Adaptation de capacite dans le reseau dedie de service pour la maximisation du benefice

Service Overlay Networks (SON) are formed by placing overlay nodes over the Internet and interconnecting them with overlay links established by leasing bandwidth with Quality of Service (QoS) guarantees from a multitude of Internet Autonomous Systems (AS). Bandwidth is leased through service level agreements (SLA) with Internet Service Providers owning the AS. By covering a multitude of AS, the SON can provide end-to-end QoS to real time service connections serviced by its network.;The amount of leased bandwidth influences both the admitted traffic and network cost, affecting the network profit. This gives the network operator the opportunity to optimize the profit by adapting the network resources to changing traffic and SLA cost conditions. To realize this optimization, two elements are required: an efficient real time estimation of traffic demand, and an optimal capacity adaptation that will be fed by the demand estimation. In this thesis, we propose novel approaches that address the mentioned elements.;The approach for capacity adaptation maximizes network profit by modifying the SLA terms as needed to adapt overlay links bandwidth allocation to frequently updated traffic demand estimates. While performing capacity adaptation, the network Grade of Service, specified by connection blocking constraints, must also be maintained. Using an economic model, the link resources adaptation is integrated with the connection admission control (CAC) and routing policy in effect in the network. In our proposal, we apply a reward maximizing CAC and routing policy that is derived from the Markov Decision Process theory, although the approach can be applied to other routing policies. This integration leads to a distributed iterative algorithm for link bandwidth adaptation, where network profit sensitivity to link dimensions is calculated from the average link shadow price, itself being a routing parameter.;Approaches for measurement-based online traffic trend estimation that fit the SON capacity adaptation are next proposed. In the first approach, the smoothing parameter of the exponential smoothing (ES) model is adapted to the traffic trend. Here, the trend is estimated using measured connection arrival rate autocorrelation or cumulative distribution functions. The second approach applies a Kalman filter whose model is built from historical traffic data. In this case, availability of the estimation error distribution provided by the filter algorithm allows for better control of the network Grade of Service during capacity adaptation.;Analytical models as well as simulation of measurement based implementations of the proposed models are used to evaluate the performance of the proposed approach. Profit maximization by the capacity adaptation approach is confirmed by analysis of small network examples. Simulations on realistic network examples demonstrates higher network profit and/or better Grade of Service obtained by our adaptation proposal, when compared to fixed link capacities.;Study of traffic demand estimation methods shows that capacity adaptation performance is further improved with the use of our proposed methods. Adaptive estimation, as compared to fixed parameter estimation, provided higher profits. The proposed autocorrelation based ES estimation gives the best combined response and stability performances when compared to known ES methods. The proposed Kalman filter based approach improves capacity adaptation performance by significantly limiting network blocking increase when traffic demand increases.;Keywords: Service overlay network, resource management, capacity adaptation, traffic estimation, Markov decision process, Kalman filter.