grafeo-cuda

GPU-accelerated graph algorithms for the Grafeo graph database.

Overview

This crate provides CUDA-accelerated implementations of graph algorithms (triangle counting, BFS, PageRank, k-truss) that operate on Grafeo's graph data. It lives in a separate repository to keep GPU dependencies out of the core Grafeo project.

Supported hardware: NVIDIA GPUs with Compute Capability 7.0+ (RTX 20xx and newer).

Architecture

Grafeo (LpgStore) -> CSR export -> GPU transfer -> CUDA kernel -> Results

Grafeo's GraphStore is exported to CSR (Compressed Sparse Row) format, transferred to GPU memory via cudarc, and processed by CUDA kernels compiled from PTX. Results are transferred back to host memory.

Quick Start

Prerequisites

• Rust 1.91.1+
• CUDA Toolkit 12.x (nvcc on PATH)
• NVIDIA GPU (RTX 20xx or newer)

Build

cargo build --release

The build.rs script compiles CUDA kernels in kernels/ to PTX via nvcc. If nvcc is not available, placeholder PTX files are created and GPU features will return runtime errors.

Usage (Rust)

use grafeo_engine::GrafeoDB;
use grafeo_cuda_core::{GpuDevice, GpuGraph};
use grafeo_cuda_algos::{triangle_counting, pagerank, bfs};

let db = GrafeoDB::new_in_memory();
db.import_tsv("graph.tsv", "E", true).unwrap();

let device = GpuDevice::new(0)?;
let gpu = GpuGraph::from_grafeo(&db, &device)?;

let triangles = triangle_counting::triangle_count(&gpu)?;
let ranks = pagerank::pagerank(&gpu, 0.85, 20, 1e-6)?;
let distances = bfs::bfs(&gpu, 0)?;

Benchmark Binaries

# Triangle counting
cargo run --release --bin gpu-triangles -- --dataset graph.tsv --save

# PageRank
cargo run --release --bin gpu-pagerank -- --dataset graph.tsv --save

# BFS
cargo run --release --bin gpu-bfs -- --dataset graph.tsv --save

# K-Truss decomposition
cargo run --release --bin gpu-ktruss -- --dataset graph.tsv --k 4 --save

Usage (Python)

from grafeo_cuda import GpuDevice, GpuGraph

device = GpuDevice(0)
gpu = GpuGraph.from_csr(offsets, targets, device)

result = gpu.triangle_count()
print(f"Triangles: {result['total_triangles']}")
print(f"GPU time: {result['time_seconds']:.4f}s")

Crate Structure

Crate	Purpose
grafeo-cuda-core	GPU device management, CSR export, memory transfer
grafeo-cuda-algos	Algorithm implementations + benchmark binaries
grafeo-cuda-python	Python bindings via PyO3

Algorithms

Algorithm	Status	Kernel	Benchmark Binary
Triangle Counting	Implemented	kernels/triangle_counting.cu	gpu-triangles
PageRank	Implemented	kernels/pagerank.cu	gpu-pagerank
BFS	Implemented	kernels/bfs.cu	gpu-bfs
K-Truss	Implemented	kernels/k_truss.cu	gpu-ktruss

License

Apache-2.0