fung cluster1

Our group's first dedicated local server fung-cluster1 went online on 2nd Feb 2023. It comes with 8 Nvidia A40 GPUs, each with a video memory of 48GB.

While there is no queue system on the computing resources and thus in general the "first come - first served" principle applies, good etiquette is encouraged. For instance, use GPU 0 for quick experimentation and prototyping only and leave long-running jobs to GPU 1-7.

Table of Contents

1. Pre-requisite
2. Login
3. Setting Up Conda

Pre-requisite

To connect to fung-cluster1, you need:

1. A GT account,
2. Access permission to fung-cluster1.

Locally, you need a terminal application. For Linux/MacOS, you can use the default Terminal app. For Windows, you may download Windows Terminal.

Login

SSH Connection

The most common way to connect to fung-cluster1 is via ssh, which allows you to establish a secure connection from your local computer (client) to the cluster (server).

Suppose we have an account with username georgeburdell. In general, this should be your GTID that precedes the @ symbol in your GT email.

You can connect to fung-cluster1 by:

ssh [email protected]

The password is the same as your GT account password. In general, you need to be on GT network or VPN to connect to fung-cluster1.

Password-less SSH Connection (optional)

To avoid entering your password every time when you need to connect to fung-cluster1, you can generate an SSH key pair for authentication. First, in your local computer's terminal, generate the key pair by

ssh-keygen

You will be prompted with the following line: Enter file in which to save the key (/Users/georgeburdell/.ssh/id_rsa):, which is essentially asking you to name your keys (default is id_rsa). In our case, key in fung1 and press enter. If prompted to enter passphrase, just skip by pressing Enter for now.

When the generation ends, you will find two keys: fung1 and fung1.pub under the current directory. If the current directory is not ~/.ssh/, you might want to move these two keys there by mv fung1 fung1.pub ~/.ssh.

Next, we want fung-cluster1 to host the public key (fung1.pub). To do this, we can simply do the following:

ssh-copy-id -i /path/to/fung1.pub [email protected]

You will need to key in your password for this action.

Lastly, we just need to edit ~/.ssh/config accordingly. Below is an example:

Host fung1
  User georgeburdell
  Hostname fung-cluster1.cc.gatech.edu
  IdentityFile /path/to/fung1

Note that the IdentityFile line has to point to the secret key fung1, the one without .pub extension!

Now, make sure to save the changes and in your terminal, do ssh fung1. No password should be required now.

Setting Up Conda

Moving .conda to Project Space

Before installing a copy of conda distribution, you may want to move .conda to the project space.

The project directory of 5TB storage on fung-cluster1 is

/net/csefiles/coc-fung-cluster/

Installing Miniconda

The system does not come with global conda installation. Thus, each user has to install their own copy of anaconda (you can check if there is an existing conda by doing conda -V). In the absence of an existing conda installation, miniconda is strongly encouraged compared to the full anaconda distribution.

To install miniconda, do the following:

1. Go to the webpage of the miniconda installer links.
2. Under "Latest Miniconda Installer Links", locate the link for Miniconda3 Linux 64-bit.
3. Right click on Miniconda3 Linux 64-bit and copy link address. Check that this address ends with .sh extension.
4. On fung-cluster1, download the installer by wget [copied link address]. For example:

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh

5. Install miniconda through the bash script, e.g. bash Miniconda3-latest-Linux-x86_64.sh.
6. We need to reload ~/.bashrc to let the system know that conda has been installed. To do so, run source ~/.bashrc.

Now, miniconda should be ready for use.