RE: exit timing analysis v1 - comments&discussions welcome

[Hollis Blanchard <hollisb@us.ibm.com>]

On Thu, 2008-09-25 at 17:32 +0800, Liu Yu-B13201 wrote:
> > 
> > On Wed, 2008-09-24 at 11:24 +0200, Christian Ehrhardt wrote:
> > >                      count      min         max             
> > > sum            avg        stddev     time%
> > >     EMULINST     6,423,699       12     247,582      
> > 91,732,705        
> > > 14.2804       241.200     45.36
> > >     ITLBVIRT     1,777,242       21     264,257      
> > 47,116,557        
> > > 26.5111       286.040     23.30
> > >     DTLBVIRT     1,544,241       22     263,947      
> > 41,765,447        
> > > 27.0459       218.997     20.65
> > 
> > (The max here is of course when the guest was de-scheduled in favor of
> > another host process.)
> > 
> > I think it's interesting that the min instruction emulation time is 12
> > usecs. In fact, our "null" exits where we do almost no 
> > processing (ISI,
> > DSI, syscall) are 11 usecs, which suggests we have a problem with
> > interrupt handler overhead (for all exit types).
> > 
> 
> Will you consider about moving tlb manipulation out of entering path?
> Examining the modify array may cost some time.

Yes, definitely.

That's about the only thing I can see in the 440 interrupt path that
might take significant time. Unfortunately we can't profile that code
because we have no performance counters, so finding the problem will
require some experimentation.

-- 
Hollis Blanchard
IBM Linux Technology Center