Running a Kubernetes native x86_64 application on Raspberry Pis, and why you shouldn’t!

While this guide does work, don’t do this. Emulation is slow and Raspberry Pi CPUs (even overclocked to 2GHz) are slower!

Also, you have to use x86_64 emulation (rather than KVM) because Pis use an ARM instruction set rather than x86.

CPU benchmark for slowness

This benchmark is terrible, but gives an indication of single threaded performance

#On my work machine
time dd if=/dev/urandom of=/dev/null bs=2000000 count=100

real	0m4.264s
user	0m0.000s
sys	0m4.264s

# On a VM on k3s-005
$ time dd if=/dev/urandom of=/dev/null bs=2000000 count=100

real 0m 19.08s
user 0m 0.03s
sys 0m 18.94s

# Natively on that Pi
pi@node-005:~ $ time dd if=/dev/urandom of=/dev/null bs=2000000 count=100

real	0m2.784s
user	0m0.004s
sys	0m2.769s


x86_64 Application
Kubernetes (k3s)
QEMU x86_64 emulation
Raspberry Pis

Equipment in use

  • 4 * Pi 4 8GB
  • SSD enclosure per Pi supporting UASP
  • SSD per Pi
  • Ethernet switch
  • A router
  • Power
  • misc cables

OSs in use

  • Raspbian 64 bit on the Pis
  • Alpine in the x86_64 VMs

Setting up the Pis

We installed 64 bit Raspbian on the SSDs and booted from them.

Installing the dependencies

sudo apt update && sudo apt upgrade -y && sudo apt install -y qemu nmon virtinst qemu-utils qemu-system-x86 tmux vim dnsmasq-utils dnsmasq-base iptables libvirt-daemon-system

Overclocking the Pis

The Raspberry Pi doesn’t have the fastest CPU, so we overclocked it to 2GHz and over_voltage 6 to keep the warranty. Full guide

Setting up the QEMU groups

groupadd libvirt-qemu
groupadd libvirt
groupadd libvirtd
useradd -g libvirt-qemu libvirt-qemu

Setting up swap (in case it’s need since 8GB RAM isn’t much)

# Create and mount swap
sudo fallocate -l 64G /swapfile && sudo chmod 600 /swapfile && sudo mkswap /swapfile && sudo swapon /swapfile

Next add an entry to fstab so that the swap is mounted on boot.

sudo su -c "echo '/swapfile swap swap defaults 0 0' >> /etc/fstab"

Next check that trim works on the PI, if it’s working you should get something similar to the following

sudo fstrim -av
/boot: 0 B (0 bytes) trimmed
/: 19.9 GiB (21294051328 bytes) trimmed

If this fails, you’ll need to follow

Due to the high volume of writes expected since the SSD will be used as RAM, configure TRIM to run frequently to prevent rapid deterioration of the SSD.

# trimming every 2 min:
sudo vi /lib/systemd/system/fstrim.timer
# change:

sudo systemctl daemon-reload

Setting up the Bridge networking so the VMs can connect directly (helpful for k8s)

Largely following

sudo su -c "echo '[NetDev]
' >> /etc/systemd/network/bridge-br0.netdev"

sudo su -c "echo '[Match]

' >> /etc/systemd/network/bridge-br0.netdev"
sudo systemctl enable systemd-networkd

Next ensure that eth0 is on the bridge network, and it’s the bridge that does dhcp rather than other interfaces

sudo vim /etc/dhcpcd.conf
# At top add the following, without the #
# denyinterfaces wlan0 eth0

# at the very bottom add, without the #
# interface br0

Next is to let QEMU use this bridge, largely following

sudo mkdir /etc/qemu

# Add all that into a script (run as root):

cat <<EOT > /etc/systemd/network/bridge-br0.netdev

cat <<EOT > /etc/systemd/network/


sed -i '1s/^/denyinterfaces wlan0 eth0 \n/' /etc/dhcpcd.conf
echo "interface br0" >> /etc/dhcpcd.conf

if [[ ! -d "/etc/qemu" ]]; then
  mkdir /etc/qemu
echo "allow br0" > /etc/qemu/bridge.conf

#then need to execute later
systemctl enable system-networkd

Running the x86 VM

For the VM we used Alpine as it’s a light OS and compatible with k3s. You can choose your download from


Next you’ll have to create a disk image for the Alpine VM, we created a 128GB sparse image given the size of the SSDs we used.

qemu-img create -f qcow2 alpine.qcow2 128G

Now to run the VM, this command must be run from a display because it attempts to open a window. It is possible to do this in a headless fashion, but we couldn’t get that working. Each VM needs a unique mac address for the networking to work correctly

sudo qemu-system-x86_64 --name alpine-node -drive file=/home/pi/alpine.qcow2 -smp cpus=4 -m 60G,slots=4,maxmem=61G -accel tcg,thread=multi -nic bridge,br=br0,model=virtio-net-pci,mac=52:54:00:12:34:50

Exciting options, -accel tcg,thread=multi enables the VM to use multiple host cores virtio-net-pci allows a virtual nic that works with the bridge configured.

You should ensure each of these VMs have a unique hostname, and a static IP in your router. This will make k3s configuration much easier

Configuring k3s

For our setup, one Pi ran the k3s control plane natively (and didn’t have a VM running) and the other Pis each had a VM running. Those VMs were joined to the k3s cluster as worker nodes.

Before installing on the control node, follow to set up iptables correctly and to set up cgroups correctly.

Before installing on the (virtual) worker nodes, follow

Installing the control-plane

Unfortunately, the easiest way to install k3s with systemd is running random scripts from the internet…

sudo -i
apt install -y curl
curl -sfL | sh -

Prevent pods running on control plane

It’s generally a bad idea to run workloads on your control plane, especially in this case as our workloads with be x86 but the control plane is ARM64.

kubectl taint nodes node-0001

Adding worker nodes

Pretty simple, full guide here.

apk add curl && curl -sfL | K3S_URL=https://node-001:6443 K3S_TOKEN=SomeTokenFromControlPlane sh -

With this, the x86_64 workers nodes should be added the cluster. Each node should have a unique hostname and be connectable on that hostname from all nodes.


If you’re successful, you should have a working k3s cluster with workloads running on x86 (slowly!)

pi@node-001:~ $ kubectl get nodes
NAME                            STATUS   ROLES                  AGE    VERSION
node-001                        Ready    control-plane,master   105m   v1.21.5+k3s2
k3s-002                         Ready    control-plane,master   100m   v1.21.5+k3s2
k3s-003                         Ready    control-plane,master   90m   v1.21.5+k3s2
k3s-004                         Ready    control-plane,master   80m   v1.21.5+k3s2
k3s-005                         Ready    control-plane,master   70m   v1.21.5+k3s2

Why you shouldn’t do this

Over 1.5 cores of the Pi is used by k3s, never mind the workloads we want to run!

Here’s a simple pod which didn’t manage to start, even after 5 mins due to the CPU limitations.

pi@node-001:~ $ kubectl get pod -w
NAME                                              READY   STATUS              RESTARTS   AGE
kubegres-controller-manager-75b6765589-kvr97      1/2     ContainerCreating   0          4m54s


Don’t run x86 on Pis, especially not on kubernetes! They just aren’t fast enough (yet)

Copyright © 2024 Richard Finlay Tweed. All rights reserved. All views expressed are my own