Subject: Re: v6 (was Re: -current sendmail cancer in IPv4-only kernel)
To: Sean Doran <firstname.lastname@example.org>
From: Dennis Ferguson <email@example.com>
Date: 05/08/2000 17:35:13
> Oh, no, actually I thought that decoupling the transport
> syntax (bits on wire) from the abstract syntax used by an
> addressing authority to define the structure of their
> respective addressing domain was a great feature of
> IS 8348/AS 2 and I was sad to see it die. (Although the
> argument over 6 bytes vs 8 bytes or more vs 2 bytes or
> less was very silly and typical of design-by-committee.)
I'm not sure what it is you just said, but I do persist in thinking it was
a bit odd that fields of such great topological significance as the
addressing authority identifier were included in the very highest order
bits of the address for the forwarding lookup in every router that
touched the packet to look at. "Topological CLNP addressing" had a lot
in common with "military intelligence".
> The variable length addresses were also a great feature,
> although the maximum length was too short.
If only they'd fully specified how forwarding lookups were supposed to
work, instead of leaving this as an exercise to the reader (what do you
do when the address in the packet is an exact prefix match but shorter
than a prefix in your forwarding table?). If only they'd explicitly
disowned multicast forwarding rather than leaving this as an (impossible
to get right) exercise to the reader. I suspect the address length would
have gotten fixed at 20 bytes had this gone on long enough since the attempt
to support CLNP in just the same way that IP worked could only have been
successful if you eliminated all the corner cases that didn't occur with
IP's fixed length addresses. I.e. you can't have your cake and eat it too.
> > actually very much in use). But to gain that advantage the addressing
> > plan, and precise (ie: unaligned) packet formats, and its checksum algorithm,
> > had to go along with it, and all that baggage was simply impractical.
> The alignment canard was absolutely destroyed by Tony Li's
> FPGA demo. That forwarding speed is quoted as a reason
> for CLNP's and variable-length addressing's failure is a
> tragedy. I think people lost track of things during the
> heat about the compromise of "profiling" CLNP.
I certainly agree that the alignment issue was, and is, overrated, but
I'm not so sure about the hideous header checksum. Note that interface speeds
have gone up by a factor of 192 since 1995 but ASIC, and even full-custom,
chip clock rates aren't anywhere near 192 times the clock rate of those
1995 FPGAs. Things that can't be computed by going increasingly wider
(i.e. more gates, which are available in quantities substantially exceeding
192 times 1995 FPGA densities) are really boring.
> I think IPAE was very nearly the most disgusting thing
> that was wrung through the IPng process.
There are some things we might agree upon...