Métodos de auto-configuração em aplicações móveis par-a-par não estruturadas

• Ano de defesa: 2012
• Autor(a) principal: Diego Neves da Hora
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; UFMG
• Palavras-chave em Português: Redes adaptativas; MANET; Par-a-Par; Aprendizado de máquina; Computação; Redes de computadores
• Link de acesso: http://hdl.handle.net/1843/ESBF-8UEJ68

Peer-to-Peer (P2P) applications must be configured according to the environment in which they are executed, in order to achieve the maximum performance. The identification of the ideal parameter configuration requires the characterization of the network in each deployment. Usually, P2P networks employ a generic default configuration, which is suitable to most scenarios, however its performance is worse than that of the best manual configuration. This work investigates methods to automatically configure the parameters of mobile P2P applications on runtime. We employ the MAPE architecture, proposed by IBM to create autonomic systems, in order to devise two solutions for P2P self-configuration. We propose the P-AIMD and P-ML controllers, which configure the application on run-time using the four phases of MAPE controllers, namely monitoring, analysis, planning and execution. In P-AIMD, the analysis and planning phase employ the Additive Increase and Multiplicative Decrease (AIMD) algorithm used in TCP for congestion control. In P-ML, we employ classic machine learning techniques in the analysis phase, and then employ a simple planning algorithm that uses the classification performed in the previous step to identify the need for configuration changes. The controllers were evaluated in mobile ad hoc networks running the Gnutella protocol, however the proposed solutions are applicable to any unstructured P2P protocol. We compared the performance of the proposed solutions against the Expanding Rings protocol, and the simulation results show that P-AIMD and P-ML present a success rate that is 5.18 and 2.71 percent inferior to that of the best manual configuration, respectively.