Description
The Highball Project is a research effort designed to specify and ultimately build a fast, wide area network. It is designed to operate at gigabit speeds or higher, although nothing in the design specifies a particular speed. Connections between nodes are either multi-conductor cable or fiber optic, usually occurring in pairs, one cable or fiber for each direction. Nodes are crossbar switches, capable of connecting any input to any or to many outputs.
The network is designed for high demand, bursty users. Typical users include supercomputers, real-time disk-to-disk transfers, interactive video, and visualization on remote workstations. These users can individually demand a gigabit or more, but do so for brief, bursty periods. In contrast, conventional high speed networks are usually designed for highly multiplexed aggregates of low rate users. The statistics are very different for bursty users than for a multiplexed aggregate. A network designed for multiplexed, low rate users is ill-suited for bursty, high rate users.
Because of the high speeds involved, communication bursts are "short" compared to the size of the network. For instance, even a megabit burst at 1 Gbps is only a millisecond long and flies across the country in about 30 ms. Instantiating a virtual circuit for a burst is inefficient. On the other hand, store and forward networks with acknowledgments are infeasible at these speeds.
The scheduling problem is perhaps best understood in terms of an analogy to a railroad of high speed trains. The trains have finite size and are small compared with the distance across the network. We desire to dispatch trains without collisions and without stopping a train once started. Furthermore, unlike real train networks, we cannot signal ahead because our trains are traveling as fast as the signals. The scheduling problem is when to start the trains and how and when to set the switches along the way. One complication is that we desire to transmit many trains concurrently, each going in different directions. This scheduling analogy also gives the Highball Project its name. In the vernacular of trains, highball means that the train has priority and can travel as fast as the tracks allow. We are trying to allow our trains to travel as fast as their tracks allow.
Time synchronization is a significant hurdle for this network. Bursts propagate at about two-thirds of the speed of light and are switched on-the-fly. The entire network must agree on the clock phase to within about one microsecond, while delays between nodes must be known as accurately. We believe the network clocks can be made this accurate using the methodology of NTP, which has recently been extended for just this purpose.
Although there is no explicit funding for further work in this area at this time, we believer there is considerable potential for further research in related areas, such as capacity reservation on high speed networks in which traffic bursts may be many packets and may be multiplexed with packets from other sources. This would help avoid output blocking on high capacity ATM switches, for example. Other examples may be automated highways, in which a "burst" is a train of cars travelling as a single group.
Perhaps the most significant results from this project are spinoffs for other projects. The Highball architecture has influenced our work with multicasting and autonomous configuration in other projects funded by DARPA, NSF, U.S. Army and U.S. Navy. Hardware developed in the Highball project has been adapted for use in other projects working with precision synchronization technology. Software tools and simulators developed in the Highball project continue to be used for further development of NTP and related protocols.
Selected Publications
- Mills, D.L, C.G. Boncelet, J.G. Elias, P.A. Schragger, A.W. Jackson and A. Thyagarajan. Final report on the Highball Project. Electrical Engineeing Department Report 95-4-1, University of Delaware, April 1995, 30 pp. (paper only)
- Boncelet, C.G., and D.L. Mills. A labeling algorithm for just-in-time scheduling in TDMA networks. Proc. ACM SIGCOMM 92 Symposium (Baltimore MD, September 1992), 170-195.
- Mills, D.L. Proposal on a Highball architecture for the ACTS TDMA network. Submitted to DARPA in response to BAA-92-36 (6 August 1992), Electrical Engineering Department, University of Delaware, 21 pp. [unpublished internal report - do not cite] (paper only)
- Mills, D.L., C.G. Boncelet and J.G. Elias. Interim report on performance and policy dimensions in internet routing. Electrical Engineering Department, University of Delaware, March 1992, 11 pp. [unpublished internal report - do not cite] (paper only)
- Mills, D.L, C.G. Boncelet, J.G. Elias, P.A. Schragger and A.W. Jackson. Highball: a high speed, reserved-access, wide-area network. Electrical Engineering Department Report 90-9-3, University of Delaware, September 1990, 34 pp. Abstract: PostScript | PDF, Body: PostScript | PDF
- Mills, D.L. Final report on DARPA/IPTO Contract N00140-87-C-8901:Research on Internet architectures and protocols. Electrical Engineering Department Report 90-7-2, University of Delaware, July 1990, 11 pp. (paper only)
- Mills, D.L. Proposal on performance and policy dimensions in internet routing. Submitted to DARPA in response to BAA (28 December 1988), Electrical Engineering Department, University of Delaware, May 1989, 19 pp. [unpublished internal report - do not cite] (paper only)
- Mills, D.L. Proposal on research in internet architectures and protocols. Submitted to DARPA, Electrical Engineering Department, University of Delaware, August 1986, 12 pp. [unpublished internal report - do not cite] (paper only)
