dMPI: Facilitating Debugging of MPI Programs via Deterministic Message Passing

This paper presents a novel deterministic MPI implementation (dMPI) to facilitate the debugging of MPI programs. Distinct from existing approaches, dMPI ensures inherent determinism without using any external support (e.g., logs), which achieves convenience and performance simultaneously. The basic idea of dMPI is to use deterministic logical time to solve message races and control asynchronous transmissions, thus we could eliminate the nondeterministic behaviors of the existing message passing mechanism. To avoid deadlocks introduced by dMPI, we also integrate dMPI with a lightweight deadlock checker to dynamically detect and solve these deadlocks. We have implemented dMPI and evaluated it using NPB benchmarks. The results show that dMPI could guarantee determinism with incurring modest overhead (8% on average).