Summary

  • RFC 2143, an Experimental memo from May 1997, proposed sending IPv4 datagrams in SCSI SEND MESSAGE(10) commands among close workstations. Its framing choice was precise, but it did not turn an ordinary SCSI host adapter into a peer network interface.
  • A conventional initiator receives responses to its own commands. To receive another host’s command, a host adapter needed target mode and driver support; to discover an IP neighbor, the design had to probe SCSI targets one by one because the bus had no broadcast data-link address.
  • The memo’s 64-host routed example and cluster applications were possibilities, not demonstrated installations, measured performance, security guarantees or evidence of widespread adoption.

The command could carry the packet

A SCSI cable was familiar as the short link between a computer and its peripherals. RFC 2143 asked whether the same bus could connect several TCP/IP workstations. The proposal had a concrete packet format. Of three SCSI SEND MESSAGE command lengths, Ben Elliston selected the ten-byte form. Byte zero contained opcode 0x2A; bytes seven and eight carried the IPv4 datagram length in big-endian order. The low four bits of “stream select” indicated payload type, borrowing an analogy to Ethernet’s type field. An IP datagram followed the SCSI header.

That solved a framing question, not a network question. The 16-bit transfer length fitted the IPv4 length field in the proposal’s reasoning. It did not prove that every adapter could transmit and receive as a peer, that a particular cluster achieved the memo’s suggested approach to 360 Mb/s, or that packet delivery was secure. The RFC Editor record marks the memo Experimental and explicitly not an Internet Standard.

One role speaks first; the other replies

Ordinary SCSI transactions gave different powers to an initiator and a target. The initiator issued a command; a target replied. An initiator normally received a message addressed to it only as the response to something it had already sent. Workstations trying to exchange IP packets asynchronously could not rely on that pattern. A host wishing to accept a peer’s unsolicited command needed an adapter that supported target mode, plus a driver able to hand the received message to higher layers. The RFC named this as the critical mismatch.

The distinction matters because a physical connection does not grant protocol authority. Eight connectors on a bus do not make eight equal IP interfaces. The adapter’s mode, the operating system’s driver and the application of the SCSI communication-device command all had to align. A fast data path is useful only after a receiving host is allowed and equipped to consume its messages.

Discovery had to imitate a shout

IP forwarding still needed a link address for the next hop. Ethernet’s ARP could ask all stations at once using a broadcast hardware address, as RFC 826 describes. SCSI offered no such broadcast data-link address. RFC 2143 suggested simulating the shout with round-robin transmissions to targets. Depending on the SCSI variant, an unresolved binding might take seven or more separate bus accesses. Positive ARP caching would avoid repeating known lookups; negative caching was floated as a possible way to delay repeated failed queries.

Those caches were mitigations for repeated discovery traffic, not evidence that SCSI had acquired broadcast semantics. The cost of finding a peer depended on which IDs existed, adapter behavior and cache state. The original bus used IDs 0–7 and Wide SCSI extended the range to 0–15. In the memo’s scale example, the bus reached less than ten metres. None of those numbers described an Ethernet-like local network of arbitrary size.

Elliston then sketched eight hosts with additional adapters and IP routing to form a possible 64-host arrangement. It was an architectural thought experiment: routing between short buses, not one enlarged SCSI broadcast domain. The proposed uses—distributed computation, fault-tolerant NFS, Web and database service—likewise showed why a close, high-bandwidth cluster might be attractive. The memo did not report that these applications ran on the design.

The lesson is not that the framing was trivial or doomed. It is that encapsulation answered a narrower question than Ethernet had already hidden from many applications. A SCSI command could hold an IPv4 packet. Peer reception, address discovery and topology still needed their own implementation and evidence.

Sources and limits

  • RFC 2143, especially §§2–7, supplies the proposal, command fields, role constraint, discovery workaround, topology examples and limited security discussion.
  • RFC 826 supplies the Ethernet broadcast-ARP contrast; it does not establish SCSI behavior.
  • RFC Editor status confirms Experimental classification and date.

No tested adapter, driver, packet trace, cluster benchmark or adoption survey is in this source set. The illustration is a conceptual reconstruction, not a photograph of a deployed RFC 2143 network.