Micro datacenters (μDCs) are emerging miniaturized datacenters poised to revolutionize distributed computing in edge environments by seamlessly integrating compact, modular architectures with localized computing, storage, and networking capabilities. This survey provides a complete and up-to-date review of μDC. We introduce a comprehensive three-layer analytical framework essential for understanding and evaluating μDC technologies: 1) The infrastructure layer empowers μDCs to adapt physical resources, including compute units; 2) The platform layer prioritizes middleware-driven resource orchestration, driving improvements in computational efficiency and energy optimization; 3) At the application layer, μDCs are committed to delivering performance optimizations that meet demanding requirements for latency, reliability, and energy efficiency. This work synthesizes leading-edge implementations and technology trends, clearly outlining critical trade-offs, identifying significant challenges, and defining the research frontiers vital for advancing μDC design and deployment across diverse industrial and extreme environments.

Ultra-Reliable Low-Latency Communication (URLLC) services in forthcoming 5G/6G networks require millisecond responsiveness and very high continuity. Multi-Access Edge Computing (MEC) combined with Network Function Virtualization (NFV) offers the needed proximity, yet orchestrating Service Function Chains (SFCs) at the edge must meet strict latency and reliability targets while limiting Operational Expenditure (OPEX). We introduce CERLA-SFC, a hierarchical, multi-objective orchestrator that unifies learning-based placement, topology-aware routing and event-driven resource allocation in a single control loop. An Advantage Actor–Critic (A2C) plane selects Virtual Network Function (VNF) locations, a shortest-path mapper embeds the inter-function links, and a Karush–Kuhn–Tucker (KKT) resource plane retunes CPU, memory and bandwidth slices on demand. A redundancy module adds selective replication, instance sharing, increasing fault tolerance at marginal cost. Instance reuse and proactive VNF redeployment reduces initialization latency and instance overhead. Trace-driven experiments on twenty heterogeneous edge servers plus one cloud node, with chains of three to six functions, validate the design. Compared with six representative baselines, including Genetic Algorithm (GA), Particle Swarm Optimization (PSO), Greedy, and Proximal Policy Optimization (PPO), CERLA-SFC cuts OPEX by 10% to 53% under fixed loads of five to twenty requests per slot and by 15% to 53% when the load fluctuates between ten and twenty requests. Event-driven slicing alone lowers average cost by 51.46%, and selective redundancy trims backup expense by 22.1%. The framework maintains near-zero latency violations across all urgency classes while keeping end-to-end delay in the millisecond range.

Graph Neural Networks (GNNs) have achieved remarkable success in various applications. Sampling-based GNN training, which conducts mini-batch training on sampled subgraphs, has become a promising solution for large-scale graphs. Given the resource-intensive nature of sampling-based GNN training, Neural Processing Units (NPUs), such as the Ascend AI processor, offer a promising alternative due to their high throughput and energy efficiency, making them well-suited for GNN workloads. However, the multi-stage nature of sampling-based training, which involves subgraph sampling, feature gathering, and model training, with different resource requirements and computation volume. This requires careful coordination to fully utilize the heterogeneous computation resources of CPUs and NPUs. In this work, we present AcOrch, a sampling-based GNN training system optimized for CPU-NPU heterogeneous platforms. AcOrch offers fine-grained task orchestration and adopts a two-level pipelined execution model to overlap sampling, gathering, and training. It analyzes the heterogeneous compute features of NPUs and maps tasks to AI Cube (AIC) units, AI Vector (AIV) units, and CPU cores accordingly. Moreover, the two-level pipeline enables overlapping execution not only between the CPU and NPU, but also among different types of compute units within the NPU (e.g., AIC and AIV units), thereby maximizing the utilization of available resources. Experiments on an Ascend 910B AI processor show that AcOrch achieves an average speedup of 2.31\times over the state-of-the-art NPU-native graph learning system, MindSporeGL.

For large-scale parallel programs, intricate software behavior and complex hardware architecture lead to synchronized clocks across multiple nodes, resulting in misaligned traces for each node across the timeline, which is also known as time skew. This misalignment hampers various analyses along the temporal dimension, making it challenging to effectively optimize the performance of parallel programs. Furthermore, the time alignment across a massive amount of processes imposes a substantial computational burden, rendering existing solvers inadequate in massively parallel scenarios. In this paper, we propose TLBERT, a novel approach for timeline alignment that incorporates machine learning techniques with a well-designed training methodology. Based on TLBERT, we implement STAR, a Large-Scale Synced Trace Timeline Aligner tool to tackle the time skew problem for large-scale parallel programs. Experimental results demonstrate that STAR achieves timeline alignment for traces of large-scale programs with minimal overhead, effectively mitigating the time skew problem.

Real-time health monitoring via wearable devices has become increasingly essential for personalized health management. However, existing physiological signal processing methods, particularly for EEG data, focus primarily on frequency domain features, which can lead to lower monitoring accuracy. To address these limitations, we propose a novel personalized health monitoring framework that integrates both frequency and spatio-temporal characteristics of physiological signals. Within this framework, we further propose a deep learning model called the Spectral-Spatial Attention and Frequency Feature Fusion Network (SSAFNet). SSAFNet consists of three key modules: cross-frequency feature extraction, spatio-temporal feature extraction, and feature fusion, which together analyze frequency and spatio-temporal physiological data, enabling more precise real-time monitoring and improving the effectiveness of health management. Using this framework, we conducted several experiments to identify key patterns in EEG signals that effectively reflect individual health conditions and compared them with traditional health monitoring methods. The results demonstrate significant differences in EEG patterns across individuals, and the proposed framework outperforms existing methods in personalized health monitoring, showing its effectiveness and potential for widespread applications.