DASH sobre OpenFlow: estimando métricas de QoS a partir da rede

• Ano de defesa: 2020
• Autor(a) principal: Lacerda, Marta Calasans Costa
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Uberlândia; Brasil; Programa de Pós-graduação em Ciência da Computação
• Palavras-chave em Português: DASH; Aprendizagem de máquina; HAS; OpenFlow; SDN; QoS; Machine Learning; Computação; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: https://repositorio.ufu.br/handle/123456789/30457; http://doi.org/10.14393/ufu.di.2020.305

Obtaining QoS metrics on the client is not trivial for most networked service providers. Recent work has indicated that, for OpenFlow networks, part of this information can be extracted from the network, using only aggregated traffic statistics as input data for machine learning methods capable of generating estimators. The objective of the research carried out and presented here was to proceed with this investigation, applying the same methodology to one of the most popular applications on the Internet, not yet investigated: video on demand applications with dynamically adaptable bit rates. For this, a test environment was implemented with a DASH service, whose communication between server and clients is carried out via the OpenFlow network, with physical switches. In this environment, data were collected to generate the models using two supervised machine learning methods: regression tree and random forest. The QoS metrics investigated were the frame rate per second (video component) and the \ textit {buffer} rate per second (audio component), which were evaluated according to the normalized mean absolute error (NMAE) and time training model. For the scenario under analysis, the results were extremely satisfactory with respect to the training time for the two QoS metrics investigated, staying in the order of milliseconds. With regard to the error, the audio metric showed better performance for the studied algorithms, staying around 13 \%, while the video metric had errors just above 16.5 \%. It was also found that the use of the ensemble method did not bring significant benefits to the results. In addition, the results revealed that the same knowledge about these metrics could be generated using predominantly aggregated data statistics from the switch that connects the client to the OpenFlow network. All of these results, however, need to be better investigated in future works that employ more data samples and bring the test environment closer to the real one.