Framework para investigação de mapeamentos de aplicações em arquiteturas manycore

• Ano de defesa: 2022
• Autor(a) principal: Lima, Denis Pereira
• Orientador(a): Pedrino, Emerson Carlos
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• 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: Mapeamento de tarefas; Arquiteturas manycore; Otimização multiobjetivo autoadaptativa; Tolerância a falhas; Eficiência energética; Balanceamento de carga
• Palavras-chave em Inglês: Task mapping; Manycore architectures; Self-adaptive multi-objective optimization; Fault Tolerance; Energy efficiency; Load balance
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAO
• Link de acesso: https://repositorio.ufscar.br/handle/20.500.14289/17580

This thesis proposes an implementation of a framework for mapping graphs onto manycore architectures with multi-objective metrics optimization. The aim is to propose a new approach in relation to the works found in the related literature. To validate this proposal, the following are presented: a calibration methodology and multi-objective mapping of tasks related to pattern detection in high-resolution images (binary and grayscale), and a proposal for a new self-adaptive methodology to be used in multi-objective algorithms for mapping applications for manycore architectures. The results obtained through the pattern detection and task mapping methodology on manycore architectures demonstrate a high rate of generalization and accuracy. This brings a new contribution regarding the use of the evaluated multi-objective algorithms, with the best performance obtained by the PESAII algorithm, which was not previously reported in the literature. The methodology related to the mapping and use of the self-adaptive strategy represents a complete study with the Hypervolume and IGD performance indicators, proving the greater effectiveness of PESAII for the Hypervolume metric. This also makes a new contribution regarding the NSGAIII and SPEA2 algorithms regarding the metric IGD, demonstrating the improvement of the obtained results in the use of the proposed self-adaptive strategy.