Home Archive
Archive
2026 Volume 63 Issue 9  Published: 2026-09-01
    Special Issue on Frontier Technologies and Applications of Knowledge-Enhanced Artificial Intelligence
  • Boxiang Ma , Ru Li , Shaoru Guo , Yinghao Li , Gutiérrez-Basulto Víctor
    doi: 10.7544/issn1000-1239.202660437

    Large language models (LLMs) exhibit systematic generalization limitations in scenario cognition tasks, revealing potential functional deficiencies in their knowledge memorization processes. Inspired by the pattern separation mechanism of the dentate gyrus in the human hippocampus, we hypothesize that failures in scenario cognition are associated with limitations in the internal pattern separation capability of LLMs. Based on this hypothesis, we first design a cross-entity representation diagnosis (CERD) method, which quantitatively measures entity representation confusion across different layers without additional training. Our analysis reveals that standard autoregressive training fails to induce sufficiently effective pattern separation capabilities, leaving substantial cross-entity correlation confusion within LLM representations. Building upon this observation, we further propose latent-space pattern separation (LPS), which explicitly incorporates pattern separation constraints into the training process by introducing a triplet-based contrastive regularization at target layers with severe entity confusion, thereby reducing correlations among easily confused knowledge representations. Extensive experiments across LLMs ranging from 0.8B to 9B parameters demonstrate that LPS consistently improves scenario cognition capabilities. Compared with standard autoregressive training, our method achieves up to 101.5% relative improvement in exact match (EM), providing a new perspective for understanding and enhancing scenario cognition and knowledge memorization mechanisms in LLMs.

  • Special Issue on Frontier Technologies and Applications of Knowledge-Enhanced Artificial Intelligence
  • Wei Chen , Fuzhen Zhuang
    doi: 10.7544/issn1000-1239.202660438

    In recent years, multimodal knowledge graphs have introduced auxiliary information such as entity textual descriptions and images, providing richer semantic sources for entity representation learning in knowledge graph completion. However, most existing multimodal knowledge graph completion methods focus on the alignment or fusion of features from different modalities, typically modeling such information as static external attributes of entities, while insufficiently exploiting the structural context formed by local topology. Since entity semantics are derived not only from their own multimodal attributes but also influenced by neighborhood relations, the lack of relation-constrained structural propagation may cause multimodal features to remain at the level of static semantics, making it difficult to form structured representations with reasoning capability. Particularly in scenarios where modal information contains strong noise, directly relying on initial multimodal features or prematurely fusing heterogeneous modalities may weaken the stability and discriminability of entity representations, thereby limiting the ability to predict complex relation patterns and potential facts. To address these issues, we propose a multimodal knowledge graph completion method based on a multi-view enhanced graph attention network, namely MEGAT. The proposed method models structural, visual, and textual information as three independent views under a shared graph topology, and constructs a relation-aware graph attention encoder for each view. In this way, relation-constrained neighborhood propagation and structural semantic modeling can be performed separately in different modal spaces. Subsequently, a gated late-fusion mechanism is employed to dynamically integrate multi-view representations, effectively alleviating the semantic interference caused by the direct fusion of heterogeneous modalities. Experimental results show that the proposed method enhances the ability of entity representations to capture graph structural context and complex relation patterns, achieving competitive completion performance on multiple public datasets.

  • Special Issue on Frontier Technologies and Applications of Knowledge-Enhanced Artificial Intelligence
  • Di Wang , Junping Du , Zhe Xue , Guanhua Ye , Tong Zhao , Xiaolong Meng , Haisheng Li
    doi: 10.7544/issn1000-1239.202660445

