Search

International audience; Distributed network emulators allow users to perform network evaluation by running large-scale virtual networks over a cluster of fewer machines. While they offer accessible testing environments for researchers to evaluate their contributions and for the community to reproduce its results, their use of limited physical network and compute resources can silently and negatively impact the emulation results. In this paper, we present a methodology that uses linear optimization to extract information about the physical infrastructure from emulation-level packet delay measurements, in order to pinpoint the root causes of emulation inaccuracy with minimal hypotheses. We evaluate the precision of our methodology using numerical simulations, then show how its implementation performs in a real network scenario.

International audience; Distributed Network emulators (e.g., Mininet Cluster Edition) have proven to be an attractive solution to perform extreme-scale network and systems evaluation on smaller-size testbeds and experiment platforms. They can provide contained, customisable, and scalable testing environments for researchers to evaluate their contributions and reproduce their results. The major drawback of this approach in network experimentation is the use of virtual components (hosts, network switches, etc.) that do not behave with perfect similarity to the physical components they emulate, mainly due to the concurrency in using the underlay network and computing resources. We thus present in this paper a methodology to monitor emulation fidelity by measuring the network delays of emulated packets, which relies on statistical metrics to evaluate their inaccuracy. We further dig into the possible sources of emulation inaccuracy and show how our system can detect them to avoid biased experiment results. We particularly show through a common experiment scenario how undetected network emulation errors can lead to biased results.

International audience; Ray Tracing is an electromagnetic wave propagation modeling approach used for accurate generation of Quality of Service (QoS) maps in mobile networks. Due to its complexity, current implementation of Ray Tracing fails to generate such maps in wide areas. In this paper, we propose an optimization to Ray Tracing able to accurately generate QoS maps in a reasonable time. Using a site-specific ray launching technique and an alternative to the reception test process, we divide by almost 1200 the execution time of Ray Tracing with less than 2% of memory usage as compared to baseline solutions.

International audience; 5G enhanced Mobile broadband (eMBB) aims to provide users with a peak data rate of 20 Gbps in the Radio Access Network (RAN). However, since most Congestion Control Algorithms (CCAs) rely on startup and probe phases to discover the bottleneck bandwidth, they cannot quickly utilize the available RAN bandwidth and adapt to fast capacity changes without introducing large delay increase, especially when multiple flows are sharing the same Radio Link Control (RLC) buffer. To tackle this issue, we propose RAPID, a RAN-aware proxy-based flow control mechanism that prevents CCAs from overshooting more than the available RAN capacity while allowing near optimal link utilization. Based on analysis of up-to-date radio information using Multi-access Edge Computing (MEC) services and packet arrival rates, RAPID is able to differentiate slow interactive flows from fast download flows and allocate the available bandwidth accordingly. Our simulation and experimentation results with concurrent Cubic and BBR flows show that RAPID can reduce delay increase by a factor of 10 to 50 in both Line-of-Sight (LOS) and Non-LOS (NLOS) conditions while preserving high throughput in both 4G and 5G environments.

International audience; The design and development of new network protocols, architectures, and technologies require an evaluation phase where the researcher must provide empirical evidence for the performance of their contributions. In this context, network emulation has proven to be an attractive approach as it offers more flexibility compared to traditional testing platforms, and more realism compared to simulation software. However, as emulation requires running multiple concurrent virtual network components on a limited amount of resources, it faces welldocumented limitations in its scalability and accuracy. Our work is an attempt to address these issues by developing a lightweight and highly scalable distributed emulator, and by designing a framework for fidelity monitoring through passive measurement of emulated packet delays. The preliminary evaluation shows promising results and demonstrates the high potential of our approach.

International audience; New cellular technologies combine complex software stacks and heterogeneous hardware. As a result, even if models and simulators are still essential to understand them, validation must go to the next level and rely on real software and hardware testbeds. Unfortunately, building comprehensive testbeds is expensive and this is why we propose the Sophianode, a cloud-native cellular network testbed accessible to the community. In this demonstration, we show how to run arbitrary *G experiments in this testbed by following DevOps techniques.

