A large proportion of jobs submitted to modern computing clusters and data centers are parallelizable and capable of running on a flexible number of computing cores or servers. Although allocating more servers to such a job results in a higher speed-up in the job’s execution, it reduces the number of servers available to other jobs, which in the worst case, can result in an incoming job not finding any available server to run immediately upon arrival. Hence, a key question to address is: how to optimally allocate servers to jobs such that (i) the average execution time across jobs is minimized and (ii) almost all jobs find at least one server immediately upon arrival. To address this question, we consider a system withnservers, where jobs are parallelizable up to$d^{(n)}$servers and the speed-up function of jobs is concave and increasing. Jobs not finding any available servers upon entry are blocked and lost. We propose a simple server allocation scheme that achieves the minimum average execution time of accepted jobs while ensuring that the blocking probability of jobs vanishes as the system becomes large ($n \to \infty $). This result is established for various traffic conditions as well as for heterogeneous workloads. To prove our result, we employ Stein’s method which also yields non-asymptotic bounds on the blocking probability and the mean execution time. Furthermore, our simulations show that the performance of the scheme is insensitive to the distribution of job execution times.

In this paper, we address the challenge of anomaly detection in multivariate time series data related to environments requiring stringent guarantees, notably in 5G and beyond 5G systems promising high reliability and quality. Traditional approaches to anomaly detection, including supervised, unsuper-vised, and self-supervised methods often struggle with the diver-sity and unpredictability of real-world anomalies. To overcome these limitations, we propose a novel dual forecasting model approach, DREAM (Dual foREcAsting Model Anomaly Detection), which leverages both normal and unlabeled data to enhance detection accuracy. Our approach involves training two distinct models: one on normal behavior and the other on mixed behavior, and then comparing their outputs to identify anomalies. We also introduce new evaluation method, addressing the shortcomings of traditional point-wise evaluation. Our experiments, with multiple networked system datasets, demonstrate that DREAM outper-forms traditional forecasting-based approaches, including when used in a hybrid manner with traditional forecasting algorithms.

A large proportion of jobs submitted to modern computing clusters and data centers are parallelizable and capable of running on a flexible number of computing cores or servers. Although allocating more servers to such a job results in a higher speed-up in the job's execution, it reduces the number of servers available to other jobs, which in the worst case, can result in an incoming job not finding any available server to run immediately upon arrival. Hence, a key question to address is: how to optimally allocate servers to jobs such that (i) the average execution time across jobs is minimized and (ii) almost all jobs find at least one server immediately upon arrival. To address this question, we consider a system with n servers, where jobs are parallelizable up to d(n) servers and the speed-up function of jobs is concave and increasing. Jobs not finding any available servers upon entry are blocked and lost. We propose a simple server allocation scheme that achieves the minimum average execution time of accepted jobs while ensuring that the blocking probability of jobs vanishes as the system becomes large (n → ∞). This result is established for various traffic conditions as well as for heterogeneous workloads. To prove our result, we employ Stein's method which also yields non-asymptotic bounds on the blocking probability and the mean execution time. Furthermore, our simulations show that the performance of the scheme is insensitive to the distribution of job execution times.

The microservice paradigm is widely adopted in the design of Virtualized Network Functions (VNFs). Still, 5G/6G networks require to carefully orchestrate the allocation and re-arrangement of (micro)services to avoid a largely segmented solution space while services arrive and leave the network. To support latency-effective service provisioning, we introduce a novel placement and migration strategy that chooses the microservice(s) to migrate and selects the optimal destination (data center) while considering the impact of the migration on other microservices. For this purpose, we devise fast and effective heuristics and we implement a prototype that maps the service in the deployment field. Extensive simulations show our proposed approach significantly reduces the service latency.

In this paper, we present ELoRa, an emulation tool that generates Long Range Wide Area Network (LoRaWAN) traffic for an arbitrary number of LoRa devices and in an end-to-end virtualized LoRaWAN setting. Using ns-3, we improve an existing radio access network simulator to produce traffic compatible with ChirpStack, an open-source, cloud-native, LoRaWAN network functions stack. Our tool can be used to create realistic traffic and anomalies in order to test orchestration techniques on a real, distributed infrastructure. Moreover, the LoRaWAN core network functions (bridges, network server) are agnostic to the simulation of the radio access and can change parameters of simulated devices using native LoRaWAN protocol primitives, therefore enabling live-testing of resource allocation techniques to manage the radio access network. Multiple ELoRa instances can be connected to the same LoRaWAN core, each instance being able to support 50000 devices and 7 gateways.

This paper summarizes the PhD thesis and the 10 associated publications on the optimization of network slice placement in large-scale distributed infrastructures by focusing on online heuristics and approaches based on Deep Reinforcement Learning (DRL). First, we rely on Integer Linear Programming (ILP) to propose a data model for on-Edge and on-network slice placement. Second, we leverage an approach called Power of Two Choices (P2C) to propose an online heuristic adapted to support placement on large-scale distributed infrastructures while incorporating Edge-specific constraints like latency. Finally, we investigate the use of Machine Learning (ML) methods, specifically DRL, to increase the scalability and automation of network slice placement by considering a multi-objective optimization approach to the problem. We will go through the extensive evaluation work that provide encouraging results about the advantages of the proposed approaches when used in realistic network scenarios.

To meet ever more stringent requirements in terms of latency, 5G/6G networks are evolving from centralized to distributed architectures, for which the cloud-native paradigm with services decomposed into microservices is utmost relevant. This in turn raises the issue related to the distribution of network functions. In this paper, we introduce a novel microservice placement strategy considering the internal service composition, notably the communication between microservices. We formulate the placement as an optimization problem with the aim of minimizing end-to-end service latency. We solve the optimization problem with a combination of greedy and genetic algorithms.

Motivated by the rapid deployment of applications based on connected objects, we propose in this paper an approach to differentiating Packet Delivery Ratios (PDRs) in LoRa Wide Area Networks (LoRaWAN). This type of network is simple to deploy and operate at the expense of loose commitments in terms of quality. To overcome this shortcoming, we propose an access control method for isolating clusters of devices and meeting differentiated PDR targets. Results show that our method outperforms known one in terms of PDR via improved parameter allocation and achieves high level of intra-cluster fairness, at the expense however of decreasing the maximum cell range.

Service quality differentiation is gaining popularity in IoT networks, notably in LoRaWAN, with the rapid widespread of applications on connected devices. There is clearly a business demand for quality in IoT in the context of smart cities and network operators are urged by application designer to offer quality differentiation. However, those types of networks have been designed on the basis of a best effort service model. In particular, Packet Delivery Ratio (PDR) can dramatically decrease in dense scenarios. In this paper, we propose and evaluate traffic control and channel assignment solutions for PDR differentiation in dense deployments. Several performance criteria are defined in order to analyze the gain achieved by a network operator as well as end users. Numerical results show that both players can benefit from quality differentiation with ad-hoc pricing. This proves to be effective if penalizing low requirement devices, as they can create a bottleneck in the system. Namely, we show that in high density settings we can reach a 20% better PDR with one of the proposed policies, improving mean device servicing rate by 10% and the operator gain by 7.5%.

Network Slice placement with the problem of allocation of resources from a virtualized substrate network is an optimization problem which can be formulated as a multi-objective Integer Linear Programming (ILP) problem. However, to cope with the complexity of such a continuous task and seeking for optimality and automation, the use of Machine Learning (ML) techniques appear as a promising approach. We introduce a hybrid placement solution based on Deep Reinforcement Learning (DRL) and a dedicated optimization heuristic based on the “Power of Two Choices” principle. The DRL algorithm uses the so-called Asynchronous Advantage Actor Critic (A3C) algorithm for fast learning, and Graph Convolutional Networks (GCN) to automate feature extraction from the physical substrate network. The proposed Heuristically-Assisted DRL (HA-DRL) allows for the acceleration of the learning process and substantial gain in resource usage when compared against other state-of-the-art approaches, as evidenced by evaluation results.

We consider online learning for optimal network slice placement under the assumption that slice requests arrive according to a non-stationary Poisson process. We propose a framework based on Deep Reinforcement Learning (DRL) combined with a heuristic to design algorithms. We specifically design two pure-DRL algorithms and two families of hybrid DRL-heuristic algorithms. To validate their performance, we perform extensive simulations in the context of a large-scale operator infrastructure. The evaluation results show that the proposed hybrid DRL-heuristic algorithms require three orders of magnitude of learning episodes less than pure-DRL to achieve convergence. This result indicates that the proposed hybrid DRL-heuristic approach is more reliable than pure-DRL in a real non-stationary network scenario.

We propose to demonstrate a network slice placement optimization solution based on Deep Reinforcement Learning (DRL), referred to as Heuristically-controlled DRL, which uses a heuristic to control the DRL algorithm convergence. The solution is adapted to realistic networks with large scale and under non-stationary traffic conditions (namely, the network load). We demonstrate the applicability of the proposed solution and its higher and stable performance over a non-controlled DRL-based solution. Demonstration scenarios include full online learning with multiple volatile network slice placement request arrivals.

We propose an online heuristic algorithm for the problem of network slice placement optimization. The solution is adapted to support placement on large scale networks and integrates Edge-specific and URLLC constraints. We rely on an approach called the "Power of Two Choices" to build the heuristic. The evaluation results show the good performance of the heuristic that solves the problem in few seconds under a large scale scenario. The heuristic also improves the acceptance ratio of network slice placement requests when compared against a deterministic online Integer Linear Programming (ILP) solution.

The evolution of telecommunication network towards cloud-native environments enables flexible centralization of the base band processing of radio signals. There is however a trade-off between the centralization benefits and the fronthaul cost for carrying the radio data between distributed antennas and data processing centers, which host the virtual RAN functions. In this paper, we present a specific split solution for an efficient fronthaul, which enables reducing the consumed bandwidth while being compliant with advanced cooperative radio technologies (interference reduction and data rate improvements). The proposed split has been implemented on the basis of Open Air Interface code and shows important gains in the required fronthaul bandwidth as well as significant latency reduction in the processing of radio frames.

Network Slicing has its roots in Network Function Virtualization (NFV) allowing high flexibility in the delivery of end-to-end network services. To achieve Network Slicing promises on efficiency, Network Slice Providers have to ensure optimized resource utilization and to guarantee Quality of Service when managing the life-cycle of a Network Slice. We focus in this paper on Network Slice Placement, intimately related to the VNF Placement and Chaining problem. In contrary to most studies related to VNF placement, we deal with the most complete and complex Network Slice topologies and we pay special attention to the geographic location of Network Slice Users. We propose a data model adapted to Integer Linear Programming. Extensive numerical experiments assess the relevance of taking into account the user location constraints.

We develop and demonstrate a network slice management architecture on top of the so-called Open Network Automation Platform (ONAP). A slice manager is introduced to allow a client to describe the desired slice along with the associated Service Level Agreement (SLA), and to interface with ONAP in order to deploy and maintain the negotiated slice. This approach is illustrated by considering a private mobile network based on open source software components orchestrated by ONAP.

In cellular networks, a local uplink scheduler cannot compute good estimates of the inter-cell interference even with exact channel state information (CSI). Losses (due to decoding errors) will hence frequently occur. The maximum achievable performance (MAP) can only be achieved when all the cells are scheduled simultaneously centrally so as to manage interference and power optimally. In this paper, we focus on the performance of a practical system with local schedulers (i.e., each cell is scheduled independently) and compare it to MAP. We also study how to improve the performance of practical systems when a Cloud Radio Access Network (C-RAN) is present, with simple coordination schemes. To this end, to compute MAP, we first formulate an offline system-wide scheduling problem and transform it into a more tractable signomial problem that we solve quasi-optimally using an iterative algorithm. Since this problem requires all the channel information in the system and it has a high computational complexity, it is not suitable for a real-time system. Then, we show that a practical system using an efficient local scheduler (in each cell) yields a much lower performance than MAP. To decrease this performance gap, in a system with a C-RAN, we propose a very simple and fast scheme to coordinate the scheduling in all cells and show that it improves the performance significantly (in terms of throughput and losses) even when only partial CSI is available.

New bandwidth-intensive and time-constrained services in 5G networks combined with network function virtualization is pushing network operators to deploy distributed cloud infrastructures at the edge of the network. Allocating resources in capacity-limited infrastructures raises new challenges, which have not really been so far considered in the cloud literature. In this context, we investigate placement and offloading strategies of constrained services. We set design principles of future distributed edge clouds in order to meet application requirements. We precisely introduce a costless distributed resource allocation algorithm, named CLOSE, which considers local information only. We compare via simulations the performance of CLOSE against those obtained by using mechanisms proposed in the literature, notably the Tricircle project within OpenStack. It turns out that the proposed distributed algorithm yields better performance while requiring less overhead.

