But Roberts kept turning down Taylor’s offer to come to Washington to be his deputy. He liked his job at Lincoln Laboratory and didn’t especially respect Taylor. There was also something Taylor didn’t know: a year earlier, Roberts had been offered Taylor’s job. “When Ivan was leaving, he asked me to come to IPTO as the next director, but it was a management job, and I preferred research,” he said. Having declined the top post, Roberts was not about to be Taylor’s deputy. “Forget it,” he told Taylor. “I’m busy. I’m having fun with this wonderful research.”43
There was another reason Roberts resisted, which Taylor could sense. “Larry was from MIT with a doctorate, and I was from Texas with just a master’s,” Taylor later said. “So I suspect he didn’t want to work for me.”44
Taylor, however, was a clever and stubborn Texan. In the fall of 1966, he asked Herzfeld, “Charlie, doesn’t ARPA fund 51 percent of Lincoln Laboratory?” Herzfeld confirmed that. “Well, you know this networking project that I want to do, I’m having a hard time getting the program manager that I want, and he works at Lincoln Laboratory.” Perhaps Herzfeld could call the head of the lab, Taylor suggested, and say that it would be in its interest to convince Roberts to accept the job. It was a Texas way of doing business, as the president at the time, Lyndon Johnson, would have appreciated. The lab’s chief was no dummy. “It would probably be a nice thing for all of us if you’d consider this,” he pointed out to Roberts after getting Herzfeld’s call.
So in December 1966, Larry Roberts went to work at ARPA. “I blackmailed Larry Roberts into becoming famous,” Taylor later said.45
When Roberts first moved to Washington, around Christmas, he and his wife stayed for a few weeks with Taylor while looking for a home. Even though they were not destined to be personal pals, the relationship between the two men was cordial and professional, at least during their years at ARPA.46
Roberts was not as genial as Licklider, nor as extroverted as Taylor, nor as congregational as Bob Noyce. “Larry’s a cold fish,” according to Taylor.47 Instead he had a trait that was just as useful in promoting collaborative creativity and managing a team: he was decisive. More important, his decisiveness was based not on emotion or personal favoritism but rather on a rational and precise analysis of options. His colleagues respected his decisions, even if they disagreed with them, because he was clear, crisp, and fair. It was one of the advantages of having a true product engineer in charge. Uncomfortable at being Taylor’s deputy, Roberts was able to work out an arrangement with ARPA’s top boss, Charlie Herzfeld, to be designated the agency’s chief scientist instead. “I managed contracts during the day and did my networking research at night,” he recalled.48
Taylor, on the other hand, was jocular and gregarious, sometimes to a fault. “I’m an outgoing person,” he observed. Each year he would convene a conference of the ARPA-funded researchers and another for their best graduate students, usually in fun places like Park City, Utah, and New Orleans. He made each researcher give a presentation, and then everyone could pile on with questions and suggestions. In that way he got to know the rising stars around the country, making him a magnet for talent that would later serve him well when he went to work at Xerox PARC. It also helped him accomplish one of the most important tasks in building a network: getting everyone to buy into the idea.
ARPANET
Taylor knew that he needed to sell the time-sharing network idea to the people it was intended to help, namely the researchers who were getting ARPA funding. So he invited them to a meeting at the University of Michigan in April 1967, where he had Roberts present the plan. The computer sites would be connected, Roberts explained, by leased phone lines. He described two possible architectures: a hub system with a central computer in a place like Omaha that would route information, or a weblike system that looked like a highway map with lines crisscrossing as they were spun from place to place. Roberts and Taylor had begun to favor the decentralized approach; it would be safer. The information could be passed along from node to node until it reached its destination.
Many of the participants were reluctant to join the network. “The universities in general did not want to share their computers with anybody,” Roberts said. “They wanted to buy their own machines and hide in the corner.”49 Nor did they want the valuable processing time of their computers to be nibbled away by having to handle the traffic routing that would come with being on the network. The first to dissent were Marvin Minsky of the MIT Artificial Intelligence Lab and his former colleague John McCarthy, who had moved to Stanford. Their computers, they said, were already being used to the max. Why would they want to allow others to tap into them? In addition they would have the burden of routing network traffic from computers they didn’t know and whose language they didn’t speak. “Both complained they would lose computing power and said they didn’t want to participate,” Taylor recalled. “I told them they had to, because it would let me cut my funding of computers by a factor of three.”50
Taylor was persuasive and Roberts persistent, and they pointed out to the participants that they were all being funded by ARPA. “We are going to build a network and you are going to participate in it,” Roberts declared flatly. “And you are going to connect it to your machines.”51 They would get no more funding to buy computers until they were hooked into the network.
