Neural network hot spot prediction algorithm for shared web caching system

Abstract

There are innumerable objects found on the Web. Both the popular objects that are requested frequently by the users and unpopular objects that are rarely requested exist. Hot spots are produced whenever huge numbers of objects are requested by the users. Often this situation produces excessive load on the cache server and original server, resulting in the system becoming a swamped state. Many problems arise, such as server refusals or slow operations. In this paper, a neural network prediction algorithm is suggested in order to solve the problems caused by the hot spot. The hot spot would be requested in the near future is prefetched to the proxy servers after the prediction of the hot spot; then the fast response for the users' requests and a higher efficiency for the proxy server can be achieved. The hot spots are obtained by analyzing the access logs file. A simulator is developed to validate the performance of the suggested algorithm, through which the hit rate improvement and the request among the shared proxy servers are load-balanced. © Springer-Verlag Berlin Heidelberg 2005.