Updates on extreme-scale turbulence simulations and data-processing on an exascale platform

ORAL

Abstract

Recent advancements in computer hardware and GPU programming techniques have allowed us to simulate homogeneous turbulence using 35-trillion grid points (Yeung etal. Comput. Phys. Commun. Vol 306, 109364 (2025)), using half or more of the exascale computer `Frontier'. Here we provide updates on several aspects and functionalities beyond the simulation of the velocity field. In particular, similarly high scalability has been obtained for studies of point particle dynamics (with and without effects of inertia), and of the mixing of passive scalars including moderately and weakly-diffusive regimes. Fluid particles are tracked by using cubic-spline interpolation for particle velocities. The interpolation is quite accurate even for highly intermittent quantities such as pressure gradients sampled along the particle trajectories. For turbulent mixing a dual-resolution and dual-communicator scheme is used at high Schmidt number where the scalar field demands very high resolution. Computation-heavy calculations such as for the structure functions are greatly facilitated by OpenMP offloading with parameters that reduce GPU thread contention. Sub-cubes in physical space are formed for visualization and public data use via the JHUTB.

*CAAR and INCITE projects at Oak Ridge Leadership Computing Facility, and subcontract NSF via Johns Hopkins Univ. (Grant 2103874).

Presenters

  • Pui-Kuen Yeung

    • Georgia Institute of Technology

Authors

  • Pui-Kuen Yeung

    • Georgia Institute of Technology
  • Rohini Uma-Vaideswaran

    • Georgia Tech
  • Daniel L Dotson

    • Georgia Institute of Technology