ARPANET

• 1. Introduction

  ›

  ARPANET carried its first message on the night of the 29th of October 1969. A UCLA student named Charley Kline tried to log into a computer at the Stanford Research Institute. Kline typed the command login from an SDS Sigma 7 computer sitting in Boelter Hall room 3420. The receiving machine crashed after only two letters made it through, an l and an o. About an hour later, once Stanford programmer Bill Duvall adjusted the system, Kline tried again and the full command went through cleanly. It was an unglamorous debut for a project funded by the US Department of Defense's Advanced Research Projects Agency. That agency's network eventually supplied the technical foundation for the Internet itself. Reaching that flawed first login had already taken years of arguing over funding, hardware, and how separate computers should even talk to one another. Getting from four connected machines to a global system would take two more decades. It would end with a formal shutdown and a poem written for the occasion. Who pushed the project into being, what rules governed it once it worked, and why did the agency that built it eventually take it apart?

• 2. The Terminal Problem

  ›

  Bob Taylor kept three separate computer terminals in his office, each wired to a different machine that ARPA was funding: one for the System Development Corporation's Q-32 in Santa Monica, one for Project Genie at the University of California, Berkeley, and one for Multics at MIT. Each terminal ran its own set of commands, so reaching a colleague on a different system meant getting up and walking to another machine entirely. Taylor later described the irritation directly: "For each of these three terminals, I had three different sets of user commands. So, if I was talking online with someone at S.D.C., and I wanted to talk to someone I knew at Berkeley, or M.I.T., about this, I had to get up from the S.D.C. terminal, go over and log into the other terminal and get in touch with them. I said, 'Oh Man!', it's obvious what to do: If you have these three terminals, there ought to be one terminal that goes anywhere you want to go. That idea is the ARPANET." J. C. R. Licklider, a computer scientist at Bolt Beranek and Newman, had described something close to that idea earlier, in April 1963 memoranda about an "Intergalactic Computer Network." In October 1963 Licklider was appointed head of ARPA's Behavioral Sciences and Command and Control programs. He convinced both Ivan Sutherland and Taylor that the concept mattered, though he left ARPA before any contract was signed. Sutherland and Taylor kept pushing the idea afterward, partly so ARPA-funded researchers could share scarce computers and partly to speed the spread of new software. In the early 1960s, Paul Baran at the RAND Corporation challenged the era's telephone-style circuit switching while researching how communications could survive partial destruction, including nuclear war. He developed a theoretical model of distributed adaptive message block switching, but the telecommunications establishment rejected it in favor of existing circuit-switched systems. Donald Davies at the United Kingdom's National Physical Laboratory reached a similar concept independently, in 1965. Davies would give this method a name, packet switching, in August 1968, before ARPA's own engineers had settled on how to build anything at all.

• 3. Building The Network

  ›

  In February 1966, Bob Taylor persuaded ARPA's director, Charles M. Herzfeld, to redirect one million dollars from a ballistic missile defense program into a network project. Taylor hired Larry Roberts as program manager in January 1967. Roberts met Paul Baran the following month but did not discuss networks with him at that meeting. Roberts asked Frank Westervelt to study message size and to draft a position paper on intercomputer communication, covering, in Westervelt's words, "conventions for character and block transmission, error checking and re-transmission, and computer and user identification." In April 1967, ARPA held a design session on these technical standards. Roberts initially proposed that all mainframe computers connect to one another directly, but the other investigators were reluctant to dedicate their machines to network administration. Wesley Clark then proposed using small minicomputers as go-betweens instead, and Roberts adopted the idea, naming these machines Interface Message Processors, or IMPs. Roberts presented the revised plan at the inaugural Symposium on Operating Systems Principles in October 1967. There, a colleague of Davies, Roger Scantlebury, presented the NPL packet-switching work, and Baran's ideas also drew attention. Roberts adopted Davies' packet-switching concept for ARPANET and sought Baran's advice on dynamic routing. The NPL network ran at 768 kilobits per second, and ARPA raised its own planned line speed from 2.4 to 50 kilobits per second to compete. By mid-1968, Roberts and Barry Wessler had written a final IMP specification, based on a report commissioned from the Stanford Research Institute. Roberts delivered it to Taylor on the 3rd of June 1968, who approved it on the 21st. ARPA then sent a Request for Quotation to 140 potential bidders. Most companies thought the proposal outlandish, and only twelve submitted bids; ARPA judged just four of those top-rank. By year's end the field had narrowed to two, and in January 1969 the contract went to Bolt Beranek and Newman, known as BBN. BBN's initial team numbered only seven people. Frank Heart led the group building the IMPs, while Bob Kahn led the network's theoretical design; the team also included Dave Walden, Severo Ornstein, and William Crowther. At each site, an IMP performed store-and-forward packet switching and connected to others over leased lines through modems, while host computers reached their local IMP through custom serial interfaces. BBN designed and installed the entire hardware and software system in nine months, staying in regular contact with the NPL team through meetings held in both the United States and Britain. The first-generation IMPs were built from a ruggedized Honeywell DDP-516 computer with expandable magnetic-core memory and a 16-channel direct memory access unit that wired up each host and modem. Beyond its ordinary front-panel lamps, the DDP-516 carried a special set of 24 lamps showing the status of the IMP's communication channels. Each IMP could serve up to four local hosts and reach up to six remote IMPs over early leased telephone lines. The finished network connected one computer in Utah to three in California, and the Department of Defense later allowed additional universities to join for sharing hardware and software. Those four machines couldn't yet agree on a common language for talking to each other; that problem would be tackled first through something called the 1822 protocol.

