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


LaBRI, INRIA Bordeaux - Sud-Ouest

High Performance Runtime Systems for Parallel Architectures

NAS Parallel Benchmarks

The NAS Parallel Benchmarks (NPB) are a small set of programs designed to help evaluate the performance of parallel supercomputers. The benchmarks, which are derived from computational fluid dynamics (CFD) applications, consist of five kernels and three pseudo-applications.

NAS Parallel Benchmarks can be downloaded at

Using NPB, version 3.2.1, we compared, Mad-MPI, our implementation of MPI, with MPICH over MX and TCP. The code of some of the benchmarks had to be modified, to overcome some of the Mad-MPI restrictions: a maximum value for the tag communication, and the impossibility for a node to send data to itself. The patch of the modifications is given here.

Benchmarks were run on IBM nodes, IBM eServer 325, with 2 CPUs AMD Opteron 248, 2,2 GHz / 1 MB L2 cache. The Myrinet driver is MX version 1.1.6, and the compiler Fortran is Intel Fortran version 9.0. We used MPICH-MX, version, and MPICH, version 1.2.7.