Add compile and platform yamls to compile model with FRE/2025.01 #131
Conversation
uwagura
changed the title
github-actions
bot
added
pass_CEFI_MOM6_RT_container
and removed
CEFI_MOM6_RT_container
labels
|
@uwagura, thanks for opening this PR—awesome work! Although it is still a draft, I have two quick thoughts: Would you mind adding a README file under yaml with brief instructions on how to perform bare-metal/container builds and how to run containerized builds using When you get a chance, could you also share the archive paths for both your bare-metal and containerized NWA-12 runs? |
|
@yichengt900 , I've added a README to the yaml directory, and added instructions to the main yaml on how to use You can find the history files and the pp files from this experiment here: @andrew-c-ross , you don't have to try the containerized run if you don't have the time, but I'd love to get your thoughts on some of the plots! |
There was a problem hiding this comment.
@uwagura, thank you for opening this PR. Overall, it looks good, and I only have a few comments regarding your instructions and some other minor things.
I successfully built the containerized model by following the modified instructions with fre/2025.01. Additionally, I conducted several 2-day regression test runs using both the bare-metal build and the containerized build, and the restarts were identical.
Furthermore, I compared the model results from your fully containerized run with my 2024/12 retrospective runs. Based on my analysis, the model results also appear to be identical. Relevant figures can be found here:
2024/12 bare-metal runs: /nbhome/ynt/results/CEFI_NWA12_COBALT_V1_repro/analysis/regional-mom6-diagnostics_20240418/physics/figures_2024_12
Fully containerized runs: /nbhome/ynt/results/CEFI_NWA12_COBALT_V1_repro/analysis/regional-mom6-diagnostics_20240418/physics/figures_container
Given these findings, I’m unsure what you meant by your original comment in your PR: The fully containerized run does not quite reproduce the bare-metal run results, but the differences mostly seem minimal. Are you referring to differences compared to the original NWA paper results?
|
Thanks for your review @yichengt900 , I'll make the changes you suggested soon. Regarding my comment, about the results not being quite the same, I meant that the results from the containerized run were not quite identical to the plots I produced with the data in |
@uwagura, thank you for the clarification. It seems the data you pointed to is likely the same one Andrew used for his paper. Unfortunately, it’s probably not possible to fully reproduce those results due to updates in libraries and compilers on C5, as well as changes in the model codes. Our main goal here is to confirm that the containerized version produces the same results as the bare-metal build. One good thing is even with those differences, the new model results are consistent and not worse than the ones in the paper. |
yamls/NWA12/README.md
Outdated
| @@ -68,45 +75,45 @@ frerun -x CEFI_NWA12_cobalt.xml -p ncrc5.intel23 -t prod CEFI_NWA12_COBALT_V1 -- | |||
| Like `fre make`, postprocessing is split into several subcommands to improve modularity. On PPAN, load the appropriate module: | |||
| ``` | |||
| module load fre/2025.01 | |||
| ``` | |||
| `` | |||
There was a problem hiding this comment.
@uwagura, everything looks good now except for a missing `. I will approve it after you change that. Thanks again for your contribution
yichengt900
added
CEFI_MOM6_RT_container
and removed
pass_CEFI_MOM6_RT_container
labels
github-actions
bot
added
pass_CEFI_MOM6_RT_container
and removed
CEFI_MOM6_RT_container
labels
This PR adds compile and platform yamls to allow users to compile the NWA configuration with
Fre/2025.01, and optionally produce containers to run the model as well. Both the containers and the bare metal models compile with these yamls can then be run usingfre/bronx-23.The full containerized does not quite reproduce the bare-metal run results, but the differences mostly seem minimal: