I am pleased to invite you to attend my PhD defense on 08/02/2023. The defense will start at 2h30 PM and it will be held at Hannah Arendt campus (in salle des thèses). I will be happy to see you there.
For those who are not able to attend, the BBB link is: https://v-au.univ-avignon.fr/live/bbb-soutenance-these-afaf-arfaoui-8-fevrier-2023/
Abstract: One of the fundamental components of wireless networks is Radio Access Network (RAN) delivering a wide reach wireless connectivity to the end users. The objective of RAN is to effectively manage and utilize the scarce spectrum to provide good connectivity to the end user. Over the past decade, the growth in usage of smartphones and handheld gadgets, as well as the surging adoption of smart vehicles and sensors, has caused a dramatic increase in the wireless traffic to be carried over the network. This traffic can be categorized into distinct use cases having varying requirements in terms of bandwidth, latency, reliability, etc., which are unlikely to be catered for by the current one-size-fits-all network infrastructure. Network Slicing (NS) has emerged as a promising architectural technology for building a highly flexible and dynamic network to meet the extremely diversified needs of use cases. The resulting network is much more nimble, flexible and scalable.
There is consensus that NS is a key enabler for the service-oriented 5G vision, that aims to cope with the increasing complexity of these networks. One of the major objectives of NS is to provide a different level of resource isolation, through resource abstraction and virtualization and the ability to efficiently share network resources. In the first part of this dissertation, we focus on video streaming traffic in the presence of other services with different Quality of Service (QoS) requirements. We propose a novel approach for resource sharing that provides inter-slice protection, flexibility, load-driven elasticity, and network efficiency. In particular, we design two-level multi-scale allocation schedulers for an efficient and low complexity RAN slicing by exploiting the characteristic of adaptive traffic such as video streaming service.
To enable NS in 5G networks, the underlying physical infrastructure should be virtualized by integrating Software-Defined Networking (SDN), Network Function Virtualization (NFV), and cloud computing. In a cloud environment, Virtual Network Functions (VNFs) are implemented as distributed open software applications running on machines in parallel to exploit I/O parallelism. Communication between the distributed computation tasks of these applications often result in heavy traffic transfers over the network. To tackle this issue, network researchers proposed variety of solutions to minimise flow completion times or to ensure per-flow fairness based on “flow” abstraction. However, this leads to a major discrepancy between the application goals on one hand, and network-level goals on the other, because the applications consider all their flows, but the network handles each flow independently which results in an overall loss of performance. To realign these goals, “coflow” abstraction was introduced with the aim to expose the network to the communication characteristics of applications by capturing the collective behavior of flows in distributed data-parallel applications. The main contribution in the second part of this thesis is to take advantage of the abstraction of coflows to propose a new scheduling algorithms to minimize the Coflow Completion Time (CCT). We then introduce a new rate allocation algorithim based on a distributed pricing mechanism that performs real-time rate allocation based on the actual congestion and data volumes trasferred by coflows. Furthermore, we introduce algorithms that is 2-approximation of the optimal, to make coflows meet their deadlines or, failing that, finishing as soon as possible after there deadlines, a problem commonly referred to as tardiness minimization. Previous work on tardiness minimization are based on flow abstractions, they only considers jobs running on parallel machines and provides and are m (n – 1)-approximation of the optimal. To the best of our knowledge this is the first work tackling the issue of tardiness minimization on a coflow-level. Our proposed algorithms are suitable for real environments since their implementation allows handling semi-clairvoyant and semi-distributed settings. They also yields near-optimal solutions. In fact, our scheduling algorithms are 2-approximation of the optimal and the rate allocation can achieve a 4-approximation.