On Fri, 2010-05-07 at 16:15 +1000, nathans@xxxxxxxxxx wrote:
> ----- "Ken McDonell" <kenj@xxxxxxxxxxxxxxxx> wrote:
>
> > the following additions to the derived metrics features.
> > ...
> > Thoughts, endorsements, criticisms?
>
> Bloat? Slippery slope here ... where does it end?
I'll wait for any other votes ... but I'm inclined to agree with you
(despite making the proposal).
> > ps of course it would be handy to understand how/why the expression
> > goes
> > negative, but this proposal provides an acceptable workaround for the
> > short-term that might even be useful for others.
>
> The mem.util metrics are all instantaneous and unsigned integers - both
> in PCP and being exported out of the kernel (fs/proc/meminfo.c) ...
> suggests a possible misunderstanding on how the metric values are being
> calculated in the kernel and that the derivation must be incorrect (not
> that I have the right answer, would entail time going thru the kernel mm
> code).
In these memory classification metrics, if the changes all along the
code paths from the kernel to linux pmda are not atomic, a - b could
still be negative for unsigned a and b.
|