三维自由面流动模拟中GPU并行计算技术  

Application of GPU acceleration techniques in 3D violent free surface flows

作　　者：李海州[1,2] 唐振远[1,2] 万德成[1,2] LI Haizhou TANG Zhenyuan WAN Decheng(State Key Laboratory of Ocean Engineering, School of Naval Architecture, Ocean and Civil Engineering, Shanghai Jiao Tong University, Shanghai 200240, China Collaborative Innovation Center for Advanced Ship and Deep-Sea Exploration, Shanghai 200240, China)

机构地区：[1]上海交通大学船舶海洋与建筑工程学院海洋工程国家重点实验室,上海200240 [2]高新船舶与深海开发装备协同创新中心,上海200240

出　　处：《海洋工程》2016年第5期20-29,39,共11页The Ocean Engineering

基　　金：国家自然科学基金资助项目(51379125;51490675;11432009;51579145;11272120);长江学者奖励计划(T2014099);上海东方学者岗位跟踪计划(2013022);工信部数值水池创新专项VIV/VIM项目(2016-23/09)

摘　　要：MPS(Moving Particle Semi-implicit)法能够有效地处理溃坝、晃荡等自由面大变形流动问题。在三维MPS方法中,粒子数量的急剧增加会导致其计算效率的降低并限制其在大规模流动问题中的应用。基于自主开发的MPS求解器MLParticleSJTU,本文对求解过程中耗时最多的邻居粒子搜寻和泊松方程求解两个模块采用了GPU并行加速,详细探讨了CPU+GPU策略。以三维晃荡和三维溃坝这两种典型的自由面大变形流动为例,比较了CPU+GPU相对于MLParticle-SJTU串行求解时的加速情况,结果表明CPU+GPU在邻居粒子和泊松方程这两个模块中的加速比最高能达到十倍左右。此外,采用CPU+GPU并行能够较准确地模拟溃坝、晃荡等自由面大变形问题。MPS（ Moving Particle Semi-implicit） is very suitable to deal with free surfaces flows of large deformation such as dambreaking and sloshing flows. However,the computational efficiency of the MPS method becomes a bottle neck for large-scale engineering applications with the increase of the number of particles,especially for 3D flow problems. Based on our in-house solver MLParticle-SJTU,the present work applies GPU technique in searching for neighboring particles and solving pressure Poisson equation,which are two most time-consuming parts in the MPS. The acceleration performance is analyzed according to 3D sloshing and 3D dam breaking flows. The results show that a speedup up to 10 x can be achieved by CPU ＋ GPU acceleration compared to that by only one CPU.

关 键 词：MLParticle-SJTU求解器 邻居粒子搜索 GPU并行技术 稀疏矩阵求解 溃坝 晃荡 

分 类 号：O353[理学—流体力学]