| Parallel computing, as opposed to sequential processing, plays a growing role in solving increasingly complex computational problems. Traditionally mainframes and top of the line workstations were used for scientific (high power) computing. This dissertation investigates parallel computing by means of a variable computer cluster approach. An original variable cluster, based on PC-class computers, was produced and implemented for the purpose of this research, along with the necessary algorithms and computer codes. The processing performance of the variable cluster was evaluated in the case of different computing workloads provided by high incidence computational algorithms for one and two-dimensional FFT, as well as by Laplacian field (mesh) algorithm calculations. In order to allow for comparison with other studies (for instance, Amdahl's work), SpeedUp and Efficiency served as main concepts for the analysis of collected experimental data. Performance gain and reliability depend on the type of computing problem, amount of data transferred, number of machines participating in the computation, as well as on the physical characteristics of the machines and on infrastructure. A discrete model explaining the experimental data is proposed; an additional continuous model is also developed. "Resonance" workloads are to a certain extent predicted by our modeling, and the relation with computational performance is specified. Useful insights into the appropriate match between the computational algorithm and the cluster architecture are documented in our study.; The implemented computer cluster was found to be a robust platform that could be used for development of engineering applications requiring greater computing power than regular workstations can deliver. For selected cases the processing performance of the variable cluster scaled linearly with the number of nodes involved in the computation. |
