Continuous Measurement of Web Proxy Cache Efficiency

Web proxy caches are deployed almost everywhere, at organisation boundaries as well as at ISPs. Proxy caches improve latency when retrieving Web documents while reducing the overall network traffic on the Internet. Proxy cache administrators spend some effort configuring the caching software and determining their suitable placements in order to obtain the best achievable quality of service, with respect to user demands. However, despite this effort, the efficiency of caches may decline over time. Indeed, the efficiency of a given caching infrastructure depends heavily on the actual use of the caches: number of users, type of requested documents, load of machines running caches. Since these characteristics are subject to change, it is unlikely that any particular configuration and deployment will remain effective during the whole cache life cycle. Therefore, to maximise the benefits offered by Web proxy caches, administrators should reconfigure the existing infrastructure as the traffic characteristics evolve; this implies frequent measurements of cache efficiency. This raises two distinct issues. First, how often should one perform measurements? With respect to the frequency of evaluations, it is important to be able to react to sudden modifications (flash crowds, network or servers failures), as well as slower ones with strong trends (growing user community, increasing use of multimedia documents). The appropriate frequency depends mainly on the impact of the cache’s efficiency on the actual quality of service. Hence the best choice of frequency is particular to each system. Each administrator should be given a way to adapt the measurement frequency of his caches, taking into account the objectives in terms of overall service improvement and the effect of the degradation of performance. Second, how to cope with the inherent subjectivity of the notion of efficiency? The suitable efficiency metric depends on the goals defined by the cache administrators; the same level of performance may lead to different experiences by different