Optimal resource allocation of virtualized router: a multi-objective approach

点击翻译按钮获取中文摘要

With the venue of the cloud and the large virtualized data centers, there is a huge need to provide new traffic isolation services between virtualized domains on the network. Unfortunately, the isolation available today is mainly based on the management plane and control plane and is not embedded with the data plane. This new**level of isolation will require to revisit the resource allocation strategies and redesign the network nodes (switches and routers). The targeted resources include the packet processing such as network processors, dedicated processors and multicore processors, memories such as TCAM, SRAM, etc., and interfaces such as Ethernet, etc.**Traditionally, the resource allocation strategies tried to maximize the global utility to guarantee performance and quality of service on packets. The aim of this research project is to push the sliceability on those node and ensure fairness between slices. The maximization problem is extended with new constraints that require the extension of the strategies, frameworks and algorithms related to resource allocation. The proposed research aims at solving this problem using an innovative combination of multiple approaches: Game Theory, Control Theory, Stochastic Optimization and Network Calculus. The Game Theory will be used to develop a fair distributed algorithm for resource allocation between slices. To achieve a fair allocation, we will develop a framework based on cooperative and non-cooperative game theory. For the cooperative approach, the Nash bargaining solution will be used to solve the centralized problem. We will then propose a distributed algorithm for implementing the optimal and fair resource allocation and provide conditions for its convergence.**On the other hand, the Stochastic Network Utility Maximization (NUM) is used to model a significantly large array of resource allocation problems and is well-adapted to the slicing router context. The utility function will depend on rate, latency, jitter, energy, distortion, etc and may be coupled across slices. The approach will help to understand and reduce the communication complexity of the scheduling algorithms.**We will develop a framework for the analysis of the distributed computational model. This framework will use the Network Calculus theory which is based on Min-plus algebra to develop the optimal mapping between resource and tasks/flow traffic. This approach will be used to elaborate traffic estimation methods of the backlog and the latency for each slice.**Another very important issue related to the resource allocation comes from the distributed structure of the router which is built on multiple controllers managing multiple components (Linecard, Switch fabrics, dedicated interface....). The distributed solution requires stability which will be obtained by applying control theory principles.**In conclusion, this project proposes a multi-objective approach drawing from the Game Theory, the Stochastic Optimization, the Network Calculus and the Control Theory to derive frameworks that will allow sliceability of the network node in order to provide an optimal and fair resource allocatio****