Concerning with On-Chip Network Features to Improve Cache Coherence Protocols for CMPs

Abstract

Chip multiprocessors (CMPs) with on-chip network connecting processor cores have been pervasively accepted as a promising technology to efficiently utilize the ever increasing density of transistors on a chip. Communications in CMPs require invalidating cached copies of a shared data block. The coherence traffic incurs more and more significant overhead as the number of cores in a CMP increases. Conventional designs of cache coherence protocols do not take into account characteristics of underlying networks for flexibility reasons. However, in CMPs, processor cores and the on-chip network are tightly integrated. Exposing the network features to cache coherence protocols will unveil some optimization opportunities. In this paper, we propose distance aware protocol and multi-target invalidations, which exploit the network characteristics to reduce the invalidation traffic overhead at negligible hardware cost. Experimental results on a 16-core CMP simulator showed that the two mechanisms reduced the average invalidation traffic latency by 5%, up to 8%.KeywordsProcessor CoreCache LineHome NodeDistribute Shared MemoryCache Coherence ProtocolThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.