Enhanced Interior Gateway Routing Protocol, Cisco White Paper

Variable Ordering for E?cient SAT Search by Analyzing Constraint-Variable Dependencies
Vijay Durairaj and Priyank Kalla
Department of Electrical and Computer Engineering University of Utah, Salt Lake City, UT-84112 {durairaj, kalla}@ece.utah.edu

Abstract: This paper presents a new technique to derive an initial static variable ordering for e?cient SAT search. Our approach not only exploits variable activity and connectivity information simultaneously, but it also analyzes how tightly the variables are related to each other. For this purpose, a new metric is proposed - the degree of correlation among pairs of variables. Variable activity and correlation information is modeled (implicitly) as a weighted graph. A topological analysis of this graph generates an order for SAT search. An algorithm called ACCORD (ACtivity - CORrelation - ORDering) is proposed for this purpose. While ACCORD rigorously analyzes constraint-variable dependencies, it does not account for the e?ect of decision-assignments on clause-variable dependencies. This issue motivates further re?nements to our approach using literal activity and correlation measures - giving rise to the L’ACCORD algorithm. Using e?cient implementations of the above, experiments are conducted over a wide range of benchmarks. The results demonstrate that: (i) the variable order generated by our approach signi?cantly improves the performance of SAT solvers; (ii) time to derive this order is a fraction of the overall solving time. As a result, our approach delivers faster performance (often, by orders of magnitude) as compared to contemporary approaches.



Contemporary SAT solvers have matured over the years and come a long way from the classical search procedures of Davis-Putnam (DP) [1] and Davis-LogemannLoveland (DLL) [2]. Recent approaches [3] [4] [5] etc., employ sophisticated methods such as constraint propagation and simpli?cation, con?ict analysis, learning and non-chronological backtracks [3] [4] [5] to e?ciently analyze and prune the search space. In recent past, a lot of e?ort has been invested in trying to understand the nature of the SAT problem. The works that deserve mention relate to symmetry analysis [6] [7], local search strategies [8], complexity of SAT viz-a-viz ATPG [9], relationship of BDD variable orderings and CNF search procedures [10], amount of search space analyzed [11], UNSAT core extraction [12] [13], among others [14] [15] [16]. An important aspect of CNF-SAT is to derive an ordering of variables to guide the search. The order in which variables (and correspondingly, constraints)

are resolved signi?cantly impacts the performance of SAT search procedures. Boolean functions arising in many applications represent some spatial, casual or logical dependencies (or connections) among variables. Therefore, analysis of these clause-variable dependencies provides useful information that can be exploited to guide CNF-SAT search procedures. Variable activity and clause connectivity are often considered as qualitative and quantitative metrics to model clause-variable dependencies. Activity of a variable (or literal) is de?ned as the number of its occurrence among all the clauses of a given SAT problem [17]. Most conventional SAT solvers [4] [5] [3] employ variable/literal-activity based branching heuristics to resolve the constraints. Connectivity of constraints has also been used as a heuristic approach to derive variable orderings for SAT search. Loosely speaking, two clauses are said to be ”connected” if one or more variables are common to their support. Clause connectivity can be modeled by representing CNF-SAT constraints as (hyper) graphs and, subsequently, analyzing the graph’s topological structure. Tree decomposition techniques have been proposed in literature [18] [19] for analyzing connectivity of constraints in constraint satisfaction programs (CSP). Such techniques identify decompositions with minimum tree-width, thus enabling a partitioning of the overall problem into a chain of connected constraints. Recently, such approaches have also found application in those problems that can be modeled as DPLL-based CNF-SAT search [19] [20] [21] [22] [23] [24]. Various approaches operate on such partitioned tree structures by deriving an order in which the partitioned set of constraints are resolved [20] [21] [22] [24]. MINCE [10] employs CAPO placer’s mechanism [25] to ?nd a variable order such that the clauses are resolved according to their chain of connectivity. Bjesse et. al. [21] proposed tree decomposition based approaches to guide variable selection and con?ict clause generation. Aloul et. al. have proposed a fast, heuristic based approach, FORCE [26], as an alternative to the computationally complex approach of MINCE. Recently, Durairaj et. al. [27] [28] [29] proposed hypergraph bi-partitioning based constraint decomposition scheme that employs both variable activity and clause connectivity simultaneously to derive a variable order. This paper presents a new approach to derive an initial static ordering for SAT search by rigorously analyzing constraint-variable dependencies. Experimental results demonstrate that our approach is faster and more robust than the contemporary variable ordering techniques, and it improves the performance of SAT solvers (in many cases by orders of magnitude). The paper is organized as follows. The following Section analyzes the limitation of previous work. Section 3 outlines the speci?c contributions of this research. The variable order generation approaches are presented in Sections 4 and 5, followed by experimental results and analysis in Section 6. Finally, Section 7 concludes the paper.


