REALITY: RL-PowEred AnomaLy Detection with Imprecise Computing in MulTi-Core SYstems

The Approximate Computing (AC) model is used by contemporary real-time systems to balance accuracy and system resource constraints better while completing a set of time-sensitive tasks within a deadline. These AC assignments include both necessary and optional components. To achieve satisfactory results, the optional components can be carried out entirely or partially depending on the available resources. Considering a real-time multiprocessor system, for a set of interconnected AC tasks with deadlines and an energy budget, creates a workable schedule to execute the necessary and optional components of tasks to meet an anticipated quality of service (QoS) level. However, during execution, if a malware attack or bug affects one or more processor cores, the system may stop working altogether after deployment, causing unanticipated power outages or processing delays that prevent the system from finishing its task by the deadline. The goal of our proposed methodology REALITY is to investigate the possibility of intelligently rescheduling a set of dependant AC tasks upon detection of any anomalous situation, running in multi-core systems under system-wide constraints to maintain acceptable levels of QoS. During execution, REALITY monitors the operational behaviour of the processing cores by collecting Hardware Performance Counters (HPCs) and identifying any irregularities through an ML-powered anomaly detection mechanism. Upon detection, it applies remedial actions by intelligently rescheduling the tasks leveraging the Proximal Policy Optimisation (PPO)-based Reinforcement Learning (RL) algorithm while maintaining 70% QoS.