Precise MPI performance measurement using MPIBench

Abstract

We have developed a new MPI benchmark package called MPIBench that uses a very precise and portable distributed timing mechanism to accurately measure the performance of a single MPI communications routine without having to average over many calls. This enables MPIBench to provide precise and detailed performance pro les of standard MPI routines which cannot be produced using existing MPI benchmark suites. We describe the design of MPIBench and how it overcomes the various drawbacks of existing MPI benchmarks. We present some examples of the kinds of new results that MPIBench can provide, including the distribution of timing results for message passing routines. This is useful for performance modelling and analysis of parallel programs, as well as for understanding the performance of parallel machines.