Inter-Networking

In their 1968 paper, “The Computer as a Communications Device,” written while the ARPANET was still in development, J.C.R. Licklider and Robert Taylor claimed that the linking of computers would not stop with individual networks. Such networks, they predicted, would merge into a “labile network of networks” that would bind a variety of “information processing and storage facilities” into an interconnected whole. Within less than a decade, such formerly theoretical speculations had already acquired an immediate practical interest. Because by the mid-1970s, computer networks were proliferating.

Networks Proliferate

They were proliferating across a variety of new media, institutions, and places. ALOHAnet was one of several new academic networks funded by ARPA in the early 1970s – the others being the PRNET, which connected mobile trucks with packet radio, and the satellite-based SATNET. Along similar lines, other countries, especially the U.K. and France, were developing their own research networks. Local networks, because of their smaller scale and lower cost, were multiplying even more quickly. Other than Xerox PARC’s Ethernet,  one could also find the Octopus at Lawrence Radiation Laboratory in Berkeley, California; the Ring at the University of Cambridge; and the Mark II network at the British National Physical Laboratory.

Around the same time, businesses also began to offer fee-based access to privately-funded packet networks. This enabled a new, national marketplace for on-line computer services. In the 1960s, various companies had launched businesses offering access to specialized databases (for legal or financial data) or to time-shared computers, to anyone with a terminal. But these were prohibitively expensive to access cross-country via the regular telephone network, which made it hard for such services to expand beyond local markets. A few larger services firms (Tymshare, for example) built their own internal networks, but commercial packet networks brought the costs down to a reasonable level for users of smaller services.

The first such network came about via a defection of ARPANET experts. In 1972, several employees left Bolt, Beranek, and Newman (BBN), the company in charge of ARPANET’s construction and operation, to form Packet Communications, Inc. Though that company ultimately failed, the sudden shock catalyzed BBN to form its own private network, called Telenet. With Larry Roberts, the architect of ARPANET, at its helm, Telenet operated successfully for five years before being acquired by GTE.

Given this explosion of network diversity, how could Licklider and Taylor’s vision of single unified system ever come about? Even were it organizationally feasible to simply merge all of these systems into ARPANET – which of course it was not – the incompatibilities among their protocols would have made it a technical impossibility. Yet ultimately these many heterogeneous networks (and their descendants) did interlink, into a confederated communication system that we know as the Internet. It began not with any grand, global plan, but with an obscure research project run by a middle-ranking ARPA manager named Robert Kahn.

Bob Kahn’s Problem

Kahn completed a Ph.D. thesis on electronic signal processing at Princeton in 1964, in between rounds of golf on the links adjacent to the Graduate College. After a brief stint as a professor at MIT, he took a position nearby at BBN, initially intending a short leave of absence to immerse himself in industry and learn how practical men decided which research problems were worthy of investigation. His pursuits at BBN, fortuitously, included research into the possible behavior of computer networks, for it was just a short time later that BBN received the bid request for ARPANET. Kahn became absorbed in the project, providing the bulk of the design for the system’s network architecture.

His short leave of absence became a six-year stint, with Kahn serving as the networking expert at BBN for the duration of the ramp up of ARPANET into its fully operational state. By 1972, however, he was tired of the topic, and, more importantly, tired of being buffeted by the constant politicking and jostling for advantage among the BBN division heads. So he accepted an offer from Larry Roberts (before Roberts himself had left for Telenet) to become a program manager at ARPA, heading a research program to develop automated manufacturing technology, with potentially hundreds of millions in funding at his command. He washed his hands of ARPANET and set off south for a clean start in a green field.

Then, within months of his arrival in Washington D.C., Congress quashed the automated manufacturing project. Kahn wanted to pack up and return to Cambridge immediately, but Roberts convinced him to stay on to help develop new networking projects for ARPA. And so Kahn, unable to escape the bonds of his own expertise, found himself managing PRNET, a packet radio network intended to bring the benefits of packet-switched networks to the operational military in the field.

