Estimating the service time of web clients using server logs

Abstract

This article proposes and evaluates measures for estimating the service time of a web client using server logs, only from the server side without introducing traffic into the network. The HTTP protocol is described as well as the different interactions between the web server, the communication components, and the web client application. The first measure is based on the time it takes for the web server application to deliver an object to its operating system, keeping in mind the buffer effect of the server network. The second measure also considers the inter-arrival times to the server application of the GET requests for the objects that are part of a web page. We propose formulas, validated experimentally, that relate the proposed measurements in the server with the different components that take part in a web transaction and the service time experienced by clients. We have carried out several experiments to evaluate the validity of the proposed measurements and the best measure estimates the service time of the client with an error below 20% for 90% of the requests. We observed a cyclic component in the measurements of the server that can simplify the estimation of future values. For this reason, the proposed measures may be used to know what is the service time the client perceives in each visit to a server. It can also be used by content distribution networks to choose between several replica located in different places in the Internet.