Melhorando a qos de streamig de vídeo adaptativo através de um loop de controle inteligente em redes programáveis

• Ano de defesa: 2024
• Autor(a) principal: Almeida, Leandro Cavalcanti de
• Orientador(a): Verdi, Fábio Luciano
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal de São Carlos; Câmpus São Carlos
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência da Computação - PPGCC
• Palavras-chave em Português: Plano de dados programável; Telemetria dentro da rede; Aprendizado de máquina; Aprendizado por reforço; Streaming de vídeo adaptativo
• Palavras-chave em Inglês: In-band network telemetry; Programmable data plane; Machine learning; Reinforcement learning; Adaptive video streaming
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAO::SISTEMAS DE INFORMACAO
• Link de acesso: https://repositorio.ufscar.br/handle/ufscar/19399

Video traffic constitutes a significant portion of Internet traffic, directly impacting the Quality of Service (QoS) for several applications sharing the network. Emerging on-demand video streaming technologies, like Dynamic Adaptive Streaming over HTTP (DASH), enable a degree of adaptability in video playback to match the quality levels provided by video service providers. However, from the perspective of network providers, monitoring and managing such applications pose considerable challenges due to their client-driven nature. In this work, we address these challenges and present solutions founded on two key pillars: i) contemporary programmable networks; and ii) artificial intelligence. We propose a solution that encompasses the Monitor-Analyze-Plan-Execute (MAPE) cycle, where monitoring and management mechanisms collaborate to enhance the QoS of DASH video streaming. In this work, we create a Smart Closed Loop, leveraging the capabilities of the Programmable Data Planes (PDP) and utilizing fine-grained measurements provided by In-band Network Telemetry (INT) to guide Machine Learning (ML) decisions. We designed and implemented a more precise method for estimating adaptive video service metrics, characterizing significant progress in the field of DASH service monitoring (M). Analyzing these estimates (A), the Smart Closed Loop can plan (P) execution (E) strategies within the network infrastructure that aim to deliver the video in better conditions. In this work, the preferred execution strategy is a probabilistic packet discard policy, due to DASH utilizing TCP as a congestion control approach. In this context, we revisited a well-known Active Queue Management (AQM) mechanism based on the RED algorithm, and inspired by it we developed our solution: ingress Random Early Detection (iRED). iRED is a disaggregated P4-AQM fully implemented in programmable data plane hardware (Tofino switches) that saves router resources. This algorithm not only conserves router resources but also aligns with the Low Latency, Low Loss, and Scalable throughput (L4S) framework. Considering the dynamic nature of video traffic, we design and implement a mechanism based on Deep Reinforcement Learning to fine-tune iRED parameters in real-time named Dynamic, Enhanced and Smart iRED (DESiRED). With DESiRED, we leverage the benefits attained in enhancing the quality of the DASH video service, making our solution adaptive to the dynamics of network traffic.