基于图形处理器的数据流快速聚类

Fast Clustering of Data Streams Using Graphics Processors

在数据流环境下,聚类算法不仅需要有较高的聚类质量,同时需要有实时处理速度.因而,提出了一类基于图形处理器(graphics processing unit,简称GPU)的快速聚类方法,包括基于K-means的基本聚类方法、基于GPU的数据流聚类以及数据流簇进化分析方法.这些方法的共同特点是充分利用了GPU强大的处理能力和流水线特性.与以往具有独立框架的数据流聚类算法不同,这些基于GPU的聚类算法具有同一框架和多种聚类分析功能,为数据流聚类分析提供了统一的平台.从分析可知,数据流聚类分析的核心操作实际上就是距离计算和比较.基于这一认识,利用GPU的子素向量处理功能进行距离计算.性能验证实验是在配有Pentium IV3.4G CPU和NVIDIA GeForce 6800 GT显卡的PC上进行的.综合分析和实验结果表明,基于GPU的数据流聚类算法比传统的CPU算法平均快7倍,从而为高速数据流应用提供了良好的支持.

Clustering data stream basically requires fast processing speed as well as quality clustering results. In this paper, some novel approaches are presented for such a clustering task using graphics processing units （GPUs）, e.g., K-means-based method, stream clustering method, and evolving data stream analysis method. The common characteristics of these methods are making use of the strong computational and pipeline power of GPUs. Different from the pervious clustering methods with individual framework, the methods share the same framework with multi-function, which provides a uniform platform for stream clustering. In stream clustering, the core operations are distance computing and comparison. These two operations could be implemented by using capabilities of GPUs on fragment vector processing. Extensive experiments are conducted in a PC with Pentium IV 3.4G CPU and NVIDIA GeForce 6800 GT graphic card. A comprehensive performance study is presented to prove the efficiency of the proposed algorithms. It is shown that these algorithms are about 7 times faster than the previous CPU-based algorithms. Therefore, they well support the applications of high speed data streams.