Introduction

I generated this with Astra in response to questions posed by John Day. This chronology follows the effort to separate a destination’s identity from its network attachments, from early naming proposals through implementations and the standards of the 1980s. It distinguishes specified mechanisms, contemporary implementation reports, and demonstrated behavior, and is updated as further sources are verified. Prepared with Astra.

Host multihoming concerns one communicating destination reachable through several network attachments. The questions are which entity a design names, who selects an attachment, and whether an existing conversation can continue after an attachment fails. A router forwarding a packet is a relay on its path; it acts as a destination when the communication terminates there.

Walden separated communication endpoints from sites

In 1970, David Walden was helping build the ARPANET at BBN. His RFC 61 (July), revised as RFC 62 (August), drew on that work to consider how communication could work in an ideal network. He proposed separating communication endpoints from the sites hosting them: port numbers named transferable endpoints, while separate site numbers located them. An endpoint could retain its identity while passing to another process on another host.1

Walden proposed placing the rendezvous machinery in the IMPs. He explicitly credited Ackerman and Plummer’s MIT PDP-1 work, An Implementation of a Multiprocessing Computer System (1967). His proposal was unimplemented and did not represent an official BBN position.

Walden addressed a moving endpoint; host multihoming concerns a continuing host with several attachments. The shared problem was how to identify a destination independently of where packets reached it. Direct descent of the later BBN designs from his proposal remains unestablished.

CYCLADES put the recipient beyond the attachment

In January 1973, Louis Pouzin’s addressing note for CYCLADES argued that a recipient reachable through several network exit points could not be identified by one exit position. He also questioned treating a physical computer as the communicating identity: coupled computers and virtual machines made that correspondence unreliable. His preferred object of identification was the communications software.2

The distinction had a routing mechanism behind it. The proposed transport-station names identified destinations; forwarding tables associated a destination with an output line or a weight used in choosing one. Regional aggregation was intended to reduce the table size, while local station entries supplied more specific destinations. The destination name did not have to become the number of the selected exit. The following illustrates the intended property, rather than a recovered network configuration or failover test:

Packet destination remains station E

             +-- node A -- attachment a --+
CIGALE ------+                           +--> E
             +-- node B -- attachment b --+

Routing could select another way to reach E without requiring the sender to substitute a different destination name. This did not make attachment knowledge or route maintenance disappear: those remained work for the communications subnet. CYCLADES separated naming the recipient from choosing how to reach it.2

That year, Cerf’s ITP draft explicitly allowed a host to attach to several packet switches in the same or different networks. The definition recognized multiple attachments without specifying how to select among them or preserve a connection after failure.3

These ideas circulated among the designers of internetworking. Cerf and Kahn’s 1974 paper cited Pouzin’s addressing note; INWG 48 recorded joint packet-format work by Pouzin, Cerf, and Sunshine. BBN’s own 1974 report explicitly debated CYCLADES over the division of work between hosts and the subnet, in the same treatment that discussed multiple host attachments.4 The record establishes contact and engagement. It does not establish that each later BBN or MIT mechanism descended from Pouzin’s particular proposal.

The April 1974 CYCLADES transport specification defined the machinery above packet delivery. A transport station, associated with a virtual installation, had its own identifying number. Packet context included source and destination station numbers; a flow joined ports within that context. Error control retransmitted unacknowledged letters, discarded duplicates, and terminated the session after repeated unsuccessful transmissions.5 The resulting recovery argument is conditional: a conversation could continue through another attachment if routing recovered while the station’s transport state remained valid and before its retry limit was exhausted. That is a deduction from the design, not a recorded failure experiment.

Pouzin’s 1974 account of CIGALE supplied a stronger implementation claim. It described the independent address space as implemented, allowed delivery to the same host address from different nodes, and permitted traffic over multiple host-to-node lines. Its routing discussion described propagated delay and buffer information per destination region or local host.6 CYCLADES therefore has contemporary evidence for implemented destination naming independent of attachment, beyond a proposal to connect a computer twice.

The remaining implementation questions are more specific. The material examined does not establish the exact procedures for registering and withdrawing a station’s attachments, or identify a dated test in which an existing flow survived the loss of one of two attachment nodes. Pouzin’s 1975 status report documents an operating network, with a first public demonstration in November 1973 and expansion to several switches in 1974. The first demonstration used one packet switch; it cannot establish recovery through a second switch.7

An April 1977 Waterloo progress report documents failure-recovery simulation. Its CIGALE-based simulator had been used to study route recovery after link or switch failure, citing Irland and Macdonald’s May 1976 report E-49. The investigators checked the simulator against program traces, calculated cases and other simulations, but said adequate real-network measurements were unavailable for validating their predictions. This supplies a dated simulation trail. It does not identify an existing transport conversation surviving a failed attachment or establish the exact attachment-management procedures running in CIGALE.8

Contemporary reports also document measurements on working CIGALE software. In April 1977, Gelenbe, Grangé and Mussard described two nodes running the regular CIGALE system, its TMM line protocol and an artificial traffic generator over a 19.2 kbit/s laboratory line. In June 1978, Eyries and Pujolle described experiments using MESCIG to generate packets and OBSERVER to record their timing. Their topology connected an IRIS 80 host to one switch, M3, with several paths among the switches. The experiments deliberately specified routes in packet text.9

The 1978 report supplies concrete operating details: destination-based output selection, queues limited to five packets, end-to-end recovery of overflow losses, and line protocols using acknowledgments and retransmission. Its TMM retry timeouts were 300 milliseconds on the 48 kbit/s line and 1,200 milliseconds on the other lines. These measurements establish functioning forwarding and line-control software. They do not establish the time to declare an attachment dead, withdraw its route, or preserve an existing conversation through another attachment.9

BBN and PARC developed ways to use several attachments

BBN’s January 1974 PWIN report recommended experiments with hosts attached to more than one IMP. Its October analysis treated mapping a logical host name to several physical IMP ports as straightforward in principle; coordinating sequencing and error-control state across IMPs was harder.10

BBN authors also explicitly recognized the CYCLADES advantage. In their January 1975 course notes, John McQuillan and David Walden explained that addressing hosts instead of switching nodes, together with leaving packet ordering out of the subnet, allowed a host to connect simultaneously to multiple nodes.11

At PARC, Ed Taft’s January 1975 Pup memo associated a host or process name with several addresses. Once chosen, connection addresses remained fixed. Yet a packet addressed to Maxc’s Ethernet attachment could arrive through its MCA attachment and be forwarded internally. Taft explicitly described preserving communication after network failure. This was a concrete design for separating a connection’s address from its physical delivery path; the cited memo does not date its first operational demonstration.12

At BBN, Falk and McQuillan’s September 1975 AUTODIN II study compared several ways to distribute traffic and recover after an access-circuit failure.10 The 1977 ARPANET proposal, submitted to Cerf, separated logical identities from physical connections and proposed rerouting channels after port failure. Its retained text explicitly said ARPANET did not yet support dual homing of hosts.13

The naming discussion continued across PARC and ISI. In January 1978, Shoch’s IEN 19 described multiple addresses and acknowledged BBN routing work; Cohen’s IEN 23 treated processes as communication endpoints and examined the bindings among names, addresses, and routes.14

BBN put host identity into routing

BBN’s 1979 report described TENEX hosts already running NCP, IP, and TCP on both ARPANET and RCCNET, with interface choices built into the monitor.15 The same report proposed a more general routing mechanism across networks.13

The proposed lookup order was full destination host identity first, network group second. A host entry could direct traffic through an attachment on B even when the address’s apparent location lay on A. When no host entry existed, forwarding could use the network-group route. The 1979 failure example explicitly retained the destination name while delivering through the other interface.

This made the interpretation of the address depend on routing state. The required work was to maintain reachability for the continuing host identity and select another attachment after failure. The source still called for experiments and trial implementations; operating two interfaces did not establish deployment of this routing scheme.

Cerf proposed recovery above TCP and IP

In August 1979, Cerf examined the recovery problem in TCP/IP. Citing a private note from Plummer and Tomlinson, he described how a host with two network attachments could lose a TCP connection when one interface failed: the connection used an address tied to that interface. Reconnecting through the other address would create a new TCP connection.16

            Client
       (connected to a)
              |
     +--------+--------+
     |                 |
 Network A         Network B
     |                 |
     X failed          | working
     |                 |
     a                 b
 +---+-----------------+---+
 |         Host H          |
 |      still running      |
 +-------------------------+

Cerf proposed placing recovery above TCP/IP, using name servers to supply multiple addresses. He considered keeping information about every host in every gateway prohibitively expensive. That objection applied to a particular distribution of state; stable host identity did not inherently require that arrangement. He presented his recommendations as a strawman for discussion.17

ISI and MIT developed alternatives in 1980

In March 1980, Sunshine and Postel proposed fixed virtual addresses whose location mappings could change, explicitly discussing mobility, multihoming, and TCP’s dependence on an address.14

Saltzer’s March 1980 campus proposal separated identifying the target from obtaining a route. It belonged to MIT’s local-network program and drew on source-routing and naming work from several groups.18 A routing service could choose a reachable attachment and return a source route:

target H → routing service → sequence of local-network addresses → selected attachment

A different route could be obtained without requiring the originator to enumerate H’s attachments. The work of selecting an attachment moved into the routing service; preserving a transport connection additionally depended on the identity and state used above that mechanism.19

Callon’s July 1980 BBN report explicitly discussed and cited the MIT proposal, bringing its routing-service alternative into BBN’s analysis.20

The disputed work was concrete

At the January 1981 Internet meeting, following MIT’s Noel Chiappa on growth and hierarchy, Cerf suggested a persistent host identifier alongside a hierarchical address used as a routing hint. Multihoming remained an explicit, unresolved goal.

Rosen’s June 1981 Internet design pursued logical host names independent of the destination network, translated by gateways; he also distinguished flat names from hierarchical internal routing. Compared with MIT’s routing service, this placed translation in gateways, with different requirements for distributing information.20

In July, Sunshine, Cohen and Postel accepted logical naming’s value but challenged Rosen’s architecture. They questioned requiring constituent networks to interpret common names, substantial host changes, and unresolved gateway selection. They also objected that his model could force hosts on the same local network to communicate through an intervening gateway.17 Their criticism came from participants developing alternatives of their own.

