Workshop on Research Directions for the Next Generation Internet White Paper March 27, 1997 Timothy J. Salo Director, Networks Minnesota Supercomputer Center, Inc. 1200 Washington Ave. S. Minneapolis, MN 55415 e-mail: salo@msci.com phone: 612-337-3555 fax: 612-337-3400 Workshop on Research Directions for the Next Generation Internet White Paper T. J. Salo March 27, 1997 The Role of ATM In the Next Generation Internet =============================================== This white paper focuses on research activities intended to integrate asynchronous transfer mode (ATM) wide-area networks into large, public, IP internets, including the Next Generation Internet (NGI). ATM networks are likely to be important components of the (NGI). ATM networks offer the promise of integrating data, video and voice traffic onto a common network, supporting quality of service (QOS) assurances, (such as guaranteeing at least a minimum bandwidth is available or that network delays are less than some maximum), as well as scaling to very high speeds. Perhaps most importantly, wide-area ATM services appear likely to offer cost-effective, high-bandwidth, wide-area solutions. Commercial wide-area ATM services are available at DS-3 (45 Mbps) and OC-3c (155 Mbps) speeds. Research testbeds have created wide-area OC-12c (622 Mbps) ATM networks using commercial, off-the-shelf equipment. In the near term, ATM appears to provide cost-effective, high- bandwidth, wide-area network solutions ideally suited for the demanding, bursty requirements of high-performance computing and visualization and other potential NGI applications. This white paper describes several areas of research which appear likely to benefit the NGI: o Architectures for Large, Public, IP/ATM Internets o Architectures for NGI Interconnections Architectures for Large, Public, IP/ATM Internets ================================================= A number of wide-area IP/ATM testbeds have been constructed, including the interconnected MAGIC-II (http://www.msci.magic.net), AAI and ATDnet testbeds. The interconnected MAGIC-II/AAI/ATDnet testbed is perhaps the largest wide-area IP/ATM network constructed to date, comprised of approximately 200 ATM switches. However, this is perhaps three orders of magnitude fewer than the number of routers in the Internet. While these testbeds have demonstrated that moderate-sized IP/ATM networks can be constructed with available technology, the techniques used in these early IP/ATM testbeds do not scale to networks the size of the NGI or the Internet. The IP/ATM architectures of the MAGIC-II, AAI and ATDnet testbeds are continuing to evolve into much scalable solutions which may be applicable to the NGI. The MAGIC-II testbed intends to demonstrate a wide-area IP/ATM network based on the IETF IP/ATM protocols, (classical IP and NHRP), while the AAI and ATDnet testbeds plan to use the ATM Forum protocols, (LAN Emulation, LANE and Multi-Protocol Over ATM, MPOA). Other efforts to build wide- area IP/ATM testbeds based on scalable ATM and IP technologies are also underway. These and other research efforts should address a number of open issues relevant to the NGI, including: o What are the appropriate architectures for constructing large, scalable, public IP/ATM internets? o Are current and near-term ATM products and technologies adequate to construct, for example, the NGI? What is missing? o What are the advantages and disadvantages of the IETF and ATM Forum IP/ATM solutions? Is one more appropriate for the NGI than the other? Can both solutions coexist in the NGI? o How should functions be distributed between public, wide-area ATM services and public IP internets? For example, where should IP/ATM address resolution facilities reside, in the ATM or the IP network? o What is the role of end-to-end ATM connections ("short-cut" virtual channels) in networks such as the NGI? Is it important to provide end-to-end ATM connections in order to use the underlying ATM networks to provide QOS assurances, or can these assurances be provided at the IP level? o How well do short-cut VCs scale? Obviously, separate ATM VCs cannot be established for every pair of hosts, or even every pair of sites, which communicate over the NGI. Therefore, how can the network determine which, (if any), short-cut VCs should be established? How do pricing or cost allocation strategies and other systems administration issues affect the establishment of short-cut VCs? o What support should be provided for non-IP ATM traffic? Can the NGI provide support for both non-IP ATM connections and IP traffic simultaneously? How should the IP and non-IP ATM traffic interact, particularly in the provision of QOS assurances? o How ought policy be translated between human terms, (e.g., "My class was scheduled at 9:00 weekdays before the quarter began..."), and IP or ATM QOS assurances and scheduling? Architectures for NGI Interconnections ====================================== The NGI will likely require new architectures for internet exchanges, systems which would enable autonomous pieces of the NGI or the NGI and the Internet to meet and exchange routing information and data. A number of possible methods of interconnecting portions of the NGI or the NGI and the Internet should be explored: o Bilateral Interconnections o Link-Layer Internet Exchanges o Network-Layer Internet Exchanges o Mixed-Media Link-Layer Internet Exchanges o Mixed Link-Layer/Network-Layer Internet Exchanges Bilateral Interconnections -------------------------- The simplest method of interconnecting two internets is to establish a point-to-point connection between them. An interconnection between two IP/ATM internets should allow link- layer (ATM) connections, as well as network-layer (IP) connections, thereby enabling "short-cut" VCs between the networks. These short-cut VCs could use the underlying ATM networks to provide a