Graphics processing unit (GPU) and clusters of GPUs have emerged as cost-effective general-purpose scientific computing platforms that can substantially accelerate simulation science applications. Over the last four years, we have developed a parallel incompressible flow solver with an amalgamated geometric multigrid method. The flow solver uses MPI for coarse-grain and CUDA for fine-grain parallelism, while overlapping communication with computations. We have tested our flow solver for scalability on the NCSA Lincoln and TACC Longhorn GPU clusters, and demonstrated the importance of overlapping communication with computations and network bandwidth on emerging GPU clusters. In this study we extend our flow solver to perform large-eddy simulations of turbulent flows in a channel with friction Reynolds number of 395. Our results agree well with published direct numerical simulation results and demonstrate that GPU clusters can be an effective tool in fundamental turbulence research.
Available at: http://works.bepress.com/inanc_senocak/34/