Jacques-Bernard LEKIEN
Ingénieur-Chercheur en Informatique depuis près de 30 ans, Jacques-Bernard Lekien a travaillé au sein de plusieurs équipes du CEA couvrant ainsi des thématiques autour de la simulation scientifique dans un contexte d’architecture massivement parallèle comme l’optimisation et la parallélisation de codes ou encore le stockage et l’exploitation de données distribuées avec sémantique incluant un aspect in-situ/in-transit.
Aujourd’hui, Jacques-Bernard travaille au sein de l’équipe d’analyse et de visualisation de gros volumes de données avec une attention particulière accordée aux maillages à raffinement adaptatif basés sur une grille d’arbres de décomposition. Ces R&D aboutissent à l’enrichissement des logicielles Themys/ParaView/VTK.
EuroVis 2019 - 21th EG/VGTC Conference on Visualization, 2022
abstract
Abstract
With the constant increase in compute power of supercomputers, high performance computing simulations are producing higher fidelity results and possibly massive amounts of data. To keep visualization of such results interactive, existing techniques such as Adaptive Mesh Refinement (AMR) can be of use. In particular, Tree-Based AMR methods (TB-AMR) are widespread in simulations and are becoming more present in general purpose visualization pipelines such as VTK. In this work, we show how TB-AMR data structures could lead to more efficient exploration of massive data sets in the Exascale era. We discuss how algorithms (filters) should be designed to take advantage of tree-like data structures for both data filtering or rendering. By introducing controlled hierarchical data reduction we greatly reduce the processing time for existing algorithms, sometimes with no visual impact, and drastically decrease exploration time for analysts. Also thanks to the techniques and implementations we propose, visualization of very large data is made possible on very constrained resources. These ideas are illustrated on million to billion-scale native TB-AMR or resampled meshes, with the HyperTreeGrid object and associated filters we have recently optimized and made available in the Visualisation Toolkit (VTK) for use by the scientific community.
Astronomy and Astrophysics, Volume 643, 2020
abstract
Abstract
We present the Extreme-Horizon (EH) cosmological simulation, which models galaxy formation with stellar and active galactic nuclei (AGN) feedback and uses a very high resolution in the intergalactic and circumgalactic medium. Its high resolution in low-density regions results in smaller-size massive galaxies at a redshift of z = 2, which is in better agreement with observations compared to other simulations. We achieve this result thanks to the improved modeling of cold gas flows accreting onto galaxies. In addition, the EH simulation forms a population of particularly compact galaxies with stellar masses of 10 ¹⁰⁻¹¹ M⊙ that are reminiscent of observed ultracompact galaxies at z ≃ 2. These objects form primarily through repeated major mergers of low-mass progenitors and independently of baryonic feedback mechanisms. This formation process can be missed in simulations with insufficient resolution in low-density intergalactic regions.
ISAV 18: Proceedings of the Workshop on In Situ Infrastructures for Enabling Extreme-Scale Analysis and Visualization, p. 7-12, 2018
IEEE Scientific Visualization Conference, 2018
abstract
Abstract
We present a highly efficient solution to interact with the Deep Water Impact Ensemble Data Set provided for the Scientific Visualization Contest 2018. Interactive visualization is made possible on one core of a laptop with the full resolution and the same accuracy as in the original data set, when originally 256 up to 2048 supercomputer nodes were required to generate the data. As far as we know this is the only way to achieve full-resolution exploration on a laptop. We first expose how our approach allows more efficient visualization by using the Tree-Based Adaptive Mesh Refinement grid data structure we introduced in VTK, vtkHyperTreeGrid [1], as compared to structured or unstructured approaches. Then we elaborate on the visualization capabilities offered by vtkHyperTreeGrid-optimized algorithms and the performance achieved on the limited resources available on a laptop. Next, we present how the hierarchical structure makes possible novel ways of exploring data interactively and helps achieve accelerated data exploration by hierarchically driving decimation of values. Finally, we show preliminary results of interactive volume rendering using splatting.
Chapman; Hall/CRC, p. 45-74, 2017
2017
abstract
Abstract
We present here the result of continuation work, performed to further fulfill the vision we outlined in [Harel,Lekien,Pébaÿ-2017] for the visualization and analysis of tree-based adaptive mesh refinement (AMR) simulations, using the hypertree grid paradigm which we proposed. The first filter presented hereafter implements an adaptive approach in order to accelerate the rendering of 2-dimensional AMR grids, hereby solving the problem posed by the loss of interactivity that occurs when dealing with large and/or deeply refined meshes. Specifically, view parameters are taken into account, in order to: on one hand, avoid creating surface elements that are outside of the view area; on the other hand, utilize level-of-detail properties to cull those cells that are deemed too small to be visible with respect to the given view parameters. This adaptive approach often results in a massive increase in rendering performance. In addition, two new selection filters provide data analysis capabilities, by means of allowing for the extraction of those cells within a hypertree grid that are deemed relevant in some sense, either geometrically or topologically. After a description of these new algorithms, we illustrate their use within the Visualization Toolkit (VTK) in which we implemented them. This note ends with some suggestions for subsequent work.
2017
abstract
Abstract
We present here the first systematic treatment of the problems posed by the visualization and analysis of large-scale, parallel adaptive mesh refinement (AMR) simulations on an Eulerian grid. When compared to those obtained by constructing an intermediate unstructured mesh with fully described connectivity, our primary results indicate a gain of at least 80\% in terms of memory footprint, with a better rendering while retaining similar execution speed. In this article, we describe the key concepts that allow us to obtain these results, together with the methodology that facilitates the design, implementation, and optimization of algorithms operating directly on such refined meshes. This native support for AMR meshes has been contributed to the open source Visualization Toolkit (VTK). This work pertains to a broader long-term vision, with the dual goal to both improve interactivity when exploring such data sets in 2 and 3 dimensions, and optimize resource utilization.