Re: [RFC][PATCH 5/5] perfcounter: Add support for kernel hardwarebreakpoints

[Frederic Weisbecker]

On Wed, Jul 29, 2009 at 10:28:52AM +0200, Peter Zijlstra wrote:
> On Wed, 2009-07-29 at 02:36 +0200, Frederic Weisbecker wrote:
> > On Tue, Jul 28, 2009 at 09:24:37AM +0200, Peter Zijlstra wrote:
> > > On Tue, 2009-07-28 at 03:03 +0200, Frederic Weisbecker wrote:
> > > > > Thirdly, we can multiplex perf counters beyond their hardware maximum,
> > > > > something you simply cannot do for a debug interface.
> > > > 
> > > > 
> > > > Again, I'm stuck in what you mean by multiplexing here :-)
> > > 
> > > If you'd create say 16 breakpoint counters, we'd RR them over the 4
> > > available hardware breakpoints (or less when others are taken by someone
> > > else).
> > > 
> > > Since its all statistics anyway, we can simply scale the event counts up
> > > by the time-share they received.
> > >
> > 
> > Aah, ok I understand now.
> > But I fear it may kill the accuracy of the breakpoints statistics.
> > It's fine for a theoretical linear rate of breakpoint events.
> > But what happens if we are profiling something much more unstable
> > with a rain of hits in a small window of memory between large timeframes?
> > If we have a breakpoint inside this window of memory and this breakpoint
> > is not plugged, waiting for its turn in the RR, we loose this rain of events.
> 
> You don't have to overload the breakpoint set, but we will if you create
> more than there are hardware resources available.
> 
> If the RR period is independent of the application and significantly
> shorter than the sample duration, statistics will be usable.
> 
> Sure, they'll not be perfect, but we can improve our confidence interval
> by running longer. At some point it really doesn't matter anymore.
> 
> Take your example, of small bursts of events, due to the first
> assumption - the RR period being independent of the application - we're
> bound to see those busts in a proportional number of sample windows.
> 
> If the full sample duration is in the same order of the RR period, then
> yes, you'll have a fair chance of missing the burst -- don't do that
> then :-)
> 
> If you know you'll be running very very short, use attr.pinned, to
> request a counter that'll not be multiplexed.


Ok, I'm convinced :-)

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/