Adaptive DNN Partitioning and Offloading in Heterogeneous Edge-Cloud Continuum
Researchers have proposed a new framework for adaptive Deep Neural Network (DNN) partitioning and offloading within heterogeneous edge-cloud environments. Addressing the limitations of existing static methods that ignore runtime dynamics and rely on simulations, this study introduces a system that dynamically splits neural network layers based on real-time conditions. The framework profiles models at startup, measures network link conditions between nodes, and periodically re-evaluates partitions to adapt to environmental changes. Validation was conducted using a physical testbed comprising a Raspberry Pi edge device, a laptop fog node, and a high-performance desktop PC acting as the cloud. The team evaluated the framework using three widely adopted convolutional neural networks: VGG16, AlexNet, and MobileNetV2. Results demonstrated significant performance improvements over static baselines, achieving energy reductions of 27.09–35.82% and end-to-end latency reductions of 6.34–22.92%. These findings confirm the superiority of adaptive partitioning strategies for resource-constrained IoT devices, offering a practical solution for optimizing AI workloads across distributed computing infrastructures.
Wire timeline
Adaptive DNN Partitioning and Offloading in Heterogeneous Edge-Cloud Continuum
Researchers have proposed a new framework for adaptive Deep Neural Network (DNN) partitioning and offloading within heterogeneous edge-cloud environments. Addressing the limitations of existing static methods that ignore runtime dynamics and rely on simulations, this study introduces a system that dynamically splits neural network layers based on real-time conditions. The framework profiles models at startup, measures network link conditions between nodes, and periodically re-evaluates partitions to adapt to environmental changes. Validation was conducted using a physical testbed comprising a Raspberry Pi edge device, a laptop fog node, and a high-performance desktop PC acting as the cloud. The team evaluated the framework using three widely adopted convolutional neural networks: VGG16, AlexNet, and MobileNetV2. Results demonstrated significant performance improvements over static baselines, achieving energy reductions of 27.09–35.82% and end-to-end latency reductions of 6.34–22.92%. These findings confirm the superiority of adaptive partitioning strategies for resource-constrained IoT devices, offering a practical solution for optimizing AI workloads across distributed computing infrastructures.
cs.AI updates on arXiv.org