lkml.org 
[lkml]   [2012]   [Mar]   [12]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 0/2] perf: add sort by inclusive time functionality (v2)
Hi,

2012-03-09 3:49 AM, Arun Sharma wrote:
> On 3/8/12 7:31 AM, Frederic Weisbecker wrote:
>> On Thu, Mar 08, 2012 at 08:29:01AM +0100, Ingo Molnar wrote:
>>>
>>> * Arun Sharma<asharma@fb.com> wrote:
>>>
>>>> This patch series refactors existing code a bit and adds sort by
>>>> inclusive time (time spent in the function + callees).
>>>>
>>>> Sample command lines:
>>>>
>>>> # perf record -ag -- sleep 1
>>>> # perf report -g graph,0.5,callee -n -s inclusive
>>>
>>> So I tried this out with:
>>>
>>> $ taskset 1 perf record -g git gc
>>>
>>> and got entries above 100% (in the TUI):
>>>
>>> $ perf report -g graph,0.5,callee -n -s inclusive
>>>
>>> + 321.11% 5628 [.] 0x392b609269
>>> + 142.27% 3774 [.] create_delta
>>> + 78.86% 1248 [.] lookup_object
>>> + 40.54% 1348 [k] system_call_fastpath
>>> [...]
>>>
>>> Is that expected?
>
> Yes - this is the "known bug" I noted in the cover letter
>
> The second column (samples) is still accurate and could be used for the analysis.
>
>>
>> I think this happens because of this:
>>
>> - hists->stats.total_period += h->period;
>> + if (!h->inclusive)
>> + hists->stats.total_period += h->period;
>>
>> Which I'm not sure why it is needed btw.
>
> Suppose the perf.data file had 1000 samples each with a period of 1e6 events.
> total_period would be 1e9 without -s inclusive. Further, let's say the
> callchains had an average length of 10.
>
> Now, after adding extra entries to the histogram, total_period would be 1e10,
> which screws up the percentages. I'd like to differentiate between the hist
> entries that were in the event stream vs the ones added for inclusive time
> computation. Desired end result: the total_period remains unchanged at 1e9.
>
> This is done via:
>
> + if (i != 0)
> + he->inclusive = 1;
> + else
> + orig_he = he;
>
> Either (i != 0) is not a good enough test, or the inclusive bit is not getting
> propagated properly after histogram collapsing/resorting. This is the part I
> need to better understand and debug.
>
> I tried to explain this problem in my first RFC message as well:
>
> http://thread.gmane.org/gmane.linux.kernel/1262289
>
> The problem Ingo is running into (and I've reproduced it on my end as well) is
> that total_period is smaller than without -s inclusive i.e. h->inclusive is 1
> when it shouldn't be.
>

I think it's because of the shared hist_entry. If a callchain is a subset of
another, it will be marked as inclusive so that it cannot be contributed to
total period. Say, there're two chains - X (a -> b -> c) and Y (a -> b), once
__hists__add_entry_inclusive() was called on X, we have:

a -> b -> c
a -> b (inclusive)
a (inclusive)

And then, calling the function on Y should make:

a -> b
a (inclusive)

However, since both callchains are in tree already they'll be shared and
marked *inclusive*. Thus the total period will not increased at all for Y.
Also I guess the reverse case - add Y first, and then X - will have the same
result.

Thanks,
Namhyung


\
 
 \ /
  Last update: 2012-03-12 08:19    [W:0.123 / U:14.640 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site