    To meet the demand for dynamic knowledge updating and long-term knowledge retention in knowledge-enhanced artificial intelligence, dynamic multimodal knowledge graph continual completion infers missing facts incrementally as graph snapshots evolve while keeping historical knowledge stable. It thus provides fundamental support for dynamic knowledge updates. Existing methods predominantly focus on single-modal structural information, failing to fully exploit textual and visual semantics; directly incorporating multimodal information may introduce unexpected semantic shifts across snapshots, thereby exacerbating catastrophic forgetting. To this end, we propose SAREC, a semantic alignment and adaptive replay enhanced method for dynamic multimodal knowledge graph continual completion. SAREC leverages large language models to generate entity semantic descriptions and combines them with vision-language pretrained representations to construct a multimodal discrepancy-aware fusion mechanism, enhancing the completeness and robustness of entity representations. A multimodal temporal semantic alignment module is further introduced to constrain cross-snapshot representation evolution and maintain semantic consistency. Finally, a temporal-aware adaptive replay mechanism is proposed, which dynamically allocates the replay budget according to snapshot recency and selectively replays historical knowledge to alleviate catastrophic forgetting. Experiments demonstrate that SAREC achieves superior performance under various knowledge evolution modes, providing effective support for continual knowledge injection and adaptive knowledge enhancement in open dynamic environments.

  • Special Issue on Frontier Technologies and Applications of Knowledge-Enhanced Artificial Intelligence
  • Jiejing Shao , Ming Yang , Xiaowen Yang , Yufeng Li
    doi: 10.7544/issn1000-1239.202660454

    Label scarcity remains a major obstacle to deploying machine learning in real applications, because many tasks require costly annotation by domain experts. Active learning reduces this cost by selecting valuable unlabeled instances for expert annotation. Existing methods usually design query criteria from a learner-centric view, such as uncertainty, diversity, committee disagreement, and core-set coverage. However, their gains are not stable across datasets, models, and annotation budgets, and an unsuitable strategy may even underperform random annotation. Meanwhile, conventional active learning often restricts experts to the role of label providers, lacking a collaborative medium that explains why samples are selected and what decision logic the model follows. As a result, expert feedback is difficult to elevate from individual labels into reusable knowledge rules, and the model cannot be effectively aligned with expert domain knowledge. To address these issues, we propose KALE, a knowledge-guided active learning method for explainable human-AI collaboration. KALE uses first-order logic rules as the collaborative medium between the model and experts. It first induces compact logic rules from the current model behavior, and then asks experts to judge their correctness according to domain knowledge. Correct rules are used for deductive weak-label expansion, while incorrect rules are treated as symbolic observations of knowledge-model disagreement and used to drive abductive disagreement querying. KALE further combines class-level rules, local rules, and sparse rule selection to keep expert interaction manageable. Experiments show that KALE achieves more robust improvement than existing active learning baselines, especially under small annotation budgets.

  • Special Issue on Frontier Technologies and Applications of Knowledge-Enhanced Artificial Intelligence
  • Yuehang Si , Zefan Zeng , Boyan Wang , Jing Huo , Yang Gao , Jincai Huang
    doi: 10.7544/issn1000-1239.202660408

    Temporal knowledge graph (TKG) reasoning aims to predict future events based on historical facts. However, existing reinforcement learning (RL)-based reasoning methods merely conduct path searches at the entity level, lacking high-level semantic and logical guidance for the reasoning process. Moreover, ontologies, as conceptual abstractions of graphs, have strong logicality, but their construction typically relies on static statistical rules and is difficult to adapt dynamically to varying data distributions. To effectively address these issues, we propose ODRT, an ontology-graph dual-agent reinforcement learning method for TKG reasoning. Specifically, it jointly initializes the TKG ontology by leveraging both large language model (LLM) semantic embeddings and pretrained TKG reasoning model structural embeddings. Furthermore, it dynamically optimizes the ontology structure through learnable frequency weights during the end-to-end training process. Within the proposed dual-agent RL framework, the upper-level ontology agent performs concept-level walks on the constructed ontology graph, providing effective semantic navigation for the lower-level graph agent through carefully designed shared historical states. Meanwhile, the lower-level graph agent conducts fine-grained entity-level searches on the temporal graph under the constraint of hierarchical soft reward functions. Extensive experiments on four publicly available reasoning datasets demonstrate that ODRT achieves state-of-the-art performance among existing RL-based baselines and reaches the current state-of-the-art level on two encyclopedic datasets. However, on event-based datasets with heterogeneous entity behavior patterns and predictions that rely on historical reuse, its performance still lags behind that of partially embedded methods. Comprehensive experimental analyses further verify the significant advantages of the proposed ontology-guided mechanism and the dual-agent reinforcement learning framework.

