The paper is devoted to the development of a non overlapping domain decomposition method suited to granular dynamics. The formulation and the efficiency of such a method are well established for structural mechanics. In order to extend this approach to granular systems a so-called primal splitting of the domain is chosen because it is a less intrusive method for software development.Once the interface problem is defined and the solver is slightly enriched with some extra numerical parameters, the method is tested on railway ballast simulations for improving the maintenance of railway tracks. Mathematics Subject ClassificationThe simulation of more and more realistic granular media leads to intensive computations. The size of the simulation increases both in term of number of bodies or particles and of duration of the process to be simulated. The increase of the computations concerns then both the space and the time, because a granular system often exhibits a complex dynamic behavior. Indeed dense collections of grains subjected to dynamic loadings require large-scale samples simulated on a long time to capture the local dynamic crises responsible of the global behavior of the system. Many applications involve such dense assemblies as the sand piles, dunes, blocky rocks, powders. We are here specially interested in a class of applications for which the connectivity of the grains remains almost stable during the studied process, as the masonry of monuments (before collapse) or the railway ballast.The denominations of 'discrete element methods' and 'distinct element methods' (DEM) are commonly used to refer to the pioneering approach of Cundall [11], today implemented in a large range of commercial pieces of software intended to handle non-interpenetrability. Also, because the computation techniques applied in such implementations are close to those of molecular simulations, the denomination of 'molecular dynamics' (MD) method is also used, specially in the domain of granular mechanics [16]. Such an approach is based on regularization strategies both for the dynamics and the interactions between grains. First, the non-interpenetrability constraints are replaced by some stiff 'elastic' repulsion laws which

Numerical simulation with discrete elements leads to several issues for large scale problems and long loading times, as for the granular dynamic simulations of the ballasted railway behavior. To reduce computational costs, we study the use of two strategies: domain decomposition methods and shared-memory parallelization with OpenMP. An example of a maintenance process, the tamping, on a portion of railway track with 7 sleepers, is simulated.RÉSUMÉ. La simulation numérique par éléments discrets présente des difficultés pour l'étude de problèmes de grande taille et en temps de sollicitation long, comme la dynamique des milieux granulaires pour le ballast ferroviaire. Afin de résoudre ce problème à moindre coût, on propose d'allier deux stratégies : la décomposition de domaine (DDM) et le calcul parallèle (en mémoire partagée avec OpenMP). Un exemple traitant d'un procédé de maintenance ferroviaire, le bourrage, sur une portion de voie ballastée de 7 blochets de long est étudié.

Numerical simulation with discrete elements leads to several issues for large-scale problems and long loading times, as for the granular dynamic simulations of the ballasted railway behaviour. To reduce computational costs, we study the use of two strategies: domain decomposition methods and shared-memory parallelisation with OpenMP. An example of a maintenance process, the tamping, on a portion of railway track with seven sleepers, is simulated.