How to read 95th percentile

Got a couple of newbie questions:

I want to use the 95th percentile to charge my customers usage. I am kind of familiar with it from MRTG, but it is a little different here on Orion.

In MRTG, you had to have 30 days worth of logs and then it would show only 1 graph line. On Orion it shows up right away and it shows 2 lines (In and Out).

First, is the data valid if I haven't been running it for 30 days yet? Second, do I add the In and Out of the 95th Percentile (On the avg bits per second graph) to get the usage of my customer?

One last thing, is that after running Orion for a few days or so, the 95th percentile suddenly diappeared and I add to go back in and re-enter a 95 in the field for it onthe Admin page. Is this a bug or something to do with not having enough data perhaps?

Thanks,

Brent Sadler
Web Creations & Consulting, LLC.

Find more posts tagged with

Accepted answers

All comments

MrSNMP

Brent,
95th percentile is the threshold where 95% of the data points fall below this level. While the data would certainly be more accurate from a usage standpoint if you had at least 30 days of statistics, you can calculate the 95th percentile on very small data samples.

Whether you add the in and out is up to you. Is it a half or full duplex circuit? How is the customer billed (traffic sent, received, etc..)?

On the last item, that sounds like a bug to me.--HTH

mr_sad68

We would be running a full duplex circuit. The customer is to pay for his avg badnwidth use on a month to month basis as opposed to paying for a set bandwidth limit. I just want to make sure that I am am not cheating myself or the customer by not looking at it correctly.

I would assume then, that you might suggest adding both in since it is duplex. Does anyone run Half Duplex on T1's?

Brent

quote:Originally posted by MrSNMP

Brent,
95th percentile is the threshold where 95% of the data points fall below this level. While the data would certainly be more accurate from a usage standpoint if you had at least 30 days of statistics, you can calculate the 95th percentile on very small data samples.

Whether you add the in and out is up to you. Is it a half or full duplex circuit? How is the customer billed (traffic sent, received, etc..)?

On the last item, that sounds like a bug to me.--HTH

Network_Guru

Brent,

From what you are saying, it sounds like you are trying to do 2 different things. Billing a customer for average utilization, is not the same as billing for 95th percentile utilization.

If you want average utilization, then I suggest you use Report Writer to come up with this figure for the month. It can give even give the average of both xmit & recv. combined.

As far as the 95th percentile calculation, I have requested that this also be added to the Report Writer util. I did some searching on my own for the 95th percentile calculation code, and found out it uses
the number 288. This equates to the following:

24hrs x 60min/288 = 5 minutes.

What this tells me, is they are basing this calculation on an interface statistics polling interval of 5 minutes. After 24 hours,
they throw out the top %5 of the samples (14.4), and the remaining highest figure = the 95th %ile.

In my case, I am collecting interface stats only every 15 minutes,
so in this case, the accuracy of this calculation is questionable.
Obviously once the detailed stats are rolled into hourly stats in the database, they would have to calculate the 95th percentile,
for one month, using a different formula. I'm not sure how/if this is
currently done, as if I graph the previous months stats, there is no
95th percentile calculation shown. If I graph the previous 3 months,
then I get the exact same figure as the last 30 days.

I was rolling the detailed stats into hourly stats every 45 days, but I have increased that now in an effort to try and get this stat for each of the last 2 months.

I have a support call in on this issue, and will update this thread when I get a response.

-=Cheers=-
NG

mr_sad68

OK,

I don't guess what I want to be doing is billing based on avg usage then. Avg usage would mean that my users who fill my bandwidth in the day and don't use any at night would not be paying properly.

This is what I thought, according to what the MRTG guru's say, that 95th percentil takes care of. But..then I still don't understand why my graphs have 2 95th percentile lines, while my MRTG shows 1?

I know I am missing some key fundamental issue here.

Brent

Network_Guru

There are 2 lines on each graph. One represents transmit utilization & the other - receive.
Our ISP bills for whichever one is the highest over the past month, & base it on a 5 minute polling interval.

Please be aware, these figures are highly dependant on the data statistics summarization settings. If you are polling the interface stats every 5 minutes, and roll the data into hourly stats after 7 days, then the monthly figures will most likely be incorrect.

My director accused me of using "Nortel" accounting practices ("Worldcom" for our US. neigbors) when checking the availability of one of our call centers over the past month. I was summarizing the detailed stats into hourly stats after 14 days. The effect was 288 - 5 minute samples for each of the first 14 days ((4032 total). The next 16 days only had 24 (hourly summarization) samples per day (384 total). The affect was that an outage in the oldest 16 days were "diluted" by the far more numerous samples in the first 14 days.

The polling interval must be identical over the total time period of the report, otherwise your results will be skewed. Most likely the 95th percentile is calcuated in the same manner on the monthly graphs?

I've changed the DB detailed stats summarization to 45 days now, even though this has increased the DB size dramatically & decreased the DB performance somewhat. This is the only way to get correct monthly SLA stats.

-=Cheers=-
NG