基于多GPU的千万级高维空间实时检索

Research on Ten Million of High Dimensional Data Real-time Retrieval by Multi-GPUs

海量高维数据的近邻检索一直是多媒体信息领域的重要研究课题。本文提出一种基于多GPU的并行高维空间距离检索排序算法,通过并行优化空间距离计算及排序过程,并充分利用GPU硬件特性和众多的流处理器单元,算法能实现百万级的高维数据的实时检索。在此基础上,研究并利用多GPU架构,提升并行效率,拓展实时数据查询的数据规模。实验结果表明,本文算法可达到千万级别高维数据的实时精确检索,极大地拓展了高维检索的应用范围。

Near neighbor searching is always one of important research fields in the field of multimedia information,especially for massive high-dimensional data.In this paper,we present a fast high-dimensional data retrieval method based on multi-GPUs.By parallel optimizing distance calculation and sort,and extremely utilizing the features of GPU and its numerous stream processors,our algorithm can realize real-time index search for millions of high-dimensional data.Base on this,the architecture of multi-GPUs is studied and utilized to improve its parallel efficiency and extend the searching data scale.Experiment shows that our method can realize real-time accurate searching for ten million of high-dimensional data, which will greatly extend application fields of high-dimensional retrieval.