CURRENT_MEETING_REPORT_
Reported by Jeffrey Mogul/DEC
Meeting of 12 December 1989
Held at the Digital Equipment Corporation Western Research Laboratory
Palo Alto, California
AGENDA
1. Discuss proposed MTU Discovery protocols
2. Attempt a consensus protocol
3. Find a victim to write up a draft
4. Set next meeting date
MINUTES
This was the first meeting of the MTU Discovery Working Group. We had
already made and discussed a number of proposals for a protocol design,
and had also discussed (over the mailing list) a number of technical and
non-technical constraints on such a protocol.
The most important non-technical constraint was that we wanted to devise
a protocol that would work fairly well, or at least no worse than ``no
protocol at all'', in the presence of large numbers of unconverted hosts
and routers.
Technical constraints include issues such as the cost of processing
options; the possible use of the reserved bit in the IP header [probably
not available to us]; the layers in which the MTU discovery protocol is
implemented; the need to support TOS, security, and asymmetric paths;
deciding when to send IP options; the problem of LANs with more than one
MTU [a possible consequence of the use of bridges between Ethernets and
FDDI]; the delays in propagation of information through the network; the
realization that the path MTU is first known at the wrong end of the
path;
The proposals made before the meeting basically fell into two
categories: those that probed the network to find out the path MTU (the
minimum MTU over a path), essentially by asking the routers along the
path to report the MTU, and those that asked the receiving host to
report fragmentation when it occurs, with the understanding that the
size of first fragment of a packet received probably reflects the path
MTU. In the latter case, the sender must send large packets occasionally
in order to discover if they will be fragmented.
The ``Report Fragmentation'' approach would be cleanest if we could have
used the reserved bit in the IP header as a flag to tell the receiver
that the sender is willing to receive and utilize a (new) ICMP
1
Fragmentation Occured message. However, we were told that this bit is
unlikely to be released for this purpose. ``Report Fragmentation''
(R-F) can also be done using an IP header option, again ``allowing'' the
receiver to send the ICMP report.
R-F has the advantage that it requires no changes in the routers, but
the disadvantage that it requires changes in most end-hosts before it
does much good. (If the receiving host does not implement it, the
sending host could obliviously continue to send packets that are being
fragmented and perhaps lost.) During the meeting, we discussed how to
decide when and how often to send the ICMP Fragmentation Occurred
message, and which layer should send it.
Since option-processing is expensive for routers, we believe that the RF
option cannot be sent on every packet. Thus, the receiving host should
remember that a sender has specified Report Fragmentation, and if
fragmentation does occur later on, the ICMP should be sent even if the
fragmented packet did not carry the option.
Since some protocols on the sending host (e.g., NFS) might not be able
to use the MTU information even thought others (e.g., TCP) might have
requested it, some felt that it was necessary to send fragmentation
reports in response only to those host+protocol+port tuples that had
sent the RF option; this means that the receiving host must keep a table
keyed in this way, and probably that it has to be maintained by the
transport protocols (TCP, UDP, etc.) At the same time, it was felt to
be unfortunate that the transport protocols would be burdened with this
chore; we consider it an ``implementation suggestion'' rather than a
protocol specification.
One approach might be to insist that a host not send the RF option
unless all of its protocols are willing to abide by MTU change reports.
We know that it is possible to make NFS obey certain MTUs, and perhaps
it is worth making this rule rather than complicating the discovery
protocol implementation.
Talk then turned to the ``MTU Probe Option'' approach. In this approach
(as it evolved from RFC1063 before the meeting), the sender would
(occasionally) send this option on packets flowing on the connection.
The option would have three fields: a ``next hop IP address field'', an
``OK'' field, and an ``minimum MTU'' field. The initial values would be
the first hop router address, ``true'', and the MTU of the first-hop
link. Each router along the path would set the MTU to the min of the
current value and the MTUs of the incoming and outgoing paths. If its
own address was NOT the same as the ``next hop'' field, it would set the
OK field to ``false''; otherwise, it would change the next-hop field.
In any case, the option is forwarded (but not copied on fragmentation).
The last-hop router (it should be able tell that it is such; we'll
address the FDDI-Ethernet bridge issue somehow) does the same thing to
the option as the previous routers, but if the OK field is still
``true'', it now knows the path MTU. It therefore sends a (new) ICMP
2
Path MTU message back to the sender, including the usual IP header info
for matching at the sender. In any case, the option continues to the
receiving host.
Note that the sender will not get a Path MTU message unless every router
along the way understands this protocol. This is because we could
otherwise be misled by a low-MTU link bordered by two unconverted
routers. However, we believe that routers will be converted much sooner
than most end hosts.
Since the option is also received by the ultimate receiving host, that
host can also interpret this option as ``Report Fragmentation'' flag (as
above). This is a backup mechanism; if the Path MTU message is not
generated, or if the MTU then decreases enough to cause fragmentation,
the sender will still find out. Of course, the sender cannot know that
the receiver is doing this. One partial solution is that if the
receiver gets a MTU Probe option with the OK field = false, then it
should send an Path MTU message to the sender with a code indicating
``unknown''; this tells the sender that it is OK to use the ``Report
Fragmentation'' approach. If the sender receives neither kind of Path
MTU message, then it must assume that it will not receive Fragmentation
Occurred reports, and it should stick to the current ``no more than 576
bytes if non-local'' rule.
This hybrid approach seems (to the attendees, at least) to combine the
best of both methods: it gives accurate, early results (i.e., before a
connection has to start sending big datagrams) without incurring
fragmentation if the routers cooperate, it detects fragmentation if the
receiver cooperates, and it causes conservative behavior otherwise.
There are still several issues to nail down, including how often to send
the MTU Probe option, how many times to send the Fragmentation Occurred
ICMP message, etc.
Keith McCloghrie and Rich Fox were volunteered to write up a draft RFC,
which we hope to circulate several weeks before the next IETF meeting.
We expect to meet again at the February IETF meeting.
ATTENDEES
Art Berggreen art@salt.acc.com
Noel Chiappa jnc@PTT.LCS.MIT.EDU
Farokh Deboo sun!iruucp!ntrlink!fjd
Steve Deering deering@pescadero.stanford.edu
Rich Fox sytek!rfox@sun.com
Ivan Liu iliu@orville.nas.nasa.gov
Keith Mc Cloghrie sytek!kzm@hplabs.HP.COM
Jeff Mogul mogul@decwrl.dec.com
Nuggehalli Pradeep pradeep@orville.nas.nasa.gov
Stephanie Price cmcvax!price@hub.ucsb.edu
[Noel Chiappa participated via telephone]
4