These arguments explain contemporary resistance to particular designs. They do not establish a single final decision rejecting general node identity.21

The identity being preserved could also extend beyond one machine. Rosen allowed a name to cover interchangeable machines, while Grapevine separated continuing service names from particular servers. Saltzer’s later naming-and-binding paper distinguished services, nodes, attachment points, and paths.22 These related problems help explain why the discussion repeatedly moved between host attachment, process identity, and service location.

BBN built hosts with multiple interfaces

Alongside the naming proposals, BBN extended its host software’s support for multiple attachments. In November 1981, BBN reported a multihomed UNIX VAX running on ARPANET and an internal BBN network. Its TOPS-20 Multinet software supplied a layer between IP and device drivers for multiple interfaces and protocol sharing.15

These were practical results. They establish operating multiple-interface support, while leaving open whether the broader host-group routing design was deployed and when an established TCP connection was first demonstrated to survive attachment failure.

1822L implemented the binding within ARPANET

Malis’s 1822L protocol gave a logical name H a list of physical host ports. The subnet selected among them. Its IP mapping let Internet senders retain an address encoding that logical name. The division of responsibility was:

Layer Identity or routing decision
TCP/IP endpoints Retain the IP address encoding network N and logical name H
Internet gateways Route toward destination network N
Destination ARPANET/PSN Map H to an available physical host port

The name could survive a port change while inter-network routing still depended on N. Extending this to attachments on unrelated IP networks would require another level of binding or host-specific routing beyond the destination subnet.

Specified in 1981 and revised in 1983, 1822L explicitly credited Rosen’s logical-addressing work. Logical-address assignments appeared in 1984; the 1986 PSN specification described logical addressing as an existing facility; Official Internet Protocols included the relevant specifications under recommended IP-over-ARPANET in 1987.23 These records establish implemented mapping, without establishing the first test of an existing TCP connection surviving attachment failure.

The general standards left the binding incomplete

The 1982 implementation description The DARPA Internet Gateway used routing tables indexed by destination network. It did not describe BBN’s proposed host-first lookup as the gateway’s procedure.24

Useful components nevertheless appeared in Internet specifications. RIP supported optional host routes with precedence over network routes. The 1989 host requirements permitted weak-host behavior: an address could be accepted through an interface other than the one to which it was associated. Combined with suitable routing and recovery, these capabilities could keep an IP address usable through alternate attachments; they did not automatically discover or maintain those alternatives.25

IP also supplied source-route options. Clark’s 1982 implementation guidance suggested obtaining a source route from a name server and carrying it through the TCP/IP interface. The complete MIT routing service did not become a general protocol requirement.26 These capabilities establish technical overlap; the cited records do not establish direct descent of RIP or IP’s source-route options from the BBN and MIT proposals.

By 1989, Requirements for Internet Hosts — Communication Layers explicitly assigned remote multihoming to applications. Its application companion recommended trying alternative addresses returned by name resolution.27 General recovery therefore continued to expose address selection above TCP/IP, while a continuing host-to-attachments binding remained something particular networks or additional routing arrangements had to supply.

CYCLADES source access

The following links provide the CYCLADES sources used here and the contemporary records of their reception. “Complete” describes the linked file or document, not a claim that every page supplies evidence for multihoming. Printed and PDF pagination differ as noted in the footnotes. The original etext scans are preserved; downloadable copies of the less accessible material accompany this article in its public repository.

Source Available text
Pouzin, Address Format in Mitranet, INWG 20 / MIT 507, January 1973 Complete five-page English note, extracted from CBI 123, box 1, folder 3, PDF 2–6
Pouzin, Presentation and Major Design Aspects of the CYCLADES Computer Network, 1973 Walden’s six-page scan, printed 80–85; bibliographies also give 80–87, whose corresponding original has not been acquired
Pouzin, Network Architectures and Components, SCH 516 / INWG 49, 1973 Complete 29-page scan
Pouzin, Interconnection of Packet Switching Networks, INWG 42, October 1973 Complete 19-page edition 1 scan
Zimmermann–Elie, Transport Protocol, SCH 519.1 / INWG 61, April 1974 Complete 31-page scan
Pouzin, A Proposal for Interconnecting Packet Switching Networks, SCH 527, March 1974 Complete 13-page scan
Pouzin, CIGALE, the Packet Switching Machine…, August 1974 Roger D. Moore’s HTML transcription; original proceedings pages not acquired
Pouzin, The CYCLADES Network — Present State and Development Trends, 1975 Roger D. Moore’s HTML transcription; original proceedings pages not acquired
BBN, Report 2913, October 1974 Complete 61-page scan; CYCLADES discussion and multihoming comparison
McQuillan–Walden, Introduction to Computer Networks, January 1975 Author-hosted course notes, CYCLADES §4.14, printed p. 81
Cerf’s 1973 ITP draft and associated comments Complete 69-page archival packet, CBI 123, box 1, folder 5; includes duplicate material
Morgan, Performance Measurement in Computer Networks, March 1975 Complete 251-page repository scan; CIGALE monitor feasibility distinguished from Waterloo laboratory operation
Waterloo, Network Modelling and Simulation, April 1977 Complete 125-page repository scan; §IV documents the CIGALE simulation work
Gelenbe–Grangé–Mussard, Performance Limits of the TMM Protocol, RR 230, April 1977 Complete 20-page HAL file, including wrapper and original cover; CC BY 4.0
Eyries–Pujolle, Validation and Prediction of Performance in the CIGALE Network, RR 314, June 1978 Complete 16-page HAL file, including wrapper and original cover; CC BY 4.0

The bibliography below gives the remaining original-report requests and links the contemporary bibliographies that identify them.28 Grangé’s April 3, 2012 oral history with Andrew L. Russell supplies retrospective participant context; only cached transcript passages were available in this research, and a complete local copy remains to be acquired. It does not date a conversation-preserving failover test.

