HPC benchmarks

We have developed some benchmarks usually used to check the abilities of hpc clusters. Our main focus is directed to the total MPI-bandwidth and the memory speed (random access, latency), because this are the most relevant bottlenecks in our multi-threaded applications. The benchmarks below are free available.

• mpi_stress.c, MPI-communication between all nodes with changing communication paths for clusters (since 2008)
• memspeed.c, parallel random memory access (for single SMP nodes, since 2002)

Older systems which are selected using above benchmarks can be find here: SC5832 MIPS cluster (2009), GS1280 alpha SMP (2003). Some results are available as gnuplot-files (memspeed.gpl, mpi_stress.gpl).