Implementing Parallel I/O: MPI-IO and HPSS

John May
Lawrence Livermore National Laboratory

Abstract

Moving data from memory to long-term storage has often been an afterthought in the design of large scientific applications. However, compute power is growing faster I/O rates. More and more, I/O can dominate the running time of a large scientific application, so it has become important for developers to focus on I/O issues. The Department of Energy's Accelerated Strategic Computing Initiative (ASCI) has serveral projects to address high performance I/O problems for its scientific simulation codes. This talk will briefly outline ASCI's I/O strategies, and then it will examine in detail the one component in this strategy, an implementation of the MPI-IO parallel I/O standard for the HPSS archival storage system. Both MPI-IO and HPSS are complex systems, and combining them presented a number of interesting problems in parallel software design. We will look at how some of these problems were solved and present I/O performance data for the current system.