energy model The rst three terms of computing the objective function (19) are the amount of energy consumed by block B if all the variables are assigned to memory, while the last term represents the energy saved by allocating certain variables to registers. The dierence in the value is due to the fact that in the activity based energy model, the energy consumption of register references is computed by the switched capacitance between two consecutive accesses and is captured by the cost, c(p; q), associated with corresponding arcs in the last term of (19). The values of x(p; q) are unknown and to be determined. If x(p; corresponds to a variable v, then v will be assigned to a register. The values of c(p; q) are dependent on the types of arcs associated with them, and can be categorized into the following cases. 1) For an arc from a node of type n f to another node of type n s , i.e., a(n f (u); n s (v)), the cost associated to the arc, c(n f (u); n s (v)), is computed dierently for two dierent kinds of energy model. For the static energy model, r while for the activity based energy model, r H(u; v)C R where N is the set of nodes in G. This is the amount of energy saved by reading u from a register and writing v to the same register. For an arc from a node of type n r (i) to another node of type n s , i.e., a(n r (i); n s (v)), the cost associated to the arc, c(n r (i); n s (v)), can take one of the following values. 2.1) If v is written in B, for the static energy model, c(n r (i); n s w e R while for the activity based energy model, c(n r (i); n s For the static energy model, the energy saving is the energy of a memory write, e M the energy of a register write, e R w . On the other hand, for the activity based energy model, the energy for writing v into a register is determined by the hamming distance between v and the variable u that occupied R i at the entry of block B. Since the probability for variable u occupying R i is P (u; R hamming distance would be the summation of the hamming distance between v and all the u which has non-zero probability of occupying R i in one of the parent blocks. The energy of writing variables in R i back to memory when R i is assigned to v is already included in the objective function in the second item. 2.2) If v is not written in B, for the static energy model, c(n r (i); n s where the rst term is the energy consumption if v is assigned to a memory location, the second is the energy of assigning v to R i in B if v is in memory in B 0 s parent blocks, and the last term is the energy of assigning v to R i in B if v is already in a register other than R i in B's parent blocks. For the activity based energy model, c(n r (i); n s r where the last term is the energy consumed by assigning v to R i which is determined by the hamming distance of v and the variables that occupies R i in any of B's parent blocks . For an arc from source node S to the node of type n r (i), i.e., a(S; n r (i)), the cost function simply zero. For an arc from a node of type n f to the nish node, F , i.e., a(n f (v); F ) for v 2 D g [ C g , we need to have two dierent cost functions. If v is read in B, for static energy model, c(n f (v); F r e R while for the activity based energy model, c(n f (v); F r (27) 4.2) If v is not read in B, the cost c(n f (v); F ) is simply assigned to zero in both the static and activity based energy model. 5) For an arc from a node of type n s to another node of type n f , which is the arc corresponding to the same variable, the cost associated to the arc is assigned to zero. Using the above c(p; q) denitions, the objective function for the network ow problem is uniquely dened. The constraints for the network ow problem are dened based on the number of registers available to dierent types of arcs. They are summarized in the following: Applying a network ow algorithm [21] to our network ow problem instance, the value of each x(p; q) can be obtained in O(kn 2 ) time for the block B, where k is the number of available registers and n is the number of variables whose lifetimes overlap with the lifetime of B. If the resulted x value of the arc associated with a variable is one, the variable is assigned to the appropriate register based on the ow information. The above formulation can then be applied to each basic block in the program control ow graph in the topological order. 5 Register Allocation for Loop In this section, we extend our approach to handle programs with loops. A loop can be considered as a control structure with more than one parent blocks merging into it. However, it presents some unique challenges since one of its parent blocks is itself. One way to solve the allocation problem for loops is to modify the ow problem formulation presented in the last section. Consider a loop with one basic block as shown in Figure 5, where the dash lines represent those variables that are referenced in more than one iteration of the loop execution. We refer to these variables as loop variables and the rest as non-loop variables. The graph for modeling the allocation problem of a loop block can be built following the same guidelines used for the merge case. The graph for the loop block in Figure 5 is shown in Figure 6. (Note that the existence of loop variables does not change the graph since only register nodes are needed to model parent block assignments.) Similar to the merge case, we associate certain probabilities to the variables at the boundary of the loop block and its parent blocks to capture potentially dierent assignments dictated by dierent parent blocks. For the example in Figure 5, assuming that the loop has one parent block corresponds to the loop itself. It may seem that one can now simply use the ow problem formulation for the merge case to nd the allocation for a loop. However, the diculty is that the probabilities associated with each loop variable (e.g., P are dependent on the allocation result of the loop block which is yet to be determined. a f f c d e e in-assignment pre-assignment post-assignment loop variable loop variable Figure 5: A loop with one basic block, where e and f are loop variables, and a, b, c and d are non-loop variables. The optimization problem can still be solved but it is no longer a simple ow problem. In general, assume that B p l is the loop body itself and that the probability of this block loops back to itself is P (B p l ). (P (B p l ) can be obtained from proling information.) The probability associated with a loop variable v can be expressed as follows. If v is assigned to register R i upon exiting the loop, i.e., x(n f (v); F is assigned to memory upon exiting the loop, i.e., x(n f (v); F then Consequently, the following expression can be derived: It follows that the ow problem in (19) can be transformed to an integer quadratic programming Loop Block e F a f f c d e ns(c) ns(d) nf(a) nf(c) nf(d) Figure Flow graph for the loop in Figure 5. problem. Integer quadratic programming problems are generally NP-hard and may be solved by the branch-and-bound method [26]. We propose an ecient heuristic approach, which generally produces superior allocation re- sults. Recall that our register allocation algorithm determines the variable allocation of a block based on the allocation results of its parent blocks. Consider a loop variable, it may have some initial assignments dictated by the allocation of its parent blocks (other than the loop block it- self), another assignment before