Cache Replacement Algorithms

Abstract

stantial performance gains in terms of an improved hit ratio compared with LRU for a wide range of cache sizes. ARC INTUITION ARC maintains two LRU page lists: L1 and L2. L1 maintains pages that have been seen only once, recently, while L 2 maintains pages that have been seen at least twice, recently. The algorithm actually caches only a fraction of the pages on these lists. The pages that have been seen twice within a short time may be thought of as having high frequency or as having longer term reuse potential. Hence, we say that L 1 captures recency, while L2 captures fre-quency. If the cache can hold c pages, we strive to keep these two lists to roughly the same size, c. Together, the two lists comprise a cache directory that holds at most 2c pages. ARC caches a variable number of most recent pages from both L 1 and L2 such that the total number of cached pages is c. ARC con-tinually adapts the precise number of pages from each list that are cached. To contrast an adaptive approach with a non-adaptive approach, suppose FRC p provides a fixed-replacement policy that attempts to keep in cache the p most recent pages from L 1 and the c − p most recent pages in L 2. Thus, ARC behaves like FRCp except that it can vary p adaptively. We introduce a learning rule that lets ARC adapt quickly and effectively to a variable workload. Many algorithms use recency and frequency as predictors of the likelihood that pages will be reused in the future. ARC acts as an adaptive filter to detect and track temporal locality. If either recency or frequency becomes more important at some time, ARC will detect the change and adapt its investment in each of the two lists accordingly. ARC works as well as the policy FRC p, even when that policy uses hindsight to choose the best fixed p with respect to the particular workload and the cache size. Surprisingly, ARC, which operates completely online, delivers performance compara-ble to several state-of-the-art cache-replacement policies, even when, with hindsight, these policies choose the best fixed values for their tuning para-meters. ARC matches LRU's ease of implementa-tion, requiring only two LRU lists.