Larry Roberts and ARPANET: The Network Architect

Zusammenfassung

J.C.R. Licklider dreamed of an intergalactic computer network. Bob Taylor convinced ARPA to pay for it. But the man who actually designed and built ARPANET — who made the engineering decisions, hired the contractors, managed the political battles, and turned a vision into working infrastructure — was Lawrence G. Roberts. Between 1966 and 1973, Roberts transformed a memo into a functioning four-node network, then a nationwide research backbone, and then left government to build the world’s first commercial packet-switched network. The internet that Vint Cerf and Bob Kahn later extended with TCP/IP ran on top of the infrastructure Roberts had built. He was the engineer of the revolution, not its prophet.

From Lincoln Lab to ARPA

Larry Roberts was born in 1937 in Connecticut, the son of two chemists. He earned his PhD in electrical engineering from MIT in 1963 and joined MIT Lincoln Laboratory, where he worked on computer graphics and the TX-2 computer — one of the most advanced interactive systems of its era. At Lincoln Lab, Roberts conducted what he later described as the first experiment in wide-area computer networking: in 1965, working with Thomas Marill, he linked Lincoln Lab’s TX-2 to a Q-32 at System Development Corporation in Santa Monica via a 1200-baud dial-up telephone line. The experiment worked, but barely. The telephone network was designed for voice, not data, and it showed: the connection was slow, unreliable, and expensive. Roberts and Marill published their findings, noting that the telephone circuit-switching model was fundamentally unsuited to computer communication. A different approach was needed.

That paper caught the attention of Bob Taylor, who had succeeded J.C.R. Licklider’s successor Ivan Sutherland as director of ARPA’s Information Processing Techniques Office in 1966. Taylor had three terminals in his Pentagon office, one for each of the major ARPA-funded research computers — at MIT, Berkeley, and the Systems Development Corporation — and he could not get them to talk to each other. He wanted to build a network. Roberts was, by 1966, the obvious person to build it.

The recruitment was not smooth. Roberts was happy at Lincoln Lab and had no interest in Washington bureaucracy. Taylor escalated to Charles Herzfeld, ARPA’s director, who called Lincoln Lab’s director and pointed out that a significant fraction of Lincoln Lab’s budget came from ARPA. Roberts arrived in Washington in early 1967 as chief scientist of IPTO, effectively the ARPANET program manager. He would spend the next six years turning the project into reality.

The Design Decisions That Shaped the Internet

Roberts arrived at ARPA with a specific idea already forming about how a computer network should work. The telephone network’s circuit switching — establishing a dedicated end-to-end connection before any data could flow, holding it open for the duration of the call — was inefficient for computer communication. Computers communicated in bursts: a request, a pause while computation occurred, a response. A circuit held open during the pause wasted capacity.

The alternative had been described independently by Paul Baran at RAND Corporation (in a classified 1962–1964 study) and Donald Davies at the UK’s National Physical Laboratory (in 1965–1966): packet switching. Data would be broken into discrete packets, each labeled with its source and destination, and routed independently through the network. Different packets of the same message might take different paths, reassembling at the destination. No dedicated circuit needed to be established; the network’s capacity would be shared dynamically.

Roberts adopted packet switching as the fundamental architecture. He also adopted a structural decision that would have profound consequences: the network’s core functions — routing, error detection, retransmission — would be handled by dedicated hardware at each node rather than by the host computers. This preserved the host computers’ time for research while making the network’s behavior predictable regardless of what software the hosts were running.

In 1967, Roberts presented his plan at the ACM Symposium on Operating System Principles in Gatlinburg, Tennessee. The reception was mixed — most attendees were skeptical that a packet-switched network could work reliably at scale — but the ARPA program proceeded. Roberts put out a Request for Proposals for the Interface Message Processors (IMPs): the dedicated minicomputers that would handle packet routing at each node.

The IMP Contract

Of the twelve companies that bid on the IMP contract in 1968, ARPA selected Bolt Beranek and Newman (BBN) — a small Cambridge, Massachusetts research and engineering firm with strong MIT connections. BBN’s proposal, led by Frank Heart, was technically the most sophisticated. The choice of BBN over larger defense contractors was characteristic of ARPA’s approach: bet on talent, not scale. BBN delivered the first IMP — built on a Honeywell DDP-516 minicomputer — to UCLA on August 30, 1969, five months after contract award and on schedule. The speed and precision of BBN’s execution remained legendary in networking circles for decades.

October 29, 1969: “LO”

The first ARPANET node was at UCLA, where Leonard Kleinrock ran the Network Measurement Center. The second was at the Stanford Research Institute (SRI), 350 miles north, where Douglas Engelbart — inventor of the computer mouse — ran the Augmentation Research Center. The third and fourth would follow at UC Santa Barbara and the University of Utah.

On October 29, 1969, Charley Kline, a UCLA student programmer, attempted to log in remotely to SRI’s machine. The login protocol required sending the word “LOGIN.” Kline typed “L,” and SRI’s programmer confirmed receipt. He typed “O,” and SRI confirmed receipt. He typed “G” — and the SRI machine crashed.

The first message transmitted over ARPANET was “LO.”

