With the rapid advancement of information technology and artificial intelligence, the continual generation of massive data has led to emerging demands on the quality of service for real-time processing of large-scale data. Distributed computing pushes storage and computation closer to end devices, cutting transmission delays and balancing resources with proximity, making it a key enabler for large-scale real-time analytics. Yet exponential device growth and ballooning data volumes have created storage and communication bottlenecks at computing nodes. This paper focuses on a scenario where computing nodes connect with multiple data source nodes for real-time analysis of large-scale data. By constructing a mathematical model, we formulate the problem as an optimization task aimed at minimizing average task completion time under resource, latency, and sequence constraints. We prove it is NP-hard and faces coupled resource decisions and dynamic networks. To tackle these issues, we propose a data prefetching strategy that integrates online learning prediction with a Deep Q-Network (DQN). The online learning algorithm predicts transmission rates to provide future state information to the DQN agent. Based on the system model, we design the state, action space, and reward function within the reinforcement learning framework, leveraging optimization mechanisms like experience replay and double networks. Experiments conducted in a simulated complex computing node environment demonstrate that, compared to baseline strategies, our approach reduces average task completion time, enhances system resource utilization and service quality, and exhibits strong adaptability to fluctuating network conditions.

In cloud-based disaggregated compute and storage architectures, deep learning training tasks must read data over network links between compute and storage clusters. Bandwidth bottlenecks can lead to I/O blocking, causing GPU idle time and increased training costs. Existing resource scheduling approaches either focus on compute resource allocation while ignoring caching and bandwidth coordination, or adopt heuristic strategies without theoretical guarantees, making global optimization difficult under complex multi-cluster constraints. This paper considers the elastic compute resource scenario where GPU allocations can be dynamically adjusted, and proposes COPR-CBOA, a collaborative resource allocation algorithm based on polynomial optimization and sparse semidefinite programming relaxation. The algorithm incorporates GPU quotas, cache capacities, bandwidth allocations, and data scheduling ratios into a unified optimization framework. It handles nonlinear components such as ideal throughput, utility functions, and bandwidth costs via polynomial approximation, and constructs a sparse SDP relaxation using variable coupling graphs and hierarchical clique decomposition, maximizing total social welfare under budget and resource capacity constraints. Theoretical analysis provides exponential convergence bounds on polynomial approximation errors, convergence guarantees for the sparse SDP relaxation, computational complexity, and asymptotic optimality. Extensive experiments on real production workloads show that COPR-CBOA improves total social welfare by 1.28× over the best baseline Gandiva, reduces average job completion time by 10.9% compared to the Random strategy, and maintains zero I/O blocking throughout the entire experimental period.

Video analysis models based on deep neural networks have a significant demand for GPU resources during training and inference. Especially in real-time video analysis scenarios, the models are usually deployed on edge servers close to the end-users. However, the models are vulnerable to data drift, and the inference accuracy gradually declines as the content of the video stream changes. To maintain the inference performance of the models, continuous learning is required. Due to business requirements, multiple models need to be deployed simultaneously on edge servers. In the resource-constrained environment of edge servers, it is extremely challenging to allocate GPU resources among multiple parallel models and between the inference and retraining tasks of the same model to maximize the overall benefit. To address this issue, this paper presents a first-of-its-kind algorithm based on partial ordering of Resource Fluctuation Sensitivity(RFS), and successfully constructs a GPU resource scheduling system prototype named Octopus. This system adopts the RFS-Aware Algorithm(RFSA) with the aim of optimizing the GPU resource allocation problem during the parallel inference and retraining of multiple models. Experimental results show that, compared with the baseline method of evenly allocating GPU resources, the GPU resource scheduling of the Octopus system can increase the model inference accuracy by 34.19%. Moreover, under the same experimental environment, compared with the existing advanced solutions Ekya and AdaInf, the overall performance of the Octopus system can be, on average, twice as high as theirs.

