A Shared Last-Level Cache Management Policy for Inclusive Cache

For the problem that the traditional LRU replacement is unaware of the temporal locality in inclusive cache, a shared last-level cache (SLLC) management policy was presented for inclusive cache. With a cost-less bypass buffer stored the useless data beforehand, the policy could avoid the resource competition in SLLC between these data and highly reused data, while it still maintains the inclusion property. To further find out the least reused blocks to replace, a temporal locality detector applied was helpful to evict these blocks from SLLC as early as possible. Finally, benefited from adjustment mutually between two predictors, a unified management algorithm was proposed to bypass the useless blocks and replace the less reused blocks. Test results show that the approach reduces miss rate by 21.67% on average and improves the prediction accuracy up to 72%, while requiring less than 1% overhead of SLLC.