Parallel Algorithms for Workstation Clusters

We investigate the potential of workstation clusters for use in high performance computation for some selected applications. Currently, the network speed found in most of the existing systems is quite low, but higher speed networks are already emerging in the market. We present four parallel algorithms that performed astonishingly well on a cluster of workstations connected by Ethernet. Three of these are algorithms for sorting, matrix multiplication, and all-pairs shortest path problems. The fourth algorithm solves computationally intensive numeric problems that require little communication. Since these numeric computations are easy to parallelize, they serve well for testing operating system strategies for designing a network-wide scheduler that utilizes idle workstations. These results appear to suggest that future progress in network speeds can potentially make workstation clusters serious competitors in high performance computation.