We consider in this paper the implementation of a cloud radio access network (C-RAN) on a centralized multicore system supporting the base band processing of several distributed antennas. We present a parallel processing model based on both functional and data decomposition of virtualized base band unit (BBU) functions in order to reduce their runtime. We study two scheduling strategies of parallel runnable BBU jobs, where computing resources can be allocated either per user equipments or else per code blocks. By using data obtained when running an open source RAN code (namely, OAI), we introduce a batch queuing model (the M[X ]/M/C multi-service system) to assess the needed processing capacity in a data center while meeting tight latency requirements in the down-link and up-link directions. The proposed model is validated by simulation when processing a hundred LTE-cells in a multi-core system. Results provide valuable guidelines for sizing and deploying Cloud-RAN systems.

We introduce a new, simple, yet effective scheme for reducing the impact of receiver buffer blocking in Multipath TCP (MPTCP). This blocking primarily occurs when the paths have different characteristics. This phenomenon is due to the limited size of the MPTCP receiver buffer. In a nutshell, our scheme allows, under certain conditions, the retransmission of a segment by a different interface than the one used originally and the closure of the badly behaving TCP connection. Our scheme anticipates the problem by opening multiple TCP connections for each interface, but only uses one at a given time. It detects when a segment needs to be retransmitted from another interface, closes the ongoing connection on the original interface, and starts using one of the backup connections. We show through NS3 simulations that our scheme improves MPTCP goodput, with a gain of 19% compared to other MPTCP schemes, and provides a smooth data delivery to the application layer. This last feature is of utmost importance for streaming applications. Furthermore, while other MPTCP schemes fail to perform better than the best single path TCP for some scenarios, our proposed scheme always outperforms the best single path TCP.

Congestion in today's mobile networks is mainly managed by end to end congestion control algorithms such as TCP. However, in spite of its many merits, TCP does not cope well with the high variability inherent to radio links and customer experience may sometimes be poor as TCP reveals unable to rapidly grasp such a fluctuating bandwidth. To remedy this problem, we introduce in this paper a new proposal called SloMo. Contrary to existing mechanisms, SloMo does not rely on protocol analysis and may be applied to encrypted traffic. SloMo is particularly easy to implement and deploy, robust to encryption, and its performance evaluation is quite encouraging.

In the framework of Network Function Virtualization (NFV), we address in this work the design and sizing of Cloud-RAN architectures. We concretely investigate the execution time of software-based Base Band Units (BBUs) on multi-core systems. Since Cloud-RAN requires a real-time behavior, we use parallel programming techniques in order to minimize the runtime of BBU functions. For an efficient utilization of computing resources, we investigate the relevance of resource pooling where a global scheduling algorithm allocates processing units to runnable BBU-jobs. We specifically examine the gain that can be obtained when applying data parallelism on the channel decoding BBU-function which is the most expensive one in terms of processing time. Performance results show a significant reduction in the runtime of PHY functions which enables the deployment of a fully centralized Cloud-RAN architecture.

Many new mobile applications create traffic among cellular users. We define intra-cellular traffic as the traffic from one cellular user to another user in the same cellular network. This type of traffic introduces new challenges for cellular network operators. Most work in the literature focuses on the possibility to utilize the direct links between those users, if they are close to each other (this is called device-to-device (D2D) communication), to by-pass the base station. However, implementing D2D is not easy, especially because detecting that a traffic is intracellular is difficult. In this paper, we assume that we know how to detect if a traffic is intra-cellular or not and focus on designing a type-aware scheduler (i.e., a scheduler which has the information on the type of traffic) in a case where direct communications between users is not enabled. This scheduler can be seen as the benchmark against the case where direct communications are allowed. We show that performance gain can be obtained by jointly scheduling the uplink and downlink with respect to the case where the scheduler is blind to the types. We show for a homogeneous network that when the traffic types are known to a scheduler, a significant performance gain (up to 28%) can be achieved compared to the case where the traffic types are not known. We also analyze heterogeneous networks that consist of macro cells and small cells and show that up to a 36% performance gain can be obtained by performing type-aware user association jointly with user scheduling.

