stefan-it/lrz-gpu-tutorial

Useful tips for LRZ GPU usage

★ 7Forks 0GitHub ↗Compare

README

LRZ GPU tutorial

This repository contains some hints for using the LRZ GPUs (mainly the DGX-1 supercomputer and virtual servers with a single P100).

Notice: This is not an official repository from LRZ.

SSH-key setup

In order to properly login into reserved container, it is very imporant to create a private/public SSH key pair.

Please refer to the LRZ documentation for setting up a SSH-key pair. Another excellent reference is the GitLab documentation.

Notice: It is very important that you use a strong password for your new SSH key pair. The LRZ security policies forbid any empty key! So you are forced to enter a password.

Store public SSH-key in LRZ Data Lab

Once the SSH key pair was created, the public key must be stored in your settings on the Data Lab reservation website.

Visit the LRZ Data Lab website. This website is only accessible within the MWN (Münchner Wissenschaftsnetz). If you're not in the MWN you can use a VPN connection. In this case you can find necessary information here.

Once you have logged into the LRZ Data Lab click on "My Account" and choose "Profile".

Locate your previously created public key (ending with .pub), e.g.:

cat ~/id_rsa.pub # or use cat ~/id_ed25519.pub in case of Ed25519

and copy this into the text field named "SSH Public Key" in the Data Lab. The next figure shows an example:

LRZ Data Lab

Resources

The LRZ currently provides access to the following GPU systems:

  • NVIDIA DGX-1 with 8 P100
  • Virtual server with 1 P100 and Docker
  • NVIDIA DGX-1 with 8 V100 and Docker

Thus, the DGX-1 tutorial can be found here.

The tutorial for the virtual server system can be found here.

Some nice videos from the High Performance Computing for AI workshop can be found here.

Byobu

Byobu is a nice text-based window manager and terminal multiplexer. Why should you use it? Imagine you are training a very deep neural network that usually takes several hours. When your internet connection is suddenly interrupted, your connection to the container is also capped and the training is also aborted.

Just use this little tool byobu and you can "resume" your terminal connection.

It can be installed via (this is only necessary when you working on the DGX-1):

sudo apt install byobu

Now start a new terminal session with (on DGX-1 or on the virtual server):

byobu

and everything will work fine. For more information see this tutorial.

Acknowledgments

We would like to thank the Leibniz-Rechenzentrum der Bayerischen Akademie der Wissenschaften (LRZ) for giving us access to the NVIDIA DGX-1 supercomputer.

Contact (Bugs, Feedback, Contribution and more)

Contributions to this tutorial are highly welcome!

For questions about this tutorial just open an issue on GitHub here. If you have any corrections or further improvements please open a pull request.

Contributors

stefan-it

Issues