Cross-datacenter distributed model training is increasingly important, but WAN packet loss and long propagation delays make retransmission-based recovery expensive. Existing RDMA FEC schemes are application-unaware and uniformly protect all packets, wasting inter-datacenter bandwidth. In this paper, we present SMART, a semantic-aware loss recovery scheme for cross-datacenter model training. SMART distinguishes data-parallel (DP) gradients from pipeline-parallel (PP) activations using lightweight message-level tags; it selectively protects high-priority DP packets, gives earlier PP traffic stronger FEC, zero-fills only small profiled residual PP loss, and compensates tolerated low-priority DP losses. Simulation experiments with Qwen3-3B-derived workloads show that SMART reduces average and P99 flow completion times compared with retransmission-only recovery and uniform FEC, while preserving convergence close to the no-loss baseline.

Space-air-ground integrated networks (SAGINs) augmented with mobile edge computing (MEC) provide a unified yet heterogeneous substrate for latency-sensitive services. Deploying service function chains (SFCs) over satellites, aerial platforms, and ground nodes, however, is difficult due to hierarchical resource heterogeneity, time-varying network states, and stringent end-to-end (E2E) delay requirements. In this paper, we study online SFC embedding in a three-layer SAGIN-MEC architecture under coupled computing and networking constraints. We model the deployment as a two-stage process: (i) placing each virtual network function (VNF) onto feasible nodes subject to computing-capacity constraints, and (ii) mapping inter-VNF traffic onto feasible paths subject to bandwidth and delay constraints. To achieve adaptive decisions under dynamic states, we cast the problem as graph reinforcement learning by jointly encoding the substrate topology and each SFC into a unified graph state, and propose a structure-aware PPO agent that combines graph convolution with domain features and an action-masking mechanism to eliminate infeasible placement/routing actions. Extensive experiments in dynamic large-scale scenarios show that the proposed method consistently outperforms competitive baselines, improving the acceptance rate by 6.991% under high load, reducing the average E2E delay by 13.556%, and increasing the long-term revenue-to-cost ratio by 27.763% on average.

Network testing is critical for evaluating the performance, reliability, and security of modern computer networks. A key challenge is creating an accurate, cost-effective, and high-performance network emulation environment. Network Impairment Emulators (NIEs) emulate real-world network conditions such as bandwidth constraints, latency, and packet loss, but existing CPU- and FPGA-based solutions suffer from limited performance, high costs, and poor flexibility. This paper proposes Software-Defined Network Impairment Emulation (SDNIE), a novel framework that leverages programmable switches for scalable, cost-efficient network impairment emulation. SDNIE introduces three key techniques: (1) intent-driven network impairment configuration, automating impairment modeling; (2) serial-parallel combined execution, optimizing performance; and (3) CPU-Tofino collaborative deployment, offloading complex computations. Experimental results show that SDNIE matches commercial emulators in performance while significantly reducing costs. This work demonstrates the potential of programmable switches in network testing, offering a scalable, cost-effective, and high-performance alternative for next-generation network impairment emulation.

In IoT environments, devices are often constrained by low power consumption and limited resources, making efficient connection establishment with minimal overhead crucial for real-time communication between edge devices and cloud servers. The QUIC protocol shows significant potential, but the encryption and decryption overhead is considerable. Reducing this overhead and improving connection establishment efficiency are critical to enhancing QUIC performance, especially for IoT edge devices that need to handle high-concurrency communication while maintaining low power consumption. This paper proposes CFcoQUIC, a CPU/FPGA co-design architecture that accelerates the handshake process and reduces the initial connection latency by parallelizing multiple encryption/decryption flows in high-concurrency environments. Experimental results demonstrate that the time for RSA encryption and decryption on FPGA is at least 19.1 times faster than on CPU, and the time for AES encryption and decryption is at least 5 times faster on FPGA. These results highlight the effectiveness of the proposed architecture in reducing QUIC handshake overhead in IoT environments with low-power edge devices.

Nonterrestrial networks (NTNs) enabled Internet of Things (IoT) extends connectivity to remote and underserved areas, enhances network reliability and coverage, and supports diverse IoT applications in challenging environments, such as rural, maritime, and disaster-stricken regions. As an emerging and fast-evolving IoT scheme, NTN-enabled IoT requires extensive evaluation to ensure effective deployment in real-world scenarios, such as connectivity, performance, and security evaluation. Since conducting testing in remote and diverse environments is logistically challenging and costly, we propose a generative artificial intelligence (GAI)-based synthetic traffic generation framework that facilitates comprehensive traffic analysis and performance evaluation. The proposed framework employs a GAI model to learn the traffic pattern and generate synthetic traffic from historical data. Our approach includes an embedding-based model for representing network flow attributes and a conditional generative adversarial network (CGAN) for generating traffic flows. Considering both source-destination information and statistical features achieves more comprehensive characterization of traffic flows. Finally, the simulation results demonstrate that the proposed approach can generate high quality traffic that conforms to real data distribution and shows obvious difference between multiple applications.

Drone networks are becoming increasingly significant in industrial Internet of Things (IIoT) applications. The limited resources of drones pose challenges in implementing robust security mechanisms that require substantial computation and power resources. Specifically, the inherent complexity of drone networks makes traditional intrusion detection systems (IDS) ineffective due to data imbalance and data scarcity. To address these challenges, this paper proposes a novel IDS framework that integrates conditional generative adversarial networks (CGANs) and utilizes the benefits from the systematic integration of foundation models within a federated learning (FL) paradigm. It leverages the CGANs to address the data issues ensures reliable performance and stable convergence against the foundation model. Moreover, our approach enhances data privacy relying on the differential privacy in FL and protects global model integrity through secure aggregation and updating. Simulation results show that the proposed framework achieve the accuracy rates of 91% and 99% on cyber and physical datasets, respectively. This framework achieves improvement ranging from 0.47% to 3.24% for cyber datasets and from 0.93% to 4.84% for physical datasets, which yields its superior performance in drone networks intrusion detection.

The Industrial Internet of Things (IIoT) enables communication among automation systems, machinery, and sensors in an industrial setting. To optimize critical industrial operations, a substantial volume of data concerning diverse in-factory activities and automation services is generated by IoT devices and sensors. This data are subsequently transferred to distant processing systems for analysis and decision-making. Nevertheless, a substantial latency in data transmission or any abnormality in the generated data may result in delayed or erroneous decisions, consequently impacting the efficacy of essential industrial systems. To address these challenges, we established an intelligent network architecture utilizing software-defined networking that achieves tactile latencies efficiently while handling industrial data traffic in an energy-efficient manner. To address the initial challenge, the suggested architecture utilizes the self-organized maps approach to distinguish between industrial traffic requiring tactile latencies and nontactile traffic. We utilize a binary tree-based flow table mapping method to enhance flow table matching and decrease lookup times. To address the second challenge, we employ the Support Vector Machine technique to identify anomalies in real-time industrial data traffic. The Hadoop system and Mininet emulator are utilized to evaluate the proposed architecture using the UNSW dataset. The results demonstrate the effectiveness of the suggested solution in providing energy-efficient tactile assurances and identifying anomalies in traffic.

RDMA over Converged Ethernet v2 (RoCEv2) is one of the most popular high-speed datacenter networking solutions. Watermark is the general term for various trigger and release thresholds of RoCEv2 flow control protocols, and its reasonable configuration is an important factor affecting RoCEv2 performance. In this paper, we propose ByteTuning, a centralized watermark tuning system for RoCEv2. First, three real cases of network performance degradation caused by non-optimal or improper watermark configuration are reported, and the network performance results of different watermark configurations in three typical scenarios are traversed, indicating the necessity of watermark tuning. Then, based on the RDMA Fluid model, the influence of watermark on the RoCEv2 performance is modeled and evaluated. Next, the design of the ByteTuning is introduced, which includes three mechanisms. They are (1) using simulated annealing algorithm to make the real-time watermark converge to the near-optimal configuration, (2) using network telemetry to optimize the feedback overhead, (3) compressing the search space to improve the tuning efficiency. Finally, We validate the performance of ByteTuning in multiple real datacenter networking environments, and the results show that ByteTuning outperforms existing solutions.

