Set OPENMP=true in load script by default

[xylar]

A developer can call ./conda/configure_compass_env.py with --without_openmp to prevent this variable from being set.

closes #347

[xylar]

@matthewhoffman, could you let me know if it's okay for MALI to have OPENMP=true even if you don't actually have OpenMP directives in your code? Is there overhead involved in doing that that you don't want to have?

[xylar]

@mark-petersen, are you okay with this solution? Or do you feel more comfortable with the current behavior in which users have to specify OPENMP=true?

[xylar]

@sbrus89, please take a look if you can.

[xylar]

Testing

I made the build script once without --no_openmp and verified that:

export OPENMP=true

was in the load script. I then used the load script to build MPAS-Ocean and verified that it also showed OPENMP = true at the beginning of the build. Then, I rebuilt the load script, this time with --no_openmp and verified that export OPENMP=true was no longer present. This time, when I build MPAS-Ocean, I see OPENMP = at the beginning of the build.

[mark-petersen]

@xylar I just tested this on badger with:

./conda/configure_compass_env.py --conda /usr/projects/climate/mpeterse/miconda3 --compiler intel \
  --no_openmp --env_name no_openmp
source load_no_openmp_badger_intel_impi.sh

Without the --no_openmp flag I indeed get

echo $OPENMP
true

but if I then use a load script that was created with --no_openmp and source it, OPENMP=true remains. I think OPENMP should be either be blank or (better) set to false, even though MPAS only checks if it is true. Here is my sequence:

echo $OPENMP
    (blank)
source load_dev_compass_1.1.0-alpha.1_badger_intel_impi.sh (built without --no_openmp)
echo $OPENMP
   true
source load_no_openmp_badger_intel_impi.sh (built with --no_openmp)
echo $OPENMP
  true

This is the wrong logic, because one expects a flag like --no_openmp to have a resulting action.

are you okay with this solution? Or do you feel more comfortable with the current behavior in which users have to specify OPENMP=true?

I think this is a nice feature. Users can take advantage of it or set OPENMP themselves, as they prefer.

[xylar]

@mark-petersen

This is the wrong logic, because one expects a flag like --no_openmp to have a resulting action.

Okay, I thought it was intuitive that this would not set OPENMP either way. So if you had previously sourced a load script or set OPENMP yourself, it would be unaffected. I can change the option to --without_openmp and set OPENMP=false.

In general, it is really hard to "deactivate" an compass load script. So it's not guaranteed to be safe to source one load script and then source another. (You will see some weird messages related to Spack if you source the same load script twice -- worth figuring out at some point.)

[xylar]

@mark-petersen, could you see if the modified (and rebased) version works for you and is more intuitive?

It created a sensible load script for me in a local test. That's all I had time to test so far.

[mark-petersen]

Yes, I tested this and it works perfectly now. Thanks!

[matthewhoffman]

I tested the MALI integration suite and this appears to have no effect. In my first test the integration suite took 3 minutes longer than usual after compiling with the OPENMP=true set by the compass env, but I think that was just machine variability, because I tried it a second time and it ran 2 minutes faster than usual. There is no reason to think including this flag would affect anything, because MALI has no OpenMP directives in the MPAS code at present.

[sbrus89]

@xylar, this works as expected for me too.

[xylar]

Thanks everyone! I'll resolve the conflict and then merge.

[xylar]

Rebased and "tested" in then sense of creating a load script with and without OPENMP.