Gunrock: CUDA/C++ GPU Graph Analytics

| Examples | Project Template | Documentation | GitHub Actions | |--------------|----------------------|-------------------|-------------------|

Gunrock[^1] is a CUDA library for graph-processing designed specifically for the GPU. It uses a high-level, bulk-synchronous/asynchronous, data-centric abstraction focused on operations on vertex or edge frontiers. Gunrock achieves a balance between performance and expressiveness by coupling high-performance GPU computing primitives and optimization strategies, particularly in the area of fine-grained load balancing, with a high-level programming model that allows programmers to quickly develop new graph primitives that scale from one to many GPUs on a node with small code size and minimal GPU programming knowledge.

| Branch | Purpose | Version | Status | |-----------|------------------------------------------------------------------------------------------------------------------------------------|----------------|------------| | main | Default branch, ported from gunrock/essentials, serves as the official release branch. | $\geq$ 2.x.x | Active | | develop | Development feature branch, ported from gunrock/essentials. | $\geq$ 2.x.x | Active | | master | Previous release branch for gunrock/gunrock version 1.x.x interface, preserves all commit history. | $\leq$ 1.x.x | Deprecated | | dev | Previous development branch for gunrock/gunrock. All changes now merged in master. | $\leq$ 1.x.x | Deprecated |

Quick Start Guide

Before building Gunrock make sure you have CUDA Toolkit[^2] installed on your system. Other external dependencies such as NVIDIA/thrust, NVIDIA/cub, etc. are automatically fetched using cmake.

shell git clone https://github.com/gunrock/gunrock.git cd gunrock mkdir build && cd build cmake .. make sssp # or for all algorithms, use: make -j$(nproc) bin/sssp ../datasets/chesapeake/chesapeake.mtx

Implementing Graph Algorithms

For a detailed explanation, please see the full documentation. The following example shows simple APIs using Gunrock's data-centric, bulk-synchronous programming model, we implement Breadth-First Search on GPUs. This example skips the setup phase of creating a problem_t and enactor_t struct and jumps straight into the actual algorithm.

We first prepare our frontier with the initial source vertex to begin push-based BFS traversal. A simple f->push_back(source) places the initial vertex we will use for our first iteration. cpp void prepare_frontier(frontier_t* f, gcuda::multi_context_t& context) override { auto P = this->get_problem(); f->push_back(P->param.single_source); } We then begin our iterative loop, which iterates until a convergence condition has been met. If no condition has been specified, the loop converges when the frontier is empty. ```cpp void loop(gcuda::multicontextt& context) override { auto E = this->getenactor(); // Pointer to enactor interface. auto P = this->getproblem(); // Pointer to problem (data) interface. auto G = P->get_graph(); // Graph that we are processing.

auto singlesource = P->param.singlesource; // Initial source node. auto distances = P->result.distances; // Distances array for BFS. auto visited = P->visited.data().get(); // Visited map. auto iteration = this->iteration; // Iteration we are on.

// Following lambda expression is applied on every source, // neighbor, edge, weight tuple during the traversal. // Our intent here is to find and update the minimum distance when found. // And return which neighbor goes in the output frontier after traversal. auto search = [=] host device( vertext const& source, // ... source vertext const& neighbor, // neighbor edget const& edge, // edge weightt const& weight // weight (tuple). ) -> bool { auto olddistance = math::atomic::min(&distances[neighbor], iteration + 1); return (iteration + 1 < olddistance); };

// Execute advance operator on the search lambda expression. // Uses loadbalancet::blockmapped algorithm (try others for perf. tuning.) operators::advance::execute<operators::loadbalancet::blockmapped>( G, E, search, context); } ``` include/gunrock/algorithms/bfs.hxx

How to Cite Gunrock & Essentials

Thank you for citing our work.

bibtex @article{Wang:2017:GGG, author = {Yangzihao Wang and Yuechao Pan and Andrew Davidson and Yuduo Wu and Carl Yang and Leyuan Wang and Muhammad Osama and Chenshan Yuan and Weitang Liu and Andy T. Riffel and John D. Owens}, title = {{G}unrock: {GPU} Graph Analytics}, journal = {ACM Transactions on Parallel Computing}, year = 2017, volume = 4, number = 1, month = aug, pages = {3:1--3:49}, doi = {10.1145/3108140}, ee = {http://arxiv.org/abs/1701.01170}, acmauthorize = {https://dl.acm.org/doi/10.1145/3108140?cid=81100458295}, url = {http://escholarship.org/uc/item/9gj6r1dj}, code = {https://github.com/gunrock/gunrock}, ucdcite = {a115}, }

bibtex @InProceedings{Osama:2022:EOP, author = {Muhammad Osama and Serban D. Porumbescu and John D. Owens}, title = {Essentials of Parallel Graph Analytics}, booktitle = {Proceedings of the Workshop on Graphs, Architectures, Programming, and Learning}, year = 2022, series = {GrAPL 2022}, month = may, pages = {314--317}, doi = {10.1109/IPDPSW55747.2022.00061}, url = {https://escholarship.org/uc/item/2p19z28q}, }

Copyright & License

Gunrock is copyright The Regents of the University of California. The library, examples, and all source code are released under Apache 2.0.

[^1]: This repository has been moved from https://github.com/gunrock/essentials and the previous history is preserved with tags and under master branch. Read more about gunrock and essentials in our vision paper: Essentials of Parallel Graph Analytics. [^2]: Recommended CUDA v11.5.1 or higher due to support for stream ordered memory allocators.