Predictive Worker Resource Characterization at the Extreme Edge

Abstract

Extreme Edge Computing (EEC) promises to enhance the Quality of Service of data-intensive and delay-critical IoT applications. Deploying EEC is challenging due to reliance on user-owned Extreme Edge Devices (EEDs) with heterogeneous computational and communication capabilities and dynamic user behaviors causing resource volatility. This volatility complicates the accurate estimation of EED computational capabilities, a critical step for efcient task allocation in EEC systems. Previous research characterizes EED capabilities simplistically by allocating benchmark tasks without accounting for resource volatility, using insufcient performance indicators, and reactively adjusting EEC task allocation only after volatility impacts system operations. In this thesis, we propose a comprehensive framework for predictively characterizing the computational resources of EEDs according to their dynamic resource usage to address this challenge. The framework consists of four key components that: 1) emulate dynamic user-access behaviors and generate resource usage data, 2) forecast resource usage states efciently over multi-step horizons, 3) allocate benchmark tasks based on predicted states to characterize EED capabilities, and 4) dynamically allocate EEC tasks based on adaptive worker characterization. Extensive experiments on a realistic EED testbed with an interactive monitoring dashboard demonstrate our framework’s efcacy in improving EEC task allocation performance, reducing execution latency, and increasing throughput without overloading device resources, compared to prominent existing schemes.