- Log in to post comments
Hello again,
I have run into a reproducibility issue with the HWRF_for_V16 branch -- namely if I run the mode forecast twice, I do not get the same result. I am judging reproducibility by comparing the *atcfunix files; although I also performed some "diff" commands on the output grib files and they were consistent with the reproducibility errors.
Interestingly, the results are reproducible for the first two cycles of the storm (Humberto 2019, AL09), but the third cycle I am running is where the differences occur.
Also, interestingly, I have run these first three cycles 8 times, and the results for the third cycle oscillate between two outcomes -- that is that 4/8 forecasts are identical, and the other 4/8 are identical.
I have reproduced the issue with scrubbing turned off:
The *atcfunix files are saved here for two of different experiments:
/scratch1/BMC/qosap/Peter.Marinescu/HWRF_AeoTest/noscrub/09L16_3c
/scratch1/BMC/qosap/Peter.Marinescu/HWRF_AeoTest/noscrub/09L16_3d
The scrub directories are here:
/scratch1/BMC/qosap/Peter.Marinescu/HWRF_AeoTest/scrub/09L16_3c
/scratch1/BMC/qosap/Peter.Marinescu/HWRF_AeoTest/scrub/09L16_3d
I also have the scrub directory data for one additional experiment that is identical to experiment 09L16_3c; it is experiment 09L16_3b, in case that is helpful as well.
The code base I am using is saved here: /scratch1/BMC/qosap/Peter.Marinescu/HWRF_AeoTest/H_V16
Also, note that I ran the forecasts for shorter (24 hours) to save on computational time for these experiments, but since the reproducibility issues start at time 0, I hope this will not be an issue.
Also, note that I also run 5 cycles of an Eastern Pacific Storm (Kiko 2019, EP13) 3 separate times and the results there were identical for these three runs.
I know some others have also had some reproducibility issues with various versions of HWRF on different machines, so hopefully these output can help figure out part of the problem.
Would you be able to help me resolve/understand this reproducibility issue?
Thank you for your time.
Best,
Peter