| Thesis Director | Thesis Co-Director | Student |
| Prof. Joan Serrat Fernandez | Dr. Juan Luis Gorricho | Rashid Mijumbi |
ABSTRACT
Network virtualisation has been identied as a promising technique for dealing with the resistance of the Internet to architectural changes, and enabling a novel business model in which infrastructure management is decoupled from service provision. It allows infrastructure providers (InP) who own substrate networks (SN) to lease out chunks of their physical resources to service providers (SP) who use these resources to create virtual networks (VN). The SPs can then, either lease out these resources to other SPs, or use them to provide services to end-users.
However, before physical resources are shared, the different VNs meant to share them should be initialised, in which case virtual links and nodes must be mapped to substrate nodes and paths respectively. An important consideration during the initialisation of virtual networks is the need for efficient sharing of substrate network resources. This problem can be divided into three stages: (1) mapping virtual nodes and links to substrate nodes and paths (also known as virtual network embedding), (2) managing the resources allocated to virtual networks throughout their lifetime (also known as dynamic resource allocation), and (3) provisioning the substrate network to ensure survivability of the mapped virtual networks.
The constrained virtual network embedding (VNE) problem is NP-Hard. As a result, to simplify the solution, many existing approaches propose heuristics that make assumptions (such as availability of an infinite amount of resources in the substrate network, or advance knowledge of all the virtual network resource requirements), some of which would not apply in practical environments. This thesis proposes an improvement in virtual network embedding by proposing a one-shot virtual network embedding algorithm which is based on path generation. The path generation approach starts by solving a restricted version of the problem, and thereafter refines it to obtain a final solution.
In addition current approaches are static in the sense that after the VNE stage, the mappings are not altered for the entire lifetime of the virtual network. The few proposals that do allow for adjustments in original mappings allocate a fixed amount of node and link resources to VNs through out their life time. Since network load varies with time due to changing user demands, allocating a fixed amount of resources based on peak load could lead to an inefficient utilisation of overall SN resources, whereby, during periods when some virtual nodes and/or links are lightly loaded, substrate resources are still reserved for them, while possibly rejecting new requests for such resources. The second contribution of this thesis are proposals that ensure that resources are efficiently utilised, while at the same making sure that the QoS requirements of VNs are met. For this purpose, we propose a set of self-management algorithms in which the substrate networks use time-difference learning techniques to make autonomous decisions with respect to resource allocation.
Finally, while some scientific research has already studied multi-domain VNE, the available approaches to survivable VNE have focused on the single provider environment. Since in the more practical situation a network virtualisation environment will involve multiple InPs, and because an extension of network survivability approaches from the single to multi domain environments is not trivial, this thesis proposes a distributed and dynamic approach to survivability in virtual networks. This is achieved by using a multi-agent-system that carries out negotiations aimed at resource coalitions for resource reservations and restorations. The objective is to make each of the InPs adaptive and dynamic by giving them the capacity to perform QoS aware restorations in case of failures.
KEYWORDS: Network Virtualization, Dynamic Resource Allocation, Reinforcement Learning, Artificial Neural Networks, Fuzzy Systems, Multiagent Systems, Autonomous networks, Future Internet, Artificial Intelligence.