The relationships between data sets matter. Discovering, analyzing, and learning those relationships is a central part to expanding our understand, and is a critical step to being able to predict and act upon the data. Unfortunately, these are not always simple or quick tasks. To help the analyst we introduce RAPIDS, a collection of open-source libraries, incubated by NVIDIA and focused on accelerating the complete end-to-end data science ecosystem. Graph analytics is a critical piece of the data science ecosystem for processing linked data, and RAPIDS is pleased to offer cuGraph as our accelerated graph library. Simply accelerating algorithms only addressed a portion of the problem. To address the full problem space, RAPIDS cuGraph strives to be feature-rich, easy to use, and intuitive. Rather than limiting the solution to a single graph technology, cuGraph supports Property Graphs, Knowledge Graphs, Hyper-Graphs, Bipartite graphs, and the basic directed and undirected graph. A Python API allows the data to be manipulated as a DataFrame, similar and compatible with Pandas, with inputs and outputs being shared across the full RAPIDS suite, for example with the RAPIDS machine learning package, cuML. This talk will present an overview of RAPIDS and cuGraph. Discuss and show examples of how to manipulate and analyze bipartite and property graph, plus show how data can be shared with machine learning algorithms. The talk will include some performance and scalability metrics. Then conclude with a preview of upcoming features, like graph query language support, and the general RAPIDS roadmap.

1. Brad Rees, Connected Data London, Oct 4th, 2019
cuGraph
Accelerating all your Graph Analytic Needs

2. 2
Brad
Rees
Name
NVIDIA
Sr
Manager
cuGraph
Lead
PhD
Community
Detection in
Social
Networks
> 30
years
education
experience
Cyber
SNA
works at
Graph
Computer
Science
>20years
HPC
Big
Data

3. 3
WE ARE
CONNECTED
7 degrees of Kevin Bacon
Duncan Watts & Steven Strogatz
Collective dynamics of
‘small-world’ networks - 1998
And have
always been
connected The small-world problem - 1968
Stanley Milgram (social psychologist)
1929

4. 4
CONNECTEDNESS
CAPTURED AS A
GRAPH
As well as
associated
information,
knowledge,
metadata, etc..

5. 5
AND THERE ARE
A LOT OF GRAPH
FRAMEWORKS
In lots of variations
Neo4j
TigerGraph
AnzoGraph
RedisGraph
Oracle
Product names are the property of the owners
GraphX
Pegasus
Pregel
GraphLab
Giraph
Graphulo
PowerGraph
GaloisLigra
Gunrock
GraphBLAS
Stinger
HornetcuGraph
NetworkX
NetworkX

6. 6
Why cuGraph?
More generally, why RAPIDS?
A) Graph is not an isolated function, and
needs to be part of the complete Data Science
Process.
And Graph are just cool

7. 7
Speed, UX, and Iteration
The Way to Win at Data Science
Slide borrowed from Francois Chollet

8. 8
cuDF cuIO
Analytics
GPU Memory
Data Preparation VisualizationModel Training
cuML
Machine Learning
cuGraph
Graph Analytics
PyTorch Chainer MxNet
Deep Learning
cuXfilter <> pyViz
Visualization
Enter
End-to-End Accelerated GPU Data Science
Dask
Reduce Data Movement and Keep All Processing on the GPU

9. 9
ETL - the Backbone of Data Science
cuDF is…
Python Library
● A Python library for manipulating GPU DataFrames
following the Pandas API
● Python interface to CUDA C++ library with
additional functionality
● Creating GPU DataFrames from Numpy arrays,
Pandas DataFrames, and PyArrow Tables
● JIT compilation of User-Defined Functions (UDFs)
using Numba
● String Support

10. 10
Extraction is the Cornerstone of ETL
cuIO is born
• Follows the APIs of Pandas and provide >10x
speedup
• CSV Reader - v0.2, CSV Writer v0.8
• Parquet Reader – v0.7
• ORC Reader – v0.7
• JSON Reader - v0.8
• Avro Reader - v0.9
• HDF5 Reader - v0.10
• Key is GPU-accelerating both parsing and
decompression wherever possible
Source: Apache Crail blog: SQL Performance: Part 1 - Input File Formats

11. 11
cuML Machine Learning
GPU-accelerated Scikit-Learn
Classification / Regression
Statistical Inference
Clustering
Decomposition & Dimensionality Reduction
Time Series Forecasting
Recommendations
Decision Trees / Random Forests
Linear Regression
Logistic Regression
K-Nearest Neighbors
Kalman Filtering
Bayesian Inference
Gaussian Mixture Models
Hidden Markov Models
K-Means
DBSCAN
Spectral Clustering
Principal Components
Singular Value Decomposition
UMAP
Spectral Embedding
ARIMA
Holt-Winters
Implicit Matrix Factorization
Cross Validation
More to come!
Hyper-parameter Tuning
1x V100
vs
2x 20 core CPU

12. 12
cuGraph
Accelerating your Graph needs

13. 13
GOALS AND BENEFITS OF CUGRAPH
• Seamless integration with cuDF and cuML
•Python APIs accepts and returns cuDF DataFrames
• Allows for Property Graph
• Features
• Extensive collection of algorithm, primitive, and utility functions**
• With Accelerated Performance
• Python API:
• Multiple APIs: NetworkX, Pregel**, GraphBLAS**, Frontier**
• Graph Query Language**
• C/C++
• Full featured C++ API
Focus on Features an Easy-of-Use
** On Roadmap

14. 14
Graph Technology Stack
Python
Cython
C++ cuGraph Algorithms
Prims
CUDA Libraries
CUDA
Dask cuGraph
Dask cuDF
cuDF
Numpy
Thrust
Cub
cuSolver
cuSparse
cuRand
Gunrock*
cuGraphBLAS cuHornet
nvGRAPH has been Opened Sourced and integrated into cuGraph. * Gunrock is from UC Davis
cuGraphBLAS projected release Is. 0.12

15. 15
Bringing in leading researchers
Leveraging the great work of others
cuGraphGunrock Hornet
GraphBLAS
https://news.developer.nvidia.com/graph-technology-leaders-combine-forces-to-advance-graph-analytics/
cuHornet
cuGraphBLAS

16. 16
Algorithms
(as of release 0.10)
GPU-accelerated NetworkX
Community
Components
Link Analysis
Link Prediction
Traversal
Structure
Spectral Clustering
Balanced-Cut
Modularity Maximization
Louvain
Subgraph Extraction
Triangle Counting
Jaccard
Weighted Jaccard
Overlap Coefficient
Single Source Shortest Path (SSSP)
Breadth First Search (BFS)
COO-to-CSR
Transpose
Renumbering
Multi-GPU
More to come!
Utilities
Weakly Connected Components
Strongly Connected Components
Page Rank
Personal Page Rank
Katz
Query Language
Page Rank
OpenCypher:
Find-Matches
Long list of additional algorithms to come
Symmetrize

17. 17
PageRank Speedup
cuGraph PageRank vs NetworkX PageRank
G = cugraph.Graph()
G.add_edge_list(gdf[‘src’], gdf[‘dst’], None)
df = cugraph.pagerank(G, alpha, max_iter, tol)
https://github.com/rapidsai/notebooks-extended/tree/master/advanced/benchmarks/cugraph_benchmark
SciPy

18. 18
PageRank Performance
HiBench Websearch benchmark
All times are in seconds
Vertices Edges
File Size
(GB)
Number of
GPUs
Read data
and create
DataFrame
Run
Pagerank
(20 iterations)
Write
Scores
TOTAL
runtime
50,000,000 1,980,000,000 34 3 28.6 6.8 6.2 41.6
100,000,000 4,000,000,000 69 6 33.4 11.3 12.7 57.4
200,000,000 8,000,000,000 146 12 36.8 24.4 26.7 87.9
400,000,000 16,000,000,000 300 16 58.3 42.8 53.0 154.1
Ø Process
Ø Read Data
Ø Parse CSV into DataFrame
Ø Run Page Rank
Ø Convert Data to CSR
Ø Setup
Ø Run PagePage Solver
Ø Collect Results and convert of a DataFrame
Ø Write Score

19. 19
Faster Speeds, Real-World Benefits
cuIO/cuDF –
Load and Data Preparation cuML - XGBoost
Time in seconds (shorter is better)
cuIO/cuDF (Load and Data Prep) Data Conversion XGBoost
Benchmark
200GB CSV dataset; Data prep includes
joins, variable transformations
CPU Cluster Configuration
CPU nodes (61 GiB memory, 8 vCPUs, 64-
bit platform), Apache Spark
DGX Cluster Configuration
5x DGX-1 on InfiniBand
network
8762
6148
3925
3221
322
213
End-to-End
Non-Graph

20. 20

21. 21
Deploy RAPIDS Everywhere
Focused on robust functionality, deployment, and user experience
Integration with major cloud providers
Both containers and cloud specific machine instances
Support for Enterprise and HPC Orchestration Layers
Cloud Dataproc Azure Machine Learning

22. G R A P H I S T info@graphistry.com
Data Scientist
Notebooks
Dev API For
Embedding
Analyst
Tool Suite
Automate
Investigations
Virtual Graph over
graph and tabular APIs
GPU Visual Analytics:
• 100X via GPUs:
client<>cloud
• Correlate w/ graph
• Time, histograms, …
100X Investigations with Graphistry:
Visibility & workflows for handling modern enterprise data
G R A P H I S T R Y

23. 23
Articles

24. THANK YOU
Please give us a star on GitHub
https://github.com/rapidsai/cugraph
Questions?

25. 25
PageRank Performance
HiBench Websearch benchmark
All times are in seconds
Vertices Edges
File Size
(GB)
Number of
GPUs
Read data
and create
DataFrame
Run
Pagerank
(20 iterations)
Write
Scores
TOTAL
runtime
50,000,000 1,980,000,000 34 3 28.6 6.8 6.2 41.6
100,000,000 4,000,000,000 69 6 33.4 11.3 12.7 57.4
200,000,000 8,000,000,000 146 12 36.8 24.4 26.7 87.9
400,000,000 16,000,000,000 300 16 58.3 42.8 53.0 154.1
Vertices Edges
Convert
DataFrame
to CSR
Just
PageRank
Solver
50,000,000 1,980,000,000 2.4 3.66
100,000,000 4,000,000,000 4.5 5.16
200,000,000 8,000,000,000 9.6 8.65
400,000,000 16,000,000,000 19.5 13.89
Ø Process
Ø Read Data
Ø Parse CSV into DataFrame
Ø Run Page Rank
Ø Convert Data to CSR
Ø Setup
Ø Run PagePage Solver
Ø Collect Results and convert of a DataFrame
Ø Write Score