This research explores serveral topics related to the performance achievable with the introduction of (system level) virtualization in modern data centers. Virtualization can increase the task of existing protocols, e.g. the transport layer protocol. It also brings new opportunities with fast packet processing solutions (e.g. DPDK, XDP, etc.) that might improve the performance of cloud services and applications, eg. data streaming applications ; and 2) the ability to solve specific problems related to the optimization of ressources (esp RAM and CPU) within the data centers, e.g. during maintenance events.
Scientific leaders: Guillaume Urvoy-Keller, Dino Martin Lopez-Pacheco
Energy is becoming a great concern. Networks themselves are apparently responsible for up to one third of the total IT consumption. We are tackling this issue along two axes. First, energy consumption in ISPs’ networks and end-to-end energy consumption. Here we focus either on evaluating green protocols or measuring energy consumption due to network activities through experiments or simulations. Second, we also explore the use of SDN (Software Defined Networks) as a enabler for reducing energy consumption of data center networks through a better management of traffic.
Scientific leaders: Dino Martin Lopez-Pacheco, Guillaume Urvoy-Keller.
This reserach aims to develop new operating frameworks where network monitoring and machine learning are integrated to automate the network control. The idea behind is to learn more efficient control agorithms from the exploitation of network telemetry data.
Scientific leader: Ramon Aparicio-Pardo
Machine learning for attention prediction in virtual reality, streaming control and multiledia content analysisWe design machine learning models to predict the movements of a person immersed in virtual reality content, and inform dynamic optimization strategies in streaming logics (which quality to transmit to which location in the sphere, which 360° film editing cuts to trigger). Miguel Romero's PhD thesis (2017-2021) achieved significant results with multimodal recurrent deep neural network architectures. Quentin Guimard's thesis aims to quantify the uncertainty of motion prediction to better inform control logic, with variational and emotional state aware approaches. Franz Franco Gallo's thesis extends this work conducted in 3 degrees of freedom to 6 degrees of freedom of motion, requiring 3D content analyses. The analysis of multimodal time series is also developed in the context of the European project AI4Media and the ANR project TRACTIVE coordinated by Lucile Sassatelli, the latter aiming at the analysis of film editing patterns to study the representation of characters according to their gender.
Scientific leader: Lucile Sassatelli
Current wireless networks use the radio spectrum in a very inefficient manner. Indeed, a frequency bandwidth is allocated in a fixed manner to the user, even when this user is not using the radio resource. This type of user is denoted as a “primary” user. When the primary user is not active, a “secondary” user could take advantage of the free spectrum: this is the idea behind cognitive radio. In this activity, the objective is to allow a primary user and a secondary user to transmit simultaneously. Indeed, if the interference caused by the secondary user is somehow known by the primary user, the latter can null out this interference and continue to transmit with the same throughput. To be able to perform this interference cancellation, the different users must share some information (in the extreme case, share all the transmitted messages). The goal of this activity is to explore the cooperation strategies (which information have to be shared) to enable the development of cognitive radio networks where both primary and secondary users transmit simultaneously.
Scientific leader: Luc Deneire
In an increasingly digitized and interconnected world, the volume of multidimensional, multimodal, heterogeneous and often incomplete data to be processed keeps growing at an unprecedented rate. In this context, tensors play today a central role in many fields of application, particularly in signal processing and machine learning, for the representation, compression, analysis, fusion and classification of data.
Our objective is to develop new structured and coupled tensor models, with specific algorithms for parameter estimation of large-scale data tensors. Since two decades, one of our main motivations is to design new tensor-based approaches for wireless communications. Future research work will mainly concern cooperative (multi-relay) massive MIMO millimeter-wave cellular systems, which open up new challenges for mobile communications, with in particular the consideration of reconfigurable intelligent surfaces (RIS) and massive MIMO reconfigurable antennas. These works are carried out in collaboration with several brazilian teams.
Scientific leader: Gérard Favier