Lagrangian Discrete Phase Model (Particle Tracking) on GPUs

Significant progress has been made in the adaptation of GPUs in CFD simulations in recent years, both in traditional Eulerian phase approaches like Finite Volume Methods and Lagrangian approaches like Smoothed Particle Hydrodynamics, reducing simulation times from days to hours. In this talk, we will investigate Euler-Lagrange type problems, specifically modeling a Lagrangian discrete phase with a continuous Eulerian phase, which are important in many fields, including gas turbine combustion. These simulations could take days to complete on conventional CPU clusters. With the finite volume solver in Ansys Fluent already optimized for GPUs, the discrete Lagrangian phase solver must also be optimized to fully reap the benefits of GPUs. We will go over the strategy we used to accomplish this. We will discuss the algorithmic modifications that must be done to properly leverage the GPU's architecture in comparison to a CPU implementation. Additionally, we'll address scaling this up to multiple GPUs and the particular difficulties presented by an Eulerian background grid. We will present some performance data and highlight the potential for this work to save significant time and money for critical simulations.