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


LaBRI, INRIA Bordeaux - Sud-Ouest

High Performance Runtime Systems for Parallel Architectures

Pingpong Performances

Mad-MPI, has shown that the performance of NewMadeleine can be obtained with MPI applications. Mad-MPI has been compared to implementations of MPI for specific high performance networks, MPICH-MX and OpenMPI-MX 1.1 over Myri-10G, and MPICH-Quadrics over Quadrics. This allowed us to evaluate the overhead of Mad-MPI under situations where no optimization is possible, as for example where a MPI ping-pong program exchanges single-segment messages (i.e. contiguous arrays of bytes). On both networks, Mad-MPI introduces a constant overhead of less than 0,5 micro-s and reaches 1155 Mbytes/s in bandwidth over Myri-10G and 835 Mbytes/s over Quadrics.