Dynamic performance limits of the Benko-Veres Passive TCP packet loss estimation algorithm

In this paper we evaluate and discuss the performance of a passive TCP packet loss estimation algorithm described by Benko and Veres. Our analysis is motivated by recent work modeling web traffic patterns and the network conditions experienced by TCP sessions carrying web traffic. We experimentally determine the dynamic characteristics of the Benko-Veres estimation algorithm, focusing on how quickly the algorithm converges to a reasonable estimate of packet loss rate for a given TCP session and path. The Benko-Veres algorithm is stimulated in a small IP network testbed, using repeated TCP sessions over a path with variable and controlled packet loss rates. Our results show that the Benko-Veres algorithm requires many packets and loss events before it converges to reasonably accurate loss rate estimates. Finally we identify and discuss the limited real-world scenarios where this method of passive, external estimation of packet loss rates can be applied.