Summary

  • RFC 948 described two ways to carry IP on IEEE 802.3 physical networks in June 1985. One kept Ethernet-style protocol types in a field that 802.3 ordinarily treated as length; the other used the IEEE 802.2 logical-link header. The same cable could carry both while their hosts remained unable to exchange IP frames directly.
  • Ira Winston proposed a coordinated cutover: make both formats a static choice in implementations, keep all hosts on Ethernet-compatible framing during preparation, then switch the LAN together. A dual-speaking gateway could connect prematurely split groups, but the memo warned of a significant performance penalty.

One field, two meanings

The trap sat in a familiar-looking header. Ethernet Version 2.0 and IEEE 802.3 used practically compatible physical signaling, RFC 948 said, while their data-link layers disagreed. Ethernet treated a two-byte field as a protocol type. IEEE 802.3 treated the same position as a length. A frame's outer shape could match the wire and still send the receiving software down the wrong path. Physical compatibility had purchased electrical admission, not a shared language for IP.

Winston offered an Ethernet-compatible route for installations already carrying IP. At 10 megabits per second, 802.3's permitted length value topped out at 1500. Values above that range could be used privately; the protocol-type values in the memo for IP, ARP and trailer encapsulation were above 1500. An 802.3-compatible physical interface could therefore carry Ethernet-style frames and keep talking to the existing Ethernet IP population. The route used an available numeric opening. It did not turn the two data-link conventions into the same convention.

His second route complied with the IEEE 802.2/802.3 data-link format. It placed IP in an LLC Type 1 Unnumbered Information frame with both DSAP and SSAP set to 96, the IP service access point cited in RFC 948. The 802.3 frame could add zero padding to meet a minimum size; those pad bytes were outside the IP datagram. Hosts also needed address mapping, with ARP-like dynamic discovery strongly recommended. These details made the second route buildable. They did not make a host using it understand an Ethernet-style frame sent by its neighbour.

The split the cable could not show

The conclusion of RFC 948 is unusually blunt. Both formats can be mixed on the same local network, Winston wrote, but doing so partitions it into two incompatible subnetworks. A machine that speaks both could act as a gateway between them. That adds a significant performance cost, according to the memo. It also turns a routine local exchange into a routed crossing just because two hosts disagree about the frame wrapper around the IP packet.

The memo preferred the IEEE 802.2/802.3 route because the IEEE standards were accepted nationally and internationally and the same encapsulation could be used on other IEEE 802 physical networks. Preference alone did not move the installed base. Winston's proposed migration was deliberately staged. New hosts would offer a static encapsulation choice. Existing ones would be changed to offer it too. All hosts would continue using the Ethernet-compatible method while that capability was installed. Only after the whole population could make the choice would the LAN switch to the IEEE method.

A migration announced before the last host was ready would create two networks on the same cable.

Three years later, RFC 1042 specified IP and ARP across IEEE 802 networks with an LLC/SNAP header rather than RFC 948's direct SAP-96 approach. Its DSAP and SSAP values were 170, followed by a zero organization code and an EtherType. The later specification also described the mixed-cable boundary in operational terms: link-layer broadcasts on one format would reach only computers using that format. When mixing was necessary, it recommended treating the populations as distinct IP networks with a gateway.

RFC 1042 is evidence of a later specification and the same interoperation problem, not proof that Winston's proposed cutover happened as written.

The story of RFC 948 is a warning against counting sockets instead of readers. The cable could be shared, the electrical rules could agree, and the IP packet inside could be perfectly valid. A receiver still needed the same outer grammar. Winston's static-choice plan made the operator responsible for the moment of agreement. Without that moment, the gateway had to collect the bill.

Sources