its rst reference in the loop block, and yet another assignment upon leaving the block. To ease our explanation, we refer to these assignments as pre-assignment, in-assignment, and post-assignment, respectively. In Figure 7(a), we show a possible assignment result for the block in Figure 5, where for variable e, its pre-assignment is a register, its in-assignment is memory, and its post-assignment is a register. The basis for our approach is the following observations. In general, a loop body is executed frequently, i.e., P (B p l ) is relatively close to one. If a loop variable's pre-assignment is dierent from its in-assignment, only a single switching from memory to register or vise versa is required. However, if a loop vari- able's in-assignment is dierent from its post-assignment, as many switchings as the number of loop iterations would be needed. Hence, reducing the number of loop variables with dierent in-assignment and post-assignment would tend to result in better allocations. Our algorithm employs an iterative approach. With the rst allocation attempt, we ignore the fact that the loop body is one of its parents and compute the allocation by simply using the results from the other parent blocks. That is, the value of P (v; R; B in the ow formulation in (19) are assumed to be zero if B p i is the loop body itself and v is a loop variable. We refer to as the initialization phase ( 0 ). The algorithm proceeds as follows. If there is no loop variable whose in-assignment is dierent from its post-assignment, the optimal allocation is found and the algorithm exits. Otherwise, we perform two more allocation phases: the second phase ( 1 ) and the third phase ( 2 ). In the second phase, we let the in-assignment of each loop variable to be the value obtained from 0 and solve the resultant allocation problem. Since the values of P (v; are unknown if v is a loop variable, it may seem that we still need to solve an integer quadratic programming problem. However, applying some simple manipulations, we can avoid this diculty. In particular, we modify the value of cost c(n f (v); F ) for each loop variable v. Instead of always setting it to zero as in the merge case, it is dened according to the in-assignment of the same loop variable as follows: c(n f (v); F in-assignment is register e M e R otherwise The value of P (v; R; B p l ) and P (v; M;B p l ) for each loop variable v are still assumed to be zero. The modied ow problem correctly captures the overall energy consumption when the in-assignment of each loop variable is set to that obtained from 0 . The solution to the modied ow problem leads to an improved allocation result. In the third phase ( 2 ), we x the post-assignment of each loop variable to the value obtained in 0 . Then, P (v; R; B p l ) and P (v; M;B p l ) can be computed by post-assignment is register R if v's post-assignment is memory: By substituting in (19) the above probability values, we again reduce the allocation problem to a min-cost ow problem. Solving this ow problem will result in an allocation that improves the result obtained from 0 . The better one from 1 and 2 is kept as the best allocation result found so far. Further improvement could be obtained if more iterations were carried out by treating the current best allocation as the one obtained from 0 and following the same procedure as used in In the following, we show that the optimal allocation for a loop block is obtained if our algorithm exits after the initialization phase 0 . Furthermore, we show that if the second and third phase are needed, their results always improve that from 0 . be the energy consumption by loop block B p l and E 0 be the energy consumption by l assuming that the loop variables do not loop back. Then, minE minE 0 . can be computed by setting in (19) P (v; R to zero for each loop variable v. Then, E can be expressed as r l )(e R l ))(e R where VL is the set of loop variables. It is not dicult to verify that E E 0 for any combination of x(n(R i values. Hence, minE minE 0 . 2 Based on Lemma 1, we can prove the following theorem. Theorem 1 If the allocation result from the initialization phase ( 0 ) makes the in-assignment of each loop variable to be the same as its post-assignment, then Proof: Consider a loop variable v in B p l . If its in-assignment and post-assignment are both memory, we have x(n(R i contribution to E in (31) reduces to zero. Similarly, if v's in-assignment and post-assignment are both register R i , we have v's contribution to E in (31) becomes zero. When more than one loop variable exists, the same conclusion can be obtained. Therefore, if the in-assignment of each loop variable is the same as its post-assignment, we have Ej 0 is the total energy consumption of B p l using the same allocation as that obtained from 0 . Since minE minE 0 by Lemma 1, we obtain According to Theorem 1, if the in-assignment of each loop variable is the same as its post- assignment, the allocation result from 0 is the optimal one for the loop block, and no more allocation phases are needed. When the in-assignment of some loop variable diers from its post-assignment, the result from 0 is no longer optimal. The reason is that the problem formulation used in 0 fails to account for the extra energy incurred by the loop variable switching its assignment at the loop boundary. Our second and third phases aim at reducing such energy. In the second phase 1 , the in-assignment of each loop variable is assumed to be the same as that obtained from 0 , and the best allocation based on this assumption is identied. Let Ej 0 be the energy consumption of B p l when the allocation result from 0 is used, and Ej 1 be the energy consumption of B p l when only the in-assignments of loop variables obtained from 0 are used. After 1 , we obtain an allocation result having minfEj 1 g as the energy consumption. Note that Ej 0 is simply one of the possible values of Ej 1 . Hence, minfEj 1 . That is, 1 always improves the result from 0 . In the third phase 2 , we x the post-assignments of loop variables to the values obtained from 0 and nd the corresponding best allocation. Similar to the statement made for 1 , the allocation result from 2 also improves that of 0 . If more allocation phases are carried out based on the better allocation from 1 and 2 , further improvement may be obtained. Our approach is essentially a hill-climbing technique and could settle to a suboptimal solution. However, observe that our incremental improvement is based on the optimal allocation for the loop body itself and that there are usually only a small subset of the variables in a loop are loop variables. Therefore, our approach tends to have a good chance to obtain the optimal allocation. Our experimental results also support this observation. We use the example in Figure 5 to illustrate how our algorithm works for a single-block loop. Figure 7(a) depicts the allocation result after the initialization phase. Since the in-assignment and post-assignment for loop variables e and f are dierent, the second and third phase are needed. In the second phase, cost c(n f;2 (e); F ) of the ow graph in Figure 6 is set to e M w e R r as the in-assignment of e is memory, while cost c(n f;2 (f); F ) is set to zero as the in-assignment of f is register. The allocation result after the second phase is shown in Figure 7(b). In the third phase, we let P since the post-assignment of e is register R 1 , and P (f; R; B p l