CPU-GPU协同加速Kriging插值的负载均衡方法

Kriging插值算法被广泛应用于地学各领域,有着极其重要的现实意义,但在面对大规模输出网格及大量输入采样点时,不可避免地遇到了性能瓶颈。利用Open CL和Open MP在异构平台上实现了CPU与GPU协同加速普通Kriging插值。针对Kriging插值中采样点的不规则分布及CPU和GPU由于体系结构差异对其的不同适应性,提出一种基于不同设备间计算性能的差异和数据分布特点的负载均衡方法。试验结果表明,该方法能有效提高普通Kriging插值速度,同时还能节约存储空间和提高访存效率。

A load balancing method in accelerating Kriging algorithm on CPU-GPU heterogeneous platforms

Kriging interpolation algorithm is of great practical significance and was widely applied to various fields of geoscience. However, Kriging interpolation will encounter the performance bottleneck when the output grid or input samples increase. Implemented with OpenCL and OpenMP in this study, the Ordinary Kriging interpolation was accelerated on heterogeneous platforms: GPU and CPU. By considering the performance difference of CPU and GPU on the densities of samples, a new load balancing method i.e. LBCPDC (load balancing based on computation performance and data distribution) was proposed, in which not only hardware performance but also data distribution characteristics were taken into account. The experiment results show that LBCPDD method can effectively enhance the speed of Ordinary Kriging, save memory space and improve the efficiency of memory access.