Closed alicebarthel closed 18 hours ago
ERS_Ld3.ne30pg2_r05_IcoswISC30E3r5.WCYCL1850.pm-cpu_intel.allactive-nlmaps
(BFB)PEM_Ln5.T62_oQU240wLI.GMPAS-DIB-IAF-DISMF.pm-cpu_intel
(BFB) since config_use_surface_salinity_monthly_restoring = .false.
in this test. The relevant flags in mpaso_in
are changed as expected. S_1x5_ndays
) GMPAS-IAF
effectively switches the flags in mpaso_in
on Perlmutter. I am running the same short run for master
to do a manual sum check on the logs, to confirm that a standard G-case will not be BFB.Per @jonbob's recommendation, I ran SMS_D_Ld1.TL319_IcoswISC30E3r5.GMPAS-JRA1p5-DIB-PISMF.pm-cpu_intel.mpaso-jra_1958
.
As expected, the G-case with SSS restoring leads to a non BFB change with the baseline (master
as of today):
A total number of 396 fields were compared
of which 103 had non-zero differences
...
diff_test: the two files seem to be DIFFERENT
I am happy with the behavior, this is ready for review. @jonbob let me know if you'd like me to check anything else.
@alicebarthel -- I think that's good. It proves this PR is having the impact you would expect. It might be worth comparing a longer run with this change vs a baseline, just to know what impact it has.
We expect this to be non-BFB for G-case simulations with salinity restoring (which we use for all scientifically validated configurations). So I have added the non-BFB
tag.
I think @alicebarthel is either already away or will be soon. We will probably need someone else to run the longer G-case that @jonbob requested. @vanroekel, do you have someone you can suggest? If not, I can do it.
@xylar -- @alicebarthel did run a long G-case with this change, data is here on lcrc
/lcrc/group/e3sm/ac.abarthel/E3SMv3/20240305..GMPAS-JRA.sssRestorindDt.underSeaIce.TL319_IcoswISC30E3r5.anvil/run
and mpas-analysis here -- https://web.lcrc.anl.gov/public/e3sm/diagnostic_output/ac.abarthel/E3SMv3/GMPAS.sssRestoring_yr1-5/
But digging around a bit, I'm not sure we have a corresponding baseline to compare against. @alicebarthel -- would Filipe's ICOS case be a reasonable baseline?
Yeah, I think we need that baseline.
I just asked Filipe about his ICOS run to see if it can be an appropriate baseline. I'll report back when I hear from him
@vanroekel technically the test you are pointing at does not test the exact code base in question. @xylar i'm still here and was planning on running 2 10-year simulations and posting a MPAS_Analysis comparison as @jonbob instructed. Happy to do that if that is still deemed appropriate.
@alicebarthel, if you're able to do that, that would be wonderful!
The 10-year G-case tests are here:
master baseline /pscratch/sd/a/abarthel/e3sm_scratch/pm-cpu/20240603b.master.GMPAS-JRA.TL319_IcoswISC30E3r5.pm-cpu
PR modifications /pscratch/sd/a/abarthel/e3sm_scratch/pm-cpu/20240603.prtest.GMPAS-JRA.TL319_IcoswISC30E3r5.pm-cpu
As expected, the differences to mpaso_in files are only the 2 relevant config options: config_salinity_restoring_under_sea_ice
and config_surface_salinity_monthly_restoring_compute_interval
The MPAS-Analysis comparison is here. As expected, the first 10 years show an increase in surface/near-surface salinity (SSS, 50m S), thus increasing MLD at high latitudes. Small to no changes to temperature (e.g. SST, 100m T), or sea ice concentration, with small changes likely to indirect effects. Regional changes to high latitudes include increased salinity (e.g. SO, Canada Basin) and cooler temperatures (SO, Canada Basin) (latter likely due to indirect effects). The AMOC timeseries and OHU are unchanged. 1Sv increase to Drake Passage transport.
I did not see anything unexpected or concerning. Have a look if you can. Thanks!
@alicebarthel, did the analysis not succeed? Or do you just need to change permissions? I'm not seeing the index.html
at https://portal.nersc.gov/cfs/e3sm/abarthel/E3SMv3/sssPR.testvsbaseline/
Nice, this looks like it's doing what we want and either improving biases or doing no harm.
Does this only change answers for G-cases ?
@rljacob -- yes, and C-cases as well. And only for grids that have restoring files, so not most of the testing ones like oQU480 and oQU240
Passes:
with expected NML DIFFs.
Regular DIFF for:
as expected. Merged to next
merged to master and expected NML and regular DIFFs blessed, except:
also blessed on pm-cpu integration next with intel
Work done under the ImPACTS SciDac led to the conclusion that it is generally preferable to include SSS restoring under sea ice by default if doing any SSS restoring, and to compute the restoring terms at each time step.
This is due to the fact that the intended linear taper in the sea ice region (by sea ice fraction) may change sign entirely due to the removal of the global mean. This leads to a small negative flux in the high-latitudes, which reinforces the fresh bias.
Work by Luke, Mat, Fred, Alice and others.
[NML] [non-BFB] for C- and G-cases using grids with restoring files
Plots made by Fred:
Of particular concern is the small negative value in the Arctic for example.