SMT GEOS DiscoverBuild

Build & Run GEOS on Discover

Deploying SMT stack via modules

On Discover, we have a pre-built stack that is accessible via module load.

• module use -a <Path to SMT-Nebulae Repo>/sw_stack/discover/sles15/modulefiles adds the various stack versions to your module list
• ml SMTStack/YYYY.MM.PP to load the targeted version of the stack

For development, a module SMTStack/YYYY.MM.PP-no-venv is available that doesn't pull ndsl or any other python packages.

Versions of all the software in the stack are present in /discover/nobackup/projects/geosongpu/sw_sles15/HISTORY.md

Build & Run GEOS

• Git clone the root of GEOS then using mepo, you clone the rest of the components which pull on the components.yaml at the root of GEOSgcm
  • There are two sources for a component: the default and the develop source
  • v11.5.2 is our baseline GEOS version (or tag), but we have dsl/develop branch where needed
  • We do not use develop for GEOSgcm_App or cmake since those have been setup for OpenACC but are not up-to-date for v11.5.2

git clone -b dsl/develop [email protected]:GEOS-ESM/GEOSgcm.git geos
cd geos
mepo clone --partial blobless
mepo develop env GEOSgcm GEOSgcm_GridComp FVdycoreCubed_GridComp pyFV3

• Use CMake to set up the GEOS makefile that uses the SMT stack.
• Then make install. Grab a coffee (or 2)
  • We override the compiler with their mpi counterpart to make sure libmpi is pulled properly. CMake should deal with that, but there's some failure floating around.

Single script to execute both cmake & make to install GEOS:

module use -a SW_STACK/modulefiles
ml SMTStack/2024.04.00

mkdir -p build
cd build
export TMP=GEOS_DIR/geos/build/tmp
export TMPDIR=$TMP
export TEMP=$TMP
mkdir $TMP
echo $TMP

export FC=mpif90
export CC=mpicc
export CXX=mpic++

cmake .. -DBASEDIR=$BASEDIR/Linux \
         -DCMAKE_Fortran_COMPILER=mpif90 \
         -DBUILD_PYFV3_INTERFACE=ON \
         -DCMAKE_INSTALL_PREFIX=../install \
         -DPython3_EXECUTABLE=`which python3`

make -j48 install

• Matt T. prepared us some GEOS-FP like data. Copy them from /discover/nobackup/projects/geosongpu/geos_data/geos-fp/stock-v11.5.2-1day-GNU-NH-GDATA-RRTMGP-GFDL-GF2020-OPS-cX-L137 where the X in cX is set as the resolution from {24, 180, 720}
• We need to set a soft link to the GEOSgcm.x executable from the GEOS install/bin directory, e.g.

cd <PATH_TO>/stock-v11.5.2-1day-GNU-NH-GDATA-RRTMGP-GFDL-GF2020-OPS-cX-L137
ln -s <PATH_TO_geos>/install/bin/GEOSgcm.x GEOSgcm.x

• We then need to edit the gcm_run.j to make it local to our experience directory and have the job parameters match our account and the GPU partition. The following assumes that we're running c24 GEOS-FP experiment.

gcm_run.j

#SBATCH --time=0:30:00
- #SBATCH --nodes=1 --ntasks-per-node=126
+ #SBATCH --nodes=1 --ntasks-per-node=48
- #SBATCH --job-name=stock-v11.5.2-1day-GNU-NH-GDATA-RRTMGP-GFDL-GF2020-OPS-c24-L137_RUN
+ #SBATCH --job-name=NDSL-geos-fp
- #SBATCH --constraint=mil
+ #SBATCH --constraint=rome
- #SBATCH --account=g0620
+ #SBATCH --account=j1013
- #SBATCH --partition=preops
+ #SBATCH --partition=gpu_a100
- #SBATCH --qos=benchmark
+ #SBATCH --qos=4n_a100
- #SBATCH --mail-type=ALL

