In a distributed environment, the management of resources, i.e., the load balancing policy, has ample importance, because it significantly affects user perceived performance. The load balancing policy in a cluster of Web servers assigns the stream of incoming requests to the servers of the cluster, striving for fast processing time per request and maintaining equally utilized servers. Since the load balancing policy deals with the stream of requests, its performance highly depends on the characteristics of the cluster workload.
Throughout this dissertation, we stressed that in a Web server the request interarrival process exhibits long-range dependence, and the request processing time is highly variable. Detailed analysis of the effects of these characteristics on the performance of the load balancing policy, allows us to develop policies that are aware of the cluster workload, and improve user perceived performance. In the following sections, we describe how we characterize the service process in a Web server by PH distribution and model it as a queueing system for performance analysis purposes. Such analysis guides us on the design of new load balancing policies that maintain low request slowdown even under dynamically changing workload characteristics.
This chapter is organized as follows. In Section 7.1, we give an overview of different Web cluster architectures. Section 7.2 outlines load balancing policies associated with clustered Web servers. Section 7.3 outlines the analytic models that we propose to analyze the performance of load balancing policies in Web server clusters. In Section 7.4, we describe a sized-based load balancing policy for clustered Web servers, and analyze its performance using analytic models. In Section 7.5, we propose EQUILOAD that is a refinement of the sized-based policy. In Section 7.6, we propose ADAPTLOAD, a load balancing policy that continuously adapts its parameters to the characteristics of the cluster workload. We conclude the chapter with a summary of the presented results.