Abstract EN:

Due to the large diffusion of P2P applications and especially P2P live-streaming, P2P traffic occupies an extremely large portion of overall Internet traffic. In this context, this thesis proposes new instruments to measure, identify and control P2P traffic. Regarding traffic classification, since traditional techniques have a hard time identifying P2P traffic, we propose a new behavioral classifier, Abacus, tailored for P2P live-streaming. Our experiments prove that Abacus, though based on simple counts of packets and bytes exchanged by a host, represents a lightweight and accurate solution for identifying P2P applications. Second, since the huge volume of traffic obliges operators to employ either flow-level monitors (e. G. NetFlow) or packet sampling to cut down the amount of measurement data, we evaluate the impact of data reduction on traffic characterization and classification. We show that Abacus can be adapted to this kind of data, suffering only a minor loss in accuracy and statistical classification remains possible if training and validation data are sampled at the same rate, in spite of the distortion introduced by packet sampling. Finally, we study a new transport protocol for P2P traffic, LEDBAT (Low Extra Delay Background Transport Protocol), the congestion control algorithm of the official BitTorrent client. This delay-based algorithm aims to provide an efficient, lower-than-best-effort service. Though faithful to its goals, the original design of LEDBAT appears affected by a latecomer advantage: we identify the main cause of the unfairness and propose effectives correction that restore the fairness.