GEOPM - Global Extensible Open Power Manager

Home Page | GEOPM Access Service | GEOPM Runtime Service | Reference Manual | Slack Workspace

Fine-grained low-latency batch access to power metrics and control knobs on Linux

Key Features

The Global Extensible Open Power Manager (GEOPM) provides a framework to explore power and energy optimizations on platforms with heterogeneous mixes of computing hardware.

Users can monitor their system's energy and power consumption, and safely optimize system hardware settings to achieve energy efficiency and/or performance objectives.

With GEOPM, a system administrator can:

Grant per-user or per-group access to individual metrics and controls, even when their underlying interfaces do not offer fine-grained access control
Ensure that user-driven changes to hardware settings are reverted when the user's process session terminates
Develop your own platform-specific monitor and control interfaces through the extensible plugin architecture

With GEOPM, an end user can:

Interact with hardware settings and sensors (e.g., set a CPU power limit or read a GPU's current power consumption) using a platform-agnostic interface
Generate summarized reports of power and energy behavior during execution of an application
Automatically detect MPI and OpenMP phases in an application, generating per-phase summaries within application reports
Optimize applications to improve energy efficiency or reduce the effects of work imbalance, system jitter, and manufacturing variation through built-in control algorithms
Develop your own runtime control algorithms through the extensible plugin architecture
Gather large groups of signal-reads or control-writes into batch operations, often reducing total latency to complete the operations.

GEOPM software is separated into two major components, the GEOPM Access Service and the GEOPM Runtime Service.

GEOPM Access Service: A privileged process that provides user interfaces to hardware signals and controls and admin interfaces to manage user access. C and C++ bindings to this interface are provided through [libgeopmd](libgeopmd). Python bindings are provided through the [geopmdpy](geopmdpy) package.

GEOPM Runtime Service: An unprivileged process that provides a framework to control platform settings based on feedback from signals and monitored application state. This process delegates platform interactions to the GEOPM Access Service. C and C++ bindings are provided through [libgeopm](libgeopm). Python bindings are provided through the [geopmpy](geopmpy) package.

Usage

Some simple use cases are illustrated below. See the Getting Started Guide for additional use cases in Bash, C, C++, and Python.

Read the current power consumption of all CPUs in the platform. The command will print the total power consumption (in Watts) summed across all CPUs on the board.

geopmread CPU_POWER board 0

https://github.com/geopm/geopm.github.io/assets/378319/795b297e-7a45-47a8-bf67-3fdf8be9448e

Apply a 3.0 GHz maximum-allowed CPU core frequency to each CPU on the board. This setting will be automatically reverted when the user's session ends (e.g., when exiting the current shell).

geopmwrite CPU_FREQUENCY_MAX_CONTROL board 0 3.0e9

https://github.com/geopm/geopm.github.io/assets/378319/0a4f2cf8-cebf-4556-b710-6c664568790d

Generate a CSV (comma-separated variable) trace of CPU core frequency versus time, sampling once every second for a total of 10 seconds:

echo -e 'TIME board 0\nCPU_FREQUENCY_STATUS package 0' | geopmsession -p 1.0 -t 10.0

https://github.com/geopm/geopm.github.io/assets/378319/382fbe44-5ab4-4c43-9173-982473ebccb8

Other use cases in the Getting Started Guide include:

Setting admin policies for user access to signals and controls
Exploring the platform's hardware topology
Reading other types of signals and writing other types of controls at various scopes in the topology
Repeatedly reading multiple signals in batches
Using the GEOPM Runtime Service alongside applications

How to Install GEOPM

We provide installable packages for Ubuntu, CentOS, openSUSE, and RHEL. Our build systems can also be used to install GEOPM from source.

More details are available in our Installation documentation page.

On Ubuntu

sudo add-apt-repository ppa:geopm/release
sudo apt update
sudo apt install geopm-service libgeopmd-dev libgeopmd2 python3-geopmdpy

On CentOS, openSUSE, and RHEL

Follow the installation wizards described in our installation guide to install our latest release through dnf or zypper.

From Source

Follow the documentation in our developer build guide for details about how to configure the build process. Note that some dependency packages required for the C++ builds may be missing from your system. Refer to the requirements guide for details on the packages required for build on your operating system. In the bash script below we show a simple way to build and install all of the GEOPM packages from the source repository assuming that all build dependencies are installed system wide.

# Choose install location
INSTALL_PREFIX=$HOME/build/geopm    # User install
# INSTALL_PREFIX=/usr/local         # Root install
pip install -r requirements.txt
cd libgeopmd
./autogen.sh
./configure --prefix=$INSTALL_PREFIX
make -j                             # Build libgeopmd
make install                        # Install to the --prefix location
cd ../libgeopm
./autogen.sh
./configure --prefix=$INSTALL_PREFIX
make -j                             # Build libgeopm
make install                        # Install to the --prefix location
cd ..
pip install -r geopmdpy/requirements.txt
pip install ./geopmdpy
pip install -r geopmpy/requirements.txt
pip install ./geopmpy
make -C docs man
make -C docs prefix=$INSTALL_PREFIX install_man

When building from source and configured with the --prefix option, the libraries, and binaries will not install into the standard system paths. At this point, you must modify your environment to specify the installed location.

export LD_LIBRARY_PATH=$INSTALL_PREFIX/lib:$LD_LIBRARY_PATH

export PATH=$INSTALL_PREFIX/bin:$PATH

Major GEOPM Versions

At a high level, major GEOPM releases are summarized as follows:

Version 3.0: The GEOPM runtime now also works with non-MPI applications.
Version 2.0: GEOPM is split into two components: a service that manages platform I/O, and a runtime that writes platform power management controls based on feedback from MPI application state and platform state. GEOPM now has interfaces for Intel and NVIDIA GPUs and the isst_interface driver.
Version 1.0: GEOPM is production-ready. It provides an abstraction layer for interaction with a platform's power metrics and control knobs, and offers the ability to interact with control knobs based on information from instrumented MPI applications.

Please refer to the ChangeLog.md for a more detailed history of changes in each release.

Repository Directories

.github contains definitions of this repository's GitHub actions
[docs](docs) contains web and man-page documentation for GEOPM
[geopmdpy](geopmdpy) provides Python bindings for libgeopmd
[geopmpy](geopmpy) provides Python bindings for libgeopm
[integration](integration) contains integration test automation for GEOPM
[libgeopm](libgeopm) provides the C/C++ implementation of the GEOPM Runtime Service
[libgeopmd](libgeopmd) provides the C/C++ implementation of the GEOPM Access Service
[release](release) packaging files for latest release by distro

Guide for Contributors

We appreciate all feedback on our project. See our contributing guide for guidelines on how to report bugs, request new features, or contribute new code.

Refer to the GEOPM Developer Guide for information about how to interact with our build and test tools.

License

The GEOPM source code is distributed under the 3-clause BSD license.

SEE [LICENSE-BSD-3-Clause](LICENSE-BSD-3-Clause) FILE FOR LICENSE INFORMATION.

Last Update

2024 April 10

Christopher Cantalupo chris.nosp@m.toph.nosp@m.er.m..nosp@m.cant.nosp@m.alupo.nosp@m.@int.nosp@m.el.co.nosp@m.m
Brad Geltz brad..nosp@m.gelt.nosp@m.z@int.nosp@m.el.c.nosp@m.om

Acknowledgments

Development of the GEOPM software package has been partially funded through contract B609815 with Argonne National Laboratory.