Decentralized Adaptive Flow Control of High-Speed Connectionless Data Networks

We introduce a permit-based adaptive control scheme for regulating traffic admission in high-speed connectionless data networks, such as the internet. Permits are awarded to potential customers arriving from outside and travel with them towards their destinations, where the permits are assigned to the local controllers. The controllers randomly distribute the permits among the entry gates at the nodes. Customers from outside are not allowed to enter the network unless there are permits available at the entrance node, thus the model is that of a closed queueing network if we model the dynamics of the permits. The goal is to find the permit distribution strategy that maximizes network performance subject to the restrictions of the network topology. A traffic balance approach is used to establish nonuniqueness of the optimal distribution probabilities for the decentralized operation. We exploit nonuniqueness introducing the concept of the automata actions , focusing on two strategies for the actions. For each strategy, a learning automaton is implemented at the controllers using the Kuhn–Tucker conditions for optimality. Our first learning algorithm converges weakly to a unique limit point, which is optimal, while the limit behaviour of our second learning algorithm may be suboptimal. We illustrate our results using computer simulations in order to compare the two strategies for the same network.