GGNN Documentation

GGNN performs nearest-neighbor computations on CUDA-capable GPUs. It supports billion-scale, high-dimensional datasets and can execute on multiple GPUs through sharding. When using just a single GPU, data can be exchanged directly with other code (e.g., torch tensors) without copying through CPU memory. GGNN is implemented using C++ and CUDA. It can also be used from Python (>=3.8) via its nanobind bindings.

GGNN is based on the method proposed in the paper GGNN: Graph-based GPU Nearest Neighbor Search by Fabian Groh, Lukas Ruppert, Patrick Wieschollek, and Hendrik P.A. Lensch. The original/official code corresponding to the published paper can be found in the release_0.5 branch.

The Installation section explains how to install the library, and the Using the GGNN Python Module and Using the GGNN C++ Library sections provides short tutorials and code examples.

Contents

• Home
• Installation
  • Dependencies
  • Installing the GGNN Python Module
  • Installing the GGNN C++ Library
  • Troubleshooting
• Approximate Nearest Neighbor Search
  • Search Graph Parameters
  • Query Parameters
  • Distance Measures
• Using the GGNN Python Module
  • Importing GGNN
  • Using CPU Data
  • Using GPU Data
  • Using Multiple GPUs
  • Loading Datasets (e.g. SIFT1M)
• Using the GGNN C++ Library
  • Including GGNN
  • Using CPU Data
  • Using GPU Data
  • Using multiple GPUs
  • Loading Datasets (e.g. SIFT1M)
• Benchmarking
  • Running Standardized Benchmarks
  • ANN-Benchmarks / HDF5
  • Reference Configurations
• FAQ
  • Known Issues
  • I can’t find the answer to my question here

Capabilities and Limitations

The GGNN library supports…

• Billion-scale datasets with up to \(2^{31}-1\) vectors.

• Data with up to 4096 dimensions.

• Building search graphs with up to 512 edges per node.

• Searching for up to 6000 nearest neighbors.

• Two distance measures: cosine and euclidean (L2) distance.

Citing this Project

You can use the following BibTeX entry to cite GGNN:

@ARTICLE{groh2022ggnn,
  author={Groh, Fabian and Ruppert, Lukas and Wieschollek, Patrick and Lensch, Hendrik P. A.},
  journal={IEEE Transactions on Big Data},
  title={GGNN: Graph-Based GPU Nearest Neighbor Search},
  year={2023},
  volume={9},
  number={1},
  pages={267-279},
  doi={10.1109/TBDATA.2022.3161156}
}

The official article can be found on IEEE, following this DOI: 10.1109/TBDATA.2022.3161156. Alternatively, see the ArXiV preprint.