Monday morning, I felt great despite my 12-hour jet lag and 26-hour trip. This was my last stop, my 56th city since INTEROP 91 Fall, less than 6 months ago, and I was certainly looking forward to terminating this little journey.

I had come to Madison to learn how the Internet had gone from a DoD-centered ARPANET to a collection of autonomous systems in the early 1980s. To learn this ancient lore, I had come to meet Larry Landweber, one of the key figures in the early development of the Internet.

In India, some of us had been sitting around over a beer speculating on the topic of Larry's age. He was a fairly prominent theoretician before his entry into networking, and the consensus was that Lah-Ree (as his name is pronounced in much of the world) must easily be in his 60s and close to retirement.

When I mentioned this little anecdote to Larry, he quickly established that he was well under 50 and had no immediate plans to retire. It's always good to start an interview by putting your subject in a pleasant, relaxed state of mind.

We walked down to the computer room in the Computer Science department for a quick tour. Computer Science at Wisconsin is large, with over 35 faculty, and extremely well equipped. The machine room is huge and includes such sports computers as a Thinking Machines CM5 and a now ancient 32-node HyperCube, each node equipped with a 300 Mbyte disk drive.

As we walked past old VAX 11/750s, kept in place so the university auditors didn't reclaim machine room space, we stumbled across two brand new Silicon Graphics computers. Larry called over the computer room manager, who was busy trying to get some equipment packed and shipped back to the manufacturer.

"Whose SGI machines are those?" Larry asked.

"Those are yours, Larry," the manager said with a patient smile.

At the corporate offices of Me, Inc., the arrival of a new RS-232 cable is enough to stop work for a celebration, so I couldn't imagine having so much equipment you forget about two Silicon Graphics machines.

Outside the machine room, Larry logged onto a little terminal next to a coke machine and typed in his request for a Diet Coke, which soon came slamming out of the slot. I commented on how clever it was to rig up the machine to a host.

Turned out the reason for this automated soda server was fairly practical. The campus administration had passed some sort of rule that would have removed the machine from its convenient location outside the machine room. By hooking it up to the network, a mere soda server became a full-fledged example of computer science research. The machine stayed.

Back upstairs, we sat down for a history lesson. Larry's first exposure to networking came in the summer of 1977 when he was sitting in a bar with a few other theoreticians lamenting the lack of communications among themselves. They dreamed up the idea of TheoryNet for exchanging e-mail and, luckily, an NSF official was sitting with them.

"Give me a proposal," he said.

Next thing he knew, Larry had a grant for U.S. $136,000, which for a theoretician was a lot of money, far surpassing any previous grant he had received. Soon, theoreticians all over the country were happily banging away on their Texas Instruments paper terminals, exchanging missives on NP completeness and computational complexity.

In 1977, Larry became chair of the department and started looking around for ways to improve computing facilities. He began casting covetous glances at the ARPANET, but Wisconsin didn't have nearly enough defense contracts to justify its inclusion in the elite dozen or so nodes.

Wisconsin certainly wasn't the only school not on the ARPANET, and Larry started looking around for ways to get the have-nots on the network. He was primarily interested in Computer Science departments, a logical focus given the level of penetration of networks at the time.

Kent Curtis was director of the Computer Science section of NSF and was quickly sold on the idea of a CSNET. A small group of computer scientists was convened in May 1979 to discuss the idea. First on the agenda was to see if maybe they could just get into the ARPANET. While the ARPANET had plans to grow, it wasn't clear that it would grow so fast that it would include them. Just in case, Kent Curtis encouraged the group to submit a proposal.

In November 1979 the proposal for CSNET was submitted and went out for review. The results were disastrous. Reviewers thought TCP/IP was far too complex for end-users, networking should be left to the ARPANET professionals, and who did these amateurs think they were anyway?

Not exactly a vote of confidence. Kent Curtis looked at the reviews, thought about it, and sent Larry back with U.S. $49,000 to make a study. By June 1981, a second proposal for CSNET was submitted and the reviews were again negative.

Kent Curtis looked at the reviews, and called an advisory panel together. They didn't like the idea, either. Curtis kept on supporting the idea though. What made CSNET possible was the skillful lobbying by Kent Curtis inside the government and the cooperation of DARPA. Vinton Cerf had attended a CSNET planning meeting and indicated that DARPA looked on the pro}ect favorably and would be willing to work with CSNET on ways to interconnect. January 1980, after some fairly intense lobbying, NSF awarded U.S. $5 million for five years to the CSNET project.

