Ansible project for managing a small cluster with high-performance networking backbone.

• Automatically setup SSH keys
• Configure cluster backbone network with jumbo frames and bonding
• Automated network performance testing
• Easy cluster management with Ansible

• uv - Fast Python package installer (Install uv)
• Two NVIDIA DGX Sparks connected by a QSFP cable

uv sync

This installs Ansible and all required collections.

Edit inventory/hosts.yml to match your cluster:

Update hostnames and IP addresses:

spark-0:
  ansible_host: 192.168.0.194  # Change to your node's IP/hostname
spark-1:
  ansible_host: 192.168.0.221  # Change to your node's IP/hostname

Select correct network interfaces for bonding:

The bond members should be the interfaces you want to bond together. Interface naming:

• np0 = left port
• np1 = right port

Example:

cluster_bond:
  members:
    - enp1s0f0np0    # First interface, left port
    - enP2p1s0f0np0  # Second interface, left port
  address: 172.16.0.1/30  # Cluster backbone IP
  mtu: 9000               # Jumbo frames

Verify interface names on your nodes with: ip link show

Configure SSH keys and passwordless sudo:

uv run ./playbooks/setup.yml

Enter your SSH password when prompted. This will:

• Update all packages and reboot if needed
• Generate SSH keypair in data/
• Configure passwordless SSH access
• Enable passwordless sudo

uv run ./playbooks/setup.yml -e "state=absent"

Set up network bonding with performance testing:

uv run ./playbooks/cluster-backbone.yml

This will:

• Install NetworkManager and iperf3
• Create bonded network interfaces (mode: balance-xor)
• Configure jumbo frames (MTU 9000)
• Run automated performance tests between nodes

uv run ./playbooks/cluster-backbone.yml -e "state=absent"

Make it easy to SSH to cluster nodes from your workstation:

uv run ./playbooks/configure-ssh-client.yml

After this, you can simply: ssh spark-0 or ssh spark-1

uv run ./playbooks/configure-ssh-client.yml -e "state=absent"

Keep your cluster up to date:

uv run ./playbooks/update.yml

This will:

• Update package cache
• Upgrade all packages
• Refresh firmware metadata and upgrade firmware
• Check if reboot is required
• Automatically reboot nodes that need it
• Wait for nodes to come back online

Install NVIDIA Workbench on all cluster nodes:

uv run ./playbooks/install-workbench.yml

This will:

• Download the latest nvwb-cli for your platform
• Install NVIDIA Workbench with Docker support
• Run in non-interactive mode

uv run ./playbooks/install-workbench.yml -e "state=absent"