Center for Science, Technology, and Economic Development (CSTED) > Selected Reports

The Role of NSF's Support of Engineering in Enabling Technological Innovation

IV. THE INTERNET

TECHNOLOGICAL DECOMPOSITION

s The Internet consists of a very large number of different technologies, the absence of any one of which would cause it to cease functioning. In addition to the technologies we have already listed as "defining" technologies, from a functional point of view we must include the physical communication links (satellites, microwave links, copper and fiber optic cables), computer hardware and the operating systems that control them, and support software such as e-mail, addressing systems, telnet, FTP, and so on. However, these latter technologies are not intrinsic to the Internet (defined as a system of interlinked, packet-switched, distributed networks). To bound our analysis, we have had to distinguish between "intrinsic" and "supporting" technologies. Such bounding is essential because resources to pursue the history of each element of an innovation are limited, but more importantly, chronological and intellectual bounds must also be defined. We do not, after all, wish to trace the Internet to Maxwell's equations and Ohm's law.

Intrinsic technologies are defining technologies: they are the unique features of the innovation under study that enable it to function as it does. Put differently, they are the features of the innovation that more or less uniquely account for its socioeconomic impact. Thus, in the absence of user-friendly browsers, the Internet would not have nearly its present level or variety of types of use. In the absence of technological means for seamless interconnection of networks and routing of communication, we would have no Internet. The following table summarizes the distinctions we are making.

Since drawing these boundaries will exclude NSF's contributions to supporting technologies, whatever conclusions we reach about NSF's influence on the Internet will be conservative.

Internetworking

In April 1995, NSFNET reverted to being a network serving the research community, and the larger Internet consisted of a set of interconnected network providers. Between 1986, when NSFNET was created, and this point, when it reverted, NSFNET was the backbone that supported the skeleton of the Internet. At the time of its decommissioning, NSFNET was a T3 system, operating at 44.736 Mbps. It had operated at this level since 1991, when it was upgraded from T1 (1.544 Mbps). In 1988, NSFNET was upgraded to T1 from its initial operating capacity, 56 Kbps, which was its designed capacity when it was created in 1986. These were evolutionary changes that did not depend on significant technological breakthroughs, although each upgrade pushed the envelope of technology.

The basic structure (architecture) of the Internet remained the same during this period: a backbone linking regional networks, which in turn linked campus and other local servers. Apparently, this structure was not the result of a priori design, but rather after-the-fact rationalization of the way NSF was supporting the network in 1986-87: the backbone administered by NSF, the regional networks funded either by NSF or by other institutions, and campus networks funded by the universities they served. The rationalization was offered in a 1987 memo to NSF program manager Steve Wolff from the NSFNET Networking Program Advisory Group (Mandelbaum and Mandelbaum, 1993: 65). [An alternative view is offered by Merit (1995: 11): “NSF staff in the Computer and Information Science and Engineering (CISE) directorate, created on October 1, 1986, envisioned the overall design of the new NSFNET as consisting of a logical topology, or architecture, composed of three ‘tiers’: the national backbone, various regional networks ... and campus networks."] Thus we need to look more closely at 1986 as a transition point in the Internet’s structure.

According to historical accounts and to our interviews, the conception of NSFNET built on two major streams of technology and experience: ARPANET and CSNET. Interestingly, CSNET might be considered a “dead end” in the sense that it was terminated as designed and subsumed by NSFNET. Yet the experience NSF gained from CSNET proved significant for the design and evolution of NSFNET.