Limitations of Previous Work

The computational complexity (exponential) of minimum tree-width decomposition algorithms results in large compute times to search for the variable order. The technique of Dechter et. al. [18] is shown to be time exponential in the treewidth. Algorithms for approximating tree-width with bounded error [23] are also shown to be too costly for industrial problems [21]. As a result, these techniques are impractical for large/hard CNF-SAT problems. In general, the time required to derive a variable order should be small as compared to the subsequent SAT solving time - it should certainly not exceed the solving time. Unfortunately, for large and hard SAT problems, it has been observed that MINCE [10] [30] and Amir’s tool [20] [31] require unacceptably long time just to derive the variable order. This behaviour is depicted in Table 1, which compares the time required to derive the variable order by Amir’s tool [31], MINCE [30], FORCE [26] and the hypergraph partitioning based technique (HGPart) of [27] against zCHAFF’s (2003 version) solve time. It can be observed from the table that in order to derive the variable order, both Amir’s and MINCE approach su?er from long compute times - much longer than the default SAT solving time. In contrast, FORCE and HGPart can derive the variable order much faster than the other two. This clearly demonstrates the computational limitations of Amir’s and MINCE approach; as such they are too expensive to be applicable for practical SAT problems. While, FORCE is fast, the quality of the variable order does not consistently improve the performance of SAT solvers.
Table 1. Comparison of original zCHAFF runtime, Amir’s, MINCE’s, FORCE’s and HGPart’s partitioning time Benchmark c2670 c3540 c5315 c7552 4pipe 5pipe Vars/ zCHAFF’03 Amir Clauses (sec) (sec) 2.5K/6.4K 1.48 16.54 3.4K/9.2K 20.57 23.66 5.0K/14.1K 34.62 54.56 5.5K/15.1K 105.97 71.30 5.2K/80.2K 129 505.83 9.5K/195K 196.53 >2000 MINCE FORCE HGPart (sec) (sec) (sec) 8.23 0.94 1.15 13.95 0.83 1.42 25.51 6.52 1.25 24.43 3.81 1.76 93.1 1.36 13.84 267.2 2.82 38.40

Table 2. Quality of variable order derived by FORCE BencVars/ zCHAFF’03 FORCE hmark Clauses (sec) (sec) c3540 3431/9262 20.57 36.55 c5315 4992/14151 34.62 9.25 c7552 5466/15150 105.97 136.55 4pipe q0 k 5380/69072 66.71 87.78 5pipe q0 k 10026/154409 649.9 375.37 clus set1 1200/4800 59.62 >1000

Table 2 presents some results that re?ect the inconsistency in the quality of the variable order derived by FORCE. Since MINCE and Amir’s tool have been shown to be impractical (too slow), we analyze the variable order derived by FORCE. The order derived by FORCE was given to the zCHAFF solver (2003 version) to resolve constraints - the runtime is given in Column 4. Column 3 corresponds to original zCHAFF runtime (VSIDS heuristic). From the results, it can be observed that the the variable order generated by FORCE does not consistently enhance the performance of the SAT solver1 . In fact, for some benchmarks, it degrades the SAT solver performance by an unacceptably long time. Recent approach of [27] has shown to deliver good results. Not only can the variable order be derived in a reasonable amount of time, the quality of the order also consistently improves the performance of the solver. However, as this technique relies on hypergraph partitioning, there is no direct control over the decomposition. As a result, many problems (such as the NQueens) cannot be decomposed e?ciently. Another limitation of these techniques is that while they do analyze variable activity, clause connectivity or both, however, they do not analyze how tightly the variables are connected to each other. As a result, tightly connected, hard problems may not realize the run-time improvements.


Research Contributions

As discussed earlier, e?cient CNF-SAT decision heuristics should analyze both variable activity and connectivity simultaneously for faster constraint resolution. Moreover, they should also analyze how tightly the variable are related to each other. Furthermore, the procedure to generate such an order should be fast, scalable and robust enough to handle large SAT problems. This paper proposes an e?cient variable order generation procedure that attempts to ful?ll the above criteria/requirements. The contributions of our work are as follows: 1. We present an e?cient technique that analyzes constraint-variable dependencies to derive an ordering of variables to guide SAT diagnosis. Both variable activity and connectivity information is exploited to derive the order. 2. In order to analyze how tightly two or more constraints are related (in terms of common variables), we propose a new metric called the degree of correlation among pairs of variables. 3. Variable activity and correlation information is (implicitly) modeled as a weighted graph, the topological analysis of which enables the derivation of such an order. An e?cient algorithm (ACCORD) is described for this purpose. 4. In order to analyze the e?ect of decision-assignments made by SAT solver on the variable ordering, further re?nements to the ACCORD algorithm are proposed.

