Algoritmo k-means em ambiente manycore para redução do tempo de resposta da mineração de dados

Abstract

Data mining (MD) is potentially costly, studies to decrease response time are essen- tial to deliver results in shorter times. Many solutions proposed by related work use computation in clusters of computers, an alternative to this is to use GPU computing. Programming with GPU requires a thorough knowledge of the algorithm to be worked on and an understanding of the GPU architecture that will be used. This work has the general objective to investigate the use of parallel computing in the manycore environ- ment to reduce the response time of MD algorithms. The K-means algorithm for being commonly adopted in AI tasks and its NP-Difficult feature was chosen to be paralyzed. Tools were used to identify the bottleneck of K-means, while evaluating the positives and negatives of these tools. After identifying the bottleneck of the algorithm, it was rewritten to run with GPU support, collected response times, and ended the performance gain measurement using the GPU. When using the GPU, a maximum speed of 7.09 and an efficiency of 0.65% was found to be small when compared to other studies in the literature. To circumvent this was done an increase in the database utilized by increasing the run time, thus obtaining better results with a speed up of 26.001 and a efficiency of 2.4% when using the maximum of colors of the GPU. It is concluded that it is possible to decrease the response time of data mining algorithms using GPU, without having to change the hardware of the equipment.