by Matthew Lyon
Just before the meeting ended, Wes Clark passed a note up to Roberts. It read, “You’ve got the network inside out.” Roberts was intrigued, and he wanted to hear more; however, the meeting was breaking up, and people were already leaving. Roberts, Taylor and a few others huddled around Clark afterward, and a small group decided to continue the discussion during the ride back to the airport. In the car, Clark sketched out his idea: Leave the host computers out of it as much as possible and instead insert a small computer between each host computer and the network of transmission lines. (This was, by coincidence, precisely what Davies had concluded separately in England.)
The way Clark explained it, the solution was obvious: a subnetwork with small, identical nodes, all interconnected. The idea solved several problems. It placed far fewer demands on all the host computers and correspondingly fewer demands on the people in charge of them. The smaller computers composing this inner network would all speak the same language, of course, and they, not the host computers, would be in charge of all the routing. Furthermore, the host computers would have to adjust their language just once—to speak to the subnet. Not only did Clark’s idea make good sense technically, it was an administrative solution as well. ARPA could have the entire network under its direct control and not worry much about the characteristics of each host. Moreover, providing each site with its own identical computer would lend uniformity to the experiment.
The most curious thing about the idea was that Clark thought it up. He hadn’t been paying much attention to the proceedings in Ann Arbor. In fact, he was a bit bored by it all. He had already told Roberts in no uncertain terms that he had no desire to put his computer at Washington University in St. Louis on the network. Clark was not friendly toward time-sharing, or even resource-sharing. He had been working on computers designed for individual use and saw no particular reason to share his facility with people on a network. But when he heard the discord over how the ARPA experiment should be deployed, he couldn’t help but hazard a suggestion. Perhaps it was Clark’s antipathy toward time-sharing that enabled him to think of this. By assigning the task of routing to the host computers, Roberts and others were essentially adding another time-sharing function. Clark’s idea was to spare the hosts that extra burden and build a network of identical, nonshared computers dedicated to routing.
During the ride to the airport, the discussion turned lively. Wouldn’t an entire subnetwork composed of individual computers be prohibitively expensive and defeat the original goal of saving money? And who, Roberts wanted to know, did Wes Clark think could build such a thing? “There’s only one person in the country who can do that,” Clark responded. “Frank Heart.”
Larry Roberts knew Frank Heart. The two had worked together at Lincoln Lab, and Roberts had shared an office with Heart’s wife, Jane, a programmer at Lincoln. Roberts and Heart had never worked together directly, but Roberts knew Heart to be an exacting systems engineer. He was an expert in real-time systems built for when the physical world demands a response within fractions of seconds—or at least before the next set of data arrives. Anything dealing with incoming information in a time-critical fashion, such as radar tracking data sent to the SAGE system, and seismic information generated during an earthquake, was considered a real-time system, and in the 1960s few people understood real-time systems as well as Heart did.
Roberts also knew that Heart and Clark were good friends from Lincoln, where Heart had shown Clark the ropes of programming more than a decade earlier. Now, as far as Roberts knew, Heart was at Bolt Beranek and Newman in Cambridge, where he had moved in 1966 to work on the use of computers in medicine.
The ARPA network wasn’t intended as a real-time system, not in the same sense of the word that true real-timers understand it. (Anything that takes more than 10 to 20 milliseconds, the point at which delays become humanly perceptible, is not considered real-time.) Strictly speaking, the ARPA network was to be a store-and-forward system. But data would zip in and out of the nodes so quickly, and the response time from a human perspective would be so rapid, that it qualified as a real-time problem. The system would have to cope with dozens of problems involving closely sequenced events and extremely tight timing. The status of the network would change constantly, and whoever programmed the computers that composed Clark’s proposed subnet would need to know how to make the system handle incoming and outgoing data reliably at very fast rates.
Despite the logic of Clark’s recommendation, however, Roberts couldn’t simply turn the job over to Heart. ARPA had to play by the government’s contracting rules. Over the years, most proposals requesting funding arrived at ARPA unsolicited. Seldom had the agency actually requested proposals. But this one was different. The agency had come up with the network idea internally, and in that respect it was unusual. Also, because the network would be government property controlled centrally by ARPA and wouldn’t reside on one campus, say, or at one research firm, Roberts and others decided they had to send this project out for competitive bids.
When he returned to Washington, Roberts wrote a memorandum describing Clark’s idea and distributed it to Kleinrock and others. He called the intermediate computers that would control the network “interface message processors,” or IMPs, which he pronounced “imps.” They were to perform the functions of interconnecting the network, sending and receiving data, checking for errors, retransmitting in the event of errors, routing data, and verifying that messages arrived at their intended destinations. A protocol would be established for defining just how the IMPs should communicate with host computers. After word of Clark’s idea spread, the initial hostility toward the network diminished a bit. For one thing, a separate computer to carry out the switching functions removed the burdensome kludge (a kludge is an inelegant solution to a technical problem) that could result from adding those functions to the host computer. People also saw it as a way of getting another computer to play with.
