Suporte ao controle e alocação dinâmica de computadores em java

• Ano de defesa: 2005
• Autor(a) principal: Cera, Marcia Cristina
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Santa Maria; BR; Engenharia de Produção; UFSM; Programa de Pós-Graduação em Engenharia de Produção
• Palavras-chave em Português: Java; Programação paralela; Dinamicidade; Transparência; Parallel programming; Dynamism; Transparency; CNPQ::ENGENHARIAS::ENGENHARIA DE PRODUCAO
• Link de acesso: http://repositorio.ufsm.br/handle/1/8252

This thesis presents the project and an implementation of a distributed computer allocation system, based on computer idleness. The system, called Cadeo (control and dynamic allocation of idle workstations) aims to simplify the creation of parallel applications that can be executed on distributed systems. It supports a simple programming model, similar to the model found in parallel applications that run on shared memory computers. The Cadeo s execution platform is called dynamic cluster, which is a cluster composed by momentarily avaliable computers, or idle computers. Parallel execution is obtained by concurrent task execution, implemented as asynchronous remote method invocations in Java. Mapping application onto resources happens in two levels, first, associating applications to dynamic clusters, and, second, associating application tasks to computers of such clusters. Cadeo manages these associations offering both transparent task location and cluster dynamism, thus simplifying application development. The implementation was done with the purpose of having a basic and functional structure, which should be easily adapted in the future. This version was not focused on incorporating the best known algorithms to associate applications with clusters, or tasks with computers. Nevertheless, the implemented version strictly follows the proposed model, being now possible to write applications using the transparency and assynchrony accordingly. It was possible to demonstrate that, with the version implemented, the system overhead was very low.