PhD defence: Content and Resource Management in Edge Networks
Place: Online at Zoom and Lecture hall E:B, building E, Ole Römers väg 3, Faculty of Engineering LTH, Lund University, Lund
Contact: Bjorn [dot] Landfeldt [at] eit [dot] lth [dot] se
Save event to your calendar
Author: Mohammadhassan Safavi
Faculty Opponent: Professor Paolo Bellavista, University of Bologna Italy.
Abstract: In this thesis, we investigate and develop new methods for efficient and functional use of resources in edge networks. Setting this work aside from previous work, we study User Generated Content (UGC) such as social media information and data generated in the new emerging Internet of Things systems. We present efficient solutions for placing such content and managing which network resources should be used to make the edge networks effective. By effective we for example mean; using little energy, processing data with short delay or carrying out their tasks with little load on the network. In order to achieve this, we have used a range of optimization and control theoretic tools and studied different aspects of content and resource management in operator managed content distribution networks (CDN). The main parts of the contributions of the thesis can be summarized as follows:
First, we have studied end-to-end energy usage in video delivery systems. We studied the energy usage of a sample video considering separate delivery components and created a model for overall energy usage when delivering video over the Internet. The study comprises experimental and simulated measurements of encoding with different qualities, transmissions over core and wireless access networks and decoding in user devices. We showed how video popularity affects end-to-end energy usage by codec selection.
Second, we proposed optimal and on-line placement algorithms for content placement at the edge. We focused on UGC, considering its distributed bottom-up trajectory pattern. ISP-managed CDNs are considered to be suitable caching hosts of popular UGCs. Furthermore, we proposed on-line learning algorithms to enable decision agents at the edge to predict content popularity from users' social activities.
Third, we took the data center viewpoint of a delivery system. We designed scheduling and request assignment algorithms with an energy usage objective. We showed that an energy-efficient dynamic server provisioning (DSP)-based assignment may lead to an unstable system if sufficient care has not be taken. We then investigated ways of keeping the servers stable, energy efficient and performing load balancing to provide better quality of service (QoS) for end users.
Fourth, we expanded the idea of edge placement in an IoT service offloading context. We investigated the service placement problem in a distributed 5G F-RAN (fog radio access network) architecture with an existing centralized cloud. We proposed optimal and reinforcement learning based algorithms to perform joint service scheduling and placement in fog-cloud hosts based on a utilization objective. We showed that the learning algorithm converges to an optimal policy when there are uncertainties in positioning and service demand parameters.