Configuring a compute node

Last updated on 2026-06-19 | Edit this page

Overview

Questions

  • What packages does a compute node need to join the cluster?
  • How does the compute node authenticate with the login node?

Objectives

  • Install the required packages on a compute node
  • Copy the Slurm configuration and munge key from the login node
  • Mount the shared filesystem from the login node via NFS
  • Start munge and slurmd services
  • Disable WiFi once the compute node is connected via ethernet

This section demonstrates how to set up a compute node on your Raspberry Pi, and add it to your cluster.

Flash an SD card as described in episode 2 and give it a name of node02 where node is the name that you use for all your nodes in your HPC (e.g. orange, black, green, blue, yellow).

Check the hostname (and fix if required)

Check your hostname for your compute node ends 02. We covered this in more depth in the last section, so refer back to that for more info. In short, check it with the command hostname, and edit with sudo raspi-config if it is incorrect.

No changes to /etc/hosts are needed on the compute node. Cluster-wide hostname resolution (all nodes resolving each other by name) is provided by dnsmasq on the login node and IP addresses are delivered to compute nodes via DHCP. The hostname for the compute node is all that is needed locally.

Callout

Tip

You can confirm the login node has seen this compute node and issued it a DHCP lease by running the following on the login node (try it!):

BASH 

cat /var/lib/misc/dnsmasq.leases

Each line is an active lease: expiry timestamp, MAC address, assigned IP, hostname, and client ID. Your compute node should appear with the IP you reserved for it in /etc/dnsmasq.conf by setting a dhcp-host line with the MAC address for the compute node.

This is also a useful way to check IP addresses assigned to cluster nodes if they aren’t ending up where you expected, and you can even edit the file and delete lines to clear DHCP leases for clients if they have the wrong IP address (for example, if their MAC address wasn’t added to /etc/dnsmasq.conf on the login node).

Start with an update

We need to update packages on the compute nodes, too:

BASH 

sudo apt-get update
sudo apt upgrade -y
Discussion

Network Interface Priority

When you have multiple network adapters attached to a computer, the computer needs to know which interface to route traffic over. This is selected based on the priority of the interface.

During initial setup, while both wlan0 and eth0 are connected, your Pi can get confused about which interface to use for internet traffic. If packages aren’t downloading, give wlan0 a higher interface priority. Grab the device name from nmcli:

BASH 

pi@node02:~ $ nmcli con show
NAME                              UUID                                  TYPE      DEVICE
netplan-wlan0-CarpentriesOffline  e5799f3d-8920-3080-b93f-e6e5ac4ce778  wifi      wlan0
netplan-eth0                      75a1216a-9d1a-30cd-8aca-ace5526ec021  ethernet  eth0
lo                                c4c925ab-c23d-4a84-86f3-bb9133a05b92  loopback  lo

Then give wlan0 a higher interface metric:

BASH 

sudo nmcli con mod netplan-wlan0-CarpentriesOffline ipv4.route-metric 100
sudo nmcli con down netplan-wlan0-CarpentriesOffline && sudo nmcli con up netplan-wlan0-CarpentriesOffline

Once wlan0 is disabled at the end of this tutorial, the compute node routes all traffic through eth0 to the login node, which provides internet access via its own wlan0.

Install required packages

BASH 

sudo apt-get install -y slurmd slurm-client munge ntpsec ntpsec-ntpdate lmod nfs-common vim
Callout

ntp and ntpdate are no longer available on current Raspberry Pi OS. Use ntpsec and ntpsec-ntpdate instead; they provide the same functionality.

Package Purpose
slurmd Slurm compute node daemon: executes jobs dispatched by the login node
slurm-client Slurm client tools (srun, sbatch, squeue, etc.)
munge Authentication service used by Slurm to verify inter-node messages
ntpsec NTP time synchronisation daemon: keeps node clocks in sync with the login node
ntpsec-ntpdate One-shot time sync command, useful for initial clock correction on first boot
lmod Lua-based module system for loading software environments (e.g. EESSI)
nfs-common NFS client utilities to mount the shared filesystem from the login node
vim Text editor that confuses people trying to exit it. Try :q if stuck.

