This week, the University of California in Los Angeles (UCLA) is celebrating the thirtieth anniversary of the first four nodes of the Arpanet. Over the past thirty years, the Internet has grown from 4 nodes to several hundred million; this represents a compound growth rate of over 80% p.a. every year! The equivalent growth in traffic is nearly doubling each year over the same period. This compares with a growth of the number of the telephone network of around 16% p.a. over the last 120 years. Clearly, the main advances in the network technology have come from the US. This has been greatly helped by the existence throughout this period of major networks that could be used as playthings or testbeds for experimenting with new techniques, technologies and applications. These testbeds have not been uniformly useful; that is the wrong thing to ask. If each testbed had been useful, then they would not have been ambitious enough. However even many of those that appeared not to have been successful at the time, have often led to further massive improvement in facilities in the future. It is one tenet of my talk that such testbeds are invaluable drivers of the technologies; their paucity, or at least their more limited nature, elsewhere has contributed to the relative immaturity of the activity in many other countries.

Many lessons can be drawn from the study of the Internet. Here I will concentrate on four: Openness, Heterogeneity, Availability and Globalisation. I will give

examples of how each has benefited the system that we now call the Internet, and draw some lessons on how they may continue to do so. Finally, I propose an initiative on a new, medium-scale, very heterogeneous, international network and application testbed.

Because I am more concerned at the concept of what can be learned from testbeds than the testbeds themselves, I will give only limited references. A much fuller set of references on the early experiences is given in [kirs].

There has been large-scale network activity both by industry and different segments of the US government since the sixties. Nevertheless, it is recognised universally that DARPA has been the main driving force. It is no coincidence that of the ten previous recipients of the prestigious SIGCOM award, three have been in DARPA at some time, and several more have been associated closely with Arpanet or the Internet. DARPA supported a significant number of large-scale data network testbeds, where large must clearly be considered relative to the then size of data network. These included Arpanet (68-79), Packet Satellite (75-80), Packet Radio (73-78), Wideband satellite (78-88), Internet (77-87), Gigabit (91-96), DARTNET (91-96), CAIRN (96-99), and Glomo (93-98). All of these started off as technology demonstrators; many were encouraged to develop into more generalised application platforms.

In the early 70s, many of us were using the fledgling Arpanet for long terminal sessions and even packet speech. These were not done as demonstrators; they were for serious work outside the Computer Science Area. This was rapidly replaced by the electronic mail systems of the middle 70s. By 1978, there were already a hundred UK research groups approved for regular collaboration with partners in the US. The simultaneous need to develop the technology to build a high quality network, and to provide high quality services on a significant scale were so important. They led both to the Arpanet becoming such an excellent basis for the Internet, and to the early Internet being extensible into its current commercial success. To give just one example, I remember an early meeting around 1972 where the Arpanet had had problems, and ISI had been unavailable for a week. Since all the DARPA Management Information System was at ISI, this was an intolerable disturbance. Larry Roberts said to the group "You will stop this happening again, or I will withdraw all support from Arpanet". As a result there was a massive attack on all the areas that could cause significant problems, and a step-function improvement in availability. It was this drive to provide leading edge technology, while exercising the system, which was so important in these testbeds. All these networks were research networks and yet they migrated also into networks for research.

There was an important hiatus in the mid-80s. DARPA felt it had done the basic research on networking technology; there was little need for research networks. The provision for networks for supporting research was considered the remit of NSF for the academic community. At the same time other Federal agencies were putting in networks both to support research and even operations; these included Energy, Defence and NASA. The scale of these operations provided applications drivers, even if the research networks had these only on a limited scale.

The testbeds supported by DARPA in the 90s have had less of an applications driver. Many of the Gigabit testbeds did have one or two target applications; however it was possible to tailor many aspects of the testbed to the application, and few large-scale persistent applications emerged. The requisite network components were developed, and the basis was set for another push into advanced applications. The ATDNet and ACTS networks did incorporate applications; however they were much less available to the bulk of the research community.

With the formation of the Federal Network Council (FNC) in 1995, followed by the Large Scale Networking Group (LSN) in 1997, the US seems really to have put its programme in place. With the different phases of the Next Generation Internet (NGI) programme [ngi], there are a number of research networks. With the rapid expansion of the VBNS network, followed by Abilene and the Internet-2 [int2] initiative, there are networks to support research. These have much less advanced technology, relative to the current state of the art; it follows the now well-established US successful formula of having sizeable advanced technology research networks, with their technology feeding through into yet large networks for supporting research with relatively stable technology. Moreover, Internet-2 also has a significant number of advanced development activities like IPv6, Measurement, Multicast, Network Measurement, Network Storage, Quality of Service, Routing, Security and Topology.

When we examine these testbeds by the criteria of Section 1, they do very well. Almost all were very open; being funded partially by US Federal funds, the results were normally made available widely. Their heterogeneity was limited. While many sorts of computers were often involved, the assumptions were often very similar. This was partly due to the national character of many of the testbeds, and partly on the way the funding was provided. Availability was normally good; of course technical faults might limit availability, but most were designated to be available most of the time. Globalisation was very limited; the experimental networks were mainly restricted to the US, with certain notable exceptions - some of which are discussed in Section 4.

The early support for Arpanet and Internet came almost entirely from US Federal funds. It included some industrial laboratories (e.g. Xerox Park and IBM Research) at their own expense; almost all industrial participation was under Federal contract. Even the early communications costs were largely at commercial rates. It was only with the Gigabit testbeds that there was a large-scale investment by US industry in cost-sharing arrangements. With their large investment in fibre, and a considerable surplus capacity on many routes, the US carriers have continued to participate on a shared cost basis - but almost only domestically up to now.