Mince also exhibits similar phenomenon.

5. Experiments conducted over a large and varied set of benchmarks demonstrate that our approach is fast, robust and scalable; quality of the variable order is re?ected in the impressive speed-up achieved for SAT solving. Particularly for hard instances, orders of magnitude speed-up is achieved in many cases. 6. The variable order generation time is small as compared to the overall solving time.


Variable Order Computation: Analyzing Constraint-Variable Dependencies

It is our desire to derive a variable order for SAT search by analyzing clausevariable relationships. The importance of branching on high activity variables is well understood [17]. Analyzing the connectivity of constraints is also important for constraint resolution. Contemporary techniques address both of the above issues. However, ‘how tightly are the variables related?’ - this feature too should not be overlooked. For this purpose, we propose a metric that measures how tightly the variables are connected/related. We de?ne this metric as follows: De?nition 41 Two variables xi and xj are said to be correlated if they appear together (as literals) in one or more clauses. The number of clauses in which the pair of variables (xi , xj ) appear together is termed as their degree of correlation. 4.1 Problem Modeling

In our approach, the constraint-variable relationship of a given CNF-SAT problem is modeled as a weighted graph. The variables (as opposed to literals) form the vertices, while edges denote the correlation/connectivity between them. Associated with each variable is its activity, which is modeled as an integer value within the node. The edge weights represent the degree of correlation between the variables/vertices. For example, if two variables xi , xj appear together in n clauses, then the weight of the edge eij connecting them is n. An ordering of the nodes (variable order) can be performed by analyzing the graph’s topology. We now describe our approach by means of an example corresponding to the CNF-SAT problem shown in Fig. 1. Its corresponding weighted graph is depicted in Fig. 2(a). 4.2 ACCORD: Activity-Correlation based Ordering

We begin the search by ?rst selecting the highest active variable, i.e. the node that has the highest internal weight. The variable is marked and the node is added to a set called supernode. The variable is also stored in a list (var ord list). The SAT tool should branch on this variable ?rst. It can be observed from the graph (weight within the nodes) that the activity of variables {e, u, g, v} is the

Fig. 1. An example CNF-SAT problem

Fig. 2. Weighted Graph - Edge weights denote the degree of correlation between the variables (vertices). Variable activity depicted within the node.

highest (= 5). We select one of these variables (randomly or lexicographically) as the initial branching variable. Let us select the variable u, shown in Fig. 2(a) as “First Selection.” The variable u becomes the supernode. Now we need to identify the set of variables connected to this supernode. All the nodes that share at least one edge with the supernode are identi?ed. The nodes {g, c, a, b} are connected to the supernode u, as shown in Fig. 2(b). One of these variables {g, c, a, b} is to be selected as the next branching variable. We consider the degree of correlation, modeled as edge weights, as the metric to identify this variable. The nodes that share an edge with the supernode are sorted in decreasing order of the sum of their degrees of correlation with the nodes in the supernode. In our example, variable c has the degree of correlation = 1. On the other hand, the variables {g, a, b} have the same degree of correlation with node u, which is equal to 2; hence, we consider variables {g, a, b}. In order to break this tie, we further distinguish these variables according to their activity. Since the activity of variable g is higher than that of variables {a, b}, g is selected as the next branching variable. This is shown in Fig. 2(b) as “Second Selection.” Moreover, the node g is added to the current supernode set; supernode = {u, g}. The variable order derived so far is var ord list = {u, g}. As the next step, those variables are identi?ed that share an edge with the supernode. These variables are {v, e, c, a, b} as shown in Fig.2(c) with highlighted incident edges. Consider the nodes {v, e, a, b}. They share only one edge with the supernode with weights (correlation) = 2. On the other hand, node c has two edges incident on the supernode. Therefore, the correlation modeled by both edges should be accounted for. This is computed as the sum of the degrees of correlation between c and the supernode {g, u} = 2+1 = 3. Since this sum is the highest for c, it is selected as the next branching variable. This is shown in Fig. 2(c) as third selection. The supernode set and the var ord list are correspondingly updated. The above procedure is repeated until all the nodes are ordered and included with in the supernode. The ?nal derived variable order is {u, g, c, v, e, a, b, w, d}. Our approach is inspired from the Prim’s Minimum Spanning Tree (MST) algorithm [32]. We have named it the ACCORD (ACtivity COrrelation ORDering) algorithm. The Pseudo code for the ACCORD algorithm is presented in Algorithm 1. The Variable-Clause database, mentioned in lines 1-3, is implemented using the array of arrays data-structure available within contemporary SAT solvers [4] [33]. For example, the Variable-Clause database corresponding to the CNF-SAT problem of Fig 1 is shown in Fig. 3. The variable a appears in Clauses {1, 2, 7, 9}. Hence, the ?rst element of Variable-Clause database array is an array containing {1, 2, 7, 9}. Similarly, the second element of the array is an array containing {1, 3, 8, 9} corresponding to variable b; and so on. The algorithm analyzes the clause-variable database, and for each variable, it computes its correlated variables. Using the activity and correlation measures, the variable ordering is computed.

