DReaM: Deep Reinforcement Learning for Joint Reliability and Power Management in Multicore IoT Devices

In multicore Internet of Things (IoT) devices, high reliability, low power/energy design, and real-time computing are three important design requirements. The multiple cores in these systems enable us to exploit different fault-tolerance techniques for improving system reliability. However, these techniques impose power/energy and timing overheads on the system. This paper proposes DReaM, a power- and reliability-aware approach that utilizes Deep Reinforcement Learning (DRL) to minimize power consumption in multicore systems while maintaining the desired reliability level and meeting timing constraints. DReaM controls the power consumption by adjusting the voltage and frequency levels for each task and choosing the most suitable fault-tolerant techniques between Primary/Backup (P/B), Triple Modular Redundancy (TMR), and N-Modular Redundancy (NMR). Evaluations using an ARM-based processor and applications from the MiBench benchmark suite were conducted under different fault injection scenarios. The experimental results show that DReaM reduces power consumption on average by 19.86% (up to 64.1%), with improved Quality of Service (QoS) by an average of 11.02% (up to 77%) while maintaining the required system reliability.