Re: [RFC] sched_clock: Track monotonic raw clock

[Pawel Moll <pawel.moll@arm.com>]

On Sat, 2014-07-19 at 06:02 +0100, Richard Cochran wrote:
> On Fri, Jul 18, 2014 at 06:43:39PM +0100, Pawel Moll wrote:
> > 
> > This code definitely needs more work and testing (I'm not 100%
> > sure if the Kp and Ki I've picked for the proportional and
> > integral terms are universal),
> 
> I wouldn't bet on it.

Yeah, I should have said: I'm 100% sure they are not universal.

> > but for now wanted to see
> > if this approach makes any sense whatsoever.
> 
> You are reading sched_clock and mono-raw together every so
> often. Really stupid question: Why not just place that information
> into the trace buffer and let user space do the clock correction?

That approach has been also discussed, last time in the mentioned
thread:

http://thread.gmane.org/gmane.linux.kernel/1611683/focus=1612554
http://thread.gmane.org/gmane.linux.kernel/1611683/focus=1612554

With both Ingo and John showing preference towards the clock alignment,
so that's where I looked this time (I've already done custom perf
ioctls, posix clocks... don't really know how many different ways I've
tried).

> ...
> 
> > +		/* Tune the cyc_to_ns formula */
> > +		mult_adj = sign * (error >> 2) + (cd.error_int >> 2);
> 
> So Kp = Ki = 0.25? And did you say that the sample rate is 10/second?
> 
> I guess that, while this works well on your machine, it might not
> always do so, depending on the mono-raw clock. Probably Kp/i need to
> be tunable to a particular system. Even better would be to leave this
> out of the kernel altogether.

My hope is that, given that time correlation is pretty "static" process
(the clock skew should be reasonable, otherwise the time source are
plainly rubbish), there will be a way of implementing this in a
self-tuning way. I may be proven wrong, but it seems like a noble thing
to try.

I fully agree that there is no place for manual PI regulator tuning in
the kernel.

Pawel