Performance evaluation of COWS under real parallel applications

Clusters of workstations (COWS) are often arranged as a switch-based network with irregular topology. Usually, the evaluation of interconnection networks for COWS has been carried out by simulation using synthetic traffic and by traces from real parallel applications. Although both types of traffics are used as a first approximation of the behavior of the system, a more accurate behavior can be obtained by using real parallel applications. In this paper, a new simulation framework has been developed in order to evaluate interconnection networks under real parallel applications by using an execution-driven simulator. Moreover, the new simulator can be used to evaluate the impact on the performance of the whole system of several design parameters in addition to the interconnection network. Evaluation results show that the execution time of real parallel applications can be reduced by using an effective routing algorithm. Moreover, in some cases, the achieved improvements are higher than the ones achieved by improving other design issues, such as the processor instruction issue rate, the cache size or the network bandwidth.