Bayesian Methods to Estimate Future Load in Web Farms

摘要

Web Farms are clustered systems designed to provide high availability and high performance web services. A web farm is a group of replicated HTTP servers that reply web requests forwarded by a single point of access to the service. To deal with this task the point of access executes a load balancing algorithm to distribute web request among the group of servers. The present algorithms provides a short-term dynamic configuration for this operation, but some corrective actions (granting different session priorities or distributed WAN forwarding) cannot be achieved without a long-term estimation of the future web load. On this paper we propose a method to forecast web service work load. Our approach also includes an innovative segmentation method for the web pages using EDAs (estimation of distribution algorithms) and the application of semi-naive Bayes classifiers to predict future web load several minutes before. All our analysis has been performed using real data from a world-wide academic portal.