tests_on_cluster.bash

#!/bin/bash
#
#SBATCH --job-name=oggm_tests
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=4

# Abort whenever a single step fails. Without this, bash will just continue on errors.
set -e

# We create a sub-directory for this job to store its runtime data at.
WORKDIR="/work/$USER/oggm_tests_$SLURM_JOB_ID"
mkdir -p "$WORKDIR"
cd "$WORKDIR"
echo "Workdir for this run: $WORKDIR"

# Export the WORKDIR as environment variable so our benchmark script can use it to find its working directory.
export WORKDIR

# Export the global read-only download-cache-dir
export OGGM_DOWNLOAD_CACHE=/home/data/download
export OGGM_DOWNLOAD_CACHE_RO=1

# All commands in the EOF block run inside of the container
# Adjust container version to your needs, they are guaranteed to never change after their respective day has passed.
srun -n 1 -c "${SLURM_JOB_CPUS_PER_NODE}" singularity exec docker://oggm/oggm:20181110 bash -s <<EOF
  set -e

  # Setup a fake home dir inside of our workdir, so we don't clutter the actual shared homedir with potentially incompatible stuff.
  export HOME="$WORKDIR/fake_home"
  mkdir "\$HOME"

  # Create a venv that _does_ use system-site-packages, since everything is already installed on the container.
  # We cannot work on the container itself, as the base system is immutable.
  python3 -m venv --system-site-packages "$WORKDIR/oggm_env"
  source "$WORKDIR/oggm_env/bin/activate"

  # Make sure latest pip is installed
  pip install --upgrade pip setuptools

  # Install our test OGGM, using a fixed version here, to show how this will always produce identical results.
  pip install --upgrade "git+https://github.com/OGGM/oggm.git@3809cc80bc3d82ecd1dfa5e948c66f0ea864bf4c"

  # Run the actual tests.
  pytest --run-slow --run-download --mpl-oggm --pyargs oggm

EOF

# Clean up after ourselves
rm -rf "$WORKDIR"

# Print a final message so you can actually see it being done in the output log.
echo "DONE"