A SIMULATION-BASED SCALABILITY STUDY OF PARALLEL SYSTEMS

Scalability studies of parallel architectures have used scalar metrics to evaluate their performance. Very often, it is difficult to glean t he sources of inefficiency resulting from the mismatch between the alg orithmic and architectural requirements using such scalar metrics. Low -level performance studies of the hardware are also inadequate for pre dicting the scalability of the machine on real applications. We propos e a top-down approach to scalability study that alleviates some of the se problems. We characterize applications in terms of the frequently o ccurring kernels and their interaction with the architecture in terms of overheads in the parallel system. An overhead function is associate d with each artifact of the parallel system that limits its scalabilit y. We present a simulation platform called SPASM (Simulator for Parall el Architectural Scalability Measurements) that quantifies these overh ead functions. SPASM separates the algorithmic overhead into its compo nents (such as serial and work-imbalance overheads), and interaction o verhead into its components (such as latency and contention). Such a s eparation is novel and has not been addressed in any previous study us ing real applications. We illustrate the top-down approach by consider ing a case study in implementing three NAS parallel kernels on two sim ulated message-passing platforms. (C) 1994 Academic Press, Inc.