Request for a slimmed down gcm_run.j

[mathomp4]

Johnny Chang of NAS had this to say in an email:

Also, the PBS script to run GEOSgcm.x is 583 lines long! We would like to get it to 60 lines or less, so that many folks here can easily modify the script without having to understand all the logic in those 583 lines.

Here are the only essential things that we need in the PBS script:

#PBS -l select=.....:aoe=new_OS <--- this is needed to test new_OS

module load new_compiler new_MPI etc. <--- this is needed to test new compiler, new MPI, etc.

../link_input <--- this is necessary to set the input files, etc.

set EXE=/path/to/new/executable <--- this is needed to specify which executable to run

mpiexec -np $NUM_PROCS $EXE <--- this runs the executable and the command can be easily modified to add in profilers (mpiprof, vtune, aps, allinea map, allinea report, etc.)

../validate_and_record <-- this is needed to validate the results and record the performance in our database and store some output files mainly used for validation and timing information

So, you can see that this sample script above contains only 6 essential lines. Some details are hidden in ../link_input and ../validate_and_record, but they should be minimal.

[mathomp4]

I'm not sure how easy this will be. I took a look at gcm_run.j and I tried to see what a minimal script would be for this sort of benchmark testing. I started by removing all the EMIP bits, assumed we'll be after the MERRA2 magic date, and assumed we wouldn't run multiple segments. I also stripped out the bits that tar up restarts, etc. That got me to about 430 lines.

Maybe if linkbcs was a separate script that gcm_setup made, that would remove another 50 lines?

In all I can see it maybe getting to about 250 lines at the smallest, but that would remove a lot checks and functionality.

[mathomp4]

I suppose I should bring in @lltakacs and @tclune for thoughts on this. I'm not sure we can ever get gcm_run.j to be what Johnny wants. It just has too much inertia without rewriting all of GEOSgcm_App. Hmm...