Network device replacement, relocation, and vendor migration often require existing configurations to be translated to devices from different manufacturers while preserving network policies and operational behaviors. However, manual configuration translation is labor-intensive and requires substantial expertise, especially for large-scale configurations. To address this problem, we propose ConfigTransLE, a large-language-model-enhanced network configuration translation framework. ConfigTransLE consists of four main components: parameter corresponding, command translation, command arrangement, and LLM-adviser. Specifically, for commands with parameters, we combine heuristic translation with an LLM-based command filling module to construct executable target commands under command-format and manual constraints. For commands without parameters, we use semantic retrieval and LLM-based selection to determine the target command and its keywords. To satisfy vendor-specific view requirements, the translated commands are further arranged according to the target command hierarchy. In addition, an LLM-adviser is introduced to identify remaining semantic or logical issues and perform one additional repair attempt on the translation result. We evaluate ConfigTransLE on three translation tasks, Huawei-to-Nokia, Juniper-to-Huawei, and Cisco-to-Huawei, covering multiple protocols and configuration granularities. Experimental results show that ConfigTransLE achieves high commandlevel translation accuracy across different tasks. We further compare different LLM backends and analyze their trade-offs between translation accuracy and inference cost. In addition, simulator-based deployment validation on the Juniper-to-Huawei task provides additional evidence regarding the practical deployability of translated configurations and highlights the gap between text-level correctness and actual deployment constraints. These results show that ConfigTransLE is an effective and practical framework for cross-vendor network configuration translation.

Vendor-provided router images hide detailed routing-protocol implementation code, making it challenging to implement the same routing protocol on heterogeneous devices, as well as to locate and repair the configuration errors. While some works attempt to address these issues, they often fail in largescale deployment due to various challenges, such as vast and complex configurations, difficulty in mining causal relationships, and challenges in describing and applying causal relationships. To solve these problems, we propose RTB+, a novel tool to mine the causality between network configurations and routing behaviors, in order to assist operations personnel in better pinpointing configuration errors. First, RTB+ generates test cases based on configuration features using a coverage-guided combinatorial testing algorithm. Then, RTB+ employs a differential simulator model to obtain differential datasets storing the correspondence between configurations and routing behaviors. Finally, RTB+ constructs multilayer causal graphs to reason about routing behavioral semantics and vendor-level behavioral differences exposed by the underlying execution environment, enabling three key applications: behavioral change identification, configuration error localization, and configuration logic analysis. Experimental results show that RTB+ requires approximately half of the test cases generated by Metha, while improving code coverage by 10%. Moreover, RTB+ mines configuration-routing causal relationships with an accuracy of 98%, which is 5% more efficient than existing tools. Additionally, RTB+ enhances configuration error localization efficiency by an order of magnitude with causal graphs, and can analyze configuration logic within milliseconds.

Objective Microservices transform traditional monolithic applications into lightweight, loosely coupled application components and have been widely adopted in many enterprises. Cloud platform infrastructure providers enhance the resource utilization efficiency of microservices systems by co-locating different microservices. However, this approach also introduces resource competition and interference among microservices. Designing interference-aware strategies for large-scale, co-located microservice clusters is crucial for enhancing resource utilization and mitigating competition-induced interference. These challenges are further exacerbated by unreliable metrics, application diversity, and node heterogeneity.Methods In this paper, we first analyze the characteristics of large-scale and co-located microservices clusters at Alibaba and further discuss why cycle per instruction (CPI) is adopted as a metric for interference measurement in large-scale production clusters, as well as how to achieve accurate prediction of CPI through multi-dimensional metrics. Based on CPI interference prediction and analysis, we also present the design of the C-Koordinator platform, an open-source solution utilized in Alibaba cluster, which incorporates co-location and interference mitigation strategies.Results The interference prediction models consistently achieve over 90.3% accuracy, enabling precise prediction and rapid mitigation of interference in operational environments. As a result, application latency is reduced and stabilized across all percentiles (P50, P90, P99) response time (RT), achieving improvements ranging from 16.7% to 36.1% under various system loads compared with state-of-the-art system.Conclusion These results demonstrate the system's ability to maintain smooth application performance in co-located environments.

