10:00 – 10:15  Accueil café 

10:15 – 10:45  Smart column generation solver for multi-commodity flow problems – Roberto Wolfler Calvo (LIPN), abstract , slides

10:45 – 11:15  Placement et Chainage de Fonctions de Services composées de Micro-Services : Modélisation et Résolution – Hichem MAGNOUCHE (UTT), abstract , slides

11:15 – 11:45  Robust network design and its application to slicing – Julien Lesca (Huawei), abstract , slides

11:45 – 12:15  Function Placement and Acceleration for In-Network Federated Learning Services – Nour El Houda YELLAS (CNAM), abstract

12:15 – 14:00  Pause Déjeuner

14:00 – 14:30  Deep Reinforcement Learning for network routing – A. Redha ALLICHE (I3S), abstract , slides

14:30 – 15:00  Graph Convolutional Reinforcement Learning for Collaborative Queuing Agents – Hassan Fawaz (TSP), abstract , slides

15:00 – 15:30  Monte Carlo Search Algorithms for Network Traffic Engineering – Chen Dang (Dauphine / Orange), abstract , slides

15:30 – 16:00  Pause Café

16:00 – 16:30  Scalable Damper-based Deterministic Networking – M. Yassine Naghmouchi (Huawei), abstract , slides

16:30 – 17:00  Enumération des stratégies optimales des opérateurs de réseaux mobile – Paolo Zappala (Orange), abstract , slides

17:00 – 18:00  Cocktail avec IEEE Cloudnet (amphi Pain Levé)

Smart column generation solver for multi-commodity flow problems
Roberto Wolfler Calvo (LIPN)

The multicommodity network flow problem (MFP) consists in routing a set of commodities through a capacitated network, from their source to their target, so that the capacity limits on the network arcs are respected and the overall commodity transportation cost is minimized. In the unsplittable version (UMFP) a commodity can only be routed through a single path from its source to its target. The UMFP is well known in the operations research community and has been studied for decades.

The UMFP is of practical relevance in applications spanning from transportation, e.g. routing cargos over railway or liner shipping networks, telecommunication and computer applications where commodities represent communication pairs including Internet and satellite networks. Even though the UMFP can be formulated and solved as a compact integer linear program, it is known that even computing the linear relaxation of this compact formulation may be intractable for very large networks. Rather, the linear relaxation is usually solved much more efficiently using a path-flow formulation and a delayed column generation (CG) method. However, there are still many unsolvable problems due to the large size of the instances and/or when addressing variants of UMFP such as, for example, taking in account uncertainty on data, non linear routing costs or fixed costs. The talk will present the litterature review and a first overwiev of the methods we would like to develop within a year and a half.

No, thank you. I do not want.
100% secure your website.
Powered by
Une Heuristique de Placement et Chainage de Fonctions de Services composées de Micro-Services
Hichem MAGNOUCHE (UTT)

La virtualisation des architectures en réseau présente de nombreux avantages et connaît un intérêt sans cesses croissant dans le domaine des télécommunications, notamment par le
déploiement massif de la 5G. Elle permet une plus grande flexibilité pour le déploiement des fonctions de réseau (NF), contribuant à l’émergeance de nouveaux services, à l’image de ceux
ceux nécessitant une très faible latence. Dans ce domaine, les concepts de micro-services et de programmabilité du réseau constituent des solutions prometteuses mais ils induisent de nouvelles problématiques concernant leur orchestration, et plus spécifiquement le placement des fonctions réseau et le chaînage des services. Après étudié les solutions de l’état de l’art dans
le domaine de l’orchestration de fonction réseaux virtualisées, nous avons proposé un modèle ILP qui tire profit des caractéristiques de l’approche micro-services afin de déployer des SFC
à faible latence [2]. Le modèle proposé produit des solutions optimales au bout de quelques minutes (en utilisant CPLEX) pour des instances de quelques noeuds et SFC seulement. Néanmoins, éprouver ce modèle sur des instances plus grandes en termes de noeuds et SFC s’avère impossible car le temps de résolution associé croît de manière exponentielle. Pour pallier cela, nous développons actuellement une méthode heuristique qui, en s’appuyant sur les caractéristiques de parallélisme des micro-services, produit des solutions proches de l’optimale en un
temps polynomial. L’heuristique repose sur l’algorithme d’Eppstein [1] pour calculer les k plus court chemins relatifs à une SFC puis déploie les micro-services relatifs à la SFC en tentant de
paralléliser ce qui peut être parallélisable. Dans cette présentation, nous présenterons l’avancée de ce travail à date et les premiers résultats qu’il a produit.

