Re: xennet performance collapses with multiple vCPU

[Andrew Doran <ad%netbsd.org@localhost>]

On Mon, Apr 13, 2020 at 09:17:44PM +0200, Manuel Bouyer wrote:
> On Mon, Apr 13, 2020 at 06:16:59PM +0200, Manuel Bouyer wrote:
> > Hello,
> > 
> > On Mon, Apr 13, 2020 at 04:01:39PM +0000, Andrew Doran wrote:
> > > [...]
> > > This confirms my suspicion and is why I wanted to play with HZ.  I think
> > > soft interrupt processing is being driven off the clock interrupt.  Maybe
> > > hardware interrupt processing but I think that's less likely.  Native x86
> > > configured the same way does not behave like this.  Hmm.  I will play around
> > > with it some more.
> > 
> > I did some checks in this area; including KASSERT() in idle_block() that
> > that ci_ipending is 0 before and after sleeping.
> > 
> > I wonder if it could be an issue with IPIs instead.
> 
> Adding timestamps to differents events and lowering HZ to 20, I see that it
> is the xennet's event which is delivered late. IPIs and softints works
> as expected. So it seems that the topology change interracts with events
> in some way.
> 
> Still investigating ...

Ok that's interesting.  I've been all over the IPI code a few times now, and
tried different things to exclude problems with it, and have found no issue
there.

Andrew