White Paper for CRA Workshop on Research Directions for the Next Generation Internet Dr. Craig Partridge Division Scientist BBN Corporation 7424 Rockway Ave El Cerrito CA 94530 craig@bbn.com Phone/FAX: 510-558-8675 I'm interested in attending the NGI workshop for two distinct reasons. First, I'm deeply interested in a key issue of research policy: how do we ensure that the NGI project not only builds a network, but also establishes a foundation of research results from which the NGI can continue to grow over the coming years? Second, as the leader of a project building the world's fastest IP router, I believe I could provide a helpful perspective on the future of router technology. This white paper talks briefly about both topics. Building a Strong Research Base While the NGI effort's funding is only for three years, the goal of the initiative is to strengthen US networking activities for future. Accordingly we need to give some attention to how we structure the NGI activities to ensure that when the three years are over, we will have not just built a faster network, but also have done the necessary groundwork to ensure the network (or its successors) can grow and thrive into the future. Building for the future means we need a plan to fund leading edge research. By leading edge research, I mean research intended to pay off in 3 to 10 years, as contrasted with near-term research (also called advanced development) which confronts a problem that must be solved in the very near future. Creating such a plan is a challenging problem. But we need to have a coherent inter-agency plan for at least three reasons: * First, building faster or better networks requires innovations in at least five fields: transmission technologies (how we move bits faster over wires), media layer technologies (how we interconnect wires to build networks), internet layer technologies (how we build routers to interconnect the different media layers), end-to-end protocols (how we communicate between applications over an internet), and applications. If one field fails to keep up, the other fields suffer. For instance, in the early 1990s, little research money was put into router technology, with the result that until recently, routers lagged behind other technologies. The result was that people didn't buy new higher-speed transmission technologies because they didn't have routers to plug them together. Expressed another way, one cannot just fund leading research in one field, because those results won't be usable until the other fields catch up. We need to assess the state of each field and balance the research agenda to ensure that no field gets left too far behind. * Second, like most fields, networking research has its fads or popular technologies. Fads often get more attention than they deserve. Taking an example from the late 80s and early 90s, ATM was the hot technology of the time and large amounts of Government and private research money was sunk into ATM. ATM has proved disappointing. Non-ATM technologies were neglected. Arguably 100 Mb/s Ethernet would have been available earlier, but for the money focussed on ATM. One aspect of fads that impacts the NGI initiative is that every agency wants to be seen as part of the hot research, and as a result, lots of me-too funding tends to occur (e.g., each agency had to have its own ATM project). One possible solution to this problem is to encourage research in multiple directions (i.e., fund competing technologies) and to make it easy for agencies to jointly fund multiple research efforts (so each agency can be part of a common effort rather than funding its own). * Third, research results take time. The work requires hard thought and much research depends on the availability of recent PhDs and PhD students. Research funding started in 1997 may not pay off until 2001 when the relevant Ph.D. candidate finishes her thesis. Given the goal of providing results that allow us to continue to grow the NGI in the 21st century, this time lag is perfectly acceptable. But it does mean that we need to start funding research from the start. There's a nasty and common pitfall of building the infrastructure first, and funding the research second, with the result too often being a technology gap in which the infrastructure needs to grow further but the research to make that growth possible is barely underway. Fast Routers BBN Systems and Technologies, under a research contract from DARPA, is completing a prototype multigigabit Internet router designed to handle up to 50 gigabits per second of IP datagram traffic. The router (known as the MGR) is a balanced router, in that it is capable of examining every datagram that it receives. The router also incorporates features to support quality-of-service (including providing delay guarantees to real-time traffic). The MGR supports OC-12c (622 Mb/s) interfaces and can easily support as many as eleven OC-48c (2.4 Gb/s) interfaces or over a hundred lower speed (155 Mb/s) interfaces. In the context of the NGI, where hundreds of sites will be connected at rates of 155 Mb/s and faster, several routers like the MGR will be needed in the backbone. As a result, one important question is what research in routers is needed to make high-quality commercial routers available at multigigabit speeds. Currently the high-end of the router market is in the process of moving to a new architecture, based on parallel forwarding technologies and high- performance switched backplanes. There are a number of variations on this architecture, ranging from classical routers re-implemented in the new form to routers (e.g., BBN's MGR) that use statistical techniques to try to leverage performance (e.g., the Ipsilon product). Our challenge over the next few years is twofold. We need to explore the range of choices in this new architecture and understand which ones yield the most cost-effective performance. Second, we need to work on how to scale this architecture to higher speeds. There is general agreement that the basic architecture scales to terabit speeds, but that there are a number of knotty problems, especially in memory architectures and routing algorithms, that need attention as we scale up. Some of these issues will need attention as the NGI is built.