RE: XEN 4.1 pci passthru performance regression?

[Ian Campbell <Ian.Campbell@citrix.com>]

On Wed, 2011-06-01 at 08:55 +0100, Ian Campbell wrote:
> On Wed, 2011-06-01 at 02:41 +0100, Cao, Buddy wrote:
> > Thanks Ian for you reply.
> >
> > >Are you referring only to control operation performance or also to
> > >actual guest performance?
> >
> > It does not impact actual guest performance, only make vm start and shutdown slow.
> 
> Oh good!
> 
> > >Do you find that guests without any PCI passthrough have also slowed
> > >down or is the issue purely for guests with a passed-through device?
> >
> > The guests without any PCI passthrough works as well as xen4.0.2.
> 
> Also good.
> 
> > I tired xl, but looks it always returns "libxl....." relevant error,
> > and none of the command can run properly
> 
> Well, we should also try and resolve this -- xl is expected to work in
> 4.1. Did you see: http://wiki.xen.org/xenwiki/MigrationGuideToXen4.1+ ?
> 
> Please tell us the specifics of the xl failure.
> 
> I tried starting a PV guest with a passed through device with both xm
> and xl and there was no obvious delay on startup or shutdown.
> Unfortunately I don't have a test box which can do passthrough to HVM
> guests at the moment (I'll see if I can scare one up). In the meantime
> perhaps you could add some extra logging to xend to see where the time
> is spent and/or report the xl failues as requested above.
> 
> Also, please can you try xen-unstable.hg to see if the problem still
> exists there.

I found a suitable box and tested xen-unstable.hg.

xm does seem to suffer from a delay of some sort on startup, but xl
doesn't seem to.

I'm afraid I don't have time to dig into xend any further myself.

Ian.