Re: random sample period

[John L. <le...@mo...>]

On Thu, Nov 24, 2005 at 04:15:05PM +0100, Stefan Feuerhahn wrote:

> I have a few questions regarding the sample period of oprofile. As I
> understand it the sample period can be chosen but is always fixed during
> the course of one profile. It seems obvious that when the sample period is
> fixed that certain functions that are called periodically could always or
> never fall in this period. This problem could be solved by randomizing the
> sample period.
> there is a post from Huang Ying with a patch that randomizes the sample
> period: http://sourceforge.net/mailarchive/message.php?msg_id=11612875
> But I'm missing a discussion about how important this is.
> My question is mainly why oprofile doesn't support this by default (or has
> the patch already been included?) because it seems to me so fundamental.
> Is this planed for the future or has experience shown that this is never
> an issue and is thus never needed?

Aside from timer mode, where the synchronous nature makes it inevitable,
it's not actually that important in our experience. With a very low
sample rate it'd be possible to end up in lock-step (if we're counting
CPU_CLK_UNHALTED, and using poll_idle) I suppose, but we generally have
a higher sample rate.

You might be able to construct a scenario where you end up having
serious lock-step problems, I suppose.

I'm not hugely against the idea of randomisation, but it does slow down
the fast path, and I'm unconvinced of the need for it. Some test
construction would be probably useful.

Regarding the patch above, I never heard back from Huang IIRC.

regards
john