Notes

  1. David C. Walden, A Note on Interprocess Communication in a Resource Sharing Computer Network (RFC 61, July 17, 1970); A System for Interprocess Communication in a Resource Sharing Computer Network (RFC 62, August 3, 1970), §§1–4, especially pp. 13–14, and conclusion.

    Walden’s example makes the distinction concrete. A receiver at K posts RECEIVE(M, N) to rendezvous site R; a matching SEND delivers data to K. Passing receiver port M to a process at L lets a new RECEIVE(M, N) at R direct subsequent data to L without changing the port numbers. Packets still carry site addresses. RFC 62 §1 credits Ackerman and Plummer’s An Implementation of a Multiprocessing Computer System (1967), Crocker and Crowther’s host-protocol sketch, and discussions with Crowther and Kahn. Its conclusion states that the proposed system had not been implemented. ↩︎

  2. Louis Pouzin, Address Format in Mitranet (INWG 20 / NIC 14497 / MIT 507, January 19, 1973; complete five-page English note), pp. 1, 2, 3, 4, 5. The preserved English translation is addressed to INWG and appears in CBI 123, box 1, folder 3, PDF pp. 2–6. Pages 2–3 distinguish a recipient from exit points and communications software from physical computers. Page 3 also describes internal services with instances at several nodes, among which the network can choose an accessible instance.

    Page 5 gives destination-indexed routing tables: region, external station, node, and internal-station cases lead to line or weight entries. The note does not supply a complete attachment-update or failure-detection algorithm. Pouzin’s Network Architectures and Components (SCH 516 / INWG 49 / NIC 21653, cover dated August 1973), printed p. 23 / PDF 24, also allows a host to be a distributed system, attached over several lines to different switching nodes. Pouzin’s 1973 Presentation and Major Design Aspects of the CYCLADES Computer Network, §VII, printed p. 84, also discusses station names reachable through several nodes, multiple stations on one line, and regional routing exceptions. The available scan contains printed pp. 80–85; bibliographic descriptions giving pp. 80–87 refer to a longer pagination than this copy.

    Pouzin’s distinction did not require eliminating hierarchy: p. 4 proposes network, region, and transport-station fields. Its possible 3/5/8-bit allocation differs from the conclusion’s reduced 4-bit region and 4-bit external-station components with fixed/reserved bits. These are successive proposals within the note, not a single invariant deployed header. His A Proposal for Interconnecting Packet Switching Networks (SCH 527, March 1974), §6, uses packet-switching-network names while rejecting gateway names that would bind addressing to topology. The earlier Interconnection of Packet Switching Networks (SCH 513.1 / INWG 42 / NIC 20792, edition 1, October 1973), §8, pp. 7–10, already permits gateway choice to vary while the destination remains invariant. Both interconnection proposals discuss naming areas that can overlap physical networks when naming conventions permit it. These are architectural provisions, not a demonstration of a common identity operating across arbitrary networks. Network-number fields alone therefore do not establish rejection of the recipient/attachment distinction. ↩︎ ↩︎

  3. Vinton Cerf, A Partial Specification of an International Transmission Protocol (1973), HOST definition, PDF p. 26, PDF p. 27. Charles Babbage Institute collection CBI 123, box 1, folder 5; duplicate copies of these pages appear at PDF pp. 53–54 of the preserved packet. The multiple-attachment statement is in the typewritten text, not a handwritten addition. The packet also contains Walden/McKenzie comments. This establishes early recognition, without establishing a complete selection mechanism, retained connection identity after failure, or an operating implementation. ↩︎

  4. Vinton Cerf and Robert Kahn, A Protocol for Packet Network Intercommunication (May 1974), TCP addressing discussion and reference 11, especially PDF pp. 12–13. The references include INWG 20, the CYCLADES design paper, its ST–ST specification, and Walden’s interprocess-communication paper. The acknowledgments record further exchanges among the designers.

    Cerf’s A Packet Format Proposal (INWG 48 / NIC 21596, January 24, 1974), cover and attached Pouzin Experimental Communication Protocol (SCH 515, October 1973, revised January 1974), records joint work following the Hawaii meeting and earlier discussions at Brighton and NPL. BBN Report 2913, p. 38, bibliography p. 59, explicitly engages CYCLADES over host-versus-subnet responsibility. Its multihoming analysis on p. 57 observes that host-level ordering removes the particular sequencing obstacle. These sources establish circulation and debate, without assigning authorship of BBN’s mapping scheme to Pouzin. ↩︎

  5. H. Zimmermann and M. Elie, Transport Protocol: Standard Host-Host Protocol for Heterogeneous Computer Networks (SCH 519.1 / INWG 61; complete 31-page scan), cover dated April 1974, definitions on p. 4, packet service on p. 5, error control on p. 8. These are complete pages from the preserved CBI scan.

