If you want to ask questions of the speaker you must reply to this topic. We encourage all speakers and attendees to continue fruitful discussions in replies to this topic for as long as they wish (weeks, years …). This topic will remain available for as long as this site exists.
Wrappers (caps?) have been written for a couple of different models now for ice sheet models Elmer/ice and icepack and ocean models ROMS and FVCOM - would be good to bring in MOM6 too @Hallberg-NOAA.
How easy would it be to get FISOC into the NUOPC framework do you think? Perhaps a bigger chat needed.
Hi @bgf, thanks for your question! I had a brief look at the FISOC code + paper and from what I’ve seen I think it should be pretty straightforward to port this approach to the NUOPC framework with the CMEPs mediator. There seems to a pretty clear mapping between FISOC and NUOPC:
FISOC_ISM_Wrapper_*.f90 corresponds to NUOPC caps (wrappers)
FISOC_coupler.f90 corresponds to esmFldsExchange_*_mod.F90 files (these just specify how different fields are mapped e.g. the ACCESS one we’re working on here)
I think most (maybe all?) of the other functionality exists in CMEPS
It sounds like the caps/wrappers you’re interested in already exist, so it might be a matter of just writing the esmFldsExchange_*_mod.F90, compiling the nuopc caps with CMEPs, and running. And we have build and run scripts that you should be able to use. Happy to chat further, and help with the implementation if you decide to go ahead.
Thanks for presenting to the 21st Century Weather ModSci team last week. I have a couple of questions from your presentation.
Re : Run-time regridding weight calculation? What is the advantage of this? The ESMF regridder can pre-compute weights and store them to disk. So any future regridding operation just loads the weights from disk applies it to the input data. Is run-time regridding faster than reading weights from disk?
Or are you referring to ‘integrated’ regridding as opposed to ‘offline’ regridding? See:
Or - does ‘run-time regridding’ mean the weighting grids are computed at run-time and then stored in memory for the duration of the run?
Yep we’re using “integrated” re-gridding, which is probably slightly more expensive than “offline” but it has the advantage that’s it’s easier to change resolutions and configurations, and track provenance.
The cost of “integrated” may not be worth this trade-off for high resolution runs, but it should be straightforward to support “offline” re-gridding as well if needed.
It might be computationally expensive for production runs at higher resolutions. For example, generating weights for a 0.25° grid using 288 CPU cores for the mediator can take up to 600 seconds, which represents a significant initialisation cost for high-resolution runs.
Aidan
(Aidan Heerdegen, ACCESS-NRI Release Team Lead)
8
Can the regridding weights files be cached? So it’s a one-time only cost?
