栅格地理计算并行算子对区域计算算法并行化的可用性分析——以多流向算法为例

发布时间：2018-06-01 19:53

本文选题：栅格地理计算 + 并行计算　；参考：《地球信息科学学报》2015年05期

【摘要】：栅格地理计算并行编程库的研发有助于实现对栅格地理计算算法的并行化。在现有的研究中,Qin等(2014)设计并初步研发的栅格地理计算并行算子(Pa RGO),在设计思路上能较好地隐藏与并行编程软硬件环境相关的复杂细节,实现栅格地理计算通用步骤的并行化,且较其他类似思路的编程库而言,Pa RGO能兼容多种常用的并行计算平台,具有明显优势。但Pa RGO目前在设计上仅直接支持本地、邻域及全局计算特点的栅格地理计算算法并行化,对于更为复杂的区域计算特点算法并行化的支持能力尚未探究。对此,本文选取栅格数字地形分析中具有区域计算特点、递归设计的多流向算法为算例,利用Pa RGO进行并行化设计、实现及测试,以计算时间、相对加速比和相对并行效率为定量指标。通过可运行性和并行性能进行评价,结果表明:Pa RGO虽然不能直接支持对递归的多流向算法进行并行化,但在根据多流向计算的原理将该递归算法转变为非递归的设计之后,可将算法由原区域计算改造为邻域迭代计算,就能利用Pa RGO实现并行化,并得到较好的并行效果。在集群环境下,MPI版本并行程序的并行效果优于MPI/Open MP混合版本。
[Abstract]:The research and development of grid geography computing parallel programming library is helpful to realize the parallelization of grid geography computing algorithm. In the present research, Qin et al. (2014) the parallel operator of grid geography computing, Pa RGOA, designed and developed preliminarily, can better hide the complicated details related to the hardware and software environment of parallel programming, and realize the parallelization of the general steps of grid geography calculation. Compared with other programming libraries with similar ideas, Pa RGO is compatible with many common parallel computing platforms and has obvious advantages. However, Pa RGO is only designed to directly support the parallelization of local, neighborhood and global computing algorithms, and the ability to parallelize the more complex regional computing features has not been explored. In order to solve this problem, this paper selects the region calculation characteristic in the raster digital terrain analysis, and takes the recursive multi-direction algorithm as an example. The parallel design, realization and test of Pa RGO are used to calculate the time. Relative acceleration ratio and relative parallel efficiency are quantitative indexes. The results of performance evaluation and parallelism show that although the multi-direction recursive algorithm can not be parallelized directly by RGO, the recursive algorithm is transformed into a non-recursive design based on the principle of multi-direction computation. The algorithm can be transformed from the original region computation to the neighborhood iterative computation, and the parallelization can be realized by using Pa RGO, and a better parallel effect can be obtained. The parallel effect of MPI parallel program in cluster environment is better than that of MPI/Open MP mixed version.
【作者单位】：中国科学院地理科学与资源研究所资源与环境信息系统国家重点实验室;中国科学院大学资源与环境学院;南京师范大学地理科学学院江苏省地理信息资源开发与利用协同创新中心;Department
【基金】：国家自然科学基金项目(41422109) 国家科技支撑计划项目(2013BAC08B03-4)
【分类号】：P208;TP338.6

【参考文献】