Robert,
I will make just a few comments, where it adds value to what Brian
already said.
Brian E Carpenter wrote:
>
> Robert Elz wrote:
> ...
> > It seems that the current flow label is defined to be an immutable field
> > which should be set to a PRNG to identify flows (related packet streams).
> > Right?
>
> But it is not defined in normative text, and the text we have is vague
> on the actual usage - even worse than the text on the TOS byte in RFC 791,
> for example.
>
In fact, provisions were made, to have the field mutable. For instance
transport (IPv6) checksum, and IPsec authentication skip the field, to
allow its modification en-route.
So, it does not have to be immutable.
Non-normative text in RFC 2460 defines setting the value as a PRN, as an
aid to hashing based lookups of flow state in routers. Please note, that
in absence of a "perfect hashing function", using hashing by itself,
will not deliver the lookup result, some additional steps, like a
sequential lookup in a hash bucket, may be necessary.
However, Intserv/RSVP specifications, and Diffserv specification go into
more detail of how
classifiers are used. Consequently mechanisms were developed for very
efficient classification processing, which hardly resemble the model
(vaguely) suggested in RFC 2460. Hardware classifiers for processing at
high line speeds are often implemented based on Content Addressable
Memories (CAMs), in which PRNs does not help at all. Unfortunately, some
take the use of PRN, like an axiom, which is wrong. In fact, Jarno,
Christian, myself, and maybe others showed on this thread, that the flow
label does not need be a PRN. Brian and I discussed it also in the
Carpenter/Conta draft.
However, PRNs have the negative effect, that prevent the use of
classification or filtering that do not have/use a signaling mechanism
to transmit the value of the PRN to the routers.
This is one of the impediments for the use of the RFC 2460 definition
with Diffserv, or with filtering other than Intserv/RSVP.
> >
> > To anyone who doesn't understand the diffserv algorithms, that number will
> > look random - sure, it won't pass any statistical tests on randomness, but
> > no-one is suggesting that the flow label field is required to contain that
> > kind of value, are they?
>
Well, the non-normative RFC 2460 requires PRN. As it has been said,
based on current knowledge and experience, there is no good enough
reason for using PRN.
> [...]
> Brian
Alex
S/MIME Cryptographic Signature
