Statistical modeling of power/energy of scientific kernels on a multi-GPU system

Energy efficiency of GPUs has facilitated the usage of GPUs in many complex scientific applications. Nodes with multi-GPUs along with multi-core CPUs are quite common in today's HPC landscape. This gives the flexibility to utilize CPUs or accelerators or even both according to the workload characteristics. It is not possible to measure power and energy accurately in all the cases, an alternate approach is to estimate power and energy using statistical methods. Apart from saving time and money, reasonable prediction of power/energy would lead to power saving optimizations for certain applications, without compromising performance. In this paper we employ parametric and non-parametric regression analysis to model power and energy consumption of some of the common high performance kernels (DGEMM, FFT, PRNG and FD stencils) on a multi-GPU platform. Our experiments show that using a minimal set of hardware counters and performance attributes, the average error between the measured and the predicted values of power and energy is only ~ 4%.