• 4. Four Machines, One Network

  ›

  At UCLA, Leonard Kleinrock built a Network Measurement Center to evaluate the fledgling system, using an SDS Sigma 7 as the site's first connected computer. At the Stanford Research Institute's Augmentation Research Center, Douglas Engelbart had already built NLS, an early hypertext system, and would go on to run the Network Information Center; the SDS 940 running NLS, nicknamed Genie, became that site's first host. At the University of California, Santa Barbara, the Culler-Fried Interactive Mathematics Center connected an IBM 360/75 running the OS/MVT operating system. At the University of Utah's School of Computing, Ivan Sutherland, who had since left ARPA, ran a DEC PDP-10 on the TENEX operating system. These four sites were chosen partly to hold down the cost of leased telephone lines and partly because each brought expertise the project needed. The first permanent ARPANET link went up on the 21st of November 1969, joining the IMPs at UCLA and SRI. By the 5th of December 1969, all four nodes were connected and the original network was complete. Elizabeth Feinler created the first Resource Handbook for ARPANET in 1969, work that grew into a full ARPANET directory built by Feinler and her team. That directory made it possible for users to actually find their way around the growing network. In 1968, Larry Roberts had contracted with Kleinrock to measure the network's performance and identify where it could improve. Building on his earlier work in queueing theory and message-delay optimization, Kleinrock produced mathematical models of packet-switched network performance that underpinned ARPANET's rapid growth in the early 1970s. Analytic methods only went so far, so ARPA and the NPL in Britain also turned to computer simulation studies to fill the gaps. None of this solved the more basic problem still facing those four machines: agreeing on a shared language for talking to one another.

• 5. Teaching Computers To Talk

  ›

  In 1969, computers on ARPANET spoke to their local IMP using the 1822 protocol, a message format built to work across many different machine architectures. An 1822 message consisted of little more than a message type, a numeric host address, and a data field; the IMP delivered it either to a locally connected host or onward to another IMP. Once a message reached its destination, the receiving IMP sent a Ready for Next Message acknowledgment back to confirm delivery. That simple scheme proved inadequate once a single host needed to run several applications with separate network connections at once. The fix was the Network Control Program, or NCP, which gave different processes on different host computers a standard way to open reliable, flow-controlled, bidirectional links, an early example of the layered-protocol concept the OSI model would later formalize. NCP was developed under Steve Crocker, then a graduate student at UCLA, who created and led the Network Working Group, a collection of ARPA-sponsored graduate students at universities and research labs, including Jon Postel at UCLA. Beginning in 1972, Louis Pouzin and Hubert Zimmermann at France's IRIA pioneered a simplified networking approach they called a catenet, presented at the International Conference on Computer Communications. Bob Kahn left BBN that year, briefly served as president of Telenet, then moved to DARPA, first as ARPANET's program manager under Roberts and later as director of the Information Processing Techniques Office. Crocker, now at DARPA, joined British and French colleagues in founding the International Network Working Group, and on Crocker's recommendation, Vint Cerf, then on the Stanford faculty, became its chair. Bob Metcalfe, meanwhile, developed the theory and practice behind Ethernet and the PARC Universal Packet. Research by Kahn and Cerf led to the Transmission Control Program, formulated in 1974; Cerf wrote its specification that December with Yogen Dalal and Carl Sunshine at Stanford, and researchers began circulating it as a series of Internet Experiment Notes. Testing began the following year through parallel implementations at Stanford, BBN, and University College London. The design was rebuilt as a modular protocol stack in Version 4 in 1978, drawing on ideas from the French CYCLADES project and from Metcalfe's work at Xerox PARC. That version went into production use on ARPANET on the 1st of January 1983, a changeover known as flag day, finally replacing NCP for good. The switch also fulfilled an order the Department of Defense had issued three years earlier, in 1980, making TCP/IP the standard for all of its military computer networking.

