Poster 107: Exploring Interprocess Work Stealing for Balanced MPI Communication
TimeThursday, 21 November 20198:30am - 5pm
DescriptionWorkload balance among MPI processes is a critical consideration during the development of HPC applications. However, because of many factors such as complex network interconnections and irregularity of HPC applications, fully achieving workload balance in practice is nearly impossible. Although interprocess job stealing is a promising solution, existing shared-memory techniques that lack necessary flexibility or cause inefficiency during data access cannot provide an applicable job-stealing implementation. To solve this problem, we propose a new process-in-process (PiP) interprocess job-stealing method to balance communication workload among processes on MPI layers. Our initial experimental results show PiP-based job stealing can efficiently help amortize workload, reduce imbalance, and greatly improve intra- and intersocket ping-pong performance compared with original MPI.