Block-Structured Grids for Coastal Ocean Simulation

  • Schmalfuß, Jonathan (University of Bayreuth)
  • Stahl, Julian (Friedrich-Alexander-Universität Erlangen-Nürn)
  • Grosso, Roberto (Friedrich-Alexander-Universität Erlangen-Nürn)
  • Aizinger, Vadym (University of Bayreuth)

Please login to view abstract download link

For two-dimensional ocean applications such as tidal dynamics simulations or flooding events, triangular unstructured meshes are often the preferred choice. However, numerical computations on unstructured grids often suffer from suboptimal performance due to irregular memory access patterns, whereas structured grids may achieve near-optimal computational efficiency. Block-structured grids (BSGs) offer an alternative: they consist of an unstructured collection of blocks, each containing a (usually) structured grid, and thus promise improved performance while retaining the flexibility needed to represent complex domains. In this work, we demonstrate an extension of the first BSG generator for complex ocean domains to hybrid block-structured grids: namely those consisting of structured and unstructured blocks. By allowing some of the blocks to be unstructured, our BSGs generation techniques have much better control of the local mesh resolution without compromising the element quality and can produce high-quality meshes for ocean domains with highly complex geometry and topography. For a given block size and grid resolution, the hybrid BGS generator automatically minimizes the fraction of elements in unstructured blocks. Using these two user-provided meshing parameters the computational performance can be tuned to various hardware architectures, including current generation CPUs and GPUs. The hybrid BSGs support has been integrated into UTBEST , a shallow water equations solver based on the discontinuous Galerkin method, modernized to support performance portability across heterogeneous hardware platforms via SYCL. The computational performance of hybrid BSGs has been evaluated on a range of realistic scenarios with domains of varying geometric complexity and different grid sizes.