CBL - Campus del Baix Llobregat

Projecte llegit

Títol: Design and Evaluation of AI-based resource prediction approaches for network slicing


Departament: ENTEL

Títol: Design and Evaluation of AI-based resource prediction approaches for network slicing

Data inici oferta: 15-02-2021     Data finalització oferta: 15-10-2021

Estudis d'assignació del projecte:
    MU MASTEAM 2015
Tipus: Individual
Lloc de realització: Fora UPC    
        Supervisor extern: Dr. Estefanía Coronado Calero
        Institució/Empresa: Fundació I2CAT
        Titulació del Director: PhD in Advanced Computer Technologies
Nom del segon director (UPC): Shuaib Siddiqui (i2Cat)
Departament 2n director:
Paraules clau:
5G, Network Slicing, AI
Descripció del contingut i pla d'activitats:
The intern will become a member of the Software Networks group at i2CAT. SN currently takes part in several H2020 projects developing 5G technologies, and the work of the intern would be related to one of these projects called AFFORDABLE5G.

The work of the intern will be structured in the following phases:

' Getting familiar with the concept of end-to-end network slicing, covering the resources from the RAN to the core network.

' Study the state in proactive network slicing, especially in what regards AI/ML time-series processes for the re-configuration and management of slice allocated resources to fulfill service requirements.

' Setup and deploy a full slice manager using VM or container-based environment and be able to collect periodic resource telemetry from the system.

' Design a ML or DL model that is able to use the telemetry information to provide proactive slice reconfiguration actions and anticipate to possible outages in the slices (e.g., the alarms actions can be, for instance, deploy extra VMs, or to enable a new small cell in the RAN).

' Design and perform controlled experiments on the design AI pipelines on a proof of concept, to validate the applicability of the approach.

' Compile the results in an academic publication.
Overview (resum en anglès):

The network slicing concept is fundamental for the development of the fifth generation (5G)
mobile networks. 5G approach is based in a high heterogeneity of the network creating
an increase in different Quality of Service (QoS). Additionally, with the growth in bandwidth,
the use of smaller antennas is mandatory, generating a higher densification of the
radio infrastructure and its associated transport layer. Due to an ever decreasing Average
Revenue Per User, and the fact that Mobile Network Operators (MNO) are becoming a
commodity, the investment cost of the MNO is increasing and the revenue is decreasing.
With this scenario in mind, taking full advantage of the deployed infrastructure is crucial
to guarantee an economical well being of the MNO. With this heterogeneous scenario in
mind, 5G architecture proposes the use of network slices as an isolated end-to-end network
that transports traffic with similar characteristics and usually with a given Service
Level Agreement (SLA). The main motivation for proposing this slicing isolation is to improve
the efficiency of the network utilization and enable the possibility to have ubiquitous
computing and networking. Network slicing is a logical segmentation of virtualized resources
that are deployed over physical and logical infrastructure and therefore are tied
to the capacity that said infrastructure is able to provide. However, the infrastructure is in
principle fixed while the slice traffic changes over time. For that reason, it is necessary to
provide a means so that the Network Service Provider (NSP) is able to take the corrective
measurements and guarantee the agreed SLA.
The present Master¿s Thesis (MT) objective is to develop two Machine Learning (ML) models
that allows a network operator to forecast the future slicing traffic based on current and
historical data from the network and adapt the resources accordingly. The basic mechanics
of ML are presented, including the mathematical description and the advantages and
disadvantages of each type. Two models are chosen: Long-Short Time Memory (LSTM)
and random forest. The models are then tested using a network simulator (NS3) using
four different scenarios. An additional plug-in is then used (NS3-AI) in order to be able
to access the NS3 simulation in real time and then it is changed so it can incorporate a
bidirectional exchange of communication between the simulator and the predictive model.
After the simulation is finished, the results from both forecasting models are compared in
various scenarios, where LSTM is outperformed by random forest in terms of adaptability
and metric errors.

© CBLTIC Campus del Baix Llobregat - UPC