In-network aggregation (INA) has emerged as a key technology to alleviate communication bottlenecks in large-scale distributed training, but its performance is often hindered by suboptimal routing. Existing INA-aware routing algorithms suffer from certain limitations: they either lack a global, multi-job coordination mechanism, or operate on incomplete network models that ignore key hardware constraints such as switch processing capacity. These deficiencies lead to network congestion and inefficient resource utilization, ultimately undermining the full potential of INA. To address these challenges, we present INARouting, a novel framework that holistically solves the multi-job hierarchical aggregation routing problem. We propose TINA, a hierarchical aggregation protocol that supports multi-job in-network aggregation. To address different deployment scenarios, we develop two variants: INARouting-Opt that provides optimal solutions for moderate-scale networks, and INARouting-Relax, a fast and effective heuristic using LP-relaxation and a greedy score-based rounding algorithm for large-scale deployments. Through extensive experiments on various scales of Fat-Tree and Spine-Leaf topologies, we demonstrate that INARouting significantly outperforms state-of-the-art methods. INARouting- Opt achieves provably optimal solutions, reducing average job completion time by up to 56% compared to existing methods. Meanwhile, INARouting-Relax outperforms existing algorithms while being 5× faster in solving time, enabling efficient routing in large-scale, dynamic environments.

With the progressive evolution of large language models (LLMs) and the increasing need of computing for$\mathbf{6 G}$, it becomes crucial for multiple network nodes with limited computing resources to share the need for large model inference. Model partition methods have been proposed to enable computation-intensive artificial intelligence (AI) services by splitting an AI model across multi-edge and cloud nodes. In this paper, a distributed inference optimization framework for transformer decoder-only based LLMs (DIO-LLMs) is proposed in edge-cloud collaborative networks. The partitioning and offloading strategy is determined based on the computing workload and network status. DIO-LLMs specifically accounts for the parallel execution capabilities of the transformer architecture. It employs a two-phase model partitioning strategy, comprising inter-layer and intra-layer partitions, to effectively distribute LLMs across edge and cloud nodes. Additionally, to mitigate inference latency under resource limitations, a Greedy Proximal Policy Optimization (GPPO) based algorithm has been developed to devise optimal strategies. Simulation results indicate that under memory constraints, the proposed algorithm can reduce inference latency more effectively than other baseline algorithms.

As Internet of Things (IoT)-based artificial intelligence (AI) applications grow, the surge in computational and communication demands has raised concerns about energy consumption, making it critical for 6G networks to address this challenge. This paper examines the joint optimization of AI service placement, task scheduling, and computing resource allocation in an edge-network-cloud system to minimize long-term energy consumption. These problems are interdependent: AI service placement determines service locations, influencing task scheduling, which in turn dictates computing resource allocation. The key challenge lies in the coupling of these variables and the two time-scale nature of the problem, involving long-term (AI service placement) and short-term (task scheduling and computing resource allocation) strategies. To address this, a Hierarchical Markov Decision Process (HMDP) framework is proposed for efficient and coordinated optimization across time scales. A Hierarchical Mean-Field Dueling Double Deep Q-Network (HMFD3QN) algorithm is developed within this framework, where the upper layer optimizes AI service placement, and the lower layer manages task scheduling and computing resource allocation. By integrating mean-field theory, the algorithm reduces the complexity of multi-agent interactions. The computing resource allocation problem is shown to be convex when other variables are fixed, and an optimal strategy is derived using Karush-Kuhn-Tucker (KKT) conditions to simplify the action space for reinforcement learning. Experimental results demonstrate that the proposed method can reduce energy consumption by up to 34% compared to baseline methods, significantly improve queue stability, and increase the proportion of tasks meeting QoS requirements.

