Implementation and performance evaluation of parallel computations by using cluster of networked workstations

We design, implement, and compare several parallel version codes of our shelterbelt turbulent flow model system which involves extremely strong non-linear processes, solid-air interactions, flow-pressure linkages, and turbulent feedback mechanisms. We evaluate the performance of each implementation and show that the parallel code's performance depends ors its design and coding strategies. The performance gains of the domain decomposition parallel version of our model system are quite satisfactory. For functional decomposition, we even obtained parallel performance loss. We analyze the speedup and its change with domain size and number of processors. There is tradeoff between reduced computation load and increased communication load, and the relative weight, which is related to domain size, number of processors, and problem specific features, determines the speed ratio and load balance ratio.