Skip to content

Louis Pouzin and the Datagram

Abstract

Between 1971 and 1978 a small team at the French research institute IRIA built CYCLADES, a packet network whose subnet promised nothing: no delivery guarantee, no ordering, no memory of what it had carried. Reliability was moved out to the machines at the ends. Louis Pouzin, who ran the project, argued that a network which tries to be perfect makes the hosts pay for it twice, and the argument reached Vint Cerf and Bob Kahn while TCP was still being written. The French post and telecommunications administration preferred the opposite design, funded Transpac instead, and CYCLADES was shut down in 1981. The network is gone; the argument is in every IP packet.

Bull, MIT, and the Word “Shell”

Louis Pouzin was born on 20 April 1931 in Chantenay-Saint-Imbert in central France and studied at the École Polytechnique from 1950 to 1952. He worked at Bull, the French computer manufacturer, then spent 1963 to 1965 on the staff of the MIT Computation Center, where the Compatible Time-Sharing System was being built.

There he wrote RUNCOM, a program that executed scripts of commands with argument substitution. When he sketched a command language for Multics shortly afterwards, he needed a name for the layer that sat around the system and took the user’s instructions. “I coined the word ‘shell’ to name it,” he later recalled. “It must have been at the end of 64 or beginning of 65.” He drew a flowchart, went back to France in mid-1965, and Glenda Schroeder at MIT, with a GE engineer, wrote the first Multics shell from it. Kernighan and Ritchie later described the rc in Unix file names as a fossil of the same CTSS facility, which is why configuration files are still called .bashrc and .vimrc.

The Cyclades Assignment

France in the late 1960s was spending public money on computing as a matter of sovereignty. The Plan Calcul of 1966 created a national computer champion, and in 1967 it created IRIA, the research institute at Rocquencourt west of Paris. See France’s Tech Industry. In 1971 the government gave Pouzin a network project with a specific brief: link the machines of the French research centres, and do it in a way that could later be extended abroad.

The reference design was ARPANET, which had been running since 1969 and which Pouzin had studied at close range. He took one central decision in the opposite direction.

Cigale

ARPANET’s subnet of IMPs accepted a message and undertook to deliver it, in order, to the destination host. Doing that required the switches to hold state: buffers reserved per conversation, acknowledgements between switches, retransmission inside the network. Pouzin’s subnet, called CIGALE, did none of it. It accepted a packet, routed it, and forgot it. Packets could arrive out of order, duplicated, or not at all. Sorting that out was the job of a transport protocol running on the hosts, which numbered the packets, acknowledged them and slid a window along the stream.

The name for the packets came from outside the project: Halvor Bothner-By, the CCITT rapporteur on packet switching, built the word datagram out of data and telegram in the early 1970s.

Pouzin’s case was that the network could not know what reliability any given application needed, so anything it guaranteed had to be either insufficient or redundant. “I didn’t see any real technical motivation to overlay two levels of end-to-end protocols,” he said. A file transfer wants every byte; a terminal session wants the latest keystroke and does not care about a lost one. Since the hosts had to check anyway, the switches could stay simple, cheap and stateless, and the network could be extended by anyone who implemented the protocol.

The Network

CYCLADES was demonstrated in November 1973 with three hosts and one packet switch. Seven switches were running by June 1974 and twenty nodes by 1976, with links to the National Physical Laboratory in London and to Rome.

The design travelled faster than the network. Pouzin, Steve Crocker, Donald Davies and Peter Kirstein founded the International Network Working Group in Paris in June 1972, with Cerf as chair from that October. Gérard Le Lann, one of Pouzin’s engineers, spent time in Cerf’s group at Stanford while the first Transmission Control Program was being specified, and brought the datagram argument and the sliding window with him. Cerf and Kahn credited CYCLADES in print. What the INWG could not do was agree on one internetwork protocol: the group’s compromise of 1975, INWG 96, was rejected by CCITT, by ISO and by ARPA, each of which continued with its own.

Dead End: The Administration Preferred Circuits

The French PTT was building a public data network of its own, and its engineers wanted the network to provide a reliable ordered pipe, because that is what a telephone administration sells and what it can bill. The CCITT standardised X.25 in 1976 on virtual circuits. Transpac opened in December 1978 on that model, and it worked: it carried the Minitel traffic of the 1980s and by 1987 was the largest public packet network in the world.

Two national networks with incompatible philosophies were more than the budget would carry. Government funding for CYCLADES was cut, the team dispersed, and the network was shut down in 1981. France then spent the 1980s with the most successful videotex system on earth running over virtual circuits, and arrived late to the internet.

Pouzin has been asked about this often enough to have a settled answer: the decision was administrative, not technical, and the people who made it understood exactly what a stateless network would do to a monopoly’s control over its own traffic.

Afterwards

The recognition arrived once TCP/IP had won. Pouzin received the ACM SIGCOMM Award in 1997, was made a Chevalier of the Légion d’honneur in 2003 and an Officier in 2018, entered the Internet Hall of Fame in 2012, and shared the first Queen Elizabeth Prize for Engineering in 2013 with Robert Kahn, Vinton Cerf, Tim Berners-Lee and Marc Andreessen. The citation put Kahn, Cerf and Pouzin together as the authors of the internet’s fundamental architecture.

He did not treat that as a conclusion. In 2011, at eighty, he founded Savoir-Faire, and the following year launched Open-Root, a service selling top-level domains outside the ICANN system, on the argument that a single root under American oversight is a political structure rather than a technical necessity, and he has spent years pressing for domain names in scripts other than Latin. The complaint is recognisably the same one he made in 1972: whoever holds the centre of a network holds the network.

📚 Sources