As microservice architectures become increasingly complex and system events become more frequent, Root Cause Analysis (RCA) has emerged as a critical task to ensure system reliability. However, existing deep learning-based methods often struggle with limited flexibility and a lack of interpretability when addressing complex system failures. Recent efforts to integrate large language models (LLMs) have shown promise in enhancing diagnostic transparency and reasoning capability. However, expansive search spaces, intricate workflows, and entangled constraints constrain practical adoption. We propose RCAFlow, a multi-agent framework that integrates structured workflow knowledge with hierarchical planning to address these challenges. RCAFlow transforms semi-structured documents into behavior tree-style workflows to support interpretable plan generation, employs a Git-inspired branching mechanism for modular and hierarchical task execution with path isolation, and leverages state-aware task execution with semantic analysis to improve result understanding and feedback. We evaluate RCAFlow on three benchmark datasets provided by OpenRCA. Experimental results demonstrate that RCAFlow consistently outperforms existing methods across all datasets. Further ablation studies confirm the effectiveness of each core module, highlighting the reliability, extensibility, and interpretability of RCAFlow to support complex RCA tasks within intelligent IT operations.

With the widespread deployment of large language models (LLMs) across diverse applications, optimizing their inference processes to achieve high throughput and low latency has become increasingly critical. While vLLM emerges as a high-performance inference engine that significantly accelerates processing speed, it continues to face challenges with elevated tail latency in high-concurrency scenarios. Sarathi-Serve addresses this through chunked-prefills and stall-free scheduling, achieving a balance between throughput and latency. However, our analysis reveals fundamental limitations in its static chunking strategy. This paper presents AdaptoServe, a system employing a dynamic adaptive chunk-based prefetching strategy to overcome the resource management and latency optimization limitations inherent in static chunking approaches. Our solution features intelligent chunk size adaptation through real-time monitoring of GPU utilization and work-load characteristics, ensuring optimal performance across varying resource conditions. Experimental evaluations demonstrate that AdaptoServe delivers substantial performance gains in diverse hardware configurations and work-load scenarios, effectively balancing throughput-latency tradeoffs. Compared to vLLM, AdaptoServe achieves up to 59.4 % higher request throughput, while outperforming Sarathi-Serve by$\mathbf{2 4. 4 \%}$through its dynamic chunking mechanism and enhanced multilevel feedback queue scheduling algorithm. We further investigate the strategy's applicability across different application scenarios, revealing significant optimization potential for next-generation LLM serving systems.

Time series forecasting for multiscale periodic data is critical in domains such as finance, meteorology, and energy, yet existing models struggle to effectively capture complex interactions across scales. Traditional approaches often fail to separate macro-scale periodic trends from micro-scale fluctuations, leading to limited forecast accuracy. To address these challenges, this paper introduces TimeShaper, a novel framework built upon the TimesNet architecture. TimeShaper employs sliding average techniques to decompose raw time series into macro-scale and micro-scale periodic components, thereby isolating key features and reducing noise. It further incorporates an innovative Causal Dilated Convolution module to extract intra-period and inter-period features from these decomposed components, enabling the model to utilize extensive historical periodic data while minimizing noise interference. Extensive experiments on multiple benchmark datasets demonstrate that TimeShaper consistently outperforms state-of-the-art methods in forecast accuracy, showcasing its robustness and practical applicability. By introducing this temporal decoupling and feature extraction methodology, TimeShaper provides a new perspective for tackling multiscale periodic time series forecasting tasks.

In recent years, super-resolution neural networks have achieved good results in restoring super-resolution images from low-resolution ones. However, most subsequent tasks based on super-resolution images such as object detection are done by the computer. Considering this situation, we propose a learnable downsampler–upsampler pair, which can realize both the downscaling process and the upscaling process by neural networks, and is jointly trained with the YOLOV5 network to optimize the object detection task. Thus, different from existing super-resolution networks, the entire downsampler–upsampler pair is optimized for machine perception. In addition, to further reduce the size of the downsampled images, we also propose a differentiable method for estimating image entropy and add it to the loss function. We verify the effectiveness of our method on the pothole dataset and use scale factors 2× and 4× to prove that our method is capable of diverse resizing levels. The experimental results show that using our learnable downsampler–upsampler pair as a resizing method can highly improve the detection performance compared with other resizing techniques.

In an unmanned aerial vehicle ad-hoc network (UANET), sparse and rapidly mobile unmanned aerial vehicles (UAVs)/nodes can dynamically change the UANET topology. This may lead to UANET service performance issues. In this study, for planning rapidly changing UAV swarms, we propose a dynamic value iteration network (DVIN) model trained using the episodic Q-learning method with the connection information of UANETs to generate a state value spread function, which enables UAVs/nodes to adapt to novel physical locations. We then evaluate the performance of the DVIN model and compare it with the non-dominated sorting genetic algorithm II and the exhaustive method. Simulation results demonstrate that the proposed model significantly reduces the decision-making time for UAV/node path planning with a high average success rate.

