Presentation
GPU Code Adaptations to the ACCORD Atmospheric Model within the Destination Earth On-Demand Extremes Digital Twin Project
Presenter
DescriptionThe Destination Earth On-Demand Extremes Digital Twin project will develop on-demand configurable digital twin engines (DTEs) for forecasting of extreme weather phenomena at the sub-km scale. The DTEs will be based on codes developed by the ACCORD NWP consortium. The introduction of high performance GPU-based accelerators greatly increases available computing resources, for instance through the EuroHPC Joint Undertaking, and GPU adaptation represents an important aspect of the DTE development. The approach taken here closely follows the one of the ECMWF. The codes are transformed from CPU-targeted to GPU-targeted code using source-to-source translation tools. This requires code refactoring that includes the introduction of smart data structures for fields storage, code modularization, memory usage improvement, coding standard enforcement, multiple loops re-arrangement, … While the development and testing of the source-to-source translators relies on stand-alone pieces of the code, the adaptation itself will be based on a top down approach that follows the structure of the existing model organization to define the granularity of the kernels to be offloaded to the GPUs. This will be explained in detail in this presentation and first results will be presented.
TimeTuesday, June 2717:00 - 17:30 CEST
LocationDischma
SessionMS4A - Technical Challenges and Opportunities for Digital Twins of the Earth System (Part 1/3)
Session Chair
Event Type
Minisymposium
Climate, Weather and Earth Sciences