The crash was caused by a bug in the login software, not the network itself. The IMP-to-IMP connection had worked perfectly. Within hours the problem was fixed, and the first successful remote login followed. By December 1969, all four nodes were operational and the four-node ARPANET was functional. Roberts had taken a vision from memo to working infrastructure in under three years.

Building the Backbone (1969–1972)

Roberts managed ARPANET’s expansion with the same disciplined pragmatism that had characterized its initial design. New nodes were added methodically: fifteen by 1971, thirty-seven by 1972. Each new node required installing an IMP, connecting it to the telephone network via leased lines, and integrating the host computer’s software with the Network Control Protocol (NCP) — the first ARPANET host-to-host communication protocol, developed by the Network Working Group that Roberts had established.

The NCP was Roberts’s creation in the organizational sense: he had formed the working group of graduate students and researchers from the host institutions — including Vint Cerf, Jon Postel, and Steve Crocker — who would define how hosts communicated. This structure, in which the network’s users defined its protocols through open technical discussion documented in Requests for Comments (RFCs), was itself a lasting innovation. The RFC process remains the mechanism through which internet standards are developed today.

In October 1972, Roberts organized the International Conference on Computer Communications in Washington, where ARPANET was demonstrated publicly for the first time. The demonstration, organized by Robert Kahn, connected forty terminals to ARPANET simultaneously and showed attendees that packet switching worked at scale. Skeptics who had argued the network would collapse under load watched it perform flawlessly. The conference was the moment ARPANET became real to the outside world.

The Email Accident

In 1971, BBN engineer Ray Tomlinson sent the first email between two computers on ARPANET — not because anyone had asked him to, but because he was curious whether the file transfer protocol he was working on could be adapted to send messages between users on different machines. He chose the @ sign to separate the user name from the host machine name. Roberts quickly recognized email’s value and became one of its early champions, later noting that within a year of Tomlinson’s innovation, email had become ARPANET’s dominant traffic. The network had been designed for resource sharing; its users immediately repurposed it for communication.

Leaving ARPA: The Commercial Vision

Roberts left ARPA in 1973, a year after the public demonstration of ARPANET. His departure was purposeful: he wanted to prove that packet switching could work commercially, not just in a government research context.

He became CEO of Telenet Communications Corporation, the first commercial packet-switched network in the United States, which launched service in 1974. Telenet used ARPANET’s packet-switching principles but built its own infrastructure: dedicated packet switches (similar to IMPs), leased telephone lines, and commercial pricing structures that made network access available to businesses for the first time. Telenet was eventually acquired by GTE and became the backbone of many commercial data services in the 1980s.

Meanwhile, back at ARPA, Vint Cerf and Bob Kahn — working from 1973 onward — were designing TCP/IP: the protocol suite that would eventually replace NCP and allow ARPANET to be connected to other networks, including the ones Roberts was building. The relationship was explicit: Cerf and Kahn were extending and improving the infrastructure Roberts had built. Their 1974 paper “A Protocol for Packet Network Intercommunication” described TCP as a solution to the problem of connecting ARPANET to the packet satellite and packet radio networks that ARPA was simultaneously funding. Roberts’s network was the foundation they built on. For the story of what Cerf and Kahn built next, see Vint Cerf and Bob Kahn: The Fathers of the Internet.

Later Career: A Life in Networking

Roberts spent the rest of his career in the networking industry, cycling between startups and larger companies. He founded or co-founded a succession of networking ventures: NetExpress (ATM switching), Caspian Networks (flow-based routing), and Anagran (flow-based switching). Each represented an attempt to push packet-switching technology further — to solve congestion, improve quality of service, and make networks more efficient for the video and real-time traffic that had not existed when ARPANET was designed.

His contributions were widely recognized late in his career: the IEEE W. Wallace McDowell Award (1990), the ACM SIGCOMM Award for lifetime achievement (1998), the IEEE Internet Award (2000), and the National Academy of Engineering’s Charles Stark Draper Prize (2001, shared with Cerf, Kahn, and Leonard Kleinrock). He was inducted into the Internet Hall of Fame in 2012. In the networking engineering community, his reputation as the person who had actually built the first network was never in question, even as the public narrative of the internet’s origins focused more easily on the conceptualists — Licklider, the visionary — or the protocol designers — Cerf and Kahn, the standardizers.

Roberts died on December 26, 2018, at age 81.

The Engineer and the Credit Problem

The history of ARPANET illustrates a recurring problem in the history of technology: the people who do the hardest engineering work are often least visible in the narratives that follow.

Licklider wrote the memo that described a network. Taylor provided the political will to fund it. Cerf and Kahn designed the protocols that made it universal. Roberts did the part between the vision and the protocol: he translated an idea into a procurement process, a technical architecture, a construction project, and a working system. The choices he made — packet switching, dedicated IMPs, the NCP working group, the RFC process — were not inevitable. They were engineering decisions made under uncertainty, with alternatives that might have produced a different internet or no internet at all.

The question of who “invented” the internet has generated disputes for decades. Roberts never claimed he invented it alone. He insisted, consistently, that the key architectural decision was packet switching, and that his contribution was making that decision concrete. That was enough.