面向数十万核以上异构众核平台的CFD领域性能建模与应用

Supercomputers with heterogeneous architecture containing hundreds of thousands of processor cores are becoming a mainstream hardware platform for high-performance computing, which provides a good opportunity for unprecedented large-scale CFD numerical simulation. However, efficient parallel computing for large-scale CFD application on the platform with 100,000 or more cores is also facing some challenges, such as achieving a good load balancing on the heterogeneous architecture, exploring the multi-level parallelism in the CFD applications, optimizing the memory access and improving the arithmetic intensity for the stencil computing in the typical CFD simulations, and so on. In this project, we will focus our research on the high-performance computing platform with hundreds of thousands of processor cores and the typical large-scale parallel CFD applications and try to answer partially the question of "How to get a better performance to run the large-scale CFD applications on the supercomputer with more than 100,000 cores". The topics of this research includes: (1) To build a framework to model the performance of supercomputers with hundreds of thousands of cores; (2) To design a new parallel algorithm for PDE solving in the CFD application and propose an approach to exploring the multi-level parallelism in the CFD application; (3) To improve the data locality and memory access characteristics of stencil computing to increase compute density; (4) To establish a strategy of task mapping, task scheduling and load balancing during the large-scale CFD simulation on the HPC platform with 100,000+ cores. The purpose of this project is achieving some breakthroughs in developing some more efficient, better scalable parallel algorithms of large-scale CFD application.

数十万核的异构众核计算机正在成为高性能计算领域的主流硬件平台，这为大规模CFD数值模拟提供了前所未有的良好机遇，然而大规模CFD并行模拟向数十万核并行平台上的移植还面临着平台与应用相融合的性能建模、异构体系结构下的负载平衡、CFD应用多层次并行性开发、模板计算访存性能优化等一系列挑战性问题。本项目拟以数十万处理器核心的高性能计算机为目标平台，以多区三维可压流场的CFD并行模拟为典型案例，以面向目标平台的CFD领域应用性能建模为主线，重点研究：构建典型高性能计算平台的性能分析模型，开发CFD应用问题中的多层次并行性，优化CFD典型求解计算过程的数据局部性和访存特性、提高计算密度，建立高效模拟的任务映射与负载均衡策略；以期突破若干关键算法，推动CFD应用向更大规模、更高效率、更高精度方面发展，为超大规模CFD工程应用提供新的技术途径。

数十万核的异构众核计算机正在成为高性能计算领域的主流硬件平台，这为大规模CFD数值模拟提供了前所未有的良好机遇，然而大规模CFD并行模拟向数十万核并行平台上的移植还面临着平台与应用相融合的性能建模、异构体系结构下的负载平衡、CFD应用多层次并行性开发、模板计算访存性能优化等一系列挑战性问题。本项目以数十万处理器核心的高性能计算机为目标平台，以多区三维可压流场的CFD并行模拟为典型案例，以面向目标平台的CFD领域应用性能建模为主线，重点围绕典型CFD应用的数值求解方法的模型与性能分析、同构型通用CPU众核计算平台上CFD应用数值模拟的并行与优化、以及国产异构型超级计算平台上CFD应用的协同并行优化等三方面内容展开深入研究，对数十万核的异构众核计算机上的超大规模CFD应用并行与优化问题进行了关键技术攻关，并在典型CFD求解器、多层次负载平衡、同构平台上的高效并行与优化、异构型平台上的协同并行等方面取得突破，成功实现了同构平台近20万核、异构平台约137万核场景下、真实高阶精度CFD应用的数值模拟，并行效率高，可扩展性良好。通过本项目的研究，课题组较全面地掌握了以众核、异构、宽指令为特征的主流高性能处理器平台上、开展以CFD为代表的科学工程计算数值模拟时的并行与优化关键技术，必将有力推动CFD应用向更大规模、更高效率、更高精度方面发展，为超大规模CFD工程应用提供新的技术途径；同时对航空、航天、航海等领域的一大批流体力学应用问题的高精度求解有了系统的认识与理解，这些将对其它领域的研究提供良好的借鉴。