[ih] Early internetworking ca. 1976 / First ARPANET Link Put Into Service

[John Day]

Let me add a little detail.

> On Feb 3, 2026, at 15:23, John Shoch via Internet-history <internet-history at elists.isoc.org> wrote:
> 
> Towards the end of last year there was a thread here on the first Arpanet
> link at SRI, and Barbara Denny was kind enough to share a note (11/3/25)
> from Don Nielsen also reminding us of the earliest multi-network tests of
> the TCP Internet at SRI:
> // 3. First 2-net demo of TCP was on 27 Aug 1976.? PRNET and ARPANET.
> // 4. First 3-net demo of TCP was on 22 Nov 1977.? PRNET, SATNET, and
> ARPANET.
> 
> As I recall, prior to that time all of the TCP design, implementation and
> testing had been done just on the Arpanet.  Building on the early work at
> Stanford, the effort grew to include people at SRI, BBN, and elsewhere.
> Those demonstrations in 1976 and 1977 -- with a gateway and multiple
> networks -- were  tremendously important milestones along the path to what
> became the Internet as we know it today.
> 
> In response to that post, though, I was asked about the actual state of the
> Xerox PUP internetworking effort at that time, in mid-1976.  It took me a
> while to do an archeological dig -- I wanted to find the original documents
> to ensure that we got the history right.  I found a couple of interesting
> things and wrote up a summary.  My friends at the Computer History Museum
> have long encouraged us to document some of the work, and it was suggested
> "you should share it before we lose this history."
> 
> So, some of the notes, documents, and insights on internetworking in 1976
> (as seen from a slightly different perspective):
> 
> --In 1972, about 4 years earlier, the first INWG meeting had taken place at
> the ICCC meeting in Washington DC.

It was at this meeting that Pouzin first presented where the CYCLADES clean-slate approach to the network problem after seeing the ARPANET was presented to a wide audience. CYCLADES was designed to do research on networks, rather than a production network like the ARPANET. The assumptions that CYCLADES made were considerably different than those for the ARPANET. (BBN was doing what was required by the RFQ,) A few of the differences were:
In the ARPANET, the IMPs were a front-end to the hosts. Hence the hosts were not part of the network and had to be within 10m of the IMP. (Relaxing that with DH, and VDH came later) In CYCLADES, the hosts were assumed not to be near the CIGALE ‘routers’ and could be connected to more than one.
The ARPANET was basically a reliable virtual circuit network, while CYCLADES was an unreliable ‘best-effort’ datagram network.
CYCLADES introduced end-to-end reliability (between the hosts) at the Transport Layer.

At the end of March 1973, Cerf and Kahn spent a day with Pouzin learning more about CYCLADES. It was at this meeting that Gerard LeLann presented his results on simulation of CYCLADES.

In June 1973, the next INWG meeting was held at the NCC in New York. McKenzie says that most of the discussion was between Metcalfe and Zimmermann. Cerf said he would write up the minutes. (Cerf has produced a draft protocol after this meeting and there are comments by Walden and McKenzie at least those are the ones I have seen.

> --In the summer of 1973, about 3 years earlier, there were ongoing meetings
> at Stanford, and Cerf and Kahn were drafting the TCP paper at the Cabana
> hotel in Palo Alto.

From earlier emails with John, this meeting was 17 July 1973

On 13 Sept 73, Cerf issues the minutes for the June INWG meeting (INWG39), noting in a cover letter that they had held the meetings at Stanford that summer. INWG 39 is identical to the May 1974, Cerf/Kahn paper published by IEEE.

Oct 1973  INWG 42 – L. Pouzin Interconnection of Packet Switched Networks.
INWG 43 Zimmermann/Elie  Proposed Standard. In INWG42, Pouzin, in essence, proposes the solution to internetworking as changing the name of the Transport Layer to Internet Transport Layer and treating it as an overlay, thereby avoiding any protocol translation. The only thing the supporting networks have to do is meet the minimal service required by the Internet Transport Layer.

This leads to a debate between the two proposals and eventually a synthesis of the two protocols known as INWG96, which was voted on in 1976 and passed by a 2-1 margin.

Meanwhile . . . at PARC . . .

