Traffic engineering in the era of Software Defined Networks
One of the key advantages of Software-Defined Networks (SDN) is the opportunity to integrate traffic engineering mechanisms able to dynamically optimize the network configuration as the traffic evolves. Unfortunately, the network cannot be too frequently reconfigured to follow the traffic variation, since reconfigurations can affect route stability, increase the signaling overhead, negatively affect flows dynamics, etc.
In this talk, we discuss the problem of deciding whether, when and how to reconfigure the network according to the traffic evolution. To this end, we introduce the fundamental problem of clustering traffic matrices considering their similarities both in the traffic space and in the routing domain. As we show, the temporal dimension plays also an important role in the clustering problem in order to avoid frequent reconfigurations and to provide more flexibility to online decisions on when to reconfigure the network. Results show that with a limited set of network configurations we can achieve performance that are very close to the continuous and ideal network update.
Stefano Paris is Senior Researcher at the Mathematical and Algorithmic Sciences Lab of the Paris Research Center of Huawei Technologies Co. Ltd. Before joining Huawei, he has been Assistant Professor at the Department of Computer Science of Paris Descartes University. He received his M.S. degree in Computer Engineering from University of Bergamo in 2007, and the Ph.D. in Information Engineering from Politecnico di Milano in 2010. His main research interests include topics related to optimization and game theory for the evaluation of wireless and wired networks. He currently serves as editor of Computer Communications (Elsevier). He is a member of the IEEE.
Network function virtualization: some design and performance issues for network operators
The emergence of Software Defined Networking (SDN) and Network Function Virtualization (NFV) deeply modifies the way telecommunications networks are designed and operated. Dissociating network functions from their hosting hardware introduces more flexibility for network operators to instantiate network functions, where they are needed, and on demand. Several network functions are candidates to be virtualized, notably mobile packet core and radio access network (RAN) functions. In this keynote, we describe some issues related to these two sets of functions from a virtualization perspective. The virtualization of network functions with stringent latency requirements (notably RAN functions) incite network operators to deploy distributed data centers. On the basis of three level hierarchy of data centers (namely, Main Central Offices, Core Central Offices and centralized cloud platforms), we discuss in this keynote some design principles for the instantiation of network functions. On the basis of these principles, we describe various placement algorithms, some of them being fully distributed. Recognizing that virtualized network functions (VNFs) will compete with user’s applications, in particular Multi-access Edge Computing (MEC) applications, we introduce acceptance policies based on thresholds, which can be dynamically adapted. The performance of these policies are studied via simulation for a realistic network scenario.
Fabrice Guillemin graduated from Ecole Polytechnique in 1984 and from Telecom Paris in 1989. He received the PhD degree from the University of Rennes in 1992 and defended his « habilitation » thesis in 1999 at the University Pierre et Marie Curie (LIP6), Paris. Since 1989, he has been with Orange Labs in Lannion. He has been involved in the standardization of ATM and in several projects dealing with IP traffic metrology. He is currently leading a project on the evolution of control plane of networks (fixed and cellular). He is a member of the Orange Expert community « Network of the Future » (NoF).
Multi-Armed Bandits: a novel generic optimal algorithm and applications to networking
Multi-Armed Bandits (MABs) are a powerful formalism to describe sequential decision in an uncertain environnment: a learner observes an initially unknown system and selects actions sequentially in order to maximize an expected reward. MABs are particulary relevent in networking, since most networking problems (routing, scheduling, frequency assignment, flow control) require to take decisions in a noisy environment due to bursty traffic, unreliable media, interference etc. Also, in such problems, the environment may change rapidly and the number of decisions is large (e.g. the number of routes between a source and a destination) so that one must use the problem structure to learn as quicly as possible. We present a novel, completely generic algorithm which is applicable to any structured MAB problem, and which is provably asymptotically optimal. We further illustrate its application to structures commonly found in networking.
Richard Combes is currently an assistant professor in CentraleSupélec in the Telecommunication departement. He received the Engineering Degree from Telecom Paristech (2008), the Master’s Degree in Mathematics from university of Paris VII (2009) and the Ph.D. degree in Mathematics from university of Paris VI (2013). He was a visiting scientist at INRIA (2012) and a post-doc in KTH (2013). He received the best paper award at CNSM 2011. His current research interests are machine learning, networks and probability.