Network emulator is an equipment used in the field of computer networking to replicate and simulate real-world network conditions, especially poor-quality network conditions accompanied by various damages, in a controlled environment. It plays a crucial role in the development, testing, and validation of various new network-related technologies, protocols, and applications. Compared with simulation and test-bed methods, network emulation possesses the advantages of accuracy and cost-efficiency. However, legacy network emulation methods are implemented serially, which are typically restricted in efficiency and waste computing resources. In this paper, we propose a combined network emulator, CombNE. To implement this emulator, we consider P4 programmable switches as a desirable option. CombNE consists of three logical components. First, CombNE provides a policy specification scheme to intuitively describe operator’s intents. Secondly, the CombNE parallelizer intelligently identifies the dependencies between network damages, automatically determines parallelization and generates a combination strategy. Third, CombNE will generate optimized P4 files and flow table information based on the combination strategy and deploy them to P4 programmable switches. Finally, we evaluated the performance and resources of CombNE.

Federated learning (FL) in Industrial IoT (IIoT) facilitates collaborative model training across distributed edge devices, ensuring data privacy and localized insights without centralized data aggregation. However, the networked parameter sharing mechanism in FL renders it vulnerable to exploitation by man-in-the-middle (MITM) attackers, potentially disrupting the model training process. To mitigate this threat, this article presents a novel blockchain-reinforced FL architecture aimed at enabling cooperative intrusion detection. Initially, FL is leveraged to aggregate all learned information from edge servers, thereby disseminating extracted attack characteristics to all participants through gradient sharing. Subsequently, a blockchain-based parameter verification scheme is introduced to safeguard against tampered local parameters affecting the global model. Clients record model parameters in smart contracts deployed on a private chain, and parameter servers verify parameter confidentiality before aggregation, ensuring only valid parameters are considered. Finally, extensive experiments are conducted using an edge IIoT cybersecurity data set comprising 61 features spanning ten protocol layers and five attacks targeting IIoT connectivity protocols. Simulation results demonstrate that the proposed scheme significantly enhances intrusion detection accuracy, achieving a threefold improvement when two-thirds of federated nodes are subjected to MITM attacks.

The efficient and secure management of resources within flying ad-hoc networks (FANETs) poses formidable challenges. FANETs constitute a pivotal element of the space-air–ground-integrated network (SAGIN), employing network virtualization (NV) technology in tandem with service function chain (SFC) to facilitate end-to-end network services, akin to terrestrial networks. Nonetheless, the transient, dynamic nature of FANETs coupled with their susceptibility to network attacks engenders considerable complexity in the placement of SFCs within these networks. To address the rationality and security of resource allocation for SFC placement, this article proposes a reinforcement learning algorithm that sets strict security-level restrictions on the placement process and fully extracts the key features in FANETs. Additionally, a multilayer policy network is devised to dynamically perceive alterations in the FANET environment and compute an optimal SFC placement strategy. The proposed algorithm exhibits real-time adaptability to the dynamic environment, quantifies influential factors during placement, and achieves dynamic SFC placement. To assess the efficacy of the algorithm, three evaluation metrics—namely, SFC placement success rate, long-term average revenue, and long-term revenue cost ratio—are formulated and extensively evaluated through a plethora of experiments. Comparative analysis against alternative algorithms demonstrates enhancements of 20.6%, 15.3%, and 12.1% in the aforementioned metrics, respectively. The experimental findings substantiate both the convergence and efficiency of the proposed algorithm.

In RDMA (Remote Direct Memory Access) networks, end-host networks, including intra-host networks and RNICs (RDMA NIC), were considered robust and have received little attention. However, as the RNIC line rate rapidly increases to multi-hundred gigabits, the intra-host network becomes a potential performance bottleneck for network applications. Intra-host network bottlenecks can result in degraded intra-host bandwidth and increased intra-host latency. In addition, RNIC network problems can result in connection failures and packet drops. Host network problems can severely degrade network performance. However, when host network problems occur, they can hardly be noticed due to the lack of a monitoring system. Furthermore, existing diagnostic mechanisms cannot efficiently diagnose host network problems. In this paper, we analyze the symptom of host network problems based on our long-term troubleshooting experience and propose Hostping, the first monitoring and diagnostic system dedicated to host networks. The core idea of Hostping is to conduct 1) loopback tests between RNICs and endpoints within the host to measure intra-host latency and bandwidth, and 2) mutual probing between RNICs on a host to measure RNIC connectivity. We have deployed Hostping on thousands of servers in our distributed machine learning system. Not only can Hostping detect and diagnose host network problems we already knew in minutes, but it also reveals eight problems we did not notice before.

