Abstract
High-traffic Web sites have to be based on distributed Web-server architectures that provide a single interface to the users. This paper considers a system where user requests are dispatched among multiple servers by entities that work at two different dispatching levels. Coarse-grain level dispatching is carried out by the primary Domain Name System (DNS) server of the Web site and is applied to IP-address requests. Fine-grain level dispatching is achieved through a server redirection mechanism based on the HTTP protocol and acts on Web data requests. This latter level addresses the issue of limited DNS control on client requests due to the IP-address caching. However, server redirection should be used selectively because it increases user response time. This paper proposes and compares various mechanisms that limit request redirection while balancing the server load. Specifically, our simulation results show that redirection strategies that consider some characteristics of requested objects can reduce substantially the amount of redirections without affecting system load sharing.