ABL Simulation and Analysis via the FastEddy GPU-Accelerated Model

A wide range of applications including, but not limited to, wildland fire behavior, atmospheric transport and dispersion, and wind energy could benefit from an accelerated microscale Atmospheric Boundary Layer (ABL) simulation capability. Large-eddy simulation has long been considered computationally intractable for many boundary layer turbulence modeling applications. Limited access and scalability constraints of models implemented for large and expensive, power-inefficient, Central Processing Unit (CPU)-cluster environments remains a daunting path for real-world applications aiming to incorporate high-fidelity numerical modeling of ABL processes. In 2017 the National Center for Atmospheric Research has undertaken the development of a Graphics Processing Unit (GPU)-accelerated atmospheric modeling framework, FastEddy, to address procurement cost, power efficiency, and time-to- solution constraints on highly resolved ABL modeling. Here we present the numerical framework and salient design features of FastEddy. Selected analysis of canonical-cases, complex topography, urban/rural scenarios, multiscale ABL, and multi-physics ABL applications are presented as early milestone demonstrations of maturing model skill. Additionally, quantitative performance metrics are provided for select scenarios demonstrating O(100) speedup over serial CPU execution. Advantageous performance features are shown to include a resident-GPU dynamical core, heterogeneous GPU memory utilization, GPU stream concurrency, and GPU kernel thread-occupancy. Lessons learned are shared with respect to requisite consideration of underlying algorithmic characteristics when adopting these techniques, in order to achieve a long-term goal of O(10m) ABL simulations in real time.