A system software architecture for high-end computing

Abstract

Large MPP systems can neither solve grand-c hallenge scien ti c problems nor enable large scale industrial and governmental simulations if they rely on extensions to w orkstation system softw are. A t Sandia National Laboratories we h a ve d e v eloped, with our vendors, a new system architecture for highend computing. Highest performance is achieved by p r o viding applications with a light-weight i n terface to a collection of processing nodes. Usability i s p r o vided by creating node partitions specialized for user access, net w orking, and I/O. The en tire system is glued together by a data movement i n terface which w e call portals. P ortals allo w data to o w b e t w een processing nodes with minimal system overhead while maintaining a suitable degree of protection and recon gurability.

Scalable parallel application launch on Cplant™

“…The Cplant TM machines employ the partition model of resource provision [8] ¢ that was initially developed by Intel on their early parallel platforms. This model divides the machine into several different partitions that provide specialized functionality.…”

Massively parallel computing using commodity components

“…The TFLOPS employs a partition model of resources [8], where each partition provides access to a specialized resource. Fig.…”

“…The use of the partition model of resource provision [8] allows SUs to provide a variety of resources, such as service, compute, I/O, and network. Most of the Cplant resources are likely to be devoted to creating computational nodes, i.e.…”

Enabling Department-Scale Supercomputing

“…The system software for ASCI Red partitions the nodes into three logical groups: service, I/O, and compute [13]. Since the service partition is relatively small (on the order of ten nodes) it runs a variant of OSFl/AD, which supplies a standard workstation image to users.…”