Simulations of turbulent duct flow with lattice Boltzmann method on GPU cluster

Lattice Boltzmann models are used to simulate three-dimensional pressure driven laminar and turbulent duct flows using graphics processing units. The performances of the D3Q15, D3Q19 and D3Q27 models are first validated by comparisons with laminar analytical and benchmark solutions; the D3Q27 model is shown to produce the most accurate results. Turbulent duct flows at Reτ=360 and 600 are further simulated by the D3Q27 model, and the predicted results compare favorably with the existing Navier-Stokes equation based benchmark solutions. For multi-GPU computations, the performance of 2-D domain decomposition is superior to that of 1-D domain decomposition where the performance saturates at a lower GPU count. For the present parallel performance evaluation, 1730 and 2464 MLUPS are achieved with 12 GTX Titan and 48 Tesla M2070 GPUs from nVIDIA, respectively.