    TS in the English specification corresponds to ST, station de transport, in the earlier French terminology. Subscribers and ports are separate transport entities. Section 9 applies error control to letters within a flow: T03 governs retransmission, T04 acknowledgment delay, and failure after N unsuccessful transmissions terminates the session. Neither stable station identity nor this recovery procedure preserves state lost when the destination station itself restarts. ↩︎

  6. Louis Pouzin, CIGALE, the Packet Switching Machine of the CYCLADES Computer Network (IFIP, Stockholm, August 1974, pp. 155–159), §§3.6–3.9, 3.14–3.15. The source examined is an HTML transcription, not an original page scan. Section 3.9 explicitly calls the independent address space implemented. Its host identity can also represent a logical host.

    The §3.6 header has 96 bits, including 8-bit source and destination network fields and 16-bit source and destination host fields; it is distinct from INWG 20’s earlier formats. Some congestion-control refinements and the packet-age field remain prospective. The paper should not be read as certifying deployment of every facility it discusses. ↩︎

  7. Louis Pouzin, The CYCLADES Network — Present State and Development Trends (Symposium on Computer Networks, 1975, pp. 8–13), Development, Internetwork Linkages, and Technical Changes; HTML transcription. Its chronology places design and staffing in 1972, the first public demonstration in November 1973, three switches in February 1974, and seven by June. The 1974 and 1975 accounts differ on the November host count but agree that there was one packet switch.

    This status report says the earlier liaison service was not implemented and describes transport redesign. Its bibliography dates SCH 519.1 to June 1974; the preserved cover says April. These version differences prevent attributing every proposed transport feature to one deployed transport implementation. A configuration record or test trace is still needed to connect a particular station identity, two attachment nodes, software versions, and observed recovery behavior. ↩︎

  8. University of Waterloo Computer Communications Networks Group, Network Modelling and Simulation, April 1977 progress report: printed IV.11/IV.14 (PDF 45/48), route-recovery simulation; IV.15–IV.17 (PDF 49–51), model basis and validation; IV.27–IV.28 (PDF 61–62), references to Grangé’s CIGALE Implementation, Tools and Techniques, MIT 602, September 1975, and Irland–Macdonald’s Simulation of Route Propagation Mechanism in CIGALE, CCNG E-49, May 1976. The complete progress-report scan was acquired and these pages visually checked. MIT 602 and E-49 themselves remain unexamined.

    This limitation concerns the Waterloo investigators’ model validation. D. E. Morgan’s Performance Measurement in Computer Networks (March 31, 1975), printed pp. 10–11 and 16 (PDF 13–14 and 19), describes a working Waterloo laboratory monitor while treating its use on CIGALE as a feasibility investigation. The later IRIA reports cited below provide direct accounts of controlled CIGALE measurements. ↩︎

  9. Erol Gelenbe, Jean Louis Grangé and Patrice Mussard, Performance Limits of the TMM Protocol: Modelling and Measurement, IRIA-LABORIA RR 230, April 1977 (HAL record), printed pp. 1–3 / PDF 4–6. The original cover supplies the author order, report number and month; the repository filename contains 240. TMM uses stop-and-wait acknowledgments, timeout retransmission and packet numbering to discard duplicates. The approximately two-metre laboratory line connects two nodes, each running regular CIGALE software, TMMS and ATG; typewriters control generation and collect results. The study assumes negligible error rates. This is line-protocol state, distinct from the station and flow state in SCH 519.1.

    François Eyries and Guy Pujolle, Validation and Prediction of Performance in the CIGALE Network, IRIA-LABORIA RR 314, June 1978 (HAL record), printed pp. 1–3, 7–9 / PDF 4–6, 10–12. The original cover resolves the repository filename’s 317. MESCIG runs on an IRIS 80 under SIRIS 8; OBSERVER records packet length, trip delay, intersend and receive intervals. Figure 1 shows one host attachment to M3 and a triangle of links among M3, M1 and Grenoble MG. The host–M3 line uses TMM at 19.2 kbit/s; M3–M1 uses MV8 at 19.2 kbit/s, M3–MG uses TMM at 4.8 kbit/s, and M1–MG uses TMM at 48 kbit/s. MITRA 15 nodes select output lines from destinations and routing tables, with five-packet output queues. MV8 permits eight packets before an acknowledgment; the reported TMM timeout values govern retransmission, not route withdrawal. Figures 9–12 compare measurements with the model; §5’s changed line capacities, buffer sizes and protocol mix are predictions. Neither report gives individual run dates or records conversation-preserving attachment failover. Both complete downloads retain their HAL wrappers, which state CC BY 4.0 and an October 1, 2024 deposit. ↩︎ ↩︎

  10. BBN, Final Report: Prototype WWMCCS Intercomputer Network (PWIN) (Report 2714, January 15, 1974), p. 15, recommends multiple-IMP attachment experiments as part of adapting ARPANET technology to Defense Communications System needs. This is an experimental recommendation.