Algorithm 1 Pseudo code for ACCORD
1: INPUT = CNF Clauses and Variable-Clause Database ; 2: /* Variable-Clause database is implemented as array of arrays */ 3: /* For each variable xi , the Variable-Clause database contains a list of clauses in which xi appears */ 4: OUTPUT = Variable order for SAT search ; 5: activity list = Array containing activity of each variable ; 6: var order list = Initialize variable order according to activity ; 7: connectivity list = Initialize to zero for all variables ; 8: int i = 0; 9: while i != number of variables do 10: /* Implicitly, supernode = {var order list[0], . . . , var order list[i]} */ 11: next var = var order list[i] ; 12: correlation list = ?nd variables connected to next var using Variable-Clause database ; 13: for all var ∈ correlation list do 14: connectivity list[var]++ ; /* Compute correlation */ 15: adjust variable order(var) ; 16: /* Linear sort is used to update the variable order corresponding to both connectivity list as well as activity list */ 17: end for 18: i = i+1; 19: /* At this point, {var order list[0], . . . , var order list[i]} is the current variable order*/ 20: end while 21: return(var order list) ;

Fig. 3. Variable-Clause Database


Complexity of ACCORD

A weak upper bound for the ACCORD algorithm can be found by assuming that every variable appears in every other clause. With this assumption, an upper bound on the time complexity of ACCORD can be derived as O(V ·(V ·C +V 2 )), where V represents the number of variables and C represents the number of clauses.


L’ ACCORD: Overcoming the Limitations of ACCORD

Most conventional SAT solvers employ literal-activity based branching heuristics to resolve the constraints. Also, when the variable order generated by ACCORD is given to a SAT solver, the solver will branch on a literal corresponding to the variable selected by the ACCORD. Clauses corresponding to this literal will be satis?ed due to this assignment. ACCORD does not analyze this e?ect of decision-assignments on the variable correlation. Consider the following set of clauses. (x + a + e)(x + b + f )(x + c + e) (x + d + f )(x + c) Here, the literal x has the highest activity and the SAT solver will assign x = 1. This satis?es the ?rst three clauses. The ACCORD algorithm will consider the variables {a, e, b, f, c, d} to compute their respective degrees of correlation with respect to x. However, due to the assignment x = 1, it should su?ce to look at only those clauses in which x appears, as they are unsatis?ed. This means that the correlated variables to x ({d, f, c}) and their corresponding constraints should be analyzed to derive a variable order. In order to exploit such a behaviour, we have implemented the above modi?cations to the ACCORD algorithm. Instead of analyzing variable-to-variable correlations, we now analyze correlations between a literal and its connected variables. The modi?ed algorithm is now called L’ ACCORD.


Experimental Results and Analysis

The ACCORD and L’ ACCORD algorithms have been programmed within the zCHAFF [4] solver using its native data-structures. The algorithms analyze the constraint-variable relationships of the given CNF-SAT problem and derive a variable order for SAT search. Using this as the initial order, the SAT tool (zCHAFF) performs a search for the solutions. On encountering con?icts, we allow the solver to add con?ict-induced clauses and proceed with its book-keeping and (non-chronological) backtracking procedures. In other words, ACCORD/L’ ACCORD provide only an initial static ordering. zCHAFF’s VSIDS heuristic updates this order dynamically, when con?ict clauses are added. Hence, our approach is not a replacement for VSIDS; it is to be used in conjunction with it.

Using this setup, we have conducted experiments over a large set of benchmarks that include: i) Microprocessor veri?cation benchmarks [34]; and ii) some of the hard instances speci?cally created for the SAT competition (all three categories - industrial, handmade and random). We conducted our experiments on a Linux workstation with a 2.6GHz Pentium-IV processor and 512MB RAM. 6.1 ACCORD versus L’ ACCORD comparison

