| 摘要: |
| 针对二维水流模型计算效率慢的问题，采用多个GPU并行计算以提高模型计算效率。基于CUDA计算平台，采用区域分解和交界连通的并行策略，将模型计算区域分解为多个计算子区域，每个子区域采用GPU进行并行计算，各个子区域的上下游连接处构建连通交界进行水流信息交换，实现整个计算区域的水流演进模拟。以长江干流澄通段二维水流模拟为例，采用实测资料验证了多GPU并行计算模型的可靠性，分析了不同数量级计算网格下的GPU并行加速比。与串行计算相比，单个GPU并行计算加速比最大约为191倍，四个GPU并行加速比最大约为310倍。研究表明采用区域分解和交界连通的并行策略来实现二维水流模型的多GPU并行计算可以有效地大幅提高模型计算效率，在大范围区域或精细化的水流模拟中具有良好的应用价值。 |
| 关键词: 二维水流模型 并行策略 多GPU并行计算 CUDA平台 |
| DOI： |
| 分类号:TV131.2???? |
| 基金项目:国家重点研发计划项目(2022YFC3204501)；江苏省水利科技项目(2021008) |
|
| Research on Multiple GPUs Parallel Computing of 2D Water Flow Model |
|
NIU Shuai1, LIU Jiu-fu2, LI Xiao-hong3, WANG Kai3, XIONG Liu-ming4
|
|
1.Hydrology and Water Resources Department of Nanjing Hydraulic Research Institute;2.Nanjing Research Institute of Hydrology and Water Conservation Automation,Ministry of Water Resources;3.Jiangsu Surveying and Design Institute of Water Resources Co,Ltd;4.Jiangsu Suzi Engineering Consulting Co.
|
| Abstract: |
| The research on multiple GPUs parallel computation of two-dimensional water flow model aims to address the problem of slow computational efficiency in explicit discretization of two-dimensional shallow water equations using finite volume method. Multiple GPUs are used for parallel computation to improve the computational efficiency of the model. Based on the CUDA computing platform, a parallel strategy of region decomposition and connected boundary is adopted to decompose the computational region of the model into multiple computational sub regions. Each sub region is parallelly calculated using GPU, and the upstream and downstream connections of each sub region are constructed at the connected boundary for water flow information exchange to achieve the simulation of water flow evolution in the entire computational region. Taking the two-dimensional water flow simulation of the Chengtong section of the Yangtze River as an example, the reliability of the multi GPU parallel computing model was verified using measured data, and the GPU parallel acceleration ratio was analyzed under different order of magnitude computing grids. Compared to serial computing, the acceleration ratio for parallel computing on a single GPU is approximately 191 times, while the acceleration ratio for parallel computing on four GPUs is approximately 310 times. Research has shown that using parallel methods of region decomposition and boundary connectivity to achieve multi GPU parallel computing of two-dimensional water flow models can effectively and significantly improve the computational efficiency of the model, and has good application value in large-scale or refined water flow simulations. |
| Key words: Two dimensional water flow model Parallel strategy Multiple GPUs parallel computing CUDA platform |
