2426 satellite work Pace
Funded by Vulcan and AI2, the Pace model was a pure python model based on the GT4Py DSL combined with leveraging DaCe for full program optimization. The model is a port combining the FV3 Dynamical Core and the microphysics of NOAA's GFDL. The two papers published on it are:
- Dahm, Johann et al. “Pace v0.2: a Python-based performance-portable atmospheric model.” Geoscientific Model Development (2023): n. pag. PDF
- Ben-Nun, Tal et al. “Productive Performance Engineering for Weather and Climate Modeling with Python.” SC22: International Conference for High Performance Computing, Networking, Storage and Analysis (2022): 1-14. PDF
Below we summarize the findings. For more figures and in-depth explanation read the papers.
The model ran for an idealistic baroclinic wave for 9 days at 1.8km average horizontal resolution. Validation of the model code was done numerically section by section first and was followed up with a scientifically relevant diagnostics.
Here the temperature at 850mb for the reference Fortran model and Pace:
The 9 days simulation was stable and produced rain as expected. Below is a plot of rain, cloud cover and pressure gradients after 9 days.
The 9 days simulation presented above was computed on the 4056 nodes of the Piz Daint supercomputer with 4056 P100 Nvidia GPUs. The throughput was 0.073 SYPD or 28.11 SDPD.
More complete figures for benchmarking on smaller runs are shown below.
Comparison of CPU-Fortran run vs GPU-Python runs relative to node scaling
Performance analysis of representative FV3 (dynamical core) modules
[^1]: Results presented here by a co-authors of the quoted papers