Update to use latest OM3 configuration #51
Comments
|
Once COSIMA/access-om3#250 is merged you should be good to go |
This was referenced Feb 18, 2025
Changes from updating MOMFinally have a version with updated MOM and CMEPS working in this branch, following @anton-seaice's advice to start a new branch rebasing the I've compared the water balance and flux conservation with a pre-update control using the notebooks from Martin and Kieran. Comparing the global average statistics, the two runs look similar, with flux errors that are close in magnitude. Both versions were run for 1 year. Water balanceThe following plots show the water balance in the control run
Everything looks ok between the two runs. Fluxes:The following compare the errors in various fluxes using Kieran's notebook: Water balanceNote: I think the large snow errors would be due to masking by the ice Energy balanceIce to ocean fluxesAtmosphere to ice fluxesNote that the above values are calculated using the fluxes in the mediator which have been divided by the ice fraction. If these quantities are conserved, and the ice fractions are properly mapped, then the actual fluxes should be conserved. |
Reproducibility issues after updating MOM and CMEPs versionsAfter updating the MOM6 and CMEPS components in branch cm3-0.4.0, I'm running into reproducibility issues between runs. Initially it seemed like there were no problems, as the firsy runs using a build from cm3-0.4.0 and my less organised branch cm3-update-mom6-cmeps-cdeps-ww3-share matched eachother:
Further runs using The following are from various runs after updating the versions while debugging the CMEPs history issues. The results are not completely random, as some runs match each other, e.g f3474c85caac26bcf0fd94f2073bb3cd appears three times. Last year there were run reproducibility issues (see here) that might not be resolved, however the frequency of variations was lower. I'll try rerunning my control simulation several times to check if there are any differences. |
|
Rerunning the pre-update control several times. It doesn't reproduce 100% of the time, though appears more consistent: I'll try and save timestep coupler diagnostics to see whether differences in the new build first appear in the atm wind stress exports |
|
The following table shows md5sums for mediator atmosphere output for several timesteps using the updated build, in 10 different runs:
Differences in the first timestep are only in
In timestep 2, several more fields (though not all) differ: It takes 3 timesteps for the mediator ice files to show any difference. At timestep 2, all ocean outputs match appart from runs 2 and 9, which differ in the following variables:
```
Differences 20250228-update-repro-ts-1 vs 20250303-update-repro-ts-9, timestep 2
--------------------------------------
time_bnds matches
ocnImp_Fioo_q diff: 10.657534130708429
ocnImp_So_dhdx diff: 6.00873613357248e-14
ocnImp_So_dhdy diff: 1.1019295218568587e-13
ocnImp_So_omask matches
ocnImp_So_s diff: 1.368974240634202e-07
ocnImp_So_t diff: 2.8084195946576074e-07
ocnImp_So_u diff: 2.2717652472145872e-07
ocnImp_So_v diff: 9.606853797701134e-08
ocnImp_lon matches
ocnImp_lat matches
ocnExp_Faxa_rain matches
ocnExp_Faxa_snow matches
ocnExp_Fioi_melth matches
ocnExp_Fioi_meltw matches
ocnExp_Fioi_salt matches
ocnExp_Foxx_evap matches
ocnExp_Foxx_lwnet matches
ocnExp_Foxx_rofi matches
ocnExp_Foxx_rofl matches
ocnExp_Foxx_sen matches
ocnExp_Foxx_swnet_idf matches
ocnExp_Foxx_swnet_idr matches
ocnExp_Foxx_swnet_vdf matches
ocnExp_Foxx_swnet_vdr matches
ocnExp_Foxx_taux diff: 5.232400574930573e-14
ocnExp_Foxx_tauy matches
ocnExp_Sa_pslv matches
ocnExp_Si_ifrac matches
ocnExp_So_duu10n matches
ocnExp_lon matches
ocnExp_lat matches
Med_frac_ocn_ifrac matches
Med_frac_ocn_ifrad matches
Med_frac_ocn_ofrac matches
Med_frac_ocn_ofrad matches
Med_frac_ocn_lon matches
Med_frac_ocn_lat matches
```
These two runs would have received different wind stresses from the atmosphere's first timestep, and so it makes sense that they differ.
|
|
Happy to report that zeroing out the UM exported wind stresses and imported currents (the fields which we're internally regridding in the UM) appears to restore restart determinism: (There was an additional run that crashed with a segfault, which we've been occasionally experiencing). Based on this, I think the non-determinism in the new build is the same as what Martin found with the older build last year, coming from the |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
No description provided.
The text was updated successfully, but these errors were encountered: