Performance comparison and performance tests#214
Performance comparison and performance tests#214mo-alistairp merged 13 commits intoMetOffice:mainfrom
Conversation
23420fa to
1e58ea2
Compare
ab16be2 to
824024e
Compare
|
this change is ready for sci/tech review @iboutle |
mo-alistairp
left a comment
There was a problem hiding this comment.
The changes seem to be sensible, in line with coding standards, and are passing the test suite. Just two minor requests in versions.py -
- could you update the before tag, it isn't matching the prior macro change
- your first name isn't capitalised!
Once those are changed, I'm happy to approve
my first name is always |
mo-alistairp
left a comment
There was a problem hiding this comment.
Happy with these changes now
|
There was an incorrect KGO value on main, which was requested to be fixed in this merge process. Hence, I've also updated the KGO value (rose-stem/site/meto/kgos/gungho_model/azspice/checksum_gungho_model_baroclinic-C48_MG-3panel_azspice_gnu_fast-debug-64bit.txt) despite it not being relevant to this change |
|
Apologies for the late comment - I was only added to the reviewer list on commit. @mo-marqh - there are quite a few configuration files updated here. For this sort of change we normally have to make an update to the jedi configurations that are hosted externally. This is a manual process at the moment and we base the updates on the processed namelist files that are stored in the example directory. While there are quite a few rose configuration updates, I dont see any example namelist configurations. Is that because this change does not effect those files? The jedi_lfric example "canned tests" are not run as part of the test suite so can be missed easily. @DanStoneMO would you mind making an lfric-jedi update for this if it is needed? |
The configuration update only affects jobs which are running the full physical model, i.e. lfric_atm - I'm not sure if that means you need a Jedi update or not. Basing the Jedi configs on the processed namelist files seems quite dangerous to me, because the upgrade macros do not process these, and so they can easily get missed - the upgrade macro is the main method of updating the configuration files. |
@iboutle - thanks. That means we have nothing to do :-) About the namelist - I agree but at the moment we have no other option. We had previously kept these files updated as part of the commit process when being notified of changes to the code/configuration. Ideally we would parse the rose configuration - that is the longer term aim. |
|
Just ran a test to be sure and I think we're all good on the lfric-jedi side |
PR Summary
Sci/Tech Reviewer: @iboutle
Code Reviewer: @mo-alistairp
This branch first adds a performance configuration for a C896 test run triggered multiple times, to enable cross run variability to be analysed. this test is not run regularly, but is derived from the weekly C896 test so should remain safe to use.
Following the adoption of #82 & #176 it is possible to compare timer based output from December's Performance analysis package branch (https://code.metoffice.gov.uk/trac/lfric_apps/browser/main/branches/pkg/Share/r15393_kpi_benchmark) with the current
main.We cannot run with complicated decompositions across panels until #138. But, we can run with 4704 32x32 ranks.
Using a 720 timestep C896 run with 4704 ranks then a run can be analysed with the svn package branch and the main.
Comparing test results for the legacy package branch and the current
mainshows a significant discrepancy in the timestepping time, within predominantlyfast_physicswhich is fixed by setting
["namelist:physics", "configure_segments"], ".true."and allowing values of 0 for all segments (mis-configured in rose metadata)
mo-marqh@1e58ea2
this requires a macro, the results of the macro running on the code base are detailed in:
https://github.com/mo-marqh/lfric_apps/tree/performanceComparePostMacro
mo-marqh@16041eb
PKG branch from svn performance analysis Dec25 *1880s*
This branch, *2006s*
This branch, with configure segments 16 *1807s*
Code Quality Checklist
Testing
trac.log
Test Suite Results - lfric_apps - performanceComparePostMacro/run2
Suite Information
Task Information
✅ succeeded tasks - 1167
Security Considerations
Performance Impact
AI Assistance and Attribution
Documentation
PSyclone Approval
Sci/Tech Review
(Please alert the code reviewer via a tag when you have approved the SR)
Code Review