EXPLOITING LARGE GRAIN PARALLELISM IN A SPARSE DIRECT LINEAR-SYSTEM SOLVER

MCSPARSE is a parallel solver based on large grain parallelism, combin ed with medium and fine grain parallelism. For the multiple CPU Gray-s ystems, the large grain parallelism can be exploited using Gray's macr o-tasking while Gray's micro-tasking facilities can be used to impleme nt the medium grain parallelism of MCSPARSE. The fine grain parallelis m can be mapped on the parallel fine grain Gray processor. In this pap er, this design together with its performance results on the Gray Y-MP 4/464 are presented. More specifically, we study the impact of an ela borate reordering scheme H on the resulting efficiency of MCSPARSE.