Working Set Requirements and Performance of Network Caches in Cluster-Based Multiprocessors

This paper evaluates network caching as a means to improve the performance of cluster-based multiprocessors. A network cache, shared by all processors on each cluster, offers the potential benefits of increased intra-cluster sharing, reduced network traffic, and useful prefetching. Using simulation, we evaluate the feasibility, structure, and performance of a network cache implementation. Five well-known parallel scientific applications are used in this study. We experimentally derive the network cache working sets of these applications, and demonstrate that the size requirements of the network cache are feasible using current technology. Using cache sizes derived from this working set analysis, we compare a conservative network cache implementation to that of an aggressive directory-based scheme without a network cache. In all five applications, the inclusion of the network cache improves performance. Finally, we examine the effect on application performance of varying the network cac..