Table 3 compares the quality of orders generated by ACCORD and L’ ACCORD. It is clear from the table that L’ ACCORD generates a better quality variable order than ACCORD, always improving the SAT solver performance.
Table 3. Run-time Comparison of ACCORD and L’ ACCORD zCHAFF ACCORD + zCHAFF L’ ACCORD + zCHAFF Solve Var. Solve Total Var. Solve Total (sec) Time(s) Time(s) Time(s) Time(s) Time(s) Time(s) >1000 0.01 746.62 746.63 0.01 666.32 666.33 >1000 0.01 >1000 – 0.01 700.85 700.86 135.54 0.01 82.82 82.83 0 0.42 0.42 581.56 0.01 534.64 534.65 0.01 98.02 98.03 102.99 0 111.39 111.39 0 78.56 78.56 247.73 0 21.8 21.8 0 20.01 20.01 >1000 0 >1000 – 0 289.17 289.17 654.16 0.01 >1000 – 0.01 237.48 237.49 39.8 0.02 27.64 27.66 0.05 24.9 24.95 39.73 0.01 114.45 114.46 0.03 5.58 5.61 284.48 0 8.53 8.53 0 57.96 57.96 302.79 0 33.39 33.39 0 26.53 26.53 >1000 0 >1000 – 0 612.96 612.96 >1000 0 >1000 – 0 320.62 320.62 >1000 0 >1000 – 0 198.47 198.47 >1000 0 515.24 515.241 0 151.25 151.25 >1000 0 350.12 350.12 0 141.95 141.95 68.88 0.14 43.03 43.17 0.09 4.1 4.19 7.43 0.15 0.86 1.01 0.1 0.02 0.12 504.47 12.79 464.85 477.64 15.69 381.09 396.78

Benchmark clus-2020-1 clus-1010-3 conn-n600-939 conn-n600-945 icos-stretch marg-33-add8ch marg-35-1452 mm-1x10-1488 mm-2x2-50-149 qwh-35-405 urqh1c3x3 urqh2x4 urqh2x5 urqh1c3x4 unif-r4-2 unif-r4-7 unif-r4-9 3bitadd 31 3bitadd 32 9dlx-vliw-iq1


L’ ACCORD versus Contemporary Techniques

Table 4 presents some results that compares the quality of the variable order derived by L’ ACCORD with those of zCHAFF (latest version developed in 2004) and hypergraph partitioning (HGPart) based order [27]. For a fair comparison, zCHAFF is used as the base SAT solver for all experiments - just the

variable orders are di?erent. As compared to zCHAFF and HGPart, the variable order generated by L’ ACCORD results in a orders of magnitude speedup in SAT solving. Particularly for more di?cult problems, our approach signi?cantly outperforms the other two.
Table 4. Run-time Comparison of L’ ACCORD with zCHAFF and Hypergraph Partitioning based Technique zCHAFF HGPart + zCHAFF L’ ACCORD + zCHAFF Solve Var. Solve Total Var. Solve Total (sec) Time(s) Time(s) Time(s) Time(s) Time(s) Time(s) 68.88 3.023 0.55 3.573 0.09 4.1 4.19 7.43 3.417 0.67 4.087 0.1 0.02 0.12 >1000 1.421 >1000 – 0.01 666.32 666.33 688.19 1.81 9.12 10.93 0.04 37.17 37.21 975.83 1.846 43.85 45.696 0.02 630.55 630.57 >1000 1.701 704.66 706.361 0.01 700.85 700.86 129.57 0.375 19.01 19.385 0 25.05 25.05 135.54 1.196 28.96 30.156 0 0.42 0.42 581.56 1.145 >1000 – 0.01 98.02 98.03 102.99 0.199 93.84 94.039 0 78.56 78.56 247.73 0.071 142.53 142.601 0 20.01 20.01 3.44 0.074 1.36 1.434 0 1.29 1.29 >1000 0.085 >1000 – 0 289.17 289.17 654.16 1.057 >1000 – 0.01 237.48 237.49 39.8 3.815 31.37 35.185 0.05 24.9 24.95 17.32 4.386 16.48 20.866 0.08 13.55 13.63 39.73 1.869 84.59 86.459 0.03 5.58 5.61 284.48 0.078 8.92 8.998 0 57.96 57.96 302.79 0.077 71.07 71.147 0 26.53 26.53 >1000 0.117 >1000 – 0 612.96 612.96 >1000 0.143 196.78 196.923 0 320.62 320.62 >1000 0.447 >1000 – 0 198.47 198.47 884.04 0.491 91.88 92.371 0 129.79 129.79 >1000 0.451 582.53 582.981 0 151.25 151.25 >1000 0.493 711.65 712.143 0 141.95 141.95 3.54 3.549 0.73 4.279 0.04 0.07 0.11 238.5 8.189 134.78 142.969 0.09 66.51 66.6 174.7 5.888 153.26 159.148 0.28 159.53 159.81 504.47 68.83 443.79 512.62 15.69 381.09 396.78

Benchmark 3bitadd 31 3bitadd 32 clus-2020-1 clus-2020-2 clus-4530-2 clus-1010-3 color-10-3 conn-n600-939 conn-n600-945 icos-stretch marg-33-add8ch marg-33-add8 marg-35-1452 mm-1x10-1488 mm-2x2-50-1496 mm-2x3-66-1502 qwh-35-405 urqh1c3x3 urqh2x4 urqh2x5 urqh1c3x4 unif-r4-2 unif-r4-5 unif-r4-7 unif-r4-9 ferry10 ferry12 rotmul 9dlx-vliw-iq1

