Re: perf top -g -U --sort=symbol --children == lalalalala?

[Arnaldo Carvalho de Melo <acme@kernel.org>]

Em Fri, Sep 12, 2014 at 04:24:54PM +0900, Namhyung Kim escreveu:
> Hi Arnaldo and Mike,
> 
> On Thu, 11 Sep 2014 10:43:38 -0300, Arnaldo Carvalho de Melo wrote:
> > Em Thu, Sep 11, 2014 at 02:43:12PM +0200, Mike Galbraith escreveu:
> >> On Thu, 2014-09-11 at 17:09 +0900, Namhyung Kim wrote: 
> >> > On Tue, Sep 9, 2014 at 10:37 PM, Arnaldo Carvalho de Melo
> >> > <acme@ghostprotocols.net> wrote:
> >> > > Em Tue, Sep 09, 2014 at 01:54:01PM +0200, Mike Galbraith escreveu:
> >> > >> Seems the now default on --children thingy doesn't like -U much.
> >> > >
> >> > > Namhyung, can you please take a look at this?
> >> > 
> >> > So what is the problem here?
> >> 
> >> Well, if you don't see anything wrong, I guess nothing at all.
> >
> > :-)
> >
> > I think that when we decide that it is so better to change defaults like
> > we did this time, we should be required to add a big fat warning (a
> > --tui popup, use the first lines on --stdio, etc) about why the default
> > was changed and allow quick, easy opt out, restoring previous behaviour
> > after the user, being warned, knows what to expect, tries it, and then
> > is in a better position to decide if keeping the new default is what is
> > desired.
> 
> So the problem is that why it turned on --children option by default,
> right?  I thought you mentioned there's a problem with -U option and I
> couldn't figure out what it is.

Yeah, at that point I thought the problem was that -U wasn't honoured in
--children mode, but after I looked again at the output, it was all [k],
so I thought that it was something else but didn't got back to the
thread :-)
 
> >> > >> Samples: 5K of event 'cycles', Event count (approx.): 2268660922
> >> > >>   Children      Self  Symbol
> >> > >> +   46.42%     0.04%  [k] system_call_fastpath
> >  
> >> I'll just turn it off until I figure out what cool stuff this is telling
> >> me. why that symbol becomes the number one hit, and why total% > 100.
> >  
> >> To me, it looks like top smoked it's breakfast, went to lala land ;-)
> >
> > Yeah, its confusing, I'll let Namhyung explain it ;-)
> 
> Now I have three persion yell at me for this change. :)

Hey, its not "yell" or something, I just think that it is under
documented, and hey, perf is world renowned for being under developed,
we don't need any further efforts in that area :-P
 
> When this change was developping, Ingo said it'd be better if it looks
> like output of sysprof as it's more popular for most (userland?) guys.
> You can see the discussion in the following links:
> 
>   https://lkml.org/lkml/2013/10/31/97
>   https://lkml.org/lkml/2013/11/1/85
> 
> The children field is a cumulative total overhead (for its all
> children/callee) so sum of them would be more than 100%.  And as Ingo
> requested it sorts the output entries using children overhead so that
> one can easily see higher level view of performance bottle-neck.

yeah, yeah, but please have all this in the documentation :-)

- Arnaldo