• 6. Growing Past The Lab

  ›

  Howard Frank, brought in to consult on the network's topology, recommended changes meant to raise throughput and cut costs as ARPANET scaled up. By March 1970, the network reached the East Coast when an IMP at BBN's own offices in Cambridge, Massachusetts came online. IMP counts then climbed fast: 9 by June 1970-13 by December 1970, and 18 by September 1971, by which point the network already linked 23 university and government hosts. There were 29 IMPs by August 1972 and 40 by September 1973. Growth continued to 46 IMPs by June 1974 and 57 by July 1975. By 1981, host computers on the network numbered 213, with a new one joining roughly every twenty days. Engineers added support in 1970 for inter-IMP circuits running up to 230.4 kilobits per second, though cost and limited IMP processing power left that capacity largely unused. Larry Roberts wanted to link ARPANET to Britain's NPL network, and in 1970 began pursuing a satellite connection to do it. Peter Kirstein's research group at University College London was chosen in 1971 to handle the UK end instead of NPL itself. In June 1973, a transatlantic satellite link joined ARPANET to the Norwegian Seismic Array through the Tanum Earth Station in Sweden, continuing overland to a Terminal Interface Processor at UCL. UCL then served as a gateway to British academic networks, the first international resource-sharing network of its kind, and carried out some of the earliest internetworking research anywhere. Starting in 1971, ARPANET began using the lighter, non-ruggedized Honeywell 316 as an IMP, a machine that could also serve as a Terminal Interface Processor, supporting up to 63 ASCII serial terminals through a multi-line controller. The 316 was more tightly integrated than the older 516, making it cheaper and easier to maintain; a TIP configuration carried 40 kilobytes of core memory, which grew by 1973 to 32 kilobytes for ordinary IMPs and 56 kilobytes for TIPs. The network was demonstrated publicly at the International Conference on Computer Communications in October 1972. In 1975, BBN introduced IMP software running on a multiprocessor called Pluribus, though it reached only a few sites, and in 1981 BBN released IMP software for its own C/30 processor. Operational control of ARPANET passed to the Defense Communications Agency in the summer of 1975, around the time the first ARPANET encryption devices went into service to protect classified traffic. A joint BBN and DARPA report on the project, written in 1978 and published in 1981, closed by calling it fitting that ARPANET had fed back into the strength of the very field of computer science that had produced it. Access widened again in 1981, when the National Science Foundation funded the Computer Science Network, CSNET, reaching researchers beyond the original defense-funded sites. The transatlantic link to Norway and UCL eventually grew into its own system, SATNET, which was interconnected with ARPANET and a radio-based network called PRNET in 1977. Norway and UCL switched away from ARPANET altogether in 1982, adopting TCP/IP over SATNET ahead of the wider changeover. In September 1984, the network was restructured to give US military sites their own system, MILNET, for unclassified defense communication, connected to the rest of ARPANET only at a small number of controlled gateways that could be severed entirely in an emergency. Splitting off MILNET cut the 113-node ARPANET by 68 nodes; what remained kept serving researchers as an Internet backbone while being gradually phased out. Even as its IMP count kept climbing through the 1970s, most of what actually crossed that hardware had little to do with the military mission that had funded it in the first place.

