The Standard Performance Evaluation Corp. (SPEC) has just released a parallel systems benchmark based on MPI applications
SPEC MPI2007 provides performance metrics that can be used to compare different hardware architectures (SMP, NUMA, clusters) and interconnects, processors, memory hierarchy, compilers, and MPI implementations. It tests performance based on actual end-user applications, instead of synthetic workloads or parallelized versions of sequential benchmarks.
…Two metrics are provided by SPEC MPI2007: A base metric for users who prefer the relative simplicity of a single-step build process, and a peak metric for those who want to experiment with different compilers and compiler flags to achieve the best performance. Performance measurement is normalized against a reference system to produce a “bigger-is-better” metric. The current reference system is an eight-node cluster of Celestica A2210 systems, each with dual AMD Opteron processors, connected by a TCP (GigE) interconnect.
MPI2007 complements other HPC benchmarks already available from SPEC: HPC2002 and OMP2001 (details at SPEC’s website). The benchmark is available now, but you’re going to have to pony up $800 for it. Details in the release.