Gordius has left the building. Was: RE: The Gorgon's Knot.

Deepak Jain wrote:

The problem is where the RIB, or forwarding table is exported to
the line cards which frequently have less than 128MB of usable
RAM/SRAM/memory storage/etc. This essentially means that the line
cards can only directly talk to other line cards for a specified,
limited number of routing prefixes. I do not know the algorithms
used when the line card is out of memory, but in many cases this
memory is not field upgradeable beyond a certain point.

Srinivasan, Vargese, and others have demonstrated / observed that space
is not a material problem for the architecture you describe. Rather, it
is the time to complete a longest-match (or equivalent) operation for
each forwarding event, especially for high(er) interface speeds.

Regardless, 128MB is a LOT of table memory, especially when it is
virtually unknown to see anything larger than 10-20Mbit CAMs; prior to
reduction / expansion / etc, there's scarcely 11B of unique information
in an IPv4 route. Adding in some bytes for mysterious 'other stuff',
this is easily 4-8MM routes. We'll have to see another knee in route
proliferation before this becomes a problem, and space exhaust would
most certainly prevent this from happening.

Many smaller networks can and do use PCs for BGP and for
forwarding because their total forwarding needs at their core are
say sufficiently less than 800mb/s upto which PCs seem to handle.
However, the desires and models of these smaller networks don't
scale much beyond this level with currently available PC

Agreed... as regards PCs lets remember that history has offered certain
commentary on the non-role of PCs (indeed general purpose computers) in
network infrastructure... viz. there aren't many GRFs / DDP516Rs / etc.
around anymore. I doubt this is an 'interesting problem' for the op or
planning community.

Andrew Bender

Ah, but with what minimum prefix length:
a) Verio-esque
b) /24 - current 'global minimum filter length'
c) /32
d) A mixture between (a) and (b) - current situation

People need to recognize that there is defacto filtering going
on out there is everyone's network (well, nearly everyone's),
in that very few people accept longer routes than a /24. In
a CIDR world that's a different filtering rule to the Verio
one, but it's still arbitrary (in some senses - think non-RIR
assigned class A space - rather more arbitrary).

IE is the assumption space exhaust will precede the problems
you predict on router hardware precisely BECAUSE RIR allocation
rules (the more sensible ones), and the /various/ filtering,
dampening policies and the othr pro-aggregation work, has
determined a minimum useful prefix size? In which case getting
rid of the policies which make your assumption correct would
be a bad thing.