Proteus: Towards Intent-driven Automated Resource Management Framework for Edge Sensor Nodes

Abstract

Edge computing provides critical resources for various latency-sensitive applications, including, safety-critical monitoring systems that process large volumes of data from sensors and IoT devices, employing machine learning pipelines for effective and reliable analysis. Such applications are deployed on specially designed Edge Sensor Nodes (ESNs) that possess various sensors and limited computing power and support multiple data analysis tasks. ESNs encounter unique operational challenges, including intermittent power supplies, limited connectivity, and dynamic application and resource requirements, which complicate runtime management. Conventional resource management platforms like Kubernetes and KubeEdge are unsuitable for the dynamic needs of ESNs due to their reliance on centralized control and expected stable conditions. To bridge this gap, our paper introduces a data-driven resource management framework tailored for the autonomous adaptation of ESNs to diverse application and infrastructure requirements. We propose an intent-based mechanism that aligns application requirements, such as end-to-end latency, with infrastructure goals like utilization levels. This mechanism translates high-level intents into actionable low-level configurations, balancing the competing demands of various applications and resources, thereby guiding us toward a more robust and efficient application management system. We have implemented a prototype system, evaluated it on an experimental testbed, and demonstrated that our approach performs better than static-only optimization approaches with minimal impact on application performance.