Pekka,
Thanks for the reference. Most people who have analysed traces have
seen nothing but zero flow labels, but boring data like that don't
get published.
This study was reported in 2005. The hosts involved were non-representative;
the paper says that only 84 hosts (out of 1104) set the label *consistently* per
flow, and those were all "patched FreeBSD or Solaris". The ones that set it
inconsistently were FreeBSD or NetBSD. That may be a significant fraction
of hosts on a university network, but they are negligible in the network
as a whole. However, we should cite the reference.
Actually it doesn't matter so much. The worst effect of legacy hosts
setting the flow label under RFC2460 or RFC3697 rules is that the
label values are not uniformly distributed across the 2^20 possible
values. The way we're writing 3697bis and draft-ietf-6man-flow-ecmp
will tolerate this, even if it is not the recommended behaviour.
Brian
On 2011-02-23 10:36, Pekka Savola wrote:
> On Wed, 23 Feb 2011, Brian E Carpenter wrote:
>> To all intents and purposes, there is no backwards compatibility
>> problem because there are no implementations. But again, that
>> belongs on another thread.
>
> You must have some other definition of "implementation" than I do.
>
> 39% of hosts in a study set a non-zero flow label.
>
> http://www.maths.tcd.ie/~dwmalone/p/ec2nd05.pdf
>
--------------------------------------------------------------------
IETF IPv6 working group mailing list
ipv6@ietf.org
Administrative Requests: https://www.ietf.org/mailman/listinfo/ipv6
--------------------------------------------------------------------