The booming of artificial intelligence (AI) agents has brought about promising business scenarios for sixth-generation (6G) mobile networks, while simultaneously posing significant challenges to network functionalities and infrastructure. These AI agents can be deployed on end devices (e.g., intelligent robots and intelligent cars) or as digital entities (e.g., personal AI assistants). As novel service entities with autonomous decision-making and task execution capabilities, AI agents introduce potential risks of uncontrollable actions and privacy disclosures. AI agents also require new 6G capabilities beyond traditional communication, including multimodality information interaction (e.g., AI models and tokens) and support for service requirements (e.g., computing and sensing of data). In this article, we introduce the concept of AI-agent communication network (ACN), a new paradigm to enable global information interaction and on-demand capability provisioning for single or multiple AI agents. We first introduce the vision and architectural framework of ACN. Then, key technologies and future research directions related to ACN are discussed. Furthermore, we provide potential use cases to elaborate on how ACN can expand the service capabilities of 6G networks.

Segment routing (SR) is an emerging architecture that can benefit traffic engineering (TE). Nowadays, TE in SR networks (SR-TE) is often solved as an optimization problem to optimize network performance such as link utilization. As network size grows rapidly, implementing SR-TE suffers from scalability issues, including long computation time, high control overhead and expensive deployment cost. In this paper, we propose Smart Node Selection (SNS), a scalable SR-TE method with learning-based node selection (NS). NS is a recently proposed technique for reducing computation time of SR-TE. It first selects a subset of nodes as candidate intermediate nodes to route traffic, then builds linear programming (LP) models that can be solved efficiently. However, existing NS methods use simple heuristics and consider only network topology, which may lead to unsatisfying network performance. To address this problem, we for the first time formulates NS as a reinforcement learning task, which learns a selection policy to achieve better trade-offs between TE performance and computation time, considering both topology and traffic. Besides, we extend NS with additional selection policies and a customized training algorithm, making it a unified framework for scalable SR-TE, which reduces not only computation time, but also control overhead and deployment cost. Performance evaluations on various real-world topologies and traffic matrices show that SNS significantly reduces computation time and control overhead of existing LP models while offering good network performance, and can also be used in partially deployed SR networks to reduce deployment cost.

Recently, Computing Force Networks (CFNs) have emerged to deeply integrate and flexibly schedule multi-layer, multi-domain, distributed, and heterogeneous computing force resources. CFNs build a resources trading platform between consumers and providers, facilitating efficient resource sharing. Therefore, resources trading is an important issue but it faces some challenges. Firstly, because all kinds of large-scale and small-scale resource providers are distributed in a wide area and the number of consumers is larger compared with edge/cloud computing scenarios, the credibility of consumers and providers is hard to guarantee. Secondly, due to market monopolies by large resource providers, fixed pricing strategies, and information asymmetry, both consumers and providers exhibit a low willingness to engage in resources trading. To solve these challenges, the paper proposes an incentive mechanism for trust-driven resources trading to guarantee trusted and efficient resources trading. We first design a trust guarantee scheme based on reputation evaluation, blockchain, and trust threshold setting. Then, the proposed incentive scheme can dynamically adjust prices and enable the platform to provide appropriate rewards based on providers’ classified types and contributions. We formulate an optimization problem aiming at maximizing the trading platform’s utility and obtaining an optimal contract based on individual rationality and incentive compatible constraints. Simulation results verify the feasibility and effectiveness of our scheme, highlighting its potential to reshape the future of computing resource management, increase overall economic efficiency, and foster innovation and competitiveness in the digital economy.

