Analysis of algorithm selection for optimizing collective communication with MPICH for ethernet and myrinet networks

Abstract

MPICH is probably the most widely used implementation of MPI. Versions of MPICH optimize the performance of some collective communications by switching between different algorithms depending on whether the message size is greater or less than a given change-over point, which is currently hard-coded in MPICH, based on measurements on clusters with one CPU per node. We have used MPI benchmarks to find the optimum change-over points for different systems, and found that they can vary significantly for different networks and different numbers of processes per node. In some cases significant performance improvements can be obtained by enabling MPICH to be customized in this way, particularly on clusters with more than one CPU per node.