High-Resolution Power Profiling of GPU Functions Using Low-Resolution Measurement

In order to be able to minimise the energy consumption of an application program, information about the specific energy consumption is required. Modern Nvidia graphics processing units (GPUs) measure their current power consumption and the driver makes the value available to the application every 20 ms. However, for evaluating the energy consumption of GPU kernel functions, such a sampling interval might not be sufficient since the kernels may have a shorter execution time. This article proposes a method for generating high-resolution power profiles, which is the power consumption of a specific function depending on the progress of its execution. The method uses low-resolution measuring instruments offered by GPUs. Power measurements obtained during several executions of the function are combined into a single power profile. The resulting power profile contains power values in intervals which are much shorter than the sampling interval of the hardware driver so that even short-term power changes can be considered, e.g. for calculating the energy consumption of a single function. The article also shows how to extend the approach to an online generation of power profiles. Furthermore, an overview on the power profiles of some important functions, such as BLAS routines, is given.