NDSL is a middleware for climate and weather modelling developed jointly by NOAA and NASA. The middleware brings together GT4Py (the cartesian flavor), ETH CSCS's stencil DSL, and DaCE, ETH SPCL's data flow framework, both developed for high-performance and portability. On top of those pillars, NDSL deploys a series of optimized APIs for common operations (Halo exchange, domain decomposition, MPI...), a set of bespoke optimizations for the models targeted by the middleware and tools to port existing models.
Historically NDSL was developed to port the FV3 dynamical core on the cube-sphere. Therefore, the middleware ships with ready-to-execute specialization for models based on cube-sphere grid and FV-based model in particular.
Recommended Python is 3.11.x all other dependencies will be pulled during install.
NDSL submodules gt4py and dace to point to vetted versions, use git clone --recurse-submodule.
NDSL is NOT available on pypi. Installation of the package has to be local, via pip install ./NDSL (-e supported). The packages has a few options:
ndsl[test]: installs the test packages (based onpytest)ndsl[develop]: installs tools for development and tests.
Tests are available via:
pytest -x test: running CPU serial tests (GPU as well ifcupyis installed)mpirun -np 6 pytest -x test/mpi: running CPU parallel tests (GPU as well ifcupyis installed)
For CPU backends:
- 3.11.x >= Python < 3.12.x
- Compilers:
- GNU 11.2+
- Libraries:
- Boost headers 1.76+ (no lib installed, just headers)
For GPU backends (the above plus):
- CUDA 11.2+
- Python package:
cupy(latest with proper driver support see install notes)
- Libraries:
- MPI compiled with cuda support
TBD: Code/contribution guideline
TBD: Documentation
Point of Contacts:
- NOAA: Rusty Benson: rusty.benson -at- noaa.gov
- NASA: Florian Deconinck florian.g.deconinck -at- nasa.gov