Congestion control protocols play a vital role in enhancing the performance of various applications within datacenter networks. While reactive congestion control (RCC) protocols are widely deployed in commercial datacenters, the research community has actively explored token-based proactive congestion control (TCC) protocols to further push the boundaries of performance. However, despite the emergence of numerous TCC variants, there has been a lack of systematic exploration in the design space of TCC. This paper aims to bridge this gap by proposing a framework for understanding the design choices within the TCC approach. In this study, we systematically analyze different design choices of TCC approaches and leverage this understanding to develop a novel TCC protocol called ToCC. To implement ToCC, we address a set of challenges and deploy it in NP-based smart NICs. We compare ToCC with state-of-the-art TCC and RCC protocols through extensive large-scale simulations and testbed evaluations. The results demonstrate that ToCC exhibits robustness in achieving low latency across various scenarios. Additionally, ToCC effectively reduces buffer occupancy by 4.8 times compared to existing approaches, and under incast scenarios, it significantly shortens flow completion time by up to 90%. Congestion control protocols are crucial for optimizing the performance of datacenter network applications. Although reactive congestion control (RCC) protocols are commonly used in commercial datacenters, researchers have been exploring token-based proactive congestion control (TCC) protocols to further enhance network performance. Despite the development of numerous TCC variants, there has not been a thorough examination of the design space of TCC protocols until now. This paper aims to address this gap by introducing a framework for understanding the design choices within the TCC approach for TCC protocols. By analyzing various design aspects of TCC approaches, we create a novel TCC protocol called ToCC. At the central of ToCC design is that it leverages congestion control mechanisms over tokens. To implement ToCC, we tackle several challenges and integrate it into NP-based smart NICs. Comparing ToCC with state-of-the-art TCC and RCC protocols through extensive large-scale simulations and testbed evaluations, we find that ToCC consistently achieves low latency across different scenarios. Moreover, ToCC significantly reduces buffer occupancy by 4.8 times compared to existing methods, and during incast scenarios, it decreases flow completion time by up to 90%.

With the rise of new applications such as AR/VR, cloud gaming, and vehicular networks, traditional network management solutions are no longer cost-effective. Digital Twin Network (DTN) creates a real-time virtual twin of the physical network, which improves the network's stability, security, and operational efficiency. AI models have been used to model complex network environments in DTN, whose quality mainly depends on the model architecture and data. This paper proposes an automatic data generation and optimization method for DTN called AutoOPT, which focuses on generating and optimizing data for data-driven DTN AI modeling through data-centric AI. The data generation stage generates data in small networks based on scale-independent indicators, which helps DTN AI models generalize to large networks. The data optimization stage automatically filters out high-quality data through seed sample selection and incremental optimization, which helps enhance the accuracy and generalization of DTN AI models. We apply AutoOPT to the DTN performance modeling scenario and evaluate it on simulated and real network data. The experimental results show that AutoOPT is more cost-efficient than state-of-the-art solutions while achieving similar results, and it can automatically select high-quality data for scenarios that require data quality improvement.

Distributed Denial-of-Service (DDoS) attacks pose formidable threats to the security and availability of critical Internet infrastructure. In-network computing technology brings new opportunities to address DDoS attacks due to its intrinsic data plane programmability and high performance. However, existing DDoS attacks detection schemes based on in-network computing are difficult to strike a balance between true positive rate and false positive rate, especially in low-rate DDoS attacks scenarios. In response to this challenge, we propose RADD, an entropy-based method to detect DDoS attacks in real time based on in-network computing. RADD measures the distribution of network traffic from the perspective of individual IP address to discern subtle fluctuations within network traffic, hence providing early indications of potential DDoS attacks. We implement a prototype of RADD over programmable switches and results show that our proposed method significantly outperforms the state-of-the-art or has equivalent accuracy in low-rate and highrate DDoS attacks scenarios.

