Higher CPU Usage KVM/QEMU on host vs virtual machine

I switched to KVM/QEMU vs docker supervised. I’ve noticed that the host machine shows 20-30% cpu usage by qemu-system-x86_64, but the virtual machine is uses less than 5%. I found online that making these changes to the config can help, but they did not help.

  <clock offset='utc'>
    <timer name='rtc' tickpolicy='catchup'/>
    <timer name='pit' tickpolicy='delay'/>
    <timer name='hpet' present='no'/>
    <timer name='hypervclock' present='yes'/>
  </clock>

  <!-- after: this config drops to about 3% of a host CPU core -->
  <clock offset='utc'>
    <timer name='hpet' present='yes'/>
    <timer name='hypervclock' present='yes'/>
  </clock>

Home Assistant runs fine, but I’d like to reduce the host OS cpu. This is the only virtual machine I’m running.

What can I provide to help?

3 Likes

Got the same problem running homeassistant-ova on debian 11 in kvm/qemu/libvirt.

Any update suggestions on this?

1 Like

I see the same thing too… I’ve tried to assign 8GB of RAM, and 4 VCPUs, but Home Assistant continues to suck down RAM…

I tried the clock suggestion from @patmann03 … But it’s still so slow and so laggy…

Python…

I see similar behavior :frowning: hassio shows idle time 99% but host cpu shows 8-12% usage for this VM. I have other VMs they don’t show same CPU usage. I did disable all addons and removed passthru USB devices from host and hassio kvm cpu usage on host is contantly 7-12% (at ‘top’).

Hello,

I confirm. I have the same issue. 24% CPU usage by qemu on the Host, but close to 0 inside the VM.
Is there really such a big overhead due to virtualization?

Best regards,
Hendrik

Hey, I know that this thread is a little bit old, but you can find it through google when you look for the specific problem of running Home Assistant in Qemu/KVM and encounter higher cpu load on the host than on the guest.

I was running the system with 3 GB of RAM and typically had 30 % CPU on the virtualization host, although inside the load was way lower at some percent. I tried to change the tick settings, but that did not change anything.

After setting RAM to 3,5 GB I saw a decrease in CPU load, maybe related to context switches. Further changing to 4 GB did not change anything. This is what Telegraf gives me:

There is a visible decrease in CPU load. Also context switches and CPU load does usually does not change over the weeks, so I think this is not an effect from the system restart. Interestingly, Power consumption of the Intel J4205 stayed the same.

Hello @crbble ,

thanks for sharing.
I have increased the memory to 4GB and I see no improvement.
In the HA Webinterface I see ~10% CPU load. On the Host, I see 40%.

Do you have a link at hand, that describes how to get the graph of context switches?

Best regards,
Hendrik

Hi,

same problem with my installation of HAOS as a Qemu VM managed by libvirt. The last two devices I added were a Conbee II Zigbee USB stick and a Z-Wave USB stick along with ZWave integration, ZWave JS add-on, and Zigbee2Mqtt add-on. I added the USB sticks as usb-serial devices to the VM via

<serial type="dev">
  <source path="/dev/serial/by-id/usb-...-if00"/>
  <target type="usb-serial" port="2">
    <model name="usb-serial"/>
  </target>
  <alias name="serial2"/>
  <address type="usb" bus="0" port="5"/>
</serial>

HAOS is using constantly almost one CPU core on the host, but the guest is idling at a few percent.

I installed the VM with this command at first:

  virt-install --import --connect qemu:///system --name haos --memory 2048 --vcpus=2 \
    --disk haos_ova-10.5.qcow2,format=qcow2,bus=virtio  \
    --osinfo detect=on,require=off  \
    --network=bridge:br0,model=virtio \
    --boot uefi

Other values for --os-info did not work.

The above command configured a USB2 controller. To get a more ‘virtualization-friendly’ XHCI USB3 controller (USB emulation — QEMU documentation), I re-installed the VM with

  virt-install --import --connect qemu:///system --name haos --memory 2048 --vcpus=2 \
    --disk haos_ova-10.5.qcow2,format=qcow2,bus=virtio  \
    --controller usb3 \
    --osinfo detect=on,require=off  \
    --network=bridge:br0,model=virtio \
    --boot uefi

but that didn’t change the load issue on the host.

I researched for a while and found a comparable issue description here: KVM / qemu – high CPU load but low CPU usage – and thus goes by another day

I followed the same steps as described in the post, i.e. hooking up to the qemu-system-x86_64 process with strace -p <process ID>. That results in many lines like

ppoll([{fd=3, events=POLLIN}, {fd=4, events=POLLIN}, {fd=6, events=POLLIN}, {fd=7, events=POLLIN}, {fd=8, events=POLLIN}, {fd=10, events=POLLIN}, {fd=12, events=POLLIN}, {fd=13, events=POLLIN}, {fd=14, events=0}, {fd=21, events=POLLIN}, {fd=23, events=POLLIN}, {fd=36, events=POLLIN}, {fd=37, events=POLLIN}, {fd=43, events=POLLIN}, {fd=44, events=POLLIN}, {fd=46, events=POLLIN}, {fd=47, events=POLLIN}, {fd=48, events=POLLIN}, {fd=49, events=POLLIN}, {fd=50, events=POLLIN}, {fd=51, events=POLLIN}, {fd=52, events=POLLIN}, {fd=53, events=POLLIN}, {fd=54, events=POLLIN}, {fd=55, events=POLLIN}, {fd=56, events=POLLIN}, {fd=57, events=POLLIN}, {fd=58, events=POLLIN}, {fd=59, events=POLLIN}, {fd=60, events=POLLIN}, {fd=61, events=POLLIN}, {fd=62, events=POLLIN}, ...], 77, {tv_sec=0, tv_nsec=487819}, NULL, 8) = 0 (Timeout)

I checked the file descriptors with

ls -la /proc/<process ID>/fd/

but I couldn’t find a real connection to the usb-serial devices.

The article linked above reports a solution to the problem with aio=native, I tried

<disk type='file' device='disk'>
      <driver name='qemu' type='qcow2' cache='none' io='native'/>

in the libvirt qemu XML configuration file for the VM, but: no difference in host load.

Try to use host-CPU instead of kvm64/qemu64.

(qemu-system-x86_64 -cpu host)

In libvirtd’s qemu configuration I have

<cpu mode='host-passthrough' check='none' migratable='on'/>

a check with ps aux resulted in:

qemu-system-x86_64 -cpu host,migratable=on

I tried

<cpu mode='host-passthrough' check='none' migratable='off'/>

ps xau resulted in:

qemu-system-x86_64 -cpu host

I tried

<cpu mode='host-model' />

which gave

qemu-system-x86_64 -cpu Cooperlake,ss=on,vmx=on ...

All had the same load.

Over time, I’ve seen a couple of threads on this subject, and it seems almost all who post see this as well so you are not alone. I too have around 20+% CPU load on the Host and just a few% on HAOS. I use Host pass-through as well.

It is a pitty… The virtualisation overhead should be smaller…