I run zram swap on my machines. Normally this works great, but stuff like virtual machines absolutely hate having their memory compressed. They grind to a halt as kswapd works its magic.

It is possible to disable zram swap, but I’d like to keep it on while not letting my VMs swap. Contrary to the belief of a Stack Overflow answer that annoyed me, this is possible through the magic of control groups and systemd’s abstraction over it, specifically MemorySwapMax.

Direct invocation

If you invoke Qemu directly, you can use systemd-run to set resource limits on Qemu.

systemd-run --scope -p MemorySwapMax=0 --user qemu -arg -arg -arg

Overriding machine.slice

virt-manager does not spawn Qemu instances directly, so the systemd-run method is ineffective. It instead delegates to libvirt, but it would appear at first that there isn’t a way to get your hooks in. However, it does register virtual machines with systemd-machined, which means its VMs are under machine.slice.

Run systemctl edit machine.slice and add the following override:

[Slice]
MemorySwapMax=0

Shut down all your VMs and run systemctl restart machine.slice. This will disable swap for VMs and containers that register themselves with systemd-machined, though in practice I’ve only observed libvirt and systemd’s own container manager machinectl doing so. Docker and Podman containers aren’t affected by this override.

Note: How can you tell it’s working?

Qemu will use as much RAM as is allocated to it (with a negligible amount of overhead), so in your favorite process monitor, watch the RAM usage. If you allocated 16 GB to your VM and it hasn’t settled at 16 GB used, instead shrinking down to 1 GB or less, it’s probably swapping.