Performance optimization with integrated consideration of routing, flow control, and congestion control in packet-switched networks
Kongxun Wang · SMARTech Repository (Georgia Institute of Technology) · 1991
The control of network operations in a packet-switched network is traditionally studied by separately considering its different components because of the complexity involved in an integrated approach. However, as many researchers have pointed out, the separation of some important components which are tightly coupled cannot reflect the real behavior of a network; randomly using several control mechanisms together may lead to poor network performance. In this thesis, an integrated approach is taken to the problem of network control. This approach is to find the best set of control parameters for some important basic control mechanisms which together will realize the integrated functionalities of network control. These functions are routing, flow control, and congestion control. First a closed, multi-chain, queueing network model is proposed. With such a model, the control variables are identified which corresponds to the control parameters for the control mechanisms used in the communication network, specifically, window sizes for window flow control mechanisms and traffic splitting for entry-to-exit routing. Then an optimization problem is formulated based on the model. Utilizing AMVA (Approximation Mean Value Analysis), a realistic performance objective function is defined which can resolve the performance criteria conflicts. This problem is solved analytically. Necessary formulas are derived. An algorithm is developed to compute the gradient vector and Hessian matrix. Then a general search strategy can be used. The computational results have shown the advantages of this method to be computational efficiency, accuracy of results, stable and convergent behavior of the method, etc. The solution means that both window sizing and routing are optimized together. No other work has optimized both of these factors. Even optimizing window sizes alone, which is a special case of the problem solved in this study, has not been done. A new scheme to maintain routes in the entry-to-exit routing is also proposed which not only solves one of the implementation issues of this method, but also can be used in other existing control architectures.