Computing and the National Science Foundation, 1950-2016. William Aspray

and “the capacity of today’s supercomputers is several orders of magnitude too small for problems of current urgency in science, engineering and technology.” The panel recommended a program that would increase access via high bandwidth networks; increase research on computation, software, and algorithms; train personnel; and increase R&D on new supercomputer systems.

      In response to the Lax Report, NSF organized an internal working group¹³⁴ in April 1983 to help the Foundation meet the computing needs of academic science and engineering. NSF also held a workshop in May 1983 with 13 scientists from diverse disciplines to define an initiative in large-scale computing and networking. According to what became known as the Bardon-Curtis Report,¹³⁵ NSF should: (1) coordinate with other federal agencies; (2) increase support for local computing facilities; (3) encourage proposals to provide supercomputer services and access and be prepared to support 10 supercomputer systems within three years; (4) support networks linking laboratory researchers with each other and with supercomputer centers to provide access, file transfer, and scientific communication; (5) support academic research in advanced computer systems design and computational mathematics; and (6) establish an NSF advisory committee for supercomputing.

      In November 1985, the House Committee on Science and Technology held hearings¹³⁶ on supercomputer and network resources for science research. During the hearings, NSF Director Edward Knapp cited the Bardon-Curtis Report and indicated that Edward Hayes, the NSF Controller, was chairing an NSF Task Force on Advanced Scientific Computing. Knapp also indicated that NSF was gathering information from grantees about their immediate needs for access to Class IV¹³⁷ computers and would negotiate with suppliers who could provide appropriate blocks of time. Under this plan, NSF would continue to support research in the theoretical and experimental design of computers as well as on computational mathematics, software, and algorithms. NSF indicated that its networking initiative would be part of Advanced Scientific Computing. Subsequently, the 98th Congress voted $40 million to fund the recommendations of the Bardon-Curtis Report.

      NSF established an Advisory Committee on Supercomputer Access¹³⁸ chaired by Neal Lane, then Chancellor of the University of Colorado at Colorado Springs, that would become the Advisory Committee for the Office of Advanced Scientific Computing. I staffed a subcommittee¹³⁹ on networking options, chaired by Joe Wyatt, the Chancellor at Vanderbilt. At that time, NSF expected that a supercomputing network would be developed in two phases: Phase I using conventional network technology and expanding existing viable networks, and a Phase II using satellite transponder facilities and optical fiber trunks as they became available.

      In December 1983, Landweber wrote to Edward Knapp encouraging him to “proceed as quickly as possible to establish a national Science Net [and] to use existing technologies . . . such as ARPANET, CSNET, and BITNET.”¹⁴⁰ A few weeks later, Jack Schwartz (NYU) wrote¹⁴¹ to Edward Hayes asking him to have the advisory committee consider other needs for high-bandwidth communication beyond access to the supercomputing centers. Even before the Office for Advanced Scientific Computing was established, the community, in particular Landweber and Kenneth Wilson, were pushing for a national network.

      As the CER and CSNET programs grew, Kent Curtis and I had been discussing ARPANET opportunities with DARPA. In mid 1983, Curtis asked Frank Kuo of SRI for advice on expanding or duplicating ARPANET technology to support supercomputer access¹⁴² in a network called “USERNET” that might support 200 academic research institutions or 2000 college and university sites. Kuo pointed out that splitting off MILNET from ARPANET would leave only a 40-node network intended to be a “research and development” testbed. He estimated that developing a 200-site USERNET using ARPANET technology would require $7.5–11.5 million for IMP¹⁴³ hardware and $17 million in operational costs and 10 times that much for 2000 institutions. He also raised the issue of NSF competition with commercial packet networks such as TELENET, TYMNET, UNINET, or NET1000, and suggested that NSF look instead into using commercial networks for a backbone. My thought at the time was that no commercial network supported full network services and there might be alternative “tiered” approaches.

