Menu

Success Story: GPU-accelerated aerodynamics simulations with CODA

success story # highlights:

  • Keywords:
    • Computational Fluid Dynamics
    • Aerospace
    • High-Performance Computing
    • GPU
    • Linear equation systems
    • Aircraft aerodynamics
  • Industry sector: Aerospace
  • Key codes used: CODA, Spliss, FlowSimulator

Benefits:

  • Readiness for current and future large-scale GPU-based HPC systems consisting of several hundreds of GPUs
  • Simulation of aircraft aerodynamics with shorter time to solution of up to a factor of 8.6 compared to CPU systems
  • Simulation of aircraft aerodynamics with higher resolution and higher-quality results
Simulation of external aircraft aerodynamics. Credit and copyright: DLR (CC BY-NC-ND 3.0)

Organisation involved:

DLR is the German aerospace research and technology centre. At its core, DLR develops technologies for aeronautics and space, energy and transport, as well as security and defence research. A broad spectrum of results and innovations benefit industry and business, authorities and administration, as well as public stakeholders. The DLR fulfils its responsibility to society through an intensive exchange of knowledge and targeted technology transfer.

codes involved:

CFD by ONERA, DLR and Airbus (CODA) is a Computational Fluid Dynamics software based on second-order finite-volume and higher-order Discontinuous-Galerkin discretisations on unstructured grids. CODA is being developed as part of a collaboration between the French Aerospace Lab ONERA, the German Aerospace Center (DLR), Airbus, and their European research partners.

technical / scientific challenge:

Providing high-fidelity aerodynamics simulations requires enormous computational resources. High-Performance Computing (HPC) systems aim to meet the demands for increasingly complex simulations with systems built of millions of heterogeneous processing elements including hardware accelerators such as Graphics Processing Units (GPUs). However, harnessing the full potential of these HPC systems requires CFD software such as CODA to not only simultaneously utilise different hardware architectures but also efficiently scale across a large number of GPUs. This preparation for modern HPC systems is a technical challenge up and down the software stack and can even go as deep as kernels and algorithms.

Solution:

An integral part of CODA is the Sparse Linear Systems Solver (Spliss) that is used for solving linear equation systems for implicit time integration methods. Spliss is a linear solver library that, on the one hand, is tailored to the requirements of CFD applications such as CODA but, on the other hand, is independent of any particular CFD solver. Focusing on the specific task of solving linear equation systems allows for integrating more advanced, but also more complex, hardware-specific optimisations, while at the same time hiding this complexity from the CFD solver.

Spliss already allows CODA the transparent usage of large GPU systems by providing a linear solver library that efficiently scales across a large number of GPUs. However, within EXCELLERAT, the scalability of Spliss has been significantly improved to leverage the enormous computational power of current GPU-based HPC systems consisting of hundreds of GPUs.

impact:

CODA will be the successor to DLR’s TAU CFD solver, which has been in production in the European aircraft industry, research organisations, and academia for more than 20 years. TAU was, for instance, used for the Airbus A380 and A350 wing design. To be a full replacement, though, CODA must not only match and exceed TAU’s numerical capabilities, but must also efficiently utilise current and upcoming HPC systems with their increasing use of GPUs.

Through EXCELLERAT, these significant improvements have been made in performance and scalability on GPU-based HPC systems. In combination with efficient, hybrid CPU parallelisation, EXCELLERAT’s work on Spliss resulted in the performance gain achieved with a single GPU being scaled to large distributed systems consisting of hundreds of GPUs, while reducing the runtime in a distributed set-up on Nvidia A100 GPUs by up to 50%. GPU acceleration of CODA can yield a up to 8.6 times speedup over state-of-the-art CPU systems or a up to 1.9 times speedup when equated for power consumption.

With EXCELLERAT’s added support for large-scale GPU systems, CODA is now well-prepared for current and future heterogenous HPC systems including exascale systems. CODA can now run larger, more accurate simulations of aircraft aerodynamics in weeks instead of months. During the design process of future aircraft, CODA can now provide reliable insight into new aircraft technologies to reach the best overall aircraft performance. Larger and more accurate simulations enable the design of quieter, safer, and more fuel-efficient planes that play a pivotal role on the path to climate-neutral aviation.

Potential EXCELLERAT Services:

  • GPU-optimisation service for CFD applications
  • Scalability tuning and benchmarking on exascale systems
  • Co-design consulting and application porting support
  • Tailored HPC training programmes for industrial engineers
  • Production of engineering best-practice guides

unique value of each service:

  • GPU-optimisation service: Speeds up simulations by adapting CFD codes like CODA for efficient GPU use.
  • Scalability tuning and benchmarking: Ensures optimal performance of applications on large-scale and exascale HPC systems.
  • Co-design consulting: Aligns software with future HPC hardware, improving efficiency and sustainability.
  • Tailored HPC training for engineers: Builds practical skills in simulation workflows, GPU use, and HPC optimization.
  • Best-practice guides: Offer quick access to expert knowledge and methods for engineering simulations.