since the post-assignment of f is memory. The allocation result after the third phase is the same as that after 1 . Since the in-assignment of e (resp., f) is the same as its post-assignment in the allocation result, our process stops. It is obvious that the allocation result is optimal for this block. Our algorithm can be readily extended to handle the case when there are more than one block inside a loop body, i.e., the control structure inside a loop body contains branches and loops. For branches and merges, each phase in our algorithm needs to solve the allocation problem for each block in the loop body following the process discussed in Section 3 and 4. In 1 (resp., 2 ), the in-assignments (resp., post-assignments) of loop variables obtained from 0 are used to compute the necessary cost or probability values. For loops inside a loop, the sequence of the allocation is determined by the nested levels of loops. Since inner loops are executed more frequently than outer loops, the inner-most loop is allocated rst and the result is used by the outer loops. (a) (b) e a e c d f f e a e c d f f Figure 7: The allocation results of dierent allocation phases for a loop. The solid lines correspond to variables in registers, while the dash lines correspond to variables in memory. 6 Experimental Results and Discussions Graph-coloring based techniques such as the ones in [6] are most often used for solving the register allocation problem. We compare our algorithm with a graph-coloring approach. We have implemented a graph-coloring register allocator and our block-by-block register allocator in the Tiger compiler presented in [1]. The modularity of the Tiger compiler makes it easy to x other phases of the compiler, while only alternate the register allocation algorithms. The procedures for nding the basic blocks, building the trace, and computing the lifetime of each variable are kept the same in both allocators. In the implementation of the graph-coloring allocator, an interference graph is built for each program. In applying the graph-coloring approach to color the graph, if there is no node with degree less than the number of available registers in the graph, the node with the highest degree will be spilled until all the nodes are colored or spilled. In the implementation of our algorithm, a network ow graph is built for each block. Then our algorithm for dierent control structures is applied to each basic block in the topological order to nd the best allocation. We compare the allocation results of the two dierent approaches for several real programs containing complex control structures. The numbers of memory references by the two allocators are summarized in Table 3. The improvement in the energy consumption and the number of memory references by using our algorithm over the graph-coloring approach is summarized in Table 4. According to data in [3, 17, 12], it is reasonable to assume that the ratio of the average energy for memory reference over the average energy for register reference, e M =e R , is in the range of 10 to 30. In Table 5, we summarize the sizes of allocation problems that the two approaches need to solve. As stated at the beginning of the paper, the complexity of our algorithm is O(b l(n; k)), where b is the total number of blocks and l(n; k) is the time complexity for solving the register allocation problem for a single block that has n variables and k available registers. The function l(n; is either O(nlog n) for the single-read static-energy model, O(knlog n) for the multiple-read static-energy model, or O(kn 2 ) for the activity based energy model [5, 13]. Table 3: The number of memory references allocated by two approaches. # of # of Our approach Graph-coloring Example k register data # of spilled # of memory # of spilled # of memory Programs candidates references candidates references candidates references branch 14 28 67 Table 4: The improvement of results by our algorithm over graph-coloring Energy improvement (%) Improvement of # of Programs e M =e factorial 72.8 80.7 89.6 29.8 branch 68.2 77.0 87.4 23.9 Table 5: The size of problems solved by two approaches. Interference graph Network ow Benchmarks # of nodes # of edges # of blocks b max #of variables n branch 28 293 4 26 The experiment results show that our algorithm can achieve more than 60% improvement in the energy consumption due to data references of a program over a graph-coloring approach. At the same time we also saved more than 20% of the memory references which will save the execution time due to data references of a program. Furthermore, our approach is simple to use and the running time is polynomial. There are some other improved graph-coloring based algorithms, such as the one in [11]. The register allocation results by those improved algorithms are better than the simple graph-coloring approach implemented here. But again, those algorithms are all proposed to optimize the execution time, not the energy consumption, of a program. Our algorithm proposed in this paper does not take into account any look ahead information while doing register allocation. It also does not allow any backtracking. In some cases, our approach may produce suboptimal allocation results. One simple extension of the algorithm would be to change the cost c(n f (v); F ) associated with the arc a(n f (v); F ) for a cross boundary variable v according to the possible allocation results of v in child blocks. In this paper, we proposed a new algorithm to deal with those variables whose lifetimes extend beyond basic blocks for low energy register allocation. Our algorithm has much less time and space complexity than other approaches in doing register allocation for variables extending beyond basic blocks. The experimental results by using our algorithm so far are very promising. More experiments with larger code sizes are being conducted. We are also investigating on how to deal with procedure calls. Acknowledgment This research was supported in part by the National Science Foundation under Grant CCR- 9623585 and MIP-9701416, and by an External Research Program Grant from Hewlett-Packard Laboratories, Bristol, England. --R Modern Compiler Implementation in C Digital Circuits with Microprocessor Applications Computer Architecture: A Quantitative Approach Networks and Matroids Integer and --TR Combinatorial optimization: algorithms and complexity On the Minimization of Loads/Stores in Local Register Allocation The priority-based coloring approach to register allocation Register allocation via hierarchical graph coloring A global, dynamic register allocation and binding for a data path synthesis system Allocation algorithms based on path analysis On the <italic>k</italic>-coloring of intervals Register allocation and binding for low power Power minimization in IC design Demand-driven register allocation Low energy memory and register allocation using network flow Optimal and near-optimal global register allocations using 0MYAMPERSANDndash;1 integer programming The design and implementation of RAP Computer architecture (2nd ed.) Global register allocation for minimizing energy consumption Modern Compiler Implementation in C Digital Circuits with Microprocessor Applications Global Register Allocation Based on Graph Fusion Register allocation MYAMPERSANDamp; spilling via graph coloring

low energy;register allocation

510962

How to Choose Secret Parameters for RSA-Type Cryptosystems over Elliptic Curves.

