The APEX RFP calls out several MPI related requirements that can be categorized as two-sided or one-sided and the respective measures of message rate, bandwidth and latency for each. In addition, collective operations are called out.
The general philosophy for MPI benchmarking is to use publicly available micro-benchmarks were appropriate and to develop new micro-benchmarks where there are gaps in the public benchmark suites. Unless a benchmark is explicitly called out in this document, the selected benchmarks will be mutually agreed upon between APEX and the selected offeror.
Publicly available benchmark suites that will be used include:
- The Ohio State University MVAPICH benchmarks, http://mvapich.cse.ohio-state.edu/benchmarks/
- The Sandia MPI Benchmarks (SMB), http://www.cs.sandia.gov/smb/
- The multi-threaded MPI message rate and latency benchmarks are not yet available on the SMB website, they can be found here.
|02/29/2016||Added link for SMB multi-threaded MPI benchmarks|