Skip to content

Instantly share code, notes, and snippets.

@erenard

erenard/gaming_vm.md

Created November 6, 2024 08:43
Show Gist options
  • Save erenard/af84d9cb256d71381feb87d06c10739d to your computer and use it in GitHub Desktop.
Save erenard/af84d9cb256d71381feb87d06c10739d to your computer and use it in GitHub Desktop.
Download ZIP
Gaming VM with GPU passthroug, QEMU emulator, KVM hypervisor and libVirt
Raw
gaming_vm.md

Gaming VM with GPU passthroug notes

Raw
installation.md

CPU configuration

To have the lowest possible latency, cpu cores must be partitionned between host and guest OSes.

On CPUs with [ Intel Hyperthreading / AMD Simultaneous Multi-Threading] threads should not be separated to keep latency low.

First, read the installed CPU's topology

> virsh capabilities

will output -amongst other things- the capabilities/host/topology/cells/cell/cpus, here is an example with an AMD 5800X3D

<capabilities>
  <host>
    <topology>
      <cells num='1'>
        <cell id='0'>
          <cpus num='16'>
            <cpu id='0' socket_id='0' die_id='0' cluster_id='65535' core_id='0' siblings='0,8'/>
            <cpu id='1' socket_id='0' die_id='0' cluster_id='65535' core_id='1' siblings='1,9'/>
            <cpu id='2' socket_id='0' die_id='0' cluster_id='65535' core_id='2' siblings='2,10'/>
            <cpu id='3' socket_id='0' die_id='0' cluster_id='65535' core_id='3' siblings='3,11'/>
            <cpu id='4' socket_id='0' die_id='0' cluster_id='65535' core_id='4' siblings='4,12'/>
            <cpu id='5' socket_id='0' die_id='0' cluster_id='65535' core_id='5' siblings='5,13'/>
            <cpu id='6' socket_id='0' die_id='0' cluster_id='65535' core_id='6' siblings='6,14'/>
            <cpu id='7' socket_id='0' die_id='0' cluster_id='65535' core_id='7' siblings='7,15'/>
            <cpu id='8' socket_id='0' die_id='0' cluster_id='65535' core_id='0' siblings='0,8'/>
            <cpu id='9' socket_id='0' die_id='0' cluster_id='65535' core_id='1' siblings='1,9'/>
            <cpu id='10' socket_id='0' die_id='0' cluster_id='65535' core_id='2' siblings='2,10'/>
            <cpu id='11' socket_id='0' die_id='0' cluster_id='65535' core_id='3' siblings='3,11'/>
            <cpu id='12' socket_id='0' die_id='0' cluster_id='65535' core_id='4' siblings='4,12'/>
            <cpu id='13' socket_id='0' die_id='0' cluster_id='65535' core_id='5' siblings='5,13'/>
            <cpu id='14' socket_id='0' die_id='0' cluster_id='65535' core_id='6' siblings='6,14'/>
            <cpu id='15' socket_id='0' die_id='0' cluster_id='65535' core_id='7' siblings='7,15'/>
          </cpus>
        </cell>
      </cells>
    </topology>
  </host>
</capabilities>

The take away is: siblings should not be separated.

In the above example, reserving the core 0 for the host means that all the cpus having core_id='0' must be reserved togethers. That would be cpu id='0' and cpu id='8', each of them having siblings='0,8'.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment