Mad-MPI: An Efficient Implementation of MPI for Fast Networks - Performances


LaBRI, INRIA Bordeaux - Sud-Ouest

High Performance Runtime Systems for Parallel Architectures

Aggregation of small messages

We have also shown the benefits of the aggregation of small messages, by comparing the performance of a multi-segments ping-pong program, with each ``ping'' being a serie of independent MPI_Isend operations that use separate MPI communicators. We have observed that Mad-MPI is up to 70% faster than other implementations of MPI over MX-10G, and up to 50% faster that MPICH over Quadrics.