  • Special Issue on Frontier Technologies and Applications of Knowledge-Enhanced Artificial Intelligence
  • Jie Ma , Hanchi Zhang , Ning Qu , Haoquan Xue , Xinping Wang , Jun Liu
    doi: 10.7544/issn1000-1239.202660338

    Multimodal retrieval-augmented generation introduces external textual and visual knowledge, providing important support for complex document understanding, knowledge-based question answering, and step-by-step task guidance. However, existing methods usually delegate the interleaved text-image generation process to a single multimodal large language model. Specifically, the retrieved text chunks, candidate images, and image descriptions are simultaneously fed into the same model, which is required to perform text organization, visual evidence selection, image insertion position determination, and multi-image ordering within a single generation process. In complex long-context scenarios, however, the candidate content often contains weakly relevant text, visually similar images, and redundant visual evidence, which can easily lead to missing key images, selecting irrelevant images, and misordering multiple images. To address these problems, we propose a multimodal retrieval-augmented interleaved text-image generation method based on a multi-agent debate mechanism. The proposed method decomposes the generation process into three collaborative roles: a text reasoning agent, a visual perception agent, and a global judge agent. Through dual-track draft generation, set conflict and ordering conflict detection, multi-round adversarial debate, and global judgment and fusion, the method jointly optimizes image selection, dynamic image supplementation, redundant image removal, and image reordering. Experimental results show that the proposed method achieves competitive performance across the network data, academic document, and lifestyle domains of MRAMG-Bench. In particular, in the lifestyle domain, compared with the strong baseline M2IO-R1, the proposed method improves image recall and ordering consistency by 7.02 and 7.77 percentage points, respectively. In the network data domain, the proposed method improves the average performance by 7.90 percentage points compared with the best complete baseline. These results demonstrate the effectiveness of the proposed method in complex long-document interleaved text-image generation tasks. Anonymous code and reproduction instructions are available at: https://github.com/mira-ai-lab/MRAMG-MultiAgent-Debate

  • Special Issue on Frontier Technologies and Applications of Knowledge-Enhanced Artificial Intelligence
  • Xuhan Wu , Shiliang Sun
    doi: 10.7544/issn1000-1239.202660459

    Efficient scene graph generation (SGG) serves as a bridge connecting low-level visual perception and high-level semantic reasoning, playing an important role in various vision-language tasks. However, existing dedicated SGG models often produce unreasonable relation predictions due to the lack of semantic constraint mechanisms. Although multimodal large language models (MLLMs) possess rich commonsense knowledge, their inherent black-box characteristics and severe hallucination issues lead to unreliable relationship reasoning results when directly applied to SGG tasks. To address these challenges, we propose KETR (commonsense knowledge embedding Transformer), an end-to-end scene graph generation model that incorporates commonsense knowledge embeddings. The model extracts global and local image features through a multi-scale visual feature fusion encoder and employs a parallel decoder based on learnable queries for entity detection. The core innovation lies in a soft alignment knowledge fusion mechanism that dynamically aggregates knowledge information through probabilistic mapping between visual features and external commonsense knowledge graph embedding spaces, and a commonsense-aware subject-object indicator and knowledge-driven scoring constraints are introduced, integrating the TransE scoring function as a regularization term into the joint training loss to significantly improve the logical consistency of relationship reasoning. Extensive experiments on Visual Genome and Open Images V6 demonstrate that KETR achieves competitive performance on multiple standard SGG metrics. Further ablation studies show that soft-aligned knowledge fusion improves the recognition of infrequent predicates, while the commonsense-aware subject-object indicator and knowledge-driven scoring constraint reduce relation predictions that conflict with entity roles, object affordances, and spatial semantics.