The escalating volume of network traffic drives the offloading of stateful Network Functions (NFs) to Data Processing Units (DPUs) for enhanced performance and host CPU relief. However, current DPU offloading approaches, while often boosting forwarding throughput, frequently suffer from poor concurrency performance. Effectively addressing this requires overcoming challenges inherent to DPU architectures, such as the Arm subsystem’s limitations in rapid connection setup and the intricate management of hardware flow table resources, which can otherwise cap connection processing rates (e.g., below 100k connections per second (CPS) with conventional methods). To address these limitations, we present ConnSched, a high-performance software framework for efficient and robust offloading of connection processing in stateful NFs to DPUs. ConnSched incorporates two key mechanisms: ConnSched-Filter adaptively identifies and filters out short-lived connections by monitoring their lifecycle and current DPU load, thereby reserving hardware resources for long-lived connections and enhancing robustness against concurrent traffic bursts. Complementing this, ConnSched-QoS implements a prioritized offloading mechanism, managing DPU hardware flow table entries by giving precedence to critical flows, especially when resources are constrained. Experimental results demonstrate that ConnSched achieves a 6.4× higher new connection creation rate compared to standard DPU offloading methods and 38.7% higher throughput than CPU-based approaches. ConnSched simultaneously improves concurrent connection handling capability while maintaining high packet forwarding performance.

Large language model (LLM) inference is increasingly deployed in heterogeneous environments, demanding efficient serving across diverse hardware platforms and backend engines. While existing load balancers and specialized serving systems facilitate request routing, they still suffer from various limitations, manifested in metric misalignment and observability gaps of black-box backends, non-linear performance degradation under heterogeneity, and conflicting multi-objective optimization goals. To solve these problems, we propose UPServe, a backendagnostic proxy designed for stateless LLM serving, which features three key designs: (1) a phase-aware cost modeling to estimate inference costs without relying on engine internals, (2) a unified score-based scheduling mechanism that integrates phase-aware concurrency and hardware speed normalization to handle heterogeneous non-linear degradation, and (3) an online scheduling algorithm named Phase-Aware Weighted Least Connections (PA-WLC) that drives adaptive dispatching to balance tail latency and aggregate throughput. Finally, we implement UPServe as a transparent proxy and conduct experiments for evaluations in heterogeneous environments. Comparison results demonstrate that UPServe significantly outperforms existing baselines, which reduces tail latency by up to 2.6$\times$ (P99 TTFT) and 2.1$\times$ (P99 TPOT), and improves aggregate throughput by up to 16%. Besides, UPServe operates without modifying inference engine internals, and incurs a negligible scheduling overhead of less than 0.3 ms per request.

Manual network configuration tools are constrained by their reliance on extensive domain expertise and rigid, single-purpose designs, limiting their adaptability to diverse scenarios and complex applications. This paper introduces PreConfig, a novel language model-based framework for automating network configuration tasks. By framing tasks such as configuration generation, translation, analysis, and completion as text-to-text transformations, PreConfig unifies these processes under a single versatile model. Leveraging advancements in natural language processing, PreConfig eliminates the need for extensive manual re-engineering by automatically learning domain-specific patterns through continued training on a specialized network configuration corpus. To address the lack of domain knowledge in general language models, we construct a comprehensive dataset from vendor manuals and community forums and fine-tune a programming language model for robust performance across various tasks. Additionally, we propose ConfigBLEU, a novel evaluation metric that incorporates syntax-aware features to assess the accuracy of generated configurations. Experimental results demonstrate that PreConfig significantly outperforms existing tools and general-purpose language models in both syntactic accuracy and semantic correctness across diverse network configuration tasks. This work establishes a unified and adaptable approach for advancing network configuration automation.

