Beneets of Processor Clustering in Designing Large Parallel Systems: When and How? 1

Scaling the size of parallel systems while maintaining the system performance is an important problem. A default scaling approach by using larger networks to interconnect more processors works only up to a limited extent and the derived conngurations using this approach are not cost-eeective. Recent advents in VLSI and packaging technologies now ooer multiple processors on a single multi-chip module. This ooers potential to design larger systems by replacing each processor in a base system by a processor-cluster. Similarly, advents in the interconnection network technology now provide the ability to use wider and faster network channels. In this paper we analyze the scaling of systems by taking into account current interconnection and processor-integration technolgies. The impact of software overhead to send and receive messages in parallel system on its scalability is also demonstrated. Our results indicate that systems with high message overhead costs can be scaled in a cost-eeective manner by employing larger clusters as computing nodes. This can be achieved without changing the network until the messaging costs are balanced. A system with balanced messaging costs can be further scaled in a more cost-eeective manner by using larger clusters and wider network channels.