Re: [PATCH 00/11] perf tool: Add PERF_SAMPLE_READ sample read support

[Jiri Olsa <jolsa@redhat.com>]

On Sun, Oct 21, 2012 at 06:38:49PM +0200, Ingo Molnar wrote:
> 
> * Jiri Olsa <jolsa@redhat.com> wrote:
> 
> > hi,
> > adding support to read sample values through the PERF_SAMPLE_READ
> > sample type. It's now possible to specify 'S' modifier for an event
> > and get its sample value by PERF_SAMPLE_READ.
> > 
> > For group the 'S' modifier will enable sampling only for the leader
> > and read all the group member by PERF_SAMPLE_READ smple type with
> > PERF_FORMAT_GROUP read format.
> > 
> > This patchset is based on group report patches by Namhyung Kim:
> > http://lwn.net/Articles/518569/
> > 
> > Example:
> > (making sample on cycles, reading both cycles and cache-misses
> >  by PERF_SAMPLE_READ/PERF_FORMAT_GROUP)
> > 
> >   # ./perf record -e '{cycles,cache-misses}:S' ls
> >   ...
> > 
> >   # ./perf report --group --show-total-period --stdio
> >   # ========
> >   # captured on: Sat Oct 20 16:53:39 2012
> >   ...
> >   # group: {cycles,cache-misses}
> >   # ========
> >   #
> >   # Samples: 86  of event 'anon group { cycles, cache-misses }'
> >   # Event count (approx.): 34863674
> >   #
> >   #         Overhead                    Period  Command      Shared Object                            Symbol
> >   # ................  ........................  .......  .................  ................................
> 
> Might make sense to consider this column enumeration:
> 
>     #
>     #   cycles
>     #   |       cache-misses
>     #   |       |
> >   #   v       v
> >   #
> >       16.56%  19.47%       5773450         475       ls  [kernel.kallsyms]  [k] native_sched_clock          
> >       10.87%   0.74%       3789088          18       ls  [kernel.kallsyms]  [k] rtl8169_interrupt           

no problem in '--stdio' mode I guess.. not sure in '--tui/--gtk', Namhyung?


> >        9.82%  15.86%       3423364         387       ls  [kernel.kallsyms]  [k] mark_lock                   
> >        8.43%  17.75%       2938384         433       ls  ld-2.14.90.so      [.] do_lookup_x                 
> >        6.79%  20.86%       2365622         509       ls  ls                 [.] calculate_columns           
> >        6.36%   0.61%       2216808          15       ls  [kernel.kallsyms]  [k] lock_release                
> >   ...
> 
> /me wants this feature ASAP :-)
> 
> This should probably be the out of box default for perf record 
> and perf top as well - the cache miss rate is probably one of 
> the least appreciated aspects of overhead analysis.
> 
> Does it have sane output if the cache-misses event is not 
> supported? The cache-misses column should probably stay empty in 
> that case - basically falling back to today's default output.

unsupported counters fail before report:

  # ./perf record -e '{cycles,stalled-cycles-backend}:S' ls
  Error:
  The stalled-cycles-backend  event is not supported.
  ls: Terminated

similar for top, so we'd need special treatment for default

thanks,
jirka