Ideas are often sparked by the exchanges at meetings, and one popped up at the end of the Michigan session that helped to defuse opposition to the network. It came from Wes Clark, who had conceived a personal computer at Lincoln Laboratory dubbed the LINC. He was more interested in developing computers designed for individual use than he was in promoting time-sharing of large computers, so he hadn’t been paying much attention. But as the meeting was ending he realized why it was hard getting the research centers to accept the network idea. “Just before we broke up, I do remember suddenly realizing what the meta-problem was,” he said. “I passed Larry a note saying that I thought I saw how to solve the problem.”52 On the ride to the airport, in a rental car that Taylor was driving, Clark explained his idea to Roberts, along with two other colleagues. ARPA should not force the research computers at each site to handle the routing of data, Clark argued. Instead ARPA should design and give each site a standardized minicomputer that would do the routing. The big research computer at each site would then have only the simple task of establishing a connection with its ARPA-supplied routing minicomputer. This had three advantages: it would take most of the burden off the host site’s mainframe, give ARPA the power to standardize the network, and allow the routing of data to be completely distributed rather than controlled by a few big hubs.
Taylor embraced the idea right away. Roberts asked a few questions and then agreed. The network would be managed by the standardized minicomputers that Clark had suggested, which became known as Interface Message Processors, or IMPs. Later they would simply be called “routers.”
When they got to the airport, Taylor asked who should build these IMPs. Clark said it was obvious: the task should be assigned to Bolt, Beranek and Newman, the Cambridge firm where Licklider had worked. But also in the car was Al Blue, who was in charge of compliance issues at ARPA. He reminded the group that the project would have to be sent out for bids in accordance with federal contracting standards.53
* * *
At a follow-up conference in Gatlinburg, Tennessee, in October 1967, Roberts presented the revised plan for the network. He also gave it a name, ARPA Net, which later morphed into ARPANET. But one issue remained unresolved: Would communication between two places on the network require a dedicated line between them, as a phone call did? Or was there some practical way to allow multiple data streams to share lines simultaneously, sort of like a time-sharing system for phone lines? Potential specifications for such a data network had been proposed earlier that month by a committee at the Pentagon.
That is when a young engineer from England, Roger Scantlebury, got up to present a paper describing the resear
ch of his boss, Donald Davies of Britain’s National Physical Laboratory. It provided an answer: a method of breaking messages into small units that Davies had dubbed “packets.” Scantlebury added that the idea had been developed independently by a researcher named Paul Baran at RAND. After the talk, Larry Roberts and others gathered around Scantlebury to learn more, then moved on to the bar to discuss it late into the night.
PACKET SWITCHING: PAUL BARAN, DONALD DAVIES, AND LEONARD KLEINROCK
There are many ways of sending data through a network. The simplest, known as circuit switching, is the way a phone system does it: a set of switches creates a dedicated circuit for signals to go back and forth for the duration of the conversation, and the connection remains open, even during long pauses. Another method is message switching or, as the telegraph operators called it, store-and-forward switching. In this system, an entire message is given an address header, sent into the network, and then passed along from node to node as it wends its way to its destination.
An even more efficient method is packet switching, a special type of store-and-forward switching in which the messages are broken into bite-size units of the exact same size, called packets, which are given address headers describing where they should go. These packets are then sent hopping through the network to their destination by being passed along from node to node, using whatever links are most available at that instant. If certain links start getting clogged with too much data, some of the packets will be routed to alternative paths. When all the packets get to their destination node, they are reassembled based on the instructions in the headers. “It’s like breaking a long letter into dozens of postcards, each numbered and addressed to the same place,” explained Vint Cerf, one of the Internet’s pioneers. “Each may take different routes to get to the destination, and then they’re reassembled.”54
As Scantlebury explained in Gatlinburg, the person who first fully conceived a packet-switched network was an engineer named Paul Baran (pronounced BEAR-en). His family had immigrated from Poland when he was two and settled in Philadelphia, where his father opened a small grocery store. After graduating from Drexel in 1949, Baran joined Presper Eckert and John Mauchly in their new computer company, where he tested components for UNIVAC. He moved to Los Angeles, took night classes at UCLA, and eventually got a job at the RAND Corporation.
When the Russians tested a hydrogen bomb in 1955, Baran found his life mission: to help prevent a nuclear holocaust. One day at RAND he was looking at the weekly list sent by the Air Force of topics it needed researched, and he seized on one that related to building a military communications system that would survive an enemy attack. He knew that such a system could help prevent a nuclear exchange, because if one side feared that its communications system could be knocked out it would be more likely to launch a preemptive first strike when tensions mounted. With survivable communications systems, nations would not feel the need to adopt a hair-trigger posture.
Donald Davies (1924–2000).
Paul Baran (1926–2011).
Leonard Kleinrock (1934– ).
Vint Cerf (1943– ) and Bob Kahn (1938– ).
Baran came up with two key ideas, which he began publishing in 1960. His first was that the network should not be centralized; there should be no main hub that controlled all the switching and routing. Nor should it even be merely decentralized, with the control in many regional hubs, like AT&T’s phone system or the route map of a major airline. If the enemy took out a few such hubs, the system could be incapacitated. Instead control should be completely distributed. In other words, each and every node should have equal power to switch and route the flow of data. This would become the defining trait of the Internet, the ingrained attribute that would allow it to empower individuals and make it resistant to centralized control.
