Re: low latency & preemtible kernels

[Andrew Morton <akpm@zip.com.au>]

george anzinger wrote:
> 
> wwp wrote:
> >
> > Hi there,
> >
> > here's a newbie question:
> > is it UNadvisable to apply both preempt-kernel-rml and low-latency patches
> > over a 2.4.18 kernel?
> >
> > thanx in advance
> >
> > --
> I believe that the preempt kernel patch or one related to it does the
> low-latency stuff in a more economical way,

Sigh.  Not to single you out, George - I see abject misunderstanding
and misinformation about this sort of thing all over the place.

So let's make some statements:

- preemption is more expensive that explicit scheduling points.   Always
  was, always shall be.

- Anyone who has performed measurements knows that preemption is
  ineffective.  Worst-case latencies are still up to 100 milliseconds.

- preemptability is a *basis* for getting a maintainable low-latency
  kernel.   And that's the reason why I support its merge into 2.5.  Same
  with Ingo, I expect.

  But there's a lot of icky stuff to be done yet to make it effective.

> i.e. takes advantage of the
> preemption code to implement the low-latency stuff.  See the lock-break
> patch that rml has.  It should be right next to the preempt patch.

lock-break is missing the cross-SMP reschedule hack, so on SMP it'll
still have very high worst-case latencies.  If all the other parts
of the low-latency patch were included then preempt+lock-break should
give better results than low-latency.

-