Optimization Models and Learning Algorithms for Slice Reservationin Virtualized Communication Networks

Abstract

The surge of mobile users with increasing needs of data and services has brought a tremendous momentum to telecommunications. This revolution paves the way to the development of next generation mobile networks, able to accommodate many businesses with specific needs on the same network infrastructure. The network slicing paradigm, enabled by technologies such as \ac{NFV} and \ac{SDN}, allows the \ac{NO} to host multiple verticals or \acp{SP} on the same substrate network. The network slices are dedicated logical networks that are tailored to the specific needs of each service. They are essentially composed of virtualized resources embedded on the \ac{RAN} and core elements of the network.

Virtualization has been identified as a key pillar of 6G. In virtualized systems, the \acp{SP} ask for bundles of resources that the \ac{NO} allocate accordingly to its admission control policy. The interaction between the \ac{SP} and the \ac{NO} has been modeled through dynamic markets, where the \ac{SP} can make in-advance reservation and bid for resources on the spot market. These markets have already been applied in the realm of cloud computing and have been recently transferred in the network slicing domain.

The problem has been studied extensively from the perspective of the \ac{NO} with solutions about how to find an optimal admission policy, how to solve the allocation problem which is typically NP-hard (e.g. \ac{VNF} embedding), how to charge for the slices with the aim to maximize the network utilization, or the \ac{NO} revenue. However, there has been little focus on the equally important problem of how the \acp{SP} should request the slices. Indeed, the efficiency of these markets depend equally on the \acp{SP} being able to optimize their reservation decisions.

The thesis focuses on the viewpoint of an individual \ac{SP} which aims to derive its reservation policy. The developed solutions must provide decisions robust to arbitrary changes of the service demand or the prices of the requested resources. Thus, the thesis proposes decision-making algorithms relying on modern learning tools, namely machine learning techniques and \ac{OCO}.

First, the research focuses on a \ac{RAN} reservation model, where the \ac{SP} can contract bandwidth capacity from the \ac{MNO}. The proposed solutions provide a near optimal reservation but the model does not account yet for the prices of the contracted resources. Then, the \ac{SP} can reserve different kinds of network resources, thus enabling the creation of an \ac{E2E} network slice. The derived optimization problem also accounts for the varying demand of the \ac{SP} and the evolving prices of the resources. Finally, the model is enhanced to allow the incorporation of key predictions on the demand and prices which improve the quality of the reservation decisions.

The first solutions are a \ac{DNN} and a \ac{LSTM}, which use as input the feedback from the \ac{MNO} from the $q$ previous slots and output the reservations for the following $h$ slots. The solutions outperform the \ac{ARIMA} baseline with idealized knowledge of the environment, by drastically reducing the over-reservation of \ac{RAN} resources. Then, the new objective is to maximize the long-term utility of the slice from the reservations, while not exceeding the allocated budget for the purchase of resources. The problem has convex objective and constraint functions and thus falls under the scope of \ac{OCO}. Finally, based on the feedback from the \ac{NO}, the \ac{SP} can derive accurate predictions of the future. The prediction module and the reservation solution are both online learning solutions which combined provide performance guarantees, while they outperform the \ac{FTRL} baseline.

The proposed solutions provide a near optimal online reservation policy to the \ac{SP}. The \ac{DNN} and \ac{LSTM} solutions present almost zero over-reservation and under-reservation under all traffic conditions (from congested to low traffic). The \ac{OCO}-based solution converges to the performance of the optimal online reservation policy under stationary and non-stationary settings. The prediction-assisted reservation solution provides better guarantees and outperforms the well-known \ac{FTRL} algorithm against real-world data.