Install

pip install powerlog

Python 3.8 or newer, no Python dependencies.

Supported platforms

Power sources are detected at runtime. Install only what matches your hardware; anything unavailable is reported as n/a rather than failing the run.

$ powerlog --list-backends
Detected power sources:
  GPU:
    nvidia       NVIDIA GPU (nvidia-smi / NVML)
  CPU:
    rapl-sysfs   CPU package (RAPL powercap sysfs)

The first source listed for each domain is the one that will be used.

CPU measurement is Linux only. Powerlog measures the node it runs on; it does not aggregate across nodes.

Use

powerlog ./my_program --arg value

================================================================
                    POWERLOG ENERGY SUMMARY
================================================================
Command                 ./my_program --arg value
Total time (s)          12.4180  (wall clock)
CPU                     AMD EPYC 7532 32-Core Processor
GPU                     NVIDIA A100-PCIE-40GB
----------------------------------------------------------------
Domain            Energy (J)   Share (%)   Avg Power (W)
----------------------------------------------------------------
CPU                 962.4013       31.05         77.5005
GPU                2136.7742       68.95        172.0707
----------------------------------------------------------------
TOTAL              3099.1755
EDP (J*s)         38485.5614
  avg power = energy / total time,  EDP = energy x total time
----------------------------------------------------------------
GPU power (W)           min 61.20 / max 249.80
CPU power (W)           min 74.90 / max 79.30
----------------------------------------------------------------
Samples                 124
CPU source              CPU package (RAPL powercap sysfs), 2 package domain(s)
GPU source              NVIDIA GPU (nvidia-smi / NVML), 1 device(s)
================================================================

Powerlog runs unmodified binaries and samples CPU and GPU power in lockstep, so the measurement covers the whole application — I/O, transfers, kernel launches, synchronization — not just the kernels. It writes a summary CSV and a power trace, and exits with the program's own status.

Documentation

powerlog.readthedocs.io

This repository additionally carries two things the wheel does not: example GPU programs in apps/, and datalog-engine-comparison/, a study comparing the energy behaviour of five GPU-accelerated Datalog engines.

Contributing

Bug reports and patches are welcome on the issue tracker. Open an issue first for anything substantial, then send a pull request against main. Release history is in Changelog.md.

License

MIT. See LICENSE.

Citation

@inproceedings{shovon2026heterogeneous,
  title={Heterogeneous Energy Characterization of GPU-Powered Datalog Engines},
  author={Shovon, Ahmedur Rahman and Sun, Yihao and Lan, Zhiling and Perarnau, Swann and Gilray, Thomas and Micinski, Kristopher and Papka, Michael E and Kumar, Sidharth},
  booktitle={2026 IEEE/ACM Workshop on Energy Efficiency with Sustainable Performance: Techniques, Tools, and Best Practices (EESP)},
  year={2026},
  organization={IEEE}
}