Re: [PATCH 2/3] perf tools: Spare double comparison of callchainfirst entry

From: Frederic Weisbecker
Date: Thu Jan 16 2014 - 12:35:12 EST


On Thu, Jan 16, 2014 at 10:17:53AM +0900, Namhyung Kim wrote:
> I think if the sort key doesn't contain "symbol", unmatch case would be
> increased as more various callchains would go into a same entry.

You mean -g fractal,0.5,callee,address ?

Hmm, actually I haven't seen much difference there.

> >
> >>
> >> >
> >> > This results in less comparisons performed by the CPU.
> >>
> >> Do you have any numbers? I suspect it'd not be a big change, but just
> >> curious.
> >
> > So I compared before/after the patchset (which include the cursor restore removal)
> > with:
> >
> > 1) Some big hackbench-like load that generates > 200 MB perf.data
> >
> > perf record -g -- perf bench sched messaging -l $SOME_BIG_NUMBER
> >
> > 2) Compare before/after with the following reports:
> >
> > perf stat perf report --stdio > /dev/null
> > perf stat perf report --stdio -s sym > /dev/null
> > perf stat perf report --stdio -G > /dev/null
> > perf stat perf report --stdio -g fractal,0.5,caller,address > /dev/null
> >
> > And most of the time I had < 0.01% difference on time completion in favour of the patchset
> > (which may be due to the removed cursor restore patch eventually).
> >
> > So, all in one, there was no real interesting difference. If you want the true results I can definetly relaunch the tests.
>
> So as an extreme case, could you please also test "-s cpu" case and
> share the numbers?

There is indeed a tiny difference here.

Before the patchset:

fweisbec@Aivars:~/linux-2.6-tip/tools/perf$ sudo ./perf stat -r 20 ./perf report --stdio -s cpu > /dev/null

Performance counter stats for './perf report --stdio -s cpu' (20 runs):

3343,047232 task-clock (msec) # 0,999 CPUs utilized ( +- 0,12% )
6 context-switches # 0,002 K/sec ( +- 3,82% )
0 cpu-migrations # 0,000 K/sec
128 076 page-faults # 0,038 M/sec ( +- 0,00% )
13 044 840 323 cycles # 3,902 GHz ( +- 0,12% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
16 341 506 514 instructions # 1,25 insns per cycle ( +- 0,00% )
4 042 448 707 branches # 1209,211 M/sec ( +- 0,00% )
26 819 441 branch-misses # 0,66% of all branches ( +- 0,09% )

3,345286450 seconds time elapsed ( +- 0,12% )

After the patchset:

fweisbec@Aivars:~/linux-2.6-tip/tools/perf$ sudo ./perf stat -r 20 ./perf report --stdio -s cpu > /dev/null

Performance counter stats for './perf report --stdio -s cpu' (20 runs):

3365,739972 task-clock (msec) # 0,999 CPUs utilized ( +- 0,12% )
6 context-switches # 0,002 K/sec ( +- 2,99% )
0 cpu-migrations # 0,000 K/sec
128 076 page-faults # 0,038 M/sec ( +- 0,00% )
13 133 593 870 cycles # 3,902 GHz ( +- 0,12% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
16 626 286 378 instructions # 1,27 insns per cycle ( +- 0,00% )
4 119 555 502 branches # 1223,967 M/sec ( +- 0,00% )
28 687 283 branch-misses # 0,70% of all branches ( +- 0,09% )

3,367984867 seconds time elapsed ( +- 0,12% )


Which makes about 0.6% difference on the overhead.
Now it had less overhead in common cases (default sorting, -s sym, -G, etc...).
I guess it's not really worrysome, it's mostly unvisible at this scale.
--
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/