Nowadays, various network telemetry technologies are proposed to monitor the network and detect failures accurately in real-time, which can be categorized into two types, including the proactive network telemetry (NT) and the passive one. The passive NT can monitor the network with low band-width overhead, yet, cannot guarantee full network coverage. The proactive one can achieve full coverage, yet, lead to high bandwidth cost. To deal with the problem, we propose a failure-aware clustered network telemetry approach, called CNT. CNT leverages the practical objective network operating experience: different network links have various failure probabilities. It is aware of the failure probabilities and assigns the network links into two clusters accordingly. Then, based on the original network topology, CNT designs an active path planning algorithm to connect the two clusters of links into two sub-topologies, respectively. Finally, CNT performs network telemetry with different cycles. We evaluate CNT with various simulation experiments. The results show that compared to existing proactive schemes, CNT can achieve comparable network coverage with less cost.

In-band network telemetry task orchestration is to study how to reasonably select business flows to carry in-band network telemetry tasks to cover all necessary switches and ports. An inappropriate orchestration scheme not only fails to meet the requirements, but may reduce the performance of network telemetry. This paper proposes a multi-objective optimization-based in-band network telemetry task orchestration algorithm, while taking into account both the aspects of telemetry: freshness and intrusion. The experimental results show that the proposed scheme outperforms in terms of freshness and intrusion.

Network measurement provides rich data for network monitoring, control, and management. In-band network telemetry (INT) is a new network measurement technology that uses normal data packet to collect network information hop-by-hop. However, the design and implementation of INT protocol cannot do anything about packet loss: (1) The end-to-end telemetry mechanism makes INT unable to detect packet loss; (2) Since data packets may be lost due to various reasons, INT telemetry information will inevitably be lost. In summary, INT system by itself is unreliable. Incomplete telemetry data will seriously affect the performance of upper-layer network telemetry applications. In this paper, we present our successful experience in INT packet loss monitoring. We design, implement, and open source a powerful packet loss monitoring system for INT, called LossSight. The functions of LossSight include the detection of packet loss events, the deduction of the time and location of the losses, the diagnose of the root cause of the losses, and the recovery of the lost INT information. Experiment results show that LossSight provides excellent performance and extremely low overhead, including detection accuracy and diagnostic precision close to 100%, and detection latency of just milliseconds. In particular, LossSight uses a generative adversarial network to recover lost telemetry information, with excellent accuracy and reliability. LossSight has been running stably in the supercomputing interconnection environment of the National Supercomputing Center in Jinan. We suggest that all INT applications that require reliable telemetry information should be implemented based on LossSight.

At present, there is no unified network resource scheduling and business optimization system in cloud-edge collaboration services with supercomputing as the core, especially network transmission and data management. In this poster, we focus on how to optimize the quality of user experience while meeting huge computing power requirements, and design a novel network resource scheduling and traffic optimization platform centered on supercomputing system, which is called PANGU. PANGU is a multi-dimensional network resource scheduling solution for supercomputing that can guarantee collaborative network services. PANGU can make up for the lack of computing power of edge computing while solving the adaptability of supercomputing nodes to pervasive cloud edge computing. Through effective network resource management, PANGU can provide higher productivity.

QUIC provides a secure, reliable and low-latency communication foundation for HTTP. QUIC uses the connection ID to uniquely determine a connection from client to server. After the user switches the network, the server recognizes the user request according to the connection ID and continues to provide services through the connection migration technology. This paper proposes a Proactive Connection Migration (PCM) mechanism for QUIC. PCM gives QUIC the ability to select the optimal network in a heterogeneous network environment. Firstly, PCM actively perceives the different networks available to users. Then, PCM integrates the network quality exploration of different paths into the user’s multiple request actions. Finally, PCM takes response delay and jitter into account, and uses online learning to find the optimal network for current Internet service. Experimental results show that, compared with original QUIC, the average response delay of QUIC with PCM is reduced by 59.43% at most.