Re: [patch] perf_counter tools: remove expensive old debug code from perf top

[Arnaldo Carvalho de Melo <acme@redhat.com>]

Em Tue, Oct 13, 2009 at 04:41:59PM +0200, Mike Galbraith escreveu:
> On Tue, 2009-10-13 at 15:37 +0200, Ingo Molnar wrote:
> > * Mike Galbraith <efault@gmx.de> wrote:
> > 
> > > perf_counter tools: remove expensive old debug code from perf top
> > > 
> > > Calling gettimeofday() at high frequency is painful for handicapped 
> > > boxen. The spot calling gettimeofday() is old unneeded debug code, so 
> > > remove it.
> > 
> > Thanks!
> > 
> > We still seem to have a performance problem. Just running perf top on a 
> > 16-way box:
> > 
> >  Performance counter stats for 'perf top -e cycles -c 3000000':
> > 
> >      585.694831  task-clock-msecs         #      0.113 CPUs 
> >           35163  context-switches         #      0.060 M/sec
> >              17  CPU-migrations           #      0.000 M/sec
> >           20355  page-faults              #      0.035 M/sec
> >      1476952962  cycles                   #   2521.711 M/sec
> >       730770658  instructions             #      0.495 IPC  
> >        11489471  cache-references         #     19.617 M/sec
> >         2055001  cache-misses             #      3.509 M/sec
> > 
> >     5.169518576  seconds time elapsed
> > 
> > that's 11% of CPU time used on a single CPU - just displaying a measly 
> > 16K irqs/sec. Something's not quite right here.
> 
> I'll try to figure out why after I do some more wakeup preempt testing.
> 
> btw, something broke top annotation (again).  Module symbols still
> annotate, but vmlinux symbols went south.

I'll check this one, and I'm profiling perf top as well.

- Arnaldo