PROPHET: Speculative load prediction algorithm for dynamic performance scaling

Dynamic Performance Scaling is highly efficient in reducing power consumption of computers. However, it causes a tradeoff between energy efficiency and performance if based on processor utilization history. In this paper we present an algorithm called PROPHET aimed at boosting load prediction used within dynamic performance scaling systems. PROPHET is a binary search-like speculative approach for determining current processing load and setting the performance level accordingly. PROPHET guarantees that performance degradation does not exceed a worst-case boundary value. A performance hit metric is defined and utilized for determining the maximum performance hit of the algorithm. PROPHET is applied to and tested in embedded multimedia device, as well as in a simulated environment. Test results on real device show 27% to 32% energy savings in a user-interactive use case with feasible performance. Comparison to existing energy-equal performance scaling algorithms shows up to 35% smaller performance hit on real device and even more in the simulated execution environment.