The RAPIDS cuGraph library is a collection of GPU accelerated graph algorithms that process data found in GPU DataFrames. The vision of cuGraph is to make graph analysis ubiquitous to the point that users just think in terms of analysis and not technologies or frameworks. To realize that vision, cuGraph operates, at the Python layer, on GPU DataFrames, thereby allowing for seamless passing of data between ETL tasks in cuDF and machine learning tasks in cuML. Data scientists familiar with Python will quickly pick up how cuGraph integrates with the Pandas-like API of cuDF. Likewise, users familiar with NetworkX will quickly recognize the NetworkX-like API provided in cuGraph, with the goal to allow existing code to be ported with minimal effort into RAPIDS.
While the high-level cugraph python API provides an easy-to-use and familiar interface for data scientists that's consistent with other RAPIDS libraries in their workflow, some use cases require access to lower-level graph theory concepts. For these users, we provide an additional Python API called pylibcugraph, intended for applications that require a tighter integration with cuGraph at the Python layer with fewer dependencies. Users familiar with C/C++/CUDA and graph structures can access libcugraph and libcugraph_c for low level integration outside of python.
For more project details, see rapids.ai.
NOTE: For the latest stable README.md ensure you are on the latest branch.
As an example, the following Python snippet loads graph data and computes PageRank:
import cudf
import cugraph
# read data into a cuDF DataFrame using read_csv
gdf = cudf.read_csv("graph_data.csv", names=["src", "dst"], dtype=["int32", "int32"])
# We now have data as edge pairs
# create a Graph using the source (src) and destination (dst) vertex pairs
G = cugraph.Graph()
G.from_cudf_edgelist(gdf, source='src', destination='dst')
# Let's now get the PageRank score of each vertex by calling cugraph.pagerank
df_page = cugraph.pagerank(G)
# Let's look at the top 10 PageRank Score
df_page.sort_values('pagerank', ascending=False).head(10)There are 3 ways to get cuGraph :
At GTC Spring '22 we presented results of running cuGraph on the Selene supercomputer using 2,048 GPUs and processing a graph with 1.1 Trillion edges. Synthetic data created with the RMAT generator found in cuGraph.
cuGraph Scaling
cuGraph has a new multi-layer software stack that allows users and system integrators to access cuGraph at different layers.
cuGraph Software Stack
As of Release 22.06
cuGraph supports graph creation with Source and Destination being expressed as:
- cuDF DataFrame
- Pandas DataFrame
cuGraph supports execution of graph algorithms from different graph objects
- cuGraph Graph classes
- NetworkX graph classes
- CuPy sparse matrix
- SciPy sparse matrix
cuGraph tries to match the return type based on the input type. So a NetworkX input will return the same data type that NetworkX would have.
| Type | Description |
|---|---|
| Graph | An undirected Graph by default |
| directed=True yields a Directed Graph | |
| Multigraph | A Graph with multiple edges between a vertex pair |
ALL Algorithms support Graphs and MultiGraph (directed and undirected)
Italic algorithms are planned for future releases.
| Category | Algorithm | Scale | Notes |
|---|---|---|---|
| Centrality | |||
| Katz | Multi-GPU | ||
| Betweenness Centrality | Single-GPU | ||
| Edge Betweenness Centrality | Single-GPU | ||
| Eigenvector Centrality | Multi-GPU | ||
| Degree Centrality | Multi-GPU | Python only | |
| Community | |||
| Leiden | Single-GPU | ||
| Louvain | Multi-GPU | ||
| Ensemble Clustering for Graphs | Single-GPU | ||
| Spectral-Clustering - Balanced Cut | Single-GPU | ||
| Spectral-Clustering - Modularity | Single-GPU | ||
| Subgraph Extraction | Single-GPU | ||
| Triangle Counting | Multi-GPU | ||
| K-Truss | Single-GPU | ||
| Components | |||
| Weakly Connected Components | Multi-GPU | ||
| Strongly Connected Components | Single-GPU | ||
| Core | |||
| K-Core | Single-GPU | ||
| Core Number | Single-GPU | ||
| Flow | |||
| MaxFlow | --- | ||
| Influence | |||
| Influence Maximization | --- | ||
| Layout | |||
| Force Atlas 2 | Single-GPU | ||
| Linear Assignment | |||
| Hungarian | Single-GPU | README | |
| Link Analysis | |||
| Pagerank | Multi-GPU | C++ README | |
| Personal Pagerank | Multi-GPU | C++ README | |
| HITS | Multi-GPU | ||
| Link Prediction | |||
| Jaccard Similarity | Single-GPU | ||
| Weighted Jaccard Similarity | Single-GPU | ||