While deep neural networks (DNNs) are increasingly deployed on graphics processing units (GPUs) at the network edge, the performance interference caused by co-locating and executing multiple models on a single GPU is often overlooked, leading to high inference latency and excessive energy consumption. Pursuing the best performance faces several challenges, including characterizing the non-linear interference-incurred latency, dispatching the unpredictable inference workloads, and balancing the different trade-offs related to performance and accuracy. In this study, we first construct an interference-incurred latency model via real-world profiling and measurements. With insights from this model, we formulate a time-varying integer program to minimize the long-term total cost of the edge AI inference system, including the inference latency, the inference error rate, the query-dispatching communication cost, and the energy consumption, subject to resource and workload constraints. We then propose a set of polynomial-time online algorithms that continuously make fractional control decisions for each time slot based on the feedback from the previously applied decisions and strategically round these decisions into integers. We conduct a formal theoretical analysis and prove that both the time-averaged gap between the total cost incurred by our online decisions and the total cost of the series of one-shot offline optimums and the time-averaged constraint violation gradually diminish over time. Extensive experiments on real-world testbeds and datasets demonstrate that the proposed algorithms can reduce the total cost by 40

Serving text-to-image (T2I) generation on edge servers can reduce the serving latency and enhance the preservation of user privacy. Currently, mainstream T2I models generate images iteratively, and the more iterations, usually the higher the generation quality, but the longer the generation latency. Existing T2I serving approaches simply adopt a fixed iteration number or employ large iteration numbers to merely optimize the generation quality, failing to adapt to the time-varying workload at the edge. Besides, they do not consider the edge computing environment where computing capabilities and network connections of edge servers are always heterogeneous. To make up for these drawbacks, we propose the problem of latency optimization for T2I serving at the edge under the long-term generation quality constraint and design EdgeT2I to address it. EdgeT2I employs Lyapunov optimization to decompose the long-term problem into a series of real-time sub-problems and leverages a Markov approximation (MA)-based algorithm to jointly determine the iteration number and offloading scheme of requests for each sub-problem. By doing so, EdgeT2I can adaptively adjust the iteration number to handle the time-varying workload by appropriately sacrificing generation quality to reduce serving latency during the high workload and compensating for generation quality during the low workload, as well as optimizing request offloading at the heterogeneous edge. Extensive trace-driven experiments confirm that compared to baselines, EdgeT2I can decrease the serving latency by up to 99.8% while maintaining the generation quality.

Autonomous driving increasingly relies on 3D video analytics to perceive complex environments, where limited on-board computation makes edge-cloud collaboration a common practice. However, constrained bandwidth necessitates aggressive compression to mitigate transmission costs, which substantially degrades perception accuracy since conventional codecs optimized for visual fidelity often overlook task-relevant semantics. We present NEVA, a dual-function neural codec that integrates feature extraction and compression through end-to-end co-optimization with downstream detection networks. Unlike existing codecs, NEVA employs an asymmetric architecture tailored to the uneven computational resources of edge-cloud systems. Instead of optimizing compression only for geometric reconstruction or adapting bitrate alone, NEVA jointly exposes compression strength and encoder complexity as controllable operating points under a task-oriented objective. Its lightweight edge encoder converts sparse point clouds into compact pseudo-video latents with bounded edge computation, while its cloud-side decoder integrates temporal restoration into the normal reconstruction path to recover density and semantics useful for perception. An adaptive controller then selects matched encoder–decoder variants according to bandwidth, edge compute availability, and measured analytics accuracy. Experiments show that, NEVA improves system throughput by 150%–460% on the edge-device client, while satisfying the application-specified perception-accuracy requirement.

Vision-language models are increasingly deployed as computer-use agents (CUAs) that operate desktops and browsers. Top-performing CUAs are framework-based systems that decompose planning and execution, while end-to-end screenshot-to-action policies are easier to deploy but lag behind on benchmarks such as OSWorld-Verified. GUI datasets like OSWorld pose two bottlenecks: they expose only a few hundred interactive, verifiable tasks and environments, and expert trajectories must be gathered by interacting with these environments, making such data hard to scale. We therefore ask how reinforcement learning from verifiable rewards (RLVR) can best exploit a small pool of exist expert trajectories to train end-to-end policies. Na"ively mixing these off-policy traces into on-policy RLVR is brittle: even after format conversion, expert trajectories exhibit structural mismatch and distribution shift from the learner. We propose BEPA ( Bi -Level E xpert-to- P olicy A ssimilation), which turns static expert traces into policy-aligned guidance via self-rolled reachable trajectories under the base policy (LEVEL-1) and a per-task, dynamically updated cache used in RLVR (LEVEL-2). On OSWorld-Verified, BEPA improves UITARS1.5-7B success from 22.87% to 32.13% and raises a held-out split from 5.74% to 10.30%, with consistent gains on MMBench-GUI and Online-Mind2Web. Our code and data are available at an anonymous repository: https://anonymous.4open.science/r/ACL_BEPA .

As AI training moves toward geo-distributed elastic infrastructures, its bottlenecks extend beyond GPU availability to variable energy and network conditions. The key challenge is translating heterogeneous placement, scaling, congestion, and training-stage effects into useful progress. We propose Effective Compute Utility (ECU), an interpretable progress signal, and build ECU-Aware, an online scheduler that coordinates data-center placement, elastic GPU allocation, and execution timing under budget and deadline constraints. Using Lyapunov optimization, ECU-Aware balances operating cost against long-term progress. Experiments with real traces and a small-scale prototype show that ECU-Aware improves the cost-latency-SLA tradeoff. In a 500-job workload, it reduces cost by 48.9 percent relative to a static policy, lowers average JCT by 36.0 percent, and decreases deadline misses by 17.2 percentage points compared with practical baselines.

Federated domain-specific instruction tuning (FedDIT) for large language models (LLMs) aims to enhance performance in specialized domains using distributed private and limited data, yet identifying key performance drivers and optimal augmentation strategies remains challenging. We empirically establish that cross-client domain coverage, rather than data heterogeneity, is the pivotal factor. We then introduce FedDCA, an algorithm that explicitly maximizes this coverage through diversity-oriented client center selection and retrieval-based augmentation, constructing diverse, non-redundant cross-client instruction sets. Extensive experiments across multiple domains demonstrate FedDCA's superiority over eleven baselines, achieving performance gains of up to 29.19% and domain coverage improvements of 4.82%-21.36%. FedDCA maintains its effectiveness in diverse and challenging scenarios, including data selection, held-out settings where task-specific public data is scarce and various data heterogeneity, with manageable privacy risks. This work clarifies critical FedDIT dynamics and presents FedDCA as an effective, privacy-preserving, and scalable solution for advancing domain-specific LLM tuning.

To enable edge AI providers to recruit edge devices and use them to deploy AI models and provision inference services, we conduct a comprehensive mathematical and algorithmic study on a novel incentive and optimization mechanism based on repeated auctions. We first model and formulate a time-cumulative social cost optimization problem to capture the challenges of the trade-off between cost and accuracy, the dependency between adjacent auctions, and the need of achieving desired economic properties. Then, to solve this intractable non-linear integer program in an online manner, we design a set of polynomial-time algorithms that work together. Our approach dynamically chooses and switches winning bids under careful control, incorporates online learning to overcome posterior inference accuracy and workload queue dynamics, and leverages randomization to strategically convert fractional decisions of model placement and query dispatch into integers. We also allocate payments to meet the necessary and sufficient conditions for the desired economic properties. Further, we rigorously prove the constant competitive ratio, the sub-linear regret and fit, and the truthfulness and individual rationality for our proposed approach. Finally, through extensive experiments using real devices, AI models, and data traces, we have validated the substantial advantages of our proposed approach compared to the baselines and the state-of-the-art methods.

Inference Pipeline Serving Systems (IPSS) have been widely adopted to support complex deep neural network (DNN) applications at the edge, where resources are constrained and request patterns exhibit high variability. These systems execute a sequence of inference tasks across heterogeneous stages, enabling coordinated multimodel decision-making in latency-sensitive applications such as robotics and autonomous systems. However, most existing IPSS approaches rely on a priori scheduling decisions based on predicted workloads. Such strategies are vulnerable to performance degradation under unpredictable or bursty request arrivals, which are common in edge environments. To address these limitations, we propose PRAS (Post-hoc Request Adaptive Scheduling), a novel online scheduling framework tailored for edge-based IPSS. PRAS dynamically configures model selection and batch size decisions for accumulated requests in a post hoc manner, without relying on prior workload predictions. PRAS tackles the key challenge of combinatorial scheduling complexity via a multilevel decoupled strategy: (i) constructing a FIFO-style request queue ordered by remaining completion time, (ii) formulating stage-wise model configuration as a path selection task over an augmented graph, and (iii) applying Multi-Armed Bandits (MAB) for online profiling of pipeline accuracy. We provide a theoretical guarantee of the asymptotic optimality of PRAS and demonstrate through extensive edge-deployment experiments that PRAS significantly improves accuracy and goodput under real-world request traces, achieving over 3x reduction in accuracy loss compared to baseline methods.

Distributed machine learning systems train models via iterative updates between parallel workers and the pa-rameter server.To expedite the transmissions,in-network aggregation of updates along with the packet forwarding at those programmable switches decreases the network traffic over these bottleneck links.However,existing in-network ag-gregation schemas do not adequately prepare the most suitable switches for various worker distributions and fail to cap-ture the dynamic network status.Based on the status derived from in-band network telemetry,we aim to select the best switches upon the optimization we formulate with the objective of minimum transmission latency.Although the problem is actually a non-linear integer program,by adopting delicate transformations,a substitute with totally unimodular con-straints and separable convex objective is then solved to obtain the integral optimum.We implement our in-network ag-gregation protocol and reconstruct in-band network telemetry protocol upon real devices,i.e.,Barefoot Wedge100BF and Dell servers.Then,we evaluate the performance of our proposed AGG algorithm and the results indicate that the comple-tion of related coflows decreases 40％on average compared with other strategies,improving at least 30％performance,compared with the state-of-the-art.

In this work, we introduce HeFT (Head-Frequency Tracker), a zero-shot point tracking framework that leverages the visual priors of pretrained video diffusion models. To better understand how they encode spatiotemporal information, we analyze the internal representations of Video Diffusion Transformer (VDiT). Our analysis reveals that attention heads act as minimal functional units with distinct specializations for matching, semantic understanding, and positional encoding. Additionally, we find that the low-frequency components in VDiT features are crucial for establishing correspondences, whereas the high-frequency components tend to introduce noise. Building on these insights, we propose a head- and frequency-aware feature selection strategy that jointly selects the most informative attention head and low-frequency components to enhance tracking performance. Specifically, our method extracts discriminative features through single-step denoising, applies feature selection, and employs soft-argmax localization with forward-backward consistency checks for correspondence estimation. Extensive experiments on TAP-Vid benchmarks demonstrate that HeFT achieves state-of-the-art zero-shot tracking performance, approaching the accuracy of supervised methods while eliminating the need for annotated training data. Our work further underscores the promise of video diffusion models as powerful foundation models for a wide range of downstream tasks, paving the way toward unified visual foundation models.

Virtual Reality Serious Games (VR-SGs) integrate immersive virtual reality (VR) technology with instructionoriented serious games (SGs), aiming to improve the efficiency of educational and training programs. VR-SG's training effectiveness is highly contingent upon the system's continuous consistency level. The strong continuous consistency ensures the same VR-SG world among different players, enabling them to make better decisions based on the individual game world's context. Although edge computing enables a low-delay VR system for geographically dispersed players, the delay differences among players highlight the need for strong continuous consistency. Specifically, the differences in temporal and spatial dimensions among different endplayers result in significant variations in their perceived end-to-end delay, further exhibiting different game worlds. We first propose a long-term task redistribution problem to enhance the continuous consistency for edge-assisted VR-SGs while controlling the consistency loss and player-perceived delay. To solve the above time-coupled problem, we design an online polynomial-time algorithm called the Online Continuous Consistency Enhancement (OCCE) algorithm. OCCE can effectively obtain the task redistribution scheme with the integrated randomized rounding and the Constraints-Firefighter Algorithm. We prove that the continuous consistency optimality of OCCE can approximate the optimal offline solution. Finally, the extensive evaluations based on real-world datasets and preparatory measurements show that, at the player scale of 30, OCCE improves continuous consistency by at least 2.38x compared to alternatives in the average case.

Increased videos captured by widely deployed cameras are being analyzed by computer vision-based Deep Neural Networks (DNNs) on servers rather than being streamed for humans. Unfortunately, the conventional codecs (e.g., H.26x and MPEG-x) originally designed for video streaming lack content-aware feature extraction and hinder machine-centric video analytics, making it difficult to achieve the required high accuracy with tolerable delay. Neural codecs (e.g., autoencoder) now hold impressive compression performance and have been widely advocated in video streaming. While autoencoder shows transformative potential, the application in video analytics is hampered by low accuracy in detecting small objects of highresolution videos and the serious challenges posed by multivideo streaming. To this end, we propose AdaStreamer with adaptive neural codecs to enable real machine-centric highaccuracy multi-video analytics. We also investigate how to achieve optimal accuracy under delay constraints via careful scheduling in Compression Ratios (CRs, the ratio of the compressed size to the original data size) and bandwidth allocation, and further propose a Markov-based Adaptive Compression and Bandwidth Allocation algorithm (MACBA). We have practically developed a prototype of AdaStreamer, based on which extensive experiments verify its accuracy improvement (up to 15%) compared to stateof-the-art coding and streaming solutions.

Video analytics pipelines (VAPs) have been a paradigm for large-scale video analytics. Due to temporal redundancy in video, frame filtering is widely used in VAPs to reduce analysis workload. However, existing works overlook a limitation: while inference operates only on selected frames, decoders must still process many redundant frames due to codec dependencies, leading to over-decoding trap. This limitation stems from the reference-based design in modern codecs, which require decoding preceding frames to reconstruct any selected one. As a result, over-decoding has become the practical bottleneck in VAPs using modern decoders, highlighting a critical but under-explored problem. To address this issue, we propose ParaDeco, a high-throughput video analytics framework featuring a novel frame-level parallel generative decoder. Unlike traditional decoders, ParaDeco adopts a decode-what-matters approach with decoupled frame dependencies. To decode arbitrary frames independently, ParaDeco generates frame-wise features as stand-alone skeletons using compressed video metadata, then predicts pseudo frames maintaining semantic consistency with original frames. Moreover, ParaDeco identifies which frames truly matter for analysis via delicate contribution-based frame filtering. We implement ParaDeco on a cloud server and evaluate it on large-scale real-world video datasets. Our experimental results show that ParaDeco achieves a 2.76x speed-up on average compared to state-of-the-art VAPs.

Edge computing is a rapidly developing research area known for its ability to reduce latency and improve energy efficiency, and it also has a potential for green computing. Many geographically distributed edge servers are powered by renewable energy sources, due to the difficulties of using traditional power supplies or because of advancements in energy harvesting technologies. These green edge servers can cut down carbon emissions by processing tasks locally, but the inherent limitations of their computing capacity result in some tasks having to be uploaded to a data center to meet service-level agreement (SLA) requirements. To further reduce carbon emissions in cloud-edge systems, scheduling tasks to those low-carbon data centers while meeting latency constraints is highly beneficial. In this paper, we propose a low-carbon cloud-edge scheduling algorithm that utilizes Lyapunov optimization techniques and Markov approximation to address the long-term optimization problem of carbon emissions. Our algorithm guarantees provable performance, and simulation results demonstrate its effectiveness in striking a balance between carbon emissions and task latency.

Deep neural network (DNN)-based real-time video analytics service, as a core module for numerous crucial applications such as augmented reality (AR), has garnered increasing research attention, where mobile edge computing (MEC) is often leveraged to mitigate its real-time processing burden on resourceconstrained user devices. For Quality of Experience (QoE) optimization, latest works employ reinforcement learning (RL)-based methods to adaptively adjust configurations (e.g., resolution and frame rate), yet still presenting significant challenges. Firstly, we observe a substantial diversity in QoE patterns among users. Given that existing methods integrate a fixed QoE pattern in parameter training, it is intuitive to customize a policy network for each user. However, this necessitates significant training investment, failing to support on-the-fly deployment for new users. Secondly, given the dual dynamics from both the network and video content in edge video analytics system, existing methods often fall into the dilemma of fitting newly emerged and diverse system states with offline-trained fixed parameters. While it is promising to employ online learning algorithms, most of them struggle to catch up with the high dynamics. We hence propose Mystique. In real-time edge video analytics domain, it is the first meta-RL-based user-level configuration adaptation framework. Mystique establishes an initial model in offline meta training with model-agnostic meta-learning (MAML), enabling swift online adaptation to new users and system states through limited gradient updates from initial parameters. Comprehensive experiments illustrate that Mystique can improve QoE by 42% on average compared to prior works.

Fueled by the popularity of mobile devices, mobile channels have become the preferred video delivery medium. However, users often encounter a poor quality of experience (QoE) due to bandwidth limitations, despite the implementation of adaptive bitrate (ABR) techniques. Recent advancements in super-resolution (SR) models have offered a potential solution to this situation, while the process of SR on mobile devices can introduce energy overhead and latency. To tackle these issues, we present ABUV, a system designed to enhance mobile video streaming by integrating adaptive bitrate and super-resolution technologies. ABUV leverages deep reinforcement learning to dynamically adjust both the bitrate and upsample decisions jointly based on considerations such as energy overhead and available bandwidth. It employs an optimized SR model specifically tailored for mobile devices, selectively applying the upsample process to chosen frames. Additionally, ABUV incorporates user-specific streaming information and adapts to the unique network environment through online training. In our experiments, we evaluate ABUV using real network traces and a diverse collection of videos, and the results show that ABUV can save up to 59% of data consumption and improve QoE by 27% compared to other video streaming systems.