• 28 February 2020
    • preprint
    • Published by HAL in HAL
Abstract
Computational fluid dynamics (CFD) computations often require coupling, interpolation, or projection of some values on one computational domain or portion thereof to another. This is the case both for multi-physics couplings, and for the implementation of various features, such as restarting on a different mesh, boundary condition couplings, and possibly others. In all cases, this requires a mapping, or location stage, where points or other element types are located relative to other points or elements, without a priori knowledge of any neighborhood or connectivity information. As simulations often require highly refined 3D meshes, with computations run on parallel distributed memory systems, using programming paradigms such as MPI, it is often more practical to build mappings in parallel, on the same computational resources as the computation itself. Providing these features as a library callable directly by the computational code further allows avoid costly extra file read/write operations. In this article, we describe different algorithms used in the context of the code_saturne CFD tool, and the associated Parallel Location and Exchange (PLE) library, based on their design goals, performance, and API ease of use and stability. We also provide some elements on actual observed performance, and paths for improvements and extensions for increasingly massive parallelism.
All Related Versions