A Performance Evaluation of P-EDR in Different Parallel Environments

* This work has been partially supported by the Spanish CICYT under grant TIC97-0897-C04-03 Abstract This paper presents an exhaustive performance analysis of the P-EDR algorithm in some parallel machines available nowadays. EDR algorithm is a recent result of our Research Group. It constitutes an extension of the well-known Parzen method, which can be applied when dealing with uncertainty. Management of uncertainty implies heavy computational loads in terms of both calculus and storage, so a parallel version of the algorithm is more adequate to solve this problem in a practical time, especially for samples of large sizes. P-EDR represents our parallel implementation of EDR. Our tests go from an expensive and powerful MPP (as an IBM SP2 constitutes) to a cost-effective solution, as clusters of PCs represent. In these tests we have run the P-EDR algorithm in two distinct clusters of PCs each one with a different type of interconnection network: a low-cost 100 Mbit/s Fast Ethernet and a 1.28 Gbit/s Myrinet. Results obtained confirm parallel computing on networks of PCs as an attractive alternative to MPPs for cost/performance reasons.