A Global Replacement Based on Actual Set Association

Recently, to reduce conflict misses introduced by the non-uniform distribution of memory references, global replacement is explored in set-associative cache. We propose a novel low-overhead global replacement mechanism based on actual set association. In our mechanism, the global replacement is partitioned into global set selection and local LRU replacement within set allowing data resources contention among sets. For the twelve benchmarks from SPEC CPU2000 suite, our mechanism can achieve an average miss rate reduction of 20.94% close to global LRU replacement with an extra storage cost of about 4.63%, compared to a baseline set-associative cache.