TL: A high performance buffer replacement strategy for read-write splitting web applications

Abstract

With the quickly increasing pressure of users' requests, more and more Web applications adopt the read-write splitting to improve the performance, which requires new buffer replacement strategy. Flash memory has emerged as a popular storage media, and we use it to replace hard disk for read-write splitting Web applications. Traditional buffer replacement strategies are suboptimal on flash memory, since flash memory has some distinguished features as out-of-place update and read-write asymmetry. In this paper, we design a flash-aware buffer replacement strategy Tri-List (TL), which emphasizes the I/O asymmetry of flash memory. Unlike other flash-aware buffer replacement policies which are based on LRU method, TL considers both recency and frequency of page requests to make better decision of buffer replacement. Experimental results on both synthetic and benchmarking traces shows that TL has up to 30.3% improvements than state-of-the-art flash-aware buffer management policies. © 2014 Springer International Publishing Switzerland.