Approximate analysis of LRU in the case of short term correlations

One of the most widely considered cache replacement policies is Least Recently Used (LRU) based on which many other policies have been developed. LRU has been studied analytically in the literature under the assumption that the object requests are independent. However, such an assumption does not seem to be in agreement with recent studies of Web-traces, which indicate the existence of short term correlations among the requests. This paper introduces an approximate analysis that fairly accurately predicts the hit ratio of the LRU policy in the case of short term correlations. The approximation approach is based on the relation between the working set model and LRU, while the request generation process is assumed to follow a recently proposed model for Web-traces, which captures short term correlations among the requests. The accuracy of the introduced approximate analysis is validated for synthetic as well as real Web-traces.