There are three regression tests that should be run on a new model as a relatively quick check for reproducibility. In order for these to work properly, the model needs to be compiled with a restart template (in the setup portion of the xml), and the compilation should be done with dynamic memory allocation (so that the scaling test can use the same executable, set within the compile cppDefs of the experiment). The setups for all of the regression tests are in the runTime portion of the xml.
The current GFDL rules allow 70pes to be used in an interactive session, so that regression tests with the coarse model don’t need to be submitted as a batch job. Instead, the runscripts can simply be executed on the ic. This is great for debugging.
basic: a single segment of a a few days to a few months. Usually set to 8 days. If history file output is desired for runs less than 1 month, the frequency in the diag table can be altered – this often helps with debugging.
restarts: to check that the model is smoothly reproducing across segments, set this test to run the same total length as the basic test, but in 2 segments. Usually 2×4 days.
scaling: to check that the model is properly handling the decomposition of the global domain into computational domains, set this test to run the same length as the basic test, but with a different total number of pes. This also requires changing the pe configuration in the coupler, ocean_model, and atmos namelists.
The runscripts are generated with frerun by including the regression test name after the -r switch, ie.
frerun -x model.xml -r basic experimentname
Once the runs are complete, the output can be compared for consistency with the command:
frecheck -x model.xml experimentname