The "single-assignment" approach to parallel processing

Abstract

Parallel processing systems---computer systems in which more than one processor is active simultaneously---offer potential advantages over uniprocessor systems in terms of speed, flexibility, reliability, and economies of scale. However, they pose the problem of how multiple processors can be organized to cooperate on a given problem without interference. Various solutions have been proposed to this problem. Some systems require a programmer to assign units of work to the various processors, while other systems perform this assignment automatically; in some systems, the processors are linked closely together, while in others the processors are nearly independent. The system to be described here automatically detects opportunities for parallel processing in programs written in a specific, high-level language. Parallelism is detected on a very low level---even within a single algebraic expression. The system consists of many independent, asynchronous processors, all active at once in processing a single program.