Various responses included below.
From: Avi Freedman <freedman@netaxs.com>
Subject: Re: Attempt to summarize Links on the Blink
To: cook@cookreport.com
Date: Mon, 13 Nov 1995 11:16:14 -0500 (EST)
Cc: nanog@merit.eduI would like to try to understand better where this discussion seems to
have come to rest. Yesterday the suggestion was made that the major
providers add more bandwidth to their backbones. There seemed to be no
assertion as to how this could be done.1. OC-3 is not yet routable on backbones. Is that correct?
Well, I was told that the Cisco AIP card can talk point-to-point to another
AIP card at OC-3 speed using either HDLC or PPP.
It is true that AIPs can talk back to back through an OC3 link, however
it doesn't use PPP or HDLC, just sending and receiving cells (i.e. std
ATM, just no switch). So you get 155 Mb/s minus the ATM overhead,
or about 130 Mb/s (depending on what you count as data, type of traffic,
etc.
2. What is the routing impact of parrallel T-3s? Or the creation of
mesh of T-3s? I have the impression that this is not feasible because it
would expand the routing tables unacceptably or because of the questions
of how you would load balance among them??If data can be routed on parallel T3s on a per-connection basis so that
there isn't a scrambling of ordering of packets per connection, then some
benefit is achived, though no single application or site can use more
than a T3 of bandwidth.
This is the way that our boxes work (i.e. destination locked onto a
single interface). Packet reordering isn't too much of a problem
from a functionality perspective for TCP, it just ( 
) affects the
performance. It is also atypical for a single site to need more than
T3 bandwidth (well, they may want it but they can't really but it yet).
3. There seems to be some consensus that we will see an increase in the
numbers of NAP or MAE like interchange points which could cut down on the
traffic that must traverse long haul backbones. *BUT* doesn't each
additional interchange point used by all the top level providers mean
another new set of global routes crowding router memories?It depends. If routing decisions are made locally and the routes heard
at smaller or private exchange points by NSP x are not distributed to
NSP x's larger peering/route-decision routers, then possibly no. That
would mean only hearing routes at private exchange points that were also
heard elsewhere (at a major peering point).
Private interconnects (point to point) between the larger backbone
providers and new "public" interconnects can be used as appropriate
depending on what problem you are trying to solve, and what tradeoffs
you are ready to make.
4. How much help will regional NAPs like Tucson be? Their goal is to
keep local traffic local and off long haul backbones. What liklihood is
there that these will grow in numbers quickly enough to make a
difference? If the majors start showing up at these points does their
arrival mean that the problem of crowding memory in their backbone
routers will be increased?See above.
Avi
This depends on a number of factors like how well the local providers
are able to aggregate routes, and how the local interconnects link
into the rest of the world - for example of the group of local providers
gets a large CIDR block to use as a group then they can appear as a
single prefix to the rest of the net, while providing portability of
numbers for customers that might move between them.
dave