Analyzing a web-based system's performance measures at multiple time scales

Abstract

Web and e-commerce workloads are known to vary significantly from hour to hour, day to day, and week to week. The causes of these fluctuations are changes in the number of users visiting a site and the mix of services they require. Since the workloads are known to vary over time, one should not simply choose an arbitrary time interval and consider it as a reference for performance evaluation. We conclude that times scales are of great importance for operational analysis, particularly for systems with bursty loads. Service level agreements must certainly take into account measurement time scales. Similarly input parameters for predictive models are sensitive to time scale. Ultimately, a time scale should be chosen for service level requirements that best expresses the needs of end-users and the price the owner of a site is willing to pay for QoS.