It was not preordained that NSF would become responsible for the nation’s high-speed backbone that would, in turn, evolve into the backbone of the Internet. By the early 1980s, other federal agencies had networking projects-obviously the Defense Department but also the Department of Energy and NASA (Jennings, Landweber, Fuchs, and Farber, 1986). But by 1986, NSF had initiated NSFNET, which linked five NSF supercomputer centers and the National Center for Atmospheric Research, and it supported CSNET, an earlier network created to link computer scientists at universities (Merit, 1995: 11).[66] The explosive growth of the Internet began with NSFNET. “When the NCP to TCP transition occurred in 1983, there were only a couple of hundred computers on the network. As of January 1993, there are over 1.3 million computers in the system. There were only a handful of networks back in 1983; now there are over 10,000” (Cerf, 1993). Leonard Kleinrock observes that the growth of the Internet has been exponential during its entire lifetime; it has attracted public attention now that it has reached tens of millions of people (Kleinrock, personal communication, October 12, 1996).

The decision to construct a three-tiered, high-speed national network service and, in 1987, to write and issue the solicitation that created it were fundamental to the future course of computer communication. Although the original architecture of the NSFNET was continued and rationalized, the 1987 solicitation included a number of significant requirements that dramatically shaped the course of the Internet’s institutional and, to a lesser extent, its technological development. We will address these factors in our section on organizational and managerial innovations.

According to Robert Kahn, “NSF involvement began in 1980 with the CSNET. The primary players were Kent Curtis and Larry Landweber. The CSNET originally involved the connection of four supercomputers . ... The CSNET was not a great success on a technical level, but it brought together the people and created the environment which fostered the Internet” (SRI interview with Robert Kahn, April 22, 1996). Vinton Cerf echoed this assessment of CSNET: “The success of NSFNET built on the work of David Farber, Kent Smith, and Larry Landweber on CSNET. CSNET used a variety of communications protocols, primarily X.25 but also TCP/IP. CSNET never worked very well, but it set the focus for future work” (SRI interview with Vinton Cerf, April 22, 1996). Cerf confirmed NSF’s innovative role: “CSNET adopted TCP/IP, but developed a dial-up “Phone-mail" capability for electronic mail exchange among computers that were not on ARPANET, and pioneered the use of TCP/IP over the X.25 protocol standard that emerged from commercial packet switching efforts” (Vinton G. Cerf, http://www.cs.washington.edu/homes/lazowska/cra/networks.html). CSNET helped set the model for NSFNET in yet another sense: the promise of time-limited support, after which the service provider would have to be self-supporting. Supercomputer sites paid to be involved, which was the beginning of NSF’s leveraging tactics that were eventually mastered by Steve Wolff (SRI interview with Peter Ford, April 22, 1996). “This eventually led to the regional networks and commercialization. It made the bidders think about commercialization and how to fund things after NSF closed the purse” (e-mail to SRI from Elizabeth Feinler, February 7, 1996).

Steve Wolff was a program manager at NSF when the first version of NSFNET became operational and was in an excellent position to consider the role that CSNET played. He identified Larry Landweber (University of Wisconsin), Dave Farber (University of Pennsylvania, formerly at the University of Delaware), and Rick Adrion (University of Massachusetts) as the people who “did” CSNET.

Their contribution has never been brought out and has been underestimated by many. It was the first embrace of NSF into the networking business. Owing to a deal Farber struck with Kahn, it was the first case in federal communications history where two agencies shared a common communications facility-that was the chief contribution, showing that more than one organization can achieve the advantages of multiplexing and routing of one agency’s packets through another’s systems. Prior to that time, ARPANET had only ARPANET traffic on it. It was this experience that gave NSF the confidence to get into the NSFNET business and allowed research agencies to start sharing traffic. (SRI interview with Steve Wolff, April 23, 1996)

Lawrence Landweber, a computer scientist at the University of Wisconsin, sought to establish a research computer network that would link university computer science departments. One of the reasons Landweber wanted to establish a new network was that several universities with large computer facilities were not part of the ARPANET. At a 1979 meeting convened by Landweber, representatives from NSF and ARPA and computer scientists discussed the feasibility of such a network. Landweber wrote a proposal to NSF, which, after several revisions, included a provision to link the proposed CSNET with ARPANET. Vinton Cerf not only proposed the link but urged that CSNET employ the TCP/IP protocol, thus making the link transparent. A 1980 CSNET planning group meeting reached several key goals:

• All researchers would have access to CSNET.
• The cost for institutions would be proportional to the volume and level of service.
• CSNET would eventually become financially self-sufficient.