Recently, and contrary to the common belief, Rivest and Silverman argued that the use of strong primes is unnecessary in the RSA cryptosystem. This paper analyzes how valid this assertion is for RSA-type cryptosystems over elliptic curves. The analysis is more difficult because the underlying groups are not always cyclic. Previous papers suggested the use of strong primes in order to prevent factoring attacks and cycling attacks. In this paper, we only focus on cycling attacks because for both RSA and its elliptic curve-based analogues, the length of the RSA-modulus n is typically the same. Therefore, a factoring attack will succeed with equal probability against all RSA-type cryptosystems. We also prove that cycling attacks reduce to find fixed points, and derive a factorization algorithm which (most probably) completely breaks RSA-type systems over elliptic curves if a fixed point is found.

Introduction The theory of elliptic curves has been extensively studied for the last 90 years. In 1985, Koblitz and Miller independently suggested their use in cryptography [9, 19]. After this breakthrough, elliptic curve-based analogues of RSA cryptosystem were proposed [10, 4]. RSA-type systems belong to the family of public-key cryptosystems. A public-key cryptosystem is a pair of public encryption function fK and a secret decryption K indexed by a key K and representing a permutation on a finite set M of messages. The particularity of such systems is that given the encryption function fK , it is computationally infeasible to recover f \Gamma1 K . Moreover, it might be suitable that the encryption function does not let the message unchanged, i.e. given a message m 2 M, we want that fK (m) 6= m. This is known as the message- concealing problem [3]. Simmons and Norris [29] exploited this feature for possibly recovering a plaintext from the only public information. Their attack, the so- Technical Report No. TI-35/97 Technische Universit?t Darmstadt November 1997 called cycling attack, relies on the cycle detection of the ciphertext. This was later generalized by Williams and Schmid [31] (see also [7, 1]). There are basically two ways to compromise the security of cryptosystems. The first one is to find protocol failures [20] and the other one is to directly attack the underpinning crypto-algorithm. The cycling attack and its generalizations fall into the second category. So, it is important to carefully analyze the significance of this attack. For RSA, Rivest and Silverman [25] (see also [16]) concluded that the chance that a cycling attack will succeed is negligible, whatever the form of the public modulus n. For elliptic curve-based systems, the analysis is more difficult because the underlying group is not always cyclic. We will actually give some results valid for groups of any rank, but we will mainly dwell on the security of KMOV and Demytko's system. The paper is organized as follows. In Section 2, we review KMOV and Demytko's system. We extend the message-concealing problem to elliptic curves in Section 3. Then, we show how this enables to mount a cycling attack on KMOV and De- mytko's system in Section 4. We explain how the secret factors can be recovered thanks to the cycling attack in Section 5. Finally, in Section 6, we give some concluding remarks in order to help the programmer to implement "secure" RSA-type cryptosystems. 2. Elliptic curves pq be the product of two large primes p and q, and let two integers a; b such that gcd(4a 3 curve En (a; b) over the ring Zn is the set of points (x; y) 2 Zn \Theta Zn satisfying the Weierstra- equation together with a single element On called the point at infinity. be an elliptic curve defined over the prime field F p . It is well known that the chord-and-tangent rule [17, x 2.2] makes E into an Abelian group. Algebraically, we have: (i) O p is the identity element, i.e. 8P (ii) The inverse of with \Gammay 2 otherwise. The points of En (a; b) unfortunately do not form an Abelian group. But writing e En (a; b) for the group given by the direct product e since En (a; b) ae e En (a; b), we can "add" points of En (a; b) by the chord-and-tangent rule. For large p and q, the resulting point will be a point of En (a; b) with high probability [10]. It is useful to introduce some notations. Let is defined, [k]P will denote P+P+ \Delta \Delta b). The x-coordinate of P will be denoted by x(P). Moreover, since p 2 (the y-coordinate of P) is not required to compute the x-coordinate of [k]P, we will write [k] x p 1 for x([k]P). We can now define an analogue of RSA. The public encryption key e is chosen relatively prime to and the secret decryption key d is chosen according to ed j 1 (mod Nn ). To encrypt a point P 2 En (a; b), one computes the ciphertext [e]P. Then, the authorized receiver recovers P by computing his secret key d. The only problem is to imbed messages as points on a given elliptic curve without the knowledge of the secret factors p and q. A first solution was proposed by Koyama, Maurer, Okamoto and Vanstone [10]. Another one was later proposed by 2.1. KMOV KMOV cryptosystem uses a family of supersingular elliptic curves of the form The main property of this system is that if p and q are both congruent to 2 mod 3, then whatever the value of parameter b. Therefore, to encrypt a message chosen according to and the ciphertext is given by [e]M over the curve En (0; b). The plaintext is then recovered by Another possibility is to work with elliptic curves of the form En (a; and q both congruent to 3 mod 4. The first system based on En (0; b) with p; q j 2 (mod will be referred as Type 1 scheme, and the second one based on En (a; with systems were extended by Kuwakado and Koyama to form-free primes [12]. 