Traffic Engineering (TE) has been widely used by network operators to improve network performance and deliver better service quality. One major challenge for TE is providing routing strategies that can adapt to highly dynamic future traffic scenarios. Unfortunately, existing works either suffer severe performance degradation under unexpected traffic fluctuations, or sacrifice optimality to guarantee worst-case performance when traffic remains relatively stable. In this paper, we propose LARRI, a learning-based TE framework that predicts adaptive routing strategies for unknown future traffic scenarios. By integrating future demand range prediction and optimal range routing imitation into a single step, LARRI learns to generate a routing strategy that accommodates a wide range of possible future traffic matrices, thereby achieving a good trade-off between performance optimality and worst-case guarantees. Moreover, LARRI employs a scalable graph neural network architecture, which greatly facilitates both training and inference. Extensive simulations on six real-world network topologies show that LARRI achieves near-optimal load balancing in future traffic scenarios, improves worst-case performance by up to 43.3% over state-of-the-art baselines, and consistently provides the lowest end-to-end delay under dynamic traffic fluctuations.

Software-Defined Wide Area Networking (SD-WAN) has emerged as a popular solution for today’s enterprise networks, where Traffic Engineering (TE) plays a crucial role in optimizing traffic distribution across different overlay tunnels. During network congestion, traditional SD-WAN approaches often switch traffic to backup Multiprotocol Label Switching (MPLS) tunnels with high economic costs. To address this issue, we introduce Dynamic Path Switching (DPS), a novel SD-WAN TE solution that leverages underlay path diversity within an Internet overlay tunnel to maintain high Quality of Service (QoS) while substantially reducing economic costs. DPS operates on a Virtual extensible Local Area Network (VXLAN) overlay and controls the 5-tuple flow ID in the outer encapsulation header of traffic flows at SD-WAN gateways, which enables Internet Service Providers (ISPs) to dynamically switch traffic flows across multiple underlay paths based on the hashing results of their flow IDs. Moreover, we leverage extended Berkeley Packet Filter (eBPF) to implement DPS with high efficiency. Our prototype implementation, evaluated on a real Internet testbed, demonstrates that DPS can provide 99.974% service availability for enterprise traffic while reducing economic costs by 64.26% compared to traditional MPLS-based SD-WAN TE solutions.

Visual place recognition is a critical component of robust simultaneous localization and mapping systems. Conventional approaches primarily rely on RGB imagery, but their performance degrades significantly in extreme environments, such as those with poor illumination and airborne particulate interference (e.g., smoke or fog), which significantly degrade the performance of RGB-based methods. Furthermore, existing techniques often struggle with cross-scenario generalization. To overcome these limitations, we propose an RGB-thermal multimodal fusion framework for place recognition, specifically designed to enhance robustness in extreme environmental conditions. Our framework incorporates a dynamic RGB-thermal fusion module, coupled with dual fine-tuned vision foundation models as the feature extraction backbone. Experimental results on public datasets and our self-collected dataset demonstrate that our method significantly outperforms state-of-the-art RGB-based approaches, achieving generalizable and robust retrieval capabilities across day and night scenarios. The code is available at https://github.com/HITSZ-NRSL/RGB-Thermal-VPR.

Distributed Traffic Engineering (TE) aims to optimize network performance by generating individual routing strategies at each router without a global view of the network. A major challenge for these TE solutions is handling performance degradation caused by unexpected traffic fluctuations and unpredictable link failures. Recently, Machine Learning (ML) techniques have introduced new opportunities to enhance distributed TE. In this paper, we propose Path-Based Graph Neural Network (PathGNN), which leverages the emerging GNN architecture to quickly infer robust and resilient routing strategies in a distributed manner to accommodate unexpected network conditions. PathGNN adopts a novel path-link bipartite graph modeling approach to capture the dynamics of link resources shared by routing paths. It then performs efficient GNN message exchanges among routers to make adaptive local routing decisions for better load balancing. Additionally, PathGNN leverages Supervised Learning (SL) to directly learn from optimal routing strategies through efficient offline training. Evaluation results on four real-world network topologies demonstrate PathGNN’s strong generalization capability. Compared to state-of-the-art distributed TE solutions, PathGNN improves the load balancing performance by at least 24.4% with lower end-to-end delay under dynamic traffic scenarios, and also boosts performance by up to 35.3% under multiple link failures.

