MPI performance evaluation and characterization using a compact application benchmark code

Abstract

In this paper the parallel benchmark code PSTSWM is used to evaluate the performance of the vendor-supplied implementations of the MPI message-passing standard on the Intel Paragon, IBM SP2, and Cray Research T3D. This study is meant to complement the performance evaluation of individual MPI commands by providing information on the practical significance of MPI performance on the execution of a communication-intensive application code. In particular three performance questions are addressed: how important is the communication protocol in determining performance when using MPI, how does MPI performance compare with that of the native communication library, and how efficient are the collective communication routines.