The GIM2 phase III trial demonstrated the benefit of dose-dense chemotherapy in node-positive early breast cancer (eBC). To better define the dose-dense effect in the hormone receptor-positive subgroup, we evaluated its benefit through a composite measure of recurrence risk. We conducted an ancillary analysis of the GIM2 trial evaluating the absolute treatment effect through a composite measure of recurrence risk (CPRS) in patients with hormone receptor-positive HER2-negative eBC. CPRS was estimated through Cox proportional hazards models applied to the different clinicopathological features. The treatment effect was compared to the values of CPRS by using the Sub-population Treatment Effect Pattern Plot (STEPP) process. The Disease-Free Survival (DFS)-oriented STEPP analysis showed distinct patterns of relative treatment effect with respect to CPRS. Overall, 5-year DFS differed across CPRS quartiles ranging from 95.2 to 66.4%. Each CPRS quartile was characterized by a different patients’ composition, especially for age, lymph node involvement, tumor size, estrogen and progesterone receptor expression, and Ki-67. A number needed to treat of 154 and 6 was associated with the lowest and the highest CPRS quartile, respectively. Dose-dense adjuvant chemotherapy showed a consistent benefit in node-positive eBC patients with hormone receptor-positive HER2-negative disease, but its effect varied according to CPRS.

Purpose Triple-negative breast cancer (TNBC) represents a subtype of breast cancer which lacks the expression of oestrogen receptor (ER), progesterone receptor (PR) and human epidermal growth factor receptor-2 (HER2): TNBC accounts for approximately 20% of newly diagnosed breast cancers and is associated with younger age at diagnosis, greater recurrence risk and shorter survival time. Therapeutic options are very scarce. Aim of the present analysis is to provide further insights into the clinical activity of metronomic chemotherapy (mCHT), in a real-life setting. Methods We used data included in the VICTOR-6 study for the present analysis. VICTOR-6 is an Italian multicentre retrospective cohort study, which collected data of metastatic breast cancer (MBC) patients who have received mCHT between 2011 and 2016. Amongst the 584 patients included in the study, 97 were triple negative. In 40.2% of the TNBC patients, mCHT was the first chemotherapy treatment, whereas 32.9% had received 2 or more lines of treatment for the metastatic disease. 45.4% out of 97 TNBC patients received a vinorelbine (VRL)-based regimen, which resulted in the most used type of mCHT, followed by cyclophosphamide (CTX)-based regimens (30.9%) and capecitabine (CAPE)-based combinations (22.7%). Results Overall response rate (ORR) and disease control rate (DCR) were 17.5% and 64.9%, respectively. Median progression free survival (PFS) and overall survival (OS) were 6.0 months (95% CI: 4.9–7.2) and 12.1 months (95% CI: 9.6–16.7). Median PFS was 6.9 months for CAPE-based regimens (95% CI: 5.0–18.4), 6.1 months (95% CI: 4.0–8.9) for CTX-based and 5.3 months (95% CI: 4.1–9.5) for VRL-based ones. Median OS was 18.2 months (95% CI: 9.1-NE) for CAPE-based regimens and 11.8 months for VRL- (95% CI: 9.3–16.7 and CTX-based ones (95%CI: 8.7–52.8). Tumour response, PFS and OS decreased proportionally in later lines. Conclusion This analysis represents the largest series of TNBC patients treated with mCHT in a real-life setting and provides further insights into the advantages of using this strategy even in this poor prognosis subpopulation.

International audience; Emulation has become a popular approach for the validation and evaluation of network research. It provides researchers with a contained, customizable, and scalable testing environment, which can be easily packaged and published for potential readers to reproduce their results. However, as the network components are only virtual, emulation lacks the inherent realism of physical testbeds. In light of this, monitoring specific metrics of the emulated network has been proposed as a solution to mitigate to some degree inaccuracies caused by emulation. While this is not difficult to implement in a single-machine setting (e.g. with Mininet), monitoring is limited by the lack of time synchronization in scenarios where the emulation is distributed over multiple physical machines (e.g., Distrinet). In this paper we tackle the case of packet delay monitoring, to which we propose a methodology for passively measuring one-way delays with underlying assumptions about time synchronization, and round-trip delays otherwise. For an efficient implementation of our methodology, we propose an eBPF-based packet measurement tool that performs better than current packet sniffers under emulation-specific assumptions. We implement and evaluate our system in an open testbed and show that it can reach results within few microseconds of perfect accuracy and precision.