    G. Falk and J. McQuillan, AUTODIN II Study Contract: Final Report (BBN 3119, September 15, 1975), TM-24, logical addressing, pp. 158–159, 160–161, 162–163, 164; TM-31, Multiple Homing of Hosts in AUTODIN II, pp. 174–175, 176–177, 178–179, 180–181, 182–183, 184. The study compares one active access circuit with standby switchover, selection per segment, and selection per transaction. It prefers the third for AUTODIN II, examining failure detection, uncertain acknowledgement state, retransmission, duplicates, and resynchronization. It distinguishes host attachment diversity from route diversity inside the switching network; it is not evidence of a deployed TCP/IP host. ↩︎ ↩︎

  11. John M. McQuillan and David C. Walden, Applied Math 254: Computer Networks — 1. Introduction to Computer Networks (January 1975), §4.14, printed p. 81. Their explanation joins host addressing with the absence of subnet packet ordering. Some of the surrounding routing description is in the future tense; the notes establish the authors’ understanding, not a complete implementation specification or a failure-test result. ↩︎

  12. Ed Taft, Naming and Addressing Conventions for Pup (Xerox PARC memorandum, January 9, 1975), pp. 1–4, collection PDF pp. 57–60, especially p. 4. A directory can return several addresses for one host or process; local software chooses an address when establishing communication. Taft rejects putting a destination name or address list into every Pup as cumbersome, including the ambiguity when a name denotes interchangeable process instances. His XGP-to-Maxc example keeps the selected address while using the MCA path and internal forwarding. The memo describes intended behavior, not a dated failover-test report. ↩︎

  13. BBN, Interface Message Processors for the ARPA Computer Network (Report 2913, 1974), p. 57; ARPANET Routing Algorithm Improvements (P77-CSY-93A/B, 1977), p. 10 (retained page headed P77-CSY-93A); Proposal for Continuation of Research on Command and Control Related Computer Technology (P79-ISD-76, 1979), Appendix A, pp. 69, 70–71, 72–73, 74–75, 76. Linked scans are from my collection; page numbers refer to the printed originals, and photographs retain their two-page spreads where applicable.

    The August 17, 1977 cover letter to Cerf identifies the original July 1 proposal, an August 4 revision, and replacement pages for the resubmission; John McQuillan is the technical contact. Page 10 of the retained 93A text distinguishes a logical host or process from a physical connection, including several logical entities sharing one interface. Task 2 compares changes in the network with alternatives placing more work in hosts. The revision labels should not be treated as interchangeable.

    A separate November 1977 Proposal for Continuation of Research on Command and Control Related Computer Technology (P78-ISD-15), cover and pp. 60–61, scheduled a spring 1978 gateway demonstration combining dynamic recovery around a failed gateway and/or multiple attachments to a destination network. This is a planned experiment in gateway and network redundancy, not a demonstrated host failover.

    The 1979 proposal’s full-host lookup precedes its network-group fallback. Its unchanged host name can numerically resemble the failed interface’s address; the routing interpretation, rather than the address’s shape alone, supplies the stable identity. The proposal argues against making higher protocols cycle through interface addresses and recompute checksums. ↩︎ ↩︎

  14. John F. Shoch, Inter-Network Naming, Addressing, and Routing (IEN 19, January 1978), pp. 7–8, acknowledgments, and bibliography. It discusses logical groupings and multiple addresses, and credits McQuillan’s 1974 BBN routing report and Sunshine’s 1977 interconnection work. Danny Cohen, On Names, Addresses and Routings (IEN 23, January 1978), pp. 1–2, considers communication endpoints and bindings that need not be one-to-one.

    Carl Sunshine and Jon Postel, Addressing Mobile Hosts in the ARPA Internet Environment (IEN 135, March 1980), pp. 1–2 and 5–6, proposes virtual addresses whose mappings change with physical location and directly discusses multihoming and TCP state. Its request for further work does not establish deployment. This is relevant context for their July 1981 response to Rosen. ↩︎ ↩︎

  15. BBN P79-ISD-76, Appendix A, p. 69, reports TENEX’s existing multiple attachments and monitor-built interface choices. Combined Quarterly Technical Report 22 (BBN 4761, August 1981), pp. 67, 68–69, 70–71, 72–73, 74–75, describes address libraries, separation of local address entries from interface control blocks, source-address selection, and gateway routing. The extensions were coded but still being debugged; the production TCP/IP mentioned on p. 66 does not establish completion of every extension.

    Combined Quarterly Technical Report 23 (BBN 4825, November 1981), pp. 73, 74, reports the running UNIX Cost Center VAX. Both attachments then used 1822 and ACC LH/DH-11 interfaces; other access protocols and hardware needed corresponding modules or drivers. Distribution of a second beta had been delayed by a funding gap. That particular software delay is not evidence of a decision against the broader logical-addressing proposals.

    BBN P81-ISD-113 (August 1981), Appendix A, p. 14-3, describes the developed TOPS-20 layer while proposing integration and wider availability. Multinet Device Driver Enhancement Proposal (P82-ISD-79, July 1982), pp. 1-1, 1-2, 2-1, 2-2, describes existing internal software and proposed documentation, TOPS-20 release 5 integration, operational controls, and flow-control improvements. Multiple attachments served reliability and congestion needs; existing capability and proposed improvements must be distinguished. ↩︎ ↩︎

  16. Jon Postel, Transmission Control Protocol (RFC 793, September 1981), §§1.5, 2.7.

    For a connection identified by (c, client_port, a, server_port), reconnecting to address b changes the TCP peer identity. Routing the original address a through another attachment can preserve it, provided the host accepts and replies from a there, both endpoints retain their TCP state, and recovery finishes before timeout. An address can thus serve as the continuing host identity without a new identifier field; learning another address before opening a connection supplies a weaker property.