Modelling cellular traffic pattern plays a critical role to efficiently satisfy consumers’ demand, which is skewed distributed and fast-varying. Although current methods can indicate the traffic fluctuation of each cell, it is still not enough as optimisation techniques require to know the traffic distribution inside cells. In this paper, Neural Network is used to improve the resolution to intra-cell level by modelling the hotspots of geo-tagged Twitter. This paper is based on our previous work, which has already quantified the linear relationship between Tweets and mobile traffic. Here, a similarity measurement is designed to quantify how two patterns are similar to each other. We applied this measurement on the geo-tagged Tweets and found that in one of three periods (day, evening, and night), the hotspots distributions are more similar than the other periods. Then for each period, Generative Adversarial Networks (GAN) is used to train a generator for modelling the intra-cell hotspots distribution. Such a trained generator can also continuously generate convincing artificial-data to expand the data set. The similarity measurement gives high similarity (above 0.8) between generated artificial-data and the real test data.

In dense traffic unmanned aerial vehicle (UAV) ad-hoc networks, traffic congestion can cause increased delay and packet loss, which limit the performance of the networks; therefore, a traffic balancing strategy is required to control the traffic. In this study, we propose TQNGPSR, a traffic-aware Q-network enhanced geographic routing protocol based on greedy perimeter stateless routing (GPSR), for UAV ad-hoc networks. The protocol enforces a traffic balancing strategy using the congestion information of neighbors, and evaluates the quality of a wireless link by the Q-network algorithm, which is a reinforcement learning algorithm. Based on the evaluation of each wireless link, the protocol makes routing decisions in multiple available choices to reduce delay and decrease packet loss. We simulate the performance of TQNGPSR and compare it with AODV, OLSR, GPSR, and QNGPSR. Simulation results show that TQNGPSR obtains higher packet delivery ratios and lower end-to-end delays than GPSR and QNGPSR. In high node density scenarios, it also outperforms AODV and OLSR in terms of the packet delivery ratio, end-to-end delay, and throughput.

In this paper, we propose QNGPSR, a Q-network enhanced geographic routing protocol based on well-known GPSR for the ad-hoc network. By using the reinforcement learning and neighbor topology information to make next-hop selection in multiple available paths, the probability of exploiting perimeter forwarding mode that may induce network delay can be greatly reduced. We evaluate the performance of QNGPSR in comparison with OLSR, AODV and GPSR. The simulation results show that the proposed protocol obtains a higher packet delivery ratio and a lower end-to-end delay compared with the original GPSR. In high node density and high mobility scenarios, the performance of QNGPSR is better than OLSR and AODV.

Summary Storage quality‐of‐service (QoS) is a key issue for a storage cloud infrastructure. This paper presents QoSC, aQoS‐aware indiscriminate volumeStorageCloud over the dynamic network, based on the Hadoop distributed file system. QoSC employs a data redundancy policy based on indiscriminate recovery volumes and a QoS‐aware data placement strategy. We consider the QoS of a storage node as a combination of the transfer bandwidth, the availability of service, the workload (CPU utilization), the free storage space, and the failure rate of DataNodes. We have deployed QoSC on the campus network of Zhejiang University and have conducted a group of experiments and simulations on file storage and retrieval. The results show that QoSC improves the performance of file storage and retrieval and balances the workload among DataNodes, by being aware of QoS of DataNodes. Copyright © 2016 John Wiley & Sons, Ltd.

We present TLR, a traffic-light-based intelligent routing strategy for NGEO satellite IP networks. In TLR, a set of traffic lights are used to indicate the congestion status at both the current node and the next node. When a packet travels along a pre-calculated route to the destination, it may adjust the route dynamically, according to the real-time color of traffic lights at each intermediate node. Through the combination of preliminary planning and real-time adjustment, each packet can eventually get an approximately optimal transmission path. The multi-path routing mechanism in TLR can help achieve a good distribution of traffics when the network traffic increases. The Public Waiting Queue scheme in TLR can fully utilize free spaces of the buffer queues and lower the packet drop rate. While the concept of TLR has many advantages, it may result in endless-loop of routing. To eliminate this phenomenon, a defense scheme is incorporated in the design of TLR. A set of simulations are conducted using the Network Simulator (version 2) to verify the good performance of TLR, in terms of lower packet drop rate, better distribution of traffics and higher throughput, over the entire satellite constellation.

We propose an incentive model based on information-hiding to encourage peers to vote for resources in peer-to-peer (P2P) networks. The following are key motives for our model: (1) Some trust and reputation systems have been deployed in modern P2P systems, but a lot of blank rating resources exist in these P2P systems; (2) E-commerce consumer-to-consumer (C2C) websites that adopt simple rating strategies are receiving accusations that false and useless ratings are flooded. We establish an information-hiding based RRR/RIR (resource reputation rating/reputation incentive rating) voting model, which awards or punishes voters according to their behaviors. The RRR generating algorithm and the RIR generating algorithm are presented in detail, and the information-hiding mechanism is given. Experimental results showed that the incentive RRR/RIR model can effectively encourage valid voting and prevent malicious or arbitrary voting in the P2P reputation system.