As the increase of geographically distributed applications continues, the demand for high-speed, long-distance data transmission across wide area networks (WANs) has significantly increased. Remote Direct Memory Access (RDMA) is extensively deployed in data center networks (DCNs) for its high throughput, low latency, and reduced CPU utilization, and its extension to WANs is expected to fully leverage these benefits. However, existing RDMA solutions, while demonstrating superior performance in data centers, face a performance gap over WANs due to their reliance on DCNs for optimal performance and lack of optimization for WANs’ high latency and loss rates. To bridge this gap, we introduce Lossy Wide-Area RDMA (LoWAR), a high-goodput, high-reliability RDMA solution for lossy WANs. LoWAR incorporates a forward error correction (FEC) shim layer to protect RDMA messages from packet loss, thus minimizing the inefficiency of retransmissions. It also fully offloads processing to RNICs with minimal computational overhead and storage burden, operating transparently on RNICs without requiring modifications to existing applications and networks. We implement a LoWAR prototype with FPGA and evaluate its performance through testbed experiments. The results demonstrate LoWAR’s enhanced performance in lossy WANs: in WANs with 40ms RTT and 0.001% to 0.01% loss rates, LoWAR increases RDMA goodput by 2.05 to 5.01 times, reduces average flow completion times (FCTs) by 3.5% to 12.2%, and eliminates 99th percentile tail FCTs in most scenarios.

In the forthcoming sixth-generation (6G) era, edge-network-cloud collaboration is needed to support artificial intelligence as a service (AIaaS) with a strong demand for computing power. However, how to guarantee the Quality of AI Service (QoAIS) and utilize the edge-network-cloud collaboration to enhance the performance of AI service is a big challenge. In this paper, we propose an AI service management and network resource scheduling architecture based on human-like networking. Considering the Quality of Service (QoS) requirements and AI tasks, we propose a joint AI agent placement with deep neural network (DNN) deployment and dynamic bandwidth resource allocation algorithm (JAAPD-D). JAAPD-D is proposed to solve the short-term and long-term joint resource allocation problem which includes communication, computation, and memory resources in the network. We adjust the agent placement, DNN deployment, and schedule routing path to ensure effective service transmission in the long time interval and dynamically allocate bandwidth resources in the short time interval. We use Lyapunov optimization to ensure the system stability of the whole network, meet the QoS requirements of various services, and minimize the average end-to-end delay of services. Simulation results show that JAAPD-D outperforms existing algorithms in terms of delay, traffic accepted rate, network system throughput, and cost.

Under the development of computing and network convergence, considering the computing and network resources of multiple providers as a whole in a computing force network (CFN) has gradually become a new trend. However, since each computing and network resource provider (CNRP) considers only its own interest and competes with other CNRPs, introducing multiple CNRPs will result in a lack of trust and difficulty in unified scheduling. In addition, concurrent users have different requirements, so there is an urgent need to study how to optimally match users and CNRPs on a many-to-many basis, to improve user satisfaction and ensure the utilization of limited resources. In this paper, we adopt a reputation model based on the beta distribution function to measure the credibility of CNRPs and propose a performance-based reputation update model. Then, we formalize the problem into a constrained multi-objective optimization problem and find feasible solutions using a modified fast and elitist non-dominated sorting genetic algorithm (NSGA-II). We conduct extensive simulations to evaluate the proposed algorithm. Simulation results demonstrate that the proposed model and the problem formulation are valid, and the NSGA-II is effective and can find the Pareto set of CFN, which increases user satisfaction and resource utilization. Moreover, a set of solutions provided by the Pareto set give us more choices of the many-to-many matching of users and CNRPs according to the actual situation.

In the future, 6G network is expected to achieve deep integration of communication and computation, where computation-centric services will be ubiquitous in the network. There are differences in data size, computing power types (CPU/GPU), model complexity, and Quality of Service (QoS) requirements among various CPU computing services and artificial intelligence (AI) services. By providing AI as a Service (AIaaS) in 6G network, the deployment of AI models and the scheduling of task and computing resources can be accelerated. The fundamental challenge lies in the effective amalgamation of the long-term strategy of the model placement problem and the short-term strategy of the task scheduling problem to attain dynamic scheduling and management of tasks and heterogeneous computing resources. A two-timescale optimization method for joint task offloading, computing resource allocation and model placement is proposed in this article. We present an edge-network-cloud framework that configures AIaaS functional units, taking into account the heterogeneous computing requirements and QoS demands of different services. A long-term problem to minimize latency and energy consumption is formulated. To work out the coupled optimization parameters, the problem is decomposed into short-term deterministic sub-problems using Lyapunov optimization. We propose low-complexity algorithms for joint task offloading strategy based on deferred acceptance algorithm, computing resource allocation strategy based on convex optimization, and model placement strategy based on multi-armed bandits. Experimental results demonstrate that our approach outperforms reinforcement learning and other popular optimization algorithms in terms of complexity and effectiveness.

