ARPANET: Building the Network That Became the Internet

Zusammenfassung

ARPANET was the first wide-area packet-switched network, built by the US Department of Defense’s Advanced Research Projects Agency beginning in 1969. What started as an experiment connecting four university computers grew into a 37-node research network by 1972, acquired email almost by accident in 1971, and gave birth to the TCP/IP protocols that underpin the modern internet. ARPANET was decommissioned on February 28, 1990, but the network it seeded never stopped growing. For the visionary ideas that motivated it, see J.C.R. Licklider and the Intergalactic Network; for the protocols it eventually ran on, see Vint Cerf and Bob Kahn.

The Problem with Three Terminals

In 1966, Bob Taylor was running the Information Processing Techniques Office (IPTO) at ARPA from a corner office in the Pentagon. On one wall of that office sat three terminals — one connected to the time-sharing system at MIT’s Project MAC, one to a computer at the System Development Corporation in Santa Monica, and one to a machine at UC Berkeley. Each time Taylor needed to use a different system, he had to swivel his chair to a different terminal and retype his login credentials. The inefficiency galled him.

More fundamentally, Taylor recognized that the researchers ARPA funded at different universities couldn’t share programs, data, or computing resources across institutional lines. Each node in the ARPA research community was an island. If a team at MIT developed a useful piece of software, the team at Stanford couldn’t run it. The $15 million ARPA was spending annually on computing research was being partly duplicated because the same problems were being solved independently at multiple sites.

Taylor walked down the hall to Charles Herzfeld, ARPA’s director, and pitched the idea of a network connecting the ARPA-funded research computers. Herzfeld gave him a million dollars from another program’s budget. The meeting lasted twenty minutes.

To run the program, Taylor recruited Larry Roberts from MIT’s Lincoln Laboratory. Roberts had already conducted the first experiments in wide-area networking in 1965, connecting a TX-2 computer at Lincoln to a Q-32 at SDC over a dial-up telephone line — work that demonstrated both that packet switching could work and that the telephone network’s circuit-switching architecture was the wrong model for computer communication. Roberts became ARPA’s chief program manager for the network project in 1966 and spent the next two years designing what would become ARPANET.

Packet Switching and the IMP Contract

The architectural decision that made ARPANET conceptually revolutionary was the adoption of packet switching — a technique developed independently by Paul Baran at RAND Corporation and Donald Davies at the National Physical Laboratory in the UK. Rather than establishing a dedicated circuit between sender and receiver (as the telephone network did), packet switching broke messages into small chunks, each labeled with source and destination, and allowed each chunk to travel independently through whatever route was available. Networks built this way were inherently redundant: if a node failed, packets could route around it.

Baran had developed the idea explicitly for military communications that needed to survive nuclear attack. Davies coined the term “packet.” Roberts synthesized both lines of work into the ARPANET design.

In July 1968, ARPA issued a request for proposals to build the Interface Message Processors (IMPs) — the specialized computers that would sit at each network node, handling packet routing and transmission so the host computers at each site didn’t have to. Twelve companies bid. The contract was awarded in late December 1968 to Bolt Beranek and Newman (BBN) in Cambridge, Massachusetts, led by Frank Heart. BBN had roughly nine months to build and deploy the first four IMPs.

Heart’s team was small — a dozen engineers — and the schedule was brutal. They chose a Honeywell 516 minicomputer as the IMP hardware, hardened it for continuous operation, and wrote the operating software essentially from scratch. The contract specified that the IMPs had to handle messages reliably even when host computers or network links failed. Every design decision was made under the constraint that the system had to be robust against component failure.

Four Nodes, One Crash, One Message

The first IMP was shipped to UCLA on August 30, 1969, and connected to a Sigma 7 computer. Len Kleinrock at UCLA had been running the network measurement center. The second IMP arrived at the Stanford Research Institute (SRI) in early October. UCLA and SRI were Node 1 and Node 2 of ARPANET.

The first network message was sent on October 29, 1969. A UCLA graduate student named Charley Kline sat at the terminal and began typing “LOGIN” to establish a connection to SRI’s host computer. He typed “L” — SRI received “L.” He typed “O” — SRI received “O.” He typed “G” — and the SRI system crashed.

The first message ever sent on ARPANET was “LO.” Kline and the SRI team rebooted the system, and by 10:30 pm the full login worked. The crash was logged by Kleinrock’s team with the dry annotation: “Talked to SRI host to host.”

Nodes 3 and 4 — at UC Santa Barbara and the University of Utah — came online in December 1969, completing the original four-node network. By the end of 1969, the backbone of what would become the internet existed.

The Significance of Packet Switching

Circuit-switched networks (like the telephone system) dedicate a path for the duration of a call, reserving bandwidth whether or not it is being used. Packet-switched networks are statistically multiplexed: the same physical link can carry packets from many simultaneous conversations, using its capacity efficiently. This difference in architecture — not just in speed or capacity — is why the internet scales so differently from the telephone network. An ARPANET link being used at 50% capacity was available at full speed for anyone who needed it; a phone circuit being used at 50% capacity was unavailable to anyone else.

