Dynamic Voltage Scaling with Feedback Scheduling for Real-time Embedded Systems

Scheduling;Embedded;Realt-time;DVS

Zhu, Yifan ; Frank Mueller, Committee Chair,Zhu, Yifan ; Frank Mueller ; Committee Chair

University:North Carolina State University

关键词: Scheduling;    Embedded;    Realt-time;    DVS;   

Others  :  https://repository.lib.ncsu.edu/bitstream/handle/1840.16/3852/etd.pdf?sequence=1&isAllowed=y

美国|英语

来源: null

【 摘 要 】

Dynamic voltage scaling (DVS) is a promising method to reduce the power consumption of CMOS-based embedded processors. However, pure DVS techniques do not perform well for dynamic systems where the execution times of different jobs vary significantly.A novel DVS scheme with feedback control mechanisms for hard real-time systems is proposed in this work. It produces energy-efficient schedules for both static and dynamic workloads. Task-splitting, slack-passing and preemption-handling schemes are proposed to aggressively reduce the speed of each task. Different feedback control structures are integratedinto the DVS algorithm to make it adaptable to workload variations.This scheme relies strictly on operating system support. It is evaluated in simulation as well as on an embedded platform.For given task sets, simulation experiments demonstrate the benefits of this scheme with savings of up to 29\% in energy over previous work. This scheme exhibits up to 24\% additional energy savings over other DVS algorithms on the embedded platform. The feedback-based DVS scheme is further extended to be leakage aware, which considers not only dynamic but also static power consumption caused by leakage current in circuits. A combined DVS, delay and sleeping scheme is proposed for architectures where static power exceeds dynamic power in some cases. DVS is used when dynamic power dominates the total power consumption, while a sleep mode is entered when static power becomes dominant. The extended algorithm, DVSleak, shows 30\% additional energy savings on average over a pure DVS algorithm in the simulation experiment.