A scalable high-performance I/O system
Mark Henderson, Bill Nickless, Rick L. Stevens · 2002
A significant weakness of many existing parallel supercomputers is their lack of high-performance parallel I/O. This weakness has prevented, in many cases, the full exploitation of the true potential of MPP systems. As part of a joint project with IBM, we have designed a parallel I/O system for an IBM SP system that can provide sustained I/O rates of greater than 160 MB/s from collections of compute nodes to archival disk and peak transfer rates that should exceed 400 MB/s from compute nodes to I/O servers. This testbed system is being used for a number of projects, first it will provide a high-performance experimental I/O system for traditional computational science applications, second it will be used as an I/O software and development environment for new parallel I/O algorithms and operating systems support, and third it will be used as the foundation for a number of new projects designed to develop enabling technology for the National Information Infrastructure. This report describes the system under development at Argonne National Laboratory, provides some preliminary performance results, and outlines future experiments and directions.>