49ième Forum : Le PEPR Numpex « Numérique pour l’Exascale »

49ième Forum : Le PEPR Numpex « Numérique pour l’Exascale »

29 novembre 2022, Maison de la simulation

Inscription à l’adresse : https://orap49.sciencesconf.org/

9h30 – 16h45

Le PEPR Numérique pour l’Exascale (NUMPEX) a pour objectif de concevoir et développer les briques logicielles qui équiperont les futures machines exascales et de préparer les grands domaines applicatifs à exploiter pleinement les capacités de ces machines. Grands domaines applicatifs qui relèvent aussi bien de la recherche scientifique que du secteur industriel.

  • 9h Accueil

  • 9h30 - 9h50 Introduction (J.Y. Berthou, Inria)

  • 9h50-10h20 Methods and Algorithms for Exascale - Exa-MA (C. Prud'homme)

    There is a growing number of problems where experiments are impossible, hazardous, or extremely expensive. Extreme-scale computing enables the solution of vastly more accurate predictive models and the analysis of massive quantities of data thanks to AI. Combining predictive modeling with data, coupled with machine learning and AI strategies, can create new opportunities in science. In particular, move from Human-in-the-Loop towards hybrid Human and Artificial Intelligence-driven design, discovery, or evaluation. However, various scientific and technical challenges need to be met to exploit exascale computing capabilities. These bottlenecks impact methods and algorithms in a profound way on all aspects of the simulation toolchain through (i) avoidance of communication, (ii) adaptive parallel grain and more compute-intensive at node level, (iii) handling of heterogeneous hardware and data representations and (iv) self-parametrization.

    The Exa-MA project concentrates on the Exascale aspects of the numerical methods, ensuring their scalability to existing and forthcoming hardware. Furthermore, it is a transverse project, proposing methods and tools where modeling, data and AI, through algorithms, are central.

  • 10h20– 10h50 HPC software and tools – Exa-SofT (François Trahay, Telecom SudParis)

    Though significant efforts have been devoted to the implementation and optimization of several crucial parts of a typical HPC software stack, most HPC experts agree that exascale supercomputers will raise new challenges, mostly because the trend in exascale compute-node hardware is toward heterogeneity and scalability: Compute nodes of future systems will have a combination of regular CPUs and accelerators (typically GPUs), along with a diversity of GPU architectures. Meeting the needs of complex parallel applications and the requirements of exascale architectures raises numerous challenges which are still left unaddressed.

    As a result, several parts of the software stack must evolve to better support these architectures. More importantly, the links between these parts must be strengthened to form a coherent, tightly integrated software suite. Our project aims at consolidating the exascale software ecosystem by providing a coherent, exascale-ready software stack featuring breakthrough research advances enabled by multidisciplinary collaborations between researchers.

  • Pause

  • 11h10 – 11h40 Data-oriented Software and Tools for the Exascale - Exa-DoST (J. Bigot, CEA)

    The advent of future Exascale supercomputers and the need to ensure that applications will fully exploit them raises multiple major data-related challenges. Beyond the need to advance data I/O, storage, processing and analytics to the next scale, several trends are driving the need for further specific progress in this area: the emergence of new storage technologies (e.g., first generation of commercialized NVM), which can help build faster and more scalable storage solutions, while creating a more complex storage hierarchy; the requirement to support the integration of traditional HPC simulations, Big Data, and AI in a single environment leads to the need to support new I/O schemes and data access patterns; the growing presence of accelerator-based technologies in HPC systems enabling faster, accelerator-based data analytics; the need to couple heterogeneous data from many different sources (HPC facilities, cloud-based repositories, IOT/edge-originated data from the real world). All these evolutions lead to major challenges for data transfer, I/O, storage, processing and analysis.

  • 11h40 – 12h00 GYSELA roadmap for exascale gyrokinetic plasma turbulence simulations (Virginie Grandgirard, Association Euratom-CEA, CEA/DRF/IRFM Cadarache, France)


    Predicting the performance of fusion plasmas in terms of amplification factor, namely the ratio of the fusion power over the injected power, is among the key challenges in fusion plasma physics. In this perspective, turbulence and heat transport need being modeled within the most accurate theoretical framework, using first-principles non-linear simulation tools. The gyrokinetic 5D code GYSELA [V. Grandgirard et al., Comp. Phys. Com. 2017 (2016) 35] has been developed in this framework at the IRFM/CEA for 20 years. Less than a dozen codes similar to GYSELA are developed worldwide. Because of the complexity of the problem, each of them proceeds from certain approximations. In this highly competitive field, the GYSELA code has a number of advantages that make it unique. It is the only code in the EU (and one of the very few worldwide) capable of addressing core and edge interaction in the flux driven regime. With a consumption of 150 million CPU hours per year, the code already makes massive use of petascale computing capacities. Addressing ITER-size plasmas will increase the numerical needs by 1 to 2 orders of magnitude, at least, requiring exascale HPC capabilities.