• 7. Mail, Manners And Passwords

  ›

  Ray Tomlinson, an engineer at BBN, sent the first network email in 1971, a small technical event that would come to dominate ARPANET's traffic within two years. A 1973 ARPA study found that three-quarters of all network traffic consisted of email, and mail stayed a dominant share of use from then on. Telnet, the protocol for remote time-sharing access, dated to 1969, beginning with a document called RFC 15 and later extended in RFC 855. Abhay Bhushan wrote the original File Transfer Protocol specification, published on the 16th of April 1971, and by 1973 FTP was fully defined and running, moving files across the network and doubling, in its early years, as a crude way to send mail. In the 1980s, FTP-based mail gave way to the Simple Mail Transfer Protocol, and later to POP and IMAP. The Network Voice Protocol was specified in 1977 in an attempt to carry conference calls over ARPANET, but technical shortcomings meant it never worked well; real packet voice, what would become Voice over Internet Protocol, was still decades away. At the request of Larry Roberts, then head of ARPA, engineers developed the Purdy Polynomial hash algorithm in 1971 to protect user passwords. It computed a polynomial of degree 2 to the 24th power plus 17, modulo a 64-bit prime equal to 2 to the 64th power minus 59. Digital Equipment Corporation later adopted the Purdy Polynomial to hash passwords in its VMS operating system, and it is reportedly still used there for that purpose. Leonard Kleinrock claims credit for the first illegal act committed over the network: in 1973, after a meeting in England, he sent a message asking for the return of an electric razor he had left behind, at a time when personal use of ARPANET for non-government business was against the rules. In 1978, Gary Thuerk of Digital Equipment Corporation broke that rule far more visibly, sending the first mass marketing email to roughly 400 potential clients over ARPANET. Thuerk claimed the campaign generated 13 million dollars in DEC sales, an early demonstration of email's commercial potential achieved by breaking the network's own rules to get there. A 1982 handbook on computing at MIT's AI Lab spelled out the etiquette bluntly, treating friendly personal messages as generally tolerated while warning that using ARPANET for commercial profit or political purposes was, in its words, "both anti-social and illegal." The password scheme born in 1971 would go on protecting logins for decades after the rule against Thuerk's kind of message had become impossible to enforce on a network about to change hands entirely.

• 8. Retirement And Myth

  ›

  By 1985, the National Science Foundation began funding supercomputing centers at several American universities, and a year later launched the NSFNET project to link them together and to the wider network. NSFNET quickly became the backbone linking government agencies and universities, a role ARPANET had never been built to fill at that scale. ARPA formally decommissioned ARPANET in 1990, retiring the original IMPs and TIPs that had carried its traffic for two decades. A few of those machines kept running as late as July 1990, the last holdouts of a network that had outlived its own hardware generation. On the 28th of February 1990, as the shutdown took effect, Vint Cerf marked the occasion with a poem titled "Requiem of the ARPANET": "It was the first, and being first, was best, but now we lay it down to ever rest. Now pause with me a moment, shed some tears. For auld lang syne, for love, for years and years of faithful service, duty done, I weep. Lay down thy packet, now, O friend, and sleep." He signed it simply, Vinton Cerf. The packet switching, decentralized design, and communication protocols ARPANET proved out became the foundation the global Internet was built on. Its influence spread outward too, touching a web of other research projects that either shaped ARPANET's own design or branched off from it entirely. In 1988, a group chaired by Leonard Kleinrock submitted a concept for a National Research Network to Congress, and Senator Al Gore took up the idea. Gore authored the High Performance Computing and Communication Act of 1991, known as the Gore Bill, which passed on the 9th of December 1991. The act led to the National Information Infrastructure, which Gore described as an information superhighway. The ARPANET project received two IEEE Milestone honors, both dedicated in 2009. On the 17th of May 2011, Arlington County, Virginia placed two historical markers at 1400 Wilson Boulevard in its Rosslyn neighborhood, the agency's home between 1970 and 1975. Charles Herzfeld, who directed ARPA from 1965 to 1967, later pushed back hard on a persistent myth about why the project existed. "The ARPANET was not started to create a Command and Control System that would survive a nuclear attack, as many now claim," he wrote. "To build such a system was, clearly, a major military need, but it was not ARPA's mission to do this; in fact, we would have been severely criticized had we tried. Rather, the ARPANET came out of our frustration that there were only a limited number of large, powerful research computers in the country, and that many research investigators, who should have access to them, were geographically separated from them." The Internet Society later backed Herzfeld's account, tracing the nuclear-survival rumor to an earlier RAND study and noting that only later internetworking work, not ARPANET itself, focused on surviving the loss of large portions of a network. That RAND study had been conducted by Paul Baran, the same researcher whose packet-switching model the telecommunications industry had rejected years before; Baran said his work still contributed to ARPANET's eventual development. Notes taken by Elmer Shapiro of the Stanford Research Institute at an ARPANET design meeting on the 9th and the 10th of October 1967 suggest engineers did consider a version of Baran's "hot potato" routing method. By the 1970s, ARPA's own rhetoric had shifted anyway. Stephen J. Lukasik, who directed the agency from 1970 to 1975, described the goal plainly as exploiting new computer technology for military command and control against nuclear threats, for achieving survivable control of US nuclear forces, and for improving military decision making. Lukasik's language of nuclear threats and military decision making shows how far the agency's public rationale had shifted in well under a decade.