Wildly different output sizes between subsequent perf record runs.

Wildly different output sizes between subsequent perf record runs.

[Bram Stolk]

If I do subsequent runs of 'perf record', both of which are identical,
I sometimes get wildly different output sizes.


https://gist.github.com/stolk/3666b2b4d4319764fdff4a2d4c619782

The binary did the same amount of work, and took the same amount of
time to execute.
This is with using perf from Ubuntu LTS.

Re: Wildly different output sizes between subsequent perf record runs.

[Milian Wolff]

On Samstag, 4. November 2017 02:57:01 CET Bram Stolk wrote:
> If I do subsequent runs of 'perf record', both of which are identical,
> I sometimes get wildly different output sizes.
> 
> 
> https://gist.github.com/stolk/3666b2b4d4319764fdff4a2d4c619782
> 
> The binary did the same amount of work, and took the same amount of
> time to execute.
> This is with using perf from Ubuntu LTS.

I think this is the same bug that I have raised before. See https://www.mail-archive.com/linux-kernel@vger.kernel.org/msg1479712.html and below.

Can you try to run your command with a constant period (`-c 3000000` or 
similar), instead of using a dynamic period (i.e. `--freq`)? Does the issue go 
away then? If yes, then it's the same issue that I described earlier, which 
noone else from the perf team has acknowledged so far. I know multiple people 
who are affected by this though...

-- 
Milian Wolff | milian.wolff@kdab.com | Senior Software Engineer
KDAB (Deutschland) GmbH&Co KG, a KDAB Group company
Tel: +49-30-521325470
KDAB - The Qt Experts

Re: Wildly different output sizes between subsequent perf record runs.

[Bram Stolk]

Thanks Milian,

This issue is too elusive.
Today I can no longer reproduce the issue with short data runs, even
though I use the same perf binary on the same system with same
arguments.
Not sure what has changed.

I'll try to reboot and try again, see if I can trigger it on a 'colder' machine.


On Mon, Nov 6, 2017 at 4:06 AM, Milian Wolff <milian.wolff@kdab.com> wrote:
> On Samstag, 4. November 2017 02:57:01 CET Bram Stolk wrote:
>> If I do subsequent runs of 'perf record', both of which are identical,
>> I sometimes get wildly different output sizes.
>>
>>
>> https://gist.github.com/stolk/3666b2b4d4319764fdff4a2d4c619782
>>
>> The binary did the same amount of work, and took the same amount of
>> time to execute.
>> This is with using perf from Ubuntu LTS.
>
> I think this is the same bug that I have raised before. See https://www.mail-archive.com/linux-kernel@vger.kernel.org/msg1479712.html and below.
>
> Can you try to run your command with a constant period (`-c 3000000` or
> similar), instead of using a dynamic period (i.e. `--freq`)? Does the issue go
> away then? If yes, then it's the same issue that I described earlier, which
> noone else from the perf team has acknowledged so far. I know multiple people
> who are affected by this though...
>
> --
> Milian Wolff | milian.wolff@kdab.com | Senior Software Engineer
> KDAB (Deutschland) GmbH&Co KG, a KDAB Group company
> Tel: +49-30-521325470
> KDAB - The Qt Experts
>
>



-- 
Owner/Director of Game Studio Abraham Stolk Inc.
Vancouver BC, Canada
b.stolk@gmail.com