Cerfacs Enter the world of high performance ...

PhD Defense – Thibaut LUNET: Space-time parallelization strategies for numerical simulation of turbulent flow

  Tuesday 9 January 2018 at 13h30

  Phd Thesis       ISAE-SUPAERO: salle des thèses    

Nowadays, the problem size required to simulate physical phenomena may  be extremely large in order to correctly predict increasingly complex physical behaviours. This is especially true for the numerical simulations of turbulent flows, whose size has recently reached more than 10^11 degrees of freedom (Lee & Moser, 2013). In order to further develop computing capabilities, new supercomputers will appear in the coming decades, being able to compute up to 10^15 FLOPS. However, classical parallel algorithms (usually based on domain partitioning in space) become less and less efficient when used in such a massively parallel context. For the past ten years, the HPC community has been studying the possibility to exploit time decomposition as a new degree of parallelism.

This thesis aims at presenting the current state of the art for time-parallel methods, in order to identify space-time parallelism strategies that can be applied to an explicit CFD code in a massively parallel context. After a first analysis, the Parareal algorithm with spatial coarsening is selected. First, this time parallel algorithm is then studied on a simplified model of the Navier-Stokes equations (advection equation) by means of a Fourier analysis and multiple numerical experiments. This allows us to identify the important parameters influencing the accuracy and the convergence of the algorithm.
Then, secondly, a specific attention is devoted to the definition of two three-dimensional canonical turbulent problems that can serve as relevant test cases for time-parallel algorithms: the decay of isotropic homogeneous turbulence and the turbulent channel flow. Detailed studies are being conducted on both problems to determine the performance brought by Parareal with space coarsening as part of a combined space-time parallelization, as well as its ability to correctly represent the physical properties of the turbulent flows. Finally, this thesis describes the gains that can bring time parallelism for CFD, as well as its perspectives on tomorrow’s high performance computing architectures.

 

Jury :

Martin Gander              (Université de Genève)                               Rapporteur
Rolland Masson            (Université de Nice)                                    Rapporteur
Marc Massot                 (École polytechnique)                                 Rapporteur
Qiqi Wang                     (Massachussets Institute of Technology)   Jury
Daniel Ruprecht            (University of Leeds)                                  Jury
Serge Gratton                (Toulouse-INP-IRIT)                                   Directeur
Julien Bodart                 (ISAE-Supaero, DAEP)                              Encadrant
Xavier Vasseur              (ISAE-Supaero, DISC)                               Encadrant

 

 

 

 

CALENDAR

Monday

04

December

2023

Artificial Intelligence for computational physics

From Monday 4 December 2023 to Friday 8 December 2023

  Training    

Monday

11

December

2023

Multi-architecture parallelism using Kokkos/C++ library

From Monday 11 December 2023 at 14h00 to Wednesday 13 December 2023 at 17h00

  Training    

Tuesday

26

March

2024

ALL EVENTS