Ad Widget

Collapse

95th percentile : what do you want ?

Collapse
X
 
  • Time
  • Show
Clear All
new posts
  • rledisez
    Junior Member
    • Apr 2008
    • 5

    #1

    95th percentile : what do you want ?

    Hi,

    I work at a hosting company. I evaluate Zabbix to replace our (old and buggy) monitoring system (JFFNMS). One of the features we need is the ability to calculate the 95th percentile. We use it to bill the bandwith used by customers.

    First of all : is there somebody working on this ?
    I started to code on Zabbix 1.4.5. I took Cacti's documentation [1] as a reference.

    As you can see at Cacti's website, there are 10 methods of calculation. For now, my code works more or less. But there is a lot of case and I'm not sure if I have to handle all of them.

    Example, there is a graph with :
    - net.if.in[eth0]
    - net.if.out[eth0]
    - net.if.in[eth1]
    - net.if.out[eth1]
    - vm.memory.size[buffer]

    If the user ask the 95th percentile "all_max_current", should I calculate one or two 95th percentile. Two is the good answer because it's a non-sense to calculate network's traffic with memory. But two is also the more complicated way, and it seems Zabbix is not conceived for.

    More generally, should I take care of all possible case or should I trust the user choice ? What do you think ?

    Thanks.

    [1] http://www.cacti.net/downloads/docs/html/variables.html
  • stevender
    Junior Member
    • Jan 2007
    • 20

    #2
    Hi rledisez,

    Would be nice to have 95 percentile calculation in Zabbix.
    This question hangs around since a long time, but it seems that the Zabbix developers are not implementing it.

    I wrote some PHPcode to process the data in the zabbix-database, but this is only accessible in another webpage, not in zabbix.

    For calculation of 95%:
    you should calculate the average for periods of 5 minutes for both incoming and outgoing traffic. Then you take the highest value of those 2 and discard the lower one.
    After that you discard the 5% highest values and then the highest remaining value should be the 95p.
    Also: you should use an average value for every 5 minutes timespan in that month even if there is no data.

    Sounds tricky because you need data from 2 items.

    I didn't know there were different ways to calculate it, thought the one I described was the only correct one :-)

    Anyway, good luck with the development and I hope to hear about it soon :-)

    Steven

    Comment

    Working...