Energy-Efficient Data Processing at Sweet Spot Frequencies

The processing of Big Data often includes sorting as a basic operator. Indeed, it has been shown that many software applications spend up to 25% of their time sorting data. Moreover, for compute-bound applications, the most energy-efficient executions have shown to use a CPU speed lower than the maximum speed: the CPU sweet spot frequency. In this paper, we use these findings to run Big Data intensive applications in a more energy-efficient way. We give empirical evidence that data-intensive analytic tasks are more energy-efficient when CPU(s) operate(s) at sweet spots frequencies. Our approach uses a novel high-precision, fine-grained energy measurement infrastructure to investigate the energy (joules) consumed by different sorting algorithms. Our experiments show that algorithms can have different sweet spot frequencies for the same computational task. To leverage these findings, we describe how a new kind of self-adaptive software applications can be engineered to increase their energy-efficiency.