Paper at 2023 ACM/IEEE The International Conference for High Performance Computing, Networking, Storage, and Analysis (SC'23)
@inproceedings{li2023toward,
title={Toward Sustainable HPC: Carbon Footprint Estimation and Environmental Implications of HPC Systems},
author={Li, Baolin and Basu Roy, Rohan and Wang, Daniel and Samsi, Siddharth and Gadepally, Vijay and Tiwari, Devesh},
booktitle={Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis},
pages={1--15},
year={2023}
}
We set up three sets of benchmarks to evaluate the carbon and performance on three different application scenarios: CV, NLP, and HPC. These benchmarks are made available in these repositories:
The data used to generate the performance and operation carbon results are collected from running these benchmarks.
To visualize the results presented in the paper, on a Linux CPU machine, install the necessary Python packages
pip install -r requirements.txt
cd combined_carbon
The visualization of hardware upgrade impact on carbon saving are in part1.ipynb and part2.ipynb
We create a web dashboard to allow users to interactively explore the embodied carbon and operational carbon tradeoffs. To run the dashboard, execute
python dashboard.py
The dashboard is available at http://127.0.0.1:8080/
The user can provide their measurements on the old and new hardware, and the dashboard will show the carbon saving over time from the hardware upgrade. Here is what the dashboard looks like:
cd ../embodied_carbon
The embodied carbon data per HPC component is collected in csv/embodied_carbon_by_component.csv. The embodied carbon is modeled using ACT. The packaging overhead of Seagate SSD/HDDs are modeled using public sustainablity (Seagate Sustainability). More details of the embodied carbon modeling process are available in csv/embodied_carbon_modeling.xlsx
The number of devices of individual components from supercomputers are obtained from public resources:
The visualization scripts are provided in figure_<topic>.ipynb files. For example, figure_embodied.ipynb shows the visualization of embodied carbon of individual components.
cd ../regional_operational_carbon
The regional carbon intensity data are collected in the csv directory. The regional_carbon.ipynb notebook shows the visualization of regional carbon intensities.
Credit to Rohan Basu Roy for creating the Jupyter Notebook plotting scripts.
Baolin Li: https://baolin-li.netlify.app/
In addition to the characterization and analysis, we also designed a carbon-aware schduler named Clover: Code, Paper
We welcome contributions to this repository, especially additional embodied carbon data for different components in addition to our embodied carbon data in embodied_carbon_by_component.csv.