Ok, so we found a "solution" that I thought I'll share in case this thread pops up in a future search.

The problem was that our Redhat 5.x guests still run with a kernel that lacks the "tickeless" feature, meaning that they are poking the host a 1000 times per second even if they haven't anything to do.

Disabling hyperthreading and most importantly, adding the parameter "divider=10" to the kernel boot line in grub does indeed lower the cpu utilization in the hosts almost to 0% when idling.

In our case we use the command

sudo /sbin/grubby --update-kernel=ALL --args="divider=10"

To upgrade the /boot/grub/grub.conf file and then we restart the vm.

Thank you everyone for the help.

Xavier

On Mon, Nov 10, 2014 at 8:32 AM, Xavier Naveira <xnaveira@gmail.com> wrote:

We tried a minimal installation from CD of RedHat 5.10 and it is the same.

This should be fairly easy to reproduce:

- Install a RedHat 6.5 hypervisor
- Install a RedHat 5.10 guest in it
- Enjoy your overused CPU

Is there someone with a similar setup out there?

Xavier

On Fri, Nov 7, 2014 at 12:30 PM, Gianluca Cecchi <gianluca.cecchi@gmail.com> wrote:
On Fri, Nov 7, 2014 at 10:48 AM, Xavier Naveira <xnaveira@gmail.com> wrote:
Hi,

We have installed and added a new hypervisor into the ovirt cluster but this time with disabled HT.

I migrated a RedHat 5.10 machine to it and immediately the qemu-kvm process running the vm (freshly installed, just basic packages) began to consume 20-40% CPU as showed running top on the hypervisor.

Now that I have a hypervisor to run tests in, what would you suggest the next step is?

Thank you.

Xavier

If I remember correctly you had to test plain Qemu/KVM on CentOS 6.5 and see if the difference is made by oVirt itself or by the OS changed from 5.x to 6.y...
And also compare command line (eventually both in 5.x and 6.x) between plain Qemu/KVM and oVirt spawned VMs
Gianluca