高分辨率全堆粒子输运模拟的SN并行算法研究

The current trend of particle transport numerical simulation is to solve models with much higher resolution and precision, in which high performance computing has become one of the essential methods. This project aims at addressing the bottleneck in numerical stability, computational efficiency and parallel scalability faced by the high resolution full core particle transport using SN (discrete ordinates) method. Based on our previous research experience on developing the sweeping algorithm and iteration acceleration algorithm on tens of thousands of CPU cores, this project in further studies the new iteration and acceleration algorithm, the efficient design and implementation of parallel algorithms, the software development as well as the application validation. To achieve good numerical stability, efficiency and scalability on hundreds of thousands of CPU cores, this project will put more emphasis on two key scientific problems: the physics-based preconditioner, the “energy-space-angle” multilevel parallel algorithm and the optimization on modern computer architecture. The achievements of this project can provide a supporting theory and technology for the full reactor core particle transport with high resolution, which has both theoretical and practical significance in the enhancement of reactor safety and economics, the design of next-generation reactors as well as the computational physics research.

粒子输运数值模拟研究当前正朝着计算模型越来越精确、分辨率越来越高等趋势发展，高性能计算已成为其中必不可少的支撑手段。本项目面向基于SN（离散纵标）方法的高分辨率全堆粒子输运模拟，针对已有迭代加速方法和并行算法研究在数值稳定性、计算效率和并行可扩展性方面的瓶颈，以数万核并行扫描算法与迭代加速算法的前期研究成果为基础，从新型迭代求解与加速方法、并行算法与高效实现以及软件模块研制和应用验证方面开展研究。研究将重点突破基于物理的预条件子方法、“能群-空间-方向”多级并行算法以及针对多级嵌套体系结构性能优化两项关键科学问题，实现具有良好数值稳定性和计算效率的、适应于数十万处理器核的可扩展并行计算。研究成果可为高分辨率反应堆全堆输运模拟等问题提供必要支撑，对提升核电经济性与安全性、新堆型研发以及武器物理研究，具有重要意义和工程应用价值。

本项目面向基于SN(离散纵标)方法的高分辨率全堆粒子输运模拟，针对已有离散格式、迭代加速方法和并行算法研究在数值稳定性、计算效率和并行可扩展性方面的瓶颈，从新型迭代及加速方法与离散格式、并行算法与高效实现以及软件模块研制和应用验证方面开展研究。针对各向异性较强问题提出了PCR+Anderson耦合的分群加速算法，效率提升2-10倍；针对传统非线性TWD权重差分空间离散格式引入的收敛性问题，提出了基于解析基函数的ABF离散格式耦合Krylov子空间迭代算法，相较传统方法总迭代次数减少1个量级，计算效率提高4.5倍。提出了既匹配于体系结构又适用于SN计算的高效数据结构、基于粗化图的降低扫描通信和调度开销方法，基于主从模式的任务调度和消息缓存策略以及基于循环拆散和计算分块等单核性能优化，最终实现了在FT 2000+天河2超算上25.6万核的并行可扩展。研究成果可为高分辨率反应堆全堆输运模拟等问题提供必要支撑，对提升民用核电经济性与安全性、军用核动力和武器物理研究，具有重要意义和工程应用价值。