Abstract Clusters of Personal Computers have

been proposed as potential replacements for expensive compute servers. One limitation in the overall perfor-mance is the interconnection network. A possible so-lution is to use multiple processors on each node of the PC cluster. Parallel programs can then use the fast shared memory to exchange data within a node, and access the interconnection network to communi-cate across multiple nodes. The performance of such a system is likely to be influenced by the runtime support system, whether one uses a standard message passing library or a distributed shared memory system to sim-ulate a large shared memory machine. This paper presents a cost-performance evaluation of a cluster of eight dual processor PCs and a ten pro-cessor SMP workstation using two benchmark suites. We evaluate the effect of using gigabit ethernet against fast ethernet, single processor PCs versus SMP nodes, and compare the overall performance of the PC clus-ter to the SMP workstation. It is found that the SMP machine provides the best performance in only 50 % of the programs examined. The PC cluster with a Gi-gabit network is better in most of the other cases, but the improvement over FastEthernet is marginal