Progress in Massively Parallel and Performance Portable Overset Implementations in HPCMP CREATE A/V Helios
Jayanarayanan Sitaraman, Shirzad Hosseinverdi, Dylan P. Jude, Beatrice Roget, Jennifer N. Abras, Vinod K. Lakshminarayan, George Zagaris · 2024
This paper introduces a performance-portable approach for the dual-mesh overset computational fluid dynamics paradigm used in the high fidelity DoD HPCMP CREATE\textsuperscript{TM}-AV codes, designed to enable efficient execution on both CPUs and GPUs. Three components --- a near-body unstructured solver, an off-body Cartesian solver, and an overset domain connectivity framework that connects the solvers --- are implemented in a performance-portable manner. The paper describes the implementation approach and demonstrates performance for fixed and rotary-wing applications. Scalable performance is demonstrated up to 100 GPUs. The relative performance of each component is documented along with some remaining challenges in achieving performance portability for this overset multi-solver paradigm.