Memory Efficient Stream Processing for IoT Devices

Stream processors working on mass-produced IoT devices will get more important when more developers start to make applications on such devices with small memory spaces. In this study, we propose a novel task scheduling algorithm in stream processing to achieve a small and controlled memory footprint. To adhere to memory upper limit given from configurations keeping scheduling fairness, we developed two task schedulers: one is Flow-Efficient Scheduler for fairness and sufficiently low memory utilization, and the other is Memory-Reducing Scheduler for for reducing internal memory at the cost of scheduling fairness. The former is the default scheduler, and the latter is used to recover from higher memory pressure. We devised Memory State Machine to track the memory consumption and switch schedulers. We conducted experiments that imitated an in-vehicle sensor data processing. The experimental results show that our proposed methods help resilient recovery from high memory pressure caused by burst stream data. This study also aims to introduce SpringQL, our stream processor implementation for middle to high-end IoT devices. This OSS stream processor is used in the evaluation to implement the proposed schedulers.