Re: [PATCH 1/3] perf tools: Record total sampling time

[Namhyung Kim <namhyung@kernel.org>]

2013-12-02 (월), 13:57 +0100, Ingo Molnar:
> So basically, in the end I think it should be possible to have the 
> following behavior:
> 
>    perf record -a -e cycles sleep 1
> 
>    perf report stat              # Reports as if we ran: 'perf stat -a -e cycles sleep 1'
>    perf report                   # Reports the usual histogram
> 
>    perf report --stat            # Reports the perf stat output and the histogram
> 
> or so.

I don't think we need both of 'perf report stat' and 'perf report
--stat'.  At least it looks somewhat confusing to users IMHO.

For perf report stat usage, I think there's not much thing we can do for
a single event - the most case.  We can simple show total count and
elapsed (or sampled time) for the event, but it's already in the header
with this patch.

      # Samples: 4K of event 'cycles'
      # Event count (approx.): 4087481688
      # Total sampling time  : 1.001260 (sec)


If an user really want to see perf stat-like output (without the usual
histogram) for a recorded session, it'd be better to have 'perf record
--stat' do the job (like git diff --stat) IMHO.

> 
> i.e. a perf.data file would by default always carry enough information 
> to enable the extraction of the 'perf stat' data.
> 
> At that point visualizing it is purely report-time logic, it does not 
> need any record-time options.
> 
> This would work for multi-event sampling as well, if we do:
> 
>    perf record -a -e cycles -e branches sleep 1
> 
> then 'perf report stat' would output the same as:
> 
>  $ perf stat -e cycles -e branches -a sleep 1
> 
>  Performance counter stats for 'system wide':
> 
>         34,174,518      cycles                    [100.00%]
>          3,155,677      branches                                                    
> 
>        1.000802852 seconds time elapsed
> 

Yeah, it'd be good to have same output both for perf stat and perf
report --stat (or stat if you want).  But I don't think it's possible to
determine multiplexed counter values like perf stat does unless we use
PERF_SAMPLE_READ for recoding.


> Another neat feature this kind of workflo enables is the integration 
> of --repeat to perf record, so something like:
> 
>     perf record --repeat 3 -a -e cycles -e branches sleep 1
> 
> would save 3 samples after each other, and would allow extraction of 
> the statistical stability of the measurement, and 'perf report stat' 
> would print the same result as a raw perf stat run would:
> 
>  $ perf stat --repeat 3 -e cycles -e branches -e instructions -a sleep 1
> 
>  Performance counter stats for 'system wide' (3 runs):
> 
>     28,975,150,642      cycles                     ( +-  0.43% ) [100.00%]
>     10,740,235,371      branches                                                      ( +-  0.47% ) [100.00%]
>     44,535,464,754      instructions              #    1.54  insns per cycle          ( +-  0.47% )
> 
>        1.005718027 seconds time elapsed                                          ( +-  0.43% )

Yeah, but it can be used only for a new forked workload.

> 
> Or something like that. At that point we share reporting between perf 
> stat and perf report, no special ad-hoc options are needed to just 
> measure and report timestamps, it would all be a 'natural' side effect 
> of having perf stat.
> 
> What do you think?

I think it'd be better if we can share code as much as possible.  And
it'd much better if we can forget about the difference in options. :)

Thanks,
Namhyung