Closed miaoyinb closed 2 weeks ago
The broken tests are caused by randomness in triangulation. When using different numbers of cores, the generated TRI elements are slightly different. I tried to use average elem quality and size with a reasonable custom tolerances to cover all the cases (using -p 1
as the gold). I did not try -p 5
or -p 7
, which failed in this case.
To fix this, I did a comprehensive test using different -p
numbers and adopt the mean values instead of the -p 1
values as the gold.
@GiudGiud @loganharbour FYI.
Do we not just want to limit the triangulation to a fixed set of cores to run on?
You meant just using the same number in max_parallel
and min_parallel
to limit this? (and maybe have multiple tests with different parallel numbers?)
You meant just using the same number in max_parallel and min_parallel to limit this? (and maybe have multiple tests with different parallel numbers?)
this can be done. Or simply a max_parallel = 3 if for example the results for p = 1, 2, 3 are consistent enough to pass the gold
I m fine with this averaging that you have now. Time will tell if it's robust enough though
I think the motivation of fixing p
is to remove the randomness so that Exodiff can be used instead. In that case, p = 1,2,3.. would all give different exo files.
I think the current average solution is better than a series of Exodiff tests.
Also, my observation is once p>9, things are converged for this specific case and area/quality pp values no longer change.
I think the current average solution is better than a series of Exodiff tests.
If this works, I agree.
Job Documentation on ccfec93 wanted to post the following:
View the site here
This comment will be updated on new commits.
Test failures are unrelated
Job Coverage on ccfec93 wanted to post the following:
dc8a1f | #27629 ccfec9 | ||||
---|---|---|---|---|---|
Total | Total | +/- | New | ||
Rate | 85.09% | 85.09% | +0.00% | - | |
Hits | 103153 | 103154 | +1 | 0 | |
Misses | 18077 | 18076 | -1 | 0 |
Coverage did not change
This comment will be updated on new commits.
closes #27628
Reason
Design
Impact