Um mecanismo de proveniência para uma nuvem de serviços de Computação de Alto Desempenho orientada a componentes

Abstract

Scientific experiments are increasingly using large-scale data intensive computations, often requiring the availability of High Performance Computing resources. As experiments increase in scale and complexity, the ability to reproduce and understand the experiments becomes more difficult. Thus, the provenance data (data produced throughout the life cycle of the experiments) comply with a fundamental role for the transparency of experiments, evaluation of data quality, support to computational reproducibility and improvement in information management processes. The experiments, represented through scientific workflows, also benefit in the reuse, maintenance and evolution of its elements. The provenance data can be classified as prospective and retrospective. The prospective provenance represents the specification of computational tasks to achieve a result. The retrospective provenance consists of a structured and detailed history of the execution of computational tasks. The objective of this dissertation is to present an approach to the prospective and retrospective provenance in the context of HPC Shelf, aiming at the sharing and reuse of workflows in application projects and the support of the reproducibility of execution of workflows. HPC Shelf is a proposed HPC cloud service platform for the composition, deployment and execution of large-scale component-based parallel computing systems. In this way, HPC Shelf with registered provenance information obtains a better understanding of the behavior of the experiments, greater ease in problem diagnosis, interpretation of results, identification of critical regions of execution, among other uses. To validate our approach, we present a framework for building Gust systems operating on a set of graph theory problems.