Trading-off Incrementality and Dynamic Restart of Multiple Solvers in IC3

This paperaddresses the problem of SAT solver performance in IC3, one of the major recent breakthroughs in Model Checking algorithms. Unlike other Bounded and Unbounded Model Checking algorithms, IC3 is characterized by numerous SAT solver queries on small sets of problem clauses. Besides algorithmic issues, the above scenario poses serious performance challenges for SAT solver configuration and tuning. As well known in other application fields, finding a good compromise between learning and overhead is key to performance. We address solver cleanup and restart heuristics, as well as clause database minimality, based on on-demand clause loading: transition relation clauses are loaded in solver based on structural dependency and phase analysis. We also compare different solutions for multiple specialized solvers, and we provide an experimental evaluation on benchmarks from the HWMCC suite. Though not finding a clear winner, the work outlines several potential improvements for a portfolio-based verification tool with multiple engines and tunings.