Carbon (C) allocation, which refers to the partitioning of the primary products of photosynthesis into different functional pools, has important implications for plants in optimizing growth and development under variable environmental conditions. The large-scale spatiotemporal pattern of C allocation to leaves (Pleaf, the ratio of leaf C to gross primary productivity) in grasslands and its relationship with local aridity remain poorly understood. Here, we developed a remote sensing-based framework to quantify C allocation to leaves across grasslands in the Northern Hemisphere (north of 23.5 degrees N) from 2001 to 2019. By integrating two leaf area index (LAI) products (GLASS and GLOBMAP), two gross primary productivity (GPP) datasets (GLASS and FluxSat), and two global specific leaf area (SLA) maps, we derived pixel-level estimates of Pleaf. We then analyzed the spatial patterns, temporal trends, and climatic drivers of Pleaf, as well as their relationship with local aridity. Our results revealed that Pleaf ranged from 0.008 to 0.455, with significantly lower mean values in arid regions (0.067) than in humid regions (0.089). Over 60 % of grassland pixels exhibited increasing Pleaf, particularly in hyper-arid and arid regions. The impact of CO2 concentration exceeds that of temperature, precipitation, and radiation, emerging as the dominant factor driving interannual variations in Pleaf. Our results underscore the role of aridity in modulating C allocation to leaves and enhance our understanding of how climate change affects C allocation to leaves in Northern Hemisphere grasslands.

As network scale and complexity continue to increase, managing network configurations has become an increasingly challenging task. Existing configuration tools often depend on low-level, abstract intermediate representations, which require users to have substantial technical expertise. This reliance not only increases the learning curve but also heightens the risk of configuration errors. Recent advances in Large Language Models (LLMs) have demonstrated strong potential for automating tasks across various domains. However, their applications to network configuration generation remain limited due to several challenges, including hallucination, restricted context length, and insufficient adaptability to domain-specific requirements. To address these issues, we propose ConfAgent, an advanced network configuration generation system powered by a multi-model intelligent agent. ConfAgent comprises four key components: a conflict detector, an information extractor, a routing algorithm coder, and a formal synthesizer. These components collaborate to accurately interpret complex configuration intents, detect potential conflicts, and generate robust code and network configurations through intuitive natural language interactions. Extensive experiments conducted on the NetConfEval benchmark demonstrate that ConfAgent consistently outperforms existing state-of-the-art methods by margins ranging from 36 % to 100 %, particularly excelling in configuration tasks for large-scale network topologies.

As distributed deep learning training (DLT) systems scale, collective communication has become a significant performance bottleneck. While current approaches optimize bandwidth utilization and task completion time, existing communication libraries (CCLs) backends fail to efficiently manage GPU resources during algorithm execution, limiting the performance of advanced algorithms. This paper proposes ResCCL, a novel CCL backend designed for Resource-Efficient Scheduling to address key limitations in current systems. ResCCL enhances execution efficiency by optimizing scheduling at the primitive level (e.g., send and recvReduceCopy), enabling flexible thread block (TB) allocation, and generating lightweight communication kernels to minimize runtime overhead. Our approach tackles the global scheduling problem, reduces idle TB resources, and enhances communication bandwidth. Evaluation results demonstrate that ResCCL achieves up to 2.5x improvement in bandwidth performance compared to both NCCL and MSCCL. It reduces SM resource overhead by 77.8% and increases TB utilization by 41.6% while running the same algorithms. In end-to-end DLT, ResCCL boosts Megatron's throughput by up to 39%.

