(366c) Using GPUs to Run Large Eddy Simulations of Blending in Oil Storage Tanks

In large oil storage tanks, blending is typically accomplished using side-entry mixers positioned inside manholes near the bottom of the tank. These impellers have diameters that are very small when compared to the overall tank geometry and have tip speeds that are high compared to the tank-average flow velocity. This net effect of this disparity is a spatially varying flow field that is turbulent near the impeller but may be laminar in the far-field. From an analysis perspective, these complexities make a priori blend-time predictions difficult and necessitate scale model testing of individual confirmations. From a CFD perspective, transient modeling approaches must be applied to properly predict the spatially varying blending characteristics of such systems.

In this work, we show how GPUs (as opposed to CPUs) can make large eddy simulations of such systems practical and timely. We begin by introducing the concepts governing GPU-performance, as applied lattice-Boltzmann and LES simulation. We then show how, given identical CFD algorithms, a single scientific GPU can execute simulations 100x faster than a single CPU. The extension to multiple GPUs is also discussed and demonstrated. We then use this approach to simulate the start-up and steady state flow behavior of an oil storage tank agitated by a small side-entry mixer. We model both single and stratified two-fluid miscible systems, and compare predicted blend time predictions to experimental measurements.