APP: adaptively protective policy against cache thrashing and pollution

Abstract

Least Recently Used (LRU) is the most commonly used cache replacement policy; however, it suffers from two problems: i) cache thrashing, i.e., repeated references cause continuous page evictions due to a larger size of the working set than that of the cache, and ii) cache pollution, i.e., high reuse content gets evicted by items with low or no reuse from a cache. To solve these problems, prior works divide the cache into multiple segments and keeping the history of evicted pages, which impose high overhead in terms of memory. In this paper, we propose an adaptive cache replacement policy which divides the cache into two variable-sized segments: protected and unprotected. The division of cache segments is elastic in nature and can adaptively react to the workload changes without any history of evicted pages. We conduct extensive simulations using both synthetic and real workloads. Our evaluation shows that our policy can obtain the hit ratio close to the state of the art policies which keep history information of evicted pages up to multiple times of cache size.