【作者】 李丹丹；

【导师】 王群； 季晓慧；

【作者基本信息】 中国地质大学（北京） ， 地图制图学与地理信息工程， 2014， 博士

【摘要】 地下水资源的合理利用与水质监控是关系民生的大事。地下水流动系统的研究是实现地下水资源定量评价与预测的重要工具，为地下水资源的可持续利用提供了决定性支持。地下水流动规律反映的是地下水的水位随着时间与空间的变化规律，因此地下水流动系统研究的本质为地下水流动空间数据的研究。地下水流动系统研究的不断发展要求网格粒度日益精细以及研究尺度日渐扩大，庞大的数据计算量使得模拟时间过长而无法满足实际的需求。并行计算技术的出现为求解数据密集型的计算任务提供了一种非常有效的手段，因此，本文研究的目的在于借助并行计算技术实现地下水流动空间数据的快速且高效模拟。本文研究的重点是地下水流动空间数据有限差分数值模型的并行模拟，即借助并行计算的相关理论以及地下水领域已有的并行研究成果，针对不同的并行计算平台，提出地下水流动空间数据的并行模拟策略。论文的主要创新之处有：（1）基于分布式共享内存并行系统，提出了基于MPI/OpenMP混合编程模型的重叠型区域分解算法的并行模拟策略，即将一个大的地下水流动模拟问题划分为若干个小的子问题，并实现了进程与线程两层的并行粒度。基于该并行策略，网格规模为60×100×100的数值模型在SMP集群上具有4个子区域的并行模拟加速比可达到3.31。（2）基于CPU-GPU异构环境并行系统，提出了基于GPU的地下水流动空间数据并行模拟策略，即利用GPU加速地下水流动空间数据模拟流程中的模型计算。在支持GPU通用计算的并行平台，对地下水流动空间数据基于OpenMP、基于MPI/OpenMP以及基于GPU三种并行策略的模拟性能进行测试。实验结果表明，地下水流动空间数据基于GPU的并行模拟具有最佳的并行性能。（3）基于CPU-GPU异构环境并行系统，提出了基于线性求解器CUSP的MODFLOW并行策略，即利用求解器CUSP替代MODFLOW的迭代求解法。针对MODFLOW基于OpenMP与基于CUSP的两种并行策略，以TWRI_LARGE模型为例，将两种策略的并行性能进行比较。实验结果表明，基于CUSP并行的MODFLOW具有更好的并行性能。更多还原

【Abstract】 Rational use and quality monitoring of the groundwater is a major event of people’slivelihood. Research on the groundwater flow system is an important tool to achieve quantitativeevaluation and the forecast of groundwater resources. It also provides decisive supports for thesustainable use of groundwater resources. The groundwater flow pattern reflects the hydraulicheads change with the time and space. Therefore, research on the groundwater flow system canbe attributed to study the groundwater flow spatial data. The continuous development of thegroundwater flow system requires grid granularity finer and study area larger. The traditionalsimulation of groundwater flow spatial data makes simulation time too long and can not meet theactual demands. Emergence of parallel computing technology provides a very effective means tosolve data-intensive computing tasks. Therefore, the purpose of this thesis is using parallelcomputing technology to achieve a fast and efficient simulation of groundwater flow spatial data.This thesis focuses on the parallel finite difference numerical simulation of groundwaterflow spatial data. Based on the relevant theories and existing parallel research on groundwater,this thesis proposes appropriate parallel numerical simulation strategies of groundwater flowspatial data for different parallel computers. The main innovations of the thesis are as follows.I. This thesis proposes a MPI/OpenMP hybrid programming model based on overlappingdomain decomposition algorithm to parallelize the simulation of groundwater flow spatial data.That means dividing a large-scale groundwater flow spatial data problem into several smallsub-problems. This parallel strategy achieves a two-level parallelism granularity includingprocess and thread. For a numerical model with grid size60×100×100, the experimental resultsyield a speedup of3.31for four sub-regions on a SMP cluster with two nodes.II. This thesis proposes a GPU-based parallel simulation strategy of groundwater flowspatial data. That means solving the numerical model of the groundwater flow spatial data onGPU. On parallel platforms support GPU general-purpose computing, this paper compares theparallel performance of three parallel strategies which are groundwater flow spatial data basedon OpenMP, MPI/OpenMP and GPU. The experimental results yield that the GPU-based parallelstrategy has the best parallel performance. III. This thesis proposes a parallel MODFLOW based on a linear solver CUSP. That meansusing linear solver CUSP instead of the iterative methods in MODFLOW. This thesis makes aperformance comparison between parallel MODFLOW based on OpenMP and parallelMODFLOW based on CUSP. For the TWRI_LARGE numerical model, the experimental resultsyield the maximum speedup of5.31for the OpenMP-based MODFLOW when the simulation isparallelized by eight threads. On a computer with one GPU, the experimental results yield thespeedup of10.6for the CUSP-based MODFLOW.更多还原