Modeling MPI collective communications on the AP3000 multicomputer

Abstract

The performance of the communication primitives of a parallel computer is critical for the overall system performance. The performance of the message-passing routines does not only depend on the hardware of the communication subsystem, but also on their implementation. When users migrate parallel codes to a multicomputer, they are unaware of the quality of the implementation and they can find that the performance of their parallel applications makes worse. In this work, we evaluate some representative collective communication routines of the MPI message-passing library: broadcast, scatter/gather, total exchange and reduction operations. Specifically, experimental results were conducted on the Fujitsu AP3000 multicomputer using Fujitsu MPI/AP version 1.0, in order to derive models for these MPI routines.