Design Issues in High Performance Fault-Tolerant Multicomputers

Abstract

There are a number of upcoming applications in space and defense systems which require computing with very high levels of both performance and dependability. The computer architectures generally believed to be most applicable with regard to performance requirements are highly parallel structures which make use of large numbers of processors operating concurrently. This type of highly parallel approach is cost effective in VLSI implementations where relatively inexpensive microcomputer-based nodes are implemented as a small set of custom chip-types and then replicated as many copies. For many specialized applications, these systems offer the performance of supercomputers at a fraction of their cost. There has been a large commitment of resources to the development of several promising experimental systems of this type (Ultramax, RP3, Butterfly Machine, Hypercube, Warp, Connection Machine, etc.). However, none of the parallel machines currently under development are designed to provide fault tolerance and, more generally, the level of dependability needed for many critical application environments.