Re: Simple observation

[Greg Maxwell <gmaxwell@martin.fl.us>]

On Wed, 4 Apr 2001, Niall Richard Murphy wrote:

> Hi there,
> 
> On reading the multi6 multihoming requirements draft, it strikes me that 
> given the IPv4 global routing table scales as per the number of multihomed
> entities, what we really want the IPv6 global routing table to do is
> scale /as the number of entities you can multihome to/.
> 
> Would this be correct, or even near?

Define scaling.

Today, the IPv4 global routing table grows as per the number of multihomed
entities and providers. 

As the Internet becomes more important to more people, I can't see the
perportion of sites which wish to multihome for realibility signifantly
decreasing. Furthermore, with the hidensity of Internet connectivity
increasing, more groups will want to do link sharing, and other activities
that increase the IP routing burden.    

Consider a big, statically mapped, switched, arp-less, subnet that
consists of web servers and consumers who randomly communicate. As long as
you can keep the cross sectional bandwidth sufficently high, the system
scales infinatly (1 user + 1 server an additional (tps/pairs) transactions
per second).

In contrast, the IPv4 global routing table does not scale. At all. The
only way it's capacity for more routes increases is when people upgrade
the routers, adding routers doesn't help.

The more hosts you place on the network, the more multihomed networks
there will be, and more of the globally limited routing capacity will be
consumed.

Ideally, a great solution would cause the routing burden to decrease when
more hosts are added to the network. A slightly less optimistic goal would
be for it not to grow as you add hosts and insted only when you add
big providers (something that grows logarithmically with the number of
hosts/users).

This scaling has multiple dimensions. Space (route table size), Time
(router CPU, convergence), Bandwidth..

I can't conceive of a IP level multi-homing solution that improves global
route table scaling in any of those criteria without sacrificing one or
both of the others. 

Because of this, it occured to me that the only way to keep scaling
without losing end to end transparency is to decenterlize part of the
routing process by placing it into the transport layer. 

After doing some research, I came opon SCTP which does exactly this but I
was confused when I was unable to find a general discussion of transport
level multi-homing as an Internet scaling maneuver. Which is why I started
going around asking if this was being investigated.