A pair of scripts which can be used to launch Jupyter notebooks on the Vermont Advanced Computing Center

jupyter-server.sbatch

#!/bin/bash
# Specify a partition
#SBATCH --partition=bdgpu
# Request nodes
#SBATCH --nodes=1
# Request some processor cores
#SBATCH --ntasks=8
# Request a GPU
#SBATCH --gpus=0
# Request memory 
#SBATCH --mem=40G
# Maximum runtime in hours
#SBATCH --time=48:00:00
# Name of this job
#SBATCH --job-name=jupyter


# Allow for the use of conda activate
source ~/.bash_profile

# Move to submission directory
cd ${SLURM_SUBMIT_DIR}

# your job execution follows:
whoami
hostname
conda activate deep-amd-3.9
time jupyter notebook --no-browser --port=8900

start-jupyter.sh

#!/bin/bash

out_dir=$DEVDIR/tmp
mkdir -p $out_dir
jup_outfile=$out_dir/jupyter.out

# First, kill ssh forwards from previous times running this command which may be left hanging
# around without getting properly cleaned up.
# WARNING: This assumes we're only running one notebook at a time!
if [ -f $jup_outfile ]; then
    found=false
    port=$(grep "http://localhost" $jup_outfile | sed -E 's/.*localhost:([0-9]+).*/\1/' | head -1)
    while read usr pid rest; do
      if [ ! -z $pid ]; then
          echo "Killing previous connection: $usr $pid $rest"
          kill -9 $pid
          found=true
      fi
    done <<< "$(ps -ef | grep $USER'.*[0-9] ssh.*'$port)"
    if $found; then echo ""; fi
fi

>$jup_outfile  # clear the output file

# start the jupyter server on the cluster
sbatch --output=$jup_outfile $DEVDIR/jupyter-server.sbatch
echo "Jupyter notebook server starting on compute node, waiting..."
# wait for notebook server to start
while ! grep -q "http://localhost" $jup_outfile; do
    sleep 1
done

# get port
port=$(grep "http://localhost" $jup_outfile | sed -E 's/.*localhost:([0-9]+).*/\1/' | head -1)
# get token
token=$(grep "http://localhost:$port/?token=" $jup_outfile | sed 's/.*token=\(.*\)/\1/'| head -1)
# get user
user=$(head -n 1 $jup_outfile | sed 's/.*\[\(.*\)\].*/\1/' | head -1)
# get host
host=$(head -n 2 $jup_outfile | tail -n 1 | sed 's/.*\[\(.*\)\].*/\1/' | head -1)

echo "Jupyter notebook server started on compute node $user@$host"
sleep 1

# connect to notebook server
ssh -fNT -L $port:localhost:$port $user@$host & 

echo "Port forwarding started $user@$host:$port -> localhost:$port"
echo ""
echo "connect to notebook server at:"
echo "  http://localhost:$port/?token=$token"
echo ""

When you use these scripts, be sure to:

• Edit the start script to configure the output location.
• Edit the sbatch script to configure your resource request and conda env.
• Remember that you need to change the resources and environment as needed whenever you run.
• Note that the port number is (unfortunately) hard-coded, and you must use this port number when connecting to the notebook.