Object access distribution in the Web is governed by Zipf?s law, in general. This property leads to effective Web caches, which store the most popular objects and typically employ the LFU replacement policy, which achieves high, and often the highest, cache hit rates. However, Web cache design based only on Zipf?s law has two main disadvantages: (i) it does not exploit the temporal and spatial locality of user accesses on a per session basis, and (ii) LFU implementation is costly and impractical in many environments, because it requires statistics on all objects accessed since the beginning of a cache?s operation.

We consider all parameters of locality of references in the Web (temporal, spatial and popularity) and draw an analogy with processor caches. Given cache replacement policies that address different locality characteristics, we argue that there exist replacement algorithms that combine these characteristics and achieve high performance at a low cost. We describe Window-LFU (W-LFU), a policy that combines LFU and LRU and achieves better performance than LFU at lower cost. W-LFU exploits both Zipf?s law, and temporal locality by using the accesses in a recent time-window. Simulations with actual traces indicate that W-LFU provides better results than theoretically expected.