| Overlap Similarity | Single-GPU | ||
| Sorensen Coefficient | Single-GPU | Python only | |
| Local Clustering Coefficient | --- | ||
| Sampling | |||
| Random Walks (RW) | Single-GPU | Biased and Uniform | |
| Egonet | Single-GPU | multi-seed | |
| Node2Vec | Single-GPU | ||
| Neighborhood sampling | Multi-GPU | ||
| Traversal | |||
| Breadth First Search (BFS) | Multi-GPU | with cutoff support C++ README |
|
| Single Source Shortest Path (SSSP) | Multi-GPU | C++ README | |
| ASSP / APSP | |||
| Tree | |||
| Minimum Spanning Tree | Single-GPU | ||
| Maximum Spanning Tree | Single-GPU | ||
| Other | |||
| Renumbering | Multi-GPU | multiple columns, any data type | |
| Symmetrize | Multi-GPU | ||
| Path Extraction | Extract paths from BFS/SSP results in parallel | ||
| Data Generator | |||
| RMAT | Multi-GPU | ||
| Barabasi-Albert | --- | ||
Vertex IDs are expected to be contiguous integers starting from 0. If your data doesn't match that restriction, we have a solution. cuGraph provides the renumber function, which is by default automatically called when data is added to a graph. Input vertex IDs for the renumber function can be any type, can be non-contiguous, can be multiple columns, and can start from an arbitrary number. The renumber function maps the provided input vertex IDs to either 32- or 64-bit contiguous integers starting from 0.
Additionally, when using the auto-renumbering feature, vertices are automatically un-renumbered in results.
cuGraph is constantly being updated and improved. Please see the Transition Guide if errors are encountered with newer versions
The amount of memory required is dependent on the graph structure and the analytics being executed. As a simple rule of thumb, the amount of GPU memory should be about twice the size of the data size. That gives overhead for the CSV reader and other transform functions. There are ways around the rule but using smaller data chunks.
| Size | Recommended GPU Memory |
|---|---|
| 500 million edges | 32 GB |
| 250 million edges | 16 GB |
The use of managed memory for oversubscription can also be used to exceed the above memory limitations. See the recent blog on Tackling Large Graphs with RAPIDS cuGraph and CUDA Unified Memory on GPUs: https://medium.com/rapids-ai/tackling-large-graphs-with-rapids-cugraph-and-unified-virtual-memory-b5b69a065d4
Please see the Docker Repository, choosing a tag based on the NVIDIA CUDA version you’re running. This provides a ready to run Docker container with example notebooks and data, showcasing how you can utilize all of the RAPIDS libraries: cuDF, cuML, and cuGraph.
It is easy to install cuGraph using conda. You can get a minimal conda installation with Miniconda or get the full installation with Anaconda.
Install and update cuGraph using the conda command:
# CUDA 11.4
conda install -c nvidia -c rapidsai -c numba -c conda-forge cugraph cudatoolkit=11.4
# CUDA 11.5
conda install -c nvidia -c rapidsai -c numba -c conda-forge cugraph cudatoolkit=11.5
For CUDA > 11.5, please use the 11.5 environmentNote: This conda installation only applies to Linux and Python versions 3.8/3.9.
Please see our guide for building cuGraph from source
Please see our guide for contributing to cuGraph.
Python API documentation can be generated from docs directory.
(alphabetical order)
- ArangoDB - a free and open-source native multi-model database system - https://www.arangodb.com/
- CuPy - "NumPy/SciPy-compatible Array Library for GPU-accelerated Computing with Python" - https://cupy.dev/
- Memgraph - In-memory database - https://memgraph.com/
- ScanPy - a scalable toolkit for analyzing single-cell gene expression data - https://scanpy.readthedocs.io/en/stable/
The RAPIDS suite of open source software libraries aims to enable execution of end-to-end data science and analytics pipelines entirely on GPUs. It relies on NVIDIA® CUDA® primitives for low-level compute optimization but exposing that GPU parallelism and high-bandwidth memory speed through user-friendly Python interfaces.
The GPU version of Apache Arrow is a common API that enables efficient interchange of tabular data between processes running on the GPU. End-to-end computation on the GPU avoids unnecessary copying and converting of data off the GPU, reducing compute time and cost for high-performance analytics common in artificial intelligence workloads. As the name implies, cuDF uses the Apache Arrow columnar data format on the GPU. Currently, a subset of the features in Apache Arrow are supported.