In Canada, there was an attempt in the early 70s to set up a Canadian Research Network (CANUNET), under the auspices of the Department of Communications. This had the political aim of unifying Canada at that stage of its history. For various reasons CANUNET was never built. It was only in the 90s that Industries Canada was given the responsibility for constructing CANARIE [can]. This is a true testbed, and has substantial contribution from Canadian industrial organisations. Moreover, it is available both to the Canadian universities and their industries. It has truly experimental components, while at the same time providing a proving ground for advanced applications.

The Europeans research community has almost always had a more unified view of networks than the US. There has normally been only a single agency responsible for networks for research – and that has almost always been responsible also for research networks. There were some early research networks; one of the earliest European ones was the European Informatics Network (EIN). Most of these early network were either research networks, or networks to support research. It is very rare that they have been both. Thus in France the Cyclades Network was a research network; neither EIN nor Cyclades had serious users. By contrast, Euronet was based entirely on the Cyclades technology, and was put in to support researchers; it had no element in developing network technology.

Almost all the academic network put in during the 80s and early 90s were networks to support research. In general they had a very small ostensible research content. The British SuperJANET was supposed to be an exception. Its original concept was that a clear portion of the network would be made available as a research network for network researchers; it was on this basis that British Telecom provided the special rate for the network, and agreed to fund academic research using it. In the event, certain high visibility applications, in particular video conferencing, were found to be very sensitive to traffic variability; as a result, none of the network was ever really made available as a research network that could be broken.

Since 1995, two sets of networks have been made available as real testbeds. One set comprises certain national research networks - in particular in Scandinavia, portions of SURFnet in the Netherlands, and some research networks in Germany. Clearly CANARIE [can] in Canada has had this mission, and there have been some bilateral arrangements between German Telecom and Teleglobe to connect the German DFN [dfn] and CANARIE for specific experiments.

The European Union intended to provide a network to support research for its research projects under the Framework IV programme. The resulting JAMES network fulfilled this function for several years; it also allowed some very limited network research - particularly on ATM interconnection. However, because of the way the constituent components were run by competing carriers, it was difficult to get much real insight into the operation of the network, or to have outsiders become much involved in experimenting with the network itself. The successor to JAMES, the QUANTUM project [quan], is again principally a set of interconnections between the national service networks; it will allow limited experimentation outside the network, or on Virtual Private Network (VPN) components that can be derived. However these are of very limited scope; they are similar, but much less ambitious, than those envisaged on Internet-2 without the added research potential offered by the NGI testbeds [ngi].

With the growth of the Internet, the carriers and some national academic network providers have started to see the need for serious research networks. Several carriers have started activities for research networks. UNINETT/Telenor (Norway), SUNET/Telia (Sweden), DFN/German Telecom (Germany), British Telecom (the UK) and SURFnet/KPN (Netherlands) have significant projects. In most cases, the projects are very limited in both size and scope. They normally do separate the applications from the network technology; only very limited scale network topologies are available for experimentation. All are restricted to national activities - though Nordunet, DFN, SURFnet and SuperJanet have both joined Internet-2 and planned connections to it.

I separate out the British component in the above from the others for a specific reason. Most of the national activities are a combination of a carrier and the organisation responsible for networks to support researchers. The early SuperJANET was just such a collaboration. However, when it was found not to incorporate real networking research, British Telecom decided to pursue that avenue by a separate network in conjunction with a few particular universities. A Proposal for a world-class testbed are currently under consideration in a way that would include both UKERNA and private industry.

Some of these networks are experimenting with advanced access arrangements - in particular xDSL and even wireless access. Only the French and some Germans seem to have an active programme with DBS satellite; in both cases these seem to be connected to some political considerations. In the French case it is connected to links with the Francophone countries, and in Germany it is connected to links with the former CIS countries.

In Europe there has been a massive push, under the European Framework programmes, to encourage collaboration on a pan-European scale between partners in different countries both industrial and academic. In theory, the JAMES and QUANTUM networks were intended to make on-line collaboration easy in such projects. In practice, this has not yet happened. The access to JAMES and QUANTUM has been mainly through the academic networks; they have not necessarily provided either easy or low-cost communication to commercial entities. JAMES was not organised in a way that such access could be guaranteed either. QUANTUM is just an interim project until the FRAMEWORK V initiative gets properly underway; whether it will provide the right environment for the participation of the commercials is not yet clear.

So far most of the collaborations have made only very limited use of on-line networks. Even in the ACTS programme, which was designed to develop a Broadband Integrated Network Service for Europe, most of the final demonstrators were built up at single sites!

There is an additional problem with the involvement of commercial companies in the FRAMEWORK projects. Most of their research activities are only 50% funded by public funds, and the results are not necessarily published widely. Easy dissemination of software cannot be taken for granted nearly as much in these projects as in their US equivalents.

In theory, the ACTS programme envisaged the construction of National Hosts. A National Host would consist of some specific network-based research, and it was envisaged that the National Hosts would be connected together. Both TEN-34 and JAMES were supposed to connect together these National Hosts. The theory was excellent; if completely carried out, it would have rivalled the plans in the US. The national academic networks were mostly offered as National Hosts. Other important facilities like national cellular telephone testbeds, xDSL testbeds and experimental satellite facilities were offered also. Unfortunately, the reality was far different from the 1994 hopes. Most of the interesting experimental resources were not linked in to the national academic networks. Moreover, the lack of a real pan-European research network did not allow these interesting resources to be linked on a trans-European basis.