[arin-ppml] Policy Proposal 119: Globally Coordinated Transfer Policy - revised

[William Herrin]

On Wed, Nov 24, 2010 at 5:22 AM, cja at daydream.com <packetgrrl at gmail.com> wrote:
> How do folks feel about how transfers between regions would affect the
> routing table?   We would be in essence punching holes in RIR blocks.   This
> depends entirely on whether there are ISPs filtering or aggregating on
> regional boundaries based on RIR /8 block allocations.

Hi Cathy,

There's a low but non-zero probability of harm to future routing
operations from allowing inter-region transfers.

At the network edge, it's possible to do a few tricks with routing
that reduce your table by tying most outregion routes to a few
unfiltered ones. However, these tricks are not an especially large
improvement over accepting a default and simply filtering the distant
routes.

BGP RIB compression has been all but demonstrated impossible for
transit ASes (as opposed to origin-only ASes), so there's no
anticipated impact there. You just can't aggregate that way no matter
how distant; the more specific prefix rule means your downstream will
send traffic to his other ISPs if you in any way aggregate, damaging
the product you supply.

There is a RIB parallelism scheme in which address blocks transfered
out-region might experience inefficient routing. The idea is that you
only keep the full APNIC-region RIBs in your APNIC-region routers.
Your ARIN region routers hold the full ARIN region RIB but they only
hold the APNIC /8's, expecting to encapsulate packets for those
destinations in MPLS and send them to the APNIC-region routers for
further processing.  Routing for the few prefixes transferred from
APNIC to ARIN might then suffer multiple ocean crossings.

The RIB parallelism scheme only exists on a drawing board; I know of
no examples where it's in production and it would probably require
modifications to MPLS implementations which allow routers to associate
tags with particular exits instead of just particular routers.
However, inter-region transfers could complicate or even foreclose its
future use.

FIB compression could be impacted by reducing the likelihood that
nearby prefixes all point out the same interface. The idea with FIB
compression is that you figure out which exit is the destination for
most of your routes, set a default pointing at that exit, and then
keep only the FIB entries that direct packets somewhere else. Do this
recursively and you can get the today's FIB down to between half and a
quarter of the size of the RIB at a cost of losing the implicit
authoritative set of null routes.

However it's not at all clear that the way multinational backbone
networks connect hasn't already made whatever impact will be made to
FIB compression. There's an Internet Draft somewhere close to becoming
an RFC on FIB compression but I'm not aware of its use anywhere in
deployed software.

So all told, there's a low but non-zero probability of harm to future
routing operations from allowing inter-region transfers.

Regards,
Bill Herrin


-- 
William D. Herrin ................ herrin at dirtside.com  bill at herrin.us
3005 Crane Dr. ...................... Web: <http://bill.herrin.us/>
Falls Church, VA 22042-3004