Development of Computational Tools for Large Scale Wind Simulations

Large scale wind flow simulation over complex topography requires a tremendous amount of computational resources, both in terms of CPU hours and storage requirement. Use of complex turbulence models such as large eddy simulation (LES) significantly add to the computational demand. In practice, a relatively less accurate turbulence model such as Reynolds Average Navier Stokes (RANS) equations are solved due to computational resource limitations. If computing power is available, parallel simulation can be carried out to cut down simulation time without degrading the quality of flow modeling. Coarse grained parallelism in a distributed memory cluster is traditionally achieved by domain partitioning strategies. The whole domain is partitioned into smaller sub-domains which are then assigned to a processor in a cluster of computers. In this work a non-overlapping type of domain decomposition method is implemented where information such as pressure and velocity is exchanged at the boundaries of the domain through ghost cells. The solution of all sub domains goes on simultaneously. The message passing interface (MPI) is used to exchange information between sub-domains. The second focus of this work is to examine suitability of general purpose graphic processing units (GPGPU) for solving fluid flow problems. Nowadays GPUs are becoming a much cheaper alternative for speeding up CFD simulations significantly. Speed ups of up to 100 times compared to a single CPU implementation have been reported in literature for simple flow cases. In this study, NVIDIA's CUDA programming is used to exploit the fine grained parallelism offered by GPUs, and use it for practical fluid flow problems. Both methods of parallelization are combined at the solver level so that a mixed CPU-GPU computation is possible. Finally the speed up numbers obtained for different size problems are compared.