Distributed Machine Learning (DML) utilizes parallel computation on multiple training nodes to accelerate machine learning model training. Parameter Server (PS) is a typical DML enabler and is widely used in industry and academia. Existing works propose to apply the emerging In-Network Aggregation (INA) technique to improve model training efficiency by offloading the whole gradient aggregation process in PS from hosts to programmable switches. However, existing INA systems may suffer from undetermined model training efficiency and service quality, given that many gradient aggregation processes are still performed by the server under irrational gradient aggregation strategies. In this paper, we propose a Deterministic In-Network Aggregation (DINA) scheme to improve model training efficiency by enhancing the efficiency of INA utilization in DML. Our key observation is to further increase worker sending rates by reducing gradient packets’ RTT (i.e., realizing packet sub-RTT). Based on this observation, DINA rationally selects the optimal global gradient aggregation switch depending on the switches’ available memory, worker sending rate, and server processing capacity. As a result, DINA reduces the dependence of INA systems on the server, improves worker sending rates, and mitigates network traffic load. We formulate the sub-RTT-INA-based gradient aggregation problem as a mixed-integer nonlinear programming problem. To efficiently solve the problem, we simplify it by transforming the nonlinear constraints into linear constraints and propose a mixed solution that combines randomized rounding and heuristic mechanisms. Simulation results show that DINA can provide determined training by reducing communication time by 12%-17% and network load by 28%-50% compared with existing solutions, thus taking full advantage of INA and realizing a determined INA service.

Parameter Server (PS) is a typical Distributed Machine Learning (DML) enabler and widely used in industry and academia. Existing works propose to apply the emerging In-Network Aggregation (INA) technique to improve model training efficiency. However, existing INA systems may suffer from undetermined model training efficiency and service quality, given that many gradient aggregation processes are still performed by the server under irrational gradient aggregation strategies. In this paper, we propose a Deterministic In-Network Aggregation (DINA) scheme to improve model training efficiency by enhancing the efficiency of INA utilization in DML. Our key observation is to further increase worker sending rates by reducing gradient packets’ RTT. Based on this observation, DINA can rationally select the optimal global gradient aggregation switch depending on the switches’ available memory, worker sending rate, and server processing capacity. Simulation results show that DINA can provide determined training by improving worker sending rates by 16%-87% and network load by 28%-46.8% compared with existing solutions.

Traffic Matrix (TM), which records traffic volumes among network nodes, is important for network operation and management. Due to cost and operation issues, TMs cannot be directly measured and collected in real time. Therefore, many studies work on predicting future TMs based on historical TMs. However, existing works are usually accuracy-centric prediction solutions that mainly focus on improving predicting accuracy of flows’ sizes (i.e., values of elements in TMs) without considering the practical application of TMs. In this paper, we propose a novel TM prediction solution called Prophet for Traffic Engineering (TE), a typical application for TMs which takes TMs as input to optimize routing. We identify that the critical property (i.e., ratio among elements) in a TM plays an important role in TE’s performance. Based on this analysis, we adopt the matrix normalization to maintain the critical property in TMs and customize a TE-centric angle loss function to introduce scale invariance of TMs for capturing the overall relationship error. Different from the element-wise Mean Squared Error (MSE) loss function in accuracy-centric prediction solutions, our proposed TE-centric angle loss function has a clear geometric interpretation, which confines the angle between predicted TM and real TM to zero. Simulation results show that the predicted TMs from Prophet can improve the performance of link-level TE and path-level TE by up to 45.4% and 52.8%, respectively, compared to existing solutions.

Traditional Traffic Engineering (TE) usually balances the load on network links by formulating and solving a routing optimization problem based on measured Traffic Matrices (TMs). Given that traffic demands could change unexpectedly and significantly in realistic scenarios, routing strategies opti-mized based on currently measured TMs might not work well in future traffic scenarios. To compensate for the mismatch between stale routing decisions and future TMs, network operators may perform routing updates more frequently, which could introduce significant network disturbance and service disruption. Moreover, given the high routing computation overhead of TE optimization in today's large-scale networks, routing updates could experience severe delay and thus cannot accommodate future traffic changes in time. To address these challenges, we propose Roracle, a scalable learning-based TE that quickly predicts a good routing strategy for a long sequence of future TMs, while the learning process is guided by the optimal solutions of Linear Programming (LP) problems using Supervised Learning (SL). We design a scalable Graph Neural Network (GNN) architecture that greatly facilitates training and inference processes to accelerate TE in large networks. Extensive simulation results on real-world network topologies and traffic traces show that Roracle outperforms existing TE solutions by up to 36% in terms of worst-case performance under future unknown traffic scenarios. Additionally, Roracle achieves good scalability by providing at least$71\times$speedup over the most efficient baseline method in large-scale networks.

This paper proposes a modified cockcroft-walton quasi-Z-source inverter (MCW-qZSI). The proposed inverter is conceived by embedding a boost cell reorganized and derived from the Cockcroft-Walton voltage multiplier into a traditional quasi-Z-source. Compared with the existing quasi-Z-source topology, it has a stronger voltage boost capability under the same total number of turns. The proposed circuit possesses all the advantages of a quasi-Z-source, such as continuous input current and common ground between input and output ends, while achieving high voltage gain at a small duty cycle, which results in a larger modulation index and higher output quality. In addition, the steady-state principle of the proposed inverter is analyzed. Also, its boost capacity, device voltage, current stresses, and superiority in magnetic devices are studied in detail. Then compare that with two typical (quasi-)$\mathrm{Z}$sources. Finally, a prototype of the proposed MCW-qZSI has been produced with an output power of 720 W. Theoretical analysis and experimentation results are both given to verify the feasibility of the proposed MCW-qZSI.

Traffic Engineering (TE) has been widely used by network operators to improve network performance and provide better service quality to users. One major challenge for TE is how to generate good routing strategies adaptive to highly dynamic future traffic scenarios. Unfortunately, existing works could either experience severe performance degradation under unexpected traffic fluctuations or sacrifice performance optimality for guaranteeing the worst-case performance when traffic is relatively stable. In this paper, we propose LARRI, a learning-based TE to predict adaptive routing strategies for future unknown traffic scenarios. By learning and predicting a routing to handle an appropriate range of future possible traffic matrices, LARRI can effectively realize a trade-off between performance optimality and worst-case performance guarantee. This is done by integrating the prediction of future demand range and the imitation of optimal range routing into one step. Moreover, LARRI employs a scalable graph neural network architecture to greatly facilitate training and inference. Extensive simulation results on six real-world network topologies and traffic traces show that LARRI achieves near-optimal load balancing performance in future traffic scenarios with up to 43.3% worst-case performance improvement over state-of-the-art baselines, and also provides the lowest end-to-end delay under dynamic traffic fluctuations.

Emerging applications pose different Quality of Service (QoS) requirements for the network, where Traffic Engineering (TE) plays an important role in QoS provisioning by carefully selecting routing paths and adjusting traffic split ratios on routing paths. To accommodate diverse QoS requirements of traffic flows under network dynamics, TE usually periodically computes an optimal routing strategy and updates a significant number of forwarding entries, which introduces considerable network operation management overhead. In this paper, we propose QoS-RL, a Reinforcement Learning (RL)-based TE solution for QoS provisioning and load balancing with low management overhead and service disruption during routing updates. Given the traffic matrices that represent the traffic demands of high and low priority flows, QoS-RL can intelligently select and update only a few destination-based forwarding entries to satisfy the QoS requirements of high priority traffic while maintaining good load balancing performance by rerouting a small portion of low priority traffic. Extensive simulation results on four real-world network topologies demonstrate that QoS-RL provides at least 95.5 % of optimal end-to-end delay performance on average for high priority flows, and also achieves above 90 % of optimal load balancing performance in most cases by updating only 10% of destination-based forwarding entries.

Traffic Engineering (TE) is an important network operation that routes/reroutes flows based on network topology and traffic demands to optimize network performance. Recently, new emerging applications pose challenges to TE with dynamic network conditions, where frequent routing updates are required to maintain good network performance with Software-Defined Networking (SDN). However, flow rerouting operations could lead to considerable Quality of Service (QoS) degradation and service disruption, which is often neglected by existing TE solutions. In this paper, we apply a new QoS metric named network disturbance to measure the negative impact of flow rerouting operations performed by TE. To achieve near-optimal load balancing performance and mitigate network disturbance together in dynamic network scenarios, we propose a flexible and disturbance-aware TE solution called FlexDATE that combines Reinforcement Learning (RL) and Linear Programming (LP). Specifically, FlexDATE leverages RL to intelligently identify flexible numbers of critical flows for each traffic matrix and reroutes these critical flows based on LP optimization to improve network performance with low disturbance. Empowered by a customized actor-critic architecture coupled with Graph Neural Networks (GNNs), FlexDATE can generalize well to unseen traffic scenarios and remain resilient to single link failures. Extensive simulations are conducted on five real-world network topologies to evaluate FlexDATE with real and synthetic traffic traces. The results show that FlexDATE can achieve the performance target (i.e., 90% of optimal performance) in 99% of network scenarios and effectively mitigate the average and maximum network disturbance by up to 9.1% and 38.6%, respectively, compared to state-of-the-art TE solutions.

Real-time video communication is profoundly changing people's lives, especially in today's pandemic situation. However, packet loss during video transmission degrades reconstructed video quality, thus impairing users' Quality of Experience (QoE). Forward Error Correction (FEC) techniques are commonly employed in today's audio and video conferencing applications, such as Skype and Zoom, to mitigate the impact of packet loss. FEC helps recover the lost packets during transmissions at the receiver side, but the additional bandwidth consumption is also a concern. Since network conditions are highly dynamic, it is not trivial for FEC to maintain video quality with a fixed bandwidth overhead. In this paper, we propose RL-AFEC, an adaptive FEC scheme based on Reinforcement Learning (RL) to improve reconstructed video quality with an aim to mitigate bandwidth consumption for different network conditions. RL-AFEC learns to select a proper redundancy rate for each video frame, and then adds redundant packets based on the frame-level Reed-Solomon (RS) code. We also implement a novel packet-level Video Quality Assessment (VQA) method based on Video Multimethod Assessment Fusion (VMAF), which leverages Supervised Learning (SL) to generate video quality scores in real time by only extracting information from the packet stream without the need of visual contents. Extensive evaluations demonstrate the superiority of our scheme over other baseline FEC methods.

Traffic Engineering (TE) is a widely-adopted network operation to optimize network performance and resource utilization. Destination-based routing is supported by legacy routers and more readily deployed than flow-based routing, where the forwarding entries could be frequently updated by TE to accommodate traffic dynamics. However, as the network size grows, destination-based TE could render high time complexity when generating and updating many forwarding entries, which may limit the responsiveness of TE and degrade network performance. In this paper, we propose a novel destination-based TE solution called FlexEntry, which leverages emerging Reinforcement Learning (RL) to reduce the time complexity and routing update overhead while achieving good network performance simultaneously. For each traffic matrix, FlexEntry only updates a few forwarding entries calledcritical entriesfor redistributing a small portion of the total traffic to improve network performance. These critical entries are intelligently selected by RL with traffic split ratios optimized by Linear Programming (LP). We find out that the combination of RL and LP is very effective. Our simulation results on six real-world network topologies show that FlexEntry reduces up to 99.3% entry updates on average and generalizes well to unseen traffic matrices with near-optimal load balancing performance.

