# PilotANN > Memory-bounded GPU acceleration for graph-based approximate nearest neighbor search. ## Build ### RHEL/CentOS Setup ``` # Install development tools sudo dnf install @'Development Tools' # Install additional dependencies sudo dnf install epel-release sudo dnf config-manager --set-enabled crb sudo dnf install cmake git swig lapack lapack-devel ``` ### CUDA Setup ``` # Enable NVIDIA repository sudo dnf config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel9/x86_64/cuda-rhel9.repo # Install CUDA sudo dnf clean all sudo dnf -y install cuda-toolkit-11-8 sudo dnf module install nvidia-driver:latest-dkms # Set environment variables (add to ~/.bashrc): export PATH=/usr/local/cuda/bin${PATH:+:${PATH}} export LD_LIBRARY_PATH=/usr/local/cuda/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}} ``` ### Python Setup (Anaconda or Miniconda) ``` # The code is built on top of LibTorch conda create -n pytorch python=3.10 && conda activate pytorch conda install pytorch=2.1.2 torchvision torchaudio pytorch-cuda=11.8 numpy -c pytorch -c nvidia ``` ### Build FAISS (Under Current Python Env) ``` # Download source git clone https://github.com/facebookresearch/faiss/archive/refs/tags/v1.8.0.tar.gz tar -xvf v1.8.0.tar.gz && cd faiss-* # Build faiss mkdir build && cd build cmake .. -DFAISS_OPT_LEVEL='avx2' \ -DFAISS_ENABLE_GPU=OFF \ -DFAISS_ENABLE_PYTHON=ON \ -DCMAKE_BUILD_TYPE=Release # Install cd faiss/python && python3 ./setup.py install ``` ### Build PilotANN (Under Current Python Env) ``` cd pilot_ann_src python3 ./setup.py develop ``` ## Dataset Setup + Expected directory structure: ``` .datasets/ ├── deep-1m/ ├── deep-100m/ └── laion-100m/ ``` ## Run Benchmark ``` # Using 16 CPUs export OMP_THREAD_LIMIT=16 # Run benchmark python3 script/bench_1.py --dataset='laion-1m' --sample_ratio=0.25 --d_principle=128 python3 script/bench_1.py --dataset='laion-100m' --sample_ratio=0.25 --d_principle=128 ``` ### Available benchmark options + --top_k: Set `k` (any int <= 100, default 10) + --dataset: Choose dataset (deep-1m, text2img-1m, wiki-1m, laion-1m) + --sample_ratio: Set sampling ratio (any float in (0.0, 1.0]) + --d_principle: Set SVD reduction dimension (any int, prefer numbers divisible by 32) + --algorithm: Choose graph angorithm (hnsw, nsg) + --router: Entry selection method (random32, random64, router32x32, router64x64, router means FES) + --n_queries: Set number of queries to process (default 1024) Benchmark Results + Example output ``` [hnsw] dataset: laion-1m, d_model: 768 svd 128 build nsw sampling 0.25 [INFO] building entry evaluate: 0%| | 0/22 [00:00