[Spice-devel] SPICE Fedora 15 guest X running at 100%

[Alon Levy]

On Fri, Jul 01, 2011 at 04:29:50PM -0400, John A. Sullivan III wrote:
> On Thu, 2011-06-16 at 01:39 +0200, Alon Levy wrote:
> > On Wed, Jun 15, 2011 at 04:08:43PM -0400, John A. Sullivan III wrote:
> > > Since we were having some trouble as just outlined on our Windows tests,
> > > we thought we would let SPICE put its best foot forward and try a Fedora
> > > 15 guest running on a Fedora 15 KVM host.
> > 
> > Heh, seeing as windows support is much more advanced that is not quite how
> > I'd put it. Certainly we are working on putting linux on equal footing, but
> > the most work so far has gone into the windows driver, not the linux one.
> > 
> > > 
> > > When it worked, it was amazing.  However, most of the time, the system
> > > was barely responsive and the X process was consuming 100% of the CPU.
> > > We initially thought this might be from KDE4 so we installed twm and
> > > experienced the same.  We then launched a few applications without any
> > > Windows Manager at all and saw the same results.
> > > 
> > > Alon was helpful on IRC and mentioned that it was because there was no
> > > kernel module for the driver.
> > 
> > Maybe that seemed implied, but I didn't mean it like that. I just mentioned
> > this in passing, that a kernel module doesn't exist. The main thing we could
> > use a kernel module for is interrupt support. But spice works without that as
> > well, since the communication is done asynchronously most of the time from
> > host to guest (and this is the only place where an interrupt is useful - to
> > wakeup the guest occasionally).
> > 
> > > 
> > > Does this mean that there is no driver for the QXL driver and thus it
> > > runs in user space and drives up the utilization? If so, what are people
> > > doing who are running this in production?
> > > 
> > > This leads to another question.  Our understanding is that rendering is
> > > done on the client and not the guest unless the client is unable to do
> > > so (haven't read enough on the protocol to understand how this is
> > > determined).  Does this mean that, in cases where rendering is happening
> > > on the guest that a high end graphics card in the physical host would
> > > improve performance? Our experience with using NX is that the physical
> > > hardware is never involved but that is a completely different paradigm.
> > > 
> > 
> > Rendering is done on the client always. It is also done on the server if
> > the guest requires the results of rendering, which can happen for instance
> > when you do a print screen.
> > 
> > > If the rendering is taking place on the client, why is the lack of a
> > > kernel module for QXL causing a problem? Thanks - John
> > 
> > Not the problem. The problem is simply in the X driver, and perhaps you can supply
> > some more details to allow to reproduce the 100% cpu scenario?
> <snip>
> Hello, all.  This is still an issue for us.  What additional information
> can we provide to help resolve this problem? Thanks - John
> 
Well, I want to reproduce this, so I guess if you could point out something specific,
what is the client hardware, which client are you running, what is the guest doing?
Also, could you try both spicec and spicy/vinagre?