Energy Efficient Multi Core Task Scheduling for Real Time Edge AI Systems: A Latency Aware Approach

Abstract

Edge intelligence systems face dual constraints of real time computing and multicore energy efficiency. Existing research mostly relies on isolated voltage and frequency regulation mechanisms, failing to critically analyze the deep correlation between upper layer container cold starts and lower layer hardware sleep and wake up latencies, as well as the limitations of traditional methods. In view of this, this paper proposes a software and hardware collaborative scheduling framework that integrates container prefetching with C state management. During the dynamic evolution of the study, the algorithm assigns low power cores to asynchronously handle network input and output, while proactively calculating the physical wake up timing for performance cores, attempting to simultaneously mask the dual overhead of software and hardware along the timeline. Interpretation from a multi angle data perspective indicates that this mechanism significantly extends the sleep residency period of cores to a certain extent. However, tail effects triggered by extreme network fluctuations and the incidental energy consumption of the scheduler itself may introduce potential measurement bias, and the robustness of the system under complex topologies requires further study. This urges us to further consider the research depth of cross stack system optimization, providing a prospective path for future paradigm shifts toward highly dynamic scenarios such as the Internet of Vehicles.