Event-driven dynamic power management based on wavelet forecasting theory

The goal of dynamic power management is to reduce power dissipation in system level by putting system components into different states. This paper proposes a wavelet based approach that models the device behavior precisely. In spite of conventional stochastic models, this method eliminates some impractical assumptions which were shortcoming for previous methods. The stationary behavior of devices and the use of a memoryless distribution for device modelling are some of the aforementioned assumptions that have been relaxed in our model. Additionally, wavelet model can capture the local information accurately. Furthermore this algorithm is adaptive. This method has two additional benefits; firstly according to the device application a suitable wavelet basis can be used. Secondly, it has a sparse time-scale representation indicating that only a few coefficients in their wavelet representation have to be estimated. The simulation results show 95% accuracy in desktop hard disk drive (HDD) states prediction and power saving by a factor of 2.