Offline, hybrid or fully online remappers ?
The Energy, Exascale, Earth Sytem Model (E3SM) is a global climate simulation framework that exposes multiple physics components that represent the atmosphere, ocean, land, river, sea-ice and land-ice physical interactions on long timescales. However, in the well known operator decomposition, it is imperative to resolve the solution coupling between strongly coupled processes e.g., atmosphere and ocean models, in order to control the propagation of numerical approximation errors in global climate simulations. In addition to the temporal treatment, any inaccuracy in the mapping of coupled fields between different discretizations of the sphere can also lead to severe instabilities during the simulation.
In order to tackle these issues, the E3SM global climate simulator has used a verified and validated, hybrid workflow that involves an offline remapping weight generation phase, and seamless transfer of information between component models using the Model Coupling Toolkit (MCT) software infrastructure. Alternatively, we present details about ongoing work to integrate the Mesh Oriented datABase as a new coupler within E3SM, with interfaces to TempestRemap library, for flexibly and scalably computing the remapping weights directly at runtime. This new fully-online workflow maximizes the computational efficiency and increases scientific productivity, without sacrificing discretization accuracy of field data being transferred.
This talk will present details on both the parallel offline remap weight generation with MOAB tools, and the fully online E3SM simulation workflow that utilize the same underlying API. We will showcase strong scaling results of the new MOAB-based coupler for representative problems on large-scale machines, and also provide comparative performance analysis against the existing MCT-based coupler for remap application and data transfer between components. Additionally, the runtime characteristics with a MOAB based workflow will also be analyzed to demonstrate the benefits of using the fully online workflow as the problem resolutions and processor counts are increased.