The PRNET project, launched under the auspices of Stanford Research Institute (SRI), was intended to extend the basic technical kernel of packet broadcasting from ALOHANET to support repeaters and multiple stations, including mobile vans. However, it was obvious to Kahn early on that the network by itself would be sorely lacking in utility, for it was a computer network with scarcely any computers. When it became operational in 1975, it consisted of one computer at SRI and four repeater stations positioned around the San Francisco Bay. Mobile field stations could not economically support the size and power requirements of a 1970s mainframe. All of the significant computing resources available resided in the ARPANET, which used a totally different set of protocols and had no way of interpreting a message broadcast on PRNET. How, he began to wonder, could his infant network be interlinked with its far more mature cousin?

Kahn turned to an old acquaintance from the early ARPANET days for help in crafting the answer. VintonCerf had gotten interested in computers as a math undergraduate at Stanford, and decided to go back to grad school in computer science at UCLA after a couple years at IBM’s Los Angeles office. He arrived in 1967, and, with his old high school friend Steve Crocker, joined Len Kleinrock’s Network Measurement Center, the UCLA branch of ARPANET. There he and Crocker became experts in protocol design, as leading voices in the Network Working Group, which developed both the base Network Control Program (NCP) for sending messages on ARPANET and the higher level file transfer and remote login protocols.

Cerf met Kahn in early 1970, when the latter flew out the UCLA from BBN to put the network through its paces with some load testing. He generated congestion in the network with the help of software built by Cerf for generating artificial traffic. As Kahn had expected, the network collapsed under the stress, and he recommended changes to improve congestion control. In the ensuing years, Cerf continued on with what looked like a promising academic career. Around the same time that Kahn decamped from BBN for Washington D.C., Cerf traveled up the opposite coast, to take up an assistant professorship at Stanford.

Kahn knew a lot about computer networks, but had no experience with the details of protocol design – he was a signals processing guy, not a computer scientist. He knew Cerf would be perfect to supply those skills, which would be crucial to any attempt to link ARPANET and PRNET. Kahn reached out to him about inter-networking, and they met several times throughout 1973 before holing up at the Cabana Hyatt in Palo Alto to produce their seminal paper, “A Protocol for Packet Network Intercommunication,” published in the May 1974 issue of IEEE Transactions on Communications. It presented the design for a Transmission Control Program (TCP) – the P later became protocol – the cornerstone for the software of the modern Internet.

Outside Influences

Not two people or one moment are more closely identified with the invention of the Internet than Cerf and Kahn and this 1974 paper. Yet, the creation of the Internet was not truly an event that happened at a point in time, but a process that unfolded over years of development. The initial protocol described in Cerf and Kahn’s 1974 paper was tweaked and revised numerous times over the ensuing years. Not until 1977 was the first cross-network link tested; the protocol was not split into two layers – the now-ubiquitous TCP and IP – until 1978; and ARPANET did not adopt it for its own use until 1982.¹ The participants in that process of invention extended well beyond the two most well-known principals. In the early years, an organization called the International Packet Network Working Group (INWG) served as the main venue for their collaboration.

ARPANET debuted to the wider technical world in October 1972, at the first International Conference on Computer Communications, amid the swooping curves of the modernist Washington Hilton. In addition to Americans like Cerf and Kahn, several prominent European network experts attended, among them Louis Pouzin of France and Donald Davies from the U.K. At the instigation of Larry Roberts, they decided to form an international working group to discuss packet-switching systems and protocols, modeled on the Network Working Group that established the protocols for ARPANET. Cerf, a newly minted Stanford professor, agreed to serve as chair.  One of the first topics that this new International NWG took up was the problem of inter-networking.

Among the important early contributors to this discussion was Robert Metcalfe, whom we previously met as the architect of Xerox PARC’s Ethernet. Though Metcalfe could not say so to any of his colleagues, by the time of the publication of Cerf and Kahn’s paper,  he and his colleagues were already well underway with the design of their own internet protocol, the PARC Universal Packet, or PUP.

