Section: New Results

Future networks and architectures

Participants : Adlen Ksentini, Bruno Sericola, Yassine Hadjadj-Aoul, Jean-Michel Sanner, Hamza Ben Ammar.

SDN and NFV. Network Function Visualization (NFV) and Software Defined Network (SDN) currently play a key role to transform the network architecture from hardware-based to software-based.

SDN is in the process of revolutionizing the way of managing networks by providing a new way to support current and future services. However, by relocating the control functionality in a remote entity, the measurements' accuracy of the resources' utilization becomes more difficult, which complicates the decision making. Although there are previous works focusing on the problem of network management and measurement in SDN networks, only a few proposed solutions have taken into consideration the trade-off existing between statistics'€™ polling frequency (i.e. generated overhead), and the accuracy of monitoring results (i.e. optimized resources’ allocation). In [62], we proposed a new approach calculating accurately the bandwidth utilization while adapting the polling frequency according to ports/switches activity. The emulations' results under Mininet clearly demonstrate the effectiveness of the proposed solution, which proved to be scalable compared to classical approaches. The controllers'€™ placement is another important concern that emerged recently to solve the scalability and the reliability issues of SDN networks. The placement efficiency is influenced by both network operators (NO) strategy and the supported service requirements, which makes more complex the decision-making process. In particular, the need to support QoS-constrained services may lead NO to guide the controllers' placement in a way to ensure services efficiency while optimizing the underlying infrastructure. In [82] and [66], we proposed a model for the placement of network controllers, and we formulated a general optimization problem. To provide more flexibility and to avoid time-prohibitive calculations, we proposed a hierarchical clustering strategy for the controllers' placement allowing to minimize the number of network controllers while reducing the potential disparity of burden between the different controllers. Besides, the algorithms' structure makes it easy to act on other network parameters to improve the reliability of the SDN network. In [107], we proposed an improvement of such algorithms, by considering an evolutionary solution based on a genetic technique with an ad hoc cross-over operator designed to solve a mono-objective controller placement problem.

To connect the VNFs hosted in the same Data Center (DC) or across multiple DCs, virtual switches are required. Besides forwarding functions, virtual switches can be configured to mirror traffics for network management needs. Among the existing virtual switch solutions, Open vSwitch (OVS) is the most known and used. OVS is open source, and included in most of the existing Linux distributions. However, OVS performance in terms of throughput for smaller packets is very smaller than of line rate of the interface. To overcome this limitation, OVS was ported to Data Plane Development Kit (DPDK), namely OVDK. The latter achieves an impressive line rate throughput across physical interfaces. In [83], we presented the result of OVDK performance test when flow and port mirroring are activated, which was not tested so far. The performance test focuses on two parameters, throughput and latency in OVDK, allowing to validate the use of OVDK for flow forwarding and network management in the envisioned virtualized network architecture.

Mobile cloud. To cope with the tremendous growth in mobile data traffic on one hand, and the modest average revenue per user on the other hand, mobile operators have been exploring network virtualization and cloud computing technologies to build cost-efficient and elastic mobile networks and to have them offered as a cloud service. In such cloud-based mobile networks, ensuring service resilience is an important challenge to tackle. Indeed, high availability and service reliability are important requirements of carrier grade, but not necessarily intrinsic features of cloud computing. Building a system that requires the five nines reliability on a platform that may not always grant it is therefore a hurdle. Effectively, in carrier cloud, service resilience can be heavily impacted by a failure of any network function (NF) running on a virtual machine (VM). In [31], we introduce a framework, along with efficient and proactive restoration mechanisms, to ensure service resilience in carrier cloud. As restoration of a NF failure impacts a potential number of users, adequate network overload control mechanisms are also proposed. A mathematical model is developed to evaluate the performance of the proposed mechanisms. The obtained results are encouraging and demonstrate that the proposed mechanisms efficiently achieve their design goals.

Typically, maintaining a static pool of cloud resources to meet peak requirements with good service quality makes the cloud infrastructure costly. To cope with this, [58] proposes an approach that enables a cloud infrastructure to automatically and dynamically scale-up or scale-down resources of a virtualized environment aiming for efficient resource utilization and improved quality of experience (QoE) of the offered services. The QoE-aware approach ensures a truly elastic infrastructure, capable of handling sudden load surges while reducing resource and management costs. The paper also discusses the applicability of the proposed approach within the ETSI NFV MANO framework for cloud-based 5G mobile systems.

Video distribution. Due to the Internet usage evolution over these last years, the current IP-based architecture becomes heavier and less efficient for providing Internet services. In order to face this shortcoming, “Content Centric Networking” has been proposed. One of its important features is the use of in-network caching as a way of improving network performance and service scalability. However, in most of the existing CCN-based approaches several copies of the same content are present in the network, which reduce its efficiency. In [45], we proposed the “CLIque-based cooperative Caching” (CLIC) strategy, which basically consists in detecting cliques within the network topology to allocate more efficiently the content in the network. The main motivation of the proposed solution is to eliminate contents'€™ redundancy between neighboring nodes while promoting the most popular contents. This approach guarantees a sufficient number of copies of popular files within the network while maximizing the number of distinct content items. We evaluate the proposed scheme through simulation. The results show significant improvements in terms of cache management and network performance.

In [59], we make the case for opening the telco CDN infrastructure to content providers by means of network function virtualization (NFV) and cloud technologies. We design and implement a CDN-as-a-Service architecture, where content providers can lease CDN resources on demand at regions where the ISP has presence. Using open northbound RESTful APIs, content providers can express performance requirements and demand specifications, which can be translated to an appropriate service placement on the underlying cloud substrate. To gain insight which can be applied to the design of such service placement mechanisms, we evaluate the capabilities of key enabling virtualization technologies by extensive testbed experiments.

Network design using new dependability metrics. When designing a network taking into account its capabilities face to possible failures to its components, the basic theoretical framework is classical network reliability, where the system under study is represented by a graph with perfect nodes and imperfect links randomly and independently failing. The corresponding connectivity-based metrics must then be evaluated in order to quantify the robustness of the networking architecture. Recently, a new family of metrics, called diameter-constrained, have been proposed and analyzed by Dionysos' collaborators and members. In [53], we developed some elements for a factoring theory associated with these metrics. The paper is focused on the detection of irrelevant components, a key task when evaluating these types of quantities using factorization. The paper also includes a factoring algorithm, which is an up-to-date procedure exploiting all available results for implementing the pivoting idea (proved to be one of the most powerful methods in classical reliability analysis).

In [54], we consider an homogeneous network (identical and independent components). In this context, if p is the probability that each of the components works, then any reliability metric is necessarily a polynomial in p, and computing these metrics can be reduced to counting problems (counting specific classes of paths or of cuts, for instance). In the paper, we quantify, in some sense, the “degree of difficulty” of these counting processes, and we identify the situations where they are “easy”. The second contribution of the paper is to propose a fundamental problem from survivable network design, called the Network Utility Problem. The goal is to maximize network utility (defined as the opposite of the level of difficulty minus one), under a minimum edge-connectivity requirement.

Optical network design. Paper [65] presents a fast and accurate mathematical method to evaluate the blocking probability (the probability of a burst loss) in dynamic WDM networks without wavelength conversion (the present used technology). We assume that all links have the same number of wavelengths (the same capacity). The proposed model considers different traffic loads at each network connection (heterogeneous traffic). To take into account the wavelength continuity constraint, the method divides the network into a sequence of networks where all the links have capacity 1. Every network in the sequence is evaluated separately using an analytical technique. Then, a procedure combines the results of these evaluations in a way that captures the dependencies that occur in the real system due to the competition for bandwidth between the different connections. The method efficiently achieves results very close to those obtained by simulation, but orders of magnitude faster, allowing the evaluation of the blocking probability of all users (connections) for mesh network topologies.