Research On Modeling And Measurement Of Large Scale P2P TV Systems

With the advancement of technology and popularization of the Internet, the Peer-to-Peer based Television (P2P TV) has been rapidly developing. Nowadays, P2P TV plays an important role in propagating knowledge and affecting the public opinion, and has grown into a new type of mass media. However, due to the essential features of P2P technology, such as non-centralization (or weak-centralization) and self-organization, it brings couples of problems in terms of network traffic management, video content censorship and public opinion supervision. In this dissertation, several popular P2P TV systems are measured using a self-designed crawler tool. We respectively investigate topological characteristics of the overlay network, information diffusion feature, users'spatio-temporal distribution and behaviors mode in these systems. This work is meaningful not only for improving and optimizing the P2P TV application protocol under practical network environment, but also serving as strong support for network video content censorship as well as public opinion detection, guidance and control. The main contributions are summarized as follows:(1) In this work, we first propose a distributed framework for active measurement in large scale P2P TV systems. Crawlers provide a principal method of P2P TV active measurement. Taking it into consideration, we build a distributed multi-protocol P2P TV crawler named TVCrawler, which is used to measure and investigate the live channels of three P2P TV systems: PPLive, PPStream and UUSee. Based on a large volume of live measurements, we collecte a large size of crawled data which opens up vast opportunities for further research work.(2) Through analyzing features of users'spatio-temporal distribution and factors which impact the distribution, we model users'spatio-temporal distribution of P2P TV systems. Highlights of spatio-temporal distribution model include temporal evolution analysis and geographic distribution of online users in the P2P TV systems. First, based on time series analysis, we model time evolution of online users. Using the model, we can predict number of online users in P2P TV channel. Multiscale entropy (MSE) analyzing method is utilized to investigate complexity in time series of online user's number. Second, we conduct a study in online users'geographic distribution, and map online users based on Google Map. On this basis, we then compare stimulateous online user number in different provinces of China with distinct economic development levels, indicating that stimultaneous online user number is inversely proportional to the developed level.(3) We develop a taxonomic approach of P2P TV users based on online session length, and then conduct a study in comparing users'arrival rate and dynamics in different categories. The taxonomy and in-depth study based on taxonomy provide us several insights in intrinsic characteristics of P2P TV user activities. First, using mixed exponent distribution model to fit the distribution of user session length, we clsssify P2P TV users into three categories: light watchers with 3~5 minutes of mean session length, medium watchers with 20~50 minutes of mean session length, and heavy watchers with 120~200 minutes of mean session length. Second, based on the taxonomy, we elaborate the analysis of users'stimultaneously online number and arrival rate. Finally, we find that the channel population of P2P TV is highly self-correlative, indicating that there exists a relatively stable user"subset"mainly composed of medium watchers and heavy watchers.(4) Based on empirical analysis of topological characteristic of P2P TV overlay network, a local evolution model is proposed for complex network. In this dissertation, we conduct a comparative research in topology characteristics and the behavior characteristics of P2P TV online user. Our research focuses on small-world phenomenon, degree distribution, mixed pattern, clustering phenomenon and overlay robustness in the P2P TV systems. We propose a novel complex network model generated through random walk and policy attachment (RAPA). Extensive simulations demonstrate that RAPA model can reproduce not only small-world and scale-free feature, but also some non-power-law features such as exponential cutoff or saturation for small variables. In addition, RAPA model can also construct some network with evident clustering structure and assortative/disassortative mixed pattern.