The previous coarse-grain sharing optimizations track information at two granularities: the cache tracks information for fine-grain blocks, and the snoop filter tracks information for coarse-grain regions. Patel and Ghose extend this basic concept by tracking information at a third granularity. The snoop-filter table (SFT) tracks sharing information for coarse-grain regions within a larger page. For example, an entry in the SFT might track information for the four 1KB regions within a 4KB page of memory. For each of these regions, the SFT tracks whether it is shared by other processors and how many blocks it currently has in the cache. When a block from a new page is first brought into the cache, a snoop broadcast is performed and sharing information is collected for each of the regions within that page. Tracking information at two coarse granularities allows the SFT to extract the most benefit from spatial locality while minimizing the effects of false sharing.

Avadh Patel, Kanad Ghose: Energy-efficient MESI cache coherence with pro-active snoop filtering for multicore microprocessors. ISLPED 2008