Running Jupyter on a GPU node

TODO: Need to Update for Fall 2021

You can currently access a K4000 GPU on the ACI interactive nodes from the Jupyter notebook server. Students in the class can access more powerful P100 GPUs (with support for double precission arithmetic) via the CyberLAMP cluster by specifying the following PBS options:

If you would like to run Jupyter on a GPU node, then you can try the following instructions kindly provided by Justin Petucci:

  1. Download firefox singularity image:
mkdir -p ~/work/sw/singularity
cd ~/work/sw/singularity
singularity pull shub://jpetucci/firefox_icsaci
  1. Launch an Interactive Desktop session from OpenOnDemand

  2. Submit an interactive batch job (you will need to change the resources like memory and walltime based on your needs):

qsub -I -X -A cyberlamp_class -l qos=cl_class -l nodes=1:ppn=1:gpus=1 -l walltime=2:00:00
  1. After your job starts, start a screen session:
  1. Load the Python module:
module load python/3.6.3-anaconda5.0.1
  1. Start Jupyter:
jupyter notebook
  1. Copy the link for the notebook server. For example, the address might look like http://localhost:8888/?token=de99f0c76cbcfcb183693ff0491f00f278d781bb3586ea8e . Do not try to use the address given as an example.

  2. Detach the screen session: press control+A+D to detach from the screen session

  3. Launch firefox:

singularity run  ~/work/sw/singularity/jpetucci-firefox_icsaci-master-latest.simg
  1. Paste the address of the notebook server into firefox

Good luck. If you find corrections, pleaset submit a PR to improve these instructions.