Roger Bohn, University of California, San Diego
Hans-Werner Braun and Kimberly C. Claffy, SDSC
Stephen Wolff, Cisco
The characteristics of traffic on the Internet are changing dramatically, and the Internet is struggling to adjust, making it vulnerable to an impending traffic "crunch." Many researchers--including groups at the University of Michigan,1 Lawrence Berkeley Laboratory,2 Xerox PARC,3 and the Information Sciences Institute at the University of Southern California4--have begun to propose methods to cope with new workloads. In anticipation of the need for a stopgap measure before long-term solutions can be deployed uniformly, one research group has suggested an allocation scheme based on levels of service.
The Internet has long been perceived as a free highway rather than a toll road, and many users shudder at the mention--or implication--of usage-based accounting for Internet Protocol (IP) traffic. However, the proposed scheme is similar to the principles established many years ago at the national supercomputer centers. These principles--which are accepted by the researchers who use the centers' resources--are based on prioritizing supercomputer jobs as a way of charging against CPU allocations. Under these systems, the researcher specifies the priority at which to run a job. The higher the priority, the faster the job turnaround, but a correspondingly larger "charge" is deducted from the researcher's allocation.
The proposed scheme is similar in that it gives individuals, network service providers (NSPs), and other organizations that allocate computing resources (such as universities) incentives to prioritize their own traffic, thereby minimizing congestion.
The Internet's current architecture was designed to accommodate random traffic: The traffic comes from many sources, and its "burstiness" varies greatly. This architecture has worked well in the past, since most users have made relatively small, text-based demands on the network for e-mail, file transfers, and interactive login sessions aggregated across a large number of sources. As a result, all users have been able to share Internet resources more or less equitably while receiving equivalent service levels.
The emergence of high-volume, real-time applications, such as packet audio and video and rapidly changing graphics, is placing entirely new demands on the network. Rather than being random or "bursty," such applications impose continuous and often large flows on the network. Internet components, particularly routers, have little if any ability to control the volume and distribution of incoming traffic; therefore they can be overwhelmed by these flows. Furthermore, this situation allows for a relatively small number of users to commandeer large amounts of network bandwidth. Sharing, in the traditional sense, has become much harder to do.
A group of researchers has proposed an interim solution to make sharing more attractive. This solution would be used until an augmented network is implemented that can accommodate applications the Internet was not designed to handle. The researchers are Hans-Werner Braun and Kimberly C. Claffy (applied network researchers at SDSC), Stephen Wolff (former director of DNCRI at NSF and now with Cisco, a router company), and Roger Bohn (an economist in International Relations and Pacific Studies at the University of California, San Diego).
The group proposes a scheme for end users and applications to set Internet traffic priorities voluntarily by using the existing three-bit precedence field in the IP header. The scheme requires modifying Internet router software to support the field for queuing, and modifying applications software to set nonzero values in this field. This encourages three new behaviors on the Internet:
* Users and applications would use appropriate precedence values voluntarily in their outgoing transmissions according to flexible criteria.
* Internet routers would maintain multiple queues. Incoming packets would be queued by the value in the IP precedence field rather than by the customary first-in-first-out (FIFO) approach.
* NSPs and local network administrators could monitor the precedence values of all traffic entering and leaving their respective networks, and use some mechanism, such as a quota system, to give their users incentive to limit use of high-precedence values.
All three elements would be implemented gradually across the Internet, providing a smooth transition from the present system.
The group proposes implementing eight levels of service, 0-7, ranging from lowest to highest priority. They do not advocate linking precedence value with application type because, with thousands of applications already running on the Internet and users placing various "values" on any given application, such a plan would be too inflexible for such a diverse user community. Rather, they advocate linking precedence value with the "quality" of the individual flow. Aspects of a flow's quality could include:
* Time dependence (e.g., telnet), which requires reasonable interactivity. These types of flows would have higher precedence than ftp bulk file transfers and Usenet news.
* Volume: Applications requiring less bandwidth could have higher precedence than those requiring higher bandwidth (e.g., color video).
* Nature of transmission: Agency traffic across its own infrastructure would have higher precedence than game playing or other personal-interest traffic.
These three aspects could easily be inconsistent. For example, a videoconference might be very high bandwidth but would be very important to broadcast. Therefore the researchers also advocate flexibility so that users can judge the value of their own traffic and pay accordingly. Institutions might also base precedence values on other locally determined criteria--for example, user status (full professor, undergraduate student) and time of day (busy hour, late night). The default value for bulk traffic would be 1; precedence value 7 would be reserved for routing or other network management traffic considered critical.
This levels-of-service system can be implemented in a decentralized fashion and, therefore, only as needed. NSPs could enable it periodically or selectively as they feel a crunch, and disable it in the case of misuse. They could also choose whether to impose incentives for their "lower-level" customers. The incentive would be some quota on the total sum of IP precedence values used in their packets, in conjunction with some weighted average of the number of bytes and packets transmitted. Packets sent at the 0 or 1 precedence levels would always be quota-free, meaning that customers could always use unlimited Internet resources in the absence of congestion.
If a customer exceeds his or her quota, the NSP could assess a penalty (such as increasing the connection charge for a month). If the customer exceeds the quota regularly, he or she could negotiate a higher quota, establish and enforce quotas on subordinate units (as in the case of a university buying a bulk quota from an NSP, then suballocating it to departments and individual users), or continue paying the assessed penalties. The proposed scheme has several short-term advantages. It would
* Provide limited immediate relief to congested points,
* Support multiple levels of service on a single physical medium,
* Be easy to implement gradually, and
* Be socially acceptable due to its fairness.
Furthermore, the experience gained from this implementation would provide valuable knowledge upon which to build sound accounting and billing mechanisms for developing the commercial Internet.
An article on the proposed scheme was published in the Journal of High Speed Networks, and a paper is available from ftp.sdsc.edu (in pub-sdsc-anr-papers) and at http://www.sdsc.edu/0/SDSC/Research/ANR/kc/precedence/precedence.html. --SS
1 MacKie-Mason, J.K., and H. Varian, "Pricing the Internet." Public Access to the Internet, edited by B. Kahin and J. Keller. Prentice-Hall, 1994 (available at ftp://gopher.econ.lsa.umich.edu/pub/Papers).
2 Floyd, S., and V. Jacobson (Lawrence Berkeley Laboratory). "Link-sharing and Resource Management Models for Packet Networks." IEEE/ACM Transactions on Networking, July 1995 (available at ftp://ftp.ee.lbl.gov/papers/link.ps.Z).
3 Shenker, S., D.C. Clark, and L. Zhang (Xerox PARC). "A Scheduling Service Model and a Scheduling Architecture for an Integrated Services Packet Network" (online, submitted to Transactions on Networking).
4 Zhang, L., S. Deering, D. Estrin, S. Shenker, and D. Zappala (University of Southern California). "RSVP: A New Resource Reservation Protocol." IEEE Network, September 1993, 7(5):8-18.
|
