Re: [Ganglia-developers] Adding cpu_wio to universal metrics

[Brooks D. <br...@on...>]

On Thu, May 27, 2004 at 09:03:38AM -0700, Martin Knoblauch wrote:
>=20
> --- Brooks Davis <br...@on...> wrote:
> > On Thu, May 27, 2004 at 08:03:28AM -0700, Martin Knoblauch wrote:
> > > Hi,
> > >=20
> > >  after receiving overwhelming feedback on my request for discussing
> > the
> > > addition of cpu_wio as a global metric :-) I just went ahead and
> > did
> > > it.
> >=20
> > Arg, apparently I should have read my mail more carefully.  What is
> > WIO
> > and is it something we could expect to be available nearly everwhere?
> >=20
>=20
>   Linux, prior to 2.6 only reported user/nice/system/idle (actually
> idle was all cycles not accounted for). With Linux-2.6 the kernel
> reports user/nice/system/idle/wio/irq/softirq.=20
>=20
> wio - cycles spent on waiting for IO
> irq - cycles spent on serving hardware interrupts
> softirq - cycles spent on serving software interrupts
>=20
>  So, in any case something has/had to be done for 2.6.x kernels,
> otherwise the CPU-stats would not add up to 100%. The alternatives
> are/were:
>=20
> - add wio explicitely to idle and add irq/softirq to system, as it was
> in 2.4
> - export them as separate metrics
>=20
>  I went ahead and took the second path, because I think those numbers
> are interesting/useful. With the spin that irq/softirq are counted
> twice. Once as separate metrics (Linux only) and once in the global
> system metrics. The reason for counting them inside system is to avoid
> breaking the CPU-Report in the webfrontend. The "global" cpu-stats
> should add up to 100%.
>=20
>  I did put wio into the global section, because four major platforms
> report it: Linux (kernel 2.6), Solaris, HP-UX and IRIX (although up to
> today IRIX only calculated the number, but did not export it). OSF
> potentially could do it. The others likely not (not that I am very
> impressed with darwin and cygwin on CPU-stat reporting :-).
>=20
>  Hope this explains my thoughts and action.

OK, this makes sense, though I'm not 100% happy with it.  It would be
nice if we could report a value that indicated "always zero" so the
interface could not display it where appropriate in that case.  I wonder
what, if anything, would break if we returned -0.0.

-- Brooks

--=20
Any statement of the form "X is the one, true Y" is FALSE.
PGP fingerprint 655D 519C 26A7 82E7 2529  9BF0 5D8E 8BE9 F238 1AD4