[vfio-users] Significant FPS drop with AMD and GTX970

Wed Oct 12 10:36:58 UTC 2016

Checked DPC and it was ok and turned off the nested paging. Actually it
helped since now I get to 110fps, but that's not stable, it can drop as low
as 35-40. Before the ntp it couldn't go higher than 70. Any tipps on why
the fps are not stable and are dropping so low. Seems to be CPU and/or HDD
related. It mostly happens when there is HDD activity or the CPU needs to
compute more stuff. I am using the virtio SCSI controller and the hdds are
set on SCSI.

On Oct 10, 2016 12:18, "Rokas Kupstys" <rokups at zoho.com> wrote:
>
> I do not think it is powersaving. It could be the case with CPU because
host holds ultimate control of CPU, but GPU is completely controlled by VM
and driver in VM decides at what clock GPU should run.
>
> Maybe CPU is bottleneck here? After all your CPU (and i had similar one)
says it has 8 cores, but in reality every two cores share resources +
single FPU core. On top of that AMD FPUs seem to be not as good as intel
ones so right there could be bottleneck. Also i suggest checking out your
DPC latency - it does cause trouble in some cases.
>
> And i just remembered something that was not listed in first post of that
thread on archlinux forum:
https://bbs.archlinux.org/viewtopic.php?pid=1270311#p1270311 - i think this
actually improved my performance a lot.
>
> P.S. Use [Reply All] or [Reply List] button (whatever is present) when
replying to mailing list messages ;)
>
> Rokas Kupstys
>
> On 2016.10.10 13:07, Andrei Grigore wrote:
>>
>> Ok. So it's clearly not a hardware failure since on native Windows it
just works. The GPU is not the issue since there are reports that the same
GPU performs flawlessly. The CPU is also not an issue, even though most
people are using Intel. The configuration and passthrough procedure is
definatelly right (i even wiped the whole system and did everything all
over again). It's either a drawback with newer kernels/qemu for this MB and
/ or Chipset? It seems to me impossible to debug or to find the root cause.
>>
>> What is certain is that the behavior didn't change after the last couple
of updates, this issue is there for a couple of months now. Just waited and
waited for some performance improvements.
>>
>> It feels to me like the system is not used at it's full strength. I can
run 100% stress CPU tests, the GPU also is used, RAM is plenty, but the
fans are not turbo spinning as on native. I have disabled are power saving
features from the VM and set the host governor to high performance. Can it
be that some wierd power state / saving feature is causing this? Should I
disable something more from BIOS / host?
>>
>>
>> On Oct 10, 2016 10:57, "Rokas Kupstys" <rokups at zoho.com> wrote:
>>>
>>> You know.. What i am going to tell you will not sound good. My board
was Sabertooth 990-fx R2. My little story is that at first it seemed to run
ok (with r9 270X card). I sinked enough hours to titanfall and it was
great. After good period of time of non-usage (as board was in warranty
service and eventually swapped to brand new board), multiple kernel updates
and migration to archlinux i no longer could get it running as fast. Now i
do not know if i just did not feel slowness initially and it was always
there or if something changed, but i know for a fact that testing older
kernel releases did yield a little bit better performance (enough that i
could notice). So i moved back to dual-booting due to performance reasons.
In the end combination of this and r9 270X constantly causing entire PC to
go down with a red screen i simply swapped hardware for intel x99 + nvidia
gpu. Such is sad story of mine..
>>>
>>> Rokas Kupstys
>>>
>>> On 2016.10.10 11:48, Andrei Grigore wrote:
>>>>
>>>> Could it be an issue caused by the MB chipset (990-fx)? How can you
debug something like this?
>>>>
>>>>
>>>> On Oct 8, 2016 15:46, "Andrei Grigore" <andrei.grg at gmail.com> wrote:
>>>>>
>>>>> Pretty sure, yes. Under: /etc/modprobe.d , then rebuilt the initramfs
with mkinitcpio -p linux and rebooted.
>>>>>
>>>>>
>>>>> On Oct 8, 2016 15:44, "Rokas Kupstys" <rokups at zoho.com> wrote:
>>>>>>
>>>>>> Sure you put config in the right place? Because depending on how
kernel modules are compiled config go to different places.
>>>>>>
>>>>>>
>>>>>>
>>>>>> Sent from my Samsung device
>>>>>>
>>>>>>
>>>>>> -------- Original message --------
>>>>>> From: Andrei Grigore <andrei.grg at gmail.com>
>>>>>> Date: 10/8/16 16:14 (GMT+02:00)
>>>>>> To: Rokas Kupstys <rokups at zoho.com>
>>>>>> Cc: vfio-users <vfio-users at redhat.com>, Jayme Howard <
g.prime at gmail.com>
>>>>>> Subject: Re: [vfio-users] Significant FPS drop with AMD and GTX970
>>>>>>
>>>>>> Just tried. No improvement.
>>>>>>
>>>>>>
>>>>>> On Oct 8, 2016 14:43, "Rokas Kupstys" <rokups at zoho.com> wrote:
>>>>>>>
>>>>>>> See https://bbs.archlinux.org/viewtopic.php?id=162768 first post.
See if you need ignore_msrs or allow_unsafe_interrupts. I had similar issue
with and board and some module options fixed it I think.
>>>>>>>
>>>>>>>
>>>>>>> Sent from my Samsung device
>>>>>>>
>>>>>>>
>>>>>>> -------- Original message --------
>>>>>>> From: Andrei Grigore <andrei.grg at gmail.com>
>>>>>>> Date: 10/8/16 10:49 (GMT+02:00)
>>>>>>> To: Jayme Howard <g.prime at gmail.com>
>>>>>>> Cc: vfio-users <vfio-users at redhat.com>
>>>>>>> Subject: Re: [vfio-users] Significant FPS drop with AMD and GTX970
>>>>>>>
>>>>>>> Unfortunately, that didn't help. Thanks for the sugestion though!
>>>>>>>
>>>>>>>
>>>>>>> On Oct 8, 2016 03:06, "Jayme Howard" <g.prime at gmail.com> wrote:
>>>>>>>>
>>>>>>>> Check to see if you have MSI enabled.  That'd be my bet.  Alex had
a blog post that covered it, but I don't have the link handy.
>>>>>>>>
>>>>>>>> On Fri, Oct 7, 2016 at 6:39 PM, Andrei Grigore <
andrei.grg at gmail.com> wrote:
>>>>>>>>>
>>>>>>>>> Hi all,
>>>>>>>>>
>>>>>>>>> I was wondering if anyone is experiencing this and if there is a
known fix for it.
>>>>>>>>>
>>>>>>>>> I am using ArchLinux with latest stable Kernel, qemu and libvirt.
>>>>>>>>>
>>>>>>>>> My CPU is FX-8230 with a GTX 970. I've passed thorugh the GPU to
a Win10 guest.
>>>>>>>>>
>>>>>>>>> In WoW on bare metal i get 90 fps and in the same area with the
same graphic settings on the VM i get around 50. There seems to be no cpu
and or ram bottlenet and the GPU seems to be quite gully used since the
fans are starting to spin quite soon.
>>>>>>>>>
>>>>>>>>> Can anyone help here with a suggestion? I saw reports with about
3% performance drop, but I am experiencing a lot more.
>>>>>>>>>
>>>>>>>>> Thanks!
>>>>>>>>> Andrei.
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> _______________________________________________
>>>>>>>>> vfio-users mailing list
>>>>>>>>> vfio-users at redhat.com
>>>>>>>>> https://www.redhat.com/mailman/listinfo/vfio-users
>>>>>>>>>
>>>>>>>>
>>>>
>>>>
>>>> _______________________________________________ vfio-users mailing list
vfio-users at redhat.com https://www.redhat.com/mailman/listinfo/vfio-users
>>>
>>> _______________________________________________ vfio-users mailing list
vfio-users at redhat.com https://www.redhat.com/mailman/listinfo/vfio-users
>
>
> _______________________________________________
> vfio-users mailing list
> vfio-users at redhat.com
> https://www.redhat.com/mailman/listinfo/vfio-users
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listman.redhat.com/archives/vfio-users/attachments/20161012/73335833/attachment.htm>