The ecosystem services provided by grasslands depend on their biomass or leaf area index (LAI). Under the background of climate change, the impact of preseason climate and extreme weather events, such as high temperature and drought, on the spatiotemporal dynamics of grassland LAI remains unclear. Here, we constructed three interpretable machine learning models (including a Bayesian model, an interpretable neural network model, and a random forest model), to investigate the impact mechanisms of climate factors on grassland LAI in China from 2001 to 2020. The results showed that all three models performed well in simulating LAI (with R2 ranging from 0.540 to 0.963). The random forest model performed the best. Preseason climate was the most important factor driving LAI changes. The increase in preseason temperature, precipitation, and radiation could lead to an increase in grassland LAI. Regarding extreme weather events, heat events and heavy-rainfall events had positive effects on LAI, while frost events and no-rainfall events had negative impacts. CO2 showed a significant fertilization effect. Grazing intensity had a relatively small impact on LAI. The impact of precipitation on LAI was greater in spring than in autumn, whereas the impacts of temperature and radiation were greater in autumn than in spring. This study develops a climate change-adaptive framework for predicting grassland growth dynamics based on machine learning models, which can provide scientific support for grassland ecological protection and adaptive management. The random forest model performed the best in simulating grassland LAI. Preseason climate was the most important factor driving LAI changes. Heat events and heavy-rainfall events had positive effects on grassland LAI. CO2 showed a significant fertilization effect on grassland LAI. The impact of precipitation on LAI was greater in spring than in autumn.

Real-time network traffic classification plays a crucial role in ensuring Quality of Service and network security, and machine learning (ML) based methods achieve high classification accuracy but induce significant computational overhead. While SmartNIC solutions can offload classification tasks thus reducing CPU burdens, they still suffer from various limitations, manifested in limited computing capabilities, insufficiency in dynamic load handling and high latency from heterogeneous computing architectures. To solve these problems, we propose SmartTC, with three key designs: (1) SmartTC employs hardware-software co-design to optimize SmartNIC processing power, (2) SmartTC adopts a trafficaware dynamic batch submission strategy that adjusts submission policies based on real-time network load, and (3) SmartTC proposes parallel pipeline scheduling that ensures efficient task execution while minimizing communication overhead. Finally, we implement SmartTC on the BlueField-3 DPU and conduct extensive experiments for evaluations, and comparison results demonstrate that SmartTC significantly outperforms existing solutions. For example, it reduces average traffic classification time by up to 16.8% under low loads and 90.9% under high loads. Besides, SmartTC does not affect Bluefield-3 network services, and saves host CPU usage by at least two cores.

With the rapid development of the Internet of Things (IoT) and Artificial Intelligence (AI) technologies, the Artificial Intelligence of Things (AIoT) has become a key driving force for realizing intelligent and automated applications. The deployment of Service Function Chains (SFCs) is crucial in dynamic AIoT environments, where efficiently and flexibly deploying SFCs to meet real-time application demands is a research focus. However, existing SFC deployment methods often face challenges such as dynamic variations and uncertainty in contextual information, resource allocation inefficiencies, and limited adaptability to changing network conditions. To address these issues, we propose a learning-based context-aware dynamic SFC deployment method tailored for AIoT environments. Specifically, we introduce an attention-based contextual feature extraction method to capture dynamic changes (e.g., link latency variations) and prioritize key contextual information, improving the rate of served requests by 17.90% (69.60% vs. 59.03% for MADDPG) and enhancing the flexibility of SFC deployment decisions. Additionally, to address resource allocation bottlenecks and adaptability challenges in SFC deployment, we propose a distributed learning-based context-aware approach that uses collaborative learning and periodic updates (every 200 ms) to adjust SFC deployment strategies in response to topology changes and load variations and optimize system performance. Extensive experimental results demonstrate the efficacy of the proposed algorithm. Numerical results demonstrate that our algorithm reduces SFC deployment latency by 8% (46 ms vs. 50 ms for MADDPG), achieves 98.3% computational resource utilization, processes 211 Mbit/s service data volume, and improves adaptability to network changes, as validated in simulations.

Modern distributed training systems face escalating communication bottlenecks as GPU clusters scale to accommodate large models. While collective communication libraries and automated synthesizers address algorithmic efficiency, they suffer from three critical limitations including labor-intensive manual intervention requirement, overreliance on predefined input optimization, and suboptimal isolated configuration optimization. To solve these problems, we present TuCCL, a systematic framework that co-optimizes communication algorithms and runtime parameters through three innovations: Topology-Aware Sketch Generation that automatically produces high-performance primitives, Hierarchical Configuration Optimization modeling nonlinear parameter-performance relationships, and Multi-phase Resource-Aware Configuration Optimization enabling joint configuration tuning with adaptive search space pruning. Evaluations demonstrate TuCCL’s superiority over state-of-the-art systems with 1.75x–11.49x bandwidth improvements for AllGather/AllReduce on NVIDIA V100/A100 clusters, 90.2% faster configuration search than grid methods, and 1.22x-2.52x end-to-end training speedups across diverse model scales.

