Real-time communications scheduling for massively parallel processors

Abstract

Can general purpose commercial massively parallel processors (MPPs) be used for computationally intensive real-time applications that have traditionally required a custom arrangement of special-purpose computers and mainframes? If so, then the enormous lifecycle costs of many systems needed by, for instance, the Government could potentially be reduced. The components would be commercially available and continuing technological advances could more easily be incorporated into existing systems. Relevant applications have requirements not found in large-scale scientific computing, which has up to now provided most of the motivation for the development of MPPs. Perhaps the most important difference is the need for real-time processing. Depending on the application, multi-level security, fault tolerance and other features may also be necessary. There have already been hardware advances that may make such a high performance computing solution possible. However, daunting software challenges remain. We focus on one critical problem, the real-time scheduling of the communications between processing nodes. We discuss real-time MPP application benchmarks.