2.2. Demytko's system Demytko's system uses fixed parameters a and b. It has the particularity to only make use of the x-coordinate of points of elliptic curves. It relies on the fact that if a number x is not the x-coordinate of a point on an elliptic curve E p (a; b), then it will be the x-coordinate of a point of the twisted curve defined as the set of points (x; y) satisfying is a fixed quadratic non-residue modulo p, together with the point at infinity. So, Nn is given by message m is encrypted as [e] x m. Then, m is recovered from the ciphertext c by For efficiency purposes, the original scheme (see [4]) was presented with message- dependent decryption keys. The length of the decryption key is divided by a factor of 2, on average. However, in the sequel, we will use the message-independent description because this simplifies the analysis, and because we are not concerned with efficiency issues. 3. Concealing-message problem In [3], Blakley and Borosh showed that there are always at least 9 messages that are unconcealable (i.e. the ciphertext of a message is exactly the same as the cleartext) for any RSA cryptosystem. Though this problem is well-known for RSA, nothing appears in the literature about its elliptic curve-based analogues. Since unconcealed messages must be avoided, effective criteria are needed for evaluating the concealing power of these latter systems. Before analyzing the number of unconcealed messages for elliptic curve-based systems, we will first give some general group-theoretic results. G be an Abelian (multiplicatively written) finite group of order #G. Consider the map G; x 7! x k . Then - k permutes the elements of G if and only if gcd(k; Theorem 1 Let G be an Abelian (multiplicatively written) finite group of rank r whose generators are G; x 7! x k permutes the elements of G, then - k has exactly Fix(G; r Y fixed points. Proof: Write rg. So, Each equation has solutions. There are thus fixed points by the permutation map - k . Let p and q be distinct primes and let unconcealed message on RSA, we mean a message m 2 Zn so that m e j m (mod n) for a fixed integer e latter condition ensures that the exponentiation by e is a permutation map, or equivalently that RSA encryption is a permutation of Zn . pq be the RSA-modulus and let e be the RSA-encryption key. Then, the number of unconcealed messages for RSA is given by Proof: Since Z since 0 is always solution to x e j x (mod p), Theorem 1 tells that there are (gcd(e \Gamma fixed points in Z p . Moreover, since by Chinese remaindering, the proof is complete. Note that since p, q and e are odd integers, there are at least 9 unconcealed messages for the original RSA system. If we exclude to encrypt 0 and \Sigma1 (that are always unconcealable messages), there are at least 6 unconcealed messages. An elliptic curve E p (a; b) over the prime field F p is an Abelian group of rank 1 or 2 and of type (n Theorem 2.12]. Therefore, we can 1. If we call x-fixed point a point such that, when given an integer k, Theorem 1 becomes: Theorem be an elliptic curve over the prime field F p . If permutes exactly fixed points. Furthermore, - k has exactly Fix x x-fixed points, where - 2 is the number of points of order 2. Proof: The first part follows immediately from Theorem 1. b). P is a x-fixed point if and only if we \Psi , we have ae Since O p and points of order 2 are counted twice, we obtain Eq. (11). Indeed, KMOV Type 1 scheme is based on elliptic curves of the form E and 3). The underlying group is isomorphic to the cyclic group Z p+1 . Type 2 scheme uses curves of the form E In that case, the underlying group is also isomorphic to Z p+1 if a is a quadratic residue modulo p; and it is isomorphic to Zp+1\Phi Z 2 otherwise. From Eq. (10), for an odd we see that, for a given KMOV elliptic curve E p (a; b), there are at least 2 fixed points if E p (a; b) is cyclic and at least 4 fixed points otherwise. These points correspond to the point at infinity together with the points of order 2. Noting that the encryption key e is always odd for KMOV, and since the point at infinity is not used to represent messages, there are at least 1, 3 or 9 unconcealed messages on a given KMOV elliptic curve En (a; b). Consequently, the probability that a random message is unconcealed can be at least 1=n. This has to be compared with 6=n for the original RSA. Demytko's encryption works in a group of the form G (i) p \Theta G (j) where G (1) G (2) Writing G (i) , we define and similarly for G (j) q . Demytko's system only makes use of the x-coordinate. So, since the point at infinity is never used for encryption, Theorem 2 indicates that there are messages. 2 This number may be equal to 0, and we cannot give general information on the minimal number of unconcealed messages in Demytko's system. For efficiency purposes, the public encryption key e is usually relatively small (for example, are common choices). In all systems, the number of unconcealed messages depends on expressions of the form . Therefore, the maximal number of unconcealed messages is mainly bounded by (e \Sigma 1) 2 . So, if the encryption key is equal to 2 then the probability that a message is unconcealed is at most - 10 \Gamma144 for a 512- bit RSA-modulus and - 10 \Gamma299 for a 1024-bit RSA-modulus. Even if the number of unconcealed messages is small, we will see in the next section how this can be turned into an active attack. 4. Cycling attack 4.1. Previous results on RSA be the ciphertext corresponding to message m, where (e; n) is the public key. If we find an integer k that satisfies the equation then we can obviously recover the plaintext m by computing mod n. Note that we do not have to factor the public modulus n, so this might be a serious failure for the RSA cryptosystem. This attack, firstly proposed by Simmons and Norris [29], was later extended by Williams and Schmid [31] (see also [7]) in the following way. Let P(t) be a polynomial. They showed that if the ciphertext c has a period such that for some integer g, then the plaintext m can be recovered. 4.2. Generalizing the cycling attack We can generalize the results of the previous paragraph to any Abelian finite group G. Theorem 3 Let G be an Abelian (multiplicatively written) finite group. Let a message e be the corresponding ciphertext, where gcd(e; 1. 