Using surface effect measures to model parallel performance
Belgacem Ben Youssef · 2010
Many factors affect the performance of parallel programs including idle time, wait time, and communication costs. The latter often constitute a significant source of overhead. We consider in this paper the impact of such costs on the performance of parallel implementations of three domain decompositions of a three-dimensional model for tissue growth on a cluster. These are regular domain decompositions that comprise a slab decomposition, a rod decomposition, and a block decomposition. Using a set of measures that qualify each decomposition, we explore their capability to predict the parallel performance of each implementation. The results of our experiments confirm the applicability of using these measures as predictors of performance for our application, and potentially for other similar ones.