A pronounced imbalance in GPU resources exists on campus, where some laboratories own underutilized servers while others lack the compute needed for AI research. GPU sharing can alleviate this disparity, while existing platforms typically rely on centralized oversight and persistent allocation models, conflicting with the voluntary and autonomous nature of academic resource ownership. We present GPUnion, a campus-scale GPU sharing platform enabling voluntary participation while preserving full provider autonomy. GPUnion incorporates three core mechanisms: i) container-based task dispatching and execution, ii) resource provider-first architecture, and iii) resilient execution featuring automatic checkpointing and migration. Case studies across multiple campus scenarios demonstrate 30% more GPU utilization improvement, 40% increase in interactive sessions, and 94% successful workload migration during provider departures.

The prominent paradigm of federated learning (FL) is increasingly being applied to emerging and cross-silo applications, particularly with edge computing systems serving as pivotal agents. However, this shift also renders FL training more energy and carbon intensive. To this end, we propose GreenFL, a carbon-aware FL training framework designed to systematically navigate the trade-offs between carbon emission, training accuracy, and training efficiency. GreenFL employs a hybrid training strategy that combines inter-group asynchronous training and intra-group synchronous training to mitigate the straggler effect caused by inefficient participants. In the overall design of the framework, we promote the participation of edge computing nodes with abundant renewable energy sources and implement strategic participant selection to balance carbon emissions and training accuracy. We prove the solvability of optimizing the selection strategy and provide an online greed-based solution based on penalty values and bipartite greedy algorithms. Through extensive data-driven experiments, we demonstrate that GreenFL can significantly improve the carbon efficiency of the entire FL procedure, while maintaining or exceeding state-of-the-art levels of training accuracy and efficiency.

The geographically distributed edge servers can naturally draw power from nearby renewable energy (RE) generators. Complemented by the dynamic scheduling of energy storage batteries, edge service providers (ESPs) can thus build low- or even zero-carbon edge computing systems. Nevertheless, the distributed and heterogeneous nature of edge computing systems, as well as the limited information sharing among ESPs, leads to a more complex battery planning problem than that in cloud computing. The unpredictability of RE resources further complicates the problem, making conventional model-based approaches ineffective. To this end, we propose a multi-agent deep reinforcement learning (MADRL) approach for the independent decision making of individual ESPs. Particularly, MADRL takes privacy into account by ensuring that no sensitive information is disclosed among ESPs. For better model training, we further customize the invalid action masking and develop action transformation techniques based on segmented linear optimization. Extensive experiments demonstrate that, with our proposed approach, the overall carbon emission of edge computing systems can be significantly reduced (by over 60%) while maintaining acceptable operation costs in battery scheduling.

Geo-distributed edge data centers (EDCs) are expected to handle a large portion of tasks offloaded from cloud data centers for various emerging edge services. However, the high energy consumption and cost add a huge burden to edge service providers (ESPs). This presents a unique challenge as traditional energy-saving strategies applicable in cloud data centers fail to apply to EDCs, given the latency-sensitive nature of edge services. In response, we put forward an innovative electric vehicle (EV)-assisted edge computing architecture that leverages idle computing resources and stored energy of EVs. Our design aims to decrease energy expenditures for ESPs by choosing EVs with more economical service costs to handle a portion of the edge services during critical periods. We construct an energy cost-aware workload offloading model and discretize the original model into multiple small-scale solvable forms in both temporal and spatial dimensions. Furthermore, we reconfigure the Kuhn-Munkres algorithm to produce an online joint matching solution to counter QoS decline, generating a mutually advantageous situation for ESPs and EV participants. Upon experimentation with real-world traces, our design demonstrates a significant reduction in total energy cost (up to 31%) and offers considerable incentives for EV participants.

