Suivez

la liste

Context

Computations of turbulent flows has been undergoing two breakthrough over the last three decades, concerning both numerical methods and computational architectures. On one hand, computing clusters are reaching a size of many hundred of thousands of core, which allows to overcome the RANS modelling for considering unstationary LES or even DNS simulations. This kind of method has required to develop high order numerical methods, which were also improved a lot recently. These last years, heterogeneous massively multicore architectures have been manufactured. These kind of architecture need to redesign implementations and algorithmic in order to be efficiently used.

In the framework of a partnership between INRIA, Airbus, CEA and PPrime institute within the "HPC scalable ecosystem" project co-funded by Région Nouvelle Aquitaine, two goal types are set. On one hand, methodological challenge, in order to improve scheduling tools and support systems for large scale high performance communications and the big data analysis. On the other hand, the application of these improvements in the framework of applications.

The hired postdoctoral fellow will contribute to the applicative challenge which will consist in the development of task based algorithm for the simulation of compressible flows with high order discontinuous Galerkin methods.

An access to the high performance computing experimental bench PlaFRIM (https://www.plafrim.fr/) will be provided.

Assignment

If the exact architecture of the next exascale computing machines is not known yet, the development of heterogeneous architectures (including accelerators or coprocessors like GPU or Xeon-Phi) seems to be an upward trend. Adapting a code to each architecture is a highly demanding task, and can be very intrusive in the code. Moreover, it can strongly depend on the target architecture. For overcoming these problems, runtime schedulers such as StarPU (http://starpu.gforge.inria.fr/), or ParSEC (http://icl.utk.edu/parsec/) have been developed. Using these runtime schedulers allows to develop a code with improved portability. However, this requires to develop a task based implementation. The original algorithm shall then be split into tasks. The runtime scheduler will manage the workflow and will execute the tasks on the available devices (CPU, GPU,…).

The aim of this work is to propose task based implementations of the discontinuous Galerkin method in order to use runtime schdulers. The starting point of this work will be a prototype [1] developed within a project in 2016 for the finite volume approximation of Euler equations. The postdoc will participate to the development of a three dimensional Navier-Stokes solver on structured meshes based on the discontinuous Galerkin method. This code will be based on the runtime scheduler StarPU [2].

The postdoc will participate in the porting and scalability tests on different parallel machines, such as PlaFRIM, the machines of the technology watch group of GENCI, and also the largest European platforms (PRACE).

During his stay, the postdoc will interact with the other members of the project "HPC Scalable ecosystem", in order to benefit of their skills on parallelisation and scalability of algorithm, and in order to test the methodological improvements that will be implemented in the runtime scheduler StarPU during the project.

[1] Mohamed Essadki, Jonathan Jung, Adam Larat, Milan Pelletier and Vincent Perrier, A Task-Driven Implementation of a Simple Numerical Solver for Hyperbolic Conservation Laws, ESAIM: Proceedings, 63 (2018) 228-247
[2] Cédric Augonnet, Samuel Thibault, Raymond Namyst, and Pierre-André Wacrenier
StarPU: A Unified Platform for Task Scheduling on Heterogeneous Multicore Architectures. Concurrency and Computation: Practice and Experience, Special Issue: Euro-Par 2009, 23:187-198, February 2011

Main activities

  • State of the art on the algorithm of high order methods and their implementation
  • Development of task based algorithm for the discontinuous Galerkin method
  • Implementation based on the runtime scheduler StarPU.
  • Scalability tests
  • Dissemination of the results (conferences, publications)

Remuneration

2653€ / month (before taxs)

Link to the Inria post-doctoral position

https://jobs.inria.fr/public/classic/en/offres/2019-01721