[vfio-users] cpu usage in guest != cpu usage in host, even with exclusive pinning

Okky Hendriansyah okky.htf at gmail.com
Mon Apr 4 18:17:37 UTC 2016

On Mon, Apr 4, 2016 at 8:20 AM, Jayme Howard <g.prime at gmail.com> wrote:

> Okay wow, Alex.  That made a huge difference immediately.  That dropped me
> from 90% to 33% on the main thread, and all the children are below 10%.

> On Sun, Apr 3, 2016 at 8:07 PM, Alex Williamson <
> alex.l.williamson at gmail.com> wrote:
>> If this is the issue I think it is, the unraid folks bisected this and
>> found the culprit as:
>> aca6ff2 KVM: dynamic halt-polling
>> (
>> https://github.com/torvalds/linux/commit/aca6ff29c4063a8d467cdee241e6b3bf7dc4a171
>> )
>> It seems to be a poor interaction of the halt polling interval vs the
>> timer ticks on win10.  You can pick a different polling interval with the
>> kvm module option halt_poll_ns.  The default is 500000.  It seemed that
>> setting this to 400000 or lower resolves the issue.  You can do this via a
>> modprobe entry, 'options kvm halt_poll_ns=400000' or on the kernel command
>> line with kvm.halt_poll_ns=400000.  You can also change it dynamically via
>> 'echo 400000 > /sys/module/kvm/parameters/halt_poll_ns'
I can also confirm that tuning the halt_poll_ns kvm module parameter on
kernel 4.4.5 drops the host CPU usage! Haven't done any game benchmark but
from listening to music the CPU load is as low as it is on 4.1.20. Although
reducing the value to what Alex (400000) had suggested is probably
sufficient, I'm currently trying out the value of 0 for it just like what
Unraid 6.20 beta 20 changelog had it set on default [1]. Thanks for the
info, Alex.


Best regards,
Okky Hendriansyah
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listman.redhat.com/archives/vfio-users/attachments/20160405/d1232fa3/attachment.htm>

More information about the vfio-users mailing list