| Description |
Existing multiprocessors and multicomputers require the programmer or compiler to perform data dependence analysis at compile time. We propose a parallel computer that performs this task at runtime. In particular, the Virtual Time Machine (VTM) detects violations of data dependence constraints as they occur, and automatically recovers from them. A sophisticated memory system that is addressed using both a spatial and a temporal coordinate is used to efficiently implement this mechanism. Initially targeted for discrete event simulation applications, many of the ideas used in the machine architecture have direct application in the more general realm of parallel computation. The long term goal of this work is to develop a general purpose parallel computer that will support a wide range of parallel programming paradigms. This paper outlines the motivations behind the V TM architecture, the underlying computation model, a proposed implementation, and initial performance results. A recurring theme that pervades the entire paper is our contention that existing shared memory and message-base machines do not pay adequate attention to the dimension of time. We argue that this architectural deficiency is the underlying reason behind many difficult problems in parallel computation today. |
