Training large-scale deep learning models, particularly large language models (LLMs), demands extensive computational resources. As dedicated GPU clusters become increasingly cost-prohibitive, decentralized training that leverages underutilized or consumer-grade hardware across diverse locations has emerged as a promising alternative. However, this paradigm faces significant challenges due to hardware heterogeneity and the highly variable latency of geo-distributed communication routes. In this paper, we propose HeteroDL, a lightweight distributed training framework specifically designed for these decentralized heterogeneous settings. HeteroDL first introduces a computation graph partitioner that generates fine-grained model shards to achieve superior load balancing across diverse devices. Then, HeteroDL incorporates DecPipe, a pipeline parallelism runtime engine optimized for geo-distributed servers with heterogeneous networking topologies. In particular, DecPipe enhances existing pipeline schedules through a novel micro-batch communication scheduler, which effectively hides communication latency while maximizing GPU utilization in decentralized scenarios. Meanwhile, it also optimizes the execution plan generation for HeteroDL. Finally, we evaluate HeteroDL using real-world geo-distributed clusters and network conditions. Our results demonstrate that HeteroDL achieves near-linear throughput scaling as the cluster size increases and a speedup of up to $\mathbf {3.3\times }$ compared to existing state-of-the-art distributed deep learning frameworks.
This case-study examines the integration of energy awareness into Kubernetes scheduling on a small, real-world Edge testbed composed of ARM-based devices. A sub-component of the Kubernetes CODECO framework captures node-level energy metrics via the Kubernetes-based Efficient Power Level Exporter and injects this information into the Kubernetes scheduling process through a cross-layer “greenness” cost. The study formalizes compute and network energy metrics, describes their aggregation, and analyzes the applicability and limitations of Kepler's estimation models in ARM architectures. The evaluation compares energy-sensitive placement with the default Kubernetes scheduler under moderate- and high-load conditions. Configuration files, scripts, and datasets are provided to support replication and further experimentation. The results achieved indicate reduced estimated cluster energy consumption under high-load scenarios and mixed results under moderate loads, highlighting the potential of energy-informed orchestration to improve sustainability in Edge-Cloud environments.
Edge-end collaborative architectures equipped with neural processing units (NPUs) provide an effective paradigm for real-time video analytics, where the end device performs low-latency NPU inference and selectively offloads frames to an edge server for accuracy enhancement. However, excessive offloading may violate latency service-level objectives (SLOs), whereas overly conservative offloading may sacrifice potential accuracy gains. Balancing this trade-off under joint accuracy and latency uncertainties is a key challenge. Unlike prior approaches that assume fixed edge latency or rely on heuristics, we pro pose a joint accuracy and latency uncertainty-aware offloading framework to maximize the expected long-term average accuracy gain under latency SLOs. We experimentally verify that frame level inference accuracy exhibits a Markov property under the same deep neural network (DNN). By jointly characterizing accuracy and latency uncertainties, we formulate the online offloading problem as a finite Markov decision process (MDP). We theoretically prove that the MDP is unichain and that its induced Markov chains are aperiodic, ensuring the convergence of relative value iteration (RVI). We then apply RVI to obtain an optimal stationary offloading strategy under the undiscounted average-reward criterion. Extensive experiments on real-world video workloads demonstrate that our approach outperforms state-of-the-art strategies and remains robust under edge service rate mismatches.
The rapid advancement of IoT technology enables edge devices to generate vast amounts of facial data for machine learning. Traditional centralized methods require sensitive data to be uploaded to a semi-trusted cloud, increasing privacy risks. Federated learning addresses this by keeping data local and reducing computational demands, but the uploaded model gradients are vulnerable to attacks that can reconstruct images. Current privacy protection methods struggle to balance communication efficiency with attribute privacy protection, especially for identity-related features. Existing methods often provide insufficient protection against sensitive facial-attribute inference while attempting to preserve the usability of protected data. To address these issues, this work proposes FedDP, a federated learning based reversible face differential privacy protection method. It integrates a pre-trained StyleGAN2 generator and a lightweight latent space encryptor to reversibly encrypt facial identity in the latent space, allowing dynamic data recovery with a key and adjustable privacy strength via the privacy budget. In addition, FedDP reduces the risk of sensitive facial-attribute inference by applying image-specific latent-space encryption and a unified local differential privacy mechanism to the complete encrypted latent representation. FedDP does not identify semantic attributes or assign separate attribute-level privacy budgets during the protection process. Clients apply latent space encryption and local differential privacy noise, while the server adds noise to aggregated updates, ensuring end-to-end protection. Experiments show that FedDP achieves 99% feature reversibility, allows the overall privacy level to be adjusted through the unified privacy budgets, and limits the ResNet-based multi-attribute inference attack success rate to 2.15% when all 10 target attributes are correctly inferred, effectively balancing usability and privacy.
Energy-efficient container migration must reduce IT and cooling energy while controlling thermal risk, resource pressure, and migration overhead. A sequential, search-space-coupled framework with two components is presented. The Fuzzy Dual-Threshold (FDT) trigger uses a fully specified Mamdani fuzzy system and sliding-window confirmation to identify persistent overload, thermal overheat, and underload states. Its output defines the pending-container set searched by the Dueling-DQN-guided Branch-and-Bound (DBD) scheduler. A dueling Q-network ranks feasible branches, whereas explicit checks and a safe lower bound govern the search. The framework further introduces a continuous lifetime-imbalance penalty and a placement-dependent cooling setpoint derived from the heat-recirculation model. The evaluation combines repeated CloudSim experiments using Google Cluster Trace workloads with a 16-node Kubernetes–containerd deployment. In the evaluated simulations, relative to a static threshold, FDT reduces migration count by 39.8%, IT energy by 14.0%, and the operational threshold-exceedance ratio by 91.9%. Relative to no migration, DBD reduces total energy by 13.8% and modeled cooling energy by 25.5%.
