Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

remove unneeded fields for dev/ufs-weather-model #1317

Draft
wants to merge 40 commits into
base: dev/ufs-weather-model
Choose a base branch
from

Conversation

DeniseWorthen
Copy link
Contributor

@DeniseWorthen DeniseWorthen commented Nov 8, 2024

Pull Request Summary

Removes un-needed allocations and time-interpolation of import fields when CMEPS is used. Note this PR is built on top of the PIO PR. Changes are minimal when compared to that PR.

Description

Introduces a logical flag (use_cmeps) which is false by default. When true, arrays that are not needed at the global domain resolution are minimally allocated and time-interpolation w/in the update routines is skipped.

To test the impact, three cases were run using the same DATM-S2SW configuration used for testing the PIO work on Gaea. Each case was run out 48 hours on 816 WAV tasks, with WW3 restarts written every 2 hours.

The ProfileMemory configuration flag was turned on and memory statistics for the WW3 PETs were pulled out for every 50th PE (18 total traces). Three values were plotted: VmPeak, VmRSS and CommittedPercent. The first is the max resident size, the second is the mean resident size and the third is (as defined here) an interesting number I happened to notice in the output. I'm not sure of this last value, but it seems relevant---for example, if I use too few tasks for WAV, the model crashes and I also see values like CommittedPercent: 166.939 %.

Each of the following figures show the current dev/ufs-weather-model branch (top), the feature/pio4ww3 branch (middle) and this feature branch (bottom):

VmPeak:

vmpeak plot

VmRss:

Note here that the highest trace on the top figure shows the values for the last (naprst) PET.

vmrss plot

CommittedPercent:

Note here that the lowest trace on all figures is the first PET.

pct plot

My interpretation of these results is that a) the PIO feature itself reduces memory usage by reducing the high-memory cost of the naprst task and b) this feature has a measurable impact on the memory footprint of the model. However, no consistent impact on model execution time was found in my limited testing.

Issue(s) addressed

Commit Message

  • minimally allocate unneeded arrays when CMEPS is in use. Skip time-interpolation for imported fields.

Check list

Testing

  • How were these changes tested?
  • Are the changes covered by regression tests? (If not, why? Do new tests need to be added?)
  • Have the matrix regression tests been run (if yes, please note HPC and compiler)?
  • Please indicate the expected changes in the regression test output, (Note the list of known non-identical tests.)
  • Please provide the summary output of matrix.comp (matrix.Diff.txt, matrixCompFull.txt and matrixCompSummary.txt):

I ran this PR against a baseline created with PR #1303. All baselines passed.

DeniseWorthen and others added 30 commits August 28, 2024 08:06
* remove wav_grdout routine, now moved into wav_history_mod
* remove more cap stuff from w3iorsmd. only ww3 native filenaming
is possible w/ binary restarts
* remove ifdef w3_ascii from w3wavemd, since the ascii commit is not
yet present in mesh cap branch
* clean up config variable logic for filenaming
* nrqrs can be non-zero if also using the restart from binary
option
* flpart needs to be set for either historync or not
* move restart and history nc blocks outside of ww3 time testing
block.
* add log print messages for reading and writing restarts
* get logging working correctly for ufs
* fix noclobber variable and file name in wav_history
* clarify some comments
* make binary history files match when restartnc=true
* rework wav_restart_mod, which originally was designed to be able
to read and write restarts for testing purposes from inside wav_comp_nuopc.
* verboselog is true by default, but can be set false by config
* the header text for ww3 logging as it steps through time is now
turned off in w3init and placed into the mesh cap. this allows the
mesh cap to order the logging correctly
* move block where addrstflds was set to before call to w3init
since restarts are read in w3init
* ensure that if nml file lacks a specification of extra fields,
the default value of "unset" will not be returned as a field name
* only ice is added for now
* tab cleanup in w3grid
* need to send explicit array bounds for ice array since
it is 0:nsea
* all baselines b4b against f9531d0
* intialize floutg and floutg2 which are can be unintialized when
waves are in slow loop and historync is true
* local arrays dimension nx or greater are locally allocated
* deallocations added, even though not required for user clarity
* global_input and _output can be allocatable
DeniseWorthen and others added 10 commits October 8, 2024 08:30
* tested by using nco to alter the nth,nk values in a restart
file. The model correctly aborted after finding nth did not
match the expected value
* delay the deallocation of gindex arrays until after the diagnose_mesh
routine is called. This only impacts cases where dbug_flag>5, compiled
with gnu+debug
* minimally allocate fields used in w3updtmd when
using cmeps
* use fields provided by cmeps w/o fake time-interpolation
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

1 participant