The MPLS WG Archive

[David Charlap <david.charlap@marconi.com>]

Markus Jork wrote:
> 
> There is now a new section 2.6 in draft-ietf-mpls-rsvp-lsp-tunnel-06.
> 
> It basically says that an ingress LSR has to keep track of the Path
> MTU for each LSP (an information that is gathered by RSVP) and
> fragment all packets it sends into that LSP accordingly.
> 
> Where does this requirement come from now and why is it a "MUST"
> requirement? I think this puts too much burden on the ingress LSR. So
> this really should be a "MAY".  A midpoint LSR has to be able to do
> the fragmentation if required.

There are several reasons.  A big reason is that transit routers might
not be capable of fragmenting labeled traffic.

Another is that it is difficult (if not impossible) to guarantee end-to-
end bandwidth and/or delay reservations if transit routers are allowed
to fragment packets.

The reason is simple - fragmentation increases the amount of bandwidth
used - by adding more headers to the stream.

If the amount of traffic entering the LSP is at (or close to) the
reserved capacity, fragmentation will cause this traffic to violate the
reservation after the node where the fragmentation occurred.  This may
cause the packet to be dropped by a transit node in the tunnel, which
would be bad.

I think it is generally assumed that traffic entering a tunnel,
compliant with that tunnel's reservation, will emerge from the other
end.  (Not counting failures or best-effort tunnels, of course.)

If all fragmentation is done by the ingress node, then it can make sure
that all traffic entering the tunnel is compliant to the reservation.

Additionally, doing all the fragmentation in one place means that less
fragments will be generated, and that most of the fragments will be of
maximal size.

For instance, imagine a case where an 8192-byte packet hits an MTU-1500
link, then an MTU-1009 link, then an MTU-576 link.  If each router does
its own fragmentation, you end up with 17 fragments; if this packet is
fragmented directly to 576, however, you end up with 15 fragments.  You
save the bandwidth of two IP headers, MPLS headers and layer-2 headers. 
If you get a lot of oversized packets, this can be significant.  (It can
be even more significant if there is a wider variety of MTUs in use in
the network.)

-- David