Given the emerging needs of artificial intelligence in both commercial and consumer uses, the rapid increase of energy consumption has become a major concern for data centers. Workload scheduling plays a crucial role in reducing the energy costs associated with large compute-intensive workloads in data centers. However, there has not been much research on reducing the energy costs for GPU-intensive workloads specifically. In this paper, we proposed two thermal-ware scheduling algorithms (GPUThermalAwareMinTemp and GPUThermalAware) for GPU-intensive workloads. Through the integration of machine learning models into a modified version of the GPUCloudSimPlus simulator, we compare the performance of our algorithms with three baseline algorithms by simulating cloud computing environments using a subset of real-world workload traces from Alibaba PAI. Our experimental results show that, compared with the worst-case algorithms, the GPU thermal-aware scheduling algorithms achieved 4.72% lower average host minimum temperatures and saved total energy by 1.17%.
To overcome signal blockage and extend coverage to both indoor and outdoor users, simultaneous transmitting and reflecting reconfigurable intelligent surfaces (STAR-RIS) are employed. In addition, Non-Orthogonal Multiple Access (NOMA) is integrated to improve spectral efficiency and support multipleuser access. Given this setup, it is essential to investigate the Physical Layer Security (PLS) performance of STAR-RIS-assisted NOMA systems, particularly with multiple eavesdroppers. To evaluate the signals' interception risks, we derive closed-form expressions for the Secrecy Outage Probability (SOP) and the Intercept Probability (IP) for both system setups at the near users and colluding/non-colluding eavesdroppers.
Satellite communication networks - including Low Earth Orbit (LEO) and Geostationary Earth Orbit (GEO) systems - are becoming increasingly vital for global connectivity, particularly in remote and underserved areas. Despite their growing importance, individual satellite links often suffer from high latency, limited bandwidth, and fluctuating reliability, which can significantly degrade application performance. Multipath Quick UDP Internet Connection (MPQUIC) has emerged as a promising solution for aggregating multiple satellite paths and improving network utilization. However, two major challenges limit the effectiveness of existing MPQUIC schedulers: the heterogeneity of link characteristics across satellite types and the periodicity of LEO link behavior. To address these issues, this paper proposes Logr alpha, a novel MPQUIC scheduler that integrates Deep Reinforcement Learning (DRL) to dynamically optimize data distribution and manage adaptation speed. Logr alpha intelligently adjusts path usage to account for link diversity and adaptively adjusts the growth rate of the congestion window to respond quickly to periodic variations in LEO links. Experimental results show that Logr alpha significantly enhances adaptation speed and reduces file transfer time compared to traditional scheduling strategies.
A neural architecture for user-centric vehicular caching is introduced, termed as CacheFormer. It leverages autoformer for cache content prediction that positions vehicles as autonomous micro-servers that operate independently within personal area networks. The architecture employs multi-horizon prediction with five-minute slot granularity. It enables lookahead through decomposition of application patterns into trend and seasonal components. Integration of sparse auto-correlation mechanisms yields in reduced complexity while maintaining prediction accuracy. Further, an age-aware eviction mechanism combines forecasted demand with cache staleness metrics, which eliminates separate prediction and eviction modules. Experimental evaluation on 3,463 application sessions demonstrates convergence within 50 epochs in 27.5 minutes on consumer hardware, at high hit rates post-convergence across all popularity quartiles. Moreover, bimodal correlation distribution reveals successful temporal pattern extraction for 75% of applications. Results validate that adopting autoformer improves cache efficiency while preserving data sovereignty and eliminating dependencies on external cloud infrastructure during user mobility.
With the increasing adoption of Unmanned Aerial Vehicles (UAVs) and their technological advancements, particularly in terms of autonomy and computational capabilities, their role as active component in wireless networks could be groundbreaking. Such potential is especially relevant in constrained environments that lack proper telecommunication infrastructure or adequate network coverage. These contexts are where the Internet of Things (IoT) has emerged as a vital technology, capable of facilitating sensing operations and data retrieval by leveraging communication protocols like LoRa for long-range, low-energy transmissions. In this paper, we introduce the Traveland-Offload Problem (TOP), a new combinatorial optimization problem, formulated both as a mixed-integer nonlinear programming (MINLP) model and as a mixed-integer linear programming (MILP) model. Its primary objective is to minimize the energy consumption associated with UAV operations and data offloading tasks. Our formulation, which incorporates realistic models for energy consumption and communication, aims to optimally determine the UAV’s trajectory, hovering points and durations, as well as critical transmission parameters of LoRa-enabled sensors, while ensuring compliance with the transmission policies defined by the LoRaWAN network protocol.
