A Comparative Study of the Performance of a CFD program across different L inux Cluster A rchitectures

A bstr act L inux PC Clusters are a cost effective platform for parallel computational dynamics (CFD) applications. In this paper we compare the performance of one CFD application on multiple uid clusters. Different networking architectures, like channel-bonding and Flat Neighborhood Networks (FNN), that use inexpensive switched Fast E thernet along with one high performance network (Myrinet) are considered in our performance evaluation. T he cost effectiveness and overall performance of the clusters are compared using a large CFD code, L E STool. In this study single processor performance as well as the overall parallel performance using different sized benchmark cases are compared and evaluated to nd the most cost effective parallel platform. 1 I ntroduction L inux clusters can provide a viable and more cost effective alternative to conventional supercomputers. In order to obtain an optimal price/performance ratio it is important to carefully engineer the cluster toward the application needs. Since clusters are build from low-cost, commodity components it is feasible to have the cluster tailored to only one major application. In this scenario the cluster behaves like a piece of lab equipment designed for one speci c task. T he performance characteristics of the main application on a single node and on the whol