Thread: CDR vs 95th %tile - overselling?

Yes
Yes
Yes
Yes

basically you're correct on all parts.

Until I get time to write a better response:

http://inconcepts.biz/cr/95th.html

A lot of ISPs won't even tie you to a CDR (though they might give you better pricing if you agree to one) or, if you are on one, be too strict if you go over it. There are a few folks on this board who can testify that we're certainly not too harsh on any overages.

I think you've slightly misunderstood, it's not the top level minus five percent...5 minute averages of your port usage are taken and recorded; at the end of the month these values are arranged in ascending order, and the top 5% discarded. So basically you need to go over your CDR for approx 36 hours in the month to go into overages.

Originally Posted by reactive-uk

Does this mean that your lowest possible bill is for the x Mbits, but your actual bill will be for the transfer rate 5% off the peak level (or indeed the original x is you have no peaks?).

Your bill will be x for the CDR and y for the overages if there have been any, at what you reached less the top 5% of the results
i.e. if you were measured every hour for a month, there would be ~21600 samples sorted high-low, the 1080 highest would be thrown out, you'd pay at the level reached on sample 20520

Some ISPs do it on 95%ile, some on 98%ile, some on 90%ile, some on average, some on 100% peak ... everyone's playing a complex numbers game.

Originally Posted by reactive-uk

This presumably means that your bandwidth bill is effectively open ended (up to port speed) by nature of the fact you cant (initially) tell where and how big your peaks will be?

Sometimes. It is possible (and likely) that your upstream will apply some level of limitation based on your CDR or other arbitrary number. For example Abovenet-US rate limit you to 150% of your CDR so even on a 1000Mb/s port on a 10Mb CDR you're "capped" at 15Mb/s as standard.

Similarly there can be other factors in play - daisychained switches is a common "trick" employed - where you might have a 100Mb connection, but you +47other 100Mb/s connections are on a 1Gb/s uplink so immediately contented, and all those 1Gb/s uplinks go in into a switch with a 20Gb/s backplan so contending you again, and then the particular route your raffic wants to go might be on over an upstream supplier doing the same thing.

You can also apply your own limiting, port metric adjustments , preferences of transit etc, to "even" out - depends on how much you wat to shape and manage the traffic - with many suppliers they just shove it over whomever is cheapest, and artificially keep the expensive/quality transits as low as possible.

Originally Posted by reactive-uk

Also does this mean that the ISPs can effectively sell more megabits per month then they have committed themselves, on the basis that my peak wont be at the same time as customer y's peak

Yes, but not all do so.

Originally Posted by reactive-uk

Presumably the (decent!) ISP estimates peak loading over time and provides a decent amount of bandwidth thats in excess of the total customer CDR

Plenty dont even buy in enough to cover their clienst CDRs gambling on the idea that not every uses everything they pay for and for those that do, not everyone uses it all at the same time. There are many "consumer" oriented ISPs that regularly have problemsdue to the level of overselling on their network.

Originally Posted by reactive-uk

Presumably the ISP only really has to provide as a minimum the total customer CDR, or indeed might even provide less if he assumed that customers wouldnt (simultaneously) use all the CDR they had purchased?

The more "corporate" oriented ISPs generally have CDRs in excess of their commitement to clients.