Estimating the Heavy-Tail Index for WWW Traces

Heavy-tailed behavior of WWW traffic has serious implications for the design and performance analysis of computer networks. This behavior gives rise to rare events which could be catastrophic for the QoS of an application. Thus, an accurate detection and quantification of the degree of thickness of a distribution is required. In this paper we detect and quantify the degree of tail-thickness for the file size and transfer times distributions of several WWW traffic traces. For accomplishing the above, the behavior of four estimators in real WWW traces characteristics is studied. We show that Hill-class estimators present varying degrees of accuracy and should be used as a first step towards the estimation of the tail-index. The QQ estimator, on the other hand, is shown to be more robust and adaptable, thus giving rise to more confident point estimates