Traditional network management solutions can not easily meet the requirements of new applications (such as AR/VR, cloud gaming, and vehicular networks) at a reasonable cost. Digital Twin Network (DTN) builds real-time mirrors of physical networks, which can enhance the simulation, optimization, verification, and control capabilities that physical networks lack. AI models have been used to model complex network environments, which helps build real-time, lightweight, and high-precision DTN. This paper proposes a data generation and optimization method for data-driven DTN AI modeling through Data-Centric AI called AutoOPT. First, AutoOPT generates data in small networks based on scale-independent indicators, which helps the model generalize to large networks. Then, AutoOPT automatically filters out high-quality data through seed sample selection and incremental optimization, which helps the model be effectively trained to enhance accuracy and generalization. We apply AutoOPT to the DTN performance modeling scenario and test on GNNet Challenge datasets. The experimental results show that AutoOPT is more cost-efficient than the winning solutions of the challenge but can obtain similar results.

Recently, Computing Force Networks (CFN) is emerging to deeply integrate and flexibly schedule multi-layer, multi-domain, distributed, and heterogeneous computing force resources among the cloud, edge network, and end devices. In CFN, the market monopoly of large resource providers results in a lack of bargaining power for small-sized providers. Existing resources pricing strategies ignore dynamic market factors affecting prices. Moreover, there is information asymmetry between resource consumers and providers. These problems destroy the fairness of the trading market and damage the benefits of consumers and providers, making them have a low degree of willingness to participate in resources trading. Therefore, this paper proposes a contract theory-based incentive mechanism to solve the above problems and motivate resource consumers and providers to join CFN. The proposed scheme classifies resource commodities into different types and enables the trading platform to provide appropriate rewards based on commodities' types and contributions. More specifically, we formulate an optimization problem aiming at maximizing the trading platform's utility and obtain an optimal contract scheme based on the individual rationality and incentive compatible constraints. Simulation results verify the feasibility and effectiveness of our scheme.

Off-the-shelf x86 servers are widely deployed as middleboxes in edge and public clouds, such as cloud gateways and load balancers. They follow the “run-to-completion” model and achieve parallel traffic processing by distributing packet flows across multiple CPU cores using the RSS (receive side scaling) capability of NICs. However, RSS can cause inter-core load imbalance as it conducts stateless hashing without considering the CPU core utilization. As a result, multiple heavy-hitter flows can potentially overload a single CPU core when they are hashed onto that core. In this research, we propose P4RSS, a load-aware intra-server load balancing solution that leverages the P4 data plane. Specifically, a P4 ASIC is placed in front of the CPU to perform stateful traffic load balancing among multiple CPU cores based on real-time monitoring of core utilization. In addition, flow affinity maintenance and heavy hitter throttling are also offloaded to the P4 ASIC to free up valuable CPU computing resources. P4RSS can be implemented in the form of either hyper-converged server switches or P4-based SmartNICs. Evaluation results demonstrate that P4RSS reduces the standard deviation of CPU core utilization by 22%~53% compared to RSS. This not only improves the stability of middleboxes but also allows for higher CPU utilization without overprovisioning.

