An Experimental Study into the Effect of Multiprocessor Interconnection Networks on System Performance

点击翻译按钮获取中文摘要

The long term goal of this project is to investigate the relationship between the structure of parallel algorithms and the structure of parallel computing systems. It seems clear that it will be necessary to match the structure of an algorithm to the structure of the parallel computer on which it executes in order to realize the computing power of the parallel system. Whether this matching is done by the applications programmer or a compiler, a basic understanding of the structural relationship between parallel algorithms and parallel computer organization will be needed. This investigation proposed here will provide needed insight into this process. During the phase of the project covered in this proposal, the investigation will concentrate on the basic question of how the structure and design of a parallel computing system's interconnection network affects the overall performance of the system. The interconnection network provides communication paths between the modules of a parallel system and is a major source of system overhead. For Parallel systems composed of a large number of processors, this overhead can significantly reduce the system's performance relative to its theoretical maximum. This project will evaluate the effect of this overhead using the Rice Parallel Processing Testbed, a sophisticated simulation system developed for predicting the performance of real parallel programs executing on parallel processors.