项目名称: 面向气动CFD非线性求解的GPU/CPU混合并行JFNK算法研究
项目编号: No.11272352
项目类型: 面上项目
立项/批准年度: 2013
项目学科: 数理科学和化学
项目作者: 张理论
作者单位: 中国人民解放军国防科学技术大学
项目金额: 82万元
中文摘要: 气动CFD在工程应用中经常遇到非线性流动问题,在本质上需要采用非线性求解方法。JFNK(Jacobian-Free Newton-Krylov)方法是近几年发展起来的非线性求解方法,该方法具有存储量小、数据依赖关系小等特点,非常适合在GPU平台上应用。本项目拟以航空航天CFD典型应用为背景,研究基于GPU/CPU混合体系结构高性能计算机的JFNK并行方法,设计该方法的高效并行预条件子,并针对并行预条件JFNK方法,发展多区结构网格CFD混合并行负载平衡算法。通过本项目研究,设计出适用于气动CFD应用的预条件JFNK方法大规模混合并行算法,并在我国典型的混合异构高性能计算机平台上实现千万亿次以上计算性能的大规模并行实际应用。
中文关键词: 计算流体力学;非线性求解;无雅克比牛顿Krylov子空间方法;预条件子;GPU/CPU混合并行算法
英文摘要: Nonlinear fluid is common in aerodynamical engineering, and nonlinear solver is essentially important. Jacobian-Free Newton-Krylov is a newly developed nonlinear method with the advantage of small storage and loose data coupling. It is potential for GPU computing. With the background of CFD application in aviation and aerospace engineering, this project is aimed to research efficient JFNK method on GPU/CPU hybrid platform ,and develop suitable parallell preconditioners for it. Based on preconditioned JFNK method, load balancing algorithm for typical multiblock CFD will also be designed . Hybrid parallel load-balanced preconditioned JFNK will find application in large-scale aerodynamical CFD problem. The total goal of this project can concluded as to design hybrid preconditioned JFNK algorithm for typical aerodynamical CFD applicitions and achieve more than petascale performance on demostic-made hetergenerous high performance computer.
英文关键词: CFD;Nonlinear solver;Jacobian-free Newton-Krylov subspace method;preconditioner;GPU/CPU hybrid parallel algorithm