At the end of 1967 another computer conference, this one in Gatlinburg, Tennessee, helped advance the network plan. This symposium was sponsored by the Association for Computing Machinery, the oldest and most prestigious of professional organizations for the growing computer industry. Though small in number, the attendees represented the highest levels of the computer science establishment.
Gatlinburg was a perfect venue for Roberts to present his first paper on what he called the “ARPA net.” In his presentation, Roberts focused on the reasons for the network and described the subnet of IMPs, but said little else about how the network would actually work. One big puzzle still to solve was the question of how the data would actually be transmitted—over what kind of channel. Ever mindful of cost, Roberts closed his presentation with a brief discussion of what he called “communication needs.” He was thinking of using the same type of telephone lines that he and Marill had used for their small TX-2 experiment: full-duplex, four-wire lines. Talk of the matter ended on a note of frustration. Ordinary dial-up phone lines (as opposed to dedicated, leased lines) were slow, and holding a line open was wasteful. Roberts still hadn’t found a highly efficient means of carrying the data.
Whereas the Ann Arbor meeting months earlier had been the intellectual equivalent of a barroom brawl, Gatlinburg was high tea. People were politely coming around to the idea of a network. Roberts’s presentation was generally well received, even greeted enthusiastically by some.
Another paper was presented by Roger Scantlebury. It came from Donald Davies’ team at the National Physical Laboratory and discussed the work going on in England. His paper presented a detailed design study for a packet-switched network. It was the first Roberts had heard of it. Afterward, Roberts and a few others approached Scantlebury, and they began to discuss the NPL work. The discussion continued at the hotel bar and went late into the night. Scantlebury raised the issue of line speed with Roberts. He said that he and Davies were planning to use lines that operated much faster than the 2,000 bits per second speed Roberts was proposing. He suggested Rob
erts build the ARPA network with fewer lines carrying data at speeds more than twenty times higher to improve the response time.
Roberts also learned from Scantlebury, for the first time, of the work that had been done by Paul Baran at RAND a few years earlier. When Roberts returned to Washington, he found the RAND reports, which had actually been collecting dust in the Information Processing Techniques Office for months, and studied them. Roberts was designing this experimental network not with survivable communications as his main—or even secondary—concern. Nuclear war scenarios, and command and control issues, weren’t high on Roberts’s agenda. But Baran’s insights into data communications intrigued him nonetheless, and in early 1968 he met with Baran. After that, Baran became something of an informal consultant to the group Roberts assembled to design the network. The Gatlinburg paper presented by Scantlebury on behalf of the British effort was clearly an influence, too. When he visited Roberts during the design of the ARPA network, Davies said, “I saw that our paper had been used so much that its pages were falling apart.”
Roberts thought the network should start out with four sites—UCLA, SRI, the University of Utah, and the University of California at Santa Barbara—and eventually grow to around nineteen. UCLA was chosen as the first site because Len Kleinrock’s Network Measurement Center was there. At each of the other sites, ARPA-sponsored research that would provide valuable resources to the network was already under way. Researchers at UCSB were working on interactive graphics. Utah researchers were also doing a lot of graphics work as well as investigating night vision for the military. Dave Evans, who, with Ivan Sutherland, later started Evans and Sutherland, a pioneering graphics company, was at Utah putting together a system that would take images and manipulate them with a computer. Evans and his group were also interested in whether the network could be used for more than just textual exchanges.
Stanford Research Institute (later it severed its ties to Stanford and became just SRI) had been chosen as one of the first sites because Doug Engelbart, a scientist of extraordinary vision, worked there. Several years earlier, when Bob Taylor was at NASA he had funded Engelbart’s invention of the first computer mouse (Engelbart received a patent for the device as an “X-Y position indicator for a display system”), and for years afterward Taylor pointed with pride to his support of Engelbart’s mouse.
Engelbart had been in attendance at the 1967 Ann Arbor meeting of ARPA’s principal investigators when Taylor and Larry Roberts announced that a dozen or so of them would be expected to tie their computers together over an experimental network and that each site would be expected to make its computer resources available on the network. While others had responded skeptically to the plan, Engelbart had been delighted with it. At the time, he was directing an SRI computer research lab. Not unlike Licklider, Engelbart was interested in using computers to augment human intellect. Under a contract from ARPA, he was developing a system (called NLS, for oNLine System) that depended on computer-literate communities. He saw the ARPA experimental network as an excellent vehicle for extending NLS to a wide area of distributed collaboration. “I realized there was a ready-made computer community,” Engelbart recalled. “It was just the thing I was looking for.”
Part of the strength of NLS was its usefulness in creating digital libraries and in storing and retrieving electronic documents. Engelbart also saw NLS as a natural way to support an information clearinghouse for the ARPA network. After all, if people were going to share resources, it was important to let everyone know what was available. At the Michigan meeting, Engelbart volunteered to put together the Network Information Center, which came to be known as the NIC (pronounced “nick”). Engelbart also knew that his research group back home in Menlo Park would be equally enthusiastic about the network. His colleagues were talented programmers who would recognize an interesting project when they saw it.
