多图形处理器上Lattice-Boltzmann方法的加速  

Accelerating Lattice-Boltzmann Method with Multi-GPUs

作　　者：吴亮[1] 钟诚文[1,2] 郑彦奎[1] 刘沙[2] 卓丛山[2] 陈效鹏[3] 

机构地区：[1]西北工业大学高性能计算研究与发展中心,西安710072 [2]西北工业大学翼型、叶栅空气动力学国防科技重点实验室,西安710072 [3]西北工业大学力学与土木建筑学院,西安710072

出　　处：《计算机辅助设计与图形学学报》2010年第11期1932-1939,共8页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(10902087);航天科技创新基金(2009200066)

摘　　要：为了提高计算流体领域中复杂流动现象模拟计算的高效性和准确性,充分利用图形硬件的并行性,提出一种在单机多图形处理器下基于CUDA架构的Lattice Boltzmann方法(LBM)的模拟算法.采用区域划分策略将域上的LBM网格平均分配到不同的GPU设备上,在分区边界处搭接一层网格以方便计算该处网格的迁移过程,减少GPU间的通信量,并合理地利用CUDA存储层次架构中的全局内存和纹理内存为计算网格分配设备空间;采用多线程技术,用每个线程控制不同的GPU设备,同时引入线程同步机制信号量实现线程间的数据通信同步控制,按照LBM方程组的求解过程实现模拟计算.实验结果表明,双GPU将计算加速到单GPU的1.77倍左右,同时将流场计算网格规模从单GPU下的4160×4160扩大到双GPU下的6144×6144.To improve the efficiency and accuracy of simulating complex flow phenomena,an accelerated algorithm of Lattice-Boltzmann method based on CUDA on multi-GPUs desktop platform is proposed.The computational space is evenly divided into cells,which compose multiple sub-domains in accordance with the number of available GPUs.Each GPU stores only the data of its respective sub-domain in global memory and texture memory.To facilitate the particle interaction at the partitioned boundary between adjacent sub-domains,an additional layer of cells is attached to the boundary of each sub-domain to store the data computed by the neighboring GPU hence reducing the communication overhead between GPUs.Then each CPU thread controls an available GPU using multithreading technology.And semaphore is used to synchronize between threads to exchange data in the iteration process for solving the LBM equations.Experimental results show that,the dual-GPU performs 1.77 times faster than that of the single GPU approximately and the resolution of the computational space can be increased to 6144×6144 instead of that of single GPU which is 4160×4160.

关 键 词：CUDA 多线程技术 LATTICE BOLTZMANN方法 多图形处理器 信号量 并行计算 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]