Sensitivity methods for congestion control in computer networks.

摘要

As computer networks move towards handling diverse traffic types with different service requirements, there is a need for advanced network control mechanisms that can regulate the network traffic to meet the users' service requirements. Service guarantees can be provided if the network makes resource reservations (bandwidth, buffers, priority scheduling, etc.) on behalf of each connection. A problem arises, however, in the control of connections that do not reserve network resources, and hence are not given any performance guarantees by the network. One network control scenario is for these low priority connections to adapt to changing network conditions in order to achieve their data transfer goals and to minimize network congestion. Congestion in this case is a result of a mismatch between the network resources and the amount of traffic admitted for transmission. Consequently, congestion control can be interpreted as the problem of matching the admitted traffic to the network resources. This, in turn, could be viewed as a classical problem of feedback control (i.e., matching the output to the input of a dynamic system). This thesis considers sensitivity methods for the control of congestion in computer networks using neural network models of the system dynamics and system performance sensitivity derivatives. The control methods proposed are used to determine how a feedback-based rate controlled source can satisfy its data transfer requirements by adapting its data rate to changes in the network state.