I’m always fascinated/impressed/surprised (maybe all of the above!) by how the Internet came to be, and how it was never planned to be this successful. I’m working on something related to IP address space, so let me share one of my favourite examples.

In 1977, Vint Cerf had to decide how many addresses the Internet would need. He was running the Defense Department’s Internet research program. The engineers had argued about address size for a year without agreeing, so Cerf picked one himself: 32-bit numbers, about 4.3 billion addresses. That seemed plenty for what was still an experiment.

When that format was published as RFC 791 in September 1981, the ARPANET had 213 host computers. “The problem is the experiment never ended,” Cerf said in 2010.

The addresses were handed out as if they would never run short, and some of the biggest blocks went to some interesting organisations — Ford among them. More on that later.

Cerf has told the story many times, and always takes the blame. At the University of Maryland in 2009, describing the moment he ended the debate, he said: “Let’s go do something. Here we are. My fault.”

Then… Vint Cerf actually replied

I ended the LinkedIn version of this by asking Vint Cerf to correct me if I’d got anything wrong. He replied with the fuller history, so here it is, in his words:

Bob Kahn and I settled on 24 bits in our 1974 paper: 256 networks and 65K “tcps” (hosts). RFC675 (dec 74) had 4 bits for network, 16 bits for host and 24 bits for port number. By 1977 we had settled on 8 bits for network and 24 bits for host and 16 bits for port number. By 1981 RFC [791] specified IP addresses as 7 bits for network, 24 bits for host and 16 bits for port number. Addresses were Class A, B or C - eventually classes D and E were defined. As the internet grew, [classless] IP addressing was introduced to allow more flexibility between bits associated with network and bits associated with hosts. By 1992, it was clear we needed more addresses and there ensued the IPng debates leading to IPv6 with its 128 bit addresses. There had been some discussion in the 1977-1978 period about larger than 32 bits. One alternative was 128 (!) and one was for variable length. As a practical matter, we wanted to start implementations on as many operating systems as possible and stuck with 32 bits.

Looking back, I wish we’d given more consideration to the larger sized choice!

vint

So 128 bits, the size IPv6 eventually adopted, was already on the table in 1977. What won was getting it running.

And when another commenter said reading it felt like hearing God discuss Genesis (nice pun), Cerf replied:

I always wanted to ask G-d how that worked out in His view...

When one person handed out the addresses

Deciding how many addresses there would be was one problem. Deciding who got them was another.

In the early days, it was one person that got to assign the addresses (true story). According to ARIN, Jon Postel did it as a volunteer, first at UCLA and later at the University of Southern California’s Information Sciences Institute. That work became IANA, and the registry job passed through two contractors before landing at Network Solutions in 1991.

The largest unit handed out was a block of 16,777,216 addresses, then called a “class A” and now written as a /8. There are only 256 of them in the whole space. Universities, government agencies and big companies received whole /8s, and IANA’s registry still records some of those decisions. Apple got 17.0.0.0/8 in July 1992, Daimler got 53.0.0.0/8 in October 1993, and Ford got 19.0.0.0/8 in May 1995. Twelve /8s are listed under US defense organizations, more than 200 million addresses between them.

The early approach did not last. Regional registries took over the work in their parts of the world. RIPE NCC, for Europe, the Middle East and parts of Central Asia, started in 1992, and APNIC, for Asia-Pacific, in 1993. ARIN opened for North America in December 1997. LACNIC, for Latin America and the Caribbean, was recognized in 2002, and AFRINIC, for Africa, in 2005.

The fixed sizes went too. In 1993, classless addressing replaced the old classes, so blocks could be cut to the size an organization actually needed.

Then the addresses ran out. IANA, which supplies the registries, handed out its last five /8s on February 3, 2011. ARIN ran out on September 24, 2015, and RIPE NCC made its final regular allocation on November 25, 2019.

But many of the addresses handed out in those early days never went anywhere. Which of those original blocks are still with the organizations that got them in the 1990s is the next story I’m going to write. A small spoiler, and perhaps with no surprise given how the ARPANET came to be: it involves the US military.