Abstract
Accurate remapping strategies are critical components of coupled, multi-mesh climate problems. For high-resolution cases, the computational cost of remapping can become a severe bottleneck. The new remapping interfaces available in the Mesh Oriented datABase (MOAB) are being used in the E3SM v2 climate simulator.
For MPI-parallel solution transfer, source and target meshes must be distributed in a balanced way, while still minimizing the communication needed to compute the mesh intersection for conservative remapping. Typical approaches using independent partitions for component meshes do not necessarily align the geometric domains in a processor.
We present geometric partitioning strategies with Recursive Coordinate Bisection (RCB) in the Zoltan library to accelerate computation of remapping weights at scale. We use a primary-secondary partitioning approach in which we generate a primary partition of one component using RCB and infer the other component’s partition from the primary partition. The resulting partitions assign elements in the geometric neighborhood of both meshes to the same processor, resulting in lower communication overhead. We also compare performance using full 3D partitioning of spherical meshes to using 2D Gnomonic projections for high-resolution model problems. The scalability of the approach, and further performance improvements when using task mapping for varying mesh resolutions and arbitrary process layouts will also be presented.
