The parallel effective I/O bandwidth benchmark: b eff io

ABSTRACT. The parallel effective I/O bandwidth benchmark (b_eff_io) is aimed at producing a characteristic average number of the I/O bandwidth achievable with parallel MPI-I/O appli-cations exhibiting various access patterns and using various buffer lengths. It is designed so that 15 minutes should be sufficient for a first pass of all access patterns. First results of the b_eff_io benchmark are given for the IBM SP, Cray T3E, Hitachi SR 8000, and NEC SX-5 sys-tems, and a discussion follows about problematic issues of our current approach. We show how a redesign of our time-driven approach allows for rapid benchmarking of I/O bandwidth with various compute partition sizes. Next, we present how implementation specific file hints can be enabled selectively on a per access pattern basis, and we illustrate the benefit that hints can provide using the latest version of the IBM MPI-IO/GPFS prototype