基于GPU加速的分布式水文模型并行计算性能  

Parallel computing performance of distributed hydrological model accelerated by GPU

作　　者：庞超 周祖昊[2] 刘佳嘉[2] 石天宇 杜崇[1] 王坤[2] 于新哲 PANG Chao;ZHOU Zuhao;LIU Jiajia;SHI Tianyu;DU Chong;WANG Kun;YU Xinzhe(School of Hydraulic&Electric-power,Heilongjiang University,Harbin 150080,China;State Key Laboratory of Simulation and Regulation of River Basin Water Cycle,China Institute of Water Resources and Hydropower Research(IWHR),Beijing 100038,China)

机构地区：[1]黑龙江大学水利电力学院,哈尔滨150080 [2]中国水利水电科学研究院流域水循环模拟与调控国家重点实验室,北京100038

出　　处：《南水北调与水利科技（中英文）》2024年第1期33-38,共6页South-to-North Water Transfers and Water Science & Technology

基　　金：国家重点研发计划项目(2021YFC3000205);江西省水利科技重大项目(202023ZDKT03);水利部重大科技项目(SKR-2022056)。

摘　　要：针对具有物理机制的分布式水文模型对大流域、长序列模拟计算时间长、模拟速度慢的问题,引入基于GPU的并行计算技术,实现分布式水文模型WEP-L(water and energy transfer processes in large river basins)产流过程的并行化。选择鄱阳湖流域为实验区,采用计算能力为8.6的NVIDIA RTX A4000对算法性能进行测试。研究表明:提出的基于GPU的分布式水文模型并行算法具有良好的加速效果,当线程总数越接近划分的子流域个数(计算任务量)时,并行性能越好,在实验流域WEP-L模型子流域单元为8712个时,加速比最大达到2.5左右;随着计算任务量的增加,加速比逐渐增大,当实验流域WEP-L模型子流域单元增加到24897个时,加速比能达到3.5,表明GPU并行算法在大尺度流域分布式水文模型计算中具有良好的发展潜力。With the development of distributed hydrological models towards larger watersheds and finer granularity,computational efficiency gradually became a bottleneck,and parallel computing technology emerged as an effective solution to this challenge.In the realm of parallel computing for distributed hydrological models,most of the existing studies have primarily focused on CPU-based parallel techniques,with relatively limited research on GPU-based parallel methods.Furthermore,investigations on distributed hydrological models incorporating physical mechanisms remain scarce.This study centered around the physically-based distributed hydrological model WEP-L(water and energy transfer processes in large river basins)and explored the utilization of GPU-based parallel computing techniques.From a spatial perspective,the WEP-L model divides the watershed into numerous sub-basin units,where each unit's runoff calculations are independent,offering spatial parallelism.The interdependencies between simulation units were taken into account while allocating jobs to several computer units for parallel execution.Consequently,the runoff process of the model was parallelized based on sub-basins,dividing the Poyang Lake basin into 8,712 sub-units,and employing GPU threads to execute parallel computations through kernel functions.It is founded that the distributed hydrological model's suggested GPU-based parallel approach significantly accelerated the process.With an increase in GPU thread count,the parallel computing time steadily reduced.The parallel performance was most efficient when the total thread count closely approached the number of divided sub-basins.In the experimental Poyang Lake basin with 8,712 sub-basin units in the WEP-L model,the maximum speedup reached around 2.5.Secondly,the performance of GPU parallel computing was influenced not only by the degree of parallelism but also by the computational workload.With an increase in computational workload,both serial and parallel computation times increased.However,due to the sma

关 键 词：基于GPU的并行算法 物理机制 分布式水文模型 WEP-L模型 计算性能 

分 类 号：P333.9[天文地球—水文科学] TV121.7[水利工程—水文学及水资源]