[rrd-users] Experiance in HW forcast

[Tobias Oetiker]

Michael,

Yesterday Michael Schwartzkopff wrote:

> >
> > take some high resolution sample data worth several seasons (weeks)
> > and then try out the different parameters ... note some are very
> > sensitive ... the best way to get your config right is to tune with
> > real data ...
>
> Yes. I realized the sensitivity. that confirms my findings that the fit
> depends on gamma on some fractal way.
>
> But anyway you need some automated fit before you can create the RRD. My idea
> is to write some wrapper to make it is easy as possible to create the correct
> RRD.

> > > 2) Is there any possibility to optimize the fitting parameters to recent
> > > data? My idea would be to optimize the fitting parameters, create the
> > > RRDs and to optimize the RRDs from time to time. Any ideas or comments?
> >
> > you can rrdtool tune the parameters ...
>
> I read about that already. Perhaps my wrapper could also be user to tune the
> fit parameters every week or so.

The only reliable way I have found, is to collect 2 month real data
at high resolution and then run it through rrdtool with different
HW paramerts set, plotting the resulting curves. This gives a nice
visual way of tuneing the parameters ...

I don't think adjusting parameters as 'you go' is a good thing
(except once you know exacly what you are doing) ... as the system
is 'learning' from past data, it takes some time to settle, these
processes just take too long for experimentation to work in real
time ... use archived data ... much more satisfying ...

cheers
tobi

-- 
Tobi Oetiker, OETIKER+PARTNER AG, Aarweg 15 CH-4600 Olten, Switzerland
http://it.oetiker.ch tobi at oetiker.ch ++41 62 775 9902 / sb: -9900