MultiRegion and NonHydrostaticModel

[francispoulin]

Is it possible to use the MultiRegion code for the NonHydrostaticModel?

[simone-silvestri]

No, there is no implementation of the pressure solver as of now. PR #2795 was attempting to adapt the NonhydrostaticModel for a MultiRegionGrid by introducing a distributed pressure solver. I think it was working at some point, but it has been stale for a very long time so I think trying to bring it back up to speed would be quite difficult.

However, if you just need to use multiple GPUs you can use the Distributed architecture with GPU() as a first argument.
There is an example here that shows how to use distributed computing:
https://github.com/CliMA/Oceananigans.jl/blob/main/validation/distributed_simulations/distributed_nonhydrostatic_turbulence.jl

This is multiple CPUs, to use multiple GPUs, you need to pass the GPU argument instead of CPU in the Distributed constructor.
You can look at help in the REPL for Distributed and Partition to see how to configure the partitioning of the domain

help?> Distributed

and

help?> Partition

[francispoulin]

Thank you @simone-silvestri ! That is very good to know. We are trying out HydrostaticFreeSurfaceModel but will come back to this later, depending on how that goes.

We did have problems with construct-output, but someone is going to start a new discussion on that.

[glwagner]

Just to emphasize --- use Distributed architecture, and don't try to use MultiRegion.