3 If we find an integer P such that c G, then the plaintext m can be recovered by computing is the smallest integer such that m G. By Lagrange's Theorem, t j #G and since gcd(e; implies that t j eP and thus t j P . Therefore, Z such that letting and c We call this theorem the generalized cycling attack. This theorem indicates that KMOV and Demytko's system are also susceptible to the cycling attack. Detecting the integer P is equivalent to the problem of finding a polynomial P(t) and an integer P(g). Moreover, the relation c equivalent to If i denotes the prime decomposition of group order #G and since (c) divides #G, Eq. (16) can be reduced to for all primes p i dividing ordG (c). Here, we must check that these relations hold by picking up a random polynomial P(t) and a random integer g. This means that the cycling attack depends on the distribution of such polynomial and of the order of ciphertext c. Roughly speaking, if the order of G is smooth, we can expect that there are many elements c 2 G with small order. So, primes p i in Eq. (17) will be small, and polynomial P will be more easily found. Consequently, it might be desirable to impose that #G contains at least one large prime in order to make harder the cycling attack. We will now analyze in more details this assumption for elliptic curve-based systems. 4.3. Application to elliptic curve systems As previously mentioned, an elliptic curve E p (a; b; ) over the prime field F p is not necessarily cyclic, but isomorphic to Zn1 \PhiZ Therefore, for analyzing the cycling attack over elliptic curves, we have to estimate the number of points in E p (a; b) of a given order. If is a cyclic group), then the number of elements of order d is given by the Euler's totient function, namely OE(d). For the general case, we have: Proposition be an elliptic curve over the prime field F p . If we then the number of elements of order d is equal to Y where\Omega d;n2 is the set of primes is the power of p i which appears in the prime decomposition of n. Furthermore, given the prime factorization of gcd(#E p (a; can be computed in probabilistic polynomial time. Note that then we take Proof: The first part of the proposition is proved in Appendix A. The second part follows from Miller's probabilistic polynomial time algorithm for finding n 1 and n 2 (see [17, x5.4]). We can now derive a lower bound on the number of elements whose order is divisible by a large prime factor of the order of E p (a; b). Proposition be an elliptic curve over the prime field F p . Suppose that #E p (a; b) is exactly divisible by a prime factor l p . If F div (l p ) denotes the number of elements of order divisible by l p , then l p Proof: See Appendix B. This proposition indicates that if we randomly pick up an element in E p (a; b), it has order divisible by l p with probability OE(l p )=l . When l p is large, this probability is non-negligible (i.e. really "nearly 1"). RSA-type cryptosystems over elliptic curves are constructed on groups of the form En (a; b), which can be considered as E p (a; b) \Theta E q (a; b) by Chinese remaindering. In the sequel, we will suppose that #E p (a; b) (resp. #E q (a; b)) contains a large prime factor l p (resp. l q ). With high probability, a random point will have order divisible by l p (resp. l q ). Therefore a random point P on En (a; b) (represented by by Chinese remaindering) will have order divisible by l p and l q with high probability. As we discussed in Paragraph 4.2, the cycling attack is reduced to find a polynomial P and an integer g with c this attack becomes "Find a polynomial P and an integer g so that [P(g)]C = On for some ciphertext C 2 En (a; b)". Equivalently, this can be formulated by an expression of the form of Eq. (17). Since the order of ciphertext C is supposed to be divisible by l p and l q with high probability, we must have to mount a successful cycling attack. Williams and Schmid [31] estimated that these relations are rarely fulfilled except when k. So, we have thus to take care whether or not and similarly for prime q. Letting ord l p (e) for the smallest integer satisfying Eq. (20), k must be a multiple of ord l p (e). Consequently, the cycling attack will be useless if ord l p (e) is large. Note 1 In his fast generation algorithm of secure keys, Maurer [15] suggested to verify that e (l p \Gamma1)=r ip 6j 1 (mod l p) for p is the prime decomposition of l p \Gamma 1. This criteria implies that ord l p (e) must be large and the cycling attack is not applicable. Another method is to impose that l contains a large prime factor rp . The probability that ord l p (e) is divisible by rp will be then Proof: Let l be the prime decomposition of l p \Gamma 1. The number of elements in Z l p whose order divisible by rp is given by rp rp rp l p . This is known as the strong primes criteria. Through this section, we have proven some conditions to preclude cycling attacks. Putting all together, we have: Theorem 4 The cycling attack does not apply against KMOV if the secret prime p has the following properties: (i) p+1 has a large prime factor l p , and (ii) ord l p (e) is large; and similarly for prime q. Theorem 5 The cycling attack does not apply against Demytko's system if the elliptic curves over F p have the following properties: (i) #E p (a; b) has a large prime factor l p and #E p (a; b) has a large prime factor l 0 are large; and similarly for prime q. 5. Factoring the RSA-modulus 5.1. Relation between unconcealed message and cycling attack For a given ciphertext C 2 En (a; b), the cycling attack detects an integer k satisfying This is equivalent to the message-concealing problem where the message is now a ciphertext instead of a cleartext. If E , from Theorem 2, we know that there are Fix(En (a; b); e k unchanged ciphertexts C via encryption by e k . Moreover, by Eq. (20), [e k yields l p prime l p dividing #E p (a; similarly for prime q. So the number of unchanged ciphertexts C is larger than l p l q . Suppose that primes p and q were chosen so that both #E p (a; b) and #E q (a; b) contain a large prime factor l p and l q , respectively. Then, there may be many ciphertexts C such that [e k and the corresponding cleartexts can be re- covered. This means that a cycling attack is really effective when applicable. To prevent this attack, the designer has also to verify that ord l p (e) (resp. ord l q (e)) is large (see Theorems 4 and 5). 