Dynamic Traffic Control in Multiservice Networks - Applications of Decision Models

Sammanfattning: In this thesis, we develop and examine decision models intended for traffic engineering analysis of multiservice networks. Two important issues are addressed, namely call admission and routing control in multiservice broadband networks, and overload control in distributed-memory systems. Multiservice broadband networking generally refers to the ability to handle widely disparate traffic streams such as voice, data and multimedia services on a single platform architecture. We study call admission and routing control in such networks adopting the well known multi-rate circuit-switched network model. The various services are distinguished by their bandwidth requirements, call characteristics, and performance objectives. The optimal control problem is formulated and examined using the theory of Markov Decision Processes. Some novel structural properties of the optimal call admission policy are revealed, but it is concluded that the general control problem is impractical to solve for any realistic system. Therefore, based on the structural knowledge of the optimal policy, a novel algorithm is formulated through which a reasonably accurate call admission policy can be calculated. It is demonstrated how this algorithm may be embedded into a fixed point network traffic approximation, leading to an efficient state-dependent least cost routing strategy for multiservice broadband networks. The second study in this thesis is devoted to overload control in distributed systems. The architectures considered are distributed-memory, message-passing systems, which essentially are collections of computers that communicate by sending explicit messages to one another. Typically, in order to serve a call, one or several, possibly physically distributed resources will be needed. From an overload control point of view, this poses a complex problem, since it is generally unknown in advance which resources the call will have to involve in order to complete successfully. We formulate and examine the associated optimal control problem using the theory of Markov Decision Processes. Again, it is concluded that the general problem is impractical to solve for any realistic system, and more efficient, yet accurate approximations are therefore developed and examined.