search

ACCESS-NRI / accessdev-Trac-archive

Archive accessdev Trac contents as issues
Apache License 2.0
0 stars 0 forks source link

Restarting UM with changed STASH #368

Open penguian opened 5 years ago

penguian commented 5 years ago

| by mrd599@nci.org.au

The UM stores running partial sums of all time averaged diagnosics in the restart file in case the dump frequency doesn't match the mean frequency. A continued run expects to find these fields and gives an error if the list of STASH diagnostics has changed.

With our configuration these aren't actually necessaary because we use monthly means and monthly dumps but we're still affected the check at the start.

With a careful selection of options it's possible to restart as an NRUN without changing the model evolution and allowing changes to the STASH.

Issue migrated from trac:368 at 2024-01-31 18:34:44 +1100

penguian commented 5 years ago

@martin.dix@anu.edu.au set owner to mrd599

penguian commented 5 years ago

@martin.dix@anu.edu.au commented

The diagnostic fields in the dump are not required when starting a new run from step 0 (an NRUN) but are required in a continued run (CRUN). Normally an NRUN does some extra initialisation which also affects reproducibility. However it's possible to turn this off so that the combination of an NRUN + NRUN is the same as a normal NRUN + CRUN. The Met Office rose-stem tests include a test for this as well as a test that an NRUN + CRUN is the same as a single longer run. Note that issues with CABLE initialisation mean that ACCESS-CM2 doesn't pass this latter test at the moment.

The necessary flags for NRUN reproducibility are

[namelist:run_stochastic]
stphseed=2

[namelist:run_cloud]
l_pc2_check_init=.false.

[namelist:run_dyn]
alpha_relax_type=1

In vn11.1 this was replaced by a single switch https://code.metoffice.gov.uk/trac/um/ticket/4040.

The branch of u-bf481, https://code.metoffice.gov.uk/svn/roses-u/b/f/4/8/1/nrun_test tests this by comparing a copy of u-bf481 restarted from 070101 with the original.

There are no such complications with MOM or CICE so they can use the normal WARM_RESTART approach.