specified quality of service. Such a facility might be useful for instructional video or other flows which have demanding QOS requirements. The implications of simultaneously providing link-layer and network-layer interconnections between two networks need to be better understood. The MAGIC-II and AAI testbeds plan to implement an IP/ATM-based interconnection which should provide a good model for bilateral IP/ATM interconnections. P-NNI will be used to announce aggregated ATM address information between the two networks, while an exterior IP routing protocol will announce aggregated IP address information. NHRP will enable short-cut VCs to be established between the testbeds. It is hoped that this solution will provide an interconnection with the scalability of IP which also supports direct ATM connections for "meritorious" applications. Link-Layer Internet Exchanges ----------------------------- Today, internet exchanges are typically based on link-layer technology. The Sprint Network Access Point (NAP), typical of these exchanges, is build around a pair of FDDI switches and provides link-layer connectivity between collocated routers owned by the NSPs which receive service from the NAP. Link-layer internet exchanges have become widespread because they do not impose the policy constraints that a network-layer solution, (e.g., router), would. In particular, a network-layer solution would force all traffic to a specific destination to use the same path, (in the absence of a router which routes on origination and destination address, rather than only the destination address). For example, if the internet exchange shown below were based on a router, then all traffic from Router1 and Router2 would both be transmitted by the internet exchange across either Path1 or Path2; traffic from Router1 could not use Path1 while traffic from Router2 uses Path2. ------------ Router1--| |--Path1--\ | Internet | \ Destination | Exchange | / Router2--| |--Path2--/ ------------ Some NAPs use ATM wide-area networks as the link-layer media. These NAPs appear to provide scaling opportunities, in terms of both speed and the number of connections, which are not as easy to realize with FDDI-based solutions. ATM NAPs also appear likely to offer opportunities to more easily integrate wide-area ATM networks. Network-Layer Internet Exchanges -------------------------------- Few network-layer internet exchanges exist, largely because of the policy which they impose on attached networks. These network- layer exchanges might be viewed as implementing a buying cooperative where the attached networks, (Router1 and Router2 above), agree to all contract with the same inter-exchange provider, (they all agree to contract to use Path1 and never Path2 in diagram above). Mixed-Media Link-Layer Internet Exchanges ----------------------------------------- It is important to develop a method of gracefully migrating, for example, FDDI-based internet exchanges to ATM-based internet exchanges without using a router, (which would impose a next-hop policy as described above), to connect the different media. The ATM Forum's Multi-Protocol Over ATM (MPOA) and LAN Emulation (LANE) protocols appear likely to offer a solution, but this solution needs to be demonstrated in a production internet exchange. To the extent that ATM wide-area networks are used in the NGI, mixed-media, FDDI/ATM internet exchanges are likely to become important in interconnecting the NGI with the existing Internet. Mixed Link-Layer/Network-Layer Internet Exchanges ------------------------------------------------- Some mixed link-layer/network-layer exchanges have been proposed. As currently described, the Gigapop includes both link-layer (an ATM switch) and network-layer (a router) components. The applicability of this architecture needs to be more fully explored to better understand whether this mixed link-/network-layer solution combines the advantages of both link- and network-layer solutions, or whether it provides the worst of both worlds. Evolving Roles for Wide-Area Network Testbeds ============================================= The existing wide-area network testbeds, specifically those which include wide-area ATM networks, are ideal, cost-effective platforms for research applicable to the NGI. Efforts to interconnect these testbeds should promote and stress the scalability of IP/ATM network architectures, thereby increasing the likelihood that they will be directly applicable to the NGI. The Minnesota Supercomputer Center, Inc. ======================================== The Minnesota Supercomputer Center, Inc. (MSCI), one of the largest technical computing centers in the world, has provided supercomputing services to industry, government, and academia since 1982. MSCI also provides housing and administration for supercomputers of the Army High Performance Computing Research Center (AHPCRC). The MSCI Advanced Networking Group conducts research into new communications technologies and supports MSCI's production networking team. The four members of the group have over seventy years total experience and have been working together for more than five years. All four group members worked on the MAGIC Gigabit Testbed and are currently working on the MAGIC-II project. Timothy J. Salo has twenty years experience researching, designing, developing, marketing and operating computer communications networks. Mr. Salo is Principal Investigator for MSCI's portion of the MAGIC-II project and was the PI for MSCI's portion of the ARPA-funded MAGIC Gigabit Testbed. In these roles, he is responsible for research on the integration of high-speed, wide-area ATM networks into large, public, IP internets. In addition, Mr. Salo is an Additional Principal Investigator for one of the NSFNET Network Access Point (NAP) projects. His area of interest is the use of ATM within NAPs and elsewhere in the Internet.