سازوكار تركيبي براي مقياس‌دهي كوبرنتيز بر پايه يادگيري ماشين و پردازش لبه/مه در شبكه‌هاي اينترنت اشيا

A Hybrid Mechanism for Kubernetes Scaling Based on Machine Learning an‎d Edge/Fog Computing in IoT Networks

مهندسي كامپيوتر

With the ever-expan‎ding proliferation of the Internet of Things (IoT) an‎d the generation of massive volumes of data by edge devices, traditional centralized cloud architectures are no longer capable of meeting the requirements of latency- sensitive applications due to challenges such as high latency an‎d core network ban‎dwidth limitations. In response to these constraints, fog an‎d edge computing have emerged as complementary paradigms to cloud computing, enhancing system efficiency by bringing computational resources closer to data sources. Container orchestration using the Kubernetes platform in these distributed environments enables flexible resource management; however, Kubernetesʹ default scaling mechanisms (such as HPA) suffer from inefficiency an‎d resource over-provisioning when confronted with heterogeneous an‎d dynamic IoT workloads, due to their reliance on static thresholds. The present research proposes an intelligent hybrid mechanism for resource management in cloud/fog architectures based on Kubernetes. This solution comprises two key components. First, a Reinforcement Learning (RL)-based horizontal pod autoscaling agent utilizing the Proximal Policy Optimization (PPO) algorithm, which dynamically adjusts the optimal number of pods by analyzing a multidimensional state space including resource consumption an‎d Quality of Service (QoS) metrics. Second, a fog-aware intelligent proxy that optimally distributes traffic between cloud an‎d fog layers by leveraging metrics extracted from the Prometheus monitoring system. Experimental eva‎luation of the system was conducted using the Locust tool an‎d reconstruction of realistic traffic patterns for four heterogeneous applications (CoAP, REST, face recognition, an‎d video streaming). Experimental results demonstrated that the proposed method, compared to Kubernetesʹ default mechanism, improves average response latency (6 to 17 percent) an‎d exhibits greater stability in resource management. Furthermore, the fog-aware proxy successfully reduced end-to-end latency (1.82 to 2.8-fold) in computationally intensive applications compared to the Round-Robin approach. The findings of this thesis indicate that the integration of machine learning with container orchestration at the fog layer can effectively enhance Quality of Service levels an‎d optimize resource consumption in IoT networks.

پژوهشگاه ارتباطات و فناوري اطلاعات

6

162567