Dynamic mesh adaptation for unsteady flows within a true parallel environment

Abstract

The complete algorithm from data distribution through the successive adaptation cycles and optimization of the individual meshes to the solution process is executed on a network of processors. Most parallelization strategies using unstructured meshes for solving CFD problems are based implicitly on a so-called “master–slave” paradigm. For efficient handling of dynamic, auto-adaptive mesh strategies required by unsteady flow simulation, up to 500 different meshes may be created for a complete computational cycle. A completely parallel “slaves-only” concept has thus been adopted, whereby the partitioning, mesh adaptation, and all data manipulation take place through the network. The common implementation of flow solvers using unstructured meshes on parallel computers uses domain decomposition techniques for partitioning the computational grid, so that interprocessor communication is reduced to a minimum. Mesh-adaptation algorithms coupled to parallel flow solvers are commonly based upon a master–slave concept, whereby the refinement, derefinement, and optimization techniques are performed in serial on the master. The input file is a global mesh that respects the boundary definitions of the problem, without any subdomain structure to maintain compatibility with former approaches.