Adaptive, Lightweight, Unified Metrics.

ALUMET is a modular tool that measures energy consumption and performance metrics. It offers a new standard framework for all your measurements, with a very low overhead. Learn more on the website.

Why should I use it?

• I want to estimate the energy consumption of the software I run.
• I want to measure the energy consumption of my CPU or GPU, accurately¹.
• I need to export my measurements to local files or to a database.
• I would like to choose the acquisition frequency and to be able to use frequencies above 1000 Hz.
• I don't want the measurement tool to eat up my CPU and consume too much power.
• I don't want to setup a different tool for each hardware component and software environment I have (laptops, Edge devices with GPUs, bare-metal HPC servers, K8S clusters, ...).

If you answer yes to any of these questions, Alumet is for you!

We also have extra features (see the documentation).

How to use

Please read the Alumet user book to learn how to install and use the Alumet "agent" (the program that performs the measurements).

If you have a question, feel free to ask on the Discussions page.

Extending Alumet

The alumet crate provides a library with a plugin system. With plugins, you can extend Alumet in the following ways:

• read new sources of measurements
• apply arbitrary transformations to the data (such as energy attribution models)
• export the data to new outputs
• perform actions on startup and shutdown

Please read the Alumet developer book to learn how to make plugins.

Contributing

Alumet is a joint project between the LIG (computer science laboratory of Grenoble) and Eviden (Atos HPC R&D). It is also open to external volunteers like you!

Please go to the contributing guide to get started.

License

Copyright 2024 Guillaume Raffin, BULL SAS, CNRS, INRIA, Grenoble INP-UGA. Licensed under the EUPL-1.2 or later.

You can find more information about the EUPL here. The EUPL is compatible with many other open source licenses and shares some principles with the well-known LGPL.

Footnotes

1. See the following research paper for a detailed analysis of some common errors in RAPL-based measurement tools: Guillaume Raffin, Denis Trystram. Dissecting the software-based measurement of CPU energy consumption: a comparative analysis. 2024. ⟨hal-04420527v2⟩. ↩