Simulation de problèmes de sédimentation sur machine massivement parallèle

De cnsc
Aller à : navigation, rechercher

Jeudi 10 mai 2012 à 14h00, Petit Amphithéâtre de Mathématiques, UFR Math-Info, 7 rue Descartes, Université de Strasbourg.
Séminaire organisé par Jan Dusek, responsable de l'équipe Mécanique et Environnement de l'Institut de Mécanique des Fluides et des Solides (IMFS) de l'Université de Strasbourg.

Inauguration du Campus Numérique des Systèmes Complexes

avec Paul Bourgine Directeur honoraire du Réseau National des Systèmes Complexes.

Séminaire

Prof. Markus Uhlmann,
Institut für Hydromechanik, Karlruhe Institute of Technology

Massively parallel simulation of sedimentation problems involving many spherical particles and wake effects
The interaction between turbulent flow and suspended solid particles is of relevance in a considerable number of technical applications (e.g. civil and chemical engineering, combustion) as well as natural processes (meteorology, blood flow, ...). Reliable flow data is, however, still scarce due in part to measurement difficulties in these multi-phase flow systems.
Traditionally, suspensions involving large numbers of particles have been described computationally (at best) by a point-particle approximation. However, when the size of the particles is comparable to or larger than the smallest flow scales, this ansatz loses its validity. The same is true when the Reynolds number of the flow around individual particles is not negligibly small. We are performing numerical simulations for finite particle sizes, where the computational particles are larger than the grid and their interface is resolved.
For this purpose we resort to an immersed-boundary technique which allows for an efficient representation of submerged solid bodies in arbitrary motion across a fixed computational mesh. Parallelism is achieved through 3D Cartesian domain decomposition, with mostly nearest-neighbor communication of 'ghost-cell' data and particle-related communication protocols.
In this talk the need for rigorous validation and benchmarking will be stressed, and computational requirements for high-fidelity simulations are discussed. Targeting systems with O(10^5) particles at dilute solid volume fractions, allowing for sufficient small-scale resolution as well as large-scale sampling (box size), meshes with O(10^10) points are necessary. Typical runs are performed on O(10^4) cores of the IBM BlueGene system at Jülich Supercomputing Center where reasonable weak scaling has been demonstrated in tests of up to O(10^5) processor cores.