Path Maximum Transmission Unit Discovery Pre-WG (pmtud)
Thursday, 17-July-2003 from 13:00 to 15:00
The meeting was chaired by Matt Mathis and Matt Zekauskas. Al Morton,
Itojun Hagino and Matt Z. took notes that were assembled into these
minutes by the chairs.
* Preliminaries: Blue sheets, Note takers, etc
* WG Status
* Short history and work to date
* Robustness Issues
* Other Stakeholders
Matt Mathis led off the meeting presenting the new co-chair, the agenda, the
changes to the proposed charter, and the aggressive milestones.
The group status is that some parts of the administrative
preparation did not get done, but IESG has approved, hence "Pre WG". This
will be a fast development, silence will be acceptance (at the start as
sections are integrated). PMTUD is a re-activation of path MTU WG, which
was a very similar effort. To participate, you must subscribe to IETF list
(firstname.lastname@example.org). Matt Zekauskas volunteered to Co-chair the group. The
charter was broadened so as not to restrict to a single method.
Milestones are aggressive, and need for implementation to test is clear.
No one disagreed, nor were there suggestions for other methods to study.
Matt Mathis will make another editorial pass at the document (which was a
rewrite, starting from RFC 1981, instead of an update to the previous BOF
input). Sections will be added based on mailing list comments and any
input from stakeholder communities.
Matt Mathis sketched the previous algorithm and noted some of the
problems. He then sketched the new algorithm, noting that there is just a
small amount of MUST/SHOULD language: under what circumstances can losses be
ignored as a congestion signal. The rest is heuristics; it doesn't need to
be the same for every application, and permits vendor diversity.
Question: when can the algorithm be used by TCP? Just after 3-way
handshake, or before real communication? Matt responded that it uses live
payload data, and the draft has a recommendation not to attempt the
algorithm unless the congestion window is at least twenty packets, so the
connection is well established before the algorithm starts. Thus, this
could slow down tiny files -- the exact algorithm is a heuristic, so you
could choose to perform it differently. There are tradeoffs.
The request for collaborators in the IPsec & security area led to a big
discussion on tunnel issues. People were positive about the method, but
there are corner cases to consider. Input has been promised to the
mailing list. (In the IPv4 world, lack of PMTUD is noted as a major
problem with IPsec VPNs and providing services.)
This started around slide 11, "Plans for the Next Draft". One of the
collaborators was folks from the multicast area; one possibility is a
generalization of the algorithm for reliable delivery. This would solve an
ICMP implosion problem if the current MTU discovery technique was used.
Dave Thaler noted that the behavior as specified in IPv4 and IPv6 was
different, in 6 you respond, in v4 you don't.
Another collaborating group would be IPsec; currently the security
architecture document has major sections dealing with the
interaction of MTU discovery and IPsec (because tunnels are created); the
new technique might obsolete many of those sections.
Itojun Hagino noted that the interaction between IPsec and and TCP
depends on if the TCP stack is aware of IPsec. If the TCP stack does not
take care of the IPsec header size, the algorithm would need to be
Matt M. responded that the detail in the draft needs to be resolved in a
consistent way. You can count the IPsec header as part of the IP header or
TCP header. The really nasty cases involve additional layers, for
example IPsec on a VPN, ICMP messages could go back to the wrong place.
Michael Richardson expanded on this as an IPsec implementer. The worst
case is common at meetings such as this -- you have a corporate address on
your laptop, and a VPN back to the corporate space, so all traffic goes
back to HQ. Try to visit a bank, and they have ICMP filters. Your
gateway is sending out ICMP messages to the bank, and they drop them. This
proposed algorithm should work really well. Many times VPNs are blamed
(since they are the newest element in the path), when the problem is
really a bad ICMP filter.
However, there is a problem, if you raise the MTU and the tunnels do not
toss large messages but fragment them anyway, you will end up always
fragmenting. Michael noted that his Linux implementation (he's the
FreeS/WAN technical lead) did not honor the DF bit by default. Having a
poor(er) performing implementation was better than one that didn't work at
all. ("Poor performance is better than no security.") Perhaps there could be
a heuristic that worked for a short term solution so these mechanisms
don't interact badly... the endpoint would need to be updated for this
algorithm, so IPsec tunneling could be updated at the same time. This
behavior is often a kernel option, too. [In reviewing the minutes,
Michael related that:
"The key point is that I, the IPsec developer can't control:
1) the ICMP filter.
2) the TCP on the machine behind it.
I can have *some* influence on the TCP at the receiving end of the flow, but
not a lot. *If* the IPsec tunnel terminates on the same machine as the TCP,
then in theory, the TCP can learn about the reduced MTU, and set the MSS
appropriately. In practice, probably only KAME, Microsoft and Sun are well
enough integrated to do this right now. Probably Linux 2.6 will be able to do
so as well. We do have an option to hack the MSS on the
encapsulator's side already, alas."]
Perhaps you could fragment into tunnel, but retain the DF bit, and if set
don't do anything weird. Itojun related KAME experience; there they ended up
not setting the DF bit on output header when IPsec tunnels are created.
Another point was that IPv6 on IPv4 tunnels have the same issue. IPv6
tunnels should have a MTU of 1280 by default so a minimum MTU can be
Matt M. mentioned that he's aware that a large number of tunneling
implementations don't copy the DF bit from inner packet to outer header.
