GPU Power Modeling of HPC Applications for the Simulation of Heterogeneous Clouds

Hardware accelerators have been widely used in the scientific community, as the gain in the performance of HPC applications is significant. Hardware accelerators have been used in cloud computing as well, though existing cloud simulation frameworks do not support modeling and simulation of such hardware. Models for the estimation of the power consumption of accelerators have been proposed by many researchers, but they require large number of inputs and computations, making them unsuitable for hyper scale simulations. In previous work, a generic model for the estimation of the power consumption of accelerators has been proposed, that can be combined with generic CPU power models suitable for integration in hyper scale simulation environments. This paper extends this work by providing models for the energy consumption of GPUs and CPU-GPU pairs, that are experimentally validated with the use of different GPU hardware models and GPU intensive applications. The relative error between the actual and the estimated energy consumption is low, thus the proposed models provide accurate estimations and can be efficiently integrated into cloud simulation frameworks.