B4b reproducibility #134
B4b reproducibility #134
Conversation
initialize variable for TYPE SGRD update matrix_datarmor remove unexpected NOPA switch
see regtests : mww3_test03/work_PR2_UQ_MPI_e mww3_test_03/work_PR3_UQ_MPI_e_c
|
Hi @aliabdolali this is great progess, thanks to @mickaelaccensi . @aliabdolali have you run twice the matrix for the PR branch and compared it to itself to check it b4b is achieved with your machine given the current cmplr.env settings? |
There was a problem hiding this comment.
This PR was originated from @mickaelaccensi and was verified by @aliabdolali and myself. After talking to Ali, and having his commitment that the matrix output from the subsequent runs made with the B4b branch showing the tests are now b4b, we decided to push this through ahead of the UKMO PR because it potentially solves the b4b issues that were plaguing our regtests. @ukmo-ccbunney please update your develop branch and rerun the retest matrix for you PR #132 . We will get this PR #134 first to allow that additional testing which will be very helpful.
|
I'm afraid to report that I am still getting differences in some of the mww3_test_03 regtests on the develop branch. I ran the regtests twice on the on the same code (NOAAs develop branch). Any suggestions? ********************* non-identical cases **************************** mww3_test_03/./work_PR3_UQ_MPI_d2 (12 files differ) |
|
@ukmo-ccbunney how does this compare to the previous output from matrix? Did it solve b4b in many cases, or the results are the same? The cases you list have load balancing (communicator fraction for each grid) set separately for each grid within a group. Those cases are not expected to be b4b with @mickaelaccensi fix. The target here is other cases where the communicator fraction is not set. Eg, grdset d: grdset d2: |
|
@ajhenrique Ah - I hadn't appreciated that it was only fixing those tests not using load balancing. However, I still have differences with the tests using I am going to rerun my mww3_test_03 regtests and compare again. |
|
@ajhenrique @mickaelaccensi - I am getting similar results to my regtest before the b4b fix. From the above comments, this would be expeced for In all of my tests, it is only the |
Hi @ajhenrique @mickaelaccensi @ukmo-ccbunney
The regression test went well, the comparison with develop branch/ NOAA repo is attached.
comp.zip
Thanks Mickael for the great work you've done.
AA