5.2. Factoring by means of fixed points In Section 4, we explained how the cycling attack can recover a plaintext. Here, we will show that the knowledge of a unchanged ciphertext enables still more, i.e. to completely break the system by factoring the RSA-modulus This can be illustrated by the elliptic curve factoring method (ECM) [13] introduced by Lenstra. It can basically be described as follows. Suppose that n is the product of two primes p and q. Consider an elliptic curve En (a; b) over the ring Zn . Assume that #E p (a; b) or #E q (a; b) is B-smooth. Then define and choose a random P 2 En (a; b)-note that [r]P in En (a; b) (and not in E p (a; b) \Theta E q (a; b) because p and q are unknown). As mentioned in Section 2, some points are not "realizable" because En (a; b) is not a group. During the computation of [r]P, at step i, three situations can oc- cur: (i) [r i and [r i In cases (i) and (ii), the denominator of - in the chord-and-tangent formulas (see Eq. (2)) will have a non-trivial factor with n. So n is factored. In case (iii), [r]P is correctly computed, we obtain of n is found and we then re-iterate the process with another point P or with other parameters a and b. Let ord p (P) and ord q (P) be the order of point P in E p (a; b) and E q (a; b), re- spectively. Let - be a prime. We can write ord p Hence, if we know an integer r of the form must have On in En (a; b). If f p 6= f q , or without loss of generality f we define r equivalently and we find a non-trivial factor of n similarly as in ECM. The message-concealing problem or the cycling attack is due to the presence of fixed points P 2 En (a; b) such that message-concealing problem, and for the cycling attack. The knowledge of a fixed point P gives [r \Gamma On . We are then in the conditions of ECM and the RSA-modulus can be factored with some probability as follows. [Step Choose a prime power factor - of r \Gamma 1, i.e. - t [Step 3] Compute [r 0 ]P in En (a; b). If an error occurs (i.e. Eq. (22) is satisfied 4 ), then n is factored. Otherwise, go to Step 2i; if then go to Step 1. The next theorem says more about the probability of factoring the RSA-modulus using one iteration of this method. Theorem 6 Consider KMOV or Demytko's system. Let E and - is prime. Let fl - denotes the probability that F p +F q - 2 know a fixed point P 6= On such that [r \Gamma On and if r \Gamma 1 is divisible by -, then we can factor the RSA-modulus with probability at least fl - Proof: See in Appendix C. Assume for example that 0:5 and that we know a point P such that On . If 2 (which is the most probable case), then our algorithm will find the secret factors of n with probability at least 15%. Otherwise, we re-iterate the algorithm with another prime factor - of r \Gamma 1. 5.3. Remark on efficiency Reconsider the cycling attack [e k n). From Eq. (20), k must be a multiple of both ord l p (e) and ord l q (e) to apply the attack. However, what we ultimately need to factor the modulus n is to find an integer r 0 such that, for example, O q (mod q) (see Eq. (22)); or equivalently, such that q). This means that a cycling attack just modulo p (or modulo q) rather than modulo both primes simultaneously enables to factor n. Therefore, k needs to be just a multiple of ord l p (e) or of ord l q (e), not of both of them. This results in a higher probability of success. 6. Concluding remarks In Section 4, we proved that if the conditions of Theorems 4 and 5 are fulfilled, then cycling attacks are useless for elliptic curve-based RSA systems. This is the elliptic version of the well-known strong primes criteria. For RSA, Rivest and Silverman [25] claimed that this criteria is not required. They said: "Strong primes offer little protection beyond that offered by random primes." We will now analyze more accurately how valid this assertion is, and if it remains valid for elliptic curve-based systems. The analogue of Theorems 4 and 5 for original RSA is: Theorem 7 Let pq be the RSA modulus and let e be the public encryption exponent. The cycling attack does not apply against RSA if the secret prime p has the following properties: (i) has a large prime factor l p , and (ii) l large prime factor r p (cf Note 1); and similarly for prime q. A prime p satisfying conditions (i) and (ii) of the previous theorem is said to be a strong prime. Some authors also recommend that (iii) p + 1 has a large prime factor. Condition (iii) is required in order to protect against the algorithm [30]. In their paper, Rivest and Silverman only consider the primes p and q. They did not take into account the second condition of Theorem 7. 5 Our analysis is based on a previous work of Knuth and Trabb-Pardo [11] (see also [22, pp. 161-163]), whom rigorously calculated the distribution of the largest, second largest, . prime factors of random numbers. Also, they have tabulated: Table 1. Proportion ae(ff) of (large) numbers N whose largest prime factor is - N 1=ff . ff 1.5 2.0 2.5 3.0 4.0 5.0 6.0 8.0 We can now more precisely quantify what "large" means in Theorem 7 in order to prevent cycling attacks. A cycling attack remains to find an integer k such that (mod n) for some ciphertext c, where e is the public encryption key and is the RSA-modulus. From k, the plaintext m corresponding to c is then given by mod n. However, we noticed in x5.3 that it just suffices to mount a cycling attack modulo p (instead of modulo n) to factor the RSA-modulus. For RSA, the secret prime factors are recovered as follows. Suppose that there exists an integer k such that c e k will give p; and hence Knowing p and q, the secret key d is computed as and the plaintext m is then given by From Eqs (17) and (20), if l p denotes the largest prime