5. 2. Problem: RDBMS for
Knowledge Graphs● Using traditional database systems.
● SPARQL using database query processing.
● Use of 6-way indexes on the RDF data.
● Large intermediate join results.
Scalability of the query processor still remains a challenge.
5
6. 2. Solution: In-Memory Graph
Databases● In-memory data structure.
● Data structures that are suitable for disk storage as well as
for main memory.
● Suitable for fast data retrievals.
● Small intermediate join results.
My research project is based on two projects:
1. BitMat
6
9. 3. Knowledge Graphs
How do I represent the following fact:
“Pluto has been discovered in 1930”
in an intuitive way?
9
10. BitMat
● A compressed bit-matrix structure for
storing huge RDF graphs.
● A scalable lightweight join query
processor for RDF data.
● Employs a pruning technique to avoid
building intermediate join tables,
followed by a variable-binding
matching algorithm on in-memory
BitMats.
● Author: Medha Atre, Rensselaer
Polytechnic Institute Troy, USA.
SparqlSim
● A software prototype to compute graph
algorithms on RDF data.
● Reads SPARQL queries but computes
simulations as answers
● Was originally designed to compute pruning
for SPARQL query processing [ICDE’19]
based on graph pattern matching principles,
namely dual simulation.
● Author: Stephan Mennicke, Technische
Universität Braunschweig, Germany.
10
11. Actors
Keanu Reeves
Tom Cruise
Movies
The Matrix
John Wick
MI
Directors
Lana Wachowski
Brad Bird
Starring DirectedBy
11
SELECT ?actor ?movie ?directors WHERE
{
?actor <Starring> ?movie .
?movie <DirectedBy> ?director .
}
Y = [1, 0, 1, 0, 1, 0]
14. Rdf Bridge
● A high performance tool written in Python & Go to
generate BitMat structure based databases.
● Takes knowledge base as input.
● Generates BitMat database and BitMat configuration
file.
● Highly distributed and tested for SWAT Project - the
Lehigh University Benchmark (LUBM) with 1.3B triples.
Built using Python, Go, Redis & available at: 14
15. BitMat Interface
● A GUI tool that allows executing SPARQL queries
directly into the BitMat tool.
● Tool has internal parser that converts SPARQL into
BitMat query.
● Outputs query result with pretty print directly into the
interface, along with query statistics.
● Tested with SWAT Project - the Lehigh University
Benchmark (LUBM) with 1.3B triples.
15
16. Screening Graph in BitMat
● BitMat extension to support method of
inequalities over existing pruning approach.
● Implemented such that both original and method
of inequalities pruning can be evaluated on
same queries.
● Results into less intermediate join query results:
In some queries we were able to prune more
than 65M triples on Lubm dataset.
Written in C++ & available at:
16
18. Evaluation
We tested both pruning approaches with following datasets
● Lubm 1 - 103K Triples - Sub: 17K Pre: 18 Obj: 13K
● Lubm Full - 1.3B Triples - Sub: 223M Pre: 18 Obj: 167M
● DBPedia Sample - 97.5MTriples - Sub: 25.4K Pre: 31.4K Obj: 25.3M
● WikiData Sample - 318.5M Triples - Sub: 62M Pre: 4.6K Obj: 68.9M
Evaluation was performed on:
18
19. Results Comparison
Dataset
Avg. Prune Time (secs)
BitMat
Avg. Prune Time (secs)
SparqlSim
Lubm 1
103K Triples with 23 Queries
0.000753 0.204330
Lubm Full
1.3B Triples with 23 Queries
6.601197 2.802187
DBPedia Sample
97.5M Triples with 28 Queries
0.002939 0.033740
WikiData Sample
318.5M Triples with 43 Queries
0.031471 0.020458
19
20. Future Focus
● SparqlSim: Pruning as a service.
By creating Pruning service, we would like to evaluate BitMat performance.
Currently under development with SparqlSim.
● Dynamic programing for Sparql queries.
Query classification can be formed to automatically decide which type of queries are
suitable for which tool. 20
21. Thanks!
Any questions?
Or else you can reach me or find more details at:
● https://github.com/waqar-alamgir/BitMat/ - Waqar Alamgir, TU Braunschweig,
Germany
● Fast Dual Simulation Processing of Graph Database Queries - Stephan Mennicke,
TU Braunschweig, Germany
● BitMat: A Main Memory Bit-matrix of RDFTriples - Medha Atre, Rensselaer
Polytechnic Institute Troy, USA
20/20