Vars/ SAT/ Clauses UNSAT 8432/21210 S 8704/32316 S 1200/4800 S 1200/4800 S 1200/4800 S 1200/4919 S 300/6475 S 576/6864 S 596/7157 S 45/352 U 41/272 U 41/224 U 61/280 U 1120/7220 S 60/32000 U 1698/48771 U 1597/10658 S 41/204 U 42/336 U 53/432 U 58/476 U 500/2000 S 500/2000 S 500/2000 S 500/2000 S 2958/20791 S 4222/32199 S 5980/35229 U 24604/261473 U

The table 5 depicts some results for the microprocessor pipeline veri?cation benchmarks. These experiments were conducted using both 2003 and 2004 versions of the zCHAFF SAT solver. For both experiments, the same variable order derived by L’ ACCORD is used. Note that, using the L’ ACCORD’s order,

zCHAFF-2003 tool is able to improve the performance signi?cantly. zCHAFF2004 follows upon its earlier versions by implementing the e?cient con?ict analysis procedures. As a result, these benchmarks can be quickly solved by the 2004 version of the solver and the impact of L’ ACCORD is minimal.
Table 5. Run-time Comparison of L’ ACCORD with zCHAFF’03 and zCHAFF’04 zCHAFF’03 L’ACCORD+zCHAFF’03 zCHAFF’04 L’ ACCORD+zCHAFF’04 Solve Var. Solve Total Solve Solve Total (sec) Time(s) Time(s) Time(s) Time(s) Time(s) Time(s) 129 0.56 56.1 56.66 8.02 11.96 12.52 196.53 2.36 54.46 56.82 18.34 16.59 18.95 208.81 0.6 40.85 41.45 8.34 14.53 15.13 871.79 2.25 388.32 390.57 40.75 31.29 33.54 66.71 0.48 73.86 74.34 6.37 8.59 9.07 649.9 1.53 214.89 216.42 25.21 26.46 27.99 >2000 6.53 827.06 833.59 132.82 144.72 151.25 105.52 8.76 141.45 150.21 71.49 57.64 66.4 104.48 4.88 82.44 87.32 43.37 46.65 51.53 >2000 12.45 1523.16 1535.61 284.97 265.3 277.75 >2000 31.16 >2000 – 819.13 892.32 923.48

Benchmark 4pipe 5pipe 4pipe k 5pipe k 4pipe q0 5pipe q0 6pipe 6pipe k 6pipe q0 7pipe q0 8pipe q0

Vars/ Clauses 5237/80213 9471/195452 5095/79489 9330/189109 5380/69072 10026/154409 15800/394739 15346/408792 16775/315960 26512/536414 39434/887706


Conclusions and Future Work

This paper has advocated the need to analyze constraint-variable relationships to derive an ordering of variables to guide SAT diagnosis. To analyze the tightness of the connectivity between variables, we have a proposed the degree of correlation as a qualitative and quantitative metric. Our technique models the constraint variable dependencies on a weighted graph and analyzes the graph’s topological structure to derive the order. Our approach is fast, robust, scalable and can handle a large set of variables and constraints. The variable order derived by our procedure improves the performance of the solver by one or more orders of magnitude. As part of future work, we are exploring a dynamic variable order update strategy to be employed when con?ict clauses are added to the database.

1. M. Davis and H. Putnam, “A Computing Procedure for Quanti?cation Theory”, Journal of the ACM, vol. 7, pp. 201–215, 1960. 2. M. Davis, G. Logemann, and D. Loveland, “A machine program for theorem proving”, in Communications of the ACM, 5:394-397, 1962. 3. J. Marques-Silva and K. A. Sakallah, “GRASP - A New Search Algorithm for Satis?ability”, in ICCAD’96, pp. 220–227, Nov. 1996.

