DESCRIPTION :
The CONCACE project, jointly run by Airbus, Cerfacs and Inria, proposes to tackle this problem by taking advantage of modern development tools and languages to design high-level expressions of complex parallel algorithms. While the traditional approach to HPC is to fully exploit the hardware, our complementary approach will enable a richer composability of numerical methods, allowing existing and new numerical algorithms to be fully exploited. In this study, we propose to attempt an initial implementation of this methodology in the H-matrix solver (used by the ASERIS and ACTIPOLE software packages mentioned above) for multi-arithmetics and compression., * the state of the art in this field of activity;
* how SZ positions itself with respect to the other lossy compression methods existing;
* Implement the coupling between SZ and the H-matrix solver, at least in the sequential case;
* How does this impact the accuracy and performance of the H-matrix library;
* What is the computation cost for this new approach.
Deliverable :
* A bibliographic report on lossy compression in scientific computing applications ;
* An updated sequential H-Matrix library using SZ (in the github project hmat-oss).
part 2 : Multi-arithmetics experimentation
Multi-arithmetics as proposed in the literature is a new promising way of writing algorithms using all the possible accuracies offered by modern processors in order to maximize performance and minimize storage without impacting accuracy. In the H-matrix solver, we must evaluate the way to implement it in a complex parallel algorithm, and see the impact on computation time and accuracy.
Deliverable :
* A scientific report presenting experimentation of H-matrix with multi-arithmetics ;
* An updated software version of h-matrix., * The candidate will have to appropriate the theoretical and practical aspects of the H-matrix method, and the computer code implementing it at Airbus.
* Study and understand the different improvement paths proposed, understand the theoretical aspect, carry out the computer implementation and practical tests of these, ensure validation on industrial cases and documentation. The 4 tracks proposed are Task-based distributed parallelism, mixed arithmetic, randomized algorithms, heterogeneous computation. This list is not fixed and may evolve during the course of the thesis.
* The candidate must be able to design an experimental plan and conduct a scientific study, make scientific presentations to a specialist or generalist audience, and write scientific articles and technical reports.
Temps partiel / Temps plein : Plein temps
Type de contrat : Contrat à durée indéterminée (CDI)
Compétences : C (Langage de Programmation), C ++ (Langage de Programmation), Outils de Programmation, Fortran (Programming Language), Github, Python (Langage de Programmation), Shell Unix, Parallel Computing, Informatique Scientifique, Enterprise Application Platform, High Performance Computing, Programming Languages, Algorithmes, Arithmétique, Expérimentation, Méthodes par Éléments Finis, Algèbre Linéaire, Analyse Numérique, Documentation Scientifique, Recherche Scientifique, Capacités de Démonstration, Randomisation (randomness)
Courriel :
Guillaume.Sylvand@inria.fr
Téléphone :
0524574000
Type d'annonceur : Employeur direct