[procps] Re: C-States handling - new switch?
- From: Jaromir Capik <jcapik@xxxxxxxxxx>
- To: procps@xxxxxxxxxxxxx
- Date: Wed, 08 Feb 2012 11:50:12 -0500 (EST)
> > If I understand that correctly, CPU frequency changes
> > permanently distort the values and that also makes
> > the stats unusable for monitoring systems since CPUs
> > with halved frequencies show 2times higher load, right?
>
> Not any more, since we're now based on actual elapsed tics.
>
> Whatever tics are available, they'll be scaled to the range of
> 0-100%. Assuming, of course, a category exceeds the arbitrary
> (maybe future-tunable) 10% threshold.
Sorry ... but this sounds to me like an exact opposite.
If we scale the values according to the total available ticks,
then the same load in case of halved frequency must look 2x higher.
Because the process steals the same amount of ticks, but the
total ticks are halved.
So ... if I understand that correctly, this kind of scaling
produces the distortion. I believe, that majority of users aren't
interested in load related to the actual frequency, but they're
interested in load related to the maximum frequency, since it
defines how much of speed reserves the CPU actually has.
And I can also imagine two groups of people each wanting
a different behavior. And that smells like need for a new switch.
>
> Jim
Jaromir.
Other related posts:
