• Transport Research
• Publications close
• Computer Networks close

Bundling of multiple access technologies is currently being standardized by 3GPP in the 5G access traffic steering, switching and splitting (ATSSS) framework, with the goal to increase robustness, resiliency and capacity of wireless access. A key part of an ATSSS framework is the packet scheduler, which decides the access network over which each packet is to be transmitted. As wireless channels are highly dynamic, a challenge for any scheduler is to correctly estimate the capacity of each path, and thereby avoid congesting the paths. In this paper, we further develop a recent packet scheduler that exploits cross-layer information from the congestion control state of individual transport layer tunnels when making scheduling decisions. Our aim is to achieve good path utilization while keeping the congestion delay low. Extensive emulations show that our approach reduces the excess delay at the bottleneck to as little as 34%. We furthermore show that our approach improves the performance of end-to-end applications including WebRTC and YouTube compared to state-of-the art. 

Abstract Machine learning has seen a significant surge and uptake across many diverse applications. The high flexibility, adaptability, and computing capabilities it provides extend traditional approaches used in multiple fields including network operation and management. Numerous surveys have explored machine learning algorithms in the context of networking, such as traffic engineering, performance optimization, and network security. Many machine learning approaches focus on clustering, classification, regression, and reinforcement learning. The innovation of this research, and the contribution of this paper lies in the detailed summary and comparison of learning-based congestion control approaches. Compared with traditional congestion control algorithms which are typically rule-based, capabilities to learn from historical experience are highly desirable. From the literature, it is observed that reinforcement learning is a crucial trend among learning-based congestion control algorithms. In this paper, we explore the performance of reinforcement learning-based congestion control algorithms and present current problems with reinforcement learning-based congestion control algorithms. Moreover, we outline challenges and trends related to learning-based congestion control algorithms.

Abstract As modern communication networks become more complicated and dynamic, designing a good Traffic Engineering (TE) policy becomes difficult due to the complexity of solving the optimal traffic scheduling problem. Traditional methods usually design a fixed model of the network traffic and solve an objective function to get a TE policy, which cannot ensure the solution efficiency. The emerging Deep Reinforcement Learning (DRL) together with the Software-Defined Networking (SDN) technologies provide us with a chance to design a model-free TE scheme through Machine Learning (ML). However, existing DRL-based TE solutions are all faced with a scalability problem, i.e., the solution cannot be applied to large networks. In this paper, we propose to combine the control theory and DRL technology to achieve an efficient network control scheme for TE. The proposed scheme ScaleDRL employs the idea from the pinning control theory to select a subset of links in the network and name them critical links. Based on the traffic distribution information collected by the SDN controller, we use a DRL algorithm to dynamically adjust a set of link weights for the critical links. Through a weighted shortest path algorithm, the forwarding paths of the network flows can be dynamically adjusted using the dynamic link weights. The packet-level simulation shows that ScaleDRL reduces the average end-to-end transmission delay by up to 39% compared to the state-of-the-art DRL-based TE scheme in different network topologies.

Abstract The market prediction for network deployments has positioned Software-Defined Networks (SDN) as the first of the options for changing local, transport, or cloud networks. Since the OpenFlow protocol gained traction and evolved in the last versions, the possibilities for expanding network capabilities to deploy custom services have risen considerably. With next-generation SDN (NG-SDN), flexibility has grown as data plane programming languages, such as P4, and Data-Control Plane Interface (DCPI) protocols like P4Runtime have appeared. Furthermore, the ability to program the data plane has opened the possibilities to develop new network telemetry approaches, such as In-band Network Telemetry (INT). A transition to partially incorporated SDN, also known as hybrid SDN, often involves considerable complexity, especially when legacy devices implement non-open standards and protocols. Therefore, incorporating programmable SDN devices and deploying network telemetry protocols on top of existing legacy devices is still challenging. This research focuses on deploying and integrating the INT protocol using programmable P4 switches over a hybrid SDN network. We describe and implement the required control plane applications and data plane configuration, and discuss the constraints that need to be managed so that P4 programmable switches can interact with the rest of the MPLS legacy devices. In this sense, we discuss P4 switch placement alternatives to maximize their performance and usability in a hybrid SDN network. Then, we validate the INT-based monitoring system by ensuring traffic forwarding using several INT header placements. In these tests, we demonstrate the feasibility of merging P4 switches running INT traffic and legacy devices, presenting the requirements to accomplish hybrid next-generation SDN (HNG-SDN) architectures. Besides, we provide new monitoring features, such as MPLS label verification, and we also use telemetry data to feed back traffic forwarding applications for traffic engineering purposes. We finally show the time that packets spend in the pipeline comparing different parsing and actions performed in different cases.

© 2021 Elsevier B.V. The main objective of Egress Peer Engineering (EPE) is to steer traffic exiting one Autonomous System (AS) to another in the most cost-effective way by assigning network traffic flows of different destinations to internal routes specific to every AS. The traffic assignment process is carried out to satisfy network operators’ objectives, which include optimizing resource utilization, minimizing monetary costs and avoiding overloading the peer links. Due to network traffic dynamicity and unpredictability, traffic assignments should be constantly updated so that the aforementioned objectives are satisfied. Each of these updates results in traffic assignment changes that transition the network to a more optimized state. Executing these changes all at once is detrimental to the internal network infrastructure. To tackle this issue, this work targets finding execution plans involving several intermediate steps whereby each step includes a subset of traffic assignment changes. While executing these steps, the network operator's objectives need to be guaranteed. To that end, an oracle algorithm that generates all the possible balanced subsets of traffic changes as execution plans is formulated. This algorithm is compared to two heuristics that are designed based on an analytical study of the problem itself. To effectively evaluate these algorithms, evaluation criteria are devised that encompass several technical and design quality metrics of the desired execution plan. These three approaches are evaluated on network configurations of small size networks, and the results obtained show that one of the heuristics outperforms the oracle implementation in terms of execution time while producing comparable results based on the evaluation criteria. For big networks, the best performing heuristic satisfies the quality metrics and generates its best execution plan in a short period of time.

Abstract In the Industrial Internet of Things (IIoT) in the 5G era, the growth of smart devices will generate a large amount of data traffic, bringing a huge challenge of network traffic classification, which is the prerequisite of IIoT traffic engineering, quality of service (QoS), cyberspace security, etc. It is difficult for current traffic classification methods to distinguish encrypted dataflow and design effective handcraft features. In this paper, a novel identification scheme of encrypted traffic, TSCRNN, is proposed to automatically extract features for efficient traffic classification, which is based on spatiotemporal features. TSCRNN includes the preprocessing phase and the classification phase. In the preprocessing phase, raw traffic data are processed with flow segmentation, sampling, and vectorization, etc. To solve the classification problem of long time flow, sampling strategies are used to collect samples from the middle of the long-lived flow. In the classification phase, TSCRNN extracts abstract spatial features by CNN and then introduces stack bidirectional LSTM to learn the temporal characteristics. The experiments were performed on the dataset ISCXTor2016. The experimental results show that TSCRNN outperforms other typical methods in all scenarios, which achieves the accuracy up to 99.4% and 95.0% respectively in Tor/nonTor binary classification tasks and sixteen classification tasks. Furthermore, TSCRNN is applied to other real network datasets obtained the satisfactory performance, which validates its feasibility and universality. It means that TSCRNN can effectively identify encrypted and anonymous traffic, provide a fine-grained traffic characterization mechanism, which will support the development of core technologies in the Industrial Internet of Things.

Abstract 5G networks aim to support a vast amount of data exchange, through network densification, including the integration of multiple radio access technologies (RATs) under a unified radio access and core network. This integration creates a heterogeneous environment where end devices dynamically select the most suitable RAT for consuming a service via a new or even an ongoing communication session. This paper proposes a distributed and adaptive network selection mechanism to address this challenge. The proposed mechanism comprises two co-operating algorithms, one located at the user equipment (UE), and the other at the core network. Its main objective is to satisfy user preferences regarding monetary cost, quality of service, security, energy consumption etc., while safeguarding an operator's traffic engineering policy to avoid congestion. A key feature of the mechanism is the use of a dynamic threshold used to find the sweet spot between a well-balanced access network and maximizing the number of user sessions placed into their most preferred RAT. This threshold is adjusted in real time according to the experienced network conditions. Extensive network performance and quality of experience simulations show that the proposed mechanism accomplishes its objectives and can be used to provide efficient traffic steering decisions.

Abstract Existing solutions of flow-level traffic engineering (TE) usually depend on the deployment of SDN or MPLS. In this paper, we design a flow-level and efficient TE scheme based on the conventional hop-by-hop routing protocol, i.e., OSPF. Motivated by the analysis and modeling on the real Internet traffic, we propose to detect and schedule a few large flows in real-time. The rerouting paths for large flows are computed in a centralized server and are distributed through extended OSPF. A few ACL entries are used for flow-level forwarding. We formalize the link weight assignment-based large flow scheduling problem and prove the problem is NP-hard. We propose to precompute several candidate paths to reduce decision computation overhead and path stretch. We develop an algorithm with performance bounds to allocate large flows to paths, and two algorithms to reduce extra LSA number for different system designs. Experiment results show our scheme can reroute large flows within 0.5 s. Simulation results show our scheme gets congestion metric values (i.e., performance ratios) 10% worse than the optimal for source and destination addresses-based flows. Our optimization mechanisms reduce the extra LSA number and computation time by 87% and 83% respectively for our scheme with pre-computed paths.