(This is a follow-on from #4859 .)
There have been a couple of bugs in CESM recently (at least on branches... as far as I'm aware, none that have made it to the main branch) where writing a restart file caused answer changes. Based on discussion in today's CSEG meeting, we'd like to change the behavior of the ERS test for CESM so that it does what the ERS2 test used to do:
- Case 1 runs all the way through without writing any restart file
- Case 2 starts from the beginning, writes a restart file mid-way through the full length and stops, then restarts from this restart file to finish the full run length
This will involve reinstating the functionality of the ERS2 test that I removed in #4863 - though as noted in #4859 , the ERS test has diverged, e.g., in adopting a more robust calculation of rest_n, so we should merge the functionalities of the two tests.
This will lead to a more robust restart test that better mimics the production workflow and will catch issues with answer changes emerging from the writing of restart files in addition to catching problems with actually restarting from restart files. The downside is extra cost: an extra model initialization, and somewhat more than 1/3 additional run time.
CESM's preference is to migrate all of our ERS tests to this new, more robust restart test. So from our perspective, it would work well to change the behavior of the existing ERS test. However, if E3SM and/or NorESM prefer to stick with the current ERS behavior, we can instead create a new test type and change our test suites to use this new test.
Thoughts? @jasonb5 @jgfouca @rljacob @mvertens
(This is a follow-on from #4859 .)
There have been a couple of bugs in CESM recently (at least on branches... as far as I'm aware, none that have made it to the main branch) where writing a restart file caused answer changes. Based on discussion in today's CSEG meeting, we'd like to change the behavior of the ERS test for CESM so that it does what the ERS2 test used to do:
This will involve reinstating the functionality of the ERS2 test that I removed in #4863 - though as noted in #4859 , the ERS test has diverged, e.g., in adopting a more robust calculation of rest_n, so we should merge the functionalities of the two tests.
This will lead to a more robust restart test that better mimics the production workflow and will catch issues with answer changes emerging from the writing of restart files in addition to catching problems with actually restarting from restart files. The downside is extra cost: an extra model initialization, and somewhat more than 1/3 additional run time.
CESM's preference is to migrate all of our ERS tests to this new, more robust restart test. So from our perspective, it would work well to change the behavior of the existing ERS test. However, if E3SM and/or NorESM prefer to stick with the current ERS behavior, we can instead create a new test type and change our test suites to use this new test.
Thoughts? @jasonb5 @jgfouca @rljacob @mvertens