Session: 34-04 Computing architectures and solvers

Paper Number: 152583

Numerical Design of Experiments for Repeating Low-Pressure Turbine Stages Part I: Computational Opportunities and Methodology 

The current trend towards more compact and efficient low-pressure turbine (LPT) designs can substantially benefit from advanced numerical predictive tools. The complex transitional and turbulent nature of unsteady flows seen in LPTs often demands high-order methods such as Large Eddy Simulations (LES) for accurate predictions in turbine efficiency and loss generation. Integrating high-fidelity simulations into design cycles, predominately driven by rapid Unsteady Reynolds-Averaged Navier-Stokes (URANS) calculations, requires cutting-edge numerical tools able to leverage modern high-performance computing architectures. In Part 1 of this paper, a multi-fidelity simulation framework is presented, maximizing the computational output of the latest high-performance architectures by fully occupying the hardware with concurrent LES and URANS simulations. 
In general, the continuous rise in computational power of supercomputing facilities has primarily been driven by advances in Graphics Processing Units (GPUs). A single GPU has thousands of parallel computing threads, offering substantial parallelism and enabling the rapid processing of large datasets. With dedicated software, GPUs are ideally suited for performing LES that require extensive grid counts for adequate resolution. In contrast, modern Central Processing Units (CPUs) typically have between 32 and 64 cores, offering lower performance than GPUs, and are better suited for less compute-intensive tasks, such as Reynolds-Averaged Navier--Stokes calculations. A typical supercomputer node consists of one or two multi-core CPUs paired with up to eight GPUs. While previous studies have focused on numerical methods optimized for either CPU or GPU architectures, this paper introduces a novel multi-fidelity approach that leverages both CPUs and GPUs on the same node concurrently, an opportunity to make better use of modern supercomputers. In this framework, LES are executed on the GPUs, with each GPU requiring a dedicated CPU core to handle host operations such as data reading and writing. Typically, the remaining CPU cores on the supercomputer node would therefore be idle, which is effectively a waste of computational resources. In contrast, the proposed method allows rapid URANS simulations to run in parallel with the GPU-based LES on the otherwise-idle CPU cores on the same supercomputer node. As a result, this approach maximizes the efficiency and computational power of the entire node, and provides high-fidelity and low-fidelity results at the same time.
Furthermore, we demonstrate the use of this multi-fidelity framework within containerized environments, which are isolated images independent of the supercomputing facility's software stack. This allows use of the latest versions of compilers and optimized libraries, which are often outdated on the supercomputer itself. By employing containers, greater flexibility and access to cutting-edge software is ensured, enhancing performance and compatibility without relying on the supercomputer's default configurations. 
Lastly, the first part of this study outlines the numerical methodology for both CPU and GPU solvers including detailed grid convergence studies to demonstrate the quality of the results obtained for the repeating stage simulations at engine-relevant conditions which are discussed in the second part of the paper. 

Presenting Author: Marco Rosenzweig The University of Melbourne

Presenting Author Biography: Marco Rosenzweig is a third-year Ph.D. student at the University of Melbourne. In 2022, he obtained his M.Sc. degree in Aerospace Engineering at the Technical University of Munich and has two years of industrial work experience at MTU Aero Engines AG. Marco’s research expertise is in the turbomachinery flows critical to aircraft engines with a focus on low-pressure turbines. To date his research interests have included modeling of multi-stage component interactions, the evaluation of standard industrial design methods and scale-resolving, numerical methods. His most recent work focuses on performing multi-fidelity simulations on the latest supercomputing architectures.

Authors:

Marco Rosenzweig The University of Melbourne
Melissa Kozul The University of Melbourne
Richard Sandberg The University of Melbourne
Giovanni Giannini The University of Florence
Roberto Pacciani The University of Florence
Michele Marconcini The University of Florence
Andrea Arnone The University of Florence