Combining Reused Frequency, Most Recently Used and Program Counter Predictor as Last Level Cache Replacement Policy

Abstract

This paper proposed a last-level cache (LLC) replacement policy which evicts line based on reused frequency and most recently reused information; and inserts the line based on Program Counter (PC) Predictor prediction. The most recently used information is used to keep the most recently used (MRU) line from being evicted. The line with the lowest reused frequency value and non-MRU is the candidate of replacement. To avoid potential line from being evicted before it is being reused; a PC-based predictor is used to modify the reused frequency information of the incoming line during insertion. If a line is predicted to be reused, the proposed replacement policy will keep the line longer in the LLC memory. Simulation experiments are run using ChampSim at single-core and quad-core configurations with SPEC CPU 2006 and Cloudsuite benchmarks. The proposed replacement policy overall outperformed LRU by 0.94% while SRRIP outperformed LRU by 0.04%. The proposed replacement policy required additional 0.90% and 0.97% storage hardware compared to 2 MB and 8 MB LLC memory respectively.