[rrd-users] Holt-Winters - Seasonal value

[Tobias Oetiker]

Yesterday Koaps wrote:

> I think you want the seasonal period to be per day, not weekly.

a one day season only makes sense if your days are realy uniform
... if there is a weekly pattern, you need a week season ...


> You probably want a week or two for the row count.
>
> RRA:HWPREDICT:<row count>:<alpha>:<beta>:<period>
>
> *  row count is the number of forecasts to store before wrap-around; this number must be longer than the seasonal period. This value will also be the RRA row count for DEVPREDICT RRA.
> * alpha is the intercept adaptation parameter, which must fall between 0 and 1. The same value will be also be used for gamma.
> * beta is the slope adaptation parameter, again between 0 and 1.
> * period is the number of primary data points in the seasonal period. This value will be the RRA row count for the SEASONAL and DEVSEASONAL RRAs.
>
> So like 288 for the period at 300s(5min) will give you daily
> seasonal, so you can predict on traffic changes throughout the
> day, and 1440 for the row count will store a weeks worth of
> predictions.
>
> You might be able to bump the row count to 2800 or something for
> two weeks.
>
> Just remember, if you are storing traffic graphs for like a
> switch, say a 24 or 48 port one, depending how you setup the
> rrd's they are updating 6 graphs per port with 2 RRA's maybe
> every minute, or 6 graphs with 48 to 96 RRA's every 5 minutes.
>
> This can definitely chew up some resources so unless you really
> have a reason to predict for long periods of time, I would get
> them as short as possible and use it to just look for failures in
> the short term.

what I would suggest, if that you take about 8 weeks worth of 5
minute data, setup your HW config and then feed the data in and
look at the result, twiddle the knobs and do it again ...

fun and helps you understand the effects of the various parameters
... note that some of them are only tuneable via rrdtool tune!

cheers
tobi

>
>
>
>
>
>
> --- On Thu, 3/11/10, Lyle Brooks <brooks at deseret.com> wrote:
>
> > From: Lyle Brooks <brooks at deseret.com>
> > Subject: [rrd-users] Holt-Winters - Seasonal value
> > To: rrd-users at lists.oetiker.ch
> > Date: Thursday, March 11, 2010, 12:00 PM
> > I'm trying to wrap my brain around
> > the Holt-Winters capabilities
> > within RRD tool.
> >
> > The concept of "seasonal period" is what I'm not sure if I
> > understand
> > and how to translate it into the correct value when
> > defining my RRA.
> >
> > The data I am tracking has a step value of 5 minutes. 
> > I "think" my
> > data has a seasonal value of 1 week.  By that I mean,
> > the traffic
> > pattern is this Friday's traffic is similar to each
> > previous Friday,
> > but is not related to the previous day.
> >
> > So my questions are...
> >
> > 1) What is the appropriate setting for the SEASONAL value
> > when creating
> > an RRA with HW?
> >
> > 2) If the algorithm detects aberrant behavior based on a
> > previous season,
> > would it make sense to have more that one previous season's
> > worth of
> > values to determine what is normal behavior.  In my
> > case, would 4 or 8 weeks
> > of data be yield more accurate baseline than 1 week of
> > data?   Or am
> > I mis-understanding how the SEASONAL coefficients are
> > utilized.
> >
> > I've read some Holt-Winters related documentation and have
> > played around
> > with the various alpha/beta/period parameters, but I'm not
> > sure I have
> > a full grasp of these concepts.
> >
> > Any help would be appreciated.
> >
> > Lyle
> >
> > _______________________________________________
> > rrd-users mailing list
> > rrd-users at lists.oetiker.ch
> > https://lists.oetiker.ch/cgi-bin/listinfo/rrd-users
> >
>
>
>
>
> _______________________________________________
> rrd-users mailing list
> rrd-users at lists.oetiker.ch
> https://lists.oetiker.ch/cgi-bin/listinfo/rrd-users
>
>

-- 
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