Archived: 20 January 2004
Published in Proc. of APAC'03, Gold Coast, September 2003.
MPI is the dominant paradigm for parallel programming on cluster computers. Accurate measurements of the performance of MPI routines are invaluable for analysing and optimising the performance of MPI library implementations and for understanding and modelling the performance of parallel programs that use MPI. We have developed a benchmarking tool called MPIBench that provides very accurate measurements of MPI communication in the form of timing distributions, not just minimum or average times. Here we present some MPIBench results for MPI communication performance on the APAC National Facility, a large AlphaServer SC supercomputer featuring a high-performance QsNet communication network. We show how MPIBench results can provide useful insights into the performance of the MPI libraries for this machine, and can also be used to develop improved performance models for parallel programs.
Keywords: Parallel computing, MPI, communication performance, performance modelling, AlphaServer SC.
Postscript version (gzip compressed)