4. M. Moskewicz, C. Madigan, L. Zhao, and S. Malik, “CHAFF: Engineering and E?cient SAT Solver”, in In Proc. Design Automation Conference, pp. 530–535, June 2001. 5. E. Goldberg and Y. Novikov, “BerkMin: A Fast and Robust Sat-Solver”, in DATE, pp 142-149, 2002. 6. F. Aloul, A. Ramani, I. Markov, and K. Sakallah, “Solving Di?cult SAT Instances in the Presence of Symmetry”, in Proc. DAC, pp. 731–736, 2002. 7. J. Crawford, M. Ginsberg, E. M. Luks, and A. Roy, “Symmetry-Breaking Predicates for Search Problems”, in Proceedings of the Fifth International Conference on Knowledge Representation and Reasoning (KR ’96), 1996. 8. B. Selman, H. Kautz, and B. Cohen, “Local Search Strategies for Satis?ability Testing”, DIMACS Series in Mathematics and Theoretical Computer Science, vol. 26, pp. 521–532, 1996. 9. M. Prasad, P. Chong, and K. Keutzer, “Why is ATPG Easy?”, in Design automation Conf., pp. 22–28, June 1999. 10. F. Aloul, I. Markov, and K. Sakallah, “Mince: A static global variable-ordering for sat and bdd”, in International Workshop on Logic and Synthesis. University of Michigan, June 2001. 11. F. Aloul and et al., “Satometer: How Much Have We Searched?”, IEEE Trans. CAD, vol. 22, pp. 995–1004, 2003. 12. L. Zhang and S. Mailk, “Extracting Small Unsatis?able Cores from Unsatis?able Boolean Formula”, in Sixth International Conference on Theory and Applications of Satis?ability Testing (SAT 2003), 2003. 13. E. Goldberg and Y. Novikov, “Veri?cation of Proofs of Unsatis?ability for CNF Formulas”, in Design, Automation and Test in Europe (DATE ), 2003. 14. S. Pilarski and G. Hu, “SAT with Partial Clauses and Backleaps”, in In Proc. DAC, pp. 743–746, 2002. 15. R. Damiano and J. Kukula, “Checking Satis?ability of a Conjunction of BDDs”, in DAC, pp 818-823, 2003. 16. P. Kalla, Z. Zeng, M. J. Ciesielski, and C. Huang, “A BDD-based Satis?ability Infrastructure using the Unate Recursive Paradigm”, in Proc. Design Automation and Test in Europe, DATE’00, 2000. 17. J. P. M. Silva, “The Impact of Branching Heuristics in Propositional Satis?ability Algorithms”, in Portuguese Conf. on Arti?cial Intelligence, 1999. 18. R. Dechter and J. Pearl, “Network-based Heuristics for Constraint-Satisfaction Problems”, Arti?cial Intelligence, vol. 34, pp. 1–38, 1987. 19. E. Amir and S. McIlraith, “Partition-Based Logical Reasoning”, in 7th International Conference on Principles of Knowledge Representation and Reasoning (KR’2000), 2000. 20. E. Amir and S. McIlraith, “Solving Satis?ability using Decomposition and the Most Constrained Subproblem”, in LICS workshop on Theory and Applications of Satis?ability Testing (SAT 2001), 2001. 21. P. Bjesse, J. Kukula, R. Damiano, T. Stanion, and Y. Zhu, “Guiding SAT Diagnosis with Tree Decompositions”, in Sixth International Conference on Theory and Applications of Satis?ability Testing (SAT 2003), 2003. 22. A. Gupta, Z. Yang, P. Ashar, L. Zhang, and S. Malik, “Partition-based Decision Heuristics for Image Computation using SAT and BDDs”, in Proceedings of the 2001 IEEE/ACM international conference on Computer-aided design (ICCAD), pp. 286–292. IEEE Press, 2001. 23. E. Amir, “E?cient Approximation for Triangulation of Minimum Treewidth”, in 17th Conference on Uncertainty in Arti?cial Intelligence (UAI ’01), 2001.

24. J. Huang and A. Darwiche, “A structure-based variable ordering heuristic for SAT”, in Proceedings of the 18th International Joint Conference on Arti?cial Intelligence (IJCAI), pp. 1167–1172, August 2003. 25. A. Caldwell, A. Kahng, and I. Markov, “Improved Algorithms for Hypergraph Bipartitioning”, in Proc. Asia-Paci?c DAC, 2000. 26. F. A. Aloul, I. L. Markov, and K. A. Sakallah, “FORCE: A Fast and Easy-ToImplement Variable-Ordering Heuristic”, in Proceedings of the 13th ACM Great Lakes symposium on VLSI, pp. 116–119, 2003. 27. V. Durairaj and P. Kalla, “Guiding CNF-SAT Search via E?cient Constraint Partitioning”, in ICCAD, pp. 498 – 501, 2004. 28. V. Durairaj and P. Kalla, “Exploiting Hypergraph Partitioning for E?cient Boolean Satis?ability”, in HLDVT, pp. 141–146, 2004. 29. V. Durairaj and P. Kalla, “Dynamic Analysis of Constraint-Variable Dependencies to Guide SAT Diagnosis”, in HLDVT, pp. 135–141, 2004. 30. F. Aloul, I. Markov, and K. Sakallah, “MINCE”, http://www.eecs.umich.edu/ faloul/Tools/mince/. 31. E. Amir and S. McIlraith, “The Partitioning and Reasoning Project”, http://wwwfaculty.cs.uiuc.edu/ eyal/decomp/. 32. R. C. Prim, “Shortest connection networks and some generalizations”, The Bell System Technical Journal, vol. 36, pp. 1389–1401, 1957. 33. N. E?n and N. S¨rensson, “An Extensible SAT Solver”, in 6th International e o Conference, SAT, 2003. 34. M. Velev, “Sat benchmarks for high-level microprocessor validation”, http://www.cs.cmu.edu/ mvelev.