We consider in this paper an important Quality of Experience (QoE) indicator in mobile networks that is reneging of users due to impatience. We specifically consider a cell under heavy load conditions and compute the reneging probability by using a fluid limit analysis. By solving the fixed point equation, we obtain a new QoE perturbation metric quantifying the impact of reneging on the performance of the system. This metric is then used to devise a new pricing scheme accounting of reneging. We specifically propose several flavors of this pricing around the idea of having a flat rate for accessing the network and an elastic price related to the level of QoE perturbation induced by communications.

We report in this paper traffic measurements of YouTube traffic from Orange networks. We specifically analyze two weeks of measurements in early April 2012. We show that the popularity curves of YouTube files are constant in time and can be well approximated by truncated Zipf laws with a shape parameter less than one. In addition, there is a huge number of files which are viewed very rarely (only once or twice). Even if this may appear as an unfavorable situation with regard to caching in view of theoretical results on cache systems, we show that thanks to file request dynamics caching is very efficient for YouTube traffic. The main reason is that files are massively requested in bursts. Since these bursts represent a significant part of traffic, caching files even by using a rather small storage capacity can achieve high gains in terms of saved bandwidth. Bursts are moreover sufficiently intense so that popular files are not pushed out of the cache memory by those files viewed only once or twice. These observations are illustrated by performing trace driven simulations by using traffic traces captured in the Orange IP backbone network.

We report in this paper measurements from France Telecom commercial networks carrying traffic generated and received by ADSL and FTTH customers. By adopting a flow- based approach to traffic analysis, we show that both types of customers experience similar peer-to-peer services in that the bit rates that they see is rather low. In order to understand the origin of these similarities, we develop a mathematical model, which could be seen as an abstraction of a file sharing process between peers according to the principles of eDonkey. This model allows us to exhibit a phase transition phenomenon which is nested in the file sharing principle. We believe that this phenomenon explains why both types of customers see a congested peer-to- peer network.

We investigate in this paper the performance of a simple file sharing principle. For this purpose, we consider a system composed of N peers becoming active at exponential random times; the system is initiated with only one server offering the desired file and the other peers after becoming active try to download it. Once the file has been downloaded by a peer, this one immediately becomes a server. To investigate the transient behavior of this file sharing system, we study the instant when the system shifts from a congested state where all servers available are saturated by incoming demands to a state where a growing number of servers are idle. In spite of its apparent simplicity, this queueing model (with a random number of servers) turns out to be quite difficult to analyze. A formulation in terms of an urn and ball model is proposed and corresponding scaling results are derived. These asymptotic results are then compared against simulations.

This paper proposes a computational and memory-efficient technique for online unidimensional clustering of individual IP addresses in order to detect high-volume traffic clusters (hierarchical heavy hitters). Our technique is based on a Patricia tree and can cope with today's traffic volume. We test our algorithm by using a traffic trace composed of NetFlow records sent by a few tens of routers of the France telecom IP backbone network. We moreover show how our algorithm can be used for network anomaly detection.

On the basis of a reference model for ADSL traffic on an IP backbone link, established in an earlier study, we show that it is possible to infer the characteristics of long flows by performing a deterministic 1/N packet sampling. By using the fact that the number of active long flows can be represented by means of the number of customers in an M/G//spl infin/ queue with Weibullian service times, we derive some probabilistic properties of the sampled data. These properties are then used to infer the characteristics of the original flows. The method is illustrated by considering an actual traffic trace captured in the France Telecom IP backbone network. Experimental data show that the method proves quite efficient.

To evaluate the impact of peer-to-peer applications on traffic in wide area IP networks, we analyze measurements from a high speed backbone link carrying TCP traffic towards several ADSL areas. The first observations are that the prevalent part of traffic is due to peer-to-peer applications (almost 80% of total traffic) and that the usage of network becomes symmetric in the sense that customers are not only clients but also servers. This latter point is observed by the significant proportion of long flows mainly composed of ACK segments. When analyzing the bit rate created by long flows, it turns out that those TCP connections due to peer-to-peer applications have a rather small bit rate and that there is no evidence for long range dependence. These facts are intimately related to the way peer-to-peer protocols are running.