Armed with money and the blessings of NSF and DARPA, a group of five people was put together to manage the project. A classic exercise in distributed management, CSNET was run by Larry Landweber, Peter Denning, Dave Farber, and Tony Hearn, along with Bill Kern, the NSF project manager. The project began the task of putting together a comparatively low-cost solution that would allow a sharp increase in the size of the Internet.

One of the immediate effects of the NSF award was to get Madison and the other key development sites onto the ARPANET. These sites –Madison, Purdue, RAND, and Delaware–formed the core of CSNET. For the rest of the CSNET world, there were three levels of connectivity: exchanging mail, using a service host, and full TCP/IP connectivity.

For exchanging mail, CSNET used a technique known as PhoneNet. PhoneNet used the MMDF software developed by Dave Crocker and Dave Farber to transfer messages over telephone lines, much in the same way WCP did.

There was a big difference from UUCP, however. Mail which arrived at the PhoneNet relays at RAND and the University of Delaware could continue on into the ARPANET. DARPA had agreed that the CSNET sites, a group of hosts outside of their direct administrative control, could exchange mail with ARPANET sites. This was thus the first example of two autonomous systems connected together with gateways. Although the initial exchanges were e-mail only, this was soon expanded to IP connectivity.

The agreement between ARPANET and CSNET to allow connection was a carefully worded agreement, prohibiting sites in the CSNET from forwarding data originating from the ARPANET to "non-authorized users." Most importantly, though, the agreement delegated authority to the CSNET to administer its own network, ordy requiring that the members of CSNET abide by the rules of acceptable use of the network.

The issue of charging was also carefully dealt with, specifying that there would be no charges between the autonomous systems: you accept my traffic and I'll accept yours. Lack of settlement procedures has remained a tradition in the Internet even today. The Commercial Internet Exchange, for example, uses a system of flat fees based on the size of pipes, not on the amount or attributes of the data flowing through those pipes.

PhoneNet allowed message exchange within CSNET and to the ARPANET, but didn't give people access to services like Telnet and FTP. To provide this higher level of connectivity, CSNET provided a service host at Madison.

Users were given individual accounts and used X.25 or dialin to log onto the service host. From there, they had full access to the ARPANET. The service host also held a collection of useful documents and some nameserver software to help users find electronic mail addresses.

PhoneNet and the service host caught on like wildfire. By December 1981, the RAND and Delaware relays were up and running and Purdue, Princeton, and the NSF were PhoneNet clients. By the summer of 1982, 24 sites were on PhoneNet and two years later there were 84 sites. The population of PhoneNet peaked in 1988 at 170 sites.

Getting sites fully integrated into the ARPANET with IP access was a bit tougher. At the time, X.25 was viewed as the way to set up widely dispersed hosts and the common carriers were telling people that in the future, X.25 networks would become cheap and leased lines expensive.

There were two well-known public data networks at the time, Tymnet and Telenet. Tymnet didn't return phone calls from CSNET, so Telenet became the candidate network for interconnecting CSNET hosts at the IP level. (Note the difference between the Telenet X.25 network and the Telnet virtual terminal service.)

All the X.25 networks at the time were optimized for the holy X trinity, X.3, X.28, and X.29. These standards defined how an asynchronous terminal could connect to a Packet Assembler/Disassembler, which in turn set up virtual circuits to hosts on the X.25 network.

Optimizing for terminals made it tough on hosts. Telenet used 128-byte packets and a virtual circuit could only have two packets outstanding. This certainly slowed down potential throughput, but to make matters worse, the data link level of LAP-B enforced a limit of 7 outstanding packets for all virtual circuits sharing a link.

What this meant was that a host with a 9,600 bps link to Telenet could only get 1 to 2 kilobits of throughput on one virtual circuit. This was a best-case scenario, and during peak periods congestion reduced the throughput even further.

Douglas Comer of Purdue took on the daunting task of trying to integrate X.25 and TCP/IP. Comer and his group developed a solution using multiple X.25 circuits, one of the first examples of path splitting. Comer's group also developed the software that set up and tore down virtual circuits in a manner that was transparent to IP.

Comer's group mapped datagrams over virtual circuits. The approach of layering connectionless traffic over connection-oriented circuits was considered controversial at the time. Another group at University College London took a different approach, using protocol translation to turn TCP circuits into X.25 circuits. Comer's group went off and hacked, and came back with a working IP over X.25 demonstration between Madison and Purdue.

