基于CUDA的格子Boltzmann数值模拟加速实现

Implementation of the Acceleration Simulation with Lattice Boltzmann Method Based on CUDA

针对近年来利用CUDA技术在个人计算机显卡的GPU上实现LBM并行加速计算的研究越来越多,但对在GPU中使用不同GPU存储器进行计算的具体实现算法以及其对计算性能的影响分析研究甚少,文章实现了在GPU中使用不同存储器进行LBM并行计算,给出了具体的实现算法,并以平面Poiseuille流为算例,在普通个人计算机上,分别使用NVIDIA GeForce GTS450GPU和Intel Core i5-760 4核CPU进行计算。结果表明,两者计算结果吻合得很好,最高获得了约107倍的加速比,验证了在GPU上进行LBM并行计算的可行性以及加速性能,为在低成本的个人计算机上高效率地解决计算流体力学中的复杂计算问题提供了一种非常有效的途径。

In recent years,there are more and more research on implementing LBM accelerating computation on the GPU of a PC＇s graphics card with CUDA technology is getting more attention. But the detailed algorithm and the analysis of computational performance which compute parallelly on a GPU with different memory are rarely studied. In this paper,by using different memory on a GPU ,the parallel computing of LBM was implemented and the detailed algorithms are provided. Taking the plane Poiseuille flow as a test example,the parallel computation of the LBM is implemented on a NVIDIA GeForce GTS 450 GPU and an Intel Core i5-760 quad-core CPU on a PC respectively. Both computing results have good agreement and the highest speedup of GPU is about 107 times faster than that of CPU. The result indicates parallel computation of the LBM on GPU is completely feasible and the accelerating performance is very significant,which provides a very effective way to solve the complex problem of the modern computational flukl dynamics in a low cost computer.