Measurements from an Internet backbone link carrying TCP traffic towards different ADSL areas are analyzed. For traffic analysis, we adopt a flow based approach and the popular mice/elephants dichotomy, where mice refer to short traffic transfers and elephants to long transfers. The originality of the reported experimental data, when compared with previous measurements from very high speed backbone links, is that the commercial traffic includes a significant part generated by peer-to-peer applications. This kind of traffic exhibits some remarkable properties in terms of mice and elephants, as we describe. It turns out that by adopting a suitable level of aggregation, the bit rate of mice can be described by means of a Gaussian process. The bit rate of elephants is smoother than that of mice and can also be well approximated by a Gaussian process.

In this paper we analyze the M/G/1 processor sharing queue with heavy tailed services and with impatient customers. It is assumed that impatience depends on the value of the service required. We prove that a reduced service rate (RSR) approximation holds for estimating the sojourn time of a customer in the system, when the queue capacity is finite or infinite. This allows us to evaluate the reneging probability of customers with very large service times. We then use these results to investigate the impact of admission control on a link of a packet network. Admission control simply consists of limiting the number of simultaneous connections. It turns out that there is a real benefit for the efficiency of the system to perform admission control: It globally increases the fraction of customers, who complete their service (i.e. without being impatient). Finally, we investigate the fairness of the system and propose a criterion to assess the capacity of the system so as to allow the completion of very large service times.

In this paper, we present simple performance bounds for multiplexed regulated traffic streams, which are leaky-bucket regulated with peak, mean rate and burst size constraints. We consider independent, heterogeneous streams, which are multiplexed in a common buffer. We derive bounds on the mean delay in the deterministic context and we then obtain a simple stochastic bound, which is exact when the number of sources increases. A byproduct is a characterization of the worst case sources for mean delay, when they are leaky bucket regulated.

We study in this paper the throughput of a TCP connection performing ideal congestion avoidance. With respect to existing studies published on the same topic, we determine upper and lower bounds for the throughput of a TCP connection experiencing an arbitrary constant loss. We then derive exact convergence results when the loss probability becomes arbitrarily small. A remarkable property of the results obtained in this paper is that they justify a posteriori constants, which appear in the approximation of the throughput and which has been observed earlier via simulation or experiments. The analysis is performed by assuming a finite and infinite maximum congestion window size.

With the ATM block transfer with a delayed transmission (ABT/DT) capability, the amount of bandwidth allocated to the different cell flows of a connection (namely the CLP=0+1 and possibly the OAM cell flows) is constant over an ATM block and is referred to as the block cell rate (BCR). Successive BCR values are dynamically negotiated between the user and the network via the exchange of resource management cells. The BCR negotiation for a given ATM block may be initiated either by the source or the destination. As a basic principle, before transmitting an ATM block the source should receive an explicit authorization from the network. This paper gives the existence proof of a protocol for supporting the ABT/DT capability as specified in the May 1996 version of lTU-T Recommendation I.371.

The authors address peak rate enforcement in asynchronous transfer mode (ATM) networks by pick-up policing mechanisms. Since the network cannot rely on user's compliance when declaring his traffic parameters, the policing function has been introduced to monitor the traffic characteristics of any connection. Policing mechanisms proposed in the literature are actually pick-up mechanisms since cells can pass transparently through unless a policing action was performed according to some counting process. The authors analyze the impact of the cell multiplexing jitter on policing mechanism dimensioning. Recognizing that dimensioning is a critical point for such mechanisms. It is shown that network buffer overflow cannot be avoided in certain conditions. Therefore, a new policing scheme, called cell spacing, is introduced, and two spacing algorithms are presented.>

The authors highlight the fact that most of the policing schemes proposed in the literature neglect important information on the past cell blocking behavior of the policing mechanism. They propose a modification of these schemes in which the discarded cell information is used, with almost no overhead compared to the usual strategies. The behavior of the modified policing mechanisms is discussed. This new behavior is illustrated by using the generalized leaky bucket to police the mean rate of a source. Analysis using an on/off input process shows that the modified versions detect smaller-magnitude abuses, police the sources closer to their declared resources utilization, and have a faster response time. Results are presented showing that the generalized sliding window exhibits the same kind of behavior.>