factor of must be (with probability a multiple of ord l p (e) to apply the cycling attack modulo we thus have k - ord l p (e) with probability at least 1 \Gamma 1=l p . From Knuth and Trabb-Pardo's results, we can derive how does a typical integer k look. We note that an average case analysis makes a sense since the distribution of the largest prime factor, the second largest prime factor, . is monotone. The average size of l p is similarly, the average size of the largest prime factor r p of l (Note that we suppose that l behave like random numbers. This assumption was confirmed by experimental results using the LiDIA package [14]: over 1 000 000 random 100-bit primes l p , 423 were such that smooth number, that is, a proportion of 0:000423 - 10 \Gamma3:37 . This has to be compared with .) The average size of the second largest prime factor r 0 p of l divides (see Note 1), we have with probability at least (1 \Gamma 1=l . For a 512-bit RSA modulus probability is already greater than and is greater than modulus. In summary, we have: Table 2. Lower bound K on a typical value for k such that c e k (mod p) for a t-bit RSA modulus t 512 bits 768 bits 1024 bits Lower Albeit very high, the estimation of the bound K (see Table 2) is quite pessimistic; in practice, k will be much larger than K and a cycling attack (modulo p) will have thus fewer chances to be effective. Indeed, if we take into account the third largest prime r 00 p of l p , we have k - r p r 0 with probability at least - example, for a 1024-bit RSA modulus, we have with probability at least importantly, we only take into account the largest prime factor l p of p be the second largest prime factor of its average size is . The ciphertext c has its order divisible l p l 0 with probability at least (1 \Gamma 1=l . Therefore, from Eq. (17) (see also Eq. (20)), k is very likely (i.e., with probability a multiple of lcm(ord l p (e); ord l 0 (e)). The largest prime factor s p of l 0 has an average size of (l 0 with a probability of at least (1 \Gamma 1=p 0:210 example, for a 1024-bit RSA modulus, we have with probability at least 1 Consequently, k is expected to be very large, and a cycling attack will thus have very little chance to be successful. Hasse's Theorem [27, Theorem 1.1] indicates that #E p (a; b) 2 [p and we can thus consider that #E p (a; O(p) and p have the same bit-size. Therefore, from Theorems 4 and 5, the previous discussion still applies to elliptic curve-based cryptosystems and the conclusion of Rivest and Silverman remains valid, i.e. the use of strong primes offers (quasi) no additional security against cycling attacks. However, as remarked by Pinch [21], a user might intentionally choose a "weak" RSA-modulus. Suppose that a user chooses his public RSA-modulus that a cycling attack is possible. In that case, this user can repudiate a document by asserting that an intruder has discovered by chance (the probability of a cycling attack is negligible) the weakness. If the use of strong primes is imposed in standards [8], such arguments cannot be used for contesting documents in court. In conclusion, from a mathematical point of view, strong primes are not needed, but they may be useful for other purposes (e.g., legal issues). On the other hand, since the generation of strong primes is only just a little bit more time consuming, there is no reason to not use them. Acknowledgments We are grateful to Jean-Marc Couveignes for providing useful comments on a previous version of this work. We also thank Markus Maurer for teaching us how to use LiDIA. Notes 1. OE(n) is the Euler's totient function and denotes the number of positive integers not greater than and relatively prime to n. 2. Note that this expression slightly differs from Eq. (11). This is because Eq. (11) counts the number of x-fixed points; here we have to count the number of x-coordinates that are unchanged by Demytko's encryption. 3. This condition is equivalent to -e in G is a permutation map (see Lemma 1). 4. Or if [r 0 ]P 6= Op (mod p) and [r 0 5. See [25] on p. 17: "Suppose r does not divide ord(e) mod -(N )". Note also the typo, N should be replaced by -(N). 6. This is the probability that l p divides ord Z (c) (see Note 1). --R Factoring via superencryption. Security of number theoretic cryptosystems against random attack A new elliptic curve based analogue of RSA. Strong RSA keys. Strong primes are easy to find. Critical remarks on some public-key cryptosystems International Organization for Standardization. Elliptic curve cryptosystems. New public-key schemes based on elliptic curves over the ring Zn Analysis of a simple factorization algorithm. Efficient cryptosystems over elliptic curves based on a product of form-free primes The LiDIA Group. Fast generation of secure RSA-moduli with almost maximal diversity Fast generation of prime numbers and secure public-key cryptographic param- eters Elliptic curve public key cryptosystems. Handbook of applied cryptography Use of elliptic curves in cryptography. Protocol failures in cryptosystems. On using Carmichael numbers for public-key encryption systems Prime Numbers and Computer Methods for Factorization. Remarks on a proposed cryptanalytic attack on the M. "Critical remarks on some public-key cryptosystems" Are 'strong' primes needed for RSA. A method for obtaining digital signatures and public-key cryptosystems The Arithmetic of Elliptic Curves. Fast generation of random Preliminary comment on the M. Some remarks concerning the M. --TR Strong primes are easy to find Use of elliptic curves in cryptography Fast generation of secure RSA-moduli with almost maximal diversity Prime numbers and computer methods for factorization (2nd ed.) A new elliptic curve based analogue of RSA A method for obtaining digital signatures and public-key cryptosystems Handbook of Applied Cryptography New Public-Key Schemes Based on Elliptic Curves over the Ring Zn On Using Carmichael Numbers for Public Key Encryption Systems

RSA-type cryptosystems;strong primes;cycling attacks;elliptic curves