Padmanabhan and Mogul  have shown that HTTP/1.0 connections can increase server resource requirements, the number of network packets per request, and effective latency perceived by the client. They proposed persistent connections and pipelining of HTTP requests, which have been adopted by the HTTP/1.1 standard . The work in [19,21] shows that these techniques dramatically improve HTTP/1.0 inefficiencies. Our work provides efficient support for HTTP/1.1 on cluster based Web servers with content-based request distribution.
Heidemann  describes performance problems arising from the interactions between P-HTTP and TCP in certain situations. The work also proposes some fixes that improve performance. The proposed solutions are complimentary to our work and can be applied in our cluster environment. In fact, most of the proposed fixes are already incorporated in Apache 1.3.3 .
Much current research addresses the scalability problems posed by the Web. The work includes cooperative caching proxies inside the network, push-based document distribution, and other innovative techniques [20,7,10,16,17,27]. Our proposal addresses the complementary issue of providing support for HTTP/1.1 in cost-effective, scalable network servers.
Network servers based on clusters of workstations are starting to be widely used . Several products are available or have been announced for use as front-end nodes in such cluster servers [8,15]. To the best of our knowledge, the request distribution strategies used in the cluster front-ends are all variations of weighted round-robin, and do not take into account a request's target content. An exception is the Dispatch product by Resonate, Inc., which supports content-based request distribution . The product does not appear to use any dynamic distribution policies based on content and no attempt is made to achieve cache aggregation via content-based request distribution.
Hunt et al. proposed a TCP option designed to enable content-based load distribution in a cluster server . The design is roughly comparable in functionality to our TCP single handoff protocol, but has not been implemented.
Fox et al.  report on the cluster server technology used in the Inktomi search engine. The work focuses on the reliability and scalability aspects of the system and is complementary to our work. The request distribution policy used in their systems is based on weighted round-robin.
Loosely-coupled distributed servers are widely deployed on the Internet. Such servers use various techniques for load balancing including DNS round-robin , HTTP client re-direction , Smart clients , source-based forwarding  and hardware translation of network addresses . Some of these schemes have problems related to the quality of the load balance achieved and the increased request latency. A detailed discussion of these issues is made in the work by Goldszmidt and Hunt  and Damani et al. . None of these schemes support content-based request distribution.