International audience; Traditional loss-based Congestion Control Algorithms (CCAs) suffer from performance issues over wireless networks mostly due to their inability to distinguish wireless random losses from congestion losses. Different loss discrimination algorithms have been proposed to tackle this issue but they are not efficient for 4G networks since they do not consider the impact of various link layer mechanisms such as adaptive modulation and coding and retransmission techniques on congestion in LTE Radio Access Networks (RANs). We propose MELD (MEC-based Edge Loss Discrimination), a novel server-side loss discrimination mechanism that leverages recent advancements in Multi-access Edge Computing (MEC) services to discriminate packet losses based on real-time RAN statistics. Our approach collects the relevant radio information via MEC's Radio Network Information Service and uses it to correctly distinguish random losses from congestion losses. Our experimental study made with the QUIC transport protocol shows over 80% higher goodput when MELD is used with NewReno and 8% higher goodput when used with Cubic.

International audience; 5G enhanced Mobile broadband (eMBB) aims to provide users with a peak data rate of 20 Gbps in the Radio Access Network (RAN). However, since most Congestion Control Algorithms (CCAs) rely on startup and probe phases to discover the bottleneck bandwidth, they cannot quickly utilize the available RAN bandwidth and adapt to fast capacity changes without introducing large delay increase, especially when multiple flows are sharing the same Radio Link Control (RLC) buffer. To tackle this issue, we propose RAPID, a RANaware proxy-based flow control mechanism that prevents CCAs from overshooting more than the available RAN capacity while allowing near optimal link utilization. Based on analysis of up-todate radio information using Multi-access Edge Computing (MEC) services and packet arrival rates, RAPID is able to differentiate slow interactive flows from fast download flows and allocate the available bandwidth accordingly. Our experiments with concurrent Cubic and BBR flows show that RAPID can reduce delay increase by a factor of 10 to 50 in both Line-of-Sight (LOS) and Non-LOS (NLOS) conditions while preserving high throughput.

This deliverable aims to organise the discussions regarding the reference architecture of SLICES-RI. In this document, different scenarios are analysed and presented as a baseline for the future discussions regarding the SLICES-RI’s reference architecture. Other scenarios, in addition to the above-mentioned ones, might be considered by the partners in the future.

International audience; Many companies and organizations are moving their applications from on-premises data centers to the cloud. The cloud infrastructures can potentially provide an infinite amount of computation (e.g., Elastic Compute) and storage (e.g., Simple Service Storage). In addition, all cloud providers propose different offers: IaaS, PaaS, and SaaS. This demo focuses on the IaaS services, presenting a simple tool to measure the network delay in a virtual infrastructure built entirely in the cloud. These measurements are useful for organizations that are moving current applications to, or creating new applications in, the cloud, but have requirements on the maximum, or average, network delay that these applications can tolerate. We present CloudTrace, a simple CLI tool that creates regional and multiregional experiments to measure delay, using Amazon AWS.

International audience; To handle the ever growing demand of resource intensive experiments distributed, network emulation tools such as Mininet and Maxinet have been proposed. They automatically allocate experimental resources. In this work, we show that resources are poorly allocated, leading to resource overloading and hence to dubious experimental results. This is why we propose and implement a new placement module for distributed emulation. Our algorithms take into account both link and node resources and minimize the number of physical hosts needed to carry out the emulation. Through extensive numerical evaluations, simulations, and actual experiments , we show that our placement methods outperform existing ones and allowing to re-establish trust in experimental results.

International audience; More than ever, data networks have demonstrated their central role in the world economy, but also in the well-being of humanity that needs fast and reliable networks. In parallel, with the emergence of Network Function Virtualization (NFV) and Software Defined Networking (SDN), efficient network algorithms considered too hard to be put in practice in the past now have a second chance to be considered again. In this context, as new networks will be deployed and current ones get significant upgrades, it is thus time to rethink the network dimensioning problem with protection against failures. In this paper, we consider a path-based protection scheme with the global rerouting strategy in which, for each failure situation, there may be a new routing of all the demands. Our optimization task is to minimize the needed amount of bandwidth. After discussing the hardness of the problem, we develop two scalable mathematical models that we handle using both Column Generation and Benders Decomposition techniques. Through extensive simulations on real-world IP network topologies and on randomly generated instances, we show the effectiveness of our methods: they lead to savings of 40 to 48% of the bandwidth to be installed in a network to protect against failures compared to traditional schemes. Finally, our implementation in OpenDaylight demonstrates the feasibility of the approach. Its evaluation with Mininet shows that our solution provides sub-second recovery times, but the way it is implemented may greatly impact the amount of signaling traffic exchanged. In our evaluations, the recovery phase requires only a few tens of milliseconds for the fastest implementation, compared to a few hundreds of milliseconds for the slowest one.