Large Language Models (LLMs) have advanced rapidly, but their growing parameter scales and memory demands pose critical challenges for distributed training. Although GPU memory capacity improves steadily, model sizes expand much faster, causing frequent Out-of-Memory (OOM) errors and rising training costs. Existing memory optimization approaches, such as ZeRO-3 and offloading, alleviate per-GPU memory pressure but introduce excessive collective communication, limited computation–communication overlap, and degraded scalability. We present MAZ3, a distributed training framework that mitigates these limitations through three key techniques: (1) Collaborative CPU–GPU memory management, storing full parameters in CPU memory and broadcasting them within nodes to reduce global synchronization; (2) Fine-grained communication–computation overlap, aligning collective operations with model computation to hide latency; (3) Hierarchical aggregation operators, leveraging intra-node NVLink and inter-node NIC channels concurrently to minimize communication overhead. We implement MAZ3 on a multi-GPU cluster and evaluate it with large-scale models. Results show that MAZ3 reduces inter-node model communication (gradients and parameters) by 33%, improves training efficiency by 40.3%, and increases throughput by 67.9% compared with ZeRO-3. Moreover, MAZ3 retains the memory efficiency of ZeRO-3 while approaching the training efficiency and throughput of ZeRO-2 Offload, achieving a balanced trade-off between memory optimization and performance.

State-of-the-art deep learning models rely on large GPU clusters and various parallelism strategies, which in turn depend on collective communication (CC) operators to synchronize data. While vendor libraries (e.g., NCCL, RCCL) provide standard CC algorithms, they often suffer from bandwidth bottlenecks in imbalanced topologies. Recent synthesis-based methods improve performance but face three key limitations: poor scalability due to the combinatorial explosion of scheduling space, lack of support for multistage execution, and suboptimal communication throughput. We propose Canvas, a scalable and near-optimal CC scheduling framework that addresses these challenges. Canvas introduces: (1) Hierarchical synthesis to decompose the global scheduling problem into tractable subproblems for scalability. (2) Collective decomposition to enable structured, multi-stage algorithm generation. (3) Cross-micro-batch pipeline scheduling to parallelize communication across micro-batches and maximize link utilization. Evaluations show that Canvas achieves up to 1.98× bandwidth speedup over TACCL and 3.56× over TE-CCL, and synthesizes algorithms for 512-GPU topologies within 1.77 hours, whereas TACCL fails to produce results within 24 hours.

Distributed machine learning over geo-distributed clouds enables joint training of data located in different regions, alleviating the burden of transferring large volumes of training datasets, which greatly saves bandwidth. However, the limited capacity of WAN links slows down the inter-cloud communications, which significantly decelerates the synchronization of distributed machine learning over geo-distributed clouds. Besides, the multi-tenancy in clouds results in multiple training tasks running simultaneously, whose synchronizations consistently compete for the limited WAN bandwidth with each other, which further aggravates the training performance of each task. While existing works optimize synchronizations through techniques like gradient compression, multi-resource interleaving and so on, none of them targets at the synchronization congestion especially due to multi-tenant learning, which results in inferior training performance. To solve these problems, we propose a simple but effective scheme, SCC, for fast and efficient multi-tenant learning via synchronization congestion control. SCC monitors the cross-cloud network conditions and evaluates the synchronization congestion level based on the round-trip transmission time for each synchronization. Then SCC alleviates synchronization congestion via controlling the synchronization frequency according to the synchronization congestion level in a probabilistic way. Extensive experiments are conducted within our testbeds consisted of 16 NVIDIA V100 GPUs to evaluate the performance of SCC, and comparison results show that SCC can reduce the average training completion time and makespan by up to 28.6% and 43.2% over SAP-SGD [1]. Targeted experiments are conducted to demonstrate the effectiveness and robustness of SCC.