  • Special Issue on Frontier Technologies and Applications of Knowledge-Enhanced Artificial Intelligence
  • Mingsheng Cao , Qiaolong Ding , Wenqing Zhang , Yanbin Li , Di Lin , Hui Xu
    doi: 10.7544/issn1000-1239.202660229

    To address the limitations of large language models (LLMs) in source code vulnerability detection—namely insufficient utilization of program structural information, unstable reasoning processes, and high fine-tuning costs—we propose a vulnerability detection framework that integrates graph-based representations with LLM reasoning capabilities. The proposed method constructs data flow graphs (DFGs) from source code to model variable dependencies and semantic propagation paths, and aligns graph features with the semantic embedding space of LLMs through a cross-modal mapping module, enabling effective injection of structural knowledge into the language model. Furthermore, a progressive pretraining strategy based on code reconstruction tasks is designed to enhance the graph encoder’s ability to capture vulnerability-related structural patterns. On this basis, structured reasoning distillation and parameter-efficient fine-tuning via low-rank adaptation (LoRA) are incorporated to transfer vulnerability analysis knowledge from teacher models to student models, improving detection performance while reducing training overhead. In addition, a chain-of-thought reasoning process tailored for vulnerability detection is introduced to improve interpretability and reasoning consistency. Experimental results demonstrate that the proposed method significantly improves both the accuracy and robustness of source code vulnerability detection while maintaining low fine-tuning cost.

  • Special Issue on Frontier Technologies and Applications of Knowledge-Enhanced Artificial Intelligence
  • Jiaqi Long , Suojuan Zhang , Wenning Hao , Zejun Chen , Xinjie Sun , Yifan Liu , Enhong Chen
    doi: 10.7544/issn1000-1239.202660421

    Knowledge tracing, as the core task of intelligent tutoring systems, aims to predict students’ future learning performance by analyzing their historical behavioral data, and provides a fundamental basis for personalized learning path planning. Current mainstream deep learning methods are constrained by knowledge boundaries and rely on single-feature modeling, often exhibiting issues such as overly simplified learning process modeling, inadequate contextual adaptability, and black-box prediction mechanisms that yield results lacking interpretability and difficult to translate into actionable teaching references. To address these challenges, we propose CAKT-Agent, a context-adaptive knowledge tracing agent based on mixture-of-experts model, and construct a three-tier collaborative architecture consisting of a unified memory module, a learning agent, and a companion agent: the unified memory module realizes global storage and sharing of learning data, educational priors and intermediate model states; The learning agent defines a five-dimensional context based on educational measurement theory and cognitive science, designs a hybrid expert architecture with dimension-aware capabilities, and achieves context-adaptive integration through a two-level gating mechanism; the companion agent generates learning diagnostic reports with educational comprehensibility based on expert activation weights and educational theory rules. Experimental results on real-world datasets show that the proposed model outperforms mainstream knowledge tracing models in the prediction accuracy of students' learning performance, with excellent context-adaptive capability and result comprehensibility.

  • Special Issue on Frontier Technologies and Applications of Knowledge-Enhanced Artificial Intelligence
  • Qianyi Lü , Hanwen Luo , Jun Wang , Zhen Li , Guoxian Yu
    doi: 10.7544/issn1000-1239.202660468

    Whole slide images (WSIs) are central to cancer diagnosis and prognosis. However, mainstream survival prediction methods directly map massive tile features into a single global representation for risk readout, failing to explicitly distinguish complex histological composition, ignoring cancer-specific clinical phenotype differences, and leaving the model’s internal reasoning process opaque. To address these issues, we propose PathRoute, an active evidence routing framework for interpretable WSI survival prediction. The framework models WSI reasoning as a two-stage directional process that first forms evidence and then reads out risk: four role-specific learnable queries collaborate in a shared latent space, respectively performing tissue-composition sensing, active evidence routing, phenotype-semantic auxiliary supervision, and risk readout; a directional attention mask further constrains forward visibility between queries, so that risk prediction is structurally realized as a downstream readout of pre-formed evidence representations, rather than an end-to-end black-box mapping. On five TCGA solid tumor cohorts under unified inputs and identical patient-level 5-fold cross-validation splits, PathRoute using WSI alone achieves higher average C-index than representative multimodal and WSI-only methods, validating the effectiveness of structured evidence organization. Further analysis shows that routing queries consistently and spontaneously concentrate attention onto tumor regions across all five cancer types without any tissue-type supervision, providing direct interpretability support for the active routing mechanism. Code available at https://github.com/ZephyrGo/pathRoute.

