a radical proposal (Re: protocols that don't meet the need...)

Thu Feb 16 11:01:08 UTC 2006

Mike Leber wrote:
> In line with this... (I had to point this out in a different context on
> another list before):
> 
> Networks announce prefixes because doing so makes them money.  Networks
> that listen to these prefixes do so because that makes them money.

Indeed.  That is precisely the reason why we'll see IPv6 de-aggregates
routed in the future just the way we see it with PA space and the swamp
today.

> In light of this, the current size of the routing table is a function of
> the number of businesses or organizations that would like to multihome
> limited to those that can afford to pay for the capital costs of a router
> capable of doing BGP, the operational cost of maintaining it, and the
> service cost of buying Internet connectivity that includes BGP sessions
> with their transit providers.

Well said.  $realworld at work.

> Where ever you see that function trending is how big you can expect the
> routing table to become because of economic pressure.
> 
> While there are not as many businesses and organizations as people on the
> planet, as an exercise imagine 4 billion prefixes.
> 
> For the sake of simplicity assume a 32 bit forwarding tables (4 billion
> entries) for each interface on a router expandable to 256 interfaces, with
> a byte (256 possible forwarding destinations) per entry for forwarding (4
> GB of RAM).  Such a thing could be made now and would attribute to a very
> small fraction of the cost we currently pay for new Cisco cards.  You
> might not get as good port density for the physical form factor, however
> it is doable now.

Hence my case for perfect match.  Doing 4 billion prefixes with longest-match
is doable as well, but either more expensive when you want constant lookup
times or you have to live with varibale lookup times and the max pps is defined
by the traffic and route pattern.  Perfect match is the reason why even the
cheapest DLink gigabit ethernet switch is able to switch packets at wirespeed.

> This is separate from the convergence discussion for 4 billion prefixes,
> etc etc etc.  The link speed required to be able to converge within a
> minute is left to the reader.  heh.  :)

BGP can be made more efficient with this large number of prefixes.  Don't
worry on this.

-- 
Andre