Evaluations of Web Server Performance with Heavy-tailedness

Abstract

Providing quality of service (QoS) in a business environment requires accurate estimation of Internet traffic, especially HTTP traffic. HTTP traffic, mainly consisting of file transmission traffic, depends on the sizes of files with a heavy-tailed property. Analyzing the performance of specific Web servers and comparing server performance is important in QoS provisioning for user applications. In this paper, we present the experimental analysis of Web server performance using our active measurement. We capture activity data from 1984 diverse Web servers by sending measurement packets 20 times in order to evaluate the spatial properties, and we select 14 Web servers to send packets 2, 000 times in 5-second intervals to evaluate the temporal properties. The main contribution of our study is to provide methods of evaluating the temporal and spatial properties on any Web server by measuring from a remote observation host, and to illustrate the current activity of temporal and spatial properties in a series of figures. Crovella's well known work merely describes the self-similar properties for the total transmission time generated by a heavy-tailed file size distribution. We divided the total transmission time into network and server system dependable elements, of which the heavy-tailed properties are captured. We found that temporal properties consist of two factors: stability and activity in the Poisson process. Robustness of heavy-tailedness in terms of constructing elements of total transmission time was evident from the analysis of spatial properties. Finally, we observed the upper boundary and classified groups in a mean-variance plot of primitive elements of the Web server.