lkml.org 
[lkml]   [2009]   [Nov]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: perf stat output

* Lucas De Marchi <lucas.de.marchi@gmail.com> wrote:

> Hi all!
>
> Some questions about perf stat output. See example:
>
>
> lucas@LMS-linux:~/programming/testprograms> perf stat -e
> L1-dcache-loads -e L1-dcache-load-misses -- make -j
> gcc test_schedchanges.c -o test_schedchanges
> gcc -pthread test_taskaff1.c -o test_taskaff1
> gcc -pthread test_taskaff2.c -o test_taskaff2
> gcc -pthread test_taskaff3.c -o test_taskaff3
>
> Performance counter stats for 'make -j':
>
> 161384667 L1-dcache-loads # 0.000 M/sec
> 24853791 L1-dcache-load-misses # 0.000 M/sec
>
> 0.066893389 seconds time elapsed
>
> Why do we have both L1-dcache-loads and L1-dcache-load-misses with
> 0.000 M/sec? Also, why do we have 0 M/s when running "perf stat -a -e
> cache-misses -e cache-references" but values different than 0 when
> running "perf stat -a" without selecting the events?

You need the 'task-clock' event to be able to see M/sec metrics. I.e.:

$ perf stat -e L1-dcache-loads -e L1-dcache-load-misses -e task-clock sleep 1

Performance counter stats for 'sleep 1':

201330 L1-dcache-loads # 566.234 M/sec
29916 L1-dcache-load-misses # 84.138 M/sec
0.355560 task-clock-msecs # 0.000 CPUs

1.000621650 seconds time elapsed

I agree with you that seeing '0.000 M/sec' is confusing and incorrect as
well. One solution would be to skip the printout in that case.

You can find the latest 'perf' code in:

http://people.redhat.com/mingo/tip.git/README

( the tools/perf/ bits are backwards compatible with any perf kernel you
are running right now, so no reboot is needed. )

You can find the stats printing in tools/perf/builtin-stat.c, in the
abs_printout() function:

} else {
total = avg_stats(&runtime_nsecs_stats);

if (total)
ratio = 1000.0 * avg / total;

fprintf(stderr, " # %10.3f M/sec", ratio);

I think if runtime_nsecs_stats is zero (i.e. if no 'task-clock' events
were measured), then we might be able to skip the printout via doing
something like:

} else if (runtime_nsecs_stats.n != 0) {

Would you be interested in sending a (tested) patch for that? In theory
only that oneliner change should suffice - but i have not tested it.

Ingo


\
 
 \ /
  Last update: 2009-11-15 10:17    [W:0.050 / U:0.080 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site