He drew a network that looked like a fishnet. All of the nodes would have the power to route traffic, and they were each connected to a few other nodes. If any one of the nodes was destroyed, then the traffic would just be routed along other paths. “There is no central control,” Baran explained. “A simple local routing policy is performed at each node.” He figured out that even if each node had only three or four links, the system would have almost unlimited resilience and survivability. “Just a redundancy level of maybe three or four would permit almost as robust a network as the theoretical limit.”55
“Having figured out how to get robustness, I then had to tackle the problem of getting signals through this fishnet type of network,” Baran recounted.56 This led to his second idea, which was to break up the data into standard-size little blocks. A message would be broken into many of these blocks, each of which would scurry along different paths through the network’s nodes and be reassembled when they got to their destination. “A universally standardized message block would be composed of perhaps 1024 bits,” he wrote. “Most of the message block would be reserved for whatever type data is to be transmitted, while the remainder would contain housekeeping information such as error detection and routing data.”
Baran then collided with one of the realities of innovation, which was that entrenched bureaucracies are resistant to change. RAND recommended his packet-switched network idea to the Air Force, which, after a thorough review, decided to build one. But then the Department of Defense decreed that any such undertaking should be handled by the Defense Communications Agency so that it could be used by all of the service branches. Baran realized that the Agency would never have the desire or the ability to get it done.
So he tried to convince AT&T to supplement its circuit-switched voice network with a packet-switched data network. “They fought it tooth and nail,” he recalled. “They tried all sorts of things to stop it.” They would not even let RAND use the maps of its circuits, so Baran had to use a leaked set. He made several trips to AT&T headquarters in lower Manhattan. On one of them, a senior executive who was an old-fashioned analog engineer looked stunned when Baran explained that his system would mean that data could go back and forth without a dedicated circuit remaining open the whole time. “He looked at his colleagues in the room while his eyeballs rolled up sending a signal of his utter disbelief,” according to Baran. After a pause, the executive said, “Son, here’s how a telephone works,” and proceeded with a patronizing and simplistic description.
When Baran continued to push his seemingly preposterous notion that messages could be chopped up and skedaddle through the net as tiny packets, AT&T invited him and other outsiders to a series of seminars explaining how its system really worked. “It took ninety-four separate speakers to describe the entire system,” Baran marveled. When it was over, the AT&T executives asked Baran, “Now do you see why packet switching wouldn’t work?” To their great disappointment, Baran simply replied, “No.” Once again, AT&T was stymied by the innovator’s dilemma. It balked at considering a whole new type of data network because it was so invested in traditional circuits.57
Baran’s work eventually culminated in eleven volumes of detailed engineering analysis, On Distributed Communications, completed in 1964. He insisted that it not be classified as secret because he realized such a system worked best if the Russians had one as well. Although Bob Taylor read some of it, no one else at ARPA did, so Baran’s idea had little impact until it was brought to the attention of Larry Roberts at the 1967 Gatlinburg conference. When he returned to Washington, Roberts unearthed Baran’s reports, dusted them off, and began to read.
Roberts also got hold of the papers written by Donald Davies’s group in England, which Scantlebury had summarized in Gatlinburg. Davies was the son of a Welsh coal mine clerk who died a few months after his son was born, in 1924. Young Davies was raised in Portsmouth by his mother, who worked for Britain’s General Post Office, which ran the nation’s telephone system. He spent his childhood playing with telephone circuits, then earned degrees in math and physics at Imperial College in London. During the war he worked at Birmingham University creating alloys for nuclear weapons tubes a
s an assistant to Klaus Fuchs, who turned out to be a Soviet spy. He went on to work with Alan Turing building the Automatic Computing Engine, a stored-program computer, at the National Physical Laboratory.
Davies developed two interests: computer time-sharing, which he had learned about during a 1965 visit to MIT, and the use of phone lines for data communications. Combining these ideas in his head, he hit upon the goal of finding a method similar to time-sharing for maximizing the use of communications lines. This led him to the same concepts that Baran had developed about the efficiency of bite-size message units. He also came up with a good old English word for them: packets. In trying to convince the General Post Office to adopt the system, Davies ran into the same problem that Baran had when knocking on the door of AT&T. But they both found a fan in Washington. Larry Roberts not only embraced their ideas; he also adopted the word packet.58
* * *
A third and somewhat more controversial contributor in this mix was Leonard Kleinrock, a joyful, affable, and occasionally self-promoting expert on the flow of data in networks, who became close friends with Larry Roberts when they shared an office as doctoral students at MIT. Kleinrock grew up in New York City in a family of poor immigrants. His interest in electronics was sparked when, at the age of six, he was reading a Superman comic and saw instructions for building a crystal radio with no battery. He pieced together a toilet paper roll, one of his father’s razor blades, some wire, and graphite from a pencil, and then convinced his mother to take him on the subway to lower Manhattan to buy a variable capacitor at an electronics store. The contraption worked, and a lifelong fascination with electronics blossomed. “I still am awed by it,” he recalled of the radio. “It still seems magical.” He began scoring radio tube manuals from surplus stores and scavenging discarded radios from Dumpsters, picking apart their components like a vulture so he could build his own radios.59
The Innovators Page 27