  • Network and Information Security
  • Hangyu Yan , Renhai Chen , Yu Zhang , Wenyan Shi , Hongyue Wu , Zhiyong Feng
    doi: 10.7544/issn1000-1239.202550462

    Non-volatile memory (NVM), with the advantages of persistence after power failure, high density, and large capacity, has attracted extensive attention from both academia and industry. However, its persistence also makes NVM more vulnerable to security threats such as unauthorized access and data tampering. To ensure data confidentiality and integrity during storage and access, secure NVM systems generally employ encryption and integrity verification mechanisms. Nevertheless, directly applying these mechanisms to NVM may cause severe write amplification, resulting in increased write latency and energy consumption. To address this problem, we propose epoch-based re-encryption delay (ERED), including word-epoch-based re-encryption delay (W-ERED) and line-epoch-based re-encryption delay (L-ERED). Since the re-encryption frequency of a data line is usually dominated by its most frequently updated word, W-ERED introduces flag bits for local counters to identify frequently overflowing words. When a local counter overflows, only the overflowing word is re-encrypted with a new line counter, while non-overflowing words continue to use the old line counter, thereby avoiding unnecessary re-encryption operations. Based on this observation, L-ERED further exploits the fact that the re-encryption rate of a data page is mainly determined by the data line with the highest overflow frequency in that page. By introducing overflow flags for line counters, L-ERED ensures that only overflowing data lines are re-encrypted with the major counter upon overflow, thus significantly reducing the re-encryption granularity and overhead caused by line-counter overflow. Furthermore, to reduce write overhead and improve crash recovery capability, we propose an enhanced scheme, STAR+. STAR+ introduces a counter write-filtering mechanism that writes counters back to NVM only when the number of cache updates reaches a preset threshold, thereby reducing counter write operations. During crash recovery, STAR+ adopts the Osiris technique and uses shadow counters to record incremental updates. It reconstructs counter states by reading persisted counter values, replaying the logged increments, and verifying their correctness. In addition, STAR+ restores the integrity tree by updating the authentication values of leaf and parent nodes and recursively recalculating upper-level Hash values, thereby ensuring consistency between data and metadata. Experimental results show that, compared with SECRET, W-ERED reduces the average write bit-flip rate by 5%, write latency by 11%, and write energy by 6%, with only 8/512 additional storage overhead. In L-ERED, the number of re-encrypted lines is only 26.7% of that in BASE and 47.7% of that in RSR, with only 1/512 additional storage overhead. Moreover, compared with STAR, STAR+ reduces NVM writes by 9.8% and performance overhead by 3.1%, while achieving comparable recovery time.

  • Network and Information Security
  • Zhiqiang Zhang , Haiyan Wang , Lei Du , Chenyun Duan , Jiarui Li , Zhaoquan Gu , Binxing Fang
    doi: 10.7544/issn1000-1239.202550584

    With the deep integration of 5G, artificial intelligence, and the Internet of things, network infrastructure is facing the dual challenges of an expanding dynamic attack surface and increasing threat stealthiness. Traditional static defense mechanisms are no longer capable of effectively countering the evolving attack strategies. Although the existing active deception defense systems have enhanced threat detection capabilities through honey point technology, they still have significant shortcomings, particularly the lack of dynamic effectiveness evaluation and resource optimization mechanisms, which leads to a severe imbalance between defense costs and security benefits. Current research primarily focuses on the initial deployment strategies and architectural improvements of honey points; however, these studies have not effectively addressed the periodic attenuation of defense effectiveness in dynamic confrontation scenarios. There are also obvious deficiencies in the quantitative modeling of multidimensional security indicators and closed-loop optimization mechanisms. To address this research gap, we propose a multi-dimensional effectiveness representation mechanism for honey point cluster defense systems. The mechanism constructs a comprehensive effectiveness evaluation system with three core indicators: redundancy risk index, cross-honey point IP repetition rate, and threat entrapment effectiveness score. Specifically, the method first uses deep embedded clustering technology to obtain the prior labels of honey point effectiveness, providing basic data for subsequent effectiveness evaluation. It then introduces a weight adaptive optimization mechanism driven by genetic evolutionary strategies to dynamically adjust the weights of the core indicators according to the actual threat environment and defense requirements, thereby achieving precise and efficient effectiveness evaluation. On this basis, a closed-loop defense system of “evaluation-feedback-optimization” is further established, enabling the defense system to adjust defense strategies and resource allocation promptly based on real-time effectiveness evaluation results, thereby significantly enhancing the intelligence level of active deception defense systems. This closed-loop defense system provides a feasible theoretical paradigm for promoting the intelligent development of active deception defense systems and is expected to play an important role in the field of future network security defense.

