White Paper for CRA Workshop on Research Directions for the Next Generation Internet May 13-14 1997, Vienna Virginia New Services and Capabilities for the Next Generation Internet Mark W. Garrett Senior Research Scientist Bellcore 445 South St. Morristown NJ 07960 201 829-4439 Fax: 201 829-2504 mwg@bellcore.com New Services and Capabilities for the Next Generation Internet Mark W. Garrett Bellcore White Paper for CRA Workshop on Research Directions for the Next Generation Internet May 13-14 1997, Vienna Virginia Many aspects of the current Internet and its protocols have shown remarkable resilience to the latest order of magnitude or two of growth that the Internet has experienced. Inevitably though, the current and imminent scale of the network implies dramatic qualitative changes, such as multi- provider commercialization, which change the relative importance of network features. In the past, a major emphasis was placed on the correctness of protocol specification and implementation and the interoperation between implementations. Much less attention was paid to, or needed for, such issues as efficient resource consumption, performance, reliability, service definition and verification, inter-provider financial agreements. As the Internet grows through its next order of magnitude, the population of users will grow to include, if not be dominated by, serious corporate communications and mission-critical applications. The requirements of such users will not be met, nor follow the same economic model as the current $19.95 per month best effort service. The current network users will not be displaced by business users, no more than business telephone service excludes the market for residential service. The differences in economics, usage and service requirements does, however, imply a need for differentiated services within the Internet. If we look back at the Internet before the time when the word ``http'' adorned every billboard and magazine ad, the users of the Internet were mainly the research and education community. Because of the lack of service differentiation, their needs have been essentially swamped out by the much larger group of web-surfers, whose tolerance to bad performance is much higher. Similarly, an undifferentiated Internet dominated by business usage may provide higher quality, but the price would make it inaccessible to the current users. A number of technical solutions have emerged from the Internet research community in recent years which, although certainly useful, may not turn out to be as universally curative as hoped. These include the familiar weighted fair queueing (WFQ), resource reservation protocol (rsvp), random early dropping (RED), predictive service, and measurement based admission control. While these mechanisms are based on first-rate creative thinking, the current scale of the Internet does not permit rapid experimental deployment, as it did only several years ago. This implies that a slightly different mindset and approach will be necessary for success in modifying the Internet from now on. There are three main drivers of differentiated Internet services: the source traffic description, the transport quality requirement, and economics. The TCP/IP family of protocols have always had a bias toward uniform treatment of packets, both in terms of service (first-in-first-out) and routing (shortest path first). Some policy features are now deployed in BGP routing, and the first implementations of rsvp and fair queueing are just recently available in commercial products. These mechanisms offer some departure from the philosophy of ``a packet is a packet.'' To meet the requirements and willingness to pay for a very diverse set of users and applications, will require much more sophisticated and diverse mechanisms. For example, the Integrated Services (IS) protocols from the IETF (Guaranteed Service and Controlled-Load Service) are usually discussed in the context of the rsvp protocol, which allows a user to set up explicit resource reservations from one end host to the other. While this function is useful, the capability of end systems to use rsvp will probably not grow as rapidly as the need for other types of differentiation. The IS services can be instantiated without rsvp, and resources can be thereby controlled on a non-end-to-end basis. This is an important aspect of IP IS which has not been explored much in the literature. There are related issues concerned with bandwidth provisioning. The service provider can begin to control which sources are sharing a link, and their relative priorities. But we do not understand well enough the relationship between the description of the traffic sources aggregated together, and the impact on application-perceived quality of service as a function of the bandwidth resource they are forced to share. The issue of statistical multiplexing, or ``aggregation'' becomes much more important as the Internet matures. In the past growth depended on signing up more customers. Now the volume of traffic, and the potential for huge traffic sources such as video, forces ISPs to make efficient use of bandwidth, especially in the wide area backbones. The concept that many sources can be accommodated with less than the sum of bandwidths needed for each individually, is well-accepted. The difficulty is in finding numerical values for the parameters, and in not abstracting away the detailed consideration of the Qualtiy of Service required by the applications. The advent of differentiated services also raises questions about routing protocols. It is not clear that the policy components of current methods are always sufficient to allow providers enough flexibility to route packets with the correct service levels while retaining route redundancy and proper interoperation across ISP boundaries. ``Reliability'' is another term that will grow in stature as the Internet crosses its next power of ten. The Internet does not attain its reliability in the same fashion as the telephone network does. Largely due to the particular historical and political forces of the time, telephone equipment is designed to be about 99.99% reliable. IP routers have their own history (and funding sources) which led to a different style for ensuring that packets get delivered (e.g., end-to-end loss recovery and dynamic routing table updates). Since these mechanisms were developed for a much smaller and more operationally homogeneous Internet, we can rightly ask whether the design assumptions work in the new environment. The pressures of a more business-oriented clientele for ISPs will eventually resolve some of the identifiable reliability and performance problems that exist today (e.g. bandwidth-constrained Network Access Points, error-prone router configuration methods, etc.). However, there are certainly technical gaps to be filled in order to build an industrial strength Internet. The goal at hand is to advance Internet technology to the point where people rely on it as a component of the ``national communications infrastructure.''