Obtaining Optimal Thresholds for Processors with Speed-Scaling

Abstract

In this research we consider a processor that can operate at multiple speeds and suggest a strategy for optimal speed-scaling. While higher speeds improve latency, they also draw a lot of power. Thus we adopt a threshold-based policy that uses higher speeds under higher workload conditions, and vice versa. However, it is unclear how to select "optimal" thresholds. For that we use a stochastic fluid-flow model with varying processing speeds based on fluid level. First, given a set of thresholds, we develop an approach based on spectral expansion by modeling the evolution of the fluid queue as a semi-Markov process (SMP) and analyzing its performance. While there are techniques based on matrix-analytic methods and forward-backward decomposition, we show that they are not nearly as fast as the spectral-expansion SMP-based approach. Using the performance measures obtained from the SMP model, we suggest an algorithm for selecting the thresholds so that power consumption is minimized, while satisfying a quality-of-service constraint. We illustrate our results using a numerical example.