Cheshire Cat Computing

I'd like to say that from the perspective of a system admin rather
than a network admin, the current trending works just fine and is
a significant added benefit of using the software.

We have currently about 129 SUN/IBM/HP servers that are being tracked
for about 20 different items (busy vs idle, network collisions vs.
errors, user cpu vs. system cpu, system io vs. idle, disk space,
load average 5 vs. 15, memory used vs. free, pagescans, processes,
runq vs. blocked/swapped, swap used vs swap free, tcp established
vs. listen, tcp retransmission vs. tcp dropped, tcp time wait, users,
etc.

For most of these, trending is able to extrapolate in a meaningful
way, with a decay of 1, to show reasonable-looking future predictions.
Even for the choppier items, it still looks reasonable to most of us
here.

I encourage the network crowd to climb on board and try out the
new trending but, please, if you make radical changes, please provide
linkage back to this original version. It works AOK for system
administration data trending.

Thanks,

Stuart

Anyone wanting to try the trend.cgi trending module included with v2.13a, please feel free to do so -- but, since it is a prototype still, I will be posting updates often to the web site for people who are interested as things develop. There are significant differences just between v0.1 and v0.2. Also, please post/email your comments and experiences and suggestions.

Finally, anyone sufficiently math-savvy to work out a workable algorythm to calculate 2nd-order curve approximations, please let me know...

Steve

An engineer friend of mine has suggested usnig fourier analysis on the existing data in an attempt to predict the future, given that the normal usage pattern tends to be a repeating wave (peak at midday, low at midnight, sort of thing). This sounds intriguing except that (not being an engineer) I am not too familier with fourier analysis. He suggested a Finite Fourier Transformation, but I'm still working on how to achieve this.

Anyone out there have an opinion on whether or not this will be a good way to trend the data?

Hi, I would like to know what the algorithm is for the current
trending, in pseudo-code if possible.

The reason is that I am creating a separate script to create
a list of the graphs which are showing bad rates of growth.

This will help our staff not have to go through 2600 graphs
to find them.

I did try to figure it out from the existing code but it was too
obscure and would like to get a higher-level view in pseudo-code
for re-implementation as a script. It will examine all the RRD
files and find data that is growing too fast, predicting overgrowth
1 year out.

Thanks,

Stuart