NetworkX accelerated by NVIDIA cuGraph is a newly released backend co-developed with the NetworkX team.
NVIDIA cuGraph provides GPU acceleration for popular graph algorithms such as PageRank, Louvain, and betweenness centrality. Depending on the algorithm and graph size, it can significantly accelerate NetworkX workflows, up to 50x, even 500x over NetworkX on CPU.
In this post, we share details about NetworkX and its new backend accelerator powered by cuGraph. We also discuss how to use the cuGraph backend for NetworkX, speed up your graph analytic workflows, and overcome the limitations NetworkX presents when working with large-scale graphs.
If you’re even remotely familiar with the field of graph data science, then you’ve probably used NetworkX. NetworkX is one of the most widely known open-source graph analytics libraries today and has over 80M downloads a month (as of 10/24).
Since its initial release in 2005, NetworkX has built a reputation for its easy-to-use, intuitive API, friendly and supportive community, and wide breadth of algorithms that address almost every graph analytics use case imaginable.
However, if you’re responsible for production workflows operating on real-world–sized data, then you’ve probably realized that when a graph reaches a certain size, NetworkX must be replaced with a more performant but also more complex graph solution.
NetworkX’s pure-Python, single-threaded implementation restricts its ability to scale. Many workflows using graphs typically larger than 100K nodes and 1M edges can result in significant slowdowns that often justify abandoning NetworkX altogether.
These workloads are common with high data volumes and comparisons between millions of entities and relationships, as done in fraud detection, recommendation systems, social network analysis, supply chain optimization, and others.
Since its initial release in 2019, the cuGraph open-source project has been building solutions for data scientists looking for an easy-to-use Python graph analytics library similar to NetworkX.
Now, with the addition of the cuGraph backend for NetworkX, which is the first production-ready GPU backend, you don’t have to leave NetworkX to experience material speedups.
Unlocking speed and scale
The backend for NetworkX is powered by cuGraph, a CUDA-based library, and enables NetworkX workflows running cuGraph-supported algorithms to see a range of speedups depending on the algorithm and the graph size.
You can see up to 10x, and for other algorithms and larger graphs up to 500x faster processing. The resulting speedups not only generate results faster but unlock use cases previously considered unrealistic for NetworkX.
Now graph data scientists can finally have both ease of use and speed at scale using NetworkX accelerated by NVIDIA cuGraph.
Zero-code-change GPU acceleration
NVIDIA cuGraph accelerates NetworkX through its backend dispatching system, which provides a standard mechanism for NetworkX to use third-party accelerators.
Install the cuGraph backend nx-cugraph using the pip or conda package managers or from the source on a system with an NVIDIA GPU.
The backend is activated through an environment variable that enables NetworkX automatic function dispatching, enabling NetworkX to send supported algorithms to the GPU backend and all others to the default NetworkX implementation on the CPU.
The result is automatic GPU acceleration for many of the most common graph algorithms used by data scientists without code changes, and without workarounds for missing functionality:
pagerankbetweenness_centralitylouvain_communitiesshortest_path- And many more, approximately 60 in total
You don’t need to modify your workflows to see massive speedups on large graph data that would normally make unaccelerated NetworkX virtually unusable. Those same workflows can also be used on non-GPU systems, where smaller data and longer runtimes can be tolerated for prototyping and experimentation.
From a command line, you can experience GPU speedup by setting an environment variable in your shell, where the same algorithm call runs nearly 87x faster when cuGraph is enabled.
File: demo.ipy
Graph workloads big and small
The flexibility of NetworkX is unmatched among other graph analytics libraries, and the resulting performance is acceptable for smaller graph data used for teaching, prototyping, and generally smaller problems.
But graphs exceeding 100K nodes and 1M edges are becoming much more common as the volume of data generated continues to grow, with enterprises expected to produce 20 ZB of data by 2027.
The cuGraph backend for NetworkX lets you have both ease of use for rapid development and the performance needed for modern workloads.
High volumes, along with multimodal data in modern workloads can be used by graphs to enrich the context and depth of analysis. Use cases where NetworkX accelerated by cuGraph is an ideal choice, involve large-scale graphs and fast processing requirements, outside the bounds of NetworkX on CPUs.
Graph algorithms boosted by GPU acceleration can be applied to use cases such as community or fraud detection systems or recommendation systems based on millions of consumers and their relationships to tens of millions of consumer profiles, transaction patterns, and inventory data points.
The following examples show the speedups using NetworkX accelerated by cuGraph.
Louvain community detection run on a large network graph of Hollywood actors to find those appearing in movies together (1M nodes, 58M edges) is 60x faster than NetworkX on CPU.
Figure 1. Louvain community detection algorithm run on a large network graph of Hollywood actors to find those appearing in movies together
PageRank algorithm used to compute values for a citation graph of U.S. patents (4M nodes, 16M edges) is 70x faster than NetworkX on CPU
Betweenness-centrality algorithm used to compute values for the Live Journal social network (5M nodes, 69M edges, k=100) is 485x faster than NetworkX on CPU.
All benchmarks in this post used the following configuration:
- Software: NetworkX 3.4.1, cuGraph/nx-cugraph 24.10
- CPU: Intel Xeon w9-3495X (56 cores) 250GB
- GPU: NVIDIA A100 80-G
SAGECONNECT TECHNOLOGY GROUP CO., LIMITED
Sageconnect is a reliable provider of IT network system integration and solutions for global enterprise customers.