Enhanced Interior Gateway Routing Protocol, Cisco W....pdf

Enhanced Interior Gateway Routing Protocol, Cisco White Paper - Abstract: This paper presents a n...


(CGMP:Cisco Group Management Protocol) EIGRP:增强的内部网关路由选择协议 (EIGRP:Enhanced Interior Gateway Routing Protocol) IGRP:内部网关路由协议 (IGRP:...


CCNA实验攻略(9)Cisco+EIGRP - EIGRP(Enhanced Interior Gateway Routing Protocol,增强型内部网关路由协 议)是 Cisco ...

第7章 路由协议配置.doc

4. 增强型内部网关路由协议 EIGRP 增强型内部网关路由协议( EnhancedInterior Gateway Routing Protocol,EIGRP)即增强型的 IGRP,也是 Cisco 专用的路由协议。它使用了...


EIGRP:增强的内部网关路由选择协议 EIGRP:Enhanced Interior Gateway Routing Protocol 增强的内部网关路由选择协议 EIGRP 是增强版的 IGRP 协议。 IGRP 是思科提供的...


CISCO(思科)网络协议总结大全 ( 思科) 作者从网络、路由、数据链路、网络安全...(EIGRP:Enhanced Interior Gateway Routing Protocol) 增强的内部网关路由选择协议 ...


第一篇 cisco 各种协议配置方法 网络/路由 (Network/Routing) CGMP:思科组管理...(EIGRP:Enhanced Interior Gateway Routing Protocol) IGRP:内部网关路由协议 (...


增强型内部网关路由选择协议((Enhanced Interior Gateway Routing Protocol),也称为EIGRP或增强型IGRP,它足Cisco公司在IGRP基础上的一种新的改进型协议。其度量标准主...


EIGRP:增强的内部网关路由选择协议 EIGRP:Enhanced Interior Gateway Routing Protocol 增强的内部网关路由选择协议 EIGRP 是增强版的 IGRP 协议。IGRP 是思科提供的一...


Enhanced Interior Gateway Routing Protocol (EIGRP)....Refer to the EIGRP White Paper for more ...In Cisco IOS Software Releases earlier than 12.2...


Cisco 私有协议总结 V1 整理、收集的 Cisco 私有协议本文共分为四部分: 思科...(EIGRP:Enhanced Interior Gateway Routing Protocol) IGRP:内部网关路由协议 (...

Cisco NX-OS Software - Business-Critical Cross-Plat....pdf

White Paper Cisco NX-OS Software: Business-...(Interior Gateway Routing Protocol (EIGRP), ...Enhanced Usability and Familiar Operation Cisco IOS...


(config-keychain-key)#key-string cisco r1(config)#int f0/0 r1(config-...(Enhanced Interior Gateway Routing Protocol) 1、EIGRP 简介 1.增强型内部网关...

Protocol Expert 能识别的协议.doc

Protocol Name CDP Cisco Discovery Protocol DISL Dynamic Inter-Switch Protocol EIGRP Enhanced Interior Gateway Routing Protocol (see Internet Protocol suite) ...


EIGRP:增强的内部网关路由选择协议 EIGRP:Enhanced Interior Gateway Routing Protocol 增强的内部网关路由选择协议 EIGRP 是增强版的 IGRP 协议。 IGRP 是思科提供的...


CISCO 协议总结大全从网络、路由、数据链路、网络安全技术等 4 个方面对 Cisco ...(EIGRP:Enhanced Interior Gateway Routing Protocol) IGRP:内部网关路由协议 (...


EIGRP:Enhanced Interior Gateway Routing Protocol 即 增强网关内部路由线路协议。也 翻译为 加强型内部网关路由协议。 EIGRP 是 Cisco 公司的私有协议。Cisco 公司是...

Packet Tracer 5.0建构CCNA实验攻略(9)Cisco EIGRP....doc

EIGRP(Enhanced Interior Gateway Routing Protocol,增强型内部网关路由协议)是 Cisco 内部专有 协议,其它公司的网络产品是不会拥有该协议的。 一、配置实例拓扑图 ...


思科ME6500产品手册_IT/计算机_专业资料。Data Sheet Cisco ME 6500 Series ...scalable Enhanced Interior Gateway Routing Protocol (EIGRP) Layer 3 services....


EIGRP:增强的内部网关路由选择协议 EIGRP:Enhanced Interior Gateway Routing Protocol 增强的内部网关路由选择协议 EIGRP 是增强版的 IGRP 协议。IGRP 是思科提供的一...