Re: PI/metro/geo [Re: The state of IPv6 multihoming development]

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

On Mon, 28 Oct 2002, J. Noel Chiappa wrote:

> If people don't like some of the properties that being connectivity-based
> induces in routing-names, then the answer is simple: define another
> namespace which has characteristics you like better, and map back and forth
> between the two.

> If the price of that is too high, then you just have to deal with the
> hassles of connectivity-based names, just like we have do live with the
> hassles of friction, entropy, etc.

So what is it that makes "connectivity-based" names so good and
everything else so bad? Unless I'm much more mistaken than I thought
possible, the relationship between a connection and a name/address is
purely contingent. Things would be different if routing were
hierarchical, which it isn't. Let me draw an ASCII picture:

	     G
	    / \
	   /   \
	  E     F
	 / \   / \
	A   B C   D

If the only way to get from A to D is A -> E -> G -> F -> D everything
gets very simple and very scalable. However, not only does this not
match the way things work--it doesn't even come close. There is
absolutely no way to make real-world interdomain routing fit inside a
hierarchy.

In reality, each AS occupies a place at the top of the hierarchy. Some
ASes have many lower layers inside, some don't. But if an AS could be
hidden completely behind another AS, the AS wouldn't qualify for an AS
number and not exist in the BGP view of the world to begin with.

This means we have to live with a very large number of prefixes in our
routing tables. Fortunately, a single big thing can usually be split
into a larger number of small things to make it more manageable. Four
smaller routers that handle 2000::/5, 2800::/5, 3000::/5 and 3800::/5
respectively, are probably cheaper than one big one that handles
2000::/3. Now if a multihomed customer of two ISPs with address range
3333:abc:def::/48 were somehow to connect to routers that handle the
3000::/5 portion of the routing tables at both her ISPs, that would
certainly make things easier. And if the 3000::/5 routers of those ISPs
could talk to each other, that too. It wouldn't be necessary per se,
just easier.

Iljitsch