On Wed, Dec 4, 2019 at 6:15 PM <klaasdemter(a)gmail.com> wrote:
Hi,
I'm having performance issues with a ovirt installation. It is showing
high steal (5-10%) for a cpu intensive VM. The hypervisor however has
more than 65% of his resources idle while the steal is seen inside of
the VM.
Even when placing only a single VM on a hypervisor it still receives
steal (0-2%), even though the hypervisor is not overcommited.
Hypervisor:
2 Socket system in total 2*28(56HT) cores
VM:
30vCPUs (ovirt seems to think its a good idea to make that 15 sockets *
2 cores)
I think you can control this in oVirt.
My questions are:
a) Could it be that the hypervisor is trying to schedule all 30 cores on
a single numa node, ie using the HT cores instead of "real" ones and
this shows up as steal?
b) Do I need to make VMs this big numa-aware and spread the vm over both
numa nodes?
c) Would using the High Performance VM type help in this kind of situation?
d) General advise: how do I reduce steal in an environment where the
hypervisor has idle resources
Any advise would be appreciated.
These questions are mainly about qemu, so adding qemu-discuss.
I think it will help if you share your vm qemu command line, found in:
/var/log/libvit/qemu/vm-name.log
Nir