面向流体机械仿真的层次化并行计算模型
|
摘要
随着流体机械基础并行算法的发展,传统的单核处理器已经不能很好地满足先进流体机械研发的技术需求,为此本文深入研究了流体机械的物理模型以及高性能计算机架构特点,设计并实现了能够充分表达物理模型并行性的高效的面向流体机械仿真的层次化并行计算模型(HP2H)。HP2H模型充分考虑流体机械的多层几何结构以及高性能计算机的多层逻辑架构,深入挖掘计算平台、计算模型以及物理模型的并行性,实现从物理模型到计算资源的高效任务映射。依据具体的轴流压气机转子数值模拟的实际应用背景,结合粗粒度并行和细粒度并行对模型进行实现。对HP2H计算模型进行了功能测试和性能测试,当计算核心从36核提升到432核时,计算性能提升约12倍,并行效率达到了100%。实验结果表明,HP2H计算模型不但在正确地对流体机械进行数值模拟的前提下实现了较好的计算性能,并且由于HP2H计算模型结合了粗粒度并行与细粒度并行,因而可以在不同的计算平台上运行,还可以便捷地实现计算规模的扩展,具有良好的可移植性与可扩展性。
The traditional single-core processor can not meet the technical requirements of advanced fluid machinery research and development with researches on the basic parallel algorithms of fluid machine,while the rapid development of high-performance computer provides a solution.Physical modeles of fluid machinery and the architecture of high-performance computers are deeply studied,and an efficient parallel computing model(HP2H)that can fully express the parallelism of the physical model is designed.HP2 Htakes fully account of multi-layer geometry of a fluid machinery and multi-layered logic architecture of a high-performance computer.Moreover,HP2 Hdeeply exploits the parallelism of computing platforms,computing models and physical models to achieve efficient task mappings from physical models to computing resources.According to the actual application background of the numerical simulation of axialflow compressor rotors,the model is parallelly implemented with coarse-grained parallelism and fine-grained parallelism.Functional tests and performance tests are performed on HP2 Hcomputing model.Results show that when the computing core is upgraded from 36 cores to 432 cores,the computational performance increases by 12 times and the parallel efficiency achieves100%.Experimental results show that HP2 Hachieves a good computing performance under the premise of correctly implementing the numerical simulation of the turbomachinery.It is the use of coarse-grained parallelism and fine-grained parallelism,HP2 H can run on different computing platforms and easily realize the expansion of computing scale,which means that the HP2 H computing model has a good portability and scalability.
The traditional single-core processor can not meet the technical requirements of advanced fluid machinery research and development with researches on the basic parallel algorithms of fluid machine,while the rapid development of high-performance computer provides a solution.Physical modeles of fluid machinery and the architecture of high-performance computers are deeply studied,and an efficient parallel computing model(HP2H)that can fully express the parallelism of the physical model is designed.HP2 Htakes fully account of multi-layer geometry of a fluid machinery and multi-layered logic architecture of a high-performance computer.Moreover,HP2 Hdeeply exploits the parallelism of computing platforms,computing models and physical models to achieve efficient task mappings from physical models to computing resources.According to the actual application background of the numerical simulation of axialflow compressor rotors,the model is parallelly implemented with coarse-grained parallelism and fine-grained parallelism.Functional tests and performance tests are performed on HP2 Hcomputing model.Results show that when the computing core is upgraded from 36 cores to 432 cores,the computational performance increases by 12 times and the parallel efficiency achieves100%.Experimental results show that HP2 Hachieves a good computing performance under the premise of correctly implementing the numerical simulation of the turbomachinery.It is the use of coarse-grained parallelism and fine-grained parallelism,HP2 H can run on different computing platforms and easily realize the expansion of computing scale,which means that the HP2 H computing model has a good portability and scalability.
引文
[1]国家自然科学基金委员会.工程热物理与能源利用学科发展战略研究报告[R].北京:科学出版社,2011:72-75.
[2]中国科学技术协会.2009—2010工程热物理学科发展报告[R].北京:中国科学技术出版社,2010:24-45.
[3] GICQUEL L Y M,GOURDAIN N,BOUSSUGE J F,et al.High performance parallel computing of flows in complex geometries[J].Comptes Rendus Mecanique,2011,339(2/3):104-124.
[4] GILES M B,REGULY I.Trends in high-performance computing for engineering calculations[J].Philosophical Transactions of the Royal Society,2014,372(2022):319-332.
[5]王运涛,王光学,徐庆新,等.基于结构网格的大规模并行计算研究[J].计算机工程与科学,2012,34(8):63-68.WANG Yuntao,WANG Guangxue,XU Qingxin,et al.A study of the massively parallel computation based on structured grids[J].Computer Engineering and Science,2012,34(8):63-68.
[6] HART A,ANSALONI R,GRAY A.Porting and scaling OpenACC applications on massively-parallel,GPU-accelerated supercomputers[J].European Physical Journal Special Topics,2012,210(1):5-16.
[7] SIVANANDAN V,KUMAR V,MEHER S.Designing a parallel algorithm for heat conduction using MPI,OpenMP and CUDA[J].Parallel Computing Technologiess,2015:1-7.
[8] LONCAR V,YOUNG-S L E,SKRBIC S,et al.OpenMP,OpenMP/MPI,and CUDA/MPI C programs for solving the time-dependent dipolar GrossPitaevskii equation[J].Computer Physics Communications,2016,209:190-196.
