Efficient Simulation of Bandwidth Dynamics for Peer-To-Peer Networks, Master Thesis presentation by Alexandros Gkogkas
Supervisor: Roberto RoversoExaminer: Prof. Seif HaridiAbstract:Peer-to-peer (P2P) systems have become an increasingly popular solution to the problem of scalability in the context of content delivery. However, in such systems, participant peers have inherently diverse connectivity and bandwidth characteristics. Thus, in order to evaluate the performance of peer-to-peer applications, given the large scale and complicated network interactions, proper performance evaluation techniques are required. In particular, when evaluating P2P systems designed for content distribution, it is of vital importance to correctly simulate the bandwidth dynamics behavior of the underlying network. In this thesis, we propose a scalable and accurate flow-level simulation model which can be used to efficiently mimic real network bandwidth dynamics generated by peer-to-peer content distribution applications.
In detail, we model the way that TCP shares the bandwidth capacity between competing flows based on the max-min fairness strategy using the progressive filling algorithm. We modify such algorithm in a way that it can be applied on directed network models considering only access link capacity and avoiding simulating the structure of the core backbone network. Furthermore, we minimize the burden required to emulate the behavior of the network upon the start or the completion of a transfer. This is achieved by applying the bandwidth allocation algorithm only to the part of the simulated network which is affected by the new/finished transfer. For this purpose, we provide a generalized adaptation of an existing affected subgraph algorithm that can be used with our directed network model.
The evaluation of our solution focuses on the two most important characteristics of a P2P network simulator: scalability and accuracy. In our scalability study we compare the performance of our bandwidth allocation algorithm with the State of the art. we show that our algorithm outperforms the existing solu- tions for large scale and structured network overlays. Finally, we investigate the improvement in performance when using the affected subgraph algorithm. We show that the affected subgraph algorithm drastically reduces the simulation time and greatly increases scalability. Our accuracy study focuses on the deviation between the estimated transfer times given by our flow-level simulator and NS2. The results indicate that the the accuracy varies according to the bandwidth capacities of the nodes as well as the load of each node. The deviation is small for asymmetric bandwidth ca- pacities and for symmetric bandwidth capacities under high load. Overall, the max-min fair allocation of bandwidth follows the trends of the NS2 simulated bandwidth dynamics.