Section: New Results

Future networks and architectures

Participants : Yassine Hadjadj-Aoul, Gerardo Rubino, Quang Pham Tran Anh, Anouar Rkhami.

Machine learning for network slicing. Network Function Virtualization (NFV) provides a simple and effective mean to deploy and manage network and telecommunications' services. A typical service can be expressed in the form of a Virtual Network Function-Forwarding Graph (VNF-FG). Allocating a VNF-FG is equivalent to placing VNFs and virtual links onto a given substrate network considering resources and quality of service (QoS) constraints. The deployment of VNF-FGs in large-scale networks, such that QoS measures and deployment cost are optimized, is an emerging challenge. Single-objective VNF-FGs allocation has been addressed in existing literature; however, there is still a lack of studies considering multi-objective VNF-FGs allocation. In addition, it is not trivial to obtain optimal VNF-FGs allocation due to its high computational complexity even in the single-objective case. Genetic algorithms (GAs) have proved their ability in coping with multi-objective optimization problems, thus we propose, in [26], a GA-based scheme to solve multi-objective VNF-FGs allocation problem. The numerical results confirm that the proposed scheme can provide near Pareto-optimal solutions within a short execution time.

In [25], we explore the potential of deep reinforcement learning techniques for the placement of VNF-FGs. However, it turns out that even the most well-known learning technique is ineffective in the context of a large-scale action space. In this respect, we propose approaches to find out feasible solutions while improving significantly the exploration of the action space. The simulation results clearly show the effectiveness of the proposed learning approach for this category of problems. Moreover, thanks to the deep learning process, the performance of the proposed approach is improved over time.

The placement of services, as described above, is extremely complex. The issue is even more complex when it comes to placing a service on several non-cooperative domains, where the network operators hide their infrastructure to other competing domains. In [56], we address these problems by proposing a deep reinforcement learning based VNF-FG embedding approach. The results provide insights into the behaviors of non-cooperative domains. They also show the efficiency of the proposed VNF-FG deployment approach having automatic inter-domain load balancing.

Consistent QoS routing in SDN networks. The Software Defined Networking (SDN) paradigm proposes to decouple the control plane (decision-making process) and the data plane (packet forwarding) to overcome the limitations of traditional network infrastructures, which are known to be difficult to manage, especially at scale. Although there are previous works focusing on the problem of Quality of Service (QoS) routing in SDN networks, only few solutions have taken into consideration the network consistency, which reflects the adequacy between the decisions made and the decisions that should be taken. Therefore, we propose, in [19], a network architecture that guarantees the consistency of the decisions to be taken in an SDN network. A consistent QoS routing strategy is, then, introduced in a way to avoid any quality degradation of prioritized traffic, while optimizing resources usage. Thus, we proposed a traffic dispersion heuristic in order to achieve this goal. We compared our approach to several existing framework in terms of best-effort flows average throughput, average video bitrate and video Quality of Experience (QoE). The emulations results, which are performed using the Mininet environment, clearly demonstrate the effectiveness of the proposed methodology that outperforms existing frameworks.

Optical networks. In [20] we attack the so called Capacity Crunch crisis announced for optical networks infrastructures. This problem refers to the facts that (i) the transmission capacity of an optical fiber is not limitless, (ii) the bandwidth demand continues to increase exponentially and (iii) the limits are getting dangerously close. The cheapest and shortest-term solution is to increase efficiency, because there are several possibilities to do so. This work is a contribution in that direction. We focus on strongly improving the wavelength assignment procedure by moving to an heterogeneous and flexible process, adapting the dimensioning to the individual users' needs in QoS. In the paper we demonstrate that a non-uniform dimensioning strategy and a tighten QoS provision allows to save significant networks capacity, while simultaneously provisioning to each user the QoS established in its Service Level Agreement.

Survivability of internet services is a significant and crucial challenge in designing future optical networks. A robust infrastructure and transmission protocols are needed to handle such a situation so that the users can maintain communication despite the existence of one or more failed components in the network. For this reason, we present in [40] a generalized approach able to tolerate any failure scenario, to the extent the user can still communicate with the remaining components, where a scenario corresponds to an arbitrary set of links in a non-operational state. To assess the survivability problem, we propose a joint solution to the problems listed next. We show how to find a set of primary routes, a set of alternate routes associated with each failure scenario, and the capacity required on the network to allow communication between all users, in spite of the links' failures, while satisfying for each user a specific predefined quality of service threshold, defined in the Service Level Agreement (SLA). Numerical results show that the proposed approach not only enjoys the advantages of low complexity and ease of implementation but is also able to achieve significant resource savings compared to existing methods. The savings are higher than 30% on single link failures and more than a 100% on two simultaneous link failures scenarios as well as in more complex situations.

Network tomography. Internet tomography studies the inference of the internal network performances from end-to-end measurements. For this problem, Unicast probing can be advantageous due to the wide support of unicast and the easy deployment of unicast probing paths. In [48] we propose two statistical generic methods for the inference of additive metrics using unicast probing. Our solutions give more flexibility in the choice of the collection points placement. Moreover, the probed paths are not limited to specific topologies. Firstly, we propose the $k$-paths method that extends the applicability of a previously proposed solution called Flexicast for tree topologies. It is based on the Expectation-Maximization (EM) algorithm which is characterized by high computational and memory complexities. Secondly, we propose the Evolutionary Sampling Algorithm (ESA) that enhances the accuracy and the computing time but following a different approach. In [49] we present a different approach, targeted at link metrics inference in an SDN/NFV environment (even if it can be exported outside this field) that we called TOM (Tomography for Overlay networks Monitoring). In such an environment, we are particularly interested in supervising network slicing, a recent tool enabling to create multiple virtual networks for different applications and QoS constraints on a Telco infrastructure. The goal is to infer the underlay resources states from the measurements performed in the overlay structure. We model the inference task as a regression problem that we solve following a Neural Network approach. Since getting labeled data for the training phase can be costly, our procedure generates artificial data instead. By creating a large set of random training examples, the Neural Network learns the relations between the measures done at path and link levels. This approach takes advantage of efficient Machine Learning solutions to solve a classic inference problem. Simulations with a public dataset show very promising results compared to statistical-based methods.We explored mainly additive metrics such as delays or logs of loss rates, but the approach can also be used for non-additive ones such as bandwidth.