powerlog 0.0.2

GPU power log

Powerlog

Powerlog is a lightweight command-line tool and Python package to profile Nvidia GPU power consumption during the execution of a command-line program. It uses nvidia-smi to sample power draw at regular intervals and reports total energy usage, average power, and min/max readings.

Features

• Measures real-time GPU power draw using nvidia-smi

• Computes:

  • Total runtime
  • Total energy consumed (in Joules)
  • Average (sampled and timed), min, and max power (Watts)

• Outputs both summary and raw samples as CSV

• Simple CLI interface

• Reports Avg Power Sampled (mean of all readings) and Avg Power Timed (energy divided by total time)

Installation

Requires Python 3.6+ and NVIDIA's nvidia-smi available in your system PATH.

Project page at the Python Package Index (PyPI): https://pypi.org/project/powerlog/

Install with pip:

pip install powerlog

Usage

powerlog --output power_report.csv --gpu 1 ./my_gpu_program arg1 arg2

CLI Options

Argument	Description
--output	Base name for the output CSV files
--gpu	Number of GPUs to monitor (default: 1)
cmd	Command and arguments to run and profile

Output

If --output power.csv is specified:

• power.csv: Summary of runtime, energy, and power stats
• power_samples.csv: Raw timestamped power draw samples

Example

powerlog --output matrix_power.csv --gpu 1 nvidia-smi

This will generate matrix_power.csv and matrix_power_samples.csv

Demo content of matrix_power.csv:

Total Time (s),Total Energy (J),Avg Power Sampled (W),Avg Power Timed (W),Min Power Sampled (W),Max Power Sampled (W)
0.1001,1.2856,12.8400,12.8400,12.84,12.84

Demo content of matrix_power_samples.csv:

Timestamp (ns),Power Draw (W)
1752198440220994393,12.84

Dependencies

• Python standard library (subprocess, argparse, time, csv)
• NVIDIA GPU with drivers and nvidia-smi tool

How power and energy are calculated?

Power log collection

Powerlog uses nvidia-smi to measure GPU power draw at regular intervals (default: every 0.1 seconds). The Python wrapper script automatically runs this measurement loop from the start to the end of your program.

Total energy consumption computation

Total energy E (in Joules) is computed as $$ E = \sum_{i=1}^N P_i \cdot \Delta t_i $$

where:

• N: total number of sampling intervals
• P_i: GPU power draw (Watts) at interval i
• Δt_i: elapsed time (seconds) between sample i and sample i-1

Development

Local Testing

To test Powerlog locally during development (before releasing to PyPI), you can install your package in "editable" mode:

python3 -m venv .venv
source .venv/bin/activate
pip install -e .

Publishing to PyPI

When the package is ready to publish (or update) Powerlog on PyPI, use the following commands:

python -m pip install --upgrade build
python3 -m build
python3 -m pip install --upgrade twine
twine upload dist/*

This will build the distribution files (.tar.gz and .whl) and upload them to PyPI.

It requires API token for authentication.

Changelog

See the Changelog.md file in this repository for version history and release notes.

License

MIT License

Acknowledgments

• Developed as part of GPU power-efficiency profiling experiments in Datalog-based engines.
• Inspired by the EUMaster4HPC