Publication

Abstract : Snoop-based cache coherence protocols perform well in small-scale systems by enabling low latency cache-to-cache data transfers in just two-hop coherence transactions. However, they are not a scalable alternative as they require frequent broadcast of coherence requests. Token coherence protocols were proposed to improve the scalability of snoop-based protocols by removing a large amount of traffic due to broadcast responses. Still, broadcasting coherence requests on every cache miss represents a scalability issue for medium and large-scale systems. In this paper, we propose to reduce the number of broadcast operations in Token coherence protocols by performing an efficient fine-grain private-shared data classification and disabling broadcasts for misses to data classified as private. Our fine-grain classification is orchestrated and stored by the Translation Look-aside Buffers (TLBs), where entries are kept for a longer time than in local caches. We explore different classification granularity accounting for different storage overheads and their impact on filtering coherence traffic. We evaluate our proposals on a set of parallel benchmarks through full-system cycle-accurate simulation and show that a subpage-grain classification offers the best trade-off when accounting for storage, traffic, and performance. When running a 16-core configuration, our subpage-grain classification eliminates 40.1% of broadcast operations compared to not performing any classification and 13.7% of broadcast operations more than a page-grain data classification. This reduction translates into less network traffic (16.0%), and finally, performance improvements of 12.0% compared to not having a classification mechanism.