AS2.1. SUSTAINING, IMPROVING, OPTIMIZING AND REFACTORING LEGACY CODES AND QUANTUM, ADVANCED PROGRAMMING METHODS (DSL, GPU, NEW LANGUAGES) & TECHNOLOGY WATCH
High Performance Computing (HPC) has been central to CERFACS for over 30 years, establishing its international leadership in the field. Today, two major strategic directions guide its HPC efforts. First, maintaining and evolving legacy codes—mature, high-performance software like AVBP—is critical. These codes are widely used in both research and industry but face challenges due to rapid changes in computing hardware, which evolves every 2–4 years. Updating and refactoring these complex codes for new architectures is a demanding research task and a shared concern across CERFACS and its partners. Second, CERFACS is preparing for the future through exploration of advanced programming techniques, heterogeneous computing, and emerging architectures such as GPUs and quantum computers. Achieving performance portability—the ability for software to run efficiently across varied hardware—is essential. While current programming languages (C, C++, Fortran, Python) lack universal support for heterogeneous computing, tools like OpenMP, OpenACC, Kokkos, and domain-specific languages are being explored to bridge this gap. These frameworks aim to abstract hardware complexity, enhancing both performance and maintainability. Finally, CERFACS maintains active technology watch efforts, including developments in quantum computing, to ensure it remains at the cutting edge of HPC innovation and well-positioned for future challenges.
AS2.2 COUPLING
CERFACS conducts research that combines methodological studies with tool development in high-performance computing (HPC), aiming toward exascale computing. A central focus is on code coupling, enabling efficient, scalable simulations that integrate existing legacy codes for applications like data assimilation, multiphysics modeling, design optimization, and uncertainty quantification. CERFACS has been active in coupling research for over 20 years, developing key open-source libraries such as OASIS3-MCT (with CNRS) and OpenPALM (with ONERA), providing unified frameworks for complex coupled systems.
Recent advances include the ANR COCOA project, which demonstrated the effectiveness of the Schwarz iterative method for reducing temporal inconsistencies in atmosphere-ocean models and highlighted limitations of asynchronous coupling in climate simulations. The CWIPI coupling library plays a central role in high-fidelity CFD simulations using AVBP, particularly for turbomachinery applications. This work, recognized by awards like the 2020 Teratec Trophée, supports industrial and academic partners including SAFRAN and TOTAL.
CWIPI also supports multiphysics simulations involving fluid-solid heat transfer, relevant in hydrogen combustion studies. Major updates to OASIS3-MCT (versions 4.0 and 5.0) introduced improved performance, parallelization, regridding capabilities, and new tools. Though PALM development paused due to resource constraints, user support and collaboration with ONERA on CWIPI development and integration with the Paradigm library have continued actively.
AS2.3. HPC WORKFLOW (INCLUDING DATA MANAGEMENT)
High Performance Computing (HPC) is essential across both fundamental research and industrial design, but transferring tools from academic experts to everyday engineers remains a major challenge—often described as “crossing the chasm.” While current HPC users are typically early adopters, widespread adoption requires making simulations reliable and repeatable over many years. For engineers, a “successful” simulation must run correctly on the first try, provide valuable insights, be affordable in time and resources, and be easily reproducible later. These practical needs often outweigh pure performance gains, such as a 20% speed-up, if they prevent running incorrect or failed simulations.
With the shift toward exascale computing, challenges scale dramatically: job sizes, numbers, and data volumes increase by orders of magnitude. Users can no longer move simulations between local machines and clusters due to bandwidth limitations and instead perform the full workflow—setup, execution, and post-processing—on the cluster itself. This demands new workflows and user tools.
CERFACS focuses on improving current HPC user experience while preparing for exascale systems. This includes optimizing interfaces (graphical or command-line), reducing hidden inefficiencies, lowering support and training costs, and minimizing manual job management time. The goal is to make HPC more accessible, sustainable, and efficient for a broader engineering community.
