|
Clusters of multiprocessors have a hybrid memory model at the hardware level. One specific problem consists in choosing the programming model for these machines, according to performance and programming effort. In this paper, we consider existing MPI codes and we examine when and why the unified MPI model is better than a fine grain hybrid MPI+OpenMP model. We compare the two models for NAS 2.3 benchmarks running on a cluster of PCs and two different IBM SP3 systems. The superiority of one model depends on 1) the level of shared memory model parallelization, 2) the communication patterns and 3) the efficiency for executing the pure parallel code with each model. Daniel Etiemble is a new professor in the Computer Engineering Group at the University of Toronto. His research interests are in the intersection of microelectronics (hardware for computing and VLSI design) and computer science (performance evaluation of CPU, memory hierarchies and computer architectures). |