High Performance Cache Replacement Policy For LLC

The performance gap between the microprocessor and the memory becomes large with advances in technology.In order to mitigate the performance gap,on-chip cache is widely used.Cache works based on the principle of data locality.However,many studies show that these caches are used with low efficiency when managed by conventional cache replacement policies,only a small portion of the cache block will be accessed again before they are eventually evicted.In this paper,we analyses cache access behaviors,and propose two cache management strategy to optimize the cache performance.Recently,many studies have shown that a simple change to the insertion policy can significantly improve cache performance for memory-intensive workloads.Compared to the traditional LRU replacement policy,these policies can make more accurate prediction to the new blocks,thus reducing cache pollution made by blocks which are accessed only once.At the same time,providing more opportunity to make those blocks having high locality stay in the cache for a longer time.However,these policies perform the same on cache hits,that is,when a cache block is hit,it will be promoted to the head of the orderd chain.This can potentially degrade cache performance when a cache block is re-referenced and then never reused again.In this paper,we proposed position-based promotion policy to further improve cache performance.When the cache block is hit,this policy will block promoted to the middle of the orderd chain.It gives the cache block enough time to recieve a next hit and no longer to avoid cache pollution.This policy takes both recency and frequency information into consideration at the same time.Our evaluations show that it achieves a speedup of 0.74%over the original RRIP.Position-base promotion policy improves cache performance,but it is static and can not adapt to the running programs.Secondly,the blocks will eventually be evicted from the end of the ordered chain,there is still room for improvement in cache performance.For these two issues,we propose a partition-based cache management policy.We divide the cache set into reuse groups and non-reuse groups.We learn the cache hit distribution to determine the number of blocks for reuse and non-reuse groups.Thus,partition-based cache management policy can dynamically adapted to different programs and different phases of running programg.We have also proposed value reduction and sampler techniques for counters to reduce hardware cost.Experiments show that partition-based cache management policy achieves a speedup of 5.35% over the original RRIP.