      In April 1984, Kent Curtis and I met with DARPA’s Bob Kahn, who said that the ARPANET, as an R&D network, could only expand by an additional 20 nodes and for $1 million in capital costs and $4.8 million in annual costs. At that time, the Division of Computing Research was already supporting ARPANET sites at RAND and Delaware for CSNET and at a few CER sites.

      NSF created the Office of Advanced Scientific Computing in May 1984, with John Connolly from the Division of Materials Research as Director, Larry Lee from Mathematics as Program Director for Centers, and me (on loan from the Division of Computing Research) as Program Director for Networking. The first awards for time on Class IV supercomputers totaling almost $19 million were made July 1, 1984, to Purdue University, University of Minnesota, and Boeing Computer Services. In 1985, OASC expanded access to existing supercomputer resources, adding centers at AT&T Bell Labs, Colorado State University, and Digital Productions (an early computer animation company).

      An NSF OASC review panel¹⁴⁴ met in February 1985 to consider 22 applicants. On February 25, 1985, NSF announced¹⁴⁵ funding for four National Advanced Scientific Computing Centers: the John von Neumann Center (JVNC) at Princeton University, the San Diego Supercomputer Center (SDSC) at the University of California, San Diego, (managed initially by General Atomic), the National Center for Supercomputing Applications (NCSA) at the University of Illinois, and the Cornell Theory Center. Later, NSF named the Pittsburgh Supercomputing Center (PSC) as a fifth center. (The National Center for Atmospheric Research (NCAR) was sometimes considered a sixth center, but was always dedicated to climate researchers and never a part of the program.) Each of these centers was associated with academic and industry partners and had developed a tentative “customer” base of scientists needing access to high performance computing.

      With responsibility for the centers and a network under OASC, the OASC Networking Advisory Committee recommended the establishment of a “Sciencenet Phase 1”¹⁴⁶ using available and proven technology to implement a network as soon as possible. The preferred strategy was to expand and interconnect existing networks such as ARPANET and BITNET with selective use of commercial network services. By 1985, the Defense Communications Agency had begun to use ARPANET as an operational DoD network following the cancellation in 1983 of the new command, control, communications, and intelligence (C3I) network, AUTODIN II. DoD was looking into splitting off the “research” sites and using the ARPANET (as MILNET) only for military purposes. This action complicated any approach to leveraging ARPANET for supercomputer access and eventually accelerated the growth of NSFNET.

      The planned Sciencenet Phase 1 effort involved the development of Internet protocols, access protocols, and a management strategy for the network. David Farber and Landweber defined a Phase I strategy¹⁴⁷ to quickly enable users of existing networks (ARPANET, BITNET, CSNET, MAILNET, MFENET) to run jobs on supercomputers at the national centers. ARPANET, CSNET/X.25, and MFENET users could remotely log in to supercomputers and run interactive or batch services. BITNET, CSNET/Phonenet, and MAILNET users would have to depend on electronic mail or file transfer for batch submissions. The Landweber-Farber Report also recommended that NSF should (1) add a Sciencenet manager and management team (or contract for such services); (2) establish a working group representing the centers, networks, and NSF management; and (3) establish a permanent Technical Advisory Committee (TAC). Because someone had trademarked “Sciencenet,” the network quickly became known as NSFNET.

      After I left for my IR&D assignment to Berkeley in August 1984, I was still involved remotely with both CSNET and NSFNET. NSF was looking for a replacement in OASC and concurrently UCAR was looking for a permanent CSNET Executive Director. Landweber had met Dennis Jennings, the Director of the BITNET-based European Academic Research Network (EARN) and Computing Center Director at University College Dublin. He encouraged him to apply for the CSNET directorship. Jennings visited NSF and spoke with John Connolly in August 1984. Offered both the CSNET and NSFNET positions, Jennings accepted the NSFNET directorship and began in January 1985. As he recalled, “So when I arrived at the NSF on January 2nd, 1985, the key components were in place: The demand from key researchers; a significant budget for networking—roughly 10% of the supercomputer program budget was devoted to the network;