熔盐堆堆芯流体力学计算的GPU并行方法研究

Research on GPU parallelization of fluid dynamics process of molten salt reactor core

使用计算流体力学(Computational Fluid Dynamics,CFD)数值方法对熔盐堆堆芯的流动和热传导等相关物理问题进行模拟求解,需要大量的计算时间。利用图形处理器(Graphics Processing Unit,GPU)加速技术对开源CFD软件Code_Saturne进行二次开发,研究求解熔盐堆堆芯流场的GPU并行算法。采用OpenACC语言在GPU上实现了向量运算、矩阵向量相乘等基本线性代数运算,从而实现预处理共轭梯度法(Preconditioned Conjugate Gradients,PCG)的GPU并行算法,并使用该算法求解压力状态方程。模拟了方腔驱动流模型及带下降段的熔盐堆堆芯模型的流场分布。结果表明,GPU加速后的软件与原版软件的结果一致,但计算时间更少,证明了GPU算法的正确性及有效的加速性。

Background： The simulation of fluid dynamics process for molten salt reactor proposes a large compute complexity, which requires high performance computer systems to enhance speed and efficiency. Purpose： This study aims to achieve graphics processing unit （GPU） parallelization of fluid dynamics process of molten salt reactor core. Methods： OpenACC directives were used as the main programming model to speed up the vector and matrix linear operation. And the preconditioned conjugate gradients for solving linear equations were implemented on the GPU. Finally, the parallel implementation and general optimization strategies to the OpenACC version of Code_Saturne were tested and validated on a simplified molten salt reactor. Results： From the result of the implementation of the GPU-parallel code, it is manifested that the empirical tuning of OpenACC accelerated code sections are valid for obtaining correct results, and enhancing performance and portability. Conclusion： With OpenACC, we find that the instance of fluid dynamics process for molten salt reactor is given out using the GPU version of Code Samrne and the performance of the GPU version of Code_Saturne can be enhanced compared with that of the CPU version.