Non-uniform Partitioning of Finite Diierence Methods Running on Smp Clusters

A multicomputer or workstation cluster with multiprocessor nodes introduces signiicant need and opportunity for overlapping communication with computation. We evaluate partitioning strategies for an important application class, nite diierence methods, running on clusters of symmetric multiprocessors. Our results show that even for a regular, uniform nite diierence method, a non-uniform partitioning can give the best performance by judiciously overlapping communication and computation. We present an analytic performance model and experimental results to evaluate partitioning strategies, and discuss implications for SMP cluster programming models.