In addition to a series of computer science research grants over the period 1976-81, NSF awarded Landweber a 1980 planning grant, “CSNET-A Computer Science Research Network,” for just under $50,000. Since the CSNET was the first nonmilitary network to link to the ARPANET by using the transparent TCP/IP protocol, some regard this occasion (implemented in 1983) as the birth of the Internet (Comer, 1983).

David Farber is now Alfred S. Fitler Moore Professor of Telecommunication Systems at the University of Pennsylvania. His primary appointment is with the Department of Computer and Information Science; he holds a secondary appointment in the Department of Electrical Engineering. He was Chairman of the CSNET Executive Committee and founding Chairman of the Network Program Advisory Group (NPAG) of the National Science Foundation. While on the faculty at the University of California, Irvine, he received two significant NSF awards: “Design of Distributed Computing Systems ($553,000, 1971) and “The Overseeing of Distributed Processing Computer Systems” ($82,000, 1977). With Dennis Jennings, Larry Landweber, Ira Fuchs, and Richard Adrion, he authored "The National Research Network," Science, February 1986, and "The Convergence of Computing and Telecommunications Systems" (with Paul Baran), invited article, Science, Special Issue on Electronics, 1977.

Protocols

As we noted earlier, the TCP/IP protocol used on the Internet since its creation-and which made the seamless connection among diverse networks possible-was written by Robert Kahn and Vinton Cerf in 1973. At the time, Kahn was director of ARPA’s Information Processing Techniques Office (IPTO) and Cerf was a member of the Stanford faculty, having just received his Ph.D. in computer science from UCLA. Cerf was also a member of the ARPANET Network Working Group and chair of the International Network Working Group (CBI, 1995: 258). The breakthrough incorporated in Kahn and Cerf's protocol solved the problem of routing packets across networks linked by gateways. Previously, there was no way for a packet originating in one network to be addressed in a way that enabled it to enter another network and be routed correctly. The IP packet was embedded in a packet from one network; the gateway between this network and another removed it and reencapsulated it in a packet on the other network for transport. ARPA used TCP/IP on an experimental basis in the ARPANET after 1974, but in 1983 it required that all ARPANET host computers support TCP/IP (just at the time ARPANET and CSNET were linked). Because of the fundamental and lasting contribution that TCP/IP made, many regard Cerf and Kahn as the inventors of the Internet (e.g., David Mills, SRI interview, June 7, 1996).

Unquestionably, TCP/IP was the result of ARPA research support. Kahn described his collaboration with Cerf and the role of others this way:

We actually wrote a paper on that [the design of the Internet architecture] back in 1973, which got published by the IEEE in 1974, that laid out the strategy. Then Vint and some of his students took the lead in actually taking these ideas and turning them into detailed specifications and real protocol implementations. He led the development effort. There were actually three parallel efforts-one by Vint and his students at Stanford, one by some folks at BB&N (Bolt, Beranek & Newman), and the third effort by some folks at University College, London. The three of them then interacted together. Vint was really playing a very major lead role in this whole thing. (CBI, 1989: 16)

According to Kahn, “The NSF had little or no role prior to 1980. The Defense establishment was better suited to develop this technology” (SRI interview with Robert Kahn, April 22, 1996). Cerf agrees: in response to a question about involvement of NSF before 1980 in the development of the Internet, he replied:

None. ARPA was out on a limb. AT&T thought they were nuts. The risk factor was quite high. Not until 1980, when it became obvious that being on the ARPANET was critical for universities, did NSF start involvement. (SRI interview with Vinton Cerf, April 22, 1996)

