Mathematical Aspects of Machine Learning

View the Project on GitHub jeremy9959/Math-5800-Spring-2020

Using the cluster to get GPU access

The UConn hpc cluster offers access to a range of GPU facilities. This note describes the bare minimum you need to get access to a compute node with a GPU that can run pytorch (CUDA 10.1).

  1. Get an account on the cluster (the Storrs facility). You can use my name as advisor when asked.

  2. Make sure you have the VPN up and running. Then you can log in to the head node on the cluster:

    $ ssh
  3. You are not allowed to run jobs on the login node, so you want to start an interactive job using the scheduler. If you don’t need the GPU, you can run

     $ srun --pty bash

    This will start a bash shell on an available node. Take note of the prompt or run the hostname commmand to see which node you’ve been assigned.

     [jet08013@cn277 ~]$ hostname
  4. Probably the simplest thing to do is to install the anaconda distribution in your home directory on the cluster, just as you would on your home computer. You are given 50GB of storage for your home directory which is plenty – unless you are doing serious work with large datafiles, but in that case you will need to learn more about the cluster than this note will tell you. Alternatively, install miniconda, which is basically anaconda python and conda; from there you can install just the packages you need. Do this using wget:

      $ wget
      $ sh
  5. Once you have anaconda working, you can install pytorch using conda following the directions on the pytorch website. If you want to use the GPU, make sure to follow the instructions for CUDA on linux. The command is:

    conda install pytorch torchvision cudatoolkit=10.1 -c pytorch

    It’s best to do this in a virtual environment (if you don’t know what this means, ask me).

    One complication: if you want to run pytorch without the GPU, on a cluster node without a GPU – perhaps for development – then you should install two versions of pytorch in two different virtual environments. The CUDA version won’t run on a node without a GPU.

  6. Go back to the login node (by typing exit in your interactive shell job).

  7. Connect to a gpu node. We are using the RTX GPU, which is a single-precision, consumer-type GPU. There are fancier ones on the cluster but they are running a different version of CUDA. If you are using tensorflow, you can use these other nodes, but I won’t talk about that here.

    Notice that I activate the virtual environment that has pytorch in it.

    $ srun -p gpu_rtx --gres=gpu:1 --pty bash
    srun: job 3236057 queued and waiting for resources
    srun: job 3236057 has been allocated resources
    (base) [jet08013@gpu01 ~]$ conda activate torch
    (torch) [jet08013@gpu01 ~]$ 
  8. Load CUDA (you can put this in your startup file if you want).

     $ module load cuda/10.1
  9. Check to see that torch is A-OK.

     (torch) [jet08013@gpu01 ~]$ python
     Python 3.8.1 (default, Jan  8 2020, 22:29:32)
     [GCC 7.3.0] :: Anaconda, Inc. on linux
     Type "help", "copyright", "credits" or "license" for more information.
     >>> import torch
     >>> print(torch.cuda.is_available())
  10. To run jupyter, you first start jupyter lab on the cluster, and then create an ssh “tunnel” to it so you can use your web browser. On the cluster:

    (torch) [jet08013@gpu01 ~]$ jupyter lab --no-browser --port=8888 --ip=$(hostname)

    Take note of this line at the end of the output:

    Or copy and paste one of these URLs:

    Also note in the prompt on the cluster which gpu node you are assigned to:

    (torch) [jet08013@gpu01 ~]$ hostname
  11. Keep your terminal window on the cluster open, and go back to your laptop and open an ssh tunnel:

    (laptop)$ ssh -NL localhost:8888:gpu01:8888

    where you replace gpu01 with whatever output you got from step 10 for the gpu node name. Keep this window open to maintain the connection.

  12. Finally, paste one of the URL’s you took note of into your web browser and hopefully a window will open on your browser connected to the cluster.

  13. To quit, you can quite from the window on your browser and close down the windows with the ssh tunnel and the cluster.

Note: There are many ways to simplify this process. If you find yourself using it and want to streamline it, talk to me.