面向激光聚变模拟的大规模异构众核系统可扩展并行算法与优化方法

The fusion energy is the current important new energy development direction. The high performance computation simulation is a good research approach for fusion. However, due to the complexity of the fusion process and the development of super computer system architecture, there has substantial hardship in the heterogeneous exascale computing of the laser fusion simulation. So, based on the heterogeneous multi/many-core parallel systems of hundreds of thousands of cores of CPU+GPU+MIC, this project will do further research on the establishment of parallel grid, efficient parallel computation model of combination of structured grid and non-structured grid, and study the efficient and scalable data partition theory and method of large-scale heterogeneous many-core system for fusion simulation; In view of the high performance computation in the sparse linear algebra, three dimensional FFT solution, the parallel algebra multiple, the project exploits irregular data communication methods and dynamic load balancing algorithms for fusion numerical simulation; This project also studies five major application programs, the hundreds of thousands of CPU and GPU/MIC heterogeneous cooperative parallel computing theory, and system performance optimization and operation stabilization technique. The research results is not only the theory foundation of large-scale fusion numerical simulation software design and implementation, but also will enhance the impetus of high performance computing for the independent innovation of national major application, and provide reference for other engineering and scientific computing problems.

聚变能源是当前新能源发展的重要方向，也是我国当前和未来高性能计算的重要领域。然而，由于聚变过程的复杂性和超级计算机系统体系结构的发展，目前面向激光聚变模拟的异构E级计算仍存在实质困难。鉴此，基于CPU＋GPU＋MIC数十万核异构多/众核并行系统，本项目将对聚变数值模拟并行网格的建立、结构化网格与非结构化网格相结合的高效数值并行计算模型等展开深入研究，提出适合大规模异构多/众核计算平台的高效可扩展数据剖分理论与方法；针对聚变数值模拟中的稀疏线性代数、三维FFT解法器、并行代数多重网格等典型共性计算模块探索其非规则数据通信方法、动态负载平衡算法；研究聚变模拟五个主要应用程序，适合数十万核CPU与众核GPU/MIC异构协同并行计算理论及其系统性能优化与运行稳定化技术。项目成果将不仅为聚变数值模拟软件设计奠定基础，增强高性能计算对国家重大应用自主创新的推动力，还可资其他工程与科学计算问题借鉴。

项目资助期间，立足于高性能异构并行计算的研究前沿，充分利用并行计算、并行调度、异构计算算等理论和技术成果，侧重于新理论、新技术和新方法的研究，采用理论和实验相结合的方式，针对来源于面向激光聚变模拟的大规模异构众核系统可扩展并行算法与优化方法过程，完成了面向聚变模拟典型程序的数据剖分方法、共性快速数值并行算法库的研制、激光聚变模拟并行计算模型与优化理论、面向激光聚变模拟的异构混合系统节点内协同并行计算技术、(5)异构多/众核并行计算系统性能优化与运行稳定化技术等基础性研究工作；发表学术论文42篇，其中 IEEE/ACM Transactions 论文18篇，被 SCI 收录40篇，申请国家发明专利6项，专著2本。