Abstract
Computational wind engineering is a broad field that finds applications in air pollution, emergency response, structural engineering, renewable energy and weather forecasting. Advances in graphics processing units (GPU) and manycore computing present new and exciting opportunities to accelerate computational turn-around time of wind predictions where rapid simulations can make a tremendous impact on the current practice. In environmental and homeland security applications, forward and inverse modeling of atmospheric contaminant dispersion events relies heavily on rapid wind prediction for emergency response. In wind energy applications, power grid operators can substantially benefit from a reliable wind forecasting capability to balance load and production in the face of intermittent wind, and to increase transmission line capacity through the dynamic line rating concept. Therefore, a short-term wind forecasting capability that can accurately predict winds within the lowest 200m of the atmosphere with fine spatial and temporal resolutions is needed. To this end, emerging GPU clusters provide several opportunities to accelerate time-to-results that is critical in wind forecasting and atmospheric transport and dispersion of contaminants.

In this talk, I will describe the development of a multi-GPU parallel computational fluid dynamics (CFD) software (GIN3D). GIN3D executes on multi-GPU platforms with a mixed MPI-CUDA implementation that overlaps computations with communications. I will discuss several programming strategies, including dual- and tri-level parallelism, to demonstrate the scalability and performance of a baseline incompressible solver using as many as 256 GPUs. Performance results from a full depth amalgamated multigrid solver will be used to demonstrate the impact of algorithmic complexity on parallel performance on GPU clusters. Finally, I will discuss the prospects of in-situ visualization of CFD simulations on tiled-display GPU clusters with preliminary results.