It is well known that feedback control may be difficult in a network with long propagation delays (with large bandwidth-delay products). The focus of this article is to provide a theoretical framework of various feedback-based congestion control techniques. Two broad classes of control algorithms are presented: source- and switch (network)-intelligence models. In a source-intelligence model, traffic sources adjust their transmission rates according to congestion information provided by the network. Theoretical models based on queueing theory and delay differential equations are presented. In a network-intelligence model, network switches evaluate appropriate transmission rates of sources. It is believed that any reactive control in a network with long delays is not effective. A control theoretic framework is presented, which attempts to exercise a degree of prediction of traffic patterns in the near future. Algorithms predicting the future input rates and the topological distribution of sources are presented. Finally, a unified approach combining both source- and network-intelligence models may be of interest for further study.

In HDLC, a zero is inserted after every five consecutive "ones" in order to resolve ambiguity about the location of delimiters. In this paper, we generalize the number "five" as an arbitrary positive integernand study the impacts ofnon the link utilization for a noiseless channel and on the reliability for a noisy channel. The entropy of the sequence after this generalized zero insertion is found for a noiseless channel. An expression for the optimalnthat maximizes a quantity called the delimiter efficiency is also found in terms of the frame size. It is shown thatn = 5achieves a delimiter efficiency close (within 2 percent) to that achievable by the optimaln. For a noisy channel, the reliability of HDLC is derived. It is shown that the probability of an undetectable error in HDLC is dominated by the errors that propagate through a created flag or incorrect zero deletion, and is in the neighborhood of 10-7for typical parameters. Increasingnby 8 is shown to improve the reliability by two orders of magnitude. It is proposed to use a two byte flag and a four byte frame check sequence to improve the reliability by seven orders of magnitude.