Re: [PATCH] perf_counter: Make software counters work as per-cpu counters

[Ingo Molnar <mingo@elte.hu>]

* Zhang, Yanmin <yanmin_zhang@linux.intel.com> wrote:

> > > Signed-off-by: Paul Mackerras <paulus@samba.org>
> I tested it. The patch does fix the crash.

Great!

> > ok - i've applied your patch from email to allow Yanmin to test tip:master.
> Ingo,
> 
> Here is the patch to add system-wide collection for perfstat.c.
> 
> With the new parameter '-s', perfstat introduces very little overhead to benchmark 
> testing. [...]

Very nice. I've applied your patch to perfstat.c and have uploaded the new version 
to the usual place:

   http://redhat.com/~mingo/perfcounters/perfstat.c

> [...] If tracking performance data without '-s', sometimes perfstat has too much 
> impact on benchmark results, especially on netperf UDP-U-4k.

Yes - and we need to optimize that. It wont ever be zero-cost though, so -s makes 
sense independently of any need for optimizations. -s can also be used to measure 
the overhead of perfstat per task counters itself, via recursive inherited counters 
like this:

   perfstat -s perfstat ./workload

Thanks,

	Ingo