Discussion:
[PVE-User] KVM CPU pinning, full control
Christian Balzer
2018-04-23 03:08:53 UTC
Permalink
Hello,

We're evaluating PVE as a potential replacement for a large ganeti cluster.
As in 10 compute nodes and about 700 VMs.

One absolute requirement is CPU pinning, both to keep the VMs on one
specific core (not even in-NUMA migrations) and to keep cores on the
respective NUMA nodes free of VMs to handle system tasks, in particular
network IRQs and the vhost tasks.

My exhaustive searches seem to indicate that this isn't possible natively
as such.
The "taskset" solution and the systemd script feels very clumsy and would
scale horribly from what I can see.
https://www.jnchen.com/blog/2017/01/pinning-vm-cpus-in-proxmox

And the NUMA parameter referenced below doesn't seem to be doing what I
and the 2nd to last poster want:
https://forum.proxmox.com/threads/how-to-specify-numa-cpus-in-vm-conf.29694/

So am I missing something, like being able to put "raw" qemu parameters
somewhere in the vm.conf?

Regards,

Christian
--
Christian Balzer Network/Systems Engineer
***@gol.com Rakuten Communications
Lindsay Mathieson
2018-04-23 03:35:02 UTC
Permalink
Post by Christian Balzer
So am I missing something, like being able to put "raw" qemu parameters
somewhere in the vm.conf?
https://pve.proxmox.com/wiki/Manual:_qm.conf

args: <string>

Arbitrary arguments passed to kvm, for example:

args: -no-reboot -no-hpet
--
Lindsay
Christian Balzer
2018-04-23 04:04:41 UTC
Permalink
Post by Lindsay Mathieson
Post by Christian Balzer
So am I missing something, like being able to put "raw" qemu parameters
somewhere in the vm.conf?
https://pve.proxmox.com/wiki/Manual:_qm.conf
args: <string>
args: -no-reboot -no-hpet
Thanks for that, alas the "-vcpupin" patch clearly hasn't made it into
qemu yet, so that won't save my bacon either.

So what remains is the that taskset bit or if one isn't a fan of 700
config files, a re-write of it.

Regards,

Christian
--
Christian Balzer Network/Systems Engineer
***@gol.com Rakuten Communications
Loading...