Parallel Graph-Based Stateless Model Checking

Abstract

Stateless model checking (SMC) is an automatic technique with low memory requirements for finding errors in concurrent programs or for checking for their absence. To be effective, SMC tools require algorithms that combat the combinatorial explosion in the number of process/thread interactions that need to be explored. In recent years, a plethora of such algorithms have emerged, which can be classified broadly in those that explore interleavings (i.e., complete serializations of events) and those that explore traces (i.e., graphs of events). In either case, an SMC algorithm is optimal if it explores exactly one representative from each class of equivalent executions. In this paper, we examine the parallelization of a state-of-the-art graph-based algorithm for SMC under sequential consistency, based on the reads-from relation. The algorithm is provably optimal, and in practice spends only polynomial time per equivalence class. We present the modifications to the algorithm that its parallelization requires and implementation aspects that allow us to make it scalable. We report on the performance and scalability that we were able to achieve on C/pthread programs, and how this performance compares to that of other SMC tools. Finally, we argue for the inherent advantages that graph-based algorithms have over interleaving-based ones for achieving scalability when parallelism enters the picture.