Combining Symbolic Execution and Model Checking to Verify MPI Programs
share
Message Passing Interface (MPI) is the standard paradigm of programming in high performance computing. MPI programming takes significant effort, and is error-prone. Thus, effective tools for analyzing MPI programs are much needed. On the other hand, analyzing MPI programs itself is challenging because of non-determinism caused by program inputs and non-deterministic operations. Existing approaches for analyzing MPI programs either do not handle inputs or fail to support programs with mixed blocking and non-blocking operations. This paper presents MPI symbolic verifier (MPI-SV), the first symbolic execution based tool for verifying MPI programs having both blocking and non-blocking operations. To ensure soundness, we propose a blockingdriven matching algorithm to safely handle non-deterministic operations, and a method to soundly and completely model the equivalent behavior of a program execution path. The models of MPI program paths are generated on-the-fly during symbolic execution, and verified w.r.t. the expected properties by model checking. To improve scalability, MPI-SV uses the results of model checking to prune redundant paths. We have implemented MPI-SV and evaluated it on the verification of deadlock freedom for 108 real-world MPI tasks. The pure symbolic execution based technique can successfully verify 61 out of the 108 tasks (56 hour, while in comparison, MPI-SV can verify 94 tasks (87 On average, MPI-SV also achieves 7.25X speedup on verifying deadlock freedom and 2.64X speedup on finding deadlocks. These experimental results are promising, and demonstrate MPI-SV's effectiveness and efficiency.
READ FULL TEXT