  • Network and Information Security
  • Guifen Zhang , Qinchun Su , Hongzi Li , Lina Ge
    doi: 10.7544/issn1000-1239.202550637

    Adaptive gradient accumulation backdoor attack (AGABA) framework is designed for federated learning (FL) systems, integrating a parameterized adaptive subblock trigger (AST) with dynamic transparency and a multistage gradient accumulation (MGA) mechanism. This framework effectively addresses the persistent trade-off dilemma between stealthiness and persistence that plagues traditional backdoor attacks in the FL environment. AST decomposes the complete trigger into multiple independent components via dynamic transparency control and distributed subblock superposition technology, allowing malicious clients to collaboratively construct a global trigger pattern while maintaining a high level of attack stealthiness. MGA mechanism adopts a three-phase attack strategy including initial accumulation, gradient accumulation and attack execution, and combines it with a parameter importance-aware mechanism. It realizes the latent hiding and targeted activation of malicious updates in model aggregation through cross-round progressive gradient accumulation. Moreover, AGABA adopts momentum-accelerated gradient divergence propagation and adaptive memory factor adjustment, ensuring that all malicious gradient updates fall within the legitimate distribution interval of the FL system and thus evading statistical anomaly-based detection mechanisms effectively. Experiments show that with 20% malicious clients participating in FL training, AGABA can maintain a favorable backdoor attack success rate even under the protection of various mainstream defense mechanisms, and its overall performance outperforms single backdoor attack methods.

  • Network and Information Security
  • Jiao Zhang , Yirui Liu , Zhiyuan Kang , Junwei Liu , Lingkun Meng , Jun Xu , Xinghua Zhao , Zhongming Ji , Tao Huang
    doi: 10.7544/issn1000-1239.202550249

    Artificial intelligence data centers (AIDCs) have emerged as a new and increasingly critical form of computing infrastructure. Unlike traditional cloud data centers that primarily provide virtualization of general-purpose computing resources, AIDCs are designed to support high-performance, AI-centric workloads such as large-scale model training and inference. As AI models continue to grow and complex, AIDC networks are required to deliver unprecedented levels of bandwidth, low latency, and efficient cross-device coordination. We provide a comprehensive analysis of AIDC network characteristics, including service requirements, topology design, communication patterns, and traffic characteristics, and further examine the unique challenges that arise from these aspects. Building on a network layering perspective, we then present a structured overview of key enabling technologies for AIDC networks. These technologies include collective communication libraries, transmission control, load balancing, data-link flow control, and fault management. Representative academic studies and industrial implementations are summarized to illustrate their strengths, limitations, and suitability for real-world AIDC deployments. Finally, we outline several future development trends, such as the convergence of general-purpose and AI-oriented data centers, deeper specialization and co-design across AIDC hardware and software stacks, and the advancement of multi-tenant, shareable AI computing power. These trends are expected to play a pivotal role in shaping next-generation AI infrastructure.

