Evaluation and measurement of multiprocessor latency patterns

Network latency, the delay caused by communication between processors and memory modules over the network in a multiprocessor system, is a major source of degraded parallel computing performance. We first give an overview of an experimental metric which uses network latency to measure and evaluate the scalability of parallel programs and architectures. We put emphasis on evaluation of latency sources and their methods in program execution. We report experimental results of evaluating the scalability of several scientific computing algorithms on the KSR-1. In comparison, we also present preliminary experiments on the CM-5 multicomputer architecture. Our comparisons indicate that the CM-5 has less network latency effects and is a more scalable architecture the the KSR-1. The analyses and experiments show that the latency metric is an important method to effectively measure overheads inherent in the program and the architecture for parallel computing scalability evaluation