Talk:IPv3

Colossal fail

I have a personal memory of how we all lost the variable-length addresses of IPv3. I was not actually in the meeting where this happened (at MIT-LCS, 15-16 June 1978), but just down the hall in my office (I hadn't yet been promoted into the group); Dave Reed told me what happened a few days later, and the story has stuck with me (see below for why).

So, during that TCP meeting, Bill Plummer of BBN (who wrote the TENEX TCP) was observed talking to Vint Cerf; Dave overheard him complaining that he had a limited number of registers available at interrupt time (since he needed registers for PDP-10 byte pointers - needed to deal with variable-length fields). I don't know whether this tipped the scales on the fixed-length addresses used in IPv4, but it cannot have helped. Certainly, the use of variable-length addresses made the work involved in getting a basic experimental system up and running considerably harder.

I have always felt that the fact that I wasn't invited to attend prevented me from making what would have been my greatest contribution to internetting. I have long understood that interfaces are the things which have the greatest lifetime in any technological system. (The example I use is the screw light-bulb base: forget the technology on either side; the basic science behind the technology on either side has changed out of all recognition - fission reactors and LED lights. Those weren't even known when the base was invented; but we're still using that base.) The packet formats are a key part of interfaces in networking; look at NAT - the syntax of the 'address' fields is the same; the semantics has changed somewhat. So the adoption of the fixed-length addresses really screwed us; retention of the variable-length addresses would have been a massive boon forever.

I understand why the variable-length addresses were removed; however, I feel that had I been in the room, I could/would have suggested an intermediate path, one that gave the advantages of both. I would have suggested retaining the IPv3 packet formats (perhaps rearranged a bit, to improve alignments), but saying that for the moment, the only acceptable values for the address lengths would be '4'. In the short run, this would have made something that was as easy to implement as IPv4. In the long run, it would have allowed a slow, interoperable, upgrade to longer addresses.

Yes, sites that weren't upgraded to support them would not have been able to communicate with sites that had been given longer addresses - but that would have been reasonably workable. (Look at the transition from 32-bit to 96-bit ARPANET headers. Sites that hadn't upgraded to support 96-bit leaders could not communicate with sites that had been given such addresses - but it worked.) Jnc (talk) 19:55, 5 September 2026 (UTC)