  • Network and Information Security
  • Wenchao Li , Yinhan Sun , Tianci Yu , Liangfeng Tan , Gaohui Ma , Zhigang Sun
    doi: 10.7544/issn1000-1239.202550459

    Time sensitive networking (TSN) enhances traditional Ethernet with technologies such as high-precision time synchronization and time-aware shaping, delivering end-to-end deterministic data transmission services characterized by low latency, low jitter, and high reliability. It finds extensive applications in aerospace, industrial control, vehicle networks, and other scenarios. However, the inability of a large number of existing Ethernet end nodes to upgrade for TSN time synchronization has become a bottleneck restricting TSN deployment and application. To address this, FlyPTP method is introduced. Its core innovation decouples time synchronization from end-node architecture through a cooperative design, using lightweight software on endpoints and dedicated hardware in access switches. The theoretical validity of FlyPTP method is proven, and the hardware logic of FlyPTP and related software is developed based on FPGA. Experiments are designed to deploy FlyPTP on various heterogeneous end nodes. The results show that all tested end nodes can be deployed without modifying existing software and hardware, with a synchronization accuracy reaching 85 ns. This verifies the capability of FlyPTP to achieve TSN time synchronization for heterogeneous end nodes without altering their existing architectures. By separating time synchronization functions from the complex and diverse hardware/software environments of end nodes, FlyPTP provides a flexible and scalable solution for upgrading traditional Ethernet to TSN, overcoming the limitations of hardware dependency and kernel modification in conventional methods. This not only reduces the technical threshold for TSN deployment but also ensures compatibility and performance consistency across different platforms, laying a solid foundation for the large-scale application of TSN in industrial Internet and critical infrastructure scenarios.

  • High Performance Computing
  • Zhiheng Lin , Ke Meng , Guangming Tan
    doi: 10.7544/issn1000-1239.202550501

    Graph pattern mining, as a core task for analyzing entity relationships, finds extensive applications in fields such as community detection and drug discovery. However, due to its non-polynomial time complexity, large-scale graph processing encounters challenges related to search space explosion and low computational efficiency. Existing GPU-accelerated solutions utilize parallel computing to enhance performance but often suffer from thread idleness caused by the power-law distribution of graph data and excessive CPU intervention in multi-GPU communication, which impede the full utilization of hardware capabilities. We present Saturn, a GPU-based graph pattern mining system, which overcomes the above bottlenecks through computation-communication co-optimization. At computation level, Saturn decouples the loading and querying phases of set operations, employs fine-grained parallel batch processing to dynamically adapt Hash and binary search algorithms, and optimizes memory access efficiency via shared memory caching, significantly improving thread utilization. At communication level, it establishes a CPU-free GPU communication model, leveraging high-bandwidth NVLink interconnect and unified memory mechanism for direct cross-device data access, while reducing redundant transmission and link contention through vertex caching and message aggregation scheduling. Experimental results demonstrate that Saturn outperforms state-of-the-art GPU-based graph mining systems on real-world datasets, achieving superior computational parallel efficiency and cross-device communication effectiveness for large-scale graph pattern mining tasks.

  • High Performance Computing
  • Quanxi Li , Huimin Cui , Xiaobing Feng , Chenxi Wang
    doi: 10.7544/issn1000-1239.202550871

    Disaggregated memory architectures are rapidly emerging as a significant trend for next-generation data centers, promising to enhance resource utilization and scalability. Existing system software often employs rapid context switching between lightweight threads to generate concurrent remote data access requests, thereby masking the microsecond-level latency inherent in disaggregated memory. However, this approach introduces substantial thread scheduling and communication overhead, consuming a significant amount of CPU resources that would otherwise be dedicated to computation and thus creating a new performance bottleneck. To address this challenge, a userspace asynchronous data path for disaggregated memory is proposed. The proposed data path focuses on data prefetching and eviction, eliminating the high-overhead multi-threaded concurrency model in favor of an asynchronous I/O mechanism to achieve the overlap of computation and network I/O in fine-grained tasks. By combining adaptive batching with dynamic thread management strategies, CPU overhead is minimized while ensuring the timeliness and concurrency of remote memory requests, thereby effectively hiding the latency of disaggregated memory access. A prototype system is implemented on top of AIFM and evaluated with four typical application workloads. Experimental results show that, compared with AIFM, the proposed approach reduces the CPU overhead of data prefetching and eviction by 73% while simultaneously increasing the average end-to-end application throughput by 38%.