Update ngpus_per_node for a SCREAM/E3SM GPU run on Derecho

[sjsprecious]

In order to build and run E3SM/SCREAM on Derecho's A100 GPU, we need to set the ngpus_per_node correctly for E3SM/SCREAM.

E3SM/SCREAM does not use the GPU_TYPE and GPU_OFFLOAD options.

[jedwards4b]

I assume that you have tested that this does what you want for both cesm and e3sm on derecho - what about other systems? Should you test there as well? Any way you can think of to move this to XML instead of doing it in python?

[rljacob]

I thought GPU_TYPE and GPU_OFFLOAD were part of the machine config so if its defined for derecho, they will be used. We really want less of these model-specific hooks in the python.

[sjsprecious]

Thanks @jedwards4b and @rljacob for your comments. I see your concerns and I agree that these model- or machine-specific options are not appropriate for the python workflow.

Hmm, if we remove NGPUS_PER_NODE, GPU_OFFLOAD and GPU_TYPE from the python arguments and move them to a XML file, maybe we can ask a user to do something like:

1. create a new case
2. go to the case directory
3. ./xmlchange NGPUS_PER_NODE=xxx
4. ./xmlchange GPU_OFFLOAD=xxx
5. ./xmlchange GPU_TYPE=xxx
6. ./case.setup
7. ./case.build

If this approach is feasible, I just need to make sure that values set in the XML files can be used to set up the GPU flags, GPU node type, etc accordingly in the CMake file later. Is it possible?

[jedwards4b]

Yes, I think that this is a good approach.