Fix standalone MPAS builds under OSX

[xylar]

This merge switches to the flag for including the standard c++ library (needed by SCORPIO) under OSX to -lc++, not -lstdc++.

OSX is also the most common platform where we use gfortran and clang compilers. OpenMPI on OSX does not support flags -cc=clang and -cxx=clang++. These flags also do not seem to be needed.

[BFB]

[xylar]

This PR needs to be rebased after #4362 gets merged.

[xylar]

The gfortran-clang target was added to the Makefile in MPAS-Dev/MPAS#1087. There is no description of why the -cc=clang and -cxx=clang++ flags were needed, so I feel reasonably comfortable removing them.

[mark-petersen]

With #4362 and this merged on top, MPAS-Ocean stand alone compiles with gnu debug and optimized on grizzly and passes nightly regression suite (other than those caused by other issues).

[xylar]

@mark-petersen, do you have an OSX machine you can test on? I don't really, just a MacInCloud for a few hours a day, which makes this tricky for me. It would be nice to have the ocean test suite tested on OSX, not just Grizzly.

[mark-petersen]

On my local laptop I successfully, on compass master:

./conda/configure_compass_env.py --conda ~/miconda3
source load_dev_compass_1.0.0.sh

then on this branch:

make gfortran-clang USE_PIO2=true DEBUG=true

also tried USE_PIO2=false. But it fails on the PIO test. Do I need to compile my own PIO library? I thought it might come with the conda environment.

[xylar]

@mark-petersen, you need to create the conda environment with --mpi=mpich (preferred) or --mpi=openmpi (also fine). I just made a PR to try to clear this up in the documentation because @vanroekel had the same confusion: MPAS-Dev/compass#178

[matthewhoffman]

I compiled on my Mac and encountered no issues. These changes look appropriate.

[akturner]

Tested MPAS-Seaice and everything worked fine

[xylar]

Thanks very much @matthewhoffman and @akturner!

[mark-petersen]

@xylar it's compiling with --mpi=mpich on conda. I'll report with the results of the nightly suite soon, but it looks like it is working!

[mark-petersen]

@xylar, this is great. It almost works. So far these are successful:

./conda/configure_compass_env.py --conda ~/miconda3 --mpi=mpich
source load_dev_compass_1.0.0_mpich.sh
cd E3SM-Project/components/mpas-ocean
make gfortran-clang USE_PIO2=true DEBUG=true
cd ../../..
compass suite -s -c ocean -t nightly -w ~/runs/n/210714_test_OSX -f ocean.cfg

where I've set ocean_database_root path in ocean.cfg. Then

cd ~/runs/n/210714_test_OSX
compass run

And it tries to run through the suite, but it fails all except the mesh generation.

My problem seems to be MPI on my laptop. The simple version is this:

mpirun -n 1 ./ocean_model
Fatal error in MPI_Init: Other MPI error, error stack:
MPIR_Init_thread(159)..............:
MPID_Init(164).....................: channel initialization failed
MPIDI_CH3_Init(95).................:
MPID_nem_init(314).................:
MPID_nem_tcp_init(173).............:
MPID_nem_tcp_get_business_card(395):
GetSockInterfaceAddr(369)..........: gethostbyname failed, pn1808496.lanl.gov (errno 0)

have you seen that one before? That is probably some local problem. You should go ahead and merge this PR anyway.

[xylar]

@mark-petersen, the error message suggests to me that it might be a proxy issue. I'm not good at debugging issues with LANL proxy settings but @vanroekel may have some suggestions, since he has just been struggling with that on his new machine.

[xylar]

I was able to test this with OSX in combination with MPAS-Dev/compass#171 and it worked fine for me. The Mac I was testing on didn't have 8 cores to work with, causing some decomposition tests to fail. Other tests passed.

[mark-petersen]

From a github issue pmodels/mpich#4710 (comment) I tried this:

export MPICH_INTERFACE_HOSTNAME=localhost
mpirun -host localhost -n 4./ocean_model

and it worked. I don't know why this is a problem for me and not others, but it must be either the OS version (mine is old at 10.13) or some network setup. It was the same with VPN on or off.

Regardless, please go ahead an merge this PR. It obviously works. Thanks for doing this, @xylar.

[jonbob]

merged to next

[jonbob]

merged to master

[xylar]

Awesome, thanks @jonbob!