Accès direct au contenu

CMLA

Version anglaise

aide

le CMLA - ENS Cachan au 37e forum ORAP

Florian De Vuyst (CMLA, ENS Cachan) s'exprimait lors du dernier forum ORAP dédié au calcul haute performance, à ses infrastructures et applications dans un contexte de recherche au niveau européen.

Le 17 mars dernier, l'édition 2016 du forum Orap s'est tenue au siège du CNRS à Paris, sur le thème :
Le nouveau paysage européen du calcul haute performance (HPC). Retour d'expériences sur les accélérateurs et autres manycores.

Retrouvez l'ensemble des présentations de la journée : slides, résumés et vidéos des interventions sur le site.

La présentation de Florian De Vuyst,
Professeur à l'ENS Cachan, co-directeurs du CMLA (CNRS UMR 8536)

Qu'est-ce que le HPC a changé dans notre pratique du calcul scientifique et de la modélisation numérique ?




Multicore/manycore with a scientific computing culture: experience feedback with special focus on performance modeling and redesign of computational methods

Our first experience in GPU computing was to design real-time interactive computational 2D flow dynamics simulation, using Infra Red pointers. First experiments succedeed and showed significant speedups on NVIDIA TESLA boards, but with noteworthy variations according to the choice of data structures but also to the choice of the computational approach.

Performance issues appeared to be somewhat difficult to understand. So we decided to focus on performance modeling, on multicore architectures as a first stage. Starting from a legacy Hydrodynamics solver (of the Lagrangian remapping family), we achieved the performance modeling of the method using first a roofline-type model, then a finer Execution Cache Memory (ECM) model.

The analysis kernel-per-kernel has given strong highlights on the global performance of the numerical approach. It even gave us new ideas of redesign of alternative computational approaches. SIMD feature, arithmetic intensity and data alignment are of course key factors.

More generally, some computational approaches appear to be particularly suitable of HPC on manycores like Lattice Boltzmann approaches for example, sometimes inviting for changes of paradigm for the numerical model.


Organisation





Présentation F. De Vuyst

Florian De Vuyst

Florian De Vuyst is full-professor at Ecole Normale Supérieure de Cachan Université Paris-Saclay, co-director of Centre de Mathématiques et de leurs applications, CMLA CNRS UMR 8536, Director of the multidisciplinary Farman Institute FR 3311 and teaches at the Department of Mathematics.

Florian's fields of Research are scientific computing of PDEs, numerical modeling, numerical analysis, HPC on manycore/multicore, parallel algorithms, CFD and reduced-order modeling. He is involved in the governance of Fondation de Mathématiques Jacques Hadamard FMJH, the structure of Mathematics at Paris-Saclay University.

He was also awarded for a NVIDIA CUDA Research Center in 2013 and for the EquipEx DIGISCOPE, big infrastructur
e of large display walls and devices for visualization and interaction withing Saclay's area.