SAT solver management strategies in IC3: an experimental approach

This paper addresses the problem of handling SAT solving in IC3. SAT queries posed by IC3 significantly differ in both character and number from those posed by other SAT-based model checking algorithms. In addition, IC3 has proven to be highly sensitive to the way its SAT solving requirements are handled at the implementation level. The scenario pictured above poses serious challenges for any implementation of the algorithm. Deciding how to manage the SAT solving work required by the algorithm is key to IC3 performance. The purpose of this paper is to determine the best way to handle SAT solving in IC3. First we provide an in-depth characterization of the SAT solving work required by IC3 in order to gain useful insights into how to best handle its queries. Then we propose an experimental comparison of different strategies for the allocation, loading and clean-up of SAT solvers in IC3. Among the compared strategies we include the ones typically used in state-of-the-art model checking tools as well as some novel ones. Alongside comparing multiple versus single SAT solver implementations of IC3, we propose the use of secondary SAT solvers dedicated to handling certain types of queries. Different heuristics for SAT solver clean-up are evaluated, including new ones that follow the locality of the verification process. We also address clause database minimality, comparing different CNF encoding techniques. Though not finding a clear winner among the different sets of strategies compared, we outline several potential improvements for portfolio-based verification tools with multiple engines and tunings.