[vfio-users] Calling ALL VM experts - Seeking assistance with AMD FX CPUs

[R]

First time doing this mailing list thing, hopefully I'm doing it right.

Either way, myself and an acquaintance are both using GPU pass-through
methods to do a bit of gaming on our Linux machines, both running Arch with
qemu-KVM with a Win10 guest. We've both come across an issue with a
particular CPU, namely AMD's FX-8320. The best way to describe it would be
the "rubber banding" experienced in online gaming. We both get mostly
stable fps, however we both experience this odd effect. The best way it can
be described would be as if something hangs up, then suddenly lets go,
causing the game to rush forward to where it should be. It's very much like
latency in an online match when it suddenly spikes and you see other
players rush to where they actually are, only offline and in any game.

The current theory has to do with the 8320's FPUs and how they interact
with CPU pinning, but I think it's more some kind of I/O issue. At least,
it seems like there's a *bottleneck *somewhere in the system or virtual
environment causing this any time data is requested. By what? I don't know.
Said acquaintance is a bit more experienced in this than I am and has tried
various methods (I don't know the extent) with IRQ nonsenses, where I think
it may lie to some degree. I simply have the drunken charisma and way with
words to bring our problem to light.

Either way, it's incredibly annoying. We keep two CPUs for the host, the
other six are pinned to the VM for whatever it needs though it seems that
the less cores you apply to the guest, the less the effect is noticed. Most
everything that can be run on virtio is, and hugepages make no difference.
What's more, acquaintance is running a 970 chipset and I'm running a 990FX.
Both our boards properly support this venture. As a clutch, I made sure
some of the basics with virtio is run with MSI enabled and used, but I've
only been at this a week. Thus, I'm asking for experts in the field to help
out.

I've got this far in this, I'd hate to hear it's just the nature of the AMD
beast and have to chalk it up to time mostly well sepent. If it is indeed
an FPU issue with the CPU, it makes no sense to me that it'd only appear in
a virtual environment and not on metal.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listman.redhat.com/archives/vfio-users/attachments/20170129/f02efeef/attachment.htm>