Characterizing temporal locality and its impact on web server performance

The presence of temporal locality in web traces has been long recognized. However, the close proximity of requests for the same file in a trace can be attributed to two orthogonal reasons: long-term popularity and short-term correlation. The former reflects the fact that requests for a popular document appear frequently thus they are likely to be close an an absolute sense. The latter reflects the fact that requests for a given document might concentrate around particular points in the trace due to a variety of reasons of reasons, such as deadlines or surges in user interests, hence it focuses on relative closeness. We introduce a new measure of temporal locality, the scaled stack distance, which is insensitive to popularity and captures instead the impact of short-term correlation, and use it to parametrize a synthetic trace generator. Then, we validate the appropriateness of this quantity by comparing the file and byte miss ratios corresponding to either the original or the synthetic traces.