Energy-efficient scheduling in multi-core servers

In this paper we develop techniques for analyzing and optimizing energy management in multi-core servers with speed scaling capabilities. Our framework incorporates the processor’s dynamic power, but it also accounts for other intricate and relevant power features such as the static (leakage) power and switching overhead between speed levels. Using stochastic fluid models to capture traffic burst dynamics, we propose and study different strategies for adapting the multi-core processor speeds based on the observable buffer content, so as to optimize objective functions that balance energy consumption and performance. The strategies can be non-hysteretic (i.e., the processor speed depends on current buffer level relative to the buffer thresholds) or hysteretic (i.e., it matters in which direction the buffer thresholds are crossed). It is shown that, under rather general conditions, strategies which use more threshold levels are more efficient with respect to power consumption; however, most of the efficiency gain is achieved with 1 or 2 thresholds only. In addition, the optimal power consumptions of the different strategies are only very mildly sensitive to perturbations in the input parameters, implying the highly advantageous property that the performance is robust to estimation errors in the system’s input traffic parameters.