Remove initialization of close_ind array for 3d sphere and 3d cartesian location mods.

[hkershaw-brown]

Description:

Removes the initialization of close_ind array to -99 in get_close for 3D location_mods.
See #910 for the cost of this initialization.
The output array of close_ind is only valid from 1:num_close

Fixes issue

fixes #910

Types of changes

• Bug (Performance) fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update

Documentation changes needed?

• My change requires a change to the documentation.
  • I have updated the documentation accordingly.

Tests

Please describe any tests you ran to verify your changes.
MPAS case from profiling in #910, and get_close test harness also given in #910

Checklist for merging

• Updated changelog entry
• Documentation updated
• Update conf.py

Checklist for release

• Merge into main
• Create release from the main branch with appropriate tag
• Delete feature-branch

Testing Datasets

• Dataset needed for testing available upon request
• Dataset download instructions included
• No dataset needed

[mgharamti]

Performance check (ocean models)

Experiment	Application Type	State Vector Size	Assimilated Obs	MPI Processes	Time (sec)
ROMS (original code)	Regional	645,759,164	571	8000	54
ROMS (no initialization of close_ind)	Regional	645,759,164	571	8000	57
MOM6 (original code)	Global	77,760,000	258,669	1280	87
MOM6 (no initialization of close_ind)	Global	77,760,000	258,669	1280	85

Interpretation: Omitting close_ind initialization shows negligible runtime impact in both regional and global ocean configurations. ROMS (regional): +3 s (~5%). MOM6 (global): −2 s (~2%).

[hkershaw-brown]

Thanks! I wouldn’t necessarily expect to see run time improvements in these cases, with most of the time being in IO (obs sequence or writing the large state) -> the things that don't scale to the large core count.
Did you happen to run with output_timestamps = .true? It would be interesting to see if these timings are mostly IO.

The gotcha for this pull request is if someone is relying on testing for the -99 value in their code.

[mgharamti]

output_timestamps

Helen, unfortunately I didn't turn on output_timestamps. I could rerun filter if you think it's important.

[mgharamti]

Thanks! I wouldn’t necessarily expect to see run time improvements in these cases, with most of the time being in IO (obs sequence or writing the large state) -> the things that don't scale to the large core count. Did you happen to run with output_timestamps = .true? It would be interesting to see if these timings are mostly IO.

The gotcha for this pull request is if someone is relying on testing for the -99 value in their code.

I did run with timestamps turned on. Here is a summary:

• MOM6: The localization part accounted for ~35% of the total run time with the original code. For the code without initialization of close_ind, the localization accounted for ~31% of the run time. So, there is a small gain.
• ROMS-Rutgers: Localization in both runs accounted for less than 2% of the run time. IO is the major player in this case.

For future purposes, I placed the 2 ocean DA test cases, with the necessary files, here:

• MOM6: /glade/campaign/cisl/dares/large_dart_tests/ocean/mom6_global
• ROMS-Rutgers: /glade/campaign/cisl/dares/large_dart_tests/ocean/roms_rutgers_regional

[hkershaw-brown]

awesome thanks Moha