Internet Coordinate Systems, shortly ICS, (e.g. [9] [8]) have been proposed to allow for distance (Round-Trip Time, shortly RTT) estimation between nodes, in order to reduce the measurement overhead of many applications and overlay networks. Indeed, by embedding the Internet delay space into a metric space – an operation that only requires each node in the system to measure delays to a small set of other nodes (its neighbors), nodes are attributed coordinates that can then be used to estimate the RTT between any two nodes, without further measurements, simply by applying the distance function associated with the chosen metric space to the nodes’ coordinates. Recent works have shown how coordinate-embedding services could be vulnerable to malicious attacks, providing a potentially attractive fertile ground for the disruption or collapse of the many applications and overlays that would use these services [2]. There are actually two obvious ways to disrupt the operation of a coordinate based system. First when requested to give its coordinate for a distance estimation at the application-level, a malicious node could simply and blatantly lie. Second, a malicious node, or even a colluding group, may aim at disrupting the embedding process itself. This latter strategy is very insidious and effective as it can result in important distortions of the coordinate space which then spoils the coordinate computations of many nodes (malicious and honest alike) [2]. This chapter focuses on developing and studying generic Kalman filter-based methods to secure the coordinate embedding process. More precisely, the embedding process, regardless of the actual coordinate-based positioning system, works on the premise that nodes adjust their coordinate based on some comparison between measured and estimated distances to some other nodes. Malicious nodes can interfere with this embedding process by, amongst other things, lying about their real coordinate and/or tampering with measurement probes, to create a discrepancy between measured and estimated latencies, so that unsuspecting nodes would wrongly adjust their own coordinate in a bid to reduce the difference [1]. Because the load on the network naturally varies in time, so does latency between pair of nodes, and as a Source: Kalman Filter, Book edited by: Vedran Kordic, ISBN 978-953-307-094-0, pp. 390, May 2010, INTECH, Croatia, downloaded from SCIYO.COM

International audience; In this paper, we introduce a novel dynamic controller assignment algorithm targeting connected vehicle services and applications, also known as Internet of Vehicles (IoV). The proposed approach considers a hierarchically distributed control plane, decoupled from the data plane, and uses vehicle location and control traffic load to perform controller assignment dynamically. We model the dynamic controller assignment problem as a multi-agent Markov game and solve it with cooperative multi-agent deep reinforcement learning. Simulation results using real-world vehicle mobility traces show that the proposed approach outperforms existing ones by reducing control delay as well as packet loss. Index Terms-Internet of Vehicles (IoV), Software Defined Networking (SDN), multi-agent deep reinforcement learning, controller assignment.