>  That same summer there were early discussions at PARC
> about the proposed Ethernet (as well as an alternative design simply called
> LOCAL network), and the need to interconnect them.  A memo from Aug. 1973
> reported a discussion on "...the problems of interfacing the LOCAL and
> ETHER networks.  This memo describes a rather general proposal for
> introduction of a message format standard which emerged from the
> discussion.  The adoption of the standard would enable us to interconnect
> different networks—essentially forming a network of networks…..”  A drawing
> includes ARPA, ETHER, and LOCAL networks.
> --In the Spring of 1974 the Ethernet was maturing (while the alternative
> LOCAL net never emerged), and there were some basic Ethernet-specific
> protocols implemented (EEFTP).  Yet it was clear that there would be a need
> to interconnect Ethernet networks, both locally and across geographies.
> Metcalfe had been participating in some of the INWG meetings, but it was
> also evident that we could not wait for that effort -- we needed something
> immediately.
> --That led to Bob's initial draft memo, "A Proposed PUP -- PARC Universal
> Packet" dated March 19, 1974, which began:
> "This memo is written and should be read with caution; its purpose is to
> promote a standard.  Because there isn’t an ice cube's chance in hell that
> our (or anyone else's) standard will be adopted without interminable debate
> and revision, the memo itself is quick and dirty. This way we get the ball
> rolling early. ...  A list of the packet networks at Parc would include, in
> arbitrary order of pedigree, (1) Ethernets, (2) Localnets, (3) Arpanets,
> (4) MCAnets, and (5) EIAnets.”
> [MCANets connected Data General Novas.  "EIAnets" evolved into a backbone
> packet switching network among Gateways, made up of leased serial lines.]
> .
> --Over the next two years the further design, implementation, and
> refinement of PUP were done primarily by David Boggs and Ed Taft.  Progress
> was reported in a series of memos, initially by Metcalfe and later by Boggs
> and Taft:
>  PUP Revisited
>  PUP Converging
>  Naming and Addressing Conventions for PUP
>  A Nova Gateway
>  Implementation of PUP in Tenex
>  PUP Again
>  PUP Connection State Diagram
>  PUP Servers on Maxc
>  etc., etc., etc.
> 
> --So where did things stand, after two years of work, in mid-1976?  The
> best document I have found describing things ca. 1976 is from 6 months
> earlier, a "draft" of "PUP Overview" by Taft dated Dec. 21, 1975.  He
> reports:
> "Local communication is carried on by means of several independent Ethernets
> (passive broadcast networks operating at 3 mb/s) and two MCAs
> (Multiprocessor
> Communications Adaptors for interconnecting Nova computers, operating at
> 1.6 mb/s).
> Long-haul communication is carried on over the Arpanet (a store-and-forward
> packet
> switched network·operating at 50 kb/s). We are considering making use of
> other
> transport mechanisms, such as optical fibers for very high bandwidth local
> communication, leased phone lines for regional communication at modest
> bandwidths,
> and commercial·packet switching services such as Telenet."
> The memo goes on at length to describe Basic Principles, Levels of
> Protocol, Standard Packet Format, Inter-Network Addressing, Fragmentation,
> etc.
> I only have a draft of this memo;  I have not yet found a copy of the
> complete final version.
> 
> https://drive.google.com/file/d/106a4W2mXsi4Ii-YzRgzsTAwe9_34IqJg/view?usp=sharing
> 
> --To further describe the operational state of the PUP Internet in the
> Spring of 1976, though, I have also found a copy of the then-current text
> file used to define assigned network numbers, host numbers, well-known
> sockets, and name-to-internet-address mapping.  This file was used to
> initially configure the gateways, and to load the name server.  The Tenex
> header shows this as the 50th copy of the file that was created --
> PUP-NETWORK.TXT;50, dated March 19, 1976.
> It shows that the PUP Internet at that time included 2 Ethernets, 2 MCAs,
> and the Arpanet.
> Well-known Sockets were defined for Telnet, Gateway-Info (routing updates),
> FTP, Misc. Services (name, time, etc.), and Echo.
> The name server database allowed a machine to have multiple names (e.g.,
> MAXC = Maxc1 = Parc-Maxc), and multiple internet addresses (if connected to
> more than one network).
> I count 27 Novas on the PUP Internet at the time.  This includes 3 machines
> acting as Gateways:
>  --The Portola Gateway, on both Ethernets, one MCA, and the Arpanet,
>  --The front end to the MAXC time sharing system, on one Ethernet, both
> MCAs, and the Arpanet.
>  --A Nova on one Ethernet and one MCA.
> There were 7 Novas only on one Ethernet (including a machine for font
> design and two for laser printer units).