The need for an internet at Xerox became pressing as soon as the Alto/Ethernet network became a success. PARC had another other local network, of Data General Nova minicomputers, and there was ARPANET, of course. Looking further in the future, the PARC leadership foresaw that every Xerox site would need its own Ethernet, and these would need to be connected in some fashion (probably via Xerox’s own internal ARPANET equivalent). To enable it to masquerade as an ordinary message, the PUP packet nestled within the outer packet of whatever host network it was travelling across – the PARC Ethernet, say. When the packet reached the gateway computer between Ethernet and another net (e.g., ARPANET), that computer would unwrap the PUP packet, read its address, and re-wrap it in an ARPANET packet with the appropriate headers to send it onward to its destination.

Though Metcalfe could not directly disclose what Xerox was up to, the practical experience he had acquired there inevitably trickled back into INWG discussions, in filtered form. Evidence of his influence survives in the fact that Cerf and Kahn’s 1974 paper recognizes his contribution, and Metcalfe would later show a glimmer of resentment that he did not rate the recognition of co-authorship². PUP likely affected the design of the modern Internet again later in the 1970s, when Jon Postel instigated the decision to split TCP and IP in order to avoid having to run the intricate TCP protocol on the gateways between networks. IP (Internet Protocol) was a simplified addressing protocol with none of TCP’s complex logic for ensuring the delivery of every bit. The Xerox networking protocol, – by then publicly known and rechristened as Xerox Network Systems (XNS), had already made the same division.

Another source of influence on the early internet protocols came from Europe, especially from a network developed in the early 1970s, as an offshoot of Plan Calcul, a program set in motion by Charles de Gaulle to nurture a native French computing industry. De Gualle had long been concerned about America’s growing political, commercial, financial and cultural dominance of Western Europe. He aimed to re-establish France as an independent world power, rather than a pawn in the great Cold War game between the U.S. and the Soviet Union. Two events in the 1960s particularly threatened that independence with respect to the computer industry. First, the United States refused export licenses on its most powerful computers, which France intended to use to aid in the design of its own hydrogen bomb. Second, an American company, General Electric, became the majority owner of France’s only major manufacturer of computing machinery – Compagnie des Machines Bull³ – and then shortly thereafter discontinued several major Bull product lines. Hence the Plan Calcul, to ensure that France could provide for its own computing needs.

To oversee Plan Calcul, De Gaulle created the délégation à l’informatique (roughly translated, the “delegation on computing”), reporting directly to his Prime Minister. In early 1971, that delegation selected an engineer by the name of Louis Pouzin to oversee the creation of a french ARPANET. The delegation believed that packet networks would play a crucial role in computing in the coming years, and so native technical expertise in that field would be essential to Plan Calcul’s success.

Pouzin, a graduate of the École Polytechnique, the premier engineering school for all of France, had worked as a young engineer for France’s national telephone equipment manufacturer, and then moved to Bull. There he convinced his employers that they needed to know more about the cutting edge work happening in the United States. So he spent two-and-a-half years while a Bull employee, from 1963 to 1965, helping to build the Compatible Time-Sharing System (CTSS) at MIT. This experience made him the foremost expert on time-shared, interactive computing in all of France – likely on the entire European continent.

Pouzin called the network he was tasked to build Cyclades, after a constellation of Greek islands in the Aegean Sea. Like its namesake, each computer on the network was, to a large extent, an island entire of itself. For Cyclades’ primary contribution to networking technology was the concept of a datagram, the simplest possible variety of packet communication. The idea consisted of two complementary parts:

1. Datagrams are independent: Unlike the data in a telephone call or an ARPANET message, each datagram can be processed independently. There is no reliance on any prior messages, whether based on ordering or some protocol for establishing a connection (e.g. dialing a phone number).
2. Datagrams are host-to-host: All responsibility for ensuring a message is sent reliably to a destination rests with the sender and receiver, not with the network, which is merely a “dumb” pipe.

