CentOS - Jun 2017 - kvm/qemu and CPU load

Hi,

I have a server using its 4 physical network interfaces
bonded, with the bonding interface added to a bridge.  The
bridge has the IP, and three VMs are using the bridge.  Two
of the VMs are running Debian, one is running Windoze 7.

CPU load caused by the qemu-kvm processes is way higher than
I?m happy with.  One of the Debian machine causes around
22% while it?s basically idle, the other one is around 3%,
and the Windoze one is around 50.  They are all mostly idle.

I can observe that when some network traffic is going on with
the Windoze machine, it causes a CPU load of 200%.  "Some network
traffic" means that virt-top is showing 1M/2M RX/TX.  Considering
that the bonding interface is theoretically capable of handling
4Gbit full duplex, the 3Mbit are neglectable. Virtio drivers are
being used.

Currently, virt-top shows 1.9% CPU for the Windoze machine and top
shows 22% CPU load for the corresponding qemu-kvm process.  There
is almost not network traffic.  The VM has 4 CPUs assigned.

What may cause the high CPU load?  Something must be seriously
wrong for an idle machine causing 22% CPU load and for the
same machine, still basically idle, with a some network traffic
to cause 200% CPU load on a Xeon 5690.

On 06/02/2017 04:32 AM, hw wrote:
> What may cause the high CPU load? 

Offhand, it's hard to say.  I don't see similar behavior.  Can you post 
the libvirt XML definitions for those VMs somewhere? pastebin maybe?  
What's the output of "rpm -qa qemu\*"?

Gordon Messmer wrote:
> On 06/02/2017 04:32 AM, hw wrote:
>> What may cause the high CPU load?
>
>
> Offhand, it's hard to say.  I don't see similar behavior.  Can you
post the libvirt XML definitions for those VMs somewhere? pastebin maybe? 
What's the output of "rpm -qa qemu\*"?

qemu-img-1.5.3-126.el7_3.6.x86_64
qemu-kvm-tools-1.5.3-126.el7_3.6.x86_64
qemu-kvm-common-1.5.3-126.el7_3.6.x86_64
qemu-kvm-1.5.3-126.el7_3.6.x86_64


The definitions aren?t too long, I could post them here.  There?s
nothing special about them AFAICT; I disabled USB and am trying
to use kvmclock.

I?m finding the number of "Local timer interrupts" suspicious.  From
'cat /proc/interrupts' for CPU0:


Tue Jun  6 20:01:53 CEST 2017: 217433736
Thu Jun  8 13:23:04 CEST 2017: 350172149


That seems an awful lot of interrupts.  Is this normal?

There?s also a huge amount of "Rescheduling interrupts" (102113959
earlier,
now 209740910).  The VMs are pinned to CPUs, so what?s being rescheduled
so frequently?

I can observe that CPU load of the host goes up with increases in network
traffic of the guest.  Is it a bad idea to assign a bonding interface
to a bridge?

>
> _______________________________________________
> CentOS mailing list
> CentOS at centos.org
> https://lists.centos.org/mailman/listinfo/centos