Managing Task Execution for Unknown Workloads in Batteryless IoT: A Hardware-Agnostic Evaluation

Batteryless IoT devices, which rely on energy harvesting (e.g., solar) and supercapacitors, face significant operational challenges due to volatile energy availability. Traditional scheduling methods often rely on static voltage thresholds or pre-measured, hardware-specific task profiles. These approaches struggle when workloads are unpredictable or when the device hardware changes, leading to either inefficient energy use or frequent, undesirable power failures (OFF-states).

This paper proposes two novel, hardware-agnostic scheduling strategies that treat applications as a black box, requiring no prior energy information:

• Model-free Reinforcement Learning (RL): An agent that learns to balance task throughput and system survival by observing temporal energy trends, using domain randomization to remain adaptable across different hardware configurations.
• Approximated Prediction (AP): A lightweight, on-the-fly method that estimates future energy states to make immediate execution decisions, aiming for near-oracle performance without the computational overhead of RL.

These methods were benchmarked against an adaptive task rate approach (AsTAR) and optimized static thresholds using a custom simulation framework driven by real-world solar data and dynamic LoRa transmission profiles.

The study demonstrates that there is no universal "best" scheduler; rather, each method offers distinct operational trade-offs:

• AP provides lightweight, high-throughput performance, making it suitable for resource-constrained scenarios.
• RL offers tunable flexibility, allowing developers to prioritize either system survival or task execution based on application needs.
• AsTAR is particularly effective at pacing execution over long periods of energy scarcity.

Crucially, the authors find that while these advanced dynamic strategies are essential for severely constrained systems (e.g., those with very small capacitors), devices with larger energy buffers can often achieve sufficient performance using simpler, less computationally expensive static policies.