NVpower

NVpower is a tool to measure the energy consumption of NVIDIA GPUs.

NVpower uses NVML SDK (Nvidia Management Library) to perform a sequence of power measurements sampling on the specified NVIDIA GPU devices during the execution of a given process.

The measured energy consumption can be calculated by the following formula:

Energy_measure = (Power_sampled - Power_static) * Duration 

Installation

Make sure that you use the python3 environment. In some linux machine, python usually means python2 but not python3, and also pip.

In order to avoid unnecessary errors, it is better to install the nvidia-ml-py3 package.

requirement:

• python3
• nvidia-ml-py3

Find the python release whleel https://github.com/wildkid1024/NVpower/releases/tag/0.0.1, then

> pip install NVpower-0.0.1-py3-none-any.whl

Usage

CLI

> NVpower --job /path/to/job --gpu_id device_id --interval 1

example:

# run "./GPU_bin argv1 argv2" program with device 0 
> CUDA_VISIBLE_DEVICES=0 NVpower -j "./atm 24 1024" -d 0

Python package

CPP lib

ToDo

• Add sampling interval
• Add output file
• Add threshold power
• Add all deveice power moniter

Related Works

1. Collange, Sylvain, David Defour, and Arnaud Tisserand. "Power consumption of GPUs from a software perspective." International Conference on Computational Science. Springer, Berlin, Heidelberg, 2009.
2. Mittal, Sparsh, and Jeffrey S. Vetter. "A survey of methods for analyzing and improving GPU energy efficiency." ACM Computing Surveys (CSUR) 47.2 (2014): 1-23.

Update logs

• Version 0.0.1 2020-12-03
  • release the new built package