Abstract: Traditional dynamic power management policies of embedded system only focus on equipment workload situation from the perspective of device, and ignore the application features of the workload. An improved method called SMBSP (System message based stochastic policy) for non-stationary multi-task is proposed on the basis of analyzing workload according to system state. First, this paper introduces the dynamic power management strategy algorithm, and the reason why it needs to be improved is expounded. Then, device-requester utilization lookup table was constructed based on system message, and updated distribution according to the actual time interval. Finally, a Semi-Markov stochastic model was built, the discount criteria optimization method based on linear programming strategy was given, the moment of decision and conversion state are pointed out. Experimental results show that, while considering the conditions of the performance constraints, the algorithm has good stability; delay rate is less than 0.10; competitive rates reached 0.57. The algorithm reduces power consumption more efficiently than other policies, and is beneficial to the application in embedded systems.

Key words: Dynamic power management, Markov decision processes, Policy optimization, Embedded system