How Can We Design Better Networks for DSM Systems?

Abstract

Most DSM research in current years have ignored the impact of interconnection network altogether. Similarly, most of the interconnec- tion network research have focused on better network designs by using synthetic (uniform/non-uniform) traffic. Both these trends do not lead to any concrete guidelines about designing better networks for the emerging Distributed Shared Memory (DSM) paradigm. In this paper, we address these issues by taking a three-step approach. First, we propose a compre- hensive parameterized model to estimate the performance of an applica- tion on a DSM system. This model takes into account all key aspects of a DSM system: application, processor, cache/memory hierarchy, coherence protocol, and network. Next, using this model we evaluate the impact of different network design choices (link speed, link width, topology, ra- tio between router to physical link delay) on the overall performance of DSM applications and establish guidelines for designing better networks for DSM systems. Finally, we use simulations of SPLASH2 benchmark suites to validate our design guidelines. Some of the important design guidelines established in this paper are: 1) better performance is achieved by increasing link speed instead of link width, 2) increasing dimension of a network under constant bisection bandwidth constraint is not at all beneficial, and 3) network contention experienced by short messages is very crucial to the overall performance. These guidelines together with several others lay a good foundation for designing better networks for current and future generation DSM systems.

This research is supported in part by NSF Grant MIP-9309627 and NSF Career Award MIP-9502294.

To be presented in Workshop on Parallel Computer Routing and Communication (PCRCW’97), Atlanta, GA, June 1997