Skip to content

Files

Latest commit

 

History

History
68 lines (58 loc) · 3.05 KB

gpu_servers.md

File metadata and controls

68 lines (58 loc) · 3.05 KB

GPU Servers

We have a few GPU machines that we can use.

  1. honeydew.cs.uwaterloo.ca has an NVIDIA Titan XP with 12GB RAM. We gratefully acknowledge the support of NVIDIA Corporation with the donation of the Titan Xp GPU.
  2. beaker.cs.uwaterloo.ca has an NVIDIA Titan XP with 12GB RAM. This machine was donated by Morgan McGuire of NVIDIA.
  3. koios.cs.uwaterloo.ca has 8 NVIDIA Titan XPs, each with 12GB RAM. This machine belongs to Tyler, but he invites us to use it. Thanks Tyler!

jupyterhub on koios

This is a convenient way to use koios without having to mess around with ssh. The machine is set up with jupyterhub. To request an account, go to

koios.cs.uwaterloo.ca/hub/signup

An admin will then have to go to koios.cs.uwaterloo.ca/hub/authorize to authorize the account. After that, just point your browser to

koios.cs.uwaterloo.ca

To load your code onto it, we suggest using git. Open up a terminal, and git-away!

Public vs. Private repository

You can git clone "<git repo url>" in the terminal easily. However, this works well only if your repo is public. For the private repo, you need to grant a permission.

cd .ssh
ssh-keygen
  • For the simplicity, just enter in response to the comments.
  • Then open id_rsa.pub and copy the content of the file using vim, for example.
  • Then, on your cloud server, (github, gitlab, ...) , on your profile > settings,
  • On the left, side-bar select SSH and GPG keys then new ssh key and paste the copied key from the clipboard.
  • For example, [https://github.com/settings/keys](https://github.com/settings/ssh/new),
  • Obviously a good description helps to recall SSH->Device later.

Running jupyter notebook Remotely

  1. Login to remote machine
ssh honeydew.cs.uwaterloo.ca

If you are doing this from off-campus, you might have to use a VPN.

  1. Instantiate the machine-learning environment (if you want to use pyTorch, etc)
source activate ml
  1. On remote machine, start the notebook on a specified port, and run it in the background. The command nohup allows the process to continue even if you close the terminal window or logout of the server. Note that the output will then be directed to the file nohup.out.
nohup jupyter notebook --no-browser --port=8885 &
  1. Make note of the URL that the notebook is running on. You might have to look in the file nohup.out for this. It'll have a really long random string of characters.
  2. Log out of remote machine.
  3. Create an ssh tunnel to the chosen port
ssh -N -f -L localhost:8884:localhost:8885 honeydew.cs.uwaterloo.ca
  1. In your browser, go to the URL that you noted, but replace 8885 with 8884
http://localhost:8884

You can close the terminal window, or the browser window, but that won’t affect the notebook. You can always repeat steps 6 and 7 to continue where you left off. Amazing, right?!

GPU Monitoring

Type nvidia-smi to see the GPU load.