Network operators usually adopt Traffic Engineering (TE) to configure the routing in their networks to achieve good load balancing performance and high resource utilization. While centralized TE can effectively improve network performance with a global view of the network, distributed TE has been considered as an alternative to manage large-scale networks that are usually partitioned into multiple regions. However, it is challenging for distributed TE to reach a global optimal performance since each region can make its local routing decisions only based on partially observed network states. In this paper, we propose a novel distributed TE scheme called FedTe, which leverages supervised learning coupled with a collaborative approach to improve the overall load balancing performance for multi-region networks. FedTe learns from the global optimal routing strategy in a centralized offline manner and predicts the optimal distribution of cross-region traffic among different regions through distributed deployment in real time. The predicted cross-region traffic distribution is integrated with measured local traffic to construct each region’s optimal regional traffic matrix, which is used to perform intra-region TE optimization. FedTe can also handle dynamic traffic variation and link failures with a 2-layer hierarchical graph neural network architecture. To validate the effectiveness of the proposed scheme, we evaluate FedTe with two real-world network topologies and a large-scale synthetic topology. Extensive evaluation results show that FedTe can achieve near-optimal load balancing performance and outperform state-of-the-art distributed TE approaches by up to 28.9% on average.

Traffic Engineering (TE) has been applied to optimize network performance by routing/rerouting flows based on traffic loads and network topologies. To cope with network dynamics from emerging applications, it is essential to reroute flows more frequently than today’s TE to maintain network performance. However, existing TE solutions may introduce considerable Quality of Service (QoS) degradation and service disruption since they do not take the potential negative impact of flow rerouting into account. In this paper, we apply a new QoS metric named network disturbance to gauge the impact of flow rerouting while optimizing network load balancing in backbone networks. To employ this metric in TE design, we propose a disturbance-aware TE called DATE, which uses Reinforcement Learning (RL) to intelligently select some critical flows between nodes for each traffic matrix and reroute them using Linear Programming (LP) to jointly optimize network performance and disturbance. DATE is equipped with a customized actor-critic architecture and Graph Neural Networks (GNNs) to handle dynamic traffic and single link failures. Extensive evaluations show that DATE can outperform state-of-the-art TE methods with close-to-optimal load balancing performance while effectively mitigating the 99th percentile network disturbance by up to 31.6%.

Traditional Traffic Engineering (TE) solutions can achieve the optimal or near-optimal performance by rerouting as many flows as possible. However, they do not usually consider the negative impact, such as packet out of order, when frequently rerouting flows in the network. To mitigate the impact of network disturbance, one promising TE solution is forwarding the majority of traffic flows using Equal-Cost Multi-Path (ECMP) and selectively rerouting a few critical flows using Software-Defined Networking (SDN) to balance link utilization of the network. However, critical flow rerouting is not trivial because the solution space for critical flow selection is enormous. Moreover, it is impossible to design a heuristic algorithm for this problem based on fixed and simple rules, since rule-based heuristics are unable to adapt to the changes of the traffic matrix and network dynamics. In this paper, we propose CFR-RL (Critical Flow Rerouting-Reinforcement Learning), a Reinforcement Learning-based scheme that learns a policy to select critical flows for each given traffic matrix automatically. CFR-RL then reroutes these selected critical flows to balance link utilization of the network by formulating and solving a simple Linear Programming (LP) problem. Extensive evaluations show that CFR-RL achieves near-optimal performance by rerouting only 10%-21.3% of total traffic.

Predicting the future network traffic through big data analysis technologies has been one of the important preoccupations of network design and management. Combining Markov chains with tensors to implement predictions has received considerable attention in the era of big data. However, when dealing with multi-order Markov models, the existing approaches including the combination of states and Z-eigen decomposition still face some shortcomings. Therefore, this paper focuses on proposing a novel multivariate multi-order Markov transition to realize multi-modal accurate predictions. First, we put forward two new tensor operations including tensor join and unified product (UP). Then a general multivariate multi-order (2M) Markov model with its UP-based state transition is proposed. Afterwards, we develop a multi-step transition tensor for 2M Markov models to implement the multi-step state transition. Furthermore, an UP-based power method is proposed to calculate the stationary joint probability distribution tensor (i.e., stationary joint eigentensor, SJE) and realize SJE based multi-modal accurate predictions. Finally, a series of experiments under various Markov models on real-world network traffic datasets are conducted. Experimental results demonstrate that the proposed SJE based approach can improve the prediction accuracy for network traffic by highest up to 38.47 percentage points compared with the Z-eigen based approach.