Re: perf-stat per thread results

[Wim Heirman <wim@heirman.net>]

2011/5/14 David Ahern <dsahern@gmail.com>:
> On 05/14/11 06:49, Wim Heirman wrote:
>>> Hmm.... my mileage varies using latest kernel
>>> (446cc6345d3de6571bdd0840f48aca441488a28d)
>>>
>>> $ /tmp/build-perf/perf record --stat -fo /tmp/perf.data -p $(pidof rsyslogd)
>>> ^C[ perf record: Woken up 1 times to write data ]
>>> [ perf record: Captured and wrote 0.007 MB /tmp/perf.data (~308 samples) ]
>>>
>>> $ /tmp/build-perf/perf report -T -i /tmp/perf.data
>>> # Events: 6  cycles
>>> #
>>> # Overhead   Command      Shared Object                      Symbol
>>> # ........  ........  .................  ..........................
>>> #
>>>    97.61%  rsyslogd  libc-2.13.so       [.] __libc_disable_asynccancel
>>>     2.39%  rsyslogd  [kernel.kallsyms]  [k] native_write_msr_safe
>>>
>>>
>>> #
>>> # (For a higher level overview, try: perf report --sort comm,dso)
>>> #
>>> # PID  TID
>>>
>>>
>>> ie., I do not get the counter values. Specifying the counter with -e
>>> (e.g., -e branch-misses) does not help -- still no counter output.
>>
>> Is rsyslogd multithreaded? (Or at least, do the non-main threads
>> execute any work during your perf-record measurement) If not, then
>> what you see is consistent with what I'm getting, i.e. everything but
>> the main thread is reported.
>
> It is multithreaded, but my point is that I do not get counter output at
> the end -- the PID/TID table is empty. I do not get counters for single
> threaded processes nor for commands run by perf record -- e.g.,
> /tmp/build-perf/perf record --stat -e instructions -fo /tmp/perf.data --
> sleep 1

My guess was that none of the threads got scheduled while you did your
perf-record run (rsyslogd usually isn't exactly very CPU intensive).
And the main thread isn't ever reported, at least that's the bug I'm
seeing. Can you try with a compute-intensive, multi-threaded program?

Wim.