Deep Learning Energy Measurement and Optimization

---

Project News ⚡

• [2024/08] Perseus, an optimizer for large model training, was accepted to SOSP'24! Paper | Blog | Optimizer
• [2024/07] Added AMD GPU, CPU, and DRAM energy measurement support, and preliminary JAX support!
• [2024/05] Zeus is now a PyTorch ecosystem project. Read the PyTorch blog post here!
• [2024/02] Zeus was selected as a 2024 Mozilla Technology Fund awardee!
• [2023/07] We used the ZeusMonitor to profile GPU time and energy consumption for the ML.ENERGY leaderboard & Colosseum.

---

Zeus is a library for (1) measuring the energy consumption of Deep Learning workloads and (2) optimizing their energy consumption.

Zeus is part of The ML.ENERGY Initiative.

Repository Organization

zeus/
├── zeus/             # ⚡ Zeus Python package
│  ├── monitor/       #    - Energy and power measurement (programmatic & CLI)
│  ├── optimizer/     #    - Collection of time and energy optimizers
│  ├── device/        #    - Abstraction layer over CPU and GPU devices
│  ├── utils/         #    - Utility functions and classes
│  ├── _legacy/       #    - Legacy code to keep our research papers reproducible
│  ├── metric.py      #    - Prometheus metric export support
│  ├── show_env.py    #    - Installation & device detection verification script
│  └── callback.py    #    - Base class for callbacks during training
│
├── zeusd             # 🌩️ Zeus daemon
│
├── docker/           # 🐳 Dockerfiles and Docker Compose files
│
└── examples/         # 🛠️ Zeus usage examples

Getting Started

Please refer to our Getting Started page. After that, you might look at

• Measuring Energy
• Optimizing Energy

Docker image

We provide a Docker image fully equipped with all dependencies and environments. Refer to our Docker Hub repository and Dockerfile.

Examples

We provide working examples for integrating and running Zeus in the examples/ directory.

Research

Zeus is rooted on multiple research papers. Even more research is ongoing, and Zeus will continue to expand and get better at what it's doing.

1. Zeus (NSDI 23): Paper | Blog | Slides
2. Chase (ICLR Workshop 23): Paper
3. Perseus (SOSP 24): Paper | Blog | Slides

If you find Zeus relevant to your research, please consider citing:

@inproceedings{zeus-nsdi23,
    title     = {Zeus: Understanding and Optimizing {GPU} Energy Consumption of {DNN} Training},
    author    = {Jie You and Jae-Won Chung and Mosharaf Chowdhury},
    booktitle = {USENIX NSDI},
    year      = {2023}
}

Other Resources

1. Energy-Efficient Deep Learning with PyTorch and Zeus (PyTorch conference 2023): Recording | Slides

Contact

Jae-Won Chung (jwnchung@umich.edu)