Dynamic power management and voltage scaling

DPM/DVS techniques are concerned with changing system performance levels dynamically to reduce system energy consumption, and have been extensively investigated [15,2]. One of the key problems in DPM/DVS is resource usage prediction. Various predictive and stochastic techniques have been proposed [2]. However, as demonstrated in [21,9], most existing resource usage prediction methods may work well for computation-intensive tasks, but not for interactive tasks. Although many works [21,9,7,16] have used interactive applications as their benchmarks, they have treated these applications in the same fashion as computation-intensive applications and have not exploited any features of interactive applications to make resource usage predictions better. Some of them [7,16] have proposed to utilize the human-perceptual limit to slow down the system so that the user interface can respond within the limit. Since many user interface operations still require relatively high performance, the proposed technique missed the opportunity to further slow down the processor during its wait time.

In [25], an exponential cumulative distribution was used to model user requests for power management. This is stochastic and too high-level to accurately model actual user delays.