No, thank you. I do not want.
100% secure your website.
Powered by
Robust network design and its application to slicing
Julien Lesca (Huawei)

Network optimization plays a crucial role in telecommunication networks as it permits to carefully utilize and size infrastructures. In this talk, we consider the robust network design problem to efficiently size capacities in the context of network slicing. This problem consists in routing traffic between different pairs of nodes while taking into account the uncertainty on the traffic demands. In this context, we present algorithms for the design of network slices that guarantee a non-blocking property (all traffic can be routed) and end-to-end QoS constraints using an efficient solving method based on robust optimization.

No, thank you. I do not want.
100% secure your website.
Powered by
Function Placement and Acceleration for In-Network Federated Learning Services
Nour El Houda HYELLAS (CNAM)

Edge intelligence combined with federated learning is considered as a way to distributed learning and inference tasks in a scalable way, by analyzing data close to where it is generated, unlike traditional cloud computing where data is offloaded to remote servers.
In this talk, we address the placement of Artificial Intelligence Functions (AIF) making use of federated learning and hardware acceleration. We model the behavior of federated learning and related inference point to guide the placement decision, taking into consideration the specific constraint and the empirical behavior of a virtualized infrastructure anomaly detection use-case.
Besides hardware acceleration, we consider the specific training time trend when distributing training over a network, by using empirical piece-wise linear distributions. We model the placement problem as a MILP and we propose a variant of the problem.
Simulation results show the impact that hardware acceleration can have in the decision of the number of AIF to enable, while dividing by a relevant factor the distributed training time. We also show how our approach exacerbates the importance of monitoring an end-to-end learning system delay budget composed of link propagation delay and distributed training time in the location of AIFs.

No, thank you. I do not want.
100% secure your website.
Powered by
Deep Reinforcement Learning for network routing
A. Redha ALLICHE

In recent years, the emerging of 5G and IOT introduce a lot of challenges in term of networking infrastructure, such as the multiplication of the number of connections, the increase in throughput demand as well as stronger constraints in terms of quality of service. To support these new constraints, technologies such as Network Function Virtualization (NFV) and Software Defined Network (SDN) have been introduced. Many studies have been using artificial intelligence coupled on top of the latter technologies to enhance network capabilities and make networks’ devices decision smarter. In that context, reinforcement learning approaches were widely used, since many network problems can be modelled as a Markov Decision process. In this presentation, we will tackle the problem of network routing in two scenarios.
The first one is the distributed packet routing problem. In this problem, we present a multi-agent deep reinforcement approach inspired by the state of the art, and we study the impact of using such an approach in term of performance versus overhead [1]. The experiments are conducted on our open-source simulation ns-3-based module [2]. We formulate two signalling mechanisms between agents (value sharing and model sharing). We investigate the net gains considering in-band signalling and show that routing policies close to those provided by an oracle with full knowledge of traffic and network topology can be discovered with a control overhead of 150 % with respect to injected data packets, if neighboring agents share their Deep Neural Network models. We discuss the generality of our results to underline the importance of assessing net gains of Multi-Agent Deep Reinforcement Learning (MA-DRL)-based routing.
The second scenario concerns the routing and allocation of network slices in the context of Service Function Chaining (SFC). The goal of this study [3] is to find the best frequency to reconfigure the network (compute the paths and allocate services), as there is a trade- off between a reduced network congestion and the additional costs induced by the reconfiguration. For that, we use Deep Reinforcement Learning for decision and a Column Generation algorithm to compute the reconfiguration. We show that choosing the best time during the day to reconfigure allows to maximize the profit of the network operator while minimizing the use of network resources and the congestion of the network. Moreover, by selecting the best moment to reconfigure, our approach allows to decrease the number of needed reconfigurations compared to an algorithm doing periodic reconfigurations during the day.

