Scalable Web Server

The exponential growth of the Internet, cou led with the increasing populari of dynamically generated content on the WorlJWide Web, has created the nee2 for more and faster Web servers capable of serving the over 100 million Internet users. Server clustering has emerged as a promising technique to build scalable Web servers. In this article we examine the seminal work, early products, and a sample of contemporary commercial offerings in the field of transparent Web server clustering. We broadly classify transparent server clustering into three categories. he exponential growth of the Internet, coupled with the increasing popularity of dynamically generated content on the World Wide Web, has created the need for more and faster Web servers capable of serving the over 100 million Internet users. The only solution for scaling server capacity in the past has been to completely replace the old server with a new one. Organizations must discard their investment in the old server and purchase a new one an expensive, short-term solution. A long-term solution requires incremental scalability, which provides the ability to grow gradually with demand. A pool of servers tied together to act as a single unit, or server clustering, provides such incremental scalability. Service providers may gradually add additional low-cost computers to augment the performance of existing servers. As Internet usage has grown, so has investigation into Web server clustering. The past four years have seen the emergence of several promising experimental server clustering approaches as well as a number of commercial solutions. All Web server clustering technologies are transparent to client browsers (i.e., the client browsers are unaware of the existence of the server cluster). However, not all clustering technologies are transparent to the Web server software. Early commercial cluster-based Web servers such as Zeus and Inktomi [l] are, in many respects, continuations of the traditional approach to cluster-based computing: treat the cluster as an indissoluble whole rather than the layered architecture assumed by (fully) transparent clustering. Thus, while transparent to clients, these systems are not transparent to the server nodes and require specialized software throughout the system. For example, Inktomi has a central point of entry and exit for requests, but nodes in the cluster are specialized to perform certain operations such as image manipulation and document caching. There is a coordinator that coordinates all the nodes to service client requests. In a similar vein, the Zeus Web server provides server clustering for scalability and availability, but each server node in the cluster must be running the Zeus Web server, a specialized server software developed for this environment. The cost and complexity of developing such proprietary systems is such that while they provide improved performance over a single-server solution, they cannot provide the flexibility and low cost service providers have come to expect with the wide array of Web servers and server extensions available. For this reason, our emphasis is on solutions that allow service providers to utilize commodity hardware and software. This implies that the clustering technique must be transparent to both the Web client and the Web server since the overwhelming majority of Web servers do not have any built-in clustering capabilities. While the emphasis of this article is on clustering in a Web server context, the technology is more generally applicable. Any server application may be clustered as long as it fulfills the following two properties: The application must maintain no state on the server. Any state information that is maintained must be maintained by the client. This prevents the cluster from having to deal with distributed state consistency issues. Note that some clustering agents do provide the capacity for some stateful services, but this is done on a service-by-service basis and is very protocol-specific. Clientherver transactions should be relatively short and high in frequency. As we are interested in commodity systems (hardware and software), we cannot decompose transactions into any smaller operations. Therefore, it is required that the transactions themselves be relatively small so that we can employ stochastic distribution policies to share the load more or less equally among all servers. 38 0890-8044/00/$10.00