Re: PATCH] cpuidle: A new variant of the menu governor to boost IO performance

[Arjan van de Ven <arjan@infradead.org>]

On Sat, 12 Sep 2009 12:39:39 +0100
Matthew Garrett <mjg59@srcf.ucam.org> wrote:

> On Sat, Sep 12, 2009 at 05:26:47AM +0200, Arjan van de Ven wrote:
> > On Fri, 11 Sep 2009 23:03:09 +0100
> > Matthew Garrett <mjg59@srcf.ucam.org> wrote:
> > 
> > > When you say that a bit more power was used, is that instantaneous
> > > power draw or total power consumption over the run of the
> > > benchmark? I'd have expected that completing it 50% faster and
> > > then going idle would be a win overall.
> > 
> > I meant power, not total energy :-)
> > 
> > in terms of energy it's a win if you only do a fixed amount of
> > work...
> 
> Ok, so not really a downside.

correct.

> Not entirely relatedly, we've also seen
> io throughput issues related to P-states - using ondemand, we get
> reduced throughput until the number of clients becomes high enough to
> push the system into a higher P state. Is this something you've been
> able to measure?

so far, with a 10 msec ondemand interval, there is room for
improvement, but it's not too bleak either. With longer intervals there
clearly are problems.

I have some patches for that. However the gain is not nearly as clear
as with this C state patch, and they need some more tweaking...

-- 
Arjan van de Ven 	Intel Open Source Technology Centre
For development, discussion and tips for power savings, 
visit http://www.lesswatts.org