[rrd-users] Re: Problems with multi-DS RRD that gets asynchronous-DS updates

[steve rader]

 > From: Alex van den Bogaerdt
 > One more thing: You can allow for some unknown values when building
 > the RRA. Look at this:
 > 
 >         rrdtool create filename [--start|-b start time]
 >                 [--step|-s step]
 >                 [DS:ds-name:DST:heartbeat:min:max] [RRA:CF:xff:steps:rows]
 > 
 > You will probably have set xff to 0.5 which means that half of the
 > datapoints may be absent while still producing an other-than-unknown
 > entry in the RRA.

Yep.

 > When sampling every 300 seconds and having an RRA which stores one
 > datapoint per row, there is either 0% or 100% unknown so this won't
 > change. However, in the "weekly" graph this unknown value should not
 > be there anymore. 

Correct: only my daily or sorter graphs have the "async updates
cause graph dropout problem".

So it seems the longer (weekly and greater) RRAs "tolerate"
the U values.

As a work-around, I suspect I could gather data every five minutes,
have the daily RRA have 15 (ten?) minute steps and a .5 xff.  But that
causes a lot (imho, too much) of loss of granularity of my data.


I can't just set xff to 1!  =;)  Is there a way to define a daily
RRA at a five minute polling interval that will allow for async DSs
updates and make correct graphs with a five minute granularity?

I suspect the answer is no.  I did #define DEBUG in rrd_update.c
and got a vague understanding of "rrdtool update".  Frankly, I think
there's an inherient flaw in the RRD data structure: "last_update"
is kept for the entire RRD only, and yet "last_update" is different
for each DS.  I will post the gory details about this theory to the
developers list in a few minutes.


For now, I'm just fudging my data collection times such that all DSes
get updated at the same time.

But note that this is really a kludge: if snmpget for input happens
quickly and the snmpget for output takes 60 seconds, then the data
for input would be skewed by a full 20%!

(For my RRDtool interface, this is not a big issue because my data
gathering operations time out after N (currently set to 5) seconds
by using "alarm" and the ALRM signal in perl.)

later
steve
- - -
systems guy
wiscnet.net

--
* To unsubscribe from the rrd-users mailing list, send a message with the
  subject: unsubscribe to rrd-users-request at list.ee.ethz.ch