Verify that slurmd installed and the service unit is present:

BASH 

systemctl status slurmd

:: caution systemctl should show that Slurm is installed, but not configured yet. This is OK for now! We haven’t configured it yet, so it will be in a failure state:

systemctl shows that Slurm is installed but not configured yet ::

If slurmd is not found, the package may have been silently skipped during install. Run the install command again with only slurmd to confirm:

BASH 

sudo apt-get install -y slurmd

Create a mount point for the shared drive

BASH 

sudo mkdir /sharedfs

Copy configuration files from the login node

1. Slurm configuration files

Copy the slurm config from the login node to /etc/slurm/slurm.conf:

  • On login node: scp /etc/slurm/slurm.conf pi@node02.local:slurm.conf

  • On compute node:

    BASH 

    sudo mv slurm.conf /etc/slurm/slurm.conf
sudo chown root:root /etc/slurm/slurm.conf

2. Munge key

Copy /etc/munge/munge.key from the login node to the compute node:

On login node:

BASH 

sudo cp /etc/munge/munge.key munge.key
sudo chown pi:pi munge.key && chmod 664 munge.key
scp munge.key pi@node02.local:munge.key
rm munge.key

On compute node:

BASH 

sudo mv munge.key /etc/munge/munge.key
sudo chmod 400 /etc/munge/munge.key
sudo chown munge: /etc/munge/munge.key
Callout

Tip

munge.key is owned by root with permissions 400, so scp cannot read it directly as the pi user. The steps above copy it to the home directory first and relax the permissions just long enough to transfer it, then clean up the temporary copy.

3. Filesystem table (fstab)

Update /etc/fstab to show the following:

BASH 

# Leave these lines alone:
proc            /proc           proc    defaults          0       0
PARTUUID=3e3e7392-01  /boot/firmware  vfat    defaults          0       2
PARTUUID=3e3e7392-02  /               ext4    defaults,noatime  0       1

# Append these lines:
192.168.5.101:/sharedfs    /sharedfs    nfs    defaults   0 0
192.168.5.101:/home    /home    nfs    defaults   0 0

Then reload the modifications and mount everything:

BASH 

sudo systemctl daemon-reload
sudo mount -a

Slurm cgroups configuration

Slurm’s cgroup plugin is used by slurmd to enforce resource limits on jobs. We need to configure this on our compute nodes. Create /etc/slurm/cgroup.conf:

BASH 

sudo tee /etc/slurm/cgroup.conf << 'EOF'
CgroupPlugin=autodetect
ConstrainCores=yes
ConstrainRAMSpace=yes
EOF

And /etc/slurm/cgroup_allowed_devices_file.conf:

BASH 

sudo tee /etc/slurm/cgroup_allowed_devices_file.conf << 'EOF'
/dev/null
/dev/urandom
/dev/zero
/dev/sda*
/dev/cpu/*/*
/dev/pts/*
/dev/shm
EOF

This should be a reasonable default configuration, but for a deeper dive, see (the Slurm cgroups documentation)[https://slurm.schedmd.com/cgroups.html]

Start munge and slurmd

Now that the config files are in place, start munge first (slurmd depends on it), then slurmd:

BASH 

sudo systemctl restart munge
sudo systemctl restart slurmd
sudo systemctl status slurmd

slurmd should now show active (running). If it still fails, check the log for details:

BASH 

sudo journalctl -u slurmd -n 30
Callout

Tip

We can check our node’s status from the login node using sinfo. If the node is down (in state FAIL), use scontrol to bring it back up:

BASH 

sudo scontrol update NodeName=node02 State=RESUME

Install EESSI

BASH 

mkdir eessi
cd eessi
wget https://raw.githubusercontent.com/EESSI/eessi-demo/main/scripts/install_cvmfs_eessi.sh
sudo bash ./install_cvmfs_eessi.sh

source /cvmfs/software.eessi.io/versions/2023.06/init/lmod/bash

# Older versions of EESSI needed us to put the source line into profile:
# echo "source /cvmfs/software.eessi.io/versions/2023.06/init/bash" | sudo tee -a /etc/profile
# We don't need to do this any more

Disable WiFi and Bluetooth

Unlike the login node (which keeps wlan0 as its internet uplink), the compute node can now disable WiFi. All traffic will route through eth0 to the login node and out via its wlan0 connection.

sudo nmcli con down netplan-wlan0-CarpentriesOffline should take down the default network configured in the Raspberry Pi Imager software. However, we can permanently disable the hardware for the built-in WiFi chip in the boot configuration file.

Open /boot/firmware/config.txt and add the following two lines at the bottom in the [all] section.

INI 

dtoverlay=disable-wifi
dtoverlay=disable-bt

Save the file and reboot! You now have a configured compute node. In the next section, we’ll test our cluster by submitting jobs with slurm.

Key Points
  • The compute node must have the same munge key as the login node for Slurm authentication
  • Copy slurm.conf and munge.key from the login node before starting slurmd
  • Mount shared filesystems via NFS entries in /etc/fstab
  • Disable WiFi on compute nodes so all traffic routes through eth0 to the login node