Optimization for Hybrid MPI-OpenMP Programs with Thread-to-thread Communication

摘要

This paper proposes a middle-grain approach to construct hybrid MPI-OpenMP solutions for SMP clusters from an existing MPI algorithm. Experiments with "the High Performance Parallel Benchmark-HPL" on different cluster platforms show that, our solutions exceed the solutions based on a de-facto MPI model in most cases, and sometimes by as much as 34.5% of performance. We also prove an automatic outperformance of a thread-to-thread communication model over a classical process-to-process communication model in hybrid solutions.