HLogP: um modelo de escalonamento para a execução de aplicações MPI em grades computacionais

• Autor(a) principal: Mendes, Helder de Amorim
• Data de Publicação: 2004
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da Universidade Federal Fluminense (RIUFF)
• Texto Completo: https://app.uff.br/riuff/handle/1/17120
• topic: Ciência da computação; Processamento paralelo (Computadores); Computação paralela; Escalonamento de tarefa; Escalonamento de processo; Modelagem computacional; Aplicações MPI; Grades computacionais; Computational grids; Modeling; MPI applications; Task (process) scheduling; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::TEORIA DA COMPUTACAO::COMPUTABILIDADE E MODELOS DE COMPUTACAO

In order to execute MPI applications efficiently in grid environments, MPI processes need to be allocated appropriately the resources available. Devising efficient process allocations requires a scheduling model which is sufficiently accurate to reflect the performance characteristics of the target system. It is still unclear if existing parallel program computing performance models are appropriate for grid environment. To date no model has been proposed for scheduling parallel applications in grids, due principally to the difficulty of taking in considerations characteristics such as the heterogeneity of both computation and communication resources. This work propose one such model for scheduling MPI applications in computational grids. The parameters of the model capture these heterogeneous characteristics, enabling efficient schedules to be obtained. Due to the dynamic behavior of grid environments, the models need to bee calibrated before the application is executed (for static scheduling) and during execution (for dynamic scheduling).A tool is presented which efficiently calibrates the parameters of the proposed model for use by static and dynamic grid schedulers. In addition, this work also comments on the observation that, in some cases, the two existing grid enabled implementations of MPI exhibit quite differing behaviors.