He's not yet sure if the document needs a specific section covering
tunnels and tunnel migration; an intermediate ground that works is to let
tunnels behave this way in the interim, and discourage a mode where end
systems ignore can't fragment messages.
Dave Thaler noted that mobility might add additional headers; so a tunnel
MTU of 1280 might not be enough; 1380 would be better.
Itojun stated that he was thinking of configured tunnels and not mobile
IPv6. If you send a packet with mobile headers, the TCP stack needs to be
aware of the size of the mobile IP headers and reduce MSS
appropriately -- maintain the total MTU size.
Matt Mathis noted that there was definitely a subgroup interested in
considering tunneling and MTU discovery; he encouraged folks to join the
mailing list and contribute the various circumstances where there are
Lars Eggert mentioned that RFC2003 specifies some things related to MTU
discovery, and RFC2401 specifically prohibits some of the mechanisms in
RFC2003 for security reasons. Joe Touch also has a relevant draft:
As to other transport protocols, Matt Z. reported that he had quickly
skimmed SCTP and DCCP documents, and that SCTP looked possible, but DCCP
says specifically that the MTU can't be raised. No one that claimed to be an
SCTP expert was in the room (or at least didn't comment negatively on the
applicability to SCTP). Eddie Kohler noted that this behavior was
revised in the DCCP WG meeting this week. Matt Z. prompted him to send
some DCCP text.
Matt M. emphasized that the point with getting a draft done early is to
encourage implementation as soon as possible. The algorithm will use
specific details of other protocols, and we're dependent on the
uniformity of implementation of certain features. We need to learn what
implementations really do; ideally get a custom implementation run on
servers and real field data to feed back into the document.
Matt M also noted two cases that he's worried about (although these are
just examples; others are encouraged to consider other cases, or report
back implementation experience). First, what happens if a path is
striped across multiple links, and the MTU is not the same across the
stripes? You can require that the MTU is not raised until a certain
number of segments are received successfully. You need to understand the
interaction between random losses and whether the MTU is or is not
raised. Second, what happens if there is a parametric failure -- when
raising the MTU causes the error rate to increase? An actual case is one
particular 10G gbic; it was error free with 1500 byte packets, but not with
9000 byte packets. There is an opportiunity for different heuristics
here, for example use a smaller MSS if you cannot fill a window.
Michael thought a "brokenness test page" was needed -- a good
For hard, repeated, timeouts the first thing you want to do is reduce
congestion variables, then reduce MTU. At some point want to restart the
checks to increase MTU.
There are other possible protocol interactions, too: for example, SCTP can
use multiple endpoints. What if it changes addresses, and the new path has a
Michael felt that it was important to focus where the production
environments hurt most with the current MTU scheme. Matt M noted that
different things hurt in different environments. Michael expanded that the
most frequent case will likely be large port 80 responses to a client. And
it's the client that would decide that the path is stupid or broken, and
other A record should be tried. The web server is getting the
timeouts, not the client. This won't deploy if we can't solve the web
Matt M noted another case he had thought about, but not seen: what
happens if raising the MTU causes link stability problems (as opposed to
hard failures) -- say the link "goes away" for 10 seconds and then
returns. He's thought about using a state machine to catch this case... the
link is broken, and we don't necessarily want to fix it with an MTU
On ignoring DF bits, so that a tunnel fragments large packets: Matt M
contended it was worse for a 1500 tunnel fragmenting a 9000 packet than a
tunnel fragmenting a 1500 packet by the tunnel overhead. Michael didn't
understand this at first; Matt explained that the problem is that with many
fragments the odds are greater that you lose a fragment, and hence the
whole packet than if there are only two fragments.
In thinking about other stakeholders, Magnus Westerlund felt that the
algorithm would work for RTP over UDP with the use of RTCP extensions for
packet loss vectors.
Itojun said that we should contact email@example.com for SCTP.
In the multicast case, Dave Thaler argued that this algorithm might cause an
ACK implosion that is worse than a ICMP-message-too-big implosion, since
there are typically far more receivers than there are
Magnus commented that the document as written is very TCP specific. The
algorithm should be better separated from actual deployment. Matt M said
that's the intention.
Another question was what, exactly, is the definition of MTU?
End-to-end or link-specific. Matt M said that we were talking about IP MTU
when using a particular link layer; how the IETF uses it, not what
hardware specifications say.
Dave Thaler argued that IPv6 might not need this at all; the algorithm
could arguably make performance worse (since the MSS size would ramp up
instead of being decided once by ICMP-message-too-big). Matt M. said the
new algorithm would prevent against implementation or
configuration bugs and also work in the cases where L2 MTUs were
different on a switch.
One audience member said that this should be documented in detail.
Dave Thaler mentioned that filtering ICMPv6 has larger problems. IPv6
neighbor discovery uses ICMPv6, so if ICMP is filtered you won't get
connectivity. In addition, since v6 has no DF bit (but implied DF on all
packets) blocking ICMP definitely leads to black hole problems in the
network. Thus, there is already a natural incentive to allow PMTU using
Matt M noted that some stacks have IPv4 mimic IPv6 -- they always set DF,
even on fragments, and attempt to fragment only at the endpoints.
However, there's no requirement that routers send the too big messages in
v4, but there is a requirement in v6.
Another comment was tha the MTU in Router Advertisement messages should
solve this problem. If operational experience says this isn't
happening, it should be reflected to the v6 working group.
Matt M said that in all cases the problems with path MTU discovery are
bugs. There are a large set of problems.