Even with this clever software, X.25-based networking never really caught on in the U.S. Not only was performance poor, but the cost for running TCP/IP over X.25 could be astronomical. Remember, these networks charged by the packet and a single character typed for a Telnet session could consume 4 packets (send character, acknowledge, echo character, acknowledge).

Even though the cost was high, a few sites wanted IP access badly enough that they were willing to pay. After Purdue built and tested the software, Rice was the first X.25 customer in the fall of 1984, and was soon joined by DEC's Western Research Laboratory, and BBN. Later, as it became clear that leased lines were more economical, the Purdue group developed a low-cost leased line network called Cypress, which was offered as an alternative access path in CSNET.

BBN also became the site for the new Coordination and Information Center (CIC) and became the support center for CSNET. The service host and MMDF relays were moved, and the CIC maintained the document database and helped new users get started.

Although NSF funded CSNET, they also made sure that a clause was inserted in the project plan that the network should be self sufficient at the end of five years. At the end of five years, it was in fact self sufficient, charging universities U.S. $2,000 to $5,000 per year and industrial sites such as DEC and IBM $30,000 to participate. Self-sufficiency of CSNET was used by the NSF as an important justification for funding future projects such as the NSFNET.

In addition to self-sufficiency, CSNET helped establish a tradition in the U.S. of projects of limited scope. Instead of expanding CSNET to be a backbone, to serve other communities, and to achieve world peace, CSNET remained focused on one problem. Over time, other solutions such as regional networks became more attractive and by 1988 the number of CSNET hosts was beginning to decline. By 1989, CSNET had merged with BITnet and by 1991, PhoneNet was dead.

CSNET was not just limited to the U.S. It formed an important way, along with BITNET, to spread the Internet overseas. The international spread of the Internet was directly linked to an annual series of meetings that came to be known as the Landweber Seminars. I was interested to learn that Larry was not even present at the first Landweber Seminar.

In September 1982, Peter Kirstein of University College London convened a "meeting on U.S.-European Academic Network Collaboration." The meeting was small, but included representatives from a half-dozen countries.

At this meeting, Germany described their DFN project, the Scandinavians described active projects in Norway and Sweden, and the U.K. described the Coloured Book. There was a CSNET presentation, and CERN described the early stages of what would become HEPnet.

In other words, this was a collection of people who were actually doing something about getting together to exchange information. One of the most useful functions of such a meeting was what occurred in the halls and at dinner where synergy between different groups quickly led to new projects.

In 1983, Kirstein and Landweber jointly organized another seminar in Oslo at which Larry announced that CSNET was going to go international. Needless to say, the question of where it would go international first occupied much of the discussions at the breaks.

In February 1984, Israel became the first international member of CSNET. A carefully worded agreement was signed stating that data originating from the ARPANET would not be forwarded. International access, after all, had the potential to become a political issue.

The Israel PhoneNet connection was rapidly followed by Korea, with Professor Kilnam Chon aggressively pursuing connectivity. By September 1986, PhoneNet access had spread to Australia, Canada, France, Germany, and Japan. Israel, Korea, and the Japanese NTT soon got IP access.

The Landweber Seminars became firmly established as a place for people to meet and things like CSNET connectivity, gateways, and other projects to get started. In 1984, the Seminar convened in Paris, followed in subsequent years by Stockholm, Dublin, Princeton, Jerusalem, and Sydney.

Princeton, in 1987, was the breakthrough meeting. The invitation-only seminar had reached 100 people and the attendance list was certainly impressive, including most of the people active in spreading the Internet.

By Sydney, the meeting had grown so big that the seminar format began to outlive its usefulness. It was decided that year that the seminar should metamorphosize into a conference, and the next year INET 91 was held in Copenhagen. At INET 92, in Kobe, Japan, the conference turned into the annual meeting of the Internet Society.

After a nap to shake off jet lag, I drove across town for dinner with Larry and his wife, Jean. Also there was Tony Hearn, one of the CSNET founders, who was driving through on a college irlspection tour with his son and had stopped for the night in Madison.

We had a nice dinner of vegetarian lasagna and steaming homemade rye bread, finished by strawberries with a soy powder custard. When the second bottle of wine was cracked, a 1969 Pinot Noir, Tony Hern sniffed his glass appreciatively and smiled mischievously.

"Gee, this would taste great with meat," he teased. "Honey," Larry said, turning to his wife, "do we have any cow left?"

"I think we have a slab in the freezer."

I began to fade from jet lag and went back to my hotel, anxious not to miss my early flight the next morning back to Boulder.