Flow-level modelling of TCP traffic using GPS queueing networks

We develop flow-level models of TCP/IP networks. We first consider access network models in a very general framework where Internet subscribers can use several types of TCP-based applications and may have limited peak rates. Using multiclass networks of Generalized Processor Sharing queues (in the sense of Cohen), we derive key performance indicators in closed-form. We then give arguments in favor of a weighted max-min fair (WMM) allocation, with flow weights depending both on the number of TCP connections and on the RTT of the flow, as an approximation of TCP bandwidth-sharing. Finally, we build upon the models developed for access networks to propose an approximate model for the global network, assuming WMM fairness as bandwidth-sharing paradigm. This global network model allows to account for backbone congestion by replacing access links by virtual links whose capacities are computed in order to reflect the probability of congestion on flow path.