Resolução paralela verificada de sistemas de equações lineares : uma abordagem para eficiência energética utilizando DVFS

Detalhes bibliográficos
Ano de defesa: 2013
Autor(a) principal: Lara, Viviane Linck lattes
Orientador(a): Fernandes, Luiz Gustavo Leão lattes
Banca de defesa: Não Informado pela instituição
Tipo de documento: Dissertação
Tipo de acesso: Acesso aberto
Idioma: por
Instituição de defesa: Pontifícia Universidade Católica do Rio Grande do Sul
Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência da Computação
Departamento: Faculdade de Informática
País: Brasil
Palavras-chave em Português:
Área do conhecimento CNPq:
Link de acesso: http://tede2.pucrs.br/tede2/handle/tede/6480
Resumo: Solving Systems of Linear Equations is important in several domains. In many cases, it is necessary to employ verified computing to achieve reliable results. With the support of High Performance Computing (HPC), solve efficiently huge linear systems with Verified Computing has become possible. Recently, HPC researchers have started to investigate solutions focused not only in performance but also in energy efficiency as well. In this context, the main goal of this work is to propose the use of DVFS (Dynamic Voltage and Frequency Scaling) technique to change the CPU frequency during the execution of a solver that employs Verified Computing. Furthermore, this works intends to present a case study aiming at verifying if the use of DVFS can provide a reduction on energy consumption without perfomance and accuracy being compromised. Initially, a study about the FastPILSS solver was carried out to evaluate its accuracy, performance and energy consumption over several different input matrices. After that, we observed that the use of DVFS does not affect accuracy. Analysing the results, no reduction in energy consumption using the powersave governor was observed if compared to the energy consumption using the performance governor. This occurs due to the significant increase in execution time. When the frequency was changed in isolated steps of the solver algorithm, it was possible to reduce up to 3,29% the energy consumption for dense matrices during the approximate inverse calculation.