压气机内流多级全环数值模拟的大规模高效混合并行计算方法

Full-annulus multistage simulation in study of unsteady flow mechanism of compressor, such as rotating stall or surge, has huge demand for computing resources and performance. This has already limited the development of compressor. The modern high-performance computing has brought a turnaround to solve this bottleneck of the large-scale computing. This project will focus on the hybrid computing model of the general-purpose processor and the multi-core acceleration unit in the modern high-performance heterogeneous many-core system, to establish a high efficient MPI+X parallel computing model with some general applicability. Then, considering the logic architecture characteristics of the high performance heterogeneous many-core system and the hierarchical geometrical structure of the full-annulus multistage compressor simulation, a high scalable mapping algorithm model will be established for large-scale (million computing cores) parallel simulation of the compressor. And finally form a high-efficiency, high-scalable hybrid parallel computing method for full-annulus multistage unsteady simulation of compressor, then the corresponding large-scale parallel computing programs will be developed on the China’s supercomputer Tianhe-1 and Tianhe-2. The novelty of this project is of academic values to improve the high-performance computation theory and can provide high efficiency tools for development of high load, high efficiency and wide stability margin in compressor.

用于压气机气动失稳等复杂流动机理研究中的多级全环非定常数值模拟对计算资源及计算性能的巨大需求，已经在某种程度上限制了压气机整体性能的提升，而现代高效能并行计算系统的发展为解决这一计算瓶颈问题带来了转机。本项目将结合现代高性能异构众核系统中通用处理器与众核加速单元的混合计算模式，构建MPI+X高效混合并行计算模型；在此基础上，围绕高性能异构众核计算系统的逻辑架构特征以及压气机多级全环流动域的分层几何结构，建立压气机内流大规模（万核级）并行仿真计算的高可扩展映射算法模型。最终形成压气机多级全环非定常流动数值模拟的高效、高可扩展混合并行计算方法，并结合我国超级计算机天河一号（CPU+GPU）、天河二号（CPU+MIC）系统发展相应的大规模并行计算程序。研究成果对促进高性能计算科学的研究水平具有学术价值，同时为高负荷、高效率、宽稳定裕度压气机的研制提供高效计算流体力学工具。

压气机多级全环非定常数值模拟对计算资源以及计算性能的需求巨大，本项目结合国产高性能计算系统以及新型计算硬件架构体系开展了压气机内流大规模混合并行数值计算方法研究。现代高性能计算系统多具有异构/同构众核计算核心，其非常适合于对数值求解过程中的计算密集型过程进行加速提升。结合该硬件结构特征，通过分析数值模拟计算中的计算热点部分，将焦点集中在数值计算中线性代数方程组求解部分，提出了适用于压气机内流仿真的MPI+X混合并行计算模型，分别实现了同构众核系统以及异构加速硬件上的MPI+OPENMP、MPI+CUDA、MPI+MIC的混合异构并行计算。在此基础上，结合国产高性能计算系统的架构特征以及压气机级、叶排、叶片通道、并行子区域的分层次结构，给出了适用于大规模并行计算的多层次可扩展并行映射算法。该算法通过对映射到不同叶排上的计算进程进行分组性能优化、局部计算实现混合并行计算加速、全局信息交换采用分层次高通量通信、动静叶排实现通信域局部重构等方式，可在万核乃至数十万核心上实现对压气机亿级网格规模的高效数值仿真。在我国国产E级（浮点计算性能达到百亿亿次以上）计算原型验证机上通过对某十级压气机10亿网格的初步仿真模拟，验证了本项目提出的大规模混合并行计算模型的有效性。