Robert Kahn received his Ph.D. in electrical engineering from Princeton in 1964, where his doctoral work involved sampling and representation of signals. His dissertation adviser was John Thomas, who was working on sampling theory and nonparametric detection theory at the time. Thomas has been a regular recipient of NSF awards since the time Kahn studied under him; according to the NSF awards database, the year Kahn received his degree, Thomas was awarded a 24-month, $65,000 grant: (A) Sampling Theorem Generalizations and Extensions, (B) Sampling Theorem Limitations.[67] Kahn recalls that he was the recipient of an NSF scholarship or fellowship that covered the bulk of his graduate school costs at Princeton. It was awarded to him personally and enabled him to go to graduate school (e-mail to SRI from Kahn, August 8, 1996). He joined the faculty at MIT after graduating and worked in the Research Lab of Electronics. In 1966, he took a leave of absence to go to Bolt, Beranek & Newman, where he began working on computer networking research funded by ARPA. Until he joined ARPA in 1972, Kahn worked primarily on the development of the ARPANET at BBN (CBI, 1989: 1-6).

Vinton Cerf received his Ph.D. in computer science from UCLA under Gerald Estrin. Most of his work at UCLA was supported by ARPA, as part of the large contracts supervised by Leonard Kleinrock. He then went to Stanford until 1976, when he joined Kahn at ARPA's Information Processing Techniques Office. Although Estrin probably received most of his support from ARPA, it is significant that during and shortly before Cerf became his student, Estrin received a series of substantial research awards from NSF: $188,000 in 1970, $185,000 in 1972, and $144,000 in 1973 for “Modeling and Measurement of Computer Systems.” Probably the “UCLA Mafia” that later contributed in different but significant ways to the development of the ARPANET benefited from Estrin’s NSF grants, although most of the ARPANET-related work was supported by ARPA with Kleinrock as Principal Investigator. As Cerf describes the setting,

At UCLA, the sort of UCLA Mafia sort of consisted of Steve Crocker, who was sort of the ringleader; Jon Postel, who ultimately became the editor of the RFC series and has been ever since; Bob Braden was involved because he was at the computer center. ... Incredibly, they are all still involved in networking. We were permanently infected by packet switching. (CBI, 1990b: 8)

Packet Switching

Like TCP/IP, packet switching-the idea and its realization-was the direct result of Defense Department support. At RAND, Paul Baran was concerned about the vulnerability of military communication systems to nuclear strike, and it was in this context that the idea of a distributed system with redundant paths occurred to him. There was no NSF support for his work at RAND. Baran said in our interview with him that he was influenced by Claude Shannon and Warren Weaver, who were known for their early work in formal information theory in the late 1950s, and by Warren McCullough at MIT and others at the Brain Institute at UCLA. (The brain is the best model for the types of networks he was interested in.) His last work related to network development was a contract from ARPA in 1972, on which he worked with Cerf when Cerf was at Stanford, that explored the feasibility of commercializing part of the ARPANET. Interestingly, the report concluded that the network was commercializable but was ignored (SRI interview with Paul Baran, June 17, 1996).

Baran received a B.S. in electrical engineering from Drexel in 1949, then took a job with Eckert-Mauchly Computer Company, which later became Univac. After several jobs, he joined Hughes Aircraft and at the same time enrolled in an M.S. program at UCLA, with Gerald Estrin as his adviser. “He kept me continually challenged,” and convinced Baran to enroll in the Ph.D. program, but Baran never completed the degree. He joined RAND in 1959 in the Computer Sciences Department of the Mathematics Division. At the time, RAND worked almost exclusively for the Air Force (CBI, 1990a: 9-11).

