Measuring power and energy consumption of programs running on kepler GPUs

Graphics processing units (GPUs) are becoming im-peccable choice for the upcoming exascale computing because of improvements in performance and power efficiency. In this paper, we propose an experimental methodology for evaluating power and energy consumption of programs executing on NVIDIA Kepler GPUs. NVIDIA Tesla K40c GPU is used in the experiments as a test platform. We applied our methodology on two commonly used high-performance computing (HPC) programs, Bitonic Mergesort (a parallel sorting program), and a matrix multiplication program. Using our methodology, power profile of any program executing on NVIDIA Kepler GPUs can be obtained to measure its peak power, average power, energy, and kernel runtime.