An efficient parallelization of a real scientific application

Abstract

In the past decade the cost of computing has come down considerably making high-powered computing more easily affordable. As a result many institutions and organisations now have networks of high-powered workstations. Such networks provide a large, generally untapped, source of computing power which can be used for running large scientific applications which previously could only be run on supercomputers. This dissertation shows that a substantial improvement in performance can be achieved by the parallelization of a real scientific application for a heterogeneous network of Sun and Silicon Graphics workstations connected by an Ethernet network, but that this is affected by a number of factors. These factors include communication delays, load balancing, and the number of slaves used. This dissertation shows that performance can be improved by sending more, shorter messages, and by overlapping communication with computation. Part of this thesis concerns the difficulties involved in the evaluation of parallel performance on a heterogeneous network. This dissertation shows that conventional methods such as speedup and efficiency are not appropriate for evaluating the performance of a heterogeneous system, and that linear speed gives a much more representative indication of the actual performance achieved. We also proposed new concepts of perfect linear speed and linear efficiency, which help to evaluate the improvement in parallel performance on a heterogeneous system.