LMRU: An Efficient Buffer Management Scheme for Large Sequential and Looping Reference Pattern Applications

The buffer cache management algorithm used by almost all operating systems is known as LRU(Least Recently Used). LRU is known to perform well in applications with irregular reference pattern. However, it shows poor performance for sequential-pattern or looping-pattern applications with large-size data. For sequential or looping reference pattern, MRU(Most Recently Used) replacement policy is frequently used to improve the buffer cache performance. On the way of implementing MRU on Linux, we observed that MRU shows lower response time by up to 100% compared to LRU. Indirect blocks, which are used in the file structure of Unix family operating systems for large-size file, are the main reason of decreasing performance. Indirect blocks are fetched but immediately replaced by MRU replacement policy, even those will be soon and frequently needed again. Based on this observation, we propose a buffer replacement policy named 'LMRU'. LMRU maintains frequently-used blocks such as indirect blocks in the cache, even it manages all other blocks on buffer cache with MRU. We have designed and implemented it in a Linux kernel. LMRU improves the response time by up to 70% compared to LRU and 163% compared to MRU.