[arin-ppml] Modified Header Forwarding for scalable routing

[Robin Whittle]

Short version:   MPLS doesn't seem to be suitable for the task of
                 scalably transporting traffic packets across the
                 DFZ in a core-edge separation solution to the
                 routing scaling problem.  Even if it was suitable,
                 it involves encapsulation overhead and therefore
                 Path MTU Discovery problems.

                 AFAIK, no RRG proposals involve MPLS.


Hi Matt,

You wrote:

>> Short version:   Owen DeLong suggested a scalable routing solution
>>                  would be to have DFZ routers modified to handle
>>                  packets with modified header structure.  As I
>>                  understand it, the new packet format would contain
>>                  the ASN of the destination ISP and the DFZ routers
>>                  forward it on this basis, without looking at the
>>                  destination address, which would remain unaltered.
> 
> This isn't sounding that far off from just running MPLS in the DFZ
> core, and passing labels around, with a set of label values
> representing the peering edge routers of other ASNs.  Just
> seems like if we already have a code base for passing packets
> along based on a single tag value instead of doing address
> parsing and lookups, with routing of the internal packet being
> handled at the far end like normal after popping off the intervening
> tag, we might consider leveraging an encapsulation format that's
> been widely deployed and operated for the past decade.

I think what Owen and I are attempting to achieve, with our different
proposals, is a system which doesn't add to the length of the packet
at all - and can still be used to provide multihoming, portability
and TE for very large numbers of end-user networks, without
overburdening the DFZ.

The focus of my plans is to reduce the burden on the FIB of each DFZ
router, and more importantly the RIB and therefore the DFZ's BGP
control plane.

I think Owen wants to reduce the FIB size, but is so concerned about
the RIB and the general BGP control traffic, stability etc.

Establishing an MPLS path involves a great deal of control plane
activity, and involves significant state in the FIB.  The label value
is changed at each router, and the path is a point-to-point path.  I
can't see how point-to-point paths could perform what we need here:
getting traffic packets from ITR or ITR-like device (really from any
one of potentially millions of them) across the DFZ to some ETR or
ETR-like device which restores the packet to its original form and
somehow gets it to the destination network.  These paths need to
change quickly for the purposes of multihoming service restoration
and inbound traffic engineering.

MPLS's extra packet length represents inefficiency, especially for
small VoIP packets, and raises difficult problems with Path MTU
Discovery.

So I don't think MPLS is a suitable technique.  AFAIK, no RRG
proposal uses MPLS.


> This probably would be a better discussion on NANOG or some
> other more routing-operations focused list, rather than a numbers
> resource policy list, though.  ^_^;

Maybe so, but we just had 40 or so messages on Moore's Law and about
the DFZ.  The RRG is focussed now on discussing proposals which have
been registered.  Owen's proposal is not registered, so I think PPML
is a good place to discuss it - but if the moderators say otherwise,
that's fine.   My two proposals are part of my Ivip proposal with the
RRG, so the RRG would be a good place to discuss them.

 - Robin