ROMSOC: a regional atmosphere–ocean coupled model for CPU–GPU hybrid system architectures

Abstract

Recent years have seen significant efforts to refine the horizontal resolutions of global and regional climate models to the kilometer scale. This refinement aims to better resolve atmospheric and oceanic mesoscale processes, thereby improving the fidelity of simulations. However, these high-resolution simulations are computationally demanding, often necessitating trade-offs between resolution and simulated timescale. A key challenge is that many existing models are designed to run on central processing units (CPUs) alone, limiting their ability to leverage the full computational power of modern supercomputers, which feature hybrid architectures with both CPUs and graphics processing units (GPUs).

In this study, we introduce the newest model version of ROMSOC, a recently developed regional coupled atmosphere–ocean model. This new model version integrates the Regional Oceanic Modeling System (ROMS) in its original CPU-based configuration with the Consortium for Small-Scale Modeling (COSMO) model (v5.12), which can utilize GPU accelerators on heterogeneous system architectures. The combination efficiently exploits the hybrid CPU–GPU architecture of the Piz Daint supercomputer at the Swiss National Supercomputing Centre (CSCS), achieving a speed-up of up to 6 times compared to a CPU-only version with the same number of nodes.

We evaluated the model using a configuration focused on the northeast Pacific, where ROMS covers the entire Pacific Ocean with a telescopic grid, providing full-ocean mesoscale-resolving refinement in the California Current System (CalCS; 4 km resolution). Meanwhile, COSMO covers most of the northeast Pacific at a 7 km resolution. This configuration was run in hindcast mode for the years 2010–2021, examining the roles of different modes of air–sea coupling at the mesoscale, including thermodynamical coupling (associated with heat fluxes) and mechanical coupling (associated with wind stress and surface ocean currents).

Our evaluation indicates that the hindcast generally agrees well with observations and reanalyses. Notably, large-scale sea surface temperature (SST) patterns and coastal upwelling are well-represented, but SSTs show a small cold bias, resulting from wind forcing that is too strong and biases in the radiative forcing. Additionally, the coupled model exhibits a deeper and more realistic simulation of the ocean mixed-layer depth with a more pronounced seasonal cycle, driven by the enhanced wind-driven mixing. On the other hand, our ROMSOC simulations reveal a negative cloud cover bias and related biases in surface radiative fluxes off the coast of southern California, a common issue in climate models.