三维特征线方法能够精确求解任意几何中子输运问题，是下一代反应堆物理分析的重要研究方向之一。但是，特征线方法具有收敛缓慢、计算耗时及内存占用大的不足，用于三维几何求解时计算量变得更加庞大，这些从根本上限制了三维特征线方法的发展与应用。本文针对三维特征线方法的不足，以三维特征线方法的并行和加速方法作为研究对象，主要开展了以下研究：首先，研究了基于中央处理器（CPU）并行的特征线方法。采用角度并行方式，不引入额外计算量同时保持与串行程序的一致性；提出角度分组的方法，减少了通信，提高了并行效率。其次，将新兴的高性能计算图形处理器（GPU）用于特征线方法加速计算的研究。相比于CPU，GPU拥有更多的计算核心，计算成本更低，加速效果更好。针对其硬件构造的特殊性，采用射线并行的方式，并结合CPU并行，实现了多GPU系统计算。再次，在借助计算机硬件加速特征线方法以外，提出了双重广义粗网有限差分加速方法。该方法能够同时对任意形状的网格和能群进行加速。最后，提出插值方法处理的三维模块式建模，继承了模块式建模占用存储少建模速度快的优点，解决了传统模块式建模无法处理非基本模块的问题。基于上述研究内容，研制了具有模块式建模功能、采用双重广义粗网有限差分加速、CPU和GPU并行计算的三维特征线方法程序SPIDER。数值计算结果表明：程序具有较高计算精度，通过并行和加速后程序计算时间大大缩短；插值方法处理的模块式建模方法在减少网格划分和射线求交的时间、加快建模速度的基础上实现了对存在水隙等非模块的组件的栅元模块式建模；双重广义粗网有限差分加速方法能够同时从能群和网格上进行加速；角度分组的CPU并行方式不会增加迭代计算次数，同时能够达到较高的并行效率；GPU并行计算能够以较低的计算成本实现更高的加速比。针对三维特征线方法的特点，本文研究了适用于三维特征线方法的并行算法和加速方法，加快了收敛速度，减少了计算时间，增强了三维特征线方法及程序的工程实用性。
Since the method of characteristics (MOC) can solve neutron transport equation for arbitrary geometry accurately, it becomes one of the most promising methods for the next generation in reactor analysis. However, the MOC has some drawbacks: The convergence speed is slow and very time consuming. There are also restrictions on the development of the three-dimensional MOC such as the huge amount of computation in practical three-dimensional geometric computing. In order to resolve these disadvantages, this thesis focuses on the parallelization and acceleration of three-dimensional MOC, and the research contents are as followed.First, the MOC parallelized by central processing unit (CPU) has been studied. In order to keep consistent with serial code and not to increase extra computational burden, an angular parallel scheme is adopted. Moreover, the angular grouping method is proposed to reduce the communication and thus improve the efficiency.Then, the graphics processing unit (GPU) computing, which is a newly rising high performance computing, is applied to the MOC. As the number of the GPU computing core is much more, a ray parallel scheme is chosen in GPU computing. Combined with the CPU parallel method, the MOC is performed under multiple GPUs.Besides the parallel computing, another acceleration method called two-level generalized coarse mesh finite difference (GCMFD) is proposed. The two-level GCMFD, which can accelerate the energy group as well as the arbitrary coarse mesh, is used to accelerate the MOC in this paper.Finally, because the reactor cores are usually composed of reduplicate constructions and the water gaps exist between assemblies, an interpolated three-dimensional modular ray tracing is proposed. The modular ray tracing can reduce the memory requirement and accelerate geometric modeling while the interpolated technique can simulate the water gaps between the assemblies.According to the methods mentioned above, the SPIDER code has been developed. The numerical results show that a good accelerated performance on both modeling and computing of the MOC has been achieved under the high accuracy.This thesis focuses on the study of the parallel computing and accelerated method. As a result, it accelerates the three-dimensional MOC, decreases the number of iterations, reduces the computing time and enhances its engineering practicability.
