Generalized Processor Sharing Performance Models for Internet Access Lines

For Internet service providers (ISPs) the proper dimensioning of Internet access lines is essential. Under-dimensioning generally leads to a less than acceptable end-to-end Quality of Service (QoS) perceived by the subscribers, and over-dimensioning is generally too expensive. In this paper, we demonstrate empirically that the Generalised Processor Sharing (GPS) model can be used to accurately describe the flow-level characteristics of the traffic carried by Internet access lines, based on analysis of IP packet-level traces with hundreds of subscribers connected to the Internet over high-speed access lines. Using properties of the GPS-model, we establish simple relations between the access line capacity and the utilisation of the access lines on the one hand and the download times of Internet objects on the other hand. Subsequently, we study the impact of the TCP slow start algorithm on the download times of objects. We propose simple and fast approximations for the average object download time as a function of the object size, encompassing the effect of the round-trip times, maximum window size and modem speed. Empirical data demonstrate the accuracy of the approximations.