Geo pros and cons

[Iljitsch van Beijnum <iljitsch@muada.com>]

In the discussion about geographical aggregation the past days (and on 
earlier occasions) we tend to get into a lot of detail: which routes go 
where, how many interconnects and so on. These are important details, 
and they need to be discussed. But first we need to consider the 
viability of geographical aggregation.

Geographical aggregation has some problems. For instance, it won't 
scale in the long term if we indeed reach multihoming levels of 1 : 10. 
Another problem is that it changes the way traffic flows between ISPs 
(no more hot potato routing).

But the biggest objection against it is that network topologies don't 
match geography, so geography isn't a reasonable basis for making 
routing decisions. I think this is the one we need to tackle.

Now obviously it is possible to lease a circuit to a remote location 
and connect to "the internet" in that location rather than close to 
home. But should we consider this a feature we must support, or is it 
an exceptional situation that we can safely ignore?

Suppose the long distance connection would have been IP rather than 
ATM, SONET or fiber. In that case, a data from customer A of the long 
distance service provider to customer B of the long distance provider, 
wouldn't travel all the way to the remote location and back, but the 
data would be immediately forwarded to the right destination by a local 
router. So essentially we're expecting IP to compensate for the 
stupidity of non-IP networks...

Then there is the argument that interconnection is limited. Yes, this 
is true. Only a small percentage of all internet users live very close 
to a major exchange location (a single exchange location usually has 
one or more public exchanges and several private interconnects). On the 
other hand, the percentage of internet users that live extremely far 
from an exchange location is also fairly low. Experience has shown that 
there is little incentive to implement additional interconnects if 
there is already good interconnection within 500 - 1000 km. If there 
isn't a good interconnect within 1000 km, the need for one is roughly 
proportional to the number of internet users multiplied by the distance 
to the nearest exchange location.

Since the scalability of geographical aggregation depends on the number 
of internet users and the size of the aggregation areas, where each 
aggregation area needs at least two interconnects, it would seem that 
the scalability of geographical multihoming isn't a problem: more 
multihomers means more routes in an area, but since more end-users 
means more interconnects, the areas shrink. So the number of routes per 
area should remain fairly constant.

I'm interested to hear other views.

Iljitsch van Beijnum