95th Percentile != Lame

Mon Jun 4 02:59:31 UTC 2001

> I still fail to see how "peak bits" or "bursted bits" are more expensive
> than "regular bits".  A 100Mbit FE port costs whatever it costs, and does
> not fluctuate with usage.  This is true of almost all of your links within
> the network - excluding those where you have negotiated usage-based
> billing.  An OC3, point to point, costs as much as it costs irrelevant of
> its usage.  Therefore, every bit that crosses this circuit has a cost. 
> 
> Why not simply pass this cost on to the customer bit for bit? 

let's say that you've sold 1Gb/s of connectivity to each of 50 customers.
if you don't do anything fancy in your backbone provisioning, two things
will happen: (1) you will need 50Gb/s of peering connectivity to EVERY PEER
and on EVERY BACKBONE LINK (just in case all five customers send full blast
to the same destination at the same time, as in a DDoS or flash crowd
scenario); (2) when you try to pass the costs of this provisioned capacity
to all of your customers (equally), you will have noncompetitive rates.

what actually happens is that you provision as much capacity (including
whatever headroom your business plan calls for) as is actually needed, and
then try to upgrade it fast enough so that your customers always get their
bits through.  it is not actually possible to get 50Gb/s worth of peering
connectivity to most peers -- and if you could then they'd be dropping most
of your traffic one hop inbound on their network anyway.

charging for 95th percentile (highest of in or out, not additive) is a way
to get those customers whose traffic puts the most strain on the
provisioning plan to pay more.  but consider another model for a moment.
the power companies hereabouts use the monthly peak to set the rate (per
KwH) for the month.  so if you never hit them very hard then every KwH you
buy from them is cheaper.  this model correctly passes the provisioning
(especially headroom) costs to commercial customers who put the most stress
on the system, but also correctly rewards customers with lower costs based
on them keeping their actual usage down.  i'm not suggesting that this
would apply perfectly to the IP transit situation, but it takes more of the
relevant factors into account.