Data Partitioning for a Good Node Performance

As a consequence of recent advances in interconnection network technology for MIMD parallel computers, optimizing communications in parallel programs has become a factor of secondary importance. For example, mapping processes onto processors is currently an issue of minor importance for some up-to-date distributed memory parallel computers, because the interconnection network guarantees a fairly uniform bandwidth between any pair of nodes1, 2. While communication performance is still an important issue in parallel processing on workstation clusters, single node performance optimization has become a crucial factor in parallel computers with a fast interconnection network. This optimization requirement is due to the sustained increment of the penalty for not efficiently using single node resources, like cache memories3, 4.