As we noted earlier, Baran’s ideas were embodied in the ARPANET beginning in 1969, largely because of the influence they had on Lawrence Roberts and the conceptual design of the ARPANET (CBI, 1995: 233).[68] Roberts was hired by ARPA in 1966 to work on the problem of computer networking, and later became director of IPTO from 1969 to 1973, just as the ARPANET was created and implemented. Before joining ARPA, Roberts worked at MIT’s Lincoln Laboratory between 1963 and 1966. Roberts received his Ph.D. from MIT under Peter Elias, but no NSF awards were made to Elias during or before Roberts’ graduate work there. Roberts was a graduate student with Leonard Kleinrock, who, as we have seen, was the Principal Investigator on ARPA’s project to set up the first node of the four-node network that began the ARPANET in 1969.

Leonard Kleinrock received his Ph.D. in electrical engineering from MIT in 1963, where he was a student of Ed Arthur; Claude Shannon was on his dissertation committee. His doctoral work was on analytical models of computer communication networks. Kleinrock joined the faculty at UCLA in 1963, where he is now a professor in the computer science department. Kleinrock characterized the atmosphere at MIT in the late 1950s and early 1960s this way: “That MIT environment was crucial; it made you do good things” (SRI interview with Leonard Kleinrock, June 17, 1996). A check of the NSF awards database shows no awards to Arthur in the early 1960s, or to Kleinrock then or subsequently.

Routers

Routers (sometimes referred to as gateways or switches) are the combination of hardware and software that operate at nodal points in networks and switch paths of packets in accordance with their addresses and with levels of congestion on the connecting lines. As with most Internet hardware and software, routers were developed and evolved along with packet switching and internetwork protocols. During the ARPANET days, routers were custom made on a case-by-case basis, depending on the characteristics of the networks involved. No commercial routers were available until the mid-1980s. Bolt, Beranek & Newman developed the first routers for ARPANET but did not believe there was a commercial market for them (Cerf, 1993). As far as router technology was concerned, the challenge of NSFNET, in the mid-1980s, differed substantially from that posed by ARPANET. With ARPANET, the source and recipient were known (technologically speaking) and under ARPA’s control, but with NSFNET this was not the case. The technical complexity of routing was thus increased substantially under the open system of NSFNET.

David Mills of the University of Delaware met the challenge by designing the “fuzzball” router, acknowledged to be a significant advance in router technology, one that formed the basis for the new, more complex and uncontrolled set of networks comprising the Internet. Mills’ routers, used on the original 56-Kbps NSFNET backbone, consisted of custom-designed software that was continually changed during the first several years of operation (Mandelbaum and Mandelbaum, 1993: 68-69). Steve Wolff, the NSFNET program manager, recalls Mills’ contribution like this:

The unsung hero was Dave Mills. He single-handedly kept the NSF backbone up. Dave did fine tuning and introduced things in use by every router vendor today. (SRI interview with Steve Wolff, April 23, 1996).

Mills himself describes his contribution:

The issue had to do with whether [routing] would be provided for a single system or would be a shared technology where many would contribute. The dichotomy was apparent very early. BBN contracted for ARPANET; I thought there would be other players and built a gateway that would work with the BBN. This was done to get others involved. Routing architecture had two tiers. One was a local network (e.g., a campus), which would be a uniform internal network. Then there was the issue of how to connect and let a network talk to networks (this is the early 1980s); this was my contribution. The new player on the block, NSF, wanted to expand the Internet dramatically. That’s when fuzzballs really came in. They were the routing infrastructure for the NSF backbone. (SRI interview with David Mills, June 7, 1996)

Mills observed that NSF had supported him with grants “for years and years,” including support for travel to NSF and talks to program managers. Mills was a member of the Network Technology Advisory Group, chaired the Architecture Task Force, and was a member of the Internet Activities Board. NSF provided travel support for all of these (SRI interview with David Mills, June 7, 1996).

Mills received his Ph.D. from the University of Michigan in 1971, was on the faculty at the University of Maryland from 1971 to 1976, worked at Communication Satellite Corporation from 1976 to 1986, and has been in the computer science department at the University of Delaware since 1986. He said that ARPA has paid for his education since the 1960s, but since 1982 NSF has provided about a third of his research support. The NSF awards database shows a $29,000 grant to Mills in 1986 for “Support of the NSF Supercomputer Network Program” and a 1989 award for “Support of the National Science Foundation Networking Program” in the amount of $103,155. Presumably, Mills’ previous research support from NSF was provided indirectly through other principal investigators with large awards.

