Relax tolerance for truncating small snocan values in CanopyFluxes #2457
Conversation
…tests with it now work
… with LANDFRAC_PFT
…for the checkvarlist, and increase the size of TK_ROOF in the config file to the now needed 10
…mask used for the mesh sys test to integer in ncap2 so that it will work
…is failes at line 181
…h file to compare to, that can be seen that it looks like it's expected to for the fsurdat landmask. A question now is why the identity test needs to subtract 1, when the new test doesn't
…ntical mask rather than -1, because the file already had the -1 taken out
…d grids are being run, there are a few grids that don't test both 1850 and 2000 for example, but most cases are covered
… in the single script
… full node of processors
Add a couple features from multi over to single, so that multi can use the same code.
…le not having to test both, which allows the SystemTest to be valid for just using the single script
… default jobscript name to change
Merge ctsm5.2 branch up to ctsm5.1.dev174
Merge b4bdev 20240321 Conflicts: .git-blame-ignore-revs
…tests for ctsm5.2, that were removed in the update
Update ctsm5.2 branch to ctsm5.1.dev175
The default of 1e-13 was letting through some state updates that look like they should have been truncated to 0: each time step, snocan was decreasing by about 12 orders of magnitude. Addresses ESCOMP#2444
ekluzek
added
the
type: enhancement
There was a problem hiding this comment.
Awesome. Thanks for figuring this out @billsacks.
I'm thinking for visibility this should go to master rather than ctsm5.2 since it's an answer change, and I'd like it to be more obvious than bound up in all the ctsm5.2 changes. But, I'll ask other opinions...
That makes sense to me. I branched off of the 5.2 branch so that I could reproduce the issue and verify that this resolves it, but I agree with your approach here. |
|
I rebased to master, which means it now contains all of ctsm5.2, but that should be cleared up when ctsm5.2.0 is made on master. |
ekluzek
added
the
tag: next
wwieder
removed
the
tag: next
wwieder
moved this from In progress - master/b4b-dev
to Ready to start (or start again)
in Upcoming tags
|
@billsacks, could you give me collaborator permissions, so that I may update this PR to the latest ctsm tag in preparation for merging to master? Submitted on derecho: |
Description of changes
The default of 1e-13 was letting through some state updates that look like they should have been truncated to 0: each time step, snocan was decreasing by about 12 orders of magnitude.
It looks like what's happening is that snocan is periodically getting set to a very small but non-zero value – maybe coming from a small value of liqcan (which probably needs its own truncate_small_values, which might solve this issue, but that's a matter for another time). Then the model seems to be "trying" to evaporate all of this small snocan in CanopyFluxes in the next time step: the fluxes have values that reduce snocan by something like 12 orders of magnitude. But that reduction isn't quite enough to trigger the truncate_small_values call that @slevis-lmwg introduced, since that uses the default relative epsilon of 1e-13. By loosening the tolerance on the epsilon value, we allow the state update to achieve what seems to be its "desired" reduction of the state to exactly 0 in a single time step.
I think the reason this was causing a failure in the tracer test is that eventually the snocan state got close to the smallest representable double precision value: the test failed when snocan reached 4e-296; the tracer in question is set up to be 10 orders of magnitude smaller, so it should have had a value of 4e-306 at this point. But that's getting very close to limit of representation of double precision, which is about 1e-308, so I think the tracer value lost precision and was rounded to 0, while the bulk value was still very slightly non-zero.
My hypothesis for why this just started appearing with the change that @ekluzek found in #2444 (comment) is that the new freezing soil temperatures probably cause veg temperatures to be below freezing, which causes snocan to come into play where it maybe didn't before.
Specific notes
Contributors other than yourself, if any:
CTSM Issues Fixed (include github issue #):
Are answers expected to change (and if so in what way)? YES - roundoff level changes expected for some (maybe many) tests (even those without water tracers)
Any User Interface Changes (namelist or namelist defaults changes)? No
Testing performed, if any:
Ran SMS_D_Ld10.f10_f10_mg37.I2000Clm50BgcCrop.izumi_intel.clm-tracer_consistency, verified that it PASSes now