基于OpenCL的泛Kriging算法并行化研究

Parallelization research of Kriging algorithm based on OpenCL

针对泛Kriging插值算法在大量数据处理时的高耗时问题,该文从异构平台主机端与OpenCL设备端的交互方式入手,采用OpenCL异构平台开发语言进行泛Kriging算法并行化实现研究,解决了其在进行大数据量处理时数据存储、数据交互、多设备调度等一系列问题。在K20Xm平台上使用不同的数据集和参数对并行泛Kriging算法进行测试,实验结果表明:与Intel Xeon E5-2670CPU平台相比,并行程序插值部分加速比达到40倍以上,整体并行程序加速比达到了18倍。

Under the condition of massively Geo-spatial data processing, the seral universal Kriging interpolation algorithm is time-consuming. Point to this, the design and implementation of a parallel universal Kriging algorithm on the heterogeneous computing platforms from the point of the interactive way of the host end and the OpenCL end was presented in the paper. Also, its advantages in solving a series of problems, such as data storage, data interaction and multi-devices scheduling that arise in computing big data analysis, are also provided. We present the testing results of the parallel universal Kriging algorithm with OpenCL on an NVIDIA K20Xm platform with different datasets and parameters. The experiment results indicate that the parallel algorithm can achieve a speedup up to 40 times in the core interpolation part and 18 times for the entire parallel algorithm when compared against the performance of a standard serial implementation on an Intel Xeon E5-2670 CPU.