Message-passing performance of parallel computers

In this paper we investigate some of the important factors which affect the message-passing performance on parallel computers. Variations of low-level communication benchmarks that approximate realistic cases are tested and compared with the available Parkbench codes and benchmarking techniques. The results demonstrate a substantial divergence between message-passing performance measured by low-level benchmarks and message-passing performance achieved by real programs. In particular, the influence of different data types, the memory access patterns, and the communication structures of the NAS Parallel Benchmarks are analysed in detail, using performance measurements on the Cray-T3D in Edinburgh and the IBM-SP2 in Southampton