Notes and playbook for running my compute cluster(s).
Currently focused on experimenting with different workload schedulers on my Turing Pi board.
First, install Ubuntu on each node in the cluster using the BMC.
Then create an ansible user with the correct ssh key for accessing the node. Below is the sequence of commands I ran on each node to bootstrap the cluster.
ubuntu@node1:~$ sudo su
root@node1:/home/ubuntu# useradd -rm ansible
root@node1:/home/ubuntu# echo "ansible ALL=(ALL) NOPASSWD:ALL" > /etc/sudoers.d/ansible
root@node1:/home/ubuntu# chmod 440 /etc/sudoers.d/ansible
root@node1:/home/ubuntu# mkdir -m 700 /home/ansible/.ssh
root@node1:/home/ubuntu# echo "ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIO7kqBPUpVpPS4pTJXk5zQu2FMLZkNda2Q521btGCRDI" > /home/ansible/.ssh/authorized_keys
root@node1:/home/ubuntu# chmod 600 /home/ansible/.ssh/authorized_keys
root@node1:/home/ubuntu# chown -R ansible.ansible /home/ansible/.ssh
root@node1:/home/ubuntu# exit
exit
Finally, disable root login and password login for the ssh daemon.
- Create a new virtual environment for installing python packages (like ansible). This only needs to be done the first time setting up this repository on a computer.
python3 -m venv venv - Activate virtual environment.
source venv/bin/activate - Install dependencies.
pip install -r requirements.txt - Check that you can ping every node.
ansible cluster -m ping -i inventory.yml
Packaged as 3 sockets. 1 Socket with 4 high performance cores, 2 sockets with 2 efficiency cores each. This confuses slurm so I opted to just say "there are 8 cpus that are all the same. deal with it." by setting the config_overrides flag in the SlurmdParameters config value.