setenv SITE             NCCS
- setenv GEOSDIR          /gpfsm/dnb05/projects/p50/Models/GEOSgcm-v11.5.2-GNU-SLES15/GEOSgcm/install-Release
+ setenv GEOSDIR          PATH_TO_GEOS/install/
- setenv GEOSBIN          /gpfsm/dnb05/projects/p50/Models/GEOSgcm-v11.5.2-GNU-SLES15/GEOSgcm/install-Release/bin
+ setenv GEOSBIN          PATH_TO_GEOS/install/bin
- setenv GEOSETC          /gpfsm/dnb05/projects/p50/Models/GEOSgcm-v11.5.2-GNU-SLES15/GEOSgcm/install-Release/etc
+ setenv GEOSETC          PATH_TO_GEOS/install/etc
- setenv GEOSUTIL         /gpfsm/dnb05/projects/p50/Models/GEOSgcm-v11.5.2-GNU-SLES15/GEOSgcm/install-Release
+ setenv GEOSUTIL         PATH_TO_GEOS/install/

setenv  EXPID   stock-v11.5.2-1day-GNU-NH-GDATA-RRTMGP-GFDL-GF2020-OPS-c24-L137
- setenv  EXPDIR  /discover/nobackup/mathomp4/Experiments/RunsForFlorian/stock-v11.5.2-1day-GNU-NH-GDATA-RRTMGP-GFDL-GF2020-OPS-c24-L137
+ setenv  EXPDIR  PATH_TO_EXP_DIR/stock-v11.5.2-1day-GNU-NH-GDATA-RRTMGP-GFDL-GF2020-OPS-c24-L137
- setenv  HOMDIR  /discover/nobackup/mathomp4/Experiments/RunsForFlorian/stock-v11.5.2-1day-GNU-NH-GDATA-RRTMGP-GFDL-GF2020-OPS-c24-L137
+ setenv  HOMDIR  PATH_TO_EXP_DIR/stock-v11.5.2-1day-GNU-NH-GDATA-RRTMGP-GFDL-GF2020-OPS-c24-L137

• The pipeline of GEOSgcm calculates the numbers of process required based on the NX/NY given in AGCM.rc. Previously we modified the sbatch parameters to be --ntasks-per-node=48 to fit within a Rome node. Therefore, we need to adapt the AGCM.rc

AGCM.rc

# Atmospheric Model Configuration Parameters
# ------------------------------------------
- NX: 4
+ NX: 2
- NY: 24
+ NY: 12

• Running the simulation on the compute node is now as simple as a sbatch gcm_run.j. The environment will be loaded via PATH_TO_GEOS/@env/g5_modules. Logged will be dumped in slurm-XXXXX.out

• To run the pyFV3 version, modification to the AGCM.rc and gcm_run.j. This will run the numpy backend.

AGCM.rc

###########################################################
# dynamics options
# ----------------------------------------
DYCORE: FV3
 FV3_CONFIG: HWT
+ RUN_PYFV3: 1
AdvCore_Advection: 0

gcm_run.j

setenv RUN_CMD "$GEOSBIN/esma_mpirun -np "

setenv GCMVER `cat $GEOSETC/.AGCM_VERSION`
echo   VERSION: $GCMVER

+setenv FV3_DACEMODE           Python
+setenv GEOS_PYFV3_BACKEND     numpy
+setenv PACE_CONSTANTS         GEOS
+setenv PACE_FLOAT_PRECISION   32
+setenv PYTHONOPTIMIZE         1
+setenv PACE_LOGLEVEL          Debug

+setenv FVCOMP_DIR PATH_TO_GEOS/src/Components/@GEOSgcm_GridComp/GEOSagcm_GridComp/GEOSsuperdyn_GridComp/@FVdycoreCubed_GridComp/
+setenv PYTHONPATH $FVCOMP_DIR/python/interface:$FVCOMP_DIR/python/@pyFV3


#######################################################################
#             Experiment Specific Environment Variables
######################################################################