    The GYSELA code, developed in Fortran and based on a hybrid MPI/OpenMP parallelization, runs efficiently on more than 100’000 cores on current standard architectures. However, the new architectures planned for the exascale, based on heterogeneous accelerated computing nodes, are less favorable to applications such as GYSELA which require a lot of memory and parallel communications. Achieving performance on exascale computers is a major challenge for GYSELA but will be far from trivial. The GYSELA code produces very large amounts of data. A typical 5D mesh contains several hundreds of billions of points, which leads to 5D distribution functions of the order of 2 TB to be followed at each time iteration. Knowing that a simulation can represent several tens of chained runs of more than 10’000 iterations, it is not conceivable to store the temporal evolution of these distribution functions. They are only saved at the end of each run in temporary files to allow the restart (checkpoint-restart). In the end, out of the Petabytes of data manipulated during a GYSELA simulation, only a few Terabytes are saved due to storage capacity limits. This data reduction is based on saving at fixed time steps a number of mainly 3D fluid quantities.  Knowing that there is a growing gap between CPU performance and I/O bandwidth on large-scale systems, this post-hoc approach is already very constraining and will become even more so.

    We will present the GYSELA roadmap for the next few year to prepare exascale gyrokinetic simulations for plasma turbulence in ITER-like plasmas.


    Dr. Virginie Grandgirard, research director at CEA, is the lead developer of the 5D non-linear gyrokinetic semi-Lagrangian code GYSELA used for plasma turbulence simulations https://gyselax.github.io/ . This code is highly parallelized up to hundreds of thousands cores. Her research interests focus on numerical methods for Vlasov equations, high performance computing, tokamak plasma turbulence and more recently on Physic Informed Neural Network. She has co-authored 60 publications in peer-reviewed journals.


  • 12h00 - 12h30 Development and integration project - Exa-DIPP (J.P. Vilotte, IPGP)

    This integrated project is designed to steer and accelerate more valuable and rapid insights from a variety of science and engineering driven application demonstrators that require a much higher level of inherent effectiveness in all methods, software tools, and exascale-enabled computing, data analytics and AI to be acquired. This project is in charge of the developments of the demonstrators identified by the four integrated research projects (Exa-MA, Exa-Soft, Exa-DoST, Exa-AToW) and the identification and the development of strategic applicative demonstrators. It will be supported by a networked data science and computing team to accelerate the delivery of science-based computational applications that can effectively exploit exascale systems.

  • Pause déjeuner

  • 14h00-14h30 Architectures and Tools for Large-Scale Workflows – Exa-AToW (F. Bodin, Irisa)

    The emergence of exascale supercomputers for addressing scientific challenges (e.g. creating digital twins) is concomitant to a deluge of data from large scientific instruments, simulations, or IoT infrastructures. The new discovery process relies on combining HPDA and HPC in application workflows. These workflows must be deployed at a large scale, involving numerous multidisciplinary distributed infrastructures (HPC centers, Clouds, …) distributed over multiple geographical areas. This transversal axis aims at answering simple questions, typically raised by scientific users and application developers:

    • How to allocate data and compute resources?
    • How to ensure efficiency, productivity and reproducibility of executions?
    • How to visualize results and exploit the application’s resulting data (in a FAIR spirit)?

    While simple to express, current technologies and state-of-the-art systems architectures are making the task too complex for end-users (especially in a productive and efficient manner), and prevent a generalization of the use of HPC and HPDA. This project aims at defining a coherent architecture with all the tools needed to deploy scientific applications based on large-scale workflows.

  • 14h30 – 14h50 SKA: a global experiment for Big Science across the continuum (D. Gratadour, Observatoire de Paris)

  • 14h50-15h10 CoE ChEESE2, vers l'exascale pour la Terre solide (V. Monteiller, LMA)

  • 15h10-15h30 Présentation du projet européen Center of Excellence MaX (T. Deutsch, CEA)

  • Pause

  • 15h45-16h05 Centre d'excellence Hidalgo2 (C. Prud'homme, Cemosis)

  • 16h05-16h20 Point Europe (J.P. Nominé, CEA)

  • 16h20-16h45 Présentation du Prix Caseau

    Paul BONIOL pour sa thèse :

    « Detection of Anomalies and Identification of their Precursors in Large Data Series Collections »

    Rem-Sophia MOURADI pour sa thèse :

    « Modélisation non-linéaire de champs multidimensionnels guidée par la donnée : application aux écoulements côtiers hydro-morphodynamiques »