Artifact of the BingoGCN Paper, ISCA 2025

Jiale Yan, Hiroaki Ito, Yuta Nagahara, Kazushi Kawamura, Masato Motomura, Thiem Van Chu, Daichi Fujiki.

---

This repository provides models and programs required for the artifact evaluation of the "BingoGCN: Towards Scalable and Efficient GNN Acceleration with Fine-Grained Partitioning and SLT" paper published in ISCA 2025.

The artifact of this paper includes models and programs to reproduce the key contributions of fine-grained partitioning and SLT algorithms.

Directory Structure

• BingoGCN: Main project directory containing the core code and scripts.
• Expected_results: Directory storing expected results for artifact evaluation.
• __outputs__: Directory storing checkpoints of the baseline models.
• env: Directory for environment-related files.
• scripts: Directory containing scripts for job execution used in the project.

Requirements

• NVIDIA GPUs with more than 32 GB of memory.
• OS supporting CUDA 11.6. The code is tested on Ubuntu 20.04.
• Anaconda, PyTorch 1.13, CUDA 11.6.
• A minimum of 4 GB of free disk space.

Environmental Setup

Install CUDA 11.6 on a machine running a supported OS. Install Anaconda or Miniconda. Then, download the repository and install dependencies as follows.

git clone https://github.com/LouiValley/BingoGCN.git
cd BingoGCN

conda env create -f env/conda.yml
conda activate BingoGCN
pip install -r env/requirements.txt

Run Artifact

Use one of the scripts in the scripts directory.

To reproduce all data points:

sh scripts/jobs_all.sh

Note

This will take approximately two days with 8 GPUs and can take weeks on a single GPU. For the artifact evaluation, we recommend the other options below.

To reproduce "Ours" only:

sh scripts/jobs_ours.sh

This script evaluates the data points of our proposed design, reducing the single-GPU runtime to a few days.

To reproduce small datasets in "Ours":

sh scripts/jobs_ours_light.sh

This script reduces repetition counts and skips some evaluations on large datasets. This will only take 2~3 hours to complete.

Post-Experiment Steps

Each experiment makes a directory under ./logs. After the experiments, run:

python BingoGCN/log_to_csv.py

This script aggregates the results into CSV files within each log directory.

Finally, compare these results with the ones in the expected_results directory. You can verify the data points in the figures with them. There can be minor discrepancies in the numbers due to non-deterministic factors such as RNG states.

Experiment Customizations

See Customize.md.

Citation

If you use BingoGCN, please cite this paper:

Jiale Yan, Hiroaki Ito, Yuta Nagahara, Kazushi Kawamura, Masato Motomura, Thiem Van Chu, and Daichi Fujiki, "BingoGCN: Towards Scalable and Efficient GNN Acceleration with Fine-Grained Partitioning and SLT," In Proceedings of the 52nd Annual International Symposium on Computer Architecture (ISCA'25)

@inproceedings{bingogcn,
  title={BingoGCN: Towards Scalable and Efficient GNN Acceleration with Fine-Grained Partitioning and SLT},
  author={Yan, Jiale and Ito, Hiroaki and Nagahara, Yuta and Kawamura, Kazushi and Motomura, Masato and Chu, Thiem Van and Fujiki, Daichi},
  booktitle={Proceedings of the 52nd Annual International Symposium on Computer Architecture}, 
  year={2025}
}

Licensing

This repository is available under a MIT license.