Re: Regionally aggregatable address space for multihoming

[Brian E Carpenter <brian@hursley.ibm.com>]

Iljitsch van Beijnum wrote:
> 
> On Mon, 11 Jun 2001, Brian E Carpenter wrote:
> 
> > This is my problem with this proposal, as it has been with every geo addressing
> > proposal over the last 5 years.
> 
> > Because of this fact (you have to increase the area covered until it is
> > big enough to be flat-routable), I think that every geo area would
> > reproduce exactly the problem of entropy and default-route-table bloat
> > that we have in today's tiny little Internet.
> 
> When the number of routes per area grows, it is likely that the areas
> themselves shrink, if only for this very reason. I'm not saying geographic
> aggregation will magically make all the problems go away. However, it will
> buy us two or three orders of magnitude, which is nothing to sneeze at.

We are not communicating. You deleted the text that my initial "this"
refers to, but I think the exact opposite is true: the areas will have to
grow, since ISP interconnections are essentially random with respect to
geography. Also the cost pressure is going to be against local
exchanges and in favour of large regional exchanges. Given this (and
other points noted below) I don't see that geo addressing buys us
anything much at all, certainly not orders of magnitude.

> 
> > So I have little confidence
> > that this solution is scaleable to the 10 billion node Internet.
> 
> The number of nodes is not very relevant: you can fit 10 billion many, many
> times inside the /48 a single customer would get. All that matters is the
> number or routes and the resources needed to transmit, process and store each
> route. 

Ignoring household subnets, we can expect several orders of magnitude more
business subscribers than today - that's the significance of the total size
of the network. 

> In IPv4 aggregation isn't used to its full potential, for historical
> and address depletion reasons (besides some ignorance) and because of
> multihoming. In IPv6 it will be much simpler to agressively aggregate,
> because there is little need to give ISPs only small blocks of address space
> and have them come back for more as they grow.
> 
> So the main threat to the routing table are the multihomed networks. 

Agreed.

> ...It would
> be nice to know how many people or business will be multihomed in the future,
> but there are no real figures. We only know that the number is still
> relatively small but rising fast. I think we can get away with assuming an
> upper limit on multihoming of 1% of the population. This is something like
> 10% of all businesses. 

The only safe design target is 100% of all businesses, i.e. several orders of
magnitude bigger than today's Internet. I see nothing that would limit it
to 10% or any other particular level; it is likely that multihoming just becomes
the normal thing to do.

> Obviously anything that does better than 1% would be
> great, but if we can achieve 1% we'll have bought a lot of time to think
> about the other 99%.

Not if the solution is worse than the problem. I believe that geo addressing
simply reproduces pre-CIDR IPv4, since it is based on unrealistic
assumptions about inter-ISP connectivity.
> 
> With a world population of 10 billion looming, this would be 100 million
> routes to multihomed networks. This is too much for any single router.
> However, if we can break up the world into 100 regions, this would only be 1
> (or rather 2) million per region, which is still a lot but managable, if not
> now at least in the near future.

It would, if such a scenario was plausible, which I don't believe. 

> 
> Combined with more efficient ways to store and transmit routes, this should
> work for a long time to come.
> 
> The current routing paradigm is to store a lot of information about each
> individual route, since aggregation makes sure there aren't very many similar
> routes. But in a few years (decades) when 1% of New York City is multihomed,
> this means more than 200k routes (to 100k destinations) within a fairly small
> geographic region. If this market is serviced by 20 ISPs this means 10,000
> routes per ISP, each taking more than 100 bytes of memory, even though these
> routes are likely to be nearly identical.
> 
> If we assign all these addresses out of a single 131072 * /48 block and then
> each ISP uses a bitmap attached to the aggregated /31 to announce whether any
> of those /48's is connected to them, it only takes somewhat over 100 bytes
> for the /31 + a 16k bitmap for a total of ~16500 bytes in stead of more than
> a megabyte. Another ISP that peers with all of them will only get some 330
> kilobyte of routing information in stead of 20 MB.

Perhaps, but every time there is a change they will get it all again instead
of just one item. There are multiple tradeoffs here. I'm all in favour of
summarisation but that is not the only performance issue.

> 
> My idea for geographical aggregation can be deployed in the current IPv4/BGP4
> Internet without breaking anything. The only thing that has to be done is
> divvy up the world in regions and assign address space to multihomed networks
> within the same region from an aggregatable block. Then networks can start to
> filter out routes from blocks far away without the risk of suboptimal routing
> to multihomed networks closer by if and when they want, with potential
> benefits of new route storage paradigms in the future.

Only if you revolutionise ISP interconnection strategies and find some way
to change the underlying economics.

Don't misunderstand me - we need a radical solution. I just don't think that
geo addressing, which is fundamentally the telephone solution, is radical
enough.

   Brian