With the increased complexity of today’s networks, emulation has become an essential tool to test and validate a new proposed networking solution. As these solutions also become more and more complex with the introduction of softwarization, network function virtualization, andartificial intelligence, there is a need of scalable tools to carry out resource intensive emulations. To this end, distributed emulation has been proposed. However, distributing a network emulation over a physical platform requires to choose carefully how the experiment is run over the equipment at disposal. In this work, we evaluate the placement algorithms which were proposed for, and implemented in, existing distributed emulation tools. We show that they may lead to bad placements in which several hardware resources such as link bandwidth, CPU, and memory are overloaded. Through extensive experiments, we exhibit the impact of such placements on important network metrics such as real network bandwidth usage and emulation execution time, and show that they may lead to unreliable results and to a waste of platform resources.To deal with this issue, we propose and implement a new placement module for distributed emulation. Our algorithms take into account both link and node resources and minimize the number of physical hosts needed to carry out the emulation. Through extensive numerical evaluations, simulations, and experiments, we show that our placement methods outperform existing ones leading to reliable experiments using a minimum number of resources.; Avec la complexité croissante des réseaux actuels, l’émulation est devenue un outil essentiel pour tester et valider une nouvelle solution réseau.Comme ces solutions deviennent également de plus en plus complexes avec l’introduction des réseaux logiciels, de la virtualisation des fonctions réseau et de l’intelligence artificielle, il est nécessaire de disposer d’outils qui passent à l’échelle pour réaliser des émulations intensives en ressources. A cette fin, il a été proposé de distribuer les émulations. Cependant, la distribution d’une émulation de réseau sur une plate-forme physique nécessite de choisir avec soin la manière dont l’expérience est menée sur l’équipement à disposition. Dans ce travail, nous évaluons les algorithmes de placement qui ont été proposés et implémentés dans les outils d’émulation distribuée existants. Nous montrons qu’ils peuvent conduire à de mauvais placements dans lesquels plusieurs ressources matérielles telles que la bande passante de liens, le processeur et la mémoire peuvent être surchargés. Grâce à des expériences approfondies, nous montrons l’impact de ces placements sur des paramètres importants du réseau tels que l’utilisation réelle de la bande passante et le temps d’exécution de l’émulation, et nous montrons qu’ils peuvent conduire à des résultats peu fiables et `a un gaspillage des ressources de la plate-forme. Pour traiter cette question, nous proposons et implémentons un nouveau module de placement pour l’émulation distribuée. Nos algorithmes prennent en compte les ressources des liens et des nœuds et minimisent le nombre d’hôtes physiques nécessaires pour réaliser l’émulation. Grâce à des évaluations numériques, des simulations et des expériences poussées, nous montrons que nos méthodes de placement sont plus performantes que les méthodes existantes, ce qui permetde réaliser des expériences fiables en utilisant un nombre minimal de ressources.

With the increased complexity of today’s networks, emulation has become an essential tool to test and validate a new proposed networking solution. As these solutions also become more and more complex with the introduction of softwarization, network function virtualization, andartificial intelligence, there is a need of scalable tools to carry out resource intensive emulations. To this end, distributed emulation has been proposed. However, distributing a network emulation over a physical platform requires to choose carefully how the experiment is run over the equipment at disposal. In this work, we evaluate the placement algorithms which were proposed for, and implemented in, existing distributed emulation tools. We show that they may lead to bad placements in which several hardware resources such as link bandwidth, CPU, and memory are overloaded. Through extensive experiments, we exhibit the impact of such placements on important network metrics such as real network bandwidth usage and emulation execution time, and show that they may lead to unreliable results and to a waste of platform resources.To deal with this issue, we propose and implement a new placement module for distributed emulation. Our algorithms take into account both link and node resources and minimize the number of physical hosts needed to carry out the emulation. Through extensive numerical evaluations, simulations, and experiments, we show that our placement methods outperform existing ones leading to reliable experiments using a minimum number of resources.; Avec la complexité croissante des réseaux actuels, l’émulation est devenue un outil essentiel pour tester et valider une nouvelle solution réseau.Comme ces solutions deviennent également de plus en plus complexes avec l’introduction des réseaux logiciels, de la virtualisation des fonctions réseau et de l’intelligence artificielle, il est nécessaire de disposer d’outils qui passent à l’échelle pour réaliser des émulations intensives en ressources. A cette fin, il a été proposé de distribuer les émulations. Cependant, la distribution d’une émulation de réseau sur une plate-forme physique nécessite de choisir avec soin la manière dont l’expérience est menée sur l’équipement à disposition. Dans ce travail, nous évaluons les algorithmes de placement qui ont été proposés et implémentés dans les outils d’émulation distribuée existants. Nous montrons qu’ils peuvent conduire à de mauvais placements dans lesquels plusieurs ressources matérielles telles que la bande passante de liens, le processeur et la mémoire peuvent être surchargés. Grâce à des expériences approfondies, nous montrons l’impact de ces placements sur des paramètres importants du réseau tels que l’utilisation réelle de la bande passante et le temps d’exécution de l’émulation, et nous montrons qu’ils peuvent conduire à des résultats peu fiables et `a un gaspillage des ressources de la plate-forme. Pour traiter cette question, nous proposons et implémentons un nouveau module de placement pour l’émulation distribuée. Nos algorithmes prennent en compte les ressources des liens et des nœuds et minimisent le nombre d’hôtes physiques nécessaires pour réaliser l’émulation. Grâce à des évaluations numériques, des simulations et des expériences poussées, nous montrons que nos méthodes de placement sont plus performantes que les méthodes existantes, ce qui permetde réaliser des expériences fiables en utilisant un nombre minimal de ressources.