Paralelização da resolução de EDPs pelo método Hopscotch utilizando refinamento adaptativo e balanceamento dinâmico de carga

• Ano de defesa: 2009
• Autor(a) principal: Guedes, Mauricio Jose Machado
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Programa de Pós-Graduação em Computação; Computação
• Palavras-chave em Português: Métodos numéricos; Solução numérica de EDPs; Método das diferenças finitas; Refinamento adaptativo de malhas; Método Hopscotch; Processamento paralelo; Balanceamento de carga; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::TEORIA DA COMPUTACAO::COMPUTABILIDADE E MODELOS DE COMPUTACAO
• Link de acesso: https://app.uff.br/riuff/handle/1/18753

In this work efficient parallel numerical solution of finite-difference discretizations of partial differential equations (PDEs) with the Hopscotch method is studied. A new technique for adaptive mesh refinement and dynamic load balance is proposed, and tests are presented on a distributed computing environment. Despite the many implementations already presented for the method, some of which employ parallel processing, none of them has used adaptive mesh refinement with load balance. Thus, in this work a package is built to solve PDEs discretized by finite differences, with adaptive mesh refinement and load balance. The domain is divided in subdomains, that are refined and coarsened during the execution. The process of subdomain refinement change can be chosen by the user, which provides the trigger values. A partition method that does not seek to minimize the communications is proposed. This method achieves the balance quickly, resulting in an easy-to-use technique for parallel computers and problems that frequently change the refinement. Tests are performed using three PDE types, elliptic, parabolic and hyperbolic. The tests start with a sequential version, proceed with the parallel version and end- up with the load balanced parallel version. Metrics are defined to evaluate the package performance.