Summary

  • RFC 2067 removed three packet-format choices that RFC 1374 had permitted because no implementation in use was reported to depend on them.
  • The resulting grammar was stricter than the underlying ANSI encapsulation standards: an ANSI-valid packet could still be outside the RFC and could be accepted or ignored.
  • Interoperability was claimed only for one HIPPI-SC switch or a simple two-way point-to-point link; blocking multi-switch fabrics remained implementation-dependent.

A standards revision is often described by what it adds. RFC 2067 is more revealing when read as a deletion record.

Its predecessor, RFC 1374, had carried both IP encapsulation and address resolution onto HIPPI, a high-speed simplex channel used in pairs. By January 1997, the RFC 2067 authors could report at least ten implementations of the encapsulation and switch discipline. That experience did not validate every branch of the earlier document. It showed which branches running systems actually used.

The new Draft Standard kept the working center and cut away three legal alternatives. It also separated ARP, which lacked enough implementation experience, from the standards-track IP specification. The result was not a claim that every question had been solved. It was a narrower common surface backed by a more precise account of evidence.

Three permissions became one packet grammar

RFC 1374 permitted several ways to lay out a HIPPI frame. RFC 2067 prohibited an optional short burst at the beginning, prohibited fill bytes in the D1 area and required the D2 offset to be zero. Any short burst had to be the final burst. The D1 area size was fixed at three 64-bit words, with no D1 fill.

Those are small fields with a large combinatorial effect. A receiver facing multiple legal shapes needs code paths for each one, test cases for their interactions and a decision about malformed or unfamiliar combinations. A sender may conform to one interpretation while a receiver has only exercised another. Optionality becomes a shared cost even when almost nobody benefits from it.

RFC 2067 made the format previously recommended by RFC 1374 mandatory. Its warrant was empirical and carefully phrased: no RFC 1374 implementation in use was known to employ the removed options. The document therefore believed existing implementations would already comply and would not suffer interoperability problems from the change.

That claim has boundaries. The RFC does not name the ten implementations, publish their code lineages or show an exhaustive matrix proving every pair. “Not in use” is not “physically impossible,” and it does not rule out private or abandoned code. What the text establishes is the engineering decision: deployed experience was used to decide which freedom did not deserve to remain in the common contract.

ANSI validity and Internet interoperability were different receipts

The HIPPI-FP and HIPPI-LE standards defined a broader encapsulation space. RFC 2067 openly said its rules were more restrictive. A packet could satisfy the ANSI standards and still be illegal under the Internet memo. A destination encountering such a frame could accept it or ignore it.

This distinction matters because “standards compliant” is incomplete without naming the standard and the claim. ANSI conformance could show that a hardware-level frame was well formed. It did not show that an IP-over-HIPPI peer following RFC 2067 had to interpret every permitted ANSI variation. The narrower RFC profile selected the exact subset on which independent Internet implementations could rely.

RFC 2026, published three months earlier, explains the institutional context. Advancement to Draft Standard required at least two independent interoperable implementations and sufficient successful operational experience. Its rule reached individual options and features: absent such evidence, a feature normally had to be removed. RFC 2067 looks like that discipline applied to a real packet grammar—not consensus declaring code into existence, but code constraining what the document could responsibly promise.

ARP did not pass on the encapsulation's evidence

The split with ARP is equally instructive. RFC 1374 had specified both IP and ARP on HIPPI. RFC 2067 reported enough experience for IP encapsulation and switch discipline but not for the ARP portion. ARP was moved to a separate Informational memo, with the possibility of returning to the standards track if implementations and interest later warranted it.

That separation prevented evidence from one subsystem from laundering maturity into another. A working IP packet format did not prove that address resolution, network configuration or broadcast emulation had comparable support. RFC 2834 arrived in 2000 to clarify and expand ARP and IP broadcast over HIPPI-800. Its existence reinforces the original boundary: unresolved work had its own document and its own path.

One switch was a proof boundary, not a miniature Internet

RFC 2067 believed conforming hosts would interoperate when connected through a single HIPPI-SC switch. It made the same claim for a simple two-way point-to-point HIPPI connection with no switch. More complicated networks might also work, but their behavior depended on switch internals and interconnection.

The reason appears in the topology. One HIPPI-SC switch was treated as non-blocking: a path was available from any source to any destination. Connect switches with a shared link and contention can become visible. One source may be unable to reach a destination because a different source-destination pair occupies that shared path. Connection strategy on such a blocking fabric is no longer the host problem already tested by the simple profile.

The memo did not convert a successful lab configuration into a claim about every network built from the same parts. It named the topology whose behavior its host rules could support and left other arrangements to their implementers. That is not weakness. It is the difference between an interoperability result and an extrapolation.

A fixed header still proved very little beyond itself

A captured RFC 2067 packet can show the required D1 size, zero D2 offset, absence of D1 fill and the position of a final short burst. It cannot by itself prove that the destination accepted the datagram, that IP processing succeeded or that an application received useful data. Nor does one working connection establish that a multi-switch fabric is non-blocking, that routing is correct or that a node is authorized.

The security boundary is just as narrow. RFC 2067 said it introduced no known security problems. That is not authentication, confidentiality, integrity or permission. A deterministic packet shape reduces ambiguity; it does not turn a high-speed link into a trusted one.

Read through Lu Heng's later Minimum Initial Specification framework, RFC 2067 can be seen as a compact example of trimming a common layer to demonstrated interoperability and keeping non-common topology choices outside it. This is a later editorial comparison, not a claim about John Renwick's intent. The historical record stands without it: implementation evidence earned subtraction, and the standard's credibility depended on saying where its evidence ended.

Sources