<p>Rapidly changing heterogeneous supercomputer architectures pose a great challenge to many scientific communities trying to leverage the latest technology in high-performance computing. Implementations that simultaneously result in a good performance and developer productivity while keeping the codebase adaptable and well maintainable in the long-term are of high importance. ParFlow, a widely used hydrologic model, achieves these attributes by hiding the architecture-dependent code in preprocessor macros (ParFlow embedded Domain Specific Language, eDSL) and leveraging NVIDIA's Unified Memory technology for memory management. The implementation results in very good weak scaling with up to 26x speedup when using four NVIDIA A100 GPUs per node compared to using the available 48 CPU cores. Good weak scaling is observed using hundreds of nodes on the new JUWELS Booster system at the J&#252;lich Supercomputing Centre, Germany. Furthermore, it is possible to couple ParFlow with other earth system compartment models such as land surface and atmospheric models using the OASIS-MCT coupler library, which handles the data exchange between the different models. The ParFlow GPU implementation is fully compatible with the coupled implementation with little changes to the source code. Moreover, coupled simulations offer interesting load-balancing opportunities for optimal usage of the existing resources. For example, running ParFlow on GPU nodes, and another application component on CPU-only nodes, or efficiently distributing the CPU and GPU resources of a single node between the different application components may result in the best usage of heterogeneous architectures.</p>
Representation of heterogeneity effects in Earth system modeling: Experience from land surface modeling