A Simulator for Message Passing Based Parallel and Distributed Programs to Evaluate the Influence of Latency of Message Transfers

摘要

As parallel and distributed computing environments, like the PC cluster or the GRID, become more popular, high performance parallel and distributed programs become increasingly more necessary. Within the family of such programs, those that are based on message-passing are frequently characterized by unpredictable delays in message arrival times, which lead to significant differences in program performance. This paper reports the development of a simulator that measures program performance under parallel and distributed computing environments with variable parameters. Also, two types of numerical experiments are presented: one that was conducted on a program of matrix multiplication whose performance does not depend as much on communication delays and another experiment conducted on a program of branch and bound technique whose performance is more dependent on communication delays. Both experiments were performed in the simulator presented here, as well as in a real PC cluster environment. Through these numerical experiments, the effectiveness and problems of the simulator are discussed.