Challenges and solutions in the porting of a finite difference high-fidelity CFD code to HLRS's Hunter system¶
Tobias Gibis (HLRS)
Abstract¶
The Boundary Layer Group at the Institute of Aerodynamics and Gas Dynamics employs the in-house NS3D code to simulate large-scale transitional and turbulent boundary layer flows. Porting the code to the new Hunter system at HLRS required a dedicated effort. However, beyond the technical aspects of code migration, the adoption of modern HPC architectures—particularly those relying on accelerators—raises a fundamental question: Is the current numerical scheme still well suited to such systems? A central challenge in this context lies in NS3D’s reliance on compact finite difference schemes, which have proven effective in numerous large-scale simulations in recent years. Yet these schemes may not map efficiently onto accelerator-based platforms. This talk will explore possible future directions for spatial discretisation in NS3D. Options include transitioning to explicit finite difference schemes, which are more amenable to parallelisation on accelerators, or alternatively, optimising the existing compact schemes to better exploit the capabilities of modern heterogeneous HPC architectures.