Diffusive orthogonal load balancing for Euler–Lagrange simulations - Archive ouverte HAL Accéder directement au contenu
Article Dans Une Revue International Journal for Numerical Methods in Fluids Année : 2023

Diffusive orthogonal load balancing for Euler–Lagrange simulations

Résumé

In the context of unsteady 3D simulations of particle-laden flows, a new double-constraint load balancing strategy for Euler-Lagrange models is proposed. The method relies on an existing Eulerian partitioning and implements a Lagrangian load balancing step, which is orthogonal to the pre-existing Eulerian balancing. This orthogonality property ensures to keep a near-to-ideal Eulerian load balance while strongly improving the distribution of the Lagrangian particles on the processors. The method has been designed to handle large unstructured 3D meshes on complex geometries. Lagrangian performance measurements performed on massively parallel simulations of realistic spray cases show a CPU cost reduction up to 70% compared to the unbalanced case. K E Y W O R D S double-constraint partitioning, Euler-Lagrange, large-eddy simulation, load balancing, particle-laden flows 1 INTRODUCTION Unsteady 3D computational fluid dynamics (CFD) has become a valuable tool for the prediction of particle-laden flows despite its computational power demand. Parallel efficiency in CFD is achieved when the workload is well balanced, that is, by splitting evenly the workload among the cores minimizing idle time. In Eulerian approaches, the load balancing methods often consist in giving each core the same amount of contiguous mesh elements. It is referred to as domain decomposition and achieved by using single-constraint partitioning methods. 1,2 This form of load balancing is highly efficient for pure aerodynamic simulations. 3 This is due to the fact that aerodynamic phenomena are solved using the same set of equations at each mesh element making the cost proportional to the number of elements. However, these methods struggle with multi-physics problems that may introduce space or time irregular workload. For example, in combustion applications, the chemical source terms may be solved iteratively, thus resulting in a large cost in hot regions and negligible cost in cold regions. 4 Dispersed two-phase flows solved with an Euler-Lagrange approach are also a typical example: the carrier phase is continuous but the dispersed phase only exist locally. Particle-laden flows can be handled in two separate ways. The first one, consists in distributing the particles evenly among the cores with no regards to their spatial localization on the mesh. This is known as particle sharing algorithm 5-8 or as dual grid approach in AMReX 9 and is represented on Figure 1A. It ensures a close to perfect load balance since the Eulerian and Lagrangian constraints are fully decoupled: a particle and its containing mesh cell are not owned by the same core. As a consequence, every Euler-Lagrange interaction will require parallel communications which can result in high overheads and thus bad parallel scalability on large core counts. Mirror domain decomposition has been proposed This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited.
Fichier principal
Vignette du fichier
Numerical Methods in Fluids - 2023 - Stock - Diffusive orthogonal load balancing for Euler Lagrange simulations.pdf (4.88 Mo) Télécharger le fichier
Origine : Fichiers éditeurs autorisés sur une archive ouverte

Dates et versions

hal-04307981 , version 1 (26-11-2023)

Licence

Paternité

Identifiants

Citer

Antoine Stock, Ghislain Lartigue, Vincent Moureau. Diffusive orthogonal load balancing for Euler–Lagrange simulations. International Journal for Numerical Methods in Fluids, 2023, 95, pp.1220 - 1239. ⟨10.1002/fld.5191⟩. ⟨hal-04307981⟩
4 Consultations
2 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More