[Peering-disc] [COOK/PEERING] the absolute Crux of the Issue before us!

[Andrew Odlyzko]

Dear Colleagues,

First of all, let me apologize for not participating in the
discussion before, and not even introducing myself.  I had a
mad schedule, so I have not even been able to carefully read 
all the messages that have been distributed.  It will take
me a few more days to catch up, especially as in a few hours
I am flying off on another trip.  Still, I thought I would
toss out at least one provocative idea, to stir up the
discussion.  

The question I would like to pose is whether the transit/peering
issue matters much or not.  It seems that at least in the US,
the cost of transit is low enough that it is not quite irrelevant,
but should not be critical to the survical of an ISP that can
reach decent size.  Here are some calculations that support this.
The basic assumption is that the ISP can buy transit at something
like $150/Mbps per month (and so has enough traffic to fill an
OC3, say).  

1.  ISP serving residential broadband customers:  The highest
figure I have heard for the traffic that such users generate
today was for a group of university folks, who were running an
average of about 10 Kb/s of traffic (sum of incoming and outgoing
traffic) per subscriber over long periods, with relatively low
peak-to-average ratio (since most of the traffic was peer-to-peer
music swapping).  Well, even if we allocate 20 Kb/s for each
user, and in each direction, to accomodate growth in usage and
higher peak traffic, we can get 50 such residential users into
a 1 Mbps slice, so the cost of transit per user will be $3/month.
That is simply not that much!  (And in practice usage is far
lower, and so are average per subscriber costs.)

2.  Global computation:  Current estimates (by myself, RHK, and
others) are that US Internet backbones carry something like
100,000 TB/month, which comes to 300 Gbps (averaged over a full
month).  Suppose we send that through a transit connection that
runs at 30% average utilization (measured over a full month).
(This, I am told, is a reasonable approximation of what is
achieved in practice, but perhaps somebody on the list will correct me.)
That means the transit capacity will have to be 1,000 Gbps, 
which, at  $150/Mbps per month, will cost $150 million per month,
or $2 billion per year.  (More details to justify this:  Think
of a far simplified world with just 2 ISPs, call them X & Y, of 
equal size, with customers of X communicating exclusively with
customers of Y, and vice versa.  Suppose that X & Y both buy 
transit from a backbone B, that basically just connects them.  
The traffic will on average be 150 Gbps in each direction, 
requiring transit capacity of 500 Gbps for each of X & Y, leading 
to the cost figures above.)  Yet the US backbone Internet 
connectivity market (excluding modems, etc.) is supposed 
to have revenues of $15 billion of so per year.  Thus most of that 
revenue is coming from low bandwidth connections that are run
at low average utilizations.

Comments?

Note that if you accept this, then we have all sorts of interesting
implications for likely structure of the industry, since this says
most of the costs are at the extreme edges of the network.

Best regards,
Andrew



  -----Please note new address-----

  Andrew Odlyzko
  University of Minnesota
  Digital Technology Center
  499 Walter Library
  117 Pleasant St. SE
  Minneapolis, MN 55455

  odlyzko@umn.edu       email
  612-624-9510          voice phone
  612-625-2002          fax

  http://www.dtc.umn.edu/~odlyzko