MPISE: Symbolic Execution of MPI Programs

Abstract

Abstract. Message Passing Interfaces (MPI) plays an important role in parallel computing. Many parallel applications are implemented as MPI programs. The existing methods of bug detection for MPI programs have the shortage of providing both input and non-determinism coverage, leading to missed bugs. In this paper, we employ symbolic execution to ensure the input coverage, and propose an on-the-fly schedule algorithm to reduce the interleaving explorations for non-determinism coverage, while ensuring the soundness and completeness. We have implemented our approach as a tool, called MPISE, which can automatically detect the deadlock and runtime bugs in MPI programs. The results of the experiments on benchmark programs and real world MPI programs indicate that MPISE finds bugs effectively and efficiently. In addition, our tool also provides diagnostic information and replay mechanism to help understanding bugs.

Synchronization Error Detection of MPI Programs by Symbolic Execution

“…Hence, in principle, our analysis is not sound but complete. However, when only synchronous communications exist, the method in [4] can be used, which ensures the soundness.…”

“…Each process runs the same program, and the processes communicate via message passings, using different MPI APIs and the supporting MPI platform. In addition to the synchronous APIs [4], the asynchrony related APIs we consider in this paper include:…”

“…The scheduler is a non-preemptive one, i.e., a process would keep being executed until it encounters blocking MPI calls (the handling of which can be referred to [4]) or other explicit preemption points such as sleep. The second rule of this scheduler is: always pick the process with the smallest rank to symbolic execute, i.e., if there are n processes and the process P i is blocked, the next process to be executed is P ((i+1) mod n) if P ((i+1) mod n) is not blocked; otherwise, the scheduler continues to check the next process until an active process is encountered.…”

“…For the statements that do not affect synchronization errors, such as non-communicative statements and synchronous communication statements, the symbolic execution semantics can be referred to [7] [4]. In the following of this section, we will concentrate on the handling of synchronization error related events and the optimizations.…”

“…The framework of MPISE the latter analyzes the usages of each buffer to see if there is a synchronization error according to Algorithm 2. Same as our previous work [4], we also employ a multithread MPI library as the MPI "model" for symbolic execution. We use azequiaMPI [10], which is an MPI platform on which an MPI program is run in a multi-thread manner.…”