[ovirt-users] High CPU Usage caused by migration thread on Node

[jaumotte, styve]

Hi,

Thank you for your reply, I will disable at next reboot Numa options and cpu pass-through, and I will try vCPU pinning.

You told me to check logs to see what is the origin of migration threads, could you please help me and give me more details in the ways to investigate. I begin to work with the perf command line but it is a little bit unclear.

Styve



Réponse ou transfert de la part de Styve JAUMOTTE
De : Doron Fediuck [mailto:dfediuck at redhat.com]
Envoyé : dimanche 25 septembre 2016 08:27
À : jaumotte, styve
Cc : users at ovirt.org; Martin Sivák
Objet : Re: [ovirt-users] High CPU Usage caused by migration thread on Node



On Wed, Sep 21, 2016 at 12:28 PM, jaumotte, styve <s.jaumotte at maine-et-loire.fr<mailto:s.jaumotte at maine-et-loire.fr>> wrote:
Hi,

We’ve got a problem with a critical machine which running Oracle Database and if someone can help us it would be very nice.

Our VM is running rhel 6.5 and is configuring with 16 cores (2 virtual socket with 8 virtual cores) and 240Gb.

This virtual machine is running perfectly on the first host R620 : two cpu E5-2695 v2 - 12 cores on each cpu -  and 384Gb of memory.

We wan’t to move this virtual machine on a new server, Transtec with two  E5-2650 v3 – 10 cores on each cpu – and 256Gb of memory. We have encountered high cpu usage on this host.

Both servers are running centOS 7.2 and ovirt 3.6. We have already identified that on the node, many « migrate/N » are causing high cpu usage.

We have tried few options :

-       Pass-Through Host CPU

-       Configure NUMA : numa node count = 2, numa pinning
The first day the machine was running, everything is ok and migrate threads disapeared, but on the second day, high cpu and migration thread were back. Oracle services are being restarted everyday for cold backup but not the virtual machines.

I don’t know if looking in the Numa options is a good idea and I don’t why the everything is running ok on the R620 (perhaps the available memory)

Thank you for your advices.

Styve



_______________________________________________
Users mailing list
Users at ovirt.org<mailto:Users at ovirt.org>
http://lists.ovirt.org/mailman/listinfo/users

Hi Styve,
Pass-Through Host CPU is irrelevant unless you have a specific CPU flag you need to utilize.
So I suggest you refrain from using it.
As for NUMA, it works well as long as it's not interleaving. Here as well you  should use it only
if you understand how it should be used. Otherwise the safe option for you is vCPU pinning.
Can you please check your logs to see who is creating these threads and share it?

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.ovirt.org/pipermail/users/attachments/20160926/eb3a505a/attachment-0001.html>