The datagram concept was anathema to Pouzin’s peers at the French post, telegraph and telephone authority (PTT), which was building its own network in the 1970s based on telephone-like circuit connections and terminal-to-computer  (rather than computer-to-computer) communication, under the supervision of another Polytechnique grad, Rémi Després. Culturally, the idea of giving up on reliability within the network was repellent to the PTT mindset, molded by decades of experience in trying to make the telephone and telegraph systems as robust as possible. While economically and politically, the idea of surrendering control of all applications and services to host computers at the periphery of the network threatened to make the PTT into nothing but a fungible commodity. Nothing works better to deeply entrench one’s opinions than firm resistance to them, however, and so the needling presence of the PTT’s virtual circuits only helped to confirm Pouzin in the correctness of his datagram, host-to-host approach to protocols.

Pouzin and his fellow Cyclades engineers participated actively in INWG and the various conferences where the ideas behind TCP were hashed out, and they were not shy about putting forth their opinions on how a network of networks should function. Like Metcalfe, both Pouzin and his colleague Hubert Zimmerman earned mentions in the 1974 TCP paper, and at least one other coleague, an engineer by the name of Gerard Le Lann also helped Cerf with hashing out the protocols. Cerf later recalled that “the sliding window flow control for TCP came straight out of discussions with Louis Pouzin and his people… I remember Bob Metcalfe and Le Lann and I sort of lying down of the living room in my house in Palo Alto on this giant piece of paper, trying to sketch what the state diagrams were for these protocols.”⁴

The datagram concept mapped neatly onto the behavior of broadcast networks like Ethernet and ALOHANET, which sent their messages willy-nilly into a noisy, uncaring ether (in contrast to the more telephonic ARPANET, which required in-order delivery between IMPs across a reliable AT&T line to function). It made sense to align the protocols for inter-networking with the lowest-common-denominator datagram-like networks rather than their more elaborate cousins, and indeed that is just what Kahn and Cerf’s TCP did.

I could continue still further in this vein, by describing the British role in the early inter-networking conversations, but I don’t wish to belabor the point – that the two names most closely tied to the invention of the Internet are not the only ones that mattered.

TCP Conquers All

What happened, then, to this early promise of inter-continental collaboration? How is it that Cerf and Kahn are hailed everywhere as the fathers of the Internet, yet we hear very little about Pouzin and Zimmerman? To understand this requires, for starters, getting down into the procedural weeds of the INWG’s early years.

In keeping with the spirit established by the ARPA network working group and its Requests for Comment (RFCs), the INWG created its own system of “General Notes.” In keeping with this practice, after about a year of collaborative work, Kahn and Cerf presented the preliminary version of TCP to the IWNG as note 39 in September 1973. This was effectively the same document that they published in IEEE Transactions the following spring. In April 1974 the Cyclades team, under the authorship of  Hubert Zimmerman and Michel Elie, published a counterproposal, designated INWG 61. The differences consisted in different views on certain engineering trade-offs, mainly around how packets are subdivided and re-assembled when crossing networks with small maximum packet sizes.

This rift was minor, but the need to settle on a consensus had acquired a sudden urgency due to the plans announced by the Comité Consultatif International Téléphonique et Télégraphique (CCITT) to consider packet networking standards. CCITT, the standardization body of the International Telecommunications Union, operated on a four year cycle of Plenary Assemblies. Proposals for consideration in the 1976 assembly were due in the fall of 1975, and no further changes would be possible between then and the next assembly in 1980. A scramble of meetings within INWG led up to a final vote in favor of a new protocol drafted by the representatives of the most important institutions in the world of computer networking – Cerf from ARPANET, Zimmerman from Cyclades, Roger Scantlebury from the British National Physical Laboratory, and Alex McKenzie of BBN. The new proposal, INWG 96, split the difference between 39 and 61, and seemed likely to establish the direction for network interconnection for the foreseeable future.

