Um arcabouço baseado em componentes para computação paralela de larga escala sobre grafos

Abstract

Faced with the increasing growth of data production to be processed by computer systems, a result of the current technological context and emerging applications of both industrial and scientific interest, researchers and companies have been looking for solutions to leverage large-scale data processing and analysis capacity. In addition to the large volume, many of these data must be processed by high-complexity algorithms, highlighting the inherent difficulties of problems in large graphs (BigGraph), often used to model information from large databases. Although with limitations in graph processing, the MapReduce model has motivated the construction of several high-performance frameworks, in order to meet the demand for efficient large-scale general purpose systems. Such a context has led to the proposal of more specialized solutions, such as Pregel and GAS (Gather, Apply, Scatter), as well as MapReduce extensions to deal with graph processing. However, frameworks that implement these models still have limitations, such as multi-platform constraints and general propose programming models for graphs. In this work, we show how component-oriented parallel programming can deal with MapReduce and conventional Pregel constraints. For that, we have employed HPC shelf, a component-based cloud computing platform for HPC services. On top of this platform, we introduce Gust, a flexible, extensible and adaptable BigGraph framework based on MapReduce. Besides the gains in software architecture, due to the use of a component-oriented approach, we have obtained competitive performance results compared to the state-of-the-art through an experimental study, using estatistical methods to increase confidence.