Optimizing energy consumption under flow and stretch constraints

In embedded systems and data-center systems (systems, for short), it is widely accepted that energy consumption has become the bottleneck of system´s performance improvement and it is one of the most significant factors to optimize. Unfortunately, an effective energy-aware strategy usually has an adverse impact on a job´s flow time or stretch - two important user-perspective system performance metrics. In some cases, the more energy is saved by an energy-aware policy, the more flow time and the larger stretch occur to jobs. In this paper, we investigate the impact on job processing delay introduced by power-down energy-saving mechanisms. Specifically, we study bicriteria algorithms that minimize maximum flow time or largest stretch under a fixed energy budget and minimize total energy consumption under an upper bound of flow time or stretch. We develop optimal offline algorithms to quantitatively balance the system-perspective performance metric (energy consumption) and the user-perspective performance metric (flow time and stretch). We also develop two simple min-energy online algorithms against weakened adversaries. We prove that (1.) with appropriate extra flow time, an online algorithm can beat any non-idling algorithm, in terms of energy consumption; (2.) a deterministic online algorithm which has a bounded times of optimal stretch, is optimal in terms of competitive ratio with respect to energy consumption.