But in truth, the compromise proved the last gasp of international collaboration in inter-networking, a fact foreshadowed by the ominous abstention of Bob Kahn from the INWG vote on whether to accept it. As it happened, the vote came too late to make the CCITT deadline, and Cerf further undermined its standing at CCITT with a cover letter indicating that it lacked the full consensus support of the INWG. Any proposal from INWG was likely dead-on-arrival anyway, because the telecom authorities that dominated CCITT had no interest in the datagram networks being cooked up by computer researchers. They wanted to control the flow of traffic within the network, not delegate that power to host computers that they didn’t control. Instead they ignored inter-networking altogether, and agreed on a single-network virtual circuit protocol designated X.25⁵

The Europeans, led especially by Zimmerman, made another try via a different standards body, one less dominated by the power of the telecom authorities, the International Organization for Standardization (ISO). The Open Systems Interconnection (OSI) standard that resulted had some technical advantages over TCP/IP. Notably, it lacked IP’s limited and hierarchical addressing system, whose limitations required several cheap hacks to allow for the explosive growth of the Internet in the 1990s⁶. But for a number of reasons, the process dragged out interminably without producing working software. For one thing, ISO’s processes, well-suited to blessing already established technical practices, were not appropriate for still-nascent technology. Once the TCP/IP Internet took off in the early 1990s, OSI became irrelevant.

So much for the arena of standards setting, but what about the on-the-ground practicality of network-building? The Europeans began earnestly working on an implementation of INWG 96 to link Cyclades and the National Physical Laboratory, as part of the the European Informatics Network. But Kahn and the other leaders of the ARPA Internet project did not really care to derail the TCP train for the sake of international collaboration. Kahn had already disbursed funds for TCP implementations on ARPANET and PRNET, and he didn’t want to start over. Cerf made an attempt to rally support in the U.S. for the compromise he had forged at the INWG, but finally gave up on it. He also gave up on the stresses of life as an assistant professor, following Kahn’s footsteps to become a program manager at ARPA and withdrawing from active participation in the INWG.

Why was the desire of the Europeans to establish a unified front and an official, global standard so weakly requited? The primary reason lay in the relative position of the American and European telecom authorities. The Europeans had to face constant pressure against the datagram model from the post and telecom authorities (the PTTs), which operated as administrative departments within their national governments. Because of these pressures, they had a much stronger incentive to care about building a consensus within the official standards-making processes. The rapid demise of Cyclades, which fell out of political favor in 1975 and lost all funding in 1978, provides a case study in the power of the PTTs. Pouzin blamed Cyclades’ death on the administration of Valéry Giscard d’Estaing. d’Estaing came to power in 1974, and set up a government peopled with École nationale d’administration (ENA) types, whom Pouzin disdained – if Polytechnique was something like the MIT of France, ENA was its Harvard Business School. d’Estaing’s administration focused French information technology policy around the idea of “national champions,” and a national champion computer network required the backing of the PTT. Cyclades could never acquire that support; instead Pouzin’s rival Després led the construction of a virtual-circuit X.25 network called Transpac.

The situation in the United States was quite different. A&T did not have the political leverage of its international peers, not being part of the American administrative state,  On the contrary, it was in fact in the process of being heavily constrained and weakened by that state, barred from interference in computer networking and computer services, and soon to be be dismantled entirely. ARPA could proceed with its Internet program under the umbrella of protection from the powerful Department of Defense, without any adverse political pressure. It funded TCP implementations on a variety of computers, and used its leverage to force all of the hosts on ARPANET itself to convert to the new protocol in 1983. The most influential computing network in the world, many of whose nodes happened to be the most influential academic computing institutions in the world, thus became a TCP/IP shop.

TCP/IP thus became the foundation stone of the Internet, and not just an internet, because of the relative political and financial freedom of ARPA compared to any other computer networking organization. OSI notwithstanding, ARPA became the dog, and the rest of the network research community the indignant tail. From the perspective of 1974, one can clearly see the many lines of influence that led into Cerf and Kahn’s TCP paper, and the many potential avenues of international development that might have followed from it. But from the perspective of 1995, all roads led backward to one seminal moment, one American organization, and two revered names.

[Previous] [Next]

Further Reading

Janet Abbate, Inventing the Internet (1999)

John Day, “The Clamor Outside as INWG Debated,” IEEE Annals of the History of Computing  (2016)

Andrew L. Russell, Open Standards and the Digital Age (2014)

Andrew L. Russell and Valérie Schafer, “In the Shadow of ARPANET and Internet: Louis Pouzin and the Cyclades Network in the 1970s,” Technology and Culture (2014)