PathWeaver

A High-Throughput Multi-GPU System for Graph-Based Approximate Nearest Neighbor Search

Key Files

File Path	Description
./pathweaver/csrc/pathweaver.cu	CUDA kernel implementation for PathWeaver's graph search
./pathweaver/single_pathweaver_one.py	Python file to run PathWeaver on a single GPU
./pathweaver/single_pathweaver_all.sh	Script to run PathWeaver on a single GPU with parameters
./pathweaver/multi_pathweaver_one.py	Python file to run PathWeaver on multiple GPUs
./pathweaver/multi_pathweaver_all.sh	Script to run PathWeaver on multiple GPUs with parameters

Hardware Requirements

• Approximately 250 GB of free disk space
• 4 GPUs with at least 48 GB of device memory
• NVLink bridge across GPUs
• CUDA 12.1

Steps to Reproduce

0. Environment Setup

Install dependencies via Conda:

conda env create -f environment.yaml
conda activate pathweaver

1. Build PathWeaver Binary

cd pathweaver
bash build.sh

2. Run Single-GPU Artifact Evaluation

bash single_pathweaver_all.sh

3. Run Multi-GPU Artifact Evaluation

bash multi_pathweaver_all.sh

Citation

@inproceedings{kim2025pathweaver,
  title     = {PathWeaver: A High-Throughput Multi-GPU System for Graph-Based Approximate Nearest Neighbor Search},
  author    = {Kim, Sukjin and Park, Seongyeon and Noh, Si Ung and Hong, Junguk and Kwon, Taehee and Lim, Hunseong and Lee, Jinho},
  booktitle = {Proceedings of the 2025 USENIX Annual Technical Conference (USENIX ATC '25)},
  pages     = {1501--1517},
  year      = {2025},
  month     = jul,
  address   = {Boston, MA, USA},
  publisher = {USENIX Association},
  isbn      = {978-1-939133-48-9},
  url       = {https://www.usenix.org/conference/atc25/presentation/kim}
}

Acknowledgment

This project is built upon and extends the CAGRA method from the RAPIDS cuVS library, an efficient GPU-accelerated framework for vector search.