Cluster-partitioning approaches to mapping parallel programs onto a hypercube

Abstract

The process-to-processor mapping problem is addressed in the context of a local-memory parallel computer with a hypercube interconnection topology. Two heuristic cluster-based mapping strategies are compared - 1) a nearest-neighbor approach and 2) a recursive-clustering scheme. The nearest-neighbor strategy is shown to be very effective with planar graphs and graphs with a high degree of locality while the recursive partitioning heuristic is better with highly irregular graphs.