[rrd-users] Input values normalization

[Simon Hobson]

"ENTRESSANGLE, ERIC (ERIC)" <eric.entressangle at alcatel-lucent.com> wrote:

> Just to be sure to fully understand what's happening when poller's frequency is 60 seconds and step is 1 second : for the RRA:MAX:0.5:60:43200, from which 60 PDPs is the CDP calculated, since we have only 1 sample value ?

A read of http://www.vandenbogaerdt.nl/rrdtool/after_the_fact.php (as well as the one on normalisation) might help.

Basically, when you do an update, RRD will fill in any completed buckets with data. So taking your example, you update with time=x+0 (where x is some base time "on the minute") and RRD fills in the buckets with end times up to time x. Then 30s later you supply a new update, at which point RRD fills in the values for buckets x+1, x+2, ... x+30 = all of those will contain the value you supplied except perhaps x if your previous update wasn't timestamped exactly x. At this point nothing will happen to your consolidated data.

At some point, you'll have supplied enough data to complete a consolidated data point. For the max consolidation, you will get the highest of the values stored in your 1s buckets for the preceeding consolidation period. Bear in mind that **ALL** buckets end on an exact integer multiple of the step/consolidation period since Unix epoch (midnight, Jan1 1970). Thus if you consolidate in 60s units, each consolidated data point will end on an exact minute (ie 00 seconds). When you supply the data doesn't affect this - if you update at (say) 15 seconds and 45 seconds past the minute then the 30 one second buckets you fill in with an update will span from 15 to 45 seconds past, and 45 to 15 seconds past the minute. When consolidating to 1 minute buckets, you'll get values from 15seconds of update 1 (ending at xx:xx:15), 30 seconds of update 2, and 15 seconds of update 3 - with the remaining 15 seconds from update 3 being carried over to the next minute.