A new class of web caching strategies for content delivery

Abstract

Least recently used (LRU) is the most commonly applied strategy to update the content of caches in computing and database systems as well as for web caching. Although some case studies have shown that LRU hit rates can be low when compared to web caching strategies being aware of complete statistics of past requests, proposed alternatives seem too complex to cope with constant update effort of LRU. In this work we start with an evaluation of the hit rates for LRU as compared to statistic-based caching strategies for Zipf distributed popularity of web content, which has been confirmed manifold as the relevant access profile to content on the Internet. We conclude that LRU has more than 10% absolute hit rate deficits not only in some special cases but over the entire relevant parameter range of Zipf law access pattern. We show that a 10% gain over LRU hit rates is already realized by the variant of score-gated LRU, whose mean updating effort is comparable to pure LRU. As another main advantage, score-gated LRU avoids most of the input traffic of a pure LRU strategy, which frequently reloads objects into the cache. Score-gated LRU keeps the cache content stable in case of unchanged popularity and loads new objects only when their score is increasing until it exceeds the score of a cached object.