Optimizing an LBM Application Using CUDA Graphs

With increasing focus on scalability and performance of high performance computing applications, it has become important for the simulation softwares to be able to utilize the underlying hardware as comprehensively to its maximum performance. waLBerla is a multiphysics software framework that has achieved high scalability and performance. It achieves this excellent performance due to architecture specific code generation algorithms combined with efficient communication and parallel data structures like BlockForest. In this work, we attempt to improve the GPU utilization of an Lattice-Boltzmann Method (LBM) software.

---

Contributors:

• Milena Veneva
• Andreas Herten