    Multiple interfaces were also explicit in the IP specifications: Jon Postel, DoD Standard Internet Protocol (RFC 760, January 1980), §2.3, and Internet Protocol (RFC 791, September 1981), §§2.3, 3.2. Their recognition of several physical interfaces and logical addresses did not itself supply the general attachment-selection and recovery service. ↩︎

  17. Vint Cerf, Internet Addressing and Naming in a Tactical Environment (IEN 110, August 1979), pp. 1–4; Jon Postel, Internet Meeting Notes — 28-29-30 January 1981 (IEN 175), pp. 20–21; Carl Sunshine, Danny Cohen and Jon Postel, Comments on Rosen’s Memos (IEN 191, July 1981), pp. 2–7.

    IEN 191 p. 6 also observes that the local part of an Internet address can already be a logical name. The existence of a network component does not, by itself, force the local component to identify one physical interface. The unresolved work includes the scope of that name and the mechanism that selects an attachment. ↩︎ ↩︎

  18. D. D. Clark, J. H. Saltzer, and L. Svobodova, Annual Report: July, 1976 — June, 1977 (MIT Laboratory for Computer Science, Computer Systems Research Division, RFC 145, June 30, 1977), §III, pp. 16–19, describes Farber/Irvine-derived interfaces and a TCP-derived Data Stream Protocol. This is MIT’s local RFC series. Saltzer, Clark, and Pogran’s Why a Ring? describes the prototype ring and bypass mechanisms. Ring bypass is distinct from giving one host several independent network attachments.

    Saltzer, Reed, and Clark, Source Routing for Campus-Wide Internet Transport (September 1980 version, following the August IFIP workshop), §III.4(5), acknowledgments, and references. It credits Farber and Vittal’s 1973 DCS work for basic source routing, Hopper and Wheeler for routing-service ideas, and Cohen, Postel, Shoch, and Moon’s Chaosnet papers for ideas and terminology. This later version is distinct from Saltzer’s March IEN 144, which Callon cited in July. The complete routing service remained a subject for further design. ↩︎

  19. Jerome H. Saltzer, Source Routing for Campus-Wide Internet Transport (Local Network Note 22 / IEN 144, March 11, 1980), pp. 2–6, 19–20.

    In the proposed internet-layer packet, the source route is a sequence of local transport addresses. Each gateway uses the next address for local forwarding, avoiding a topology table for selecting the remaining path. Saltzer identifies the route itself as naming an attachment; the routing service associates that route with the requested target.

    Pages 19–20 compare three arrangements: separate attachment names selected by the originator, one name with attachment selection by gateways, and a routing server selecting an attachment and route. The third moves work into the server and makes the routing service easier to change. It does not by itself specify TCP connection identity or demonstrate an established TCP connection surviving a route change. ↩︎

  20. Ross Callon, Features of Internetwork Protocol (BBN 4466, July 1980), pp. 56–57, 66–67 (scans from my collection); Eric C. Rosen, Issues in Internetting Part 3: Addressing (IEN 188, June 1981), pp. 2–3. ↩︎ ↩︎

  21. Jon Postel, Comments on Action Items from the January Meeting (IEN 177, March 24, 1981), p. 3, responds to RSRE’s request for an addressing pause while discussing changes to the network/host division of the address. Requirements for Internet Gateways (RFC 1009, June 1987), bibliography, still cites Rosen’s logical-addressing and Internet-design notes and the ISI response. Continued citation does not establish adoption. Neither the 1979 strawman nor the 1981 critique supplies a final cancellation decision; the Plummer/Tomlinson private note, follow-up discussions, first 1822L rollouts, and dated TCP failover tests remain leads to pursue. ↩︎

  22. Rosen, Issues in Internetting Part 3: Addressing, §3.1, pp. 2–3, allows a name to denote either multiple attachments or interchangeable machines. Andrew Birrell, Roy Levin, Roger Needham, and Michael Schroeder, Grapevine: An Exercise in Distributed Computing (1982), §§3, 5, 8.3, 9, describes registry-qualified names, service groups, connect addresses, and movable server identity; initial service began in 1980.

    Michael D. Schroeder, Andrew D. Birrell, and Roger M. Needham, Experience with Grapevine: The Growth of a Distributed System (1984), pp. 4–6, describes replicated registration data, interchangeable submission servers, and primary/secondary inbox sites. These are service and delivery-path redundancy, not proof that each server had multiple interfaces. The 1982 paper’s §4.4 distinguishes alternate delivery paths from replicated contents: a stored message could remain inaccessible during a server outage while new messages used another inbox.

