Energy and Power Characterization of Parallel Programs Running on Intel Xeon Phi

Intel's Xeon Phi coprocessor has successfully proved its capability by being used in Tianhe-2 and Stampede, two of the top ten most powerful supercomputers today. It is almost certain that the popularity of Xeon Phi in heterogeneous computing will grow significantly, which requires comprehensive studies on different aspects of this newly arrived many-core chip. Despite a number of previous studies on the performance of Xeon Phi, the power and energy behavior of the coprocessor has not been fully studied. In this paper, we present the performance, power and energy results of multiple parallel programs with contrasting workloads running on Intel Xeon Phi. Several interesting findings are derived from these results: 1) the Xeon Phi thread is power-hungry even when idle and altering the number of executing threads will largely affect the power consumption, 2) performance improvement and energy savings are highly related, 3) running code in native mode yields better performance and consumes less energy, and 4) co-running programs with complementary workloads has potential to conserve energy with negligible performance influence. In addition, we discuss an incorrect way of measuring power of Xeon Phi using the on-chip power sensors and present our solutions.