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

Advances in multiprocessor interconnect technology are leading to high performance networks. However, software overheads associated with message passing are limiting the processors to get maximum performance from these networks, leading to under-utilization of network resources. Several research studies are ongoing for designing messaging protocols and hardware to reduce such overheads. However, with such protocols messaging overheads cannot be eliminated. Even though the overheads can be lowered, these will continue to be reasonably high compared to the network speed. In this paper we take an alternative approach and analyze on how to exploit emerging processor-cluster technology for designing high-performance and cost-eeective parallel systems given that messaging overheads can limit maximum performance. Though processor-clusters are being used in some systems in an ad hoc manner, there is no formal analysis in the literature to show when and how processor clusters beneet in designing high performance and scalable systems. In this paper we present a design-space-graph framework for analyzing and solving this problem by considering processor-clustering, messaging overheads, and network performance in an integrated manner. Our analysis establishes the following three design guidelines. With messaging overheads constraining performance in a system, processor clustering can be used to build a) an equal-sized system with a smaller network or b) a larger system with an equal-sized network. With messaging overheads not being a constraint, a combination of processor clustering and wider channels can be used to build a range of larger-sized systems. These guidelines are validated through simulation and experimentation. All these guidelines lead to designing cost-eeective and scalable parallel systems while delivering high performance.