[vfio-users] Help Patching Qemu for Vendor

[Okky Hendriansyah]

Hi Spam,

On Mon, May 9, 2016 at 5:28 AM, Spam House <kthxplzdie at gmail.com> wrote:

> ...
> Is it correct that if my hyperv is spoofed properly I no longer need this
> in my xml?
>
>     <kvm>
>       <hidden state='on'/>
>     </kvm>..
>
> As far as I know, we still need to hide the KVM CPUID from Windows 10
guest and to use the custom Hyper-V vendor ID in order to make NVIDIA GPU
work with recent driver version.


On Sun, May 8, 2016 at 5:10 PM, Alex Williamson <alex.williamson at redhat.com>
> wrote:
>
>> On Sun, 8 May 2016 16:35:39 -0500
>> Spam House <kthxplzdie at gmail.com> wrote:
>>
>> > Thibaut,
>> >
>> > Thanks for the info. The performance difference I am measuring with
>> 3dmark
>> > firestrike.
>> >
>> > Unraid Win10 VM (QEMU/KVM):
>> > Overall Score = 14037
>> > Graphics = 17152
>> >
>> > Arch Win10 VM (QEMU/KVM)
>> > Overall Score  = 12085
>> > Graphics = 13658
>> >
>> > I have not fully implemented huge pages yet but the CPU performance
>> seems
>> > to be about the same. The gpu specifically has greatly reduced
>> performance.
>>
>>
>> Hyper-v features are known not to make much difference for 3dmark.
>> Tune the CPU type, CPU topology, pin vCPUs, and use hugepages.
>>
>
I am using similar GPU as yours (MSI GTX 980 Ti Gaming 6G, with slightly
overclock +100MHz GPU clock and +100MHz VRAM clock) and I also use Arch
Linux as the hypervisor and I have scored 19332 in Fire Strike Graphics [1].

I am using stock Arch Linux kernel 4.5.2-1-ARCH with modified config:
1.) 1000 MHz clock resolution (eliminate mini stutters on Witcher 3) [2]
2.) Voluntary preemption scheduling (I use ZFS and it prefers this) [2][3]
3.) Madvise hugepages (skippable, but seems to help keeps RAM usage of
other host processes lower than using THP)
4.) CRIU config (skippable, absolutely not related so just skip this) [4]

If you are using kernel between 4.2.x and 4.5.1 (inclusive), you should
reduce halt_poll_ns parameter manually [5]. As Alex suggested, use explicit
hugepages to improve performance. I did not do any vCPU pinning and my CPU
topology resembles the host.

I am not using libvirt though, just plain old QEMU script. [6]

[1] http://www.3dmark.com/3dm/11946102
[2] https://bbs.archlinux.org/viewtopic.php?pid=1270311#p1270311
[3] https://github.com/zfsonlinux/zfs/issues/1701
[4] https://criu.org/Installation#Configuring_the_kernel
[5] https://www.redhat.com/archives/vfio-users/2016-April/msg00017.html
[6] http://pastebin.com/GHBCrR74

Best regards,
Okky Hendriansyah
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listman.redhat.com/archives/vfio-users/attachments/20160509/392a31c6/attachment.htm>