Modeling contention of sparse-matrix-vector multiplication (SMV) in three parallel programming paradigms

Abstract

We makes use of the LoPC model, which is inspired by both the LogP/PG and BSP models but accounts for contention for message processing resources in parallel programming models to derive a general estimate of execution cost. We carry out this cost estimate analysis for three dominate programming models: message passing, shared memory and distributed shared memory. We analyze a typical application SMV multiplication written in these programming models with irregular receiver-initiated synchronous communication. The LoPC estimate for this application is shown to be reasonably accurate when compared against measured values of runtime of actual empirical computations on an SGI O2000 multiprocessor machine.