07 out 2020
14:00 Master's Defense Fully distance
Theme
Admission Control and Resource Allocation in 5G Network Slicing
Student
William Fernando Villota Jácome
Advisor / Teacher
Nelson Luis Saldanha da Fonseca
Brief summary
The infrastructure of 5G network service providers will receive requests for the implementation of network slices (in English, network slices) requested by users with different quality of service (QoS) requirements. Considering that the resources on the network substrate are finite and the 5G use cases have heterogeneous QoS requirements, as well as specific deployment costs, service providers need to manage the admission of these requests, as well as allocate resources so that the use of these be efficient. Different approaches deal with admission control and resource allocation in 5G networks using different theoretical frameworks such as Queuing Theory, Complex Network Theory, and Optimization. However, these approaches propose making admission decisions considering individual requests, which can lead to sub-optimal decisions, since the most profitable requests that arrive in the near future after the admission of a request can be rejected due to unavailability recently allocated resources. In addition, most of these approaches do not consider the specific QoS requirements of each use case for the 5G network specification, nor the allocation of resources on core and edge nodes of the network. Several other solutions only consider resource allocation without considering admission control, thereby ignoring multiple providers' interests. In this thesis, two solutions are proposed for jointly conducting admission control and allocating resources for slicing 5G networks. The first solution is based on Reinforcement Learning, which allows the learning of the admission of network slice requests aiming at the profit of the providers. The second solution, based on Deep Reinforcement Learning, aims to further improve the achievement of the providers' objectives. The allocation of resources in these solutions is accomplished by a mapping of virtual nodes in the physical nodes of the network followed by a mapping of the virtual links in physical links. In these mappings, the QoS requirements of the eMBB, URLLC and MIoT classes of 5G technology are considered. The proposed solutions were evaluated for different traffic conditions and network topologies. The results of the evaluation corroborate that the proposals produce better results than the Always Admit Requests and Node Ranking heuristics, using profit and resource utilization as a comparison parameter. Results show the effectiveness of using Reinforcement Learning and Deep Reinforcement Learning techniques to manage network slice requests in 5G networks.
Examination Board
Headlines:
| Nelson Luis Saldanha da Fonseca | IC / UNICAMP |
| Christian Rodolfo Esteve Rothenberg | FEEC / UNICAMP |
| Edmundo Roberto Mauro Madeira | IC / UNICAMP |
Substitutes:
| Leandro Aparecido Villas | IC / UNICAMP |
| Marcelo Caggiani Luizelli | UNIPAMPA |