Dynamic AI Task Partitioning Framework for Heterogeneous Edge-Cloud Continuum

Designers face challenges in determining where computation should take place when deploying AI on IoT devices with limited resources. Traditional static partitioning techniques fail to account for runtime variations and often lack validation on actual hardware. A novel framework addresses this by dynamically distributing neural network layers across the edge-cloud spectrum, utilizing model profiling at startup and assessing network conditions among edge devices, fog nodes, and cloud servers. It regularly reassesses these partitions to adapt to fluctuations in bandwidth. To evaluate the framework, a physical testbed was created using a Raspberry Pi, a laptop, and a high-performance desktop PC, testing it with VGG16, AlexNet, and MobileNetV2, unlike previous studies that depended solely on simulations.