End-to-end network measurement and its applications in peer-to-peer networks.

摘要

End-to-end measurements between Internet hosts are important in understanding network characteristics and enhancing the performance of applications. In this thesis, we investigate various end-to-end measurement techniques to characterize network performance. A general end-to-end measurement framework, which summarizes the current research efforts and highlights the insights into the measurement practice, is proposed.We first investigate the end-to-end available bandwidth measurement problem through the use of an active probing stream. Based on the probing load, bandwidth probing methods can be classified into two categories: queueing model-based versus self-congestion probing. In the queueing model-based approach, we studied two probing sequences with different inter-arrival time distributions: Poisson and periodic. The study of the departure process of the probing stream leads to analyze M1 +M2 /GIi/1 and D + M/GIi/1 using the exact queueing analysis and the approximation methods. Of particular interest to our investigation, is the squared coefficient of variation (SCV) of the inter-departure process of the probing stream. In a real measurement system, given the measured SCV of the probing stream, the model is inverted to infer the load of the cross traffic on an end-to-end path.In the self-congestion approach, we propose a unified self-congestion probing framework for bandwidth measurement. The self-congestion probing for available bandwidth measurement and TCP congestion control can be treated as special cases of this framework. We design and evaluate a simple available bandwidth probing protocol to utilize different congestion signals experienced by the probing stream. The probing process consists of two phases: the rough estimation phase and the fine-granularity tuning phase. We demonstrate that the proposed probing method, utilizing Explicit Congestion Notification (ECN) signal provided by routers, is able to achieve accurate measurement, quick convergence speed and low probing overhead.We also explore various applications of end-to-end measurements in peer-to-peer (P2P) networks. In particular, we conduct end-to-end measurements to infer network-wide quality in mesh-pull P2P live streaming systems. Peers in these systems advertise buffer maps to each other, which summarize the chunks of the video stream that they currently have cached and made available for sharing. We demonstrate how buffer maps can be exploited to provide reasonably accurate estimates of ongoing video play-back quality throughout the network. To harvest buffer maps, we build a buffer-map crawler and also deploy passive sniffing nodes. We process the harvested buffer maps and present the measurement results for network-wide playback continuity, start-up delay, playback time lag among peers, and chunk propagation patterns.