On 10/02/2017 03:25, Brian E Carpenter wrote:
Stewart,
On 10/02/2017 04:19, Stewart Bryant wrote:
...
I wonder if we would best serve both our future and our heritage
if we declared RFC1981 as historic, and either left the idea there,
or declared it as historic and wrote a new text from a clean start?
I don't see that. It's a stable, widely deployed, interoperable
mechanism. That is rather orthogonal to the issue that has been raised,
which is that faulty ICMPv6 filtering blocks it on many, many paths
across the Internet.
I will not debate whether it is faulty or not, but it seems that in
practice the
Internet breaks the mechanism. However it breaks it is a way that seems
disruptive to some user traffic. The document is really guidance
one how hosts might use ICMP for optimization, and arguable need
not be a standard at all.
My remark about heritage is that this vintage draft is very much a
product of
its time, and really needs modernizing, and after modernizing ought to
look quite different, and thus maybe we should employ a procedure
other than a simple replacement.
...
It is concerning that the draft does not talk in any detail about
how modern ECMP works, i.e. using the five tuple, and noting that
the PMTU may be different depending on the transport layer port
numbers.
Has this problem been analysed for, say, IPv4? And does the real world
contain ECMP setups with different MTUs on different paths?
I don't know if anyone has looked. Since the mechanism is
self-correcting albeit
with some disruption to user traffic it looks to the application and the
application
user, just like the Internet not working for a few moments.
In a well managed SP network there should not be, but neither should there
be asymmetric path costs, but there are. The less well manage private
networks are less well managed.
Given that a very large fraction of packets will traverse an MPLS
network at some point, I am surprised that there is no text talking
about the importance of providing support for this feature in the
MPLS domain. RFC3988 talks to this point, but is only experimental.
I don't understand. How does the fact that there might be some MPLS
segments along the path affect end-to-end PMTUD?
The point that RFC3988 makes is that MPLS looks like a single hop to IP
and the
PE has to fragment or has to reply with an ICMP error message to support
PMTUD. MPLS has ICMP extensions, but I don't know if they integrate to
result
in the right response at the end node.
My point is that the draft is silent on the subject, and perhaps it
should not be.
However your question make me ask a further question. The draft is also
silent
on NATs. Is there any advice needed for people designing and configuring
NATs?
======
If flows [I-D.ietf-6man-rfc2460bis] are in use, an implementation
could use the flow id as the local representation of a path. Packets
sent to a particular destination but belonging to different flows may
use different paths, with the choice of path depending on the flow
id. This approach will result in the use of optimally sized packets
on a per-flow basis, providing finer granularity than PMTU values
maintained on a per-destination basis.
SB> How widely is flow-id supported in networks? I thought that the
SB> current position was that it was unreliable as an ECMP indicator
SB> and thus routers tended to glean information from the packet themselves.
This is future-proofing. Agreed, usage today is limited.
(But it would be better to call it the Flow Label for consistency with other
recent RFCs.)
Well the question is whether it is simply limited today, or broken today
in a manner that
is irrecoverable? I don't know, but I do know that the mainstream ECMP
approach
is the five-tuple. There is something akin to the flow label being
deployed in MPLS. However
what distinguishes the MPLS Entropy Label is that it is inserted (and
removed) by the
service provider and is therefore trusted by the service provider.
I think your other comments are all valuable.
Thank you.
Stewart