Re: Perf report --percentage

[Milian Wolff <milian.wolff@kdab.com>]

[-- Attachment #1: Type: text/plain, Size: 3397 bytes --]

On Sonntag, 12. Juni 2016 22:55:26 CEST Namhyung Kim wrote:
> Hi Milian,
> 
> On Sat, Jun 11, 2016 at 01:53:34PM +0200, Milian Wolff wrote:
> > On Samstag, 11. Juni 2016 17:31:48 CEST Namhyung Kim wrote:
> > > Hello,
> > > 
> > > On Wed, Jun 08, 2016 at 04:21:43PM +0000, Sebastian, Shiny wrote:
> > > > Thanks Milian. I added -g graph,0 and it did add lot more percentages
> > > > -
> > > > it's still not adding up though.
> > > > 
> > > > #perf report -C 1 --no-children -g graph,0
> > > 
> > > I think it's because you used -C option to restrict samples only from
> > > cpu 1.  One way to check this is adding 'cpu' to sort keys and compare
> > > sample count and callchain count.
> > > 
> > > What is the result of this command?
> > > 
> > >   # perf report -C 1 --no-children -g graph,0,count -n -s +cpu
> > 
> > If that is the case, how would you explain my findings? Note how I did not
> > pass `-C 1`, and still missed up to 90% of backtraces:
> > 
> > $ perf report --no-children -g fractal,0
> > 
> > -    9.47%  kwrite           ld-2.23.so                     [.]
> > do_lookup_x
> > 
> >    + 23.31% do_lookup_x
> > 
> > -    7.06%  kwrite           ld-2.23.so                     [.] strcmp
> > 
> >    + 9.06% strcmp
> > 
> > +    5.71%  kwrite           ld-2.23.so                     [.]
> > check_match
> 
> Hmm.. I have no idea.  Could you please show me the count of
> callchains with following command?
> 
>   $ perf report --no-children -g graph,0,count -n

-   12.43%           252  kwrite           ld-2.23.so                     [.]
do_lookup_x
     51 do_lookup_x
      + _dl_lookup_symbol_x
-   10.46%           223  kwrite           ld-2.23.so                     [.] 
strcmp
   + 37 strcmp
-    8.66%           175  kwrite           ld-2.23.so                     [.] 
check_match
     44 check_match
        do_lookup_x
      + _dl_lookup_symbol_x

Note that I could not reproduce that on a different machine of mine.

> Also it'd be great if you double check the output of 'perf script' to
> ensure all the samples have callchains..

Quite a few samples don't have callchains:

kwrite 14838 31462.912726:     222759 cycles:u: 

kwrite 14838 31462.913158:     207623 cycles:u: 

kwrite 14838 31462.913428:     196656 cycles:u: 

kwrite 14838 31462.913532:     194832 cycles:u: 

kwrite 14838 31462.913714:     229283 cycles:u: 

kwrite 14838 31462.913895:     239871 cycles:u: 

kwrite 14838 31462.914303:     234688 cycles:u: 

kwrite 14838 31462.914384:     223273 cycles:u: 

kwrite 14838 31462.914533:     281279 cycles:u: 

So this quite possibly is the issue in my case.

Could we improve `perf report` to show a "<no stack>" entry or similar? I.e. 
something like this:

-   12.43%           252  kwrite           ld-2.23.so                     [.]
do_lookup_x
     51 do_lookup_x
      + _dl_lookup_symbol_x
    201 <no stack>
-   10.46%           223  kwrite           ld-2.23.so                     [.] 
strcmp
   + 37 strcmp
    186 <no stack>
-    8.66%           175  kwrite           ld-2.23.so                     [.] 
check_match
     44 check_match
        do_lookup_x
      + _dl_lookup_symbol_x
    131 <no stack>

I'll have a look at implementing that. Should be simple enough, I hope.

Bye
-- 
Milian Wolff | milian.wolff@kdab.com | Software Engineer
KDAB (Deutschland) GmbH&Co KG, a KDAB Group company
Tel: +49-30-521325470
KDAB - The Qt Experts

[-- Attachment #2: smime.p7s --]
[-- Type: application/pkcs7-signature, Size: 5903 bytes --]