Laissez-Faire Caching for Parallel #SAT Solving

• Published in: Theory and Applications of Satisfiability Testing -- SAT 2015 Vol. 9340; pp. 46 - 61
• Main Authors: Burchard, Jan, Schubert, Tobias, Becker, Bernd
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2015; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Artificial intelligence; Component Tree; Computer science; Conjunctive Normal Form; Fault Injection; Mathematical theory of computation; Propositional Formula; Variable Assignment
• ISBN: 3319243179; 9783319243177
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-319-24318-4_5

The problem of counting the number of satisfying assignments of a propositional formula (#SAT) can be considered to be the big brother of the well known SAT problem. However, the higher computational complexity and a lack of fast solvers currently limit its usability for real world problems. Similar to SAT, utilizing the parallel computation power of modern CPUs could greatly increase the solving speed in the realm of #SAT. However, in comparison to SAT there is an additional obstacle for the parallelization of #SAT that is caused by the usage of conflict learning together with the #SAT specific techniques of component caching and sub-formula decomposition. The combination can result in an incorrect final result being computed due to incorrect values in the formula cache. This problem is easily resolvable in a sequential solver with a depth-first node order but requires additional care and handling in a parallel one. In this paper we introduce laissez-faire caching which allows for an arbitrary node computation order in both a sequential and parallel solver while ensuring a correct final result. Additionally, we apply this new caching approach to build countAntom, the world’s first parallel #SAT-solver. Our experimental results clearly show that countAntom achieves considerable speedups through the parallel computation while maintaining correct results on a large variety of benchmarks coming from different real-world applications. Moreover, our analysis indicates that laissez-faire caching only adds a small computational overhead.