Back to Basics: The History of IP Interface Addresses

In the previous blog post in this series, we figured out that you might not need link-layer addresses on point-to-point links. We also started exploring whether you need network-layer addresses on individual interfaces but didn’t get very far. We’ll fix that today and discover the secrets behind IP address-per-interface design.

In the early days of computer networking, there were three common addressing paradigms:

  • Companies building large-scale networks connecting mainframes or minicomputers (IBM and DEC) realized you need a unique network-wide identifier per node, not per interface. After all, you want to connect to the services running on a node, and nobody (apart from the network administrator) cares about individual node interfaces.
Assign network addresses to nodes, not links

Assign network addresses to nodes, not links

  • Companies building LAN-based client-server networks (AppleTalk, Novell Netware…)1 didn’t care about the distinction between node- and interface addresses. The devices connected to their networks had at most one interface anyway.2
Keep things simple: assign network addresses to interfaces

Keep things simple: assign network addresses to interfaces

  • Anything mentally based on the phone network paradigm assigned addresses to switch ports, and the nodes attached to the network had to use the address of the connected switch port, or they wouldn’t receive any traffic. You could say that the switches used static host routing, with a host route pointing to every edge port.

Where does IP fit into this picture, and how did we get to interface addresses where it would be much better for IP routers and multihomed hosts to use node addresses?

You have to remember that IP didn’t start as a network protocol. IP stood for Internetworking Protocol and was designed by a bunch of academics trying to make sense of an underlying network built by BBN (see BBN Report 1822 for the details). IP addresses were statically assigned to the ports of the Interface Message Processor, with part of the original IP address being the IMP identifier and the second part being the port number on the IMP3.

An IMP in a nutshell

An IMP in a nutshell

Deciding to use interface-scoped network addresses when designing a protocol to fit that environment was an obvious choice. After all, if a host connected to two IMP ports, it had to use two IP addresses, and who would want to burn two IMP ports anyway.

Eventually, the IP networks got extended to Ethernet segments. The gateways between the original BBN network and the LAN network had to use two IP addresses - one for the BBN part and another for the LAN part. When subnets were introduced to split LAN networks into smaller routable bits, the architecture was already cast in stone.

Let’s use per-link prefixes to make our lives simpler

Let’s use per-link prefixes to make our lives simpler

IPv6 was another missed opportunity. The difference between node-level addresses and interface-level network addresses was well understood at that time. DEC’s experience with building large-scale WAN and LAN networks influenced the OSI CLNP protocol design; CLNP was consequently using node-level addresses. There were even ideas to use CLNP as the basis for next-generation IP. Still, following a grassroots revolt4, IAB decided to start from scratch. The “we always did things this way” mentality quickly kicked in together with a copious amount of Second System Effect.

Does that mean that it would be better to use a CLNP-like design with node-level network addresses? Not necessarily:

However, it still doesn’t make sense to be forced to use an IP subnet for every point-to-point router-to-router5 link. Welcome to the mysterious world of unnumbered IP interfaces – the topic of the next blog post in this series.

More to Explore

Why don’t you watch the Network Addressing part of How Networks Really Work webinar to get even more details?

  1. Lovingly called desktop protocols by Cisco TAC. ↩︎

  2. Netware servers were an exception, but Novell engineers realized that too late and had to settle for the same loopback interface hack we use in IP networks today. ↩︎

  3. SAN geeks might realize that Fibre Channel still uses the same approach. ↩︎

  4. You might find a few hints about that event in the hilarious Elements of Networking Style book. ↩︎

  5. People believing in vague marketese are free to replace that with a switch-to-switch link. ↩︎

Latest blog posts in Unnumbered IPv4 Interfaces series


  1. In your "was CLNP really broken" post you state that host-based adressing doesn't scale. Could you explain why that is the case?

  2. With host-based addressing, the routers have to advertise individual host addresses in routing protocols (at least until the first summarization boundary). With subnet-based addressing, they only advertise subnets.

    Not that we couldn't solve that today (after all, we're often using host routing in IP), but 30 years ago it was a big deal.

Add comment