The conversation with Scantlebury had clarified several points for Roberts. The Briton’s comments about packet-switching in particular helped steer Roberts closer to a detailed design. In specifying the network requirements, Roberts was guided by a few basic principles. First, the IMP subnet was to function as a communications system whose essential task was to transfer bits reliably from a source location to a specified destination. Next, the average transit time through the subnet should be less than half a second. Third, the subnet must be able to operate autonomously. Computers of that era typically required several hours per week of maintenance downtime. IMPs could not afford to be dependent on a local host computer or host-site personnel; they should be able to continue operating and routing network traffic whether or not a host was running. The subnetwork also had to continue functioning when individual IMPs were down for service. This idea that maintaining reliability should be incumbent on the subnetwork, not the hosts, was a key principle. Roberts and others believed the IMPs should also attend to such tasks as route selection and acknowledgment of receipt.
By the end of July, 1968, Roberts had finished drafting the request for proposals. He sent it out to 140 companies interested in building the Interface Message Processor. The document was thick with details of what the network should look like and what the IMPs would be expected to do. It was a rich piece of technical prose, filled with an eclectic mix of ideas. Kleinrock had influenced Roberts’s earliest thoughts about the theoretical possibilities. Baran had contributed to the intellectual foundation on which the technical concept was based, and Roberts’s dynamic routing scheme gave an extra nod to Baran’s work; Roberts had adopted Davies’ term “packet” and incorporated his and Scantlebury’s higher line speeds; Clark’s subnet idea was a stroke of technical genius. “The process of technological development is like building a cathedral,” remarked Baran years later. “Over the course of several hundred years new people come along and each lays down a block on top of the old foundations, each saying, ‘I built a cathedral.’Next month another block is placed atop the previous one. Then comes along an historian who asks, ‘Well, who built the cathedral?’ Peter added some stones here, and Paul added a few more. If you are not careful, you can con yourself into believing that you did the most important part. But the reality is that each contribution has to follow onto previous work. Everything is tied to everything else.”
But in 1968 the network’s principal architect was Larry Roberts: He made the initial decisions, and he established the parameters and operational specifications. Although he would get input from others, Roberts would be the one to decide who built it.
The first responses to the request for proposals were from IBM and Control Data Corporation (CDC). IBM was then the world’s largest computer manufacturer and dominated the market for large computer systems. CDC, though dwarfed by IBM, was another company that had invested heavily in developing large systems. Both declined to bid, and their reasons were identical: The network could never be built, they said flatly, because there existed no computers small enough to make it cost-effective. For the IMP, IBM had thought about proposing a 360 Model 50 computer, a large mainframe. But at a price many times that of a minicomputer, the Model 50 was too expensive to consider buying in large quantities.
Roberts, on the other hand, was thinking small. The first computer he had thought of was the PDP-8, a minicomputer made by Digital Equipment Corp. Digital had released the PDP-8 in 1965. Not only was it the company’s first big hit but the PDP-9 also established minicomputers as the new vanguard of the computer industry. Roberts knew Ken Olsen from Lincoln, and he thought Digital might even offer a quantity discount on the machine.
When bids started coming in, the majority had chosen a Honeywell computer instead. It was a minicomputer called the DDP-516; Honeywell had just introduced it. Part of the new machine’s cachet was that it could be built to heavy-duty specifications. In its “ruggedized” version, it cost about $80,000. Shortly after the machine’s introduction, at a computer conference in Las Vegas, the hardened military version was hoisted off the showroom floor by a crane. As it swun
g from ropes attached to the crane, a Honeywell employee took a sledgehammer to it. The point of the exercise was to demonstrate that the machine was tough enough to operate on a battlefield. For bidders, the more likely appeal of the 516 was its impressive cost-performance ratio and the design of its input/output system.
More than a dozen bids were submitted, resulting in a six-foot stack of paper. Marill’s company, CCA, bid jointly with Digital. Raytheon bid, and so did Bunker-Ramo. Roberts was pleasantly surprised that several of the respondents believed they could construct a network that performed faster than the goal listed in the specifications.
Raytheon was a frontrunner. A major defense contractor in the Boston area specializing in electronic systems components, Raytheon had already proposed to build a high-speed, short-distance computer network. In the middle of December, Roberts entered into final negotiations with Raytheon for the IMP contract. Raytheon officials answered ARPA’s remaining technical questions and accepted the price.
So it surprised everyone when, just a few days before Christmas, ARPA announced that the contract to build the Interface Message Processors that would reside at the core of its experimental network was being awarded to Bolt Beranek and Newman, a small consulting firm in Cambridge, Massachusetts.
3
The Third University
When Richard Bolt and Leo Beranek started their consulting company in 1948, advanced computing was not on their minds. Beranek was an electrical engineer, Bolt an architect and physicist. Both were acousticians and members of the MIT faculty during the 1940s. Bolt had worked for the Navy in World War II on methods for using sound to detect submarines. Following the war, as head of MIT’s acoustics laboratory, Bolt did consulting work, as did Beranek. MIT began receiving requests for aid in acoustical planning for new buildings around the country and passed them on to Bolt and Beranek. Independently of each other, the two had already done quite a bit of work in what is known as airborne acoustics—the sound carried in concert halls and movie theaters—as well as in noise control and noise reduction in buildings.