Graphs are common in a variety of situation, from social networks to financial transactions. These graphs are gold mines of information, and being able to process them is key to the success of many businesses. Real graph, however, are growing larger and larger, and traditional software and algorithms cannot satisfy the performance needs of the most demanding users. Gospel, a research line at NECSTLab, aims at making graph processing faster and readily available to researchers and industry, by leveraging high-performance heterogeneous and novel computer architectures. From accelerating algorithms such as PageRank by making use of modern GPUs, to extending existing frameworks for graph analysis, Gospel offers a broad array of techniques to bring graph processing to the world of high-performance computing.

1. Alberto Parravicini
2019-05-{19-31}, NGCX
Gospel -
High-performance
graph analytics

2. Alberto Parravicini 23/05/2019
High-Performance
Graph Analytics
Graphs are a gold mine of information
● Social Networks
● Financial transactions
● Recommender Systems
Real graphs are enormous
● Facebook: 2B users, Wikipedia: 200M links
We need high-performance and scalable ways
to process graphs
2

3. Alberto Parravicini 23/05/2019
Introducing Gospel
● High-performance heterogeneous
architectures for graph analytics
● Make graph processing faster and
readily available to researchers and industry
Quick examples:
1. PageRank on Wikipedia in 0.2 sec
2. Real-time Entity Linking with >80% accuracy
3

4. Alberto Parravicini 23/05/2019
The Gospel Graph
Labs
4
GPU Algorithm Acceleration
● PageRank, Embeddings
Framework/DSL extension
● Green-Marl by Oracle Labs
Embeddings & ML apps
● Entity Linking

5. Alberto Parravicini 23/05/2019
Approximating PageRank
on GPU
● The workhorse of Google’s search,
now used in many applications
● Computation time is a
bottleneck, even with GPUs
● No need for 100% accuracy
● The ranking is what matters!
● Leverage approximate computing:
● Low precision arithmetic,
loop perforation, numerical tricks
5

6. Alberto Parravicini 23/05/2019
PageRank on graphs larger
than GPU memory
● Most real-world graphs are larger than GPU
memory (e.g. the web!)
● Adapt PR to work in these cases
● Graph partitioning
● Data compression
● Double buffering and pipelining
● These techniques can be adapted to many
algorithms, and to a multi-GPU scenario
6

7. Alberto Parravicini 23/05/2019
Accelerating embeddings
primitives on GPU
● Vertex embeddings are key to perform
machine learning on graphs
● Most algorithms have the same primitives:
random walks, vertex sampling, etc…
● Often done on CPU, and results sent to GPU
● Our directions:
● Do everything on GPU,
using modified Bread-First Visit
7

8. Alberto Parravicini 23/05/2019
Fast Entity Linking via Graph
Embeddings (1/2)
● Entity Linking (EL): connect Named Entities
to unique identities (e.g. Wikipedia Page)
● Lots of applications: search engines,
recommender systems, chat bots
8
Labs

9. Alberto Parravicini 23/05/2019
Fast Entity Linking via Graph
Embeddings (2/2)
● The first EL
algorithm to
leverage graph
embeddings
● SoA results (>80%
accuracy) with
real-time latency
(30 names / sec)
Labs

10. Alberto Parravicini 23/05/2019
Automatic GPU code
generation from graph DSL
● Writing high-performance GPU graph
algorithms is difficult
● We can extend Green-Marl, a graph DSL
developed by Oracle Labs
● Graph computation as linear algebra kernels
● We leverage GraphBlast,
GPU library from the authors of Gunrock
● Easy transparent acceleration of many
algorithms (e.g. BFS, PR)
10
Labs

11. The Gospel Folks
Alberto Parravicini
Francesco Sgherzi
Elisa Tardini
Nicolò Scipione
Ivan Montalbano
Rolando Brondolin
Davide Bartolini
Rhicheek Patra
Marco Santambrogio
2019-05-{19-31}, NGCX
alberto.parravicini@polimi.it

12. ● Approximating PageRank on GPU
● PageRank on graphs larger than GPU memory
● Accelerating embeddings primitives on GPU
● Fast Entity Linking via Graph Embeddings
● Automatic GPU code generation from Graph DSL
Thank you!
Gospel - High-performance graph analytics
Alberto Parravicini
2019-05-{19-31}, NGCX
alberto.parravicini@polimi.it