Re: perf-stat per thread results

[Wim Heirman]

On 14 May 2011 14:45, Wim Heirman <wim@xxxxxxxxxxx> wrote:
> On 13 May 2011 22:45, Arnaldo Carvalho de Melo <acme@xxxxxxxxxx> wrote:
>> Em Fri, May 13, 2011 at 10:32:58PM +0200, Wim Heirman escreveu:
>>> > It's supposed to do that if --stat is specified, and it used to work - see this
>>> > commit:
>>
>>> >  8d51327090ac: perf report: Fix and improve the displaying of per-thread event counters
>>
>>> > and the output there:
>>> >
>>> >     #  PID   TID  cache-misses  cache-references
>>> >       4658  4659        495581           3238779
>>> >       4658  4662        498246           3236823
>>> >       4658  4663        499531           3243162
>>
>>> > which appears to be roughly what Wim is asking for, AFAICT.
>>
>>> Thanks, this is exactly what I'm looking for. In 2.6.32 (Ubuntu 10.04)
>>> it works, although if I use --pid rather than the -- <command> variant
>>> the first thread seams to be missing. In 2.6.38 (Ubuntu 11.04) the
>>> first thread is missing in both use cases, and I get one column per
>>> processor (which in itself is fine).
>>
>> Can you try after applying the patches in this message:
>>
>> http://marc.info/?l=linux-kernel&m=130385067430510&w=2
>>
>> and report your results?
>
> Sorry, no improvement.
>
> $ ./perf record -e cycles --stat -- ./fft -p4 -m24 && ./perf report
> --thread | tail
> [ perf record: Woken up 5 times to write data ]
> [ perf record: Captured and wrote 1.198 MB perf.data (~52331 samples) ]
> # PID  TID   cpu-clock
>  954  958  8067423322
>  954  957  6761317556
>  954  956  6006327147
>
> $ ls /proc/$(pidof fft)/task
> 954  956  957  958

Looks like perf-report --thread is reading PERF_RECORD_READ events
from perf.data. But these are only emitted by the kernel for child
threads: in kernel/events/core.c, the only call to
perf_event_read_event() is in sync_child_event().
Should perf-record then be adapted to do something like perf-stat does
and use __perf_evsel__read to read the parent counter's final values
and add them to perf.data ? That way perf-report can subtract all
children from the final value and get the main thread's counter
values.

Regards,
Wim
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/