On 10/10/12 4:30 AM, Magnus Westerlund wrote:
On 2012-10-09 14:56, Brian Haberman wrote:
On 10/9/12 8:46 AM, Adrian Farrel wrote:
Are we in the weeds yet?
I believe we are.
Likely, but the weeds here are fascinating and possibly relevant.
Given what you say below, I am not sure how relevant it is (regardless
of the fascination factor).
Agreed. As long as there is a robust L-2, the chance of in-transit
errors is low. If MPLS is predominantly carried over Ethernet, I
suspect the Ethernet CRC is providing sufficient/significant protection.
I think you are missing one point of the following reference:
[Sigcomm2000]
Jonathan Stone and Craig Partridge , , "When the CRC and
TCP Checksum Disagree", 2000.
That is that this paper found a significant amount of erroneous packets
where the UDP or TCP checksum discarded the packet, not the L2 CRCs. The
whole point of the paper is trying to analyze what type and why their is
errors that only the transport checksums detects.
If your comment about missing the point of the above paper is directed
to me, it is quite the opposite. Given that I was involved in those
6MAN WG discussions about this draft, I have a good understanding of the
questions raised by the paper.
When the topic of UDP checksums was discussed in the WG it was
questioned if this now 12 year old paper was still relevant. People in
the WG went looking at their network interface stats and found that
several different machines has UDP checksum discard rates that was
higher than one every 1000 packets. So yes there is still a lot of
packets being discard in the destination host due to transport
checksums. If "Why" has changed we don't know because someone needs to
repeat the study Stone and Partridge did.
I would not be surprised if there is differences between IP routing,
especially at the edge and MPLS label switching in regards to error
characteristics.
Completely agreed.
I think my main point around this whole issue is that you have to be
careful to understand the environment and what mitigating factors that
may exist so that you don't see these issues.
Yes! And that is why the draft is currently restricting the UDP
checksum value of 0 to tunneling protocols. I don't see any
justification, at *this* time, to loosen the rules in general.
I am fine with updating the text to better reflect reality. However, I
don't understand this MPLS and PW area sufficiently to determine what
error factors and saving graces that happens to be built into the system.
The whole point of the argumentation and checklist is. Checksums are
there for a reason. IF you have analyzed it the implications of turning
of the checksum and what it means to use a particular protocol header
field that is now unprotected then yes you can go ahead and use it.
But don't forget how other users of the network might suffer from your
traffic.
Right. This leads me to believe that the draft should remain focused on
allowing tunneling protocols to use checksum=0 when the encapsulated
payload is protected by a checksum.
Regards,
Brian
--------------------------------------------------------------------
IETF IPv6 working group mailing list
[email protected]
Administrative Requests: https://www.ietf.org/mailman/listinfo/ipv6
--------------------------------------------------------------------