The emerging satellite internet constellation aims to deploy hundreds of low-orbit satellites to provide high-speed broadband internet services to global ground terminals. However, this poses a significant challenge for large-scale and highly dynamic satellite networking due to the traditional satellite constellations’ uniform structure. This structure is limited by four laser links per low-orbit satellite, using default connections of two intra-orbit links and two inter-orbit links, and is difficult to match with the uneven population distribution and user traffic on the ground, resulting in unnecessary overheads in propagation delay and transmission hops. In recent years, researchers have developed methods for matching structure and traffic distribution that overcome the limitations of traditional connection methods, reducing transmission delay and hops. Despite this progress, these methods still maintain the characteristics of uniform configuration. To address this issue, a new link connection design for large-scale low-orbit satellite network driven by user distribution has been proposed. This mechanism enables the dynamic matching of satellite structure and users, facilitating elastic networking under dynamic topology conditions, improving the overall capacity and utilization of satellite networks. Through simulation, this user-driven link connection design has been verified to reduce the average hop count by at least 41% in many scenarios, demonstrating its effectiveness for a variety of new large-scale low-orbit satellite networks.

Visualization of network simulation is an essential tool for network practitioners. However, the front-end of existing network simulators often fails to deliver satisfactory performance when dealing with modern network scales and interface speed. In this paper, we propose NetVision, an efficient visualization front-end of network simulation based on the Unity engine, which is commonly used for video game and virtual reality development. NetVision offers flow-level visualization of network behavior and performances. Then, through parallel optimization, NetVision supports real-time visualization for large-scale high-speed networks.

Discrete Event Simulation (DES) is an essential tool for network practitioners. Unfortunately, existing DES simulators cannot achieve satisfactory performance at the scale of modern networks. Recent work has attempted to address these challenges by reducing the traffic processed via novel approximation techniques; however, we argue in this paper that much of the slowdown of existing DES simulators is due to their underlying software architecture.

Request latency directly affects the performance of modern cloud applications. Due to various causes in hosts and networks, requests can suffer from request latency anomalies (RLAs), which may violate the Service-Level Agreement. However, existing performance monitoring tools have incomplete coverage and inconsistent semantics for monitoring requests and cannot accurately diagnose RLAs. This paper presentsBufScope, a high-coverage and low-overhead request event monitoring system, which monitorsbuffersto capture most RLA-related abnormal events with consistent request-level semantics in the end-to-end datapath of request. First,BufScopemodels the datapath of request as a buffer chain and defines events based on three properties of buffers, so as toend-to-end monitorthe root causes of RLA. Then, to achieveconsistent semanticsfor captured events,BufScopedesigns a request-level semantics injection mechanism to make events captured in networks have the victim requests’ ID. Finally,BufScopeoffloads the semantics operations and event collection in software to SmartNICs forlow CPU overhead. We have implementedBufScopeon commodity SmartNICs and programmable switches. Evaluation results show thatBufScopecan diagnose 98% RLAs with < 0.08% network bandwidth overhead and 0.6% application throughput decline.

To improve the users satisfaction requesting services in Computing Force Network(CFN), the issues and challenges of existing service scheduling mechanisms are examined firstly in this paper. Then in view of users' specific preference and expectation for services, a service optimization scheduling algorithm based on Bkd-tree is proposed, taking service response time, service scheduling cost, availability, and successability into consideration. The method covers a search algorithm and a reconstructing strategy. To improve the efficiency of selecting the service instances that satisfy user expectation, a pruning strategy is used to narrow search space. To support dynamic updates of the service instances, a reconstructing strategy is proposed. The experimental results show that the improved service scheduling algorithm can schedule service instances with the best overall performance while maintaining higher efficiency.

Facing the integrated enhancement of network and computing requirements from emerging high-computing-demand service such as XR and metaverse, Mobile Computing Force Network (MCFN) is proposed to satisfy this kind of service. MCFN is the network which achieves computing and mobile network convergence based on the perception, control, and management over computing resources. MCFN could have the global knowledge of network topology and service endpoint, which could provide the best network and service path to satisfy end-to-end requirements. The main objectives of this paper are introducing use cases, requirements of MCFN, and potential influence on 5G-Advanced and 6G network architecture as well as the key technologies to realize MCFN.