High performance computing on a cluster of workstations

The effective utilization of a cluster of workstations for the implementation of a scientific application requires a highly flexible software environment. The characteristics of such an environment are considered and two novel data distribution aspects of this environment are explored. The performance of a distributed memory multicomputer, such as a workstation cluster, is very sensitive to the strategy used to distribute data to the processors. A performance prediction scheme is presented which could be incorporated into a compiler to determine the best data distribution for a given program and system. For heterogeneous systems a data distribution strategy has been developed which takes into account the different capabilities of the processors. A number of experiments have been conducted on workstation clusters to demonstrate these software techniques