[rrd-users] rrdt and prediction

[Sven Ulland]

Darryl Lewis wrote:
> Has anyone written any code to do trend analysis and prediction on
> rrdt?

RRDtool features aberrant behaviour detection based on holt-winters
forecasting/prediction. See
http://oss.oetiker.ch/rrdtool/doc/rrdcreate.en.html
http://www.usenix.org/events/lisa2000/brutlag.html
http://cricket.sourceforge.net/aberrant/rrd_hw.htm

So yes, there is backend code in place. Frontend code is available in
some network monitoring tools, but I haven't found them to be
satisfactory. It is my impression that trend analysis and prediction
requires special tuning on a case by case basis.

> What I'd like to be able to do is do some sort of prediction based
> upon a rrd and predict what it will be at a future point in time
> based upon the past trends. I'd also like to be able to predict when
> a device will reach a preset level (or 100%) so we have an idea of
> when to update equipment.

Linear trend prediction? That's an easy exercise in programming. I am
not aware of any existing code for this, but it would simply involve
fetching values from the RRD in question, and using them -- possibly
reduced to averages over some time interval -- as input values for a
linear regression calculation. If that's all you need, there's no need
to look at aberrant behaviour detection, which is rather for detecting
anomalies than predicting long-term trends.

> Does anyone have anything like this, or do I need to invent the
> wheel?

You don't need to invent the wheel, merely build it of ideas and parts
already available.

Sven