Towards a massively parallel simulations with pfem-2

JUAN M. GIMENEZ; NORBERTO M. NIGRO

Mendoza

Simposio; VI Latin American Symposium on High Performance Computing HPCLatAm 2013; 2013

In this work an implementation of the Particle Finite Element Method Two (PFEM-2) based on the distributed-memory architecture is presented. PFEM-2 consists on a material derivative based formulation of the transport equations with an hybrid spatial discretization which uses an eulerian mesh and lagrangian particles.Strategies for the parallelization of eulerian methods based on mesh or lagrangian solutions based on particles which solve fluid-dynamics problems are widely studied separately, however not enough works treat the use of both approaches together. Typical solutions for domain-distribution on eulerian frames are not proper to balance the work-load on some lagrangian stages, then to achieve good performance must be analyzed the use of weighted decomposition to the partitioning.Performance analysis of the implementation running over a beowulf cluster are presented. The weighted partitioning can be used to improve the speed-up when the diffusion of the problem is low, on the other hand, with large diffusion a classical eulerian decomposition is the best choice. However the overall cpu-time required to solve the presented incompressible flow cases with the PFEM-2 method is lower than using classical eulerian solvers, which give auspicious future thinking in solving massively parallel simulations.