Technical Papers Archive

A Hardware Abstraction for Exascale Lattice Boltzmann Simulations: Rotor-Resolved Wind Farms on Aurora and LUMI


Authors: Adrian Kummerländer and Fedor Bukreev (Karlsruhe Institute of Technology (KIT)), Leonardo Dorneles and Marcio Dorn (Federal University of Rio Grande do Sul), and Shota Ito and Mathias J. Krause (Karlsruhe Institute of Technology (KIT))

Abstract: Developing performance-portable lattice Boltzmann method (LBM) codes for heterogeneous hardware without extensive hardware-specific duplication is challenging. We introduce a unified hardware abstraction within the open-source OpenLB framework. By expressing models against abstract cell-level concepts, this design enables compile-time binding to platform-specific data structures and common subexpression elimination across composed operators, yielding up to 2.5-fold speedups over an unoptimized baseline. We validate this approach using a rotor-resolved, wall-modeled large-eddy simulation (WMLES) of a 32-turbine wind farm, combining LBM with Brinkman penalization, moving walls and atmospheric boundary layer inflow. This complex fluid-structure interaction executes across NVIDIA, AMD, and Intel GPUs from a single-source implementation without hardware-specific modifications. At extreme scale, we execute the simulation on up to 2,048 Aurora nodes, achieving 86% weak scaling efficiency against a single node. Resolving 5.9 trillion cells at 51 trillion cell updates per second, this is the largest GPU-accelerated LBM simulation reported to date.


Presentation: file

Back to Technical Papers Archive Listing