Paralelização do algoritmo para Agrupamento de Dados MRDCA-RWL usando GPU e CUDA

Abstract

The increasing volume of data available on the Internet (Big Data) creates an urgent need for management. In this sense, generating knowledge from these data becomes an even greater computational challenge. To mitigate, techniques are applied in the Data Mining field as the Data Clustering methods. However, the performance of these techniques against Big Data is not satisfactory in relation to the execution time, since the traditional algorithms are sequential or synchronous, so this research seeks to explore new ways to accelerate the algorithms of Data Clustering as the implementation of Parallel Computing resources. Therefore, CUDA technology was selected after a review of the main parallelization techniques and what results were achieved. With the effect of increasing scalability for large data sets and performing a performance comparison, then a parallel version of the Dynamic Hard Clustering Algorithm with Relevance Weight for each Dissimilarity Matrix Estimated Locally (MRDCA-RWL) has been implemented. In this way, the experiments used ten known data sets and available in the UC Irvine repository. Therefore, the parallel version proposed by this work obtained an average acceleration of 16.7 times in the execution time, which represents a significant leap in the performance of the algorithm.