> The remaining 17 Novas had both Ethernet and MCA interfaces (but were not
> necessarily running as Gateways).  These included Novas configured as
> servers controlling the older XGP printer,  the newer EARS laser printer,
> the Woodstock File System (WFS), and others.
> There were also 67 individual Altos on the Ethernets.  Most of these were
> for personal use, but the list also included one Alto as a dedicated Data
> Line Scanner system (i.e., a TIP, for in-bound and out-bound terminal
> traffic).
> [My hand-written notes suggest that there were at least 9 other Altos in
> our group which were not in the database (including mine).  Even if no one
> could find them via the name server, they could operate as client machines
> on the PUP Internet.]
> This system was in regular use, day and night, with over 100 machines on 5
> networks of 3 types.
> 
> https://drive.google.com/file/d/1SZ6yRLp6UqpbFKxbKK4D2stB6BcBX5qq/view?usp=sharing
> 
> 
> That was the state of the first and largest operational Internet in March
> 1976, 5 months before the important 2-net SRI demo of the TCP Internet.
> 
> John
> 
> PS:  A few additional notes and quick lessons from that period.
> 1.  For some software-intensive projects a smaller team can be very
> productive.  There may have been ~6 of us helping and kibitizing, but the
> vast majority of the programming and documentation was done by Taft and
> Boggs.
> 2.  As best we could, we did try to share some of our general progress with
> others.
> --I had been full-time at PARC since 1971, but had also been encouraged to
> "try to get a quick PhD" at Stanford.  I took part in Vint's networking
> seminar.
> --In the Spring of 1976 (before the 2-network demo at SRI), three of us
> from Parc were enrolled in a Systems Programming class being taught by
> Vint.  We had to do a group programming project, so the 3 of us undertook
> implementation of the 2nd generation simple file transfer:  evolving from
> the Ethernet-based EEFTP (written in Nova assembler) to a new
> internet-capable Pup-based EFTP (written in BCPL).  We turned in the
> listings, complete with some packet traces (from an Ethernet watcher) to
> show how it worked.  We all got an A+ -- thank you, Vint!
> --Later, Vint graciously arranged for us to attend some of the TCP working
> group meetings.  In July 1977 (before the 3-network demo) Yogen Dalal (who
> had been a graduate student under Vint) and I attended a TCP Working Group
> meeting held at MIT.  Jon Postel's TCP Meeting Notes (later registered as
> IEN 65) reflected some of our overview:  "PARC has 5 different networks
> with peices [sic] in Palo Alto, Los Angeles and various places on the East
> Coast.  Approximately 14 different networks, approximately 300 hosts
> connected."
> --In the following month, Aug. 1977, we attended an Internet meeting at ISI
> -- where we mentioned our gateway routing, gateway services,
> naming/addressing/routing, etc. [IEN 3] [IEN 19] [IEN 20]
> --Later, with the help of Don Nielsen and the crew at SRI, we became users
> of the PRNet -- adding it as a network in our internet, carrying
> encapsulated Pups wirelessly between two Pup gateways.  This was reported
> in IEN 78, https://www.rfc-editor.org/ien/ien78.pdf
> 3.  At this time the TCP Internet was designed to expose a reliable byte
> stream interface.
> The PUP Internet architecture provided interfaces at multiple levels:
> --Reliable byte stream (BSP), used for Telnet, FTP, etc.
> --Reliable packet stream (EFTP), used for simple file transfer, esp. to a
> print server.
> --Raw packet access, for very simple tasks (name server, time server, echo,
> etc.)
> Two years later, in 1978, the split of TCP into TCP and IP eventually
> allowed that kind of flexibility.
> 4.  Network-relative addresses were used in both the TCP Internet and the
> PUP Internet, with a network ID and a host ID.  This worked OK in the early
> days of both TCP and PUP, but we quickly realized that a) this did not work
> well if you wanted to move a machine from one network to another, and b)
> this would not scale adequately.  We tried to learn from both the
> Experimental Ethernet and the PUP Internet.  Thus, Yogen Dalal developed
> the 48-bit flat address space -- used in what became the 2nd generation DIX
> Ethernet standard, and used in the 2nd-generation Xerox Network Systems
> (XNS) protocols.  The 48-bit Ethernet address design has scaled and endured
> for ~50 years -- great work by Yogen.
>  https://dl.acm.org/doi/epdf/10.1145/800081.802680
> -- 
> Internet-history mailing list
> Internet-history at elists.isoc.org
> https://elists.isoc.org/mailman/listinfo/internet-history
> -
> Unsubscribe: https://app.smartsheet.com/b/form/9b6ef0621638436ab0a9b23cb0668b0b?The%20list%20to%20be%20unsubscribed%20from=Internet-history