No, thank you. I do not want.
100% secure your website.
Powered by
Graph Convolutional Reinforcement Learning for Collaborative Queuing Agents
Hassan Fawaz (TSP)

In this work, we explored the use of multi-agent deep learning, as well as learning to cooperate principles, to meet strict service level agreements, in terms of throughput and end-to-end delay, for a set of classified network flows. We consider agents built on top of a weighted fair queuing algorithm that continuously set weights for three flow groups: gold, silver, and bronze. We rely on a novel graph-convolution based, multi-agent reinforcement learning approach known as DGN. As benchmarks, we propose centralized and distributed deep Q network approaches and evaluate their performances in different network, traffic, and routing scenarios, highlighting the effectiveness of our proposals and the importance of agent cooperation. We simulate the proposals in an SD-WAN environment and show that our DGN-based approach meets stringent throughput and delay requirements across different scenarios.

No, thank you. I do not want.
100% secure your website.
Powered by
Monte Carlo Search Algorithms for Network Traffic Engineering
Chen Dang (Dauphine / Orange)

The aim of Traffic Engineering is to provide routing configurations in networks such that the used resources are minimized while maintaining a high level of quality of service (QoS). Among the optimization problems arising in this domain, we address in this paper the one related to setting weights in networks that are based on shortest path routing protocols (OSPF, IS-IS). Finding weights that induce efficient routing paths (e.g that minimize the maximum congested link) is a computationally hard problem. We propose to use Monte Carlo Search for the first time for this problem. More specifically we apply Nested Rollout Policy Adaptation (NRPA). We also extend NRPA with the force exploration algorithm to improve the results. In comparison to other algorithms NRPA scales better with the size of the instance and can be easily extended to take into account additional constraints (cost utilization, delay, …) or linear/non-linear optimization criteria. For difficult instances the optimum is not known but a lower bound can be computed. NRPA gives results close to the lower bound on a standard dataset of telecommunication networks.

No, thank you. I do not want.
100% secure your website.
Powered by
Scalable Damper-based Deterministic Networking
M. Yassine Naghmouchi (Huawei)

With 5G networking, deterministic guarantees are emerging as a key enabler.
In this context, we present a scalable Damper-based architecture for Large-scale Deterministic IP Networks (D-LDN) that meets required bounds on end-to-end delay and jitter. This work extends the original LDN architecture, where flows are shaped at ingress gateways and scheduled for transmission at each link using an asynchronous and cyclic opening of gate-controlled queues. To further relax the need for clock synchronization between devices, we use dampers, that consist in jitter regulators, to control the burstiness flows to provide a constant target delay at each hop. We introduce in details how data plane functionalities are implemented at all nodes (gateways and core) and we derive how the end-to-end delay and jitter are calculated. For the control plane, we propose a column generation algorithm to quickly take admission control decisions and maximize the accepted throughput. For a set of flows, it determines acceptance and selects the best shaping and routing policy. Through a proof-of-concept implementation in simulation, we verify that the architecture meets promised guarantees and that the control plane can operate efficiently at large-scale.

No, thank you. I do not want.
100% secure your website.
Powered by
Enumération des stratégies optimales des opérateurs de réseaux mobile
Paolo Zappala (Orange)

Les opérateurs de réseaux mobile planifient leurs investissements en choisissent des stratégies optimales. Dans le contexte de la théorie des jeux, cela consiste à énumérer les équilibres de Nash d’un jeu en forme extensive, un problème connu dans la littérature. On considère les jeux en forme extensive avec deux joueurs. Les stratégies d’un jeu sont souvent exponentiels en rapport avec les possibles réalisations du jeu. On introduit une méthode pour déterminer si chaque réalisation est généré par un équilibre de Nash. Cette méthode s’avère efficient dans le cas générique et est très performante quand le jeu présente des propriétés de régularité.

No, thank you. I do not want.
100% secure your website.
Powered by