[9] JACOBSEN D A,THIBAULTY J C,SENOCAK I.An MPI-CUDA implementation for massively parallel incompressible flow computations on multi-GPU clusters[C]∥Proceedings of the 48th AIAA Aerospace Sciences Meeting Including the New Horizons Forum and Aerospace Exposition.Reston,VA,USA:American Institute of Aeronautics and Astronautics Inc.,2010:0522.
[10]刘冰,陆忠华,李新亮,等.基于GPU的多重网格Navier-Stokes解算器并行优化方法研究[J].科研信息化技术与应用,2013,4(3):56-67.LIU Bing,LU Zhonghua,LI Xinliang,et al.Implementation and optimization of a Navier-Stokes solver with multigrid methodology based on GPU[J].E-Science Technology and Application,2013,4(3):56-67.
[11]胡银丰,孔强.基于MPI+OpenMP混合编程模型的并行声纳信号处理技术研究[J].舰船电子工程,2017,37(12):130-133.HU Yinfeng,KONG Qiang.Research on parallel sonar signal processing technology based on MPI+OpenMP mixed programming model[J].Ship Electronic Engineering,2017,37(12):130-133.
[12]李春.混合并行计算在海洋水质预报系统中的应用研究[D].石家庄:河北工程大学,2014:36-50.
[13]HAH C,KRAIN H.Secondary flows and vortex motion in a high-efficiency backswept impeller at design and off-design conditions[J].Journal of Turbomachinery,112(1):7-13.
[14]曹维.大规模CFD高效CPU/GPU异构并行计算关键技术研究[D].长沙:国防科学技术大学,2014:39-50.
[15]张楚华,琚亚平.流体机械内流理论与计算[M].北京:机械工业出版社,2016:129-145.
[16]刘超群.多重网格法及其在计算流体力学中的应用[M].北京:清华大学出版社,1995:1-49.
[17]TAI C H,ZHAO Y,LIEW K M.Parallel-multigrid computation of unsteady incompressible viscous flows using a matrix-free implicit method and high-resolution characteristics-based scheme[J].Computer Methods in Applied Mechanics and Engineering,2005,194(36):3949-3983.
[2]中国科学技术协会.2009—2010工程热物理学科发展报告[R].北京:中国科学技术出版社,2010:24-45.
[3] GICQUEL L Y M,GOURDAIN N,BOUSSUGE J F,et al.High performance parallel computing of flows in complex geometries[J].Comptes Rendus Mecanique,2011,339(2/3):104-124.
[4] GILES M B,REGULY I.Trends in high-performance computing for engineering calculations[J].Philosophical Transactions of the Royal Society,2014,372(2022):319-332.
[5]王运涛,王光学,徐庆新,等.基于结构网格的大规模并行计算研究[J].计算机工程与科学,2012,34(8):63-68.WANG Yuntao,WANG Guangxue,XU Qingxin,et al.A study of the massively parallel computation based on structured grids[J].Computer Engineering and Science,2012,34(8):63-68.
[6] HART A,ANSALONI R,GRAY A.Porting and scaling OpenACC applications on massively-parallel,GPU-accelerated supercomputers[J].European Physical Journal Special Topics,2012,210(1):5-16.
[7] SIVANANDAN V,KUMAR V,MEHER S.Designing a parallel algorithm for heat conduction using MPI,OpenMP and CUDA[J].Parallel Computing Technologiess,2015:1-7.
[8] LONCAR V,YOUNG-S L E,SKRBIC S,et al.OpenMP,OpenMP/MPI,and CUDA/MPI C programs for solving the time-dependent dipolar GrossPitaevskii equation[J].Computer Physics Communications,2016,209:190-196.
[9] JACOBSEN D A,THIBAULTY J C,SENOCAK I.An MPI-CUDA implementation for massively parallel incompressible flow computations on multi-GPU clusters[C]∥Proceedings of the 48th AIAA Aerospace Sciences Meeting Including the New Horizons Forum and Aerospace Exposition.Reston,VA,USA:American Institute of Aeronautics and Astronautics Inc.,2010:0522.
[10]刘冰,陆忠华,李新亮,等.基于GPU的多重网格Navier-Stokes解算器并行优化方法研究[J].科研信息化技术与应用,2013,4(3):56-67.LIU Bing,LU Zhonghua,LI Xinliang,et al.Implementation and optimization of a Navier-Stokes solver with multigrid methodology based on GPU[J].E-Science Technology and Application,2013,4(3):56-67.
[11]胡银丰,孔强.基于MPI+OpenMP混合编程模型的并行声纳信号处理技术研究[J].舰船电子工程,2017,37(12):130-133.HU Yinfeng,KONG Qiang.Research on parallel sonar signal processing technology based on MPI+OpenMP mixed programming model[J].Ship Electronic Engineering,2017,37(12):130-133.
[12]李春.混合并行计算在海洋水质预报系统中的应用研究[D].石家庄:河北工程大学,2014:36-50.
[13]HAH C,KRAIN H.Secondary flows and vortex motion in a high-efficiency backswept impeller at design and off-design conditions[J].Journal of Turbomachinery,112(1):7-13.
[14]曹维.大规模CFD高效CPU/GPU异构并行计算关键技术研究[D].长沙:国防科学技术大学,2014:39-50.
[15]张楚华,琚亚平.流体机械内流理论与计算[M].北京:机械工业出版社,2016:129-145.
[16]刘超群.多重网格法及其在计算流体力学中的应用[M].北京:清华大学出版社,1995:1-49.
[17]TAI C H,ZHAO Y,LIEW K M.Parallel-multigrid computation of unsteady incompressible viscous flows using a matrix-free implicit method and high-resolution characteristics-based scheme[J].Computer Methods in Applied Mechanics and Engineering,2005,194(36):3949-3983.