Ììñôóööð Äó Blockin Blockinððøý Ò Ïïï Êêõùù×ø Ëøöööñ×

Abstra t Temporal lo ality of referen e in Web request streams emerges from two distin t phenomena: the popularity of Web obje ts and the temporal orrelation of requests. Capturing these two elements of temporal lo ality is important be ause it enables a he repla ement poli ies to adjust how they apitalize on temporal lo ality based on the relative prevalen e of these phenomena. In this paper, we show that temporal lo ality metri s proposed in the literature are unable to delineate between these two sour es of temporal lo ality. In parti ular, we show that the ommonly-used distribution of referen e interarrival times is predominantly determined by the power law governing the popularity of do uments in a request stream. To apture (and more importantly quantify) both sour es of temporal lo ality in a request stream, we propose a new and robust metri that enables a urate delineation between lo ality due to popularity and that due to temporal orrelation. Using this metri , we hara terize the lo ality of referen e in a number of representative proxy a he tra es. Our ndings show that there are measurable di eren es between the degrees (and sour es) of temporal lo ality a ross these tra es, and that these di eren es are e e tively aptured using our proposed metri . We illustrate the signi an e of our ndings by proposing and evaluating a novel Web a he repla ement poli y| alled GreedyDual*|whi h exploits both long-term popularity and short-term temporal orrelation in an adaptive fashion. Our tra e-driven simulation experiments show the superior performan e of GreedyDual* when ompared to other Web a he repla ement poli ies.