Faits marquants

HP-Feel++ obtient 60 Millions d’heures de calcul lors du 6ème appel Prace

HP-Feel++ coordonné par C. Prud’homme (U. Strasbourg,IRMA) a obtenu 60 millions d’heures de calcul lors du 6ème appel Prace sur SUPERMUC(GAUSS@LRZ,Germany).
Le communiqué de presse de Prace est disponible ici. HP-Feel++ est une collaboration entre U. of Strasbourg(France), U. Joseph Fourier(Grenoble, France), CNRS, U. Pierre et Maris Curie (Paris) et U. Coimbra(Portugal).

Image de SuperMUC au Gauss computing center

The HP-Feel++ (High Performance Feel++) project aims at developing two research applications that
require now access of TIER-0 computing resources : blood flow rheology and high field magnets.

Although these domains are quite different they have been thoroughly developed for the past few years within the Feel++ project (http://www.feelpp.org). They share the same mathematical kernel that encompasses a large range of numerical methods to solve partial differential equations such as (i) arbitrary order continuous and discontinuous Galerkin methods in 1D, 2D and 3D, (ii) domain decomposition methods, (iii) fictitious domain methods, (iv) level-set methods or (iv) certified reduced basis methods. These methods are developed and used easily using a domain specific language embedded in C++ mimicking the mathematical language associated to Galerkin methods. This language allows physicists, engineers and mathematicians to focus on the numerical methods as well the physics whilst it hides the computer science details (e.g. parallelism) or algebraic solvers and enables the user to ramp up very quickly from rapid prototyping numerical methods to large scale computations. Within this context, blood flow rheology and high field magnets are the two domains driving Feel++ developments.

In blood flow rheology, we are interested in simulating suspensions of red blood cells (RBC) in arteries and veins and in studying the fluid properties (i.e. the fluid apparent viscosity) either in healthy contexts (our current focus) or pathological contexts (in the longer term). Not only the RBC are deformable entities, arteries and veins deform also during blood pulse ; in both cases fluid structure interaction modeling and simulations are required. We have developed two main alternatives to tackle these problems : (i) fluid structure interaction within the so-called Arbitrary Lagrangian Eulerian framework coupled with a fictitious domain method to handle the RBC and (ii) fluid structure interaction using level-set methods. In both cases, the computational and storage costs for realistic simulations require using the TIER-0 infrastructures.

As to high field magnets (i.e. magnetic intensity greater than 24T), they are being developed by a large scale equipment laboratory (Laboratoire national des champs magnetiques intenses) and they are accessible to the international scientific community through project calls. Studies range from solid physics to applied supra-conductivity and magneto-science. The design and optimisation of these high field magnets require the solution of large scale multi-physics (and mildly multi-scale) non-linear partial differential equations. Moreover to ensure a robust design, we need to assess uncertainties through quantile estimations and sensitivity analysis. The latter is built on the former as it requires hundred or thousands evaluations of the former. We have developed the so-called certified reduced basis in this context to reduce the computational cost within the uncertainty quantification and optimisation processes from millions of degrees of freedom to a few tens or hundreds. This huge computational gain requires however the acceptance of an intensive offline stage allowing to get the independence with respect to the costly (typically finite element) underlying models and which demands now the access to TIER-0 infrastructures.