Internet Browsers

In mid-1995, when the NSFNET was decommissioned, Netscape was the most popular Internet browser. Netscape Communications, Inc., was founded by the former chairman of Silicon Graphics, Inc., James Clark, who joined forces with Marc Andreessen, who headed the team at the University of Illinois that created Netscape’s predecessor, Mosaic. Andreessen was working at the National Center for Supercomputer Applications, an NSF-funded facility[69] whose purpose was to train researchers on the use of supercomputers by having them come to the University of Illinois and use the supercomputer (Mandelbaum and Mandelbaum, 1993: 60). At the time he created Mosaic (1992-93), Andreessen was an undergraduate at the University of Illinois. He graduated in 1993 with a bachelor’s degree in computer science.

NCSA was established in 1985 with a grant from the National Science Foundation. In addition to NSF grant funds, NCSA now receives major funding from the Advanced Research Projects Agency, NASA, corporate partners, the state of Illinois, and the University of Illinois.

NSF created the Computer and Information Science and Engineering (CISE) directorate in 1986 to integrate relevant programs throughout the Foundation. Among the activities supported was Advanced Scientific Computing, which included funding for the National Supercomputer Centers at a level of $35 million for FY 1986. The Supercomputer Centers program was "the focal point for participation in the Foundation-wide Computational Science and Engineering thrust to develop innovative software, numerical methods and graphical techniques to meet the needs of the scientific and engineering communities. Included are user-friendly and transparent interfaces between the user and the computer, transportable computer programs and more efficient computational methods" (National Science Foundation, FY 1988 Budget to the Congress, p. 128).[70]

The Illinois Supercomputer Center was funded at $7 million for FY 1986 and $8.5 million for FY 1987. Andreessen was a junior majoring in computer science and was working as an undergraduate research assistant in the Center during 1992 and 1993. He was not working directly on the project that was developing some of the ideas that would be embodied in Mosaic; that project was intended for use on supercomputers rather than for more widespread use on the Internet. However, he was "just down the hall" and realized the implications of what was being developed for Internet use, and developed Mosaic (SRI interview with Robert Kahn, April 22, 1996).

When asked by an interviewer how he began the task of creating Mosaic, Andreessen replied,

I guess you’ve got to understand the context of what NCSA is: it’s basically a Federally-funded research center. When I was there, it had been around for roughly eight years or so and, at that point, it had a very large established budget-many millions of dollars a year-and a fairly large staff and, frankly, not enough to do. ... NCSA is a fairly free-form environment, and we certainly had the resources, the money, the machines and the network to be able to do interesting things. When it occurred to Eric Bina and myself that creating a graphical Web client would be an interesting thing to do, we were in an environment where we were able to go do that. ... NCSA is not a place where there are necessarily a whole lot of well-defined directions or goals. A lot of the interesting things that have happened there-in fact, most of the interesting things that have happened there-have been because one or more people decided to do something interesting, and then did it. (Marc Andreessen interview with LAN Times, Thom Stark Home Page, http://www.dnai.com/~thomst/)

Elizabeth Feinler, who was involved with the development of ARPANET and the Internet at SRI, observes that NSF's support of Gopher and Mosaic at the University of Illinois yielded more than just Netscape.

I would say that CERN and NSF laid most of the foundation for the colossal growth of the Internet because of the spread of the web (hypertext) technology. The U. of Ill. did lead to the Netscape spin-off but there is lots more than Netscape out there. The web technology has spawned a whole information revolution and industry. NSF is supporting development, libraries, and an approach which again is a govt-consortium, seed approach. This example might be NSF’s finest hour. (E-mail from Elizabeth Feinler, February 7, 1996).