Latency-aware self-adaptive micro-batching techniques for GPU-accelerated stream processing

• Ano de defesa: 2024
• Autor(a) principal: Leonarczyk, Ricardo Maciel
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Pontifícia Universidade Católica do Rio Grande do Sul; Escola Politécnica; Brasil; PUCRS; Programa de Pós-Graduação em Ciência da Computação
• Palavras-chave em Português: Stream Processing; Micro-batching; Multicores; GPUs; Processamento de Streaming; CIENCIA DA COMPUTACAO::TEORIA DA COMPUTACAO
• Link de acesso: https://tede2.pucrs.br/tede2/handle/tede/11639

Stream processing plays a vital role in applications that require continuous, lowlatency data processing. Thanks to their extensive parallel processing capabilities and relatively low cost, GPUs are well-suited to scenarios where such applications require substantial computational resources. However, micro-batching becomes essential for efficient GPU computation within stream processing systems. Micro-batching introduces the challenge of finding appropriate batch sizes to maintain an adequate level of service, particularly in cases where stream processing applications experience fluctuations in input rate and workload. Addressing this challenge requires adjusting the batch size at runtime, which can be done by using self-adaptation. In this thesis, we evaluated a set of existing and proposed self-adaptive algorithms for micro-batch size adaptation using a new GPU-accelerated stream processing application. Furthermore, we proposed a new set of metrics to help rank and compare the adaptation algorithms among themselves in terms of quality of service from different perspectives. The findings indicate that the tested application and its highly dynamic workload proved challenging for the existing algorithms previously evaluated in related work, making them 33% less effective in maintaining micro-batch latency for the most strict latency requirements. Among themselves, the algorithms attained comparable performance in maintaining latency within acceptable levels in the perspective of the stream data items. Furthermore, given the right set of parameters, one of our proposed algorithms could stay 40% closer to the target latency than others due to its ability to combine batch size fine-tuning and reactivity.