Ekman et al. propose Page-Sharing Tables (PST) to track sharing at a page granularity. A table is associated with the TLB that contains entries for all pages with blocks currently in the cache. Each entry contains the number of blocks from the page currently cached and a bit-vector with one bit per processor indicating which processors are caching blocks from that page. When a new page is brought into the cache a snoop broadcast is performed and all of the processors use a special bus with one wire dedicated to each processor to indicate which processors are sharing the page. This sharing information is then used to determine whether a broadcast is necessary for future access to that page. Although the authors suggest only using the PST to avoid unnecessary broadcasts, the bit-vector could potentially be used to send multicast messages instead of full broadcasts for pages shared by only a few cores. Because of the fixed sized of the PST structure, it is possible for the cache to contain blocks from more pages than can be tracked by the PST. When this happens, the PST and snoop filtering optimization are turned off. Saturating counters are then used to estimate the spatial locality in the cache, and when the counters indicate the PST should contain sufficient entries to track the cache contents, the cache is flushed and the PST and snoop filter optimization are turned on.

M. Ekman, F. Dahlgren, and P. Stenström: “TLB and Snoop Energy-Reduction using Virtual Caches for Low-Power Chip-Multiprocessors”. In Proc. of ACM ISLPED-2002