In this article, we explore intercity mobility in urban agglomerations by surveying people traveling across cities based on private car trajectory data. Specifically, we first adopt the statistical analysis method to mine the intercity mobility in terms of various metrics of travel trips, so as to gain a preliminary understanding of intercity mobility in urban agglomeration. Then, we utilize the tensor decomposition method to conduct in-depth study on the intercity mobility pattern from the perspectives of complexity and multidimensionality. We construct a 4-D tensor based on private car trajectory and point-of-interest (POI) datasets and define the functional similarity and geographic adjacency between regions. Finally, we design an alternating proximal gradient (APG)-based method to resolve the core tensor and factor matrix, leading to the fine-grained discovery of intercity mobility patterns on administrative divisions in the urban agglomeration. Extensive experiments are conducted to evaluate the analysis of intercity mobility, using a real-world dataset containing one-year private car trajectories from five cities in the selected urban agglomeration. The experiments show that the proposed method successfully captures 20 intercity mobility patterns, in which the factor matrices retrieve the patterns from different dimensions with core tensors characterizing correlations between patterns in factor matrices. Besides, the extracted intercity mobility patterns not only cover administrative areas with frequent intercity interactions, but also contain areas with less intercity interactions. It validates that the intercity mobility is consistent with the regional functions in urban agglomeration.

In recent years, object localization and detection methods in remote sensing images (RSIs) have received increasing attention due to their broad applications. However, most previous fully supervised methods require a large number of time-consuming and labor-intensive instance-level annotations. Compared with those fully supervised methods, weakly supervised object localization (WSOL) aims to recognize object instances using only image-level labels, which greatly saves the labeling costs of RSIs. In this article, we propose a self-directed weakly supervised strategy (SD-WSS) to perform WSOL in RSIs. To specify, we fully exploit and enhance the spatial feature extraction capability of the RSIs' classification model to accurately localize the objects of interest. To alleviate the serious discriminative region problem exhibited by previous WSOL methods, the spatial location information implicit in the classification model is carefully extracted by GradCAM $++$ to guide the learning procedure. Furthermore, to eliminate the interference from complex backgrounds of RSIs, we design a novel self-directed loss to make the model optimize itself and explicitly tell it where to look. Finally, we review and annotate the existing remote sensing scene classification dataset and create two new WSOL benchmarks in RSIs, named C45V2 and PN2. We conduct extensive experiments to evaluate the proposed method and six mainstream WSOL methods with three backbones on C45V2 and PN2. The results demonstrate that our proposed method achieves better performance when compared with state-of-the-arts.

Mobile edge computing emerges to serve mobile users with low-latency computation offloading in edge networks, which are resource-constrained with massive users and workloads. However, existing communication and computing resource allocation schemes for offloaded tasks aren't efficient enough, where finished tasks still occupy resources, wasting constrained resources. Besides, the multi-user offloading is usually for scenarios of one task per user, ignoring real-world multi-task offloading scenarios where each user has multiple tasks, lack generality and flexibility. Meanwhile, local computing resource allocation schemes in multi-task scenarios ignore resource readjustment, causing low resource utilization. To solve these problems, we propose ECO-GAME, an efficient multi-task offloading scheme, which dynamically allocates bandwidth and computing resources to unfinished tasks, resulting in high resource utilization. We initially formulate the multi-task offloading problem as the game minimizing each user's cost, which is NP-hard. Thus we re-formulate the game utilizing potential games to optimize user's objective either locally or globally, and prove the existence of its Nash equilibrium. We then design an efficient multi-task offloading algorithm to obtain an approximate solution in polynomial time, together with computational complexity analysis. We further conduct performance evaluation on ECO-GAME utilizing price of anarchy. Numerical results demonstrate the efficiency of ECO-GAME, and show ECO-GAME reduces 49.2% cost over the state-of-the-art work, and scales well with the increasing number of tasks and users.