    Jerome H. Saltzer, On the Naming and Binding of Network Destinations (1982; republished as RFC 1498 in 1993), pp. 3–4, distinguishes services, nodes, attachment points, and paths. His memorandum index dates the local LNN 28 draft to March 3, 1981. These distinctions clarify the scope and lifetime of an identity; they do not require four explicit identifier fields or establish that every earlier proposal followed this framework. ↩︎

  23. Eric C. Rosen, Logical Addressing (IEN 183, May 1981, drawing on BBN Report 4473); Andrew G. Malis, The ARPANET 1822L Host Access Protocol (RFC 802, November 1981; RFC 878, December 1983), especially RFC 878 §2.2.2 and Appendix A; PSN End-to-End Functional Specification (RFC 979, March 1986), §§3.1.1, 3.2; Joyce Reynolds and Jon Postel, Official Internet Protocols (RFC 1011, May 1987), p. 39.

    Joyce Reynolds and Jon Postel, Assigned Numbers (RFC 900, June 1984), p. 16, includes NIC-administered ARPANET logical-address assignments and an assignment for BBN gateways. Registration supplies an administrative trace, not proof of a particular multihomed host’s deployment. RFC 1011’s recommended category applies to IP on ARPANET; it does not give RFC 878 a separate modern standards-track maturity.

    Andrew G. Malis, Logical Addressing Implementation Specification (BBN 5256, May 1983), pp. 1, 2, 3, 4, 5, 29. The introduction calls the implementation upcoming. The design offers first-reachable, closest-address, and load-distribution choices. If an attachment fails, the IMP-level connection is destroyed and outstanding messages receive destination-dead indications; another message to the same name triggers a new translation and connection. These are subnet mechanics, not a statement that TCP must reset. Page 29 also permits one physical machine to appear as distinct logical hosts through different port groupings. ↩︎

  24. Hinden and Sheltzer, The DARPA Internet Gateway (RFC 823, September 1982), §3.3, pp. 7–8. ↩︎

  25. Charles Hedrick, Routing Information Protocol (RFC 1058, June 1988), §3.2; Requirements for Internet Hosts — Communication Layers (RFC 1122, October 1989), §§3.3.4.2, 4.2.3.7. ↩︎

  26. Jon Postel, Internet Protocol (RFC 791, September 1981), §3.1; David D. Clark, Name, Addresses, Ports, and Routes (RFC 814, July 1982), §5, pp. 8–9. ↩︎

  27. Requirements for Internet Hosts — Communication Layers (RFC 1122, October 1989), §3.1; Requirements for Internet Hosts — Application and Support (RFC 1123, October 1989), §2.3. ↩︎

  28. The remaining implementation originals include J.-L. Grangé and L. Pouzin, Cigale, la machine de commutation de paquets du réseau Cyclades, AFCET, Rennes, November 1973, pp. 249–263; Pouzin, Les choix de Cigale, pp. 265–274; Grangé, CIGALE Implementation, Tools and Techniques, SHARE European Association, Dublin, September 1975, MIT 602; and Irland–Macdonald’s May 1976 E-49. The 1982 volume edited by Pouzin, The Cyclades Computer Network: Towards Layered Network Architectures (ISBN 0444864822; OCLC 8689452), has a HathiTrust scan with restricted access. These are acquisition leads, not full texts examined for this account.

    Contemporary bibliographic evidence is available in Wood, Watkins and Cotton’s Annotated Bibliography of the Literature on Resource Sharing Computer Networks (NBS SP 384, revised September 1976), printed p. 28 / PDF 34, which identifies Irland’s Simulation of CIGALE 1974, Fourth Data Communications Symposium, Quebec, October 7–9, 1975, pp. 5-13–5-19, IEEE 75-CH1001-7-DATA. Its extended Waterloo report E-32 (January 1975) and Irland’s 1977 dissertation remain unacquired. Trinity’s Computer Laboratory Annual Report 1973/4, §5.3 / PDF 19, supplies the planned SEAS 75 venue. It is not the proceedings.

    Jean-Pierre Thomesse’s Connexion et coopération de deux ordinateurs CII 10070–IBM 1800 (June 27, 1974), bibliography b.5–b.6 / PDF 197–198, identifies the two-volume Rennes proceedings of November 7–9, 1973. Daniel Meyer’s Transformation d’une base de données centralisée en une base de données répartie sur un réseau d’ordinateurs (November 16, 1977), printed pp. 312 and 316 / PDF 166 and 168, confirms the Grangé–Pouzin pagination and identifies García et al.’s portable ST2 implementation specification, SCH 536.1, and ECA-Automation’s May 1975 SOCRATE/CYCLADES operations manual. Meyer’s “juin 73” date for SCH 536.1 needs an original-cover check. The cited originals were not recovered.

    The recovered measurement reports supply further exact requests: RR 230, references [1] and [3], printed p. 17 / PDF 20, identifies Grangé’s March 1976 CIGALE description MIT 617 and N. Naffah’s Spécification d’interface TMM avec le réseau CIGALE, TRA 521, December 1975. RR 314, references [2], [6], [7] and [11], printed p. 13 / PDF 16, identifies Eyries–Gien’s Eurocom 77 measurement paper, Grangé–Mussard’s Liège 1978 line-protocol measurements, Grangé–Maximovitch’s propagation-delay report MES 514 (1976), and Mussard’s Paris–Grenoble line study MES 522 (1977). Their original reports, program listings and observation records remain to be obtained. ↩︎