Re: [PATCHv3 0/3] perf tool: Add new event group management

[Namhyung Kim <namhyung@gmail.com>]

2012-07-02 (월), 15:33 +0200, Jiri Olsa:
> On Mon, Jul 02, 2012 at 10:11:02PM +0900, Namhyung Kim wrote:
> > 2012-07-02 (월), 12:15 +0200, Jiri Olsa:
> > > On Mon, Jul 02, 2012 at 10:53:44AM +0900, Namhyung Kim wrote:
> > > > Just a question, is there a way to know about the grouping at perf
> > > > report time?
> > > 
> > > nope, AFAIK only ID and perf_event_attr is stored for event
> > > grouping is known only for record time
> > > 
> > 
> > I heard that Arnaldo (or Stephane) wanted to make perf report
> > group-aware or such so that it can show related events together. But to
> > do that, it seems we need to change the data file format first, right?
> > 
> > Any idea?
> 
> The next change I'm working on is to record and report
> PERF_FORMAT_GROUP related data. Here's commit comment
> from my next patchset:
> 
> ---
> perf, tool: Enable sampling on specified event group leader
> 
> Adding the functionality to the group modifier event syntax.
> Allowing user to select leader event inside the group using
> event index (command line event position in the group).
> 
> Following example selects e2 as leader:
>   -e '{e1,e2,e3,e4}:2'
> 
> The selected event becomes group leader and is the only one
> doing samples.
> 
> The rest of the events in the group are being read on each leader
> event sample by PERF_SAMPLE_READ sample type processing.
> 
> Following example:
>   perf record -e {cycles,faults}:1 ls
> 
>   - creates a group with 'cycles' and 'faults' events
>   - 'cycles' event is group leader and has sampling enabled
>   - 'faults' event is read each time 'cycles' sample,
>     the 'faults' count is attached to the 'cycles sample
>     via PERF_SAMPLE_READ sample type.
> ---
> 
> The report does not need any new metadata about grouping, because
> the samples are generated/stored only from the group leader. The
> other events data are read from the PERF_FORMAT_GROUP leader sample
> data.
> 
> So no data file format change for my next changes, but I'm not sure
> this is the report change you mean.
> 

What I said is just displaying cycles and faults events in above example
on the same screen/table like:

 (Overhead)
cycles  faults  Command      Shared object              Symbol
......  ......  .......  .................  ..................
92.98%  60.13%  noploop  noploop            [.] main
 3.21%   5.42%  noploop  [kernel.kallsyms]  [k] __lock_acquire
 1.16%  13.13%  noploop  libc-2.11.1.so     [.] _int_malloc
 0.97%   0.33%  noploop  [kernel.kallsyms]  [k] clear_page_c
...

We could pass such information to perf report explicitly, But I guess it
'd better if perf report did it for me automagically by detecting group
relations.


-- 
Regards,
Namhyung Kim