Fast and cost effective cache invalidation in DSM

Abstract

Most distributed shared memory systems use point-to-point networks in conjunction with directory-based cache coherence protocols. A cache invalidation transaction generates a number of unicast invalidation messages and as many acknowledgment messages. This results in heavy network traffic, high latency, and high occupancy at home nodes. This paper introduces a fast cache invalidation method, called collective cache invalidation (CCI), and its simple and cost effect implementation method, called virtual bus based collective cache invalidation (VCCI). The simulation results show that we can reduce the total network traffic up to 45% and the overall execution time up to 11% by VCCI. The proposed method keeps the system scalable-the growth rate of the implementation cost if O(N/spl radic/N), and, VCCI can reduce the complexity of coherence protocol and make directory controllers simple since it does not require acknowledgment messages.