End to End Performance Modeling of Web Server Architectures

Abstract

Web server performance in a distributed Object-Oriented (OO) environment is a complex interplay between a variety of factors (e.g., hardware platform, threading model, object scope model, server operating system, network bandwidth, disk file size, caching). In this paper, we present a model-based approach to Web Server performance evaluation in terms of an end-to-end queueing model implemented in a simulation tool. We have applied this model to Active Server Page (ASP) and Common Object Model (COM) technology in Microsoft's Internet Information Server and to the Java Server Page (JSP) and JavaBean technology in both IIS and Netscape Enterprise Server (NES). Our results indicate that for the ASP Script Engine, performance predictions from the simulation model matched the performance observed in a test environment. However, for the JSP Script Engine, the model predicted higher throughput than laboratory test results at high load. This result suggests that Web Server performance can be severely limited by a software bottleneck that causes requests to be serialized. This may cause a request to wait for some resource (i.e., a lock) as opposed to consuming CPU or memory. Implications of these results for Web Server performance in general are discussed