Modeling correlations in web traces and implications for designing replacement policies

Abstract

A number of web cache-related algorithms, such as replacement and prefetching policies, rely on specific characteristics present in the sequence of requests for efficient performance. Further, there is an increasing need to synthetically generate long traces of web requests for studying the performance of algorithms and systems related to the web. These reasons motivate us to obtain a simple and accurate model of web request traces. Our Markovian model precisely captures the degrees to which temporal correlations and document popularity influence web trace requests. We describe a mathematical procedure to extract the model parameters from real traces and generate synthetic traces using these parameters. This procedure is verified by standard statistical analysis. We also validate the model by comparing the hit ratios for real traces and their synthetic counterparts under various caching algorithms. As an important by-product, the model provides guidelines for designing efficient replacement algorithms. We obtain optimal algorithms given the parameters of the model. We also introduce a spectrum of practicable, high-performance algorithms that adapt to the degree of temporal correlation present in the request sequence, and discuss related implementation concerns.