基于非结构网格流场超大规模并行计算

Large scale parallel computing for fluid dynamics on unstructured grid

大规模并行的计算流体力学已成为现代航空工业研发的核心手段之一。基于非结构混合网格和有限体积法,发展了适用于工业级复杂外形气动计算的并行流动数值模拟方法。文中首先介绍了紧致数值离散格式、基于Metis的分布式多核系统网格分区技术、并行边界虚拟单元技术和MPI并行实现等相关算法。采用网格量相对较小的旋成体构型绕流模型对比分析多核并行计算结果与单核计算结果以验证并行计算的正确性,比较了不同并行规模下并行效率和残差收敛情况。然后通过对上亿网格单元的运输机复杂构型绕流模拟,开展并行效率的测试,结果表明,本文方法并行加速性能高,直到多达18816核并行效率都保持在80%以上。

Massive parallel computational fluid dynamics is a vital tool for modern aviation industry. The parallel aerodynamics numerical simulation techniques based on unstructured mixed grid and cell-centered finite volume method (FVM) were presented in this paper. These techniques were developed for complicated configurations in industrial applications. The algorithms were discussed in detail for parallel implementation with message passing interface (MPI), including compact numerical discrete schemes, grid partition methods for multi-processors based on Metis software package, ghost cells method for both physical and parallel boundary faces, and data transfer scheme. The parallel computing was verified by using a revolving model of relatively small amount of grid cells, and the convergence characteristics of aerodynamic force and flow variables residual were discussed. A large transport airplane configuration with about a hundred millions of grid elements was used for parallel test. An excellent performance of speed up was obtained, and the efficiency was above 80% with up to 18816 processors.