Implementação paralela do metodo de resolução frontal de sistemas de equações
AUTOR(ES)
Gustavo Camargo Longhin
DATA DE PUBLICAÇÃO
2001
RESUMO
Obtaining the solution of a system of linear equations, generally results in a code elaboration which stores the matrix coeficients in the RAM memory and afterwards, some decomposition processes starts. The matrix is assembled summing up the contributions of each element from the domain to the problem s degrees of freedom. The bigger the problem, the larger the assembled matrix, therefore a higher requirement regarding the RAM memory capacity. From this, a procedure which does not assemble the matrix of coeficients prior to its decomposition would be more interesting. With that motivation, Bruce Irons developed in the beggining of the seventies a procedure which does not require an initial assembly of the global stiffness matrix. In this method a structure is defined where a totally added equation is immediatelly decomposed and the decomposition results are stored in an independent storage device. The matrix which receives the equations contribution was called frontal matrix and so was the method. On that frontal structure parallel optimization techniques are applied. Shared memory equipments are the hardware basis for the implementation and accordingly, public domain multithreading libraries based on the posix specification are used (pthread under GNU &Linux) for the multi-threading development. Results are shown comparing standart methods against the frontal solver as well as serial codes against parallel ones. Object oriented techniques are applied for the solvers development and planning. As a result, excelent degrees of modularity, extendibiIity, documentation and management are observed. The Unified Modelling Language (UML) utilization as a helpping tool for object oriented development was also very important
ASSUNTO(S)
algebra programação paralela (computação) metodo dos elementos finitos programação orientada a objetos (computação)
ACESSO AO ARTIGO
http://libdigi.unicamp.br/document/?code=vtls000260372Documentos Relacionados
- Resolução de sistemas de equações linearesprovenientes da simulação de estruturas fotonicas
- Algoritmos de Broyden combinados para resolução de sistemas de equações não-lineares
- Resolução de sistemas de equações lineares através de métodos de decomposição de domínio
- Avaliação de desempenho de método para a resolução da evolução temporal de sistemas auto-gravitantes em dois paradigmas de programação paralela:troca de mensagens e memória compartilhada
- MPI: uma ferramenta para implementação paralela