Login to SciComp GPUs

The following is how to use one of the ML scicomp machines that has 4 Titan RTX GPU cards installed.
Steps:
  1. Setting up the software environment seems to be more easily done using conda. We need to first log into jlab common environment with the below ssh command.
    ssh login.jlab.org
    You'll be prompted to enter your Jlab CUE password.
  2. We need to log into ifarm with the following ssh command.
    ssh ifarm1901
  3. Setting up Tensorflow Python environment (Needs to be done just once!)
    • The software must be set up using a computer other than sciml190X since it needs a level of outside network access not available there.
    • We recommend using Conda to manage your python packages and environments.
    • Also, the size of the installation is large enough that it won't fit easily in you home directory. Conda likes to install things in ~/.conda so that must be a link to some larger disk.
    • If ~/.conda already exists, please delete it since we are going to create a symbolic link named ~/.conda
    • Create a folder in your work directory that can be linked to "~/.conda". For me, I created a folder named condaenv in "/work/halld2/home/kishan/". You can simply achieve this by running the following commands
    • mkdir /work/<your hall>/home/<your name>/condaenv
      ln -s /work/<your hall> /home/<your name>/condaenv ~/.conda
    • You can check if symbolic link is set up by running.
    • ls -la
      you will see one of the entries as .conda -> /work/<your hall>/home/<your name>/condaenv
    • Now run the following commands to load Anaconda3 and create a virtual environment named tf-gpu with tensorflow-gpu, cudatoolkit, keras and numpy installed.
    • bash
      source /etc/profile.d/modules.sh
      module use /apps/modulefiles
      module load anaconda3/4.5.12
      conda create -n tf-gpu tensorflow-gpu cudatoolkit keras numpy
    • Activate the tf-gpu virtual environment.
    • conda activate tf-gpu
  4. Reserving the GPUs
    • To reserve 2 GPU cards
    • salloc --gres gpu:TitanRTX:2 --partition gpu --nodes 1
      srun --pty bash
    • You may need to specify amount of time and memory to reserve to train a ML model
    • salloc --gres gpu:TitanRTX:2 --partition gpu --nodes 1 --time=12:00:00 --mem=24GB
      If you with to reserve n GPU nodes, change above command to gpu:TitanRTX:n
    • Now activate your tf virtual environment by running below commands.
    • source /etc/profile.d/modules.sh
      module use /apps/modulefiles
      module load anaconda3/4.5.12
      conda activate tf-gpu
    • If you log out, of both the srun and salloc command, then the job should complete and the resource should be released. You can check this by just running the sacct command to see a list of your jobs and if there are any running:
    • sacct
    • If there is a running job that you want to kill so the resource is released, cancel the jobid via:
    • scancel jobid
    • To see available devices and their status:
    • nvidia-smi
    • To see which devices were assigned to you run this:
    • printenv CUDA_VISIBLE_DEVICES