A New Resource Configuring Scheme for Variable Workload in IoT Systems

With the growing data size and the increased amount of computational load in machine learning workloads, configuring the resources of IoT (Internet-of-Things) systems in an energy-saving way is becoming important. For example, the workload of an unmanned aerial vehicle (UAV) depends on weather and obstacle conditions, and reconfiguring the processor voltage and memory state for these conditions is critical to battery life. However, resource configuring in traditional IoT systems is fixed in advance to meet real-time constraints. In this paper, we suggest a new resource configuring scheme for real-time jobs that can handle workload variations in emerging IoT systems. The goal of our scheme is to minimize the energy consumed by proper resource configurations in response to workload fluctuations and eliminate the tardiness of real-time jobs. To handle workload fluctuations, we categorize real-time jobs into primary jobs and additional jobs, and pre-plan the resource configuring for various workload situations. Based on this, we start the IoT system with the configuration for the primary jobs, and when additional jobs are activated, we update the resource configurations promptly for the new situation. In particular, our resource configuring scheme optimizes the supplied voltage of the processor and memory configuration for all real-time job combinations, and reflects to the system instantly as additional jobs are activated. Based on simulation experiments under various workload conditions, we show that the suggested scheme saves the battery power by 32.1% without the tardiness of real-time jobs.