Accelerating Lattice-Boltzmann method with multi-GPUs

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.