Performance Modeling and Evaluation of Web Server Systems with Proxy Caching

Recently, the Internet has been expanding intensively; the number of WWW (World Wide Web) users has been growing, and the network capacity provided by network carriers and Internet service providers has been increasing for enabling high quality services. The Internet users would pay the reasonable cost for a reasonably better service. For ISP (Internet Service Provider) to provide better QoS (Quality of Services) to users, it is important to design the entire network appropriately; that is, the bottleneck of the network should be identified adequately, and a well-balanced allocation of network resources to users should be performed. However, this is not an easy task: how the amount of network resources that satisfies user's QoS expectation can be determined? For typical WWW users, a response time (i.e., time elapsed until a requested document (i.e., a WWW page) has been successfully transferred) is one of most important performance measures. However, a response time of a requested document is difficult to expect since it is affected by both the network capacity and the processing power of the requested Web server. For predicting a response time of a Web server system, a relation between incoming traffic at the Internet access link and the processing delay of the Web server should be revealed. To design a Web server system, first, it is necessary to build the model of a Web i server, furthermore, a simple mathematical model is preferred since conventional computer simulation takes a lot of time for such a complicated network system. The first objective of this thesis is to model a single Web server without network. We perform various benchmark tests for an existing Web server, and investigate its characteristics in a quantitative manner. In our benchmarking experiments, a high–speed ATM switch is used to eliminate a possibility for the network to be the bottleneck. Benchmarking results show that the performance of the Web server can be improved by preparing the helper process for the http daemon. We quantitatively show a work demand, which is defined as the processing time for a given document size on a Web server. We then propose the performance model of a Web server, which consists of a FIFO (First-In First-Out) queue as a dispatcher and a PS (Processor Sharing) queue as a single processor used by all helper processes. Accordingly , the performance model of a Web server is modeled as an M/G/1/PS queue with a …