High-volume Web servers: Traffic patterns, performance implications, and resource management

Abstract

The control, management and optimization of performance-based measures in high-volume Web sites require a fundamental understanding of the user request patterns and the performance implications of such traffic patterns. In this talk we will first present an analysis of the request patterns found at various Web server complexes, including high-volume commercial and sporting event Web sites. Our analysis demonstrates complex traffic patterns that include both short-range and long-rangedependence structures and both light-tailed and heavy-tailed behaviors. We then investigate the impact of these complex user request patterns on the response-time distribution of different Web server environments. Our analysis illustrates some of the key difficulties in the management and control of system resources to satisfy certain types of service-level agreements. We then describe aspects of the system architecture and techniques employed at various commercial and sporting event Web sites that reduce the overhead of serving dynamic content and that improve the Web server response-time distribution. We will also present research on controlling the performance of different classes of user requests, including some new forms of service-level agreements. Various applications of our analysis and results will be discussed.