High Performance I/O In SupercomputingApplications
Vittorio Zecca · WIT transactions on information and communication technologies · 1970
At the IBM European Center for Scientific and Engineering Computing a prototype has been written for transparent transformation of I/O primitives into access to processor storage, thus allowing applications to continue to use standard high level language I/O statements, but taking advantage of the full storage hierarchy provided by the underlying processor in a VM/ESA environment. The prototype for High Performance CMS I/O (HPCMSIO) may be used with high level programming languages like FORTRAN, C, PL/1, and Assembler, of course. Purpose of the work was to demonstrate the capability of VM/ESA to optimally execute an industrial code on a supercomputer exploiting its storage hierarchy, and particularly the virtual and expanded storage, with negligible intervention on the programmer's side. The present paper discusses the implementation of these concepts and provides performance data measured with the fluid dynamics industrial code El DAP (*). For I/O to memory a speed of 60 Megabytes/sec was measured on an ES/9000 model 620. Such a speed is measured when the data space resides entirely in central storage. Otherwise, page faults occur and pages are moved between central storage and slower access media like Expanded Storage or disk. In any case data space access is always faster than application disk access because of the shorter access time for Expanded Storage and of the faster paging as done by VM/ESA. For a machine with a shorter cycle time like the ES/9000 model 900 a speed of 100 Megabytes/sec can be achieved for I/O to memory. (*) FIDAP is a registered trademark of Fluid Dynamics International. Transactions on Information and Communications Technologies vol 3, © 1993 WIT Press, www.witpress.com, ISSN 1743-3517