Inefficient domain decomposition when there is a lot of land

[Titouan-Moulin]

Hi,

I'm running a model setup with an elongated channel representing the Southern Ocean (see bathymetry plot). When I use mpijax for domain decomposition along the zonal (x) direction, many of the processes end up being assigned to regions that are entirely land. These processes don't contribute to the actual ocean computation and result in inefficient use of resources. (As I would like to increase again the length of the channel, it is becoming problematic)

Is there a recommended way to improve the decomposition in such a setup ?

Thanks in advance for your help.

[dionhaefner]

We only support a fixed rectangular decomposition at the moment, and I don't see an easy way to change that.

One thing that may be possible to implement with moderate effort would be to allow unequal sizes of the rectangular chunks. For this, we'd need a dedicated API to specify that, and some modifications to get_chunk_size, validate_decomposition, and get_chunk_slices here.