Email: The Killer App Nobody Planned

By 1971, ARPANET had grown to 15 nodes (23 hosts) and was being used mostly for remote login and file transfer. Then Ray Tomlinson at BBN accidentally invented something that would change communication forever.

Tomlinson was working on two separate programs: SNDMSG, which let users on a time-sharing system leave messages for each other in mailbox files, and CPYNET, which transferred files between ARPANET hosts. In the autumn of 1971, he had an insight: if CPYNET could transfer a file between machines, and SNDMSG could append a message to a mailbox file, the two programs could be combined to send a message to a mailbox on a different machine. He modified SNDMSG to do exactly that, and needed a way to specify that the destination mailbox was on a different host rather than the local machine.

He chose the @ symbol — already on the keyboard, not used in usernames, unambiguous as a separator between user and host. The choice was made in minutes. Tomlinson later said he could not remember what the first email said, only that it was something like “QWERTYUIOP” — a test to see if the mechanism worked.

By 1973, email accounted for approximately 75% of ARPANET traffic. The network had been designed to share computing resources; what people actually used it for was sending messages to each other.

The Network Grows

ARPANET expanded rapidly through the early 1970s. By 1971, there were 15 nodes. By the time of the landmark public demonstration in October 1972, there were 37.

That demonstration — the International Conference on Computer Communication (ICCC) in Washington, DC — was ARPANET’s coming-out party. Larry Roberts organized 40 terminals in the Washington Hilton ballroom, connected to the network, where conference attendees could sit down and use time-sharing systems in Cambridge, Santa Monica, and Los Angeles. The demonstration ran smoothly, which surprised nearly everyone who had worked on the network: distributed systems had a reputation for spectacular failures, and this one chose not to fail in public.

The demonstration was important not just technically but politically. Skeptics at AT&T had dismissed packet switching as impractical; AT&T executives attended ICCC, saw the demonstration, and declined to take over ARPANET when offered the chance. They believed packet switching could never work at telephone-network scale. It was one of the most consequential decisions in telecommunications history.

NCP, TCP/IP, and Flag Day

The original ARPANET ran on NCP (Network Control Program), a protocol suite developed by Steve Crocker and others in the Network Working Group starting in 1970. NCP worked but had fundamental limitations: it assumed hosts were connected to a single network and could not accommodate the idea of multiple interconnected networks.

As ARPANET expanded and satellite links and radio packet networks were added, the need for a more flexible protocol became clear. In 1974, Vint Cerf and Bob Kahn published “A Protocol for Packet Network Intercommunication,” introducing TCP/IP — a design where networks could be interconnected by gateways, and each network didn’t need to know about the others’ internal architecture. The protocol was designed so that unreliable networks could be composed into a reliable internet.

The transition from NCP to TCP/IP was managed by Jon Postel and others over several years. The cutover date — January 1, 1983 — was called Flag Day. On that date, all ARPANET hosts switched from NCP to TCP/IP. There was no fallback, no gradual migration, no long transition period. The new protocol simply became mandatory.

NSFNET and the Transition to the Modern Internet

Through the late 1970s and early 1980s, ARPANET remained a research network, its access restricted to military and academic institutions with ARPA contracts. The broader research community needed connectivity too. In 1985, the National Science Foundation funded NSFNET, connecting five supercomputer centers with 56 kbps links and opening the network to any university with a legitimate research purpose.

NSFNET quickly overwhelmed its original capacity and was upgraded to 1.5 Mbps T1 links in 1988, then to 45 Mbps T3 links in 1991. By this point, ARPANET was increasingly redundant — a smaller, older network operating in parallel with the faster and more accessible NSFNET backbone.

ARPANET was formally decommissioned on February 28, 1990. The ceremony was quiet; most users had already migrated to NSFNET. The four original IMPs had long since been replaced. The physical network ceased to exist, but the protocols, the architecture, and the culture it had created remained.

In 1995, NSF privatized NSFNET and withdrew from backbone operation, opening the internet to full commercial use. The network that Bob Taylor had proposed to end the absurdity of three terminals in his office had become the global communications infrastructure.

The People Who Built It

ARPANET’s success was built on a specific culture: small teams with wide latitude, funded by a government agency that trusted them to make good decisions without detailed oversight. The people who built it — Roberts, Heart, Tomlinson, Kleinrock, Cerf, Kahn, Postel — were mostly young, mostly academic, and deeply motivated by the intrinsic interest of the problem rather than by commercial applications.

Jon Postel, who served as RFC editor and custodian of internet address assignments from 1969 until his death in 1998, embodied this culture. He worked out of USC’s Information Sciences Institute for his entire career, never accumulating significant personal wealth, maintaining the internet’s core administrative functions essentially as a public service. His death at 55, two weeks after undergoing heart surgery, produced a genuine outpouring of grief in the technical community — the recognition that a person who had been irreplaceable had left.

The network they built was not designed for commercial use, for security, or for global scale. It was designed to connect researchers who needed to share computing resources. That its underlying architecture turned out to be robust enough to scale to billions of users and trillions of daily transactions was not entirely foreseen. But the essential insight — that packet switching could create a resilient, extensible network out of unreliable components — was exactly right, and it has not needed to be replaced.