Re: [Qemu-devel] Re: [PATCH 2/7] Enable I/O thread and VNC threads by default

[Marcelo Tosatti]

On Tue, Jan 25, 2011 at 11:34:53AM -0200, Marcelo Tosatti wrote:
> On Tue, Jan 25, 2011 at 10:17:41AM +0100, Edgar E. Iglesias wrote:
> > On Mon, Jan 24, 2011 at 04:28:48PM -0600, Anthony Liguori wrote:
> > > On 01/24/2011 03:00 PM, Anthony Liguori wrote:
> > > > Leave the disable options for now to help with testing but these will 
> > > > be removed
> > > > once we're confident in the thread implementations.
> > > >
> > > > Disabled code bit rots.  These have been in tree long enough that we 
> > > > need to
> > > > either commit to making them work or just remove them entirely.
> > > >    
> > > 
> > > I/O thread disables icount apparently.
> > > 
> > > I'm not really sure why.  Marcelo, do you know the reason 
> > > qemu_calculate_timeout returns a fixed value in the I/O thread 
> > > regardless of icount?
> > 
> > Hi,
> > 
> > The following commit hopefully fixed that issue.
> > 
> > commit 225d02cd1a34d5d87e8acefbf8e244a5d12f5f8c
> > Author: Edgar E. Iglesias <address@hidden>
> > Date:   Sun Jan 23 04:44:51 2011 +0100
> > 
> >     Avoid deadlock whith iothread and icount
> >     
> >     When using the iothread together with icount, make sure the
> >     qemu_icount counter makes forward progress when the vcpu is
> >     idle to avoid deadlocks.
> >     
> >     Signed-off-by: Edgar E. Iglesias <address@hidden>
> > 
> > See http://lists.gnu.org/archive/html/qemu-devel/2011-01/msg01602.html
> > for more info.
> > 
> > One more thing I didn't mention on the email-thread or on IRC is
> > that last time I checked, qemu with io-thread was performing
> > significantly slower than non io-thread builds. That was with
> > TCG emulation (not kvm). Somewhere between 5 - 10% slower, IIRC.

Can you recall what was the test ?

> > Also, although -icount & iothread no longer deadlocks, icount
> > still sometimes performs incredibly slow with the io-thread (compared
> > to non-io-thread qemu). In particular when not using -icount auto but
> > a fixed ticks per insn values. Sometimes it's so slow I thought it
> > actually deadlocked, but no it was crawling :) I haven't had time
> > to look at it any closer but I hope to do soon.
> > 
> > These issues should be fixable though, so I'm not arguing against
> > enabling it per default. Just mentioning what I've seen FYI..
> 
> Right, remember seeing 20% added overhead for network copy with TCG on
> the initial iothread merge.

This is not the case anymore, network transfer speed is comparable.
Probably due to SIG_IPI delivery being reliable, which was fixed later.