Contribution
An Efficient Framework for Hybrid In-Situ Aeroacoustic Simulations Using Spectral Element CFD Code Neko
* Presenting author
Abstract:
As high-performance computational fluid dynamics (CFD) codes increasingly adopt GPUs, CPUs often remain underutilized. On the other hand, offline computational aeroacoustic (CAA) simulations require enormous amounts of data exchange between the CFD and acoustic solvers. Given the high temporal resolution often required for aeroacoustic simulations of turbulent flows, data transfer poses the main bottleneck for large-scale CAA simulations.In this work, we present an efficient framework for highly parallelized hybrid aeroacoustic simulations at exascale levels as part of the EXCELLERAT Phase 2 project. The spectral element CFD package Neko is utilized on GPUs to perform DNS of turbulent flow over a forward-facing step at a Reynolds number of 8,000 based on the step height. We employ an asynchronous data streaming approach to feed data from the flow solver to the acoustic solver, enabling efficient use of idle CPUs for concurrent in-situ aeroacoustic simulations.We discuss the performance and scalability of this approach across different problem sizes. Furthermore, we provide insights into the requirements of the flow simulation and the treatment of source terms necessary to reduce numerically induced spurious sound. Ultimately, the result of the flow and aeroacoustic simulations are shown and validated against the experiment.