Predição do nível de cooperação em sistemas par-a-par de vídeo ao vivo a partir de métricas de centralidade

• Ano de defesa: 2012
• Autor(a) principal: Glauber Dias Gonçalves
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; UFMG
• Palavras-chave em Português: Topologia; IPTV; Centralidade; Incentivo; P2P; Sistemas de transmissão de dados; Computação; Videodigital
• Link de acesso: http://hdl.handle.net/1843/ESBF-8XFMPA

P2P architecture has been used successfully to reduce costs and increase scalability of Internet live streaming systems. In a P2P live broadcast, users (peers) exchange video chunks to each other and cooperate for the system to distribute the media content. To measure peer\\\'s cooperation, these systems use only upload and download rates collected periodically from peers. However, such measures may be susceptible to malicious peers lie about their cooperation, which is a problem for incentive mechanisms that provide level of quality of service according to peer\\\'s cooperation. In this work, we investigate alternative methods to obtain peers\\\' cooperation without relying specifically on the their upload and download rates. In particular,we assess the potential benefit of exploiting topological properties of P2P overlay network to predict, with reasonable accuracy, peer\\\'s cooperation.Our study relies on data collected from one of the currently most popular P2P live applications, i.e., SopCast, using a large number of PlanetLab machines. It encompasses two main steps. We first show that centrality metrics are reasonably strongly correlated with the peer\\\'s cooperation level, which is defined by the ratio of the total upload to the total download traffic the peer exchanged with its partners. Moreover, we also show that a peer\\\'s centrality remains reasonably stable over consecutive time windows. Motivated by these findings, we then develop a regression-based model to predict the level of cooperation of a peer in the following time windows given its centrality measures collected in the last window. Using our collected data, we show that our approach can produce reasonably accurate predictions.We still exploit topological properties to enable the detection of malicious peers which collud to increase their cooperation level and receive benefits from the system improperly. In this case, we investigate the use of the metric conductance and analyze scenarios where this metric can be useful, besides, its advantages and limitations related to other approaches.