Edge computing is migrating services from remote clouds to the network edge, where a vast amount of data is also flowing into edge nodes. In this context, the Edge Data-Sharing System (EDSS) enhances service quality by enabling edge nodes to cooperate. However, the EDSS is suitable for precise search and faces the existing high overhead when many users retrieve similar data. To solve the obstacle, this paper proposes a similarity-based edge storage system, SESS, which leverages the software-defined edge network to realize efficient storage and retrieval of similarity data. We first design RealminHash, a core module of SESS, for efficient signature and and index for each data. Then, SESS calculates the storage strategy based on the similarity between data. Importantly, SESS adjusts this strategy using periodic network information to ensure load balancing. Experimental results demonstrate that SESS realizes the nearest-neighbor storage while maintaining load balancing. SESS outperforms the well-known k-means and spectral clustering methods in terms of accuracy and latency and supports millisecond similar queries.

Mobile-edge computing (MEC) is booming in recent years, as it is expected to fulfill the growing low-latency requirements of offloaded applications on large amounts of end devices. To ensure more applications finish before their deadlines, it is crucial to optimize the assignment of applications among various edge servers and the scheduling in a specific server, both of which have attracted widespread attention. As applications tend to become more complicated, each application may contain multiple interdependent tasks. We find that dependency in applications is another essential factor that would greatly impact the application’s time consumption. Unfortunately, no prior study has given a comprehensive consideration involving all three factors, leading to a severe waste of computing and network resources, thus delays application processing. In this article, we model the dependencies among all tasks of an application as a directed acyclic graph (DAG) and jointly optimize the application assigning and scheduling problems to facilitate the execution of each application. To solve this NP-hard problem, we design a novel method namedDaas. It first estimates the priority of each task and then effectively tackles the application assigning and scheduling problem based on this attribute in an online manner. Extensive evaluations show thatDaasperforms well in various experimental settings, enabling 20% more applications to meet their deadlines compared with the other baselines.

Video streaming service is one of the most important IoT applications/services at the mobile end. To provide better services and earn more customers, the mobile video service providers have paid considerable attention to enhance end users’ Quality of Experience (QoE) in video streaming. As an important aspect of the mobile device, however, the battery power and its impacts on the mobile services were seldom concerned. According to our survey over 2000+ mobile users, the low battery power of mobile phones could cause the user to give up watching videos. To quantify the relationship between the battery power and user’s video abandoning probability (VAP), we first extract the VAP model from the collected survey data, leveraging a reversed accumulative histogram approach. Then, referring to the quantified VAP model, we presentEdgeSaver, an edge-assisted video transmission framework, which aims at maintaining a sustainable overall user retention rate for the service providers by reducing the power consumption of video playback at the mobile ends. Particularly, as the core component ofEdgeSaver, a low-power video scheduler is designed to strategically select user groups, such that the most profitable outcome can be achieved under the constraints of limited edge resources. With extensive experiments using a real-world data set, we demonstrate thatEdgeSavercan help the mobile video service provider improve the user retention rate by up to 30% and increase the average user viewing time by 20%.

Vehicular cloud computing (VCC) has been widely utilized to enhance traffic management and road safety both in the static and dynamic scenarios. In VCC, the onboard resources of involved smart vehicles (SVs) would be integrated as a resource pool and collaboratively accommodate various computation tasks. In this system, SVs in a section of the road can be grouped as a vehicular cloud to complete the detection job, e.g., collecting and preprocessing the road data. Then, an SV will be selected to aggregate all SVs’ detection results in the road section and upload the results to the remote monitoring platform via base stations (BSs). However, due to the high dynamics of vehicular mobility and the poor coverage of BSs in some areas, this vision is challenging to achieve. In this article, we propose a collaborative road detection system and schedule the detection task to different SVs in the road section. Specifically, 1) we model the collaborative detection process and formulate it as a task scheduling problem to minimize the response time, which is NP-hard. An adaptive location-aware scheduling scheme is proposed for task scheduling and 2) as there are road sections without coverage of any BS, therefore we have to utilize the SVs on the opposite lane to transmit the aggregation result in these road sections. Accordingly, an uploading strategy is proposed to decide to upload the feedback through SVs on the opposite lane or in the coverage area of the next BS. Extensive experiments show that our scheme can significantly reduce the response time overall and is close to the optimal solution.