This document presents a study on streaming graph partitioning algorithms. The goals are to evaluate the benefits of more complex partitioning methods, measure partitioning overhead, and determine the effect of partitioning quality on application performance. Experiments were run on several datasets and platforms. Results show that hash partitioning has the best throughput but worse edge cuts than other algorithms. Load balancing was high for most methods. Iterative applications showed differences in performance depending on partitioning, while streaming applications behaved differently. Open questions remain around state requirements, adapting algorithms for streaming, and handling more complex graphs.

1. STREAMING GRAPH
PARTITIONING
Zainab Abbas1, Vasiliki Kalavri2,
Paris Carbone1, and Vladimir Vlassov1
1. KTH Royal Institute of Technology, Stockholm, Sweden
2. ETH Zurich, Switzerland
An Experimental Study

2. Distributed Graph Analysis
Input graph partitioning computation

3. Iterative Applications
Loading Streaming
Partitioning
3

4. Streaming Applications
Loading
4
Streaming
Partitioning

5. Streaming Partitioning
5
Ingest input as stream
Partitioning on the fly
Partial graph
knowledge
Single Pass
Streaming
Partitioning

6. Iterative Connected Components
6
1
34
5
2 6
87

7. Iterative Connected Components
7
1
34
5
2 6
87
2
4 3
1
3
1
5
1
2
54 3
7 8
8
6 6
7

8. Iterative Connected Components
8
1
34
5
2 6
87
2 4 3
1 3
15 1 2 5
4 3
7 8
86
6 7

9. Streaming Connected Components (union-find)
9
1
34
5
2
6
87
Component ID Vertices
3
2
5
4
6
7
Component ID Vertices
2 2,3

10. Streaming Connected Components (union-find)
10
1
34
5
2
6
87
3
2
5
4
6
7
Component ID Vertices
4 4,5
Component ID Vertices
2 2,3

11. Streaming Connected Components (union-find)
11
1
34
5
2
6
87
3
2
5
4
6
7
Component ID Vertices
4 4,5
Component ID Vertices
2 2,3
6 6,7

12. Streaming Connected Components (union-find)
12
1
34
5
2
6
87
7
8
1
4
3
2
5
4
6
7
Component ID Vertices
4 4,5
7 7,8
Component ID Vertices
2 2,3
6 6,7

13. Streaming Connected Components (union-find)
13
1
34
5
2
6
87
7
8
1
4
3
2
5
4
6
7
Component ID Vertices
4 4,5
7 7,8
Component ID Vertices
2 2,3
6 6,7
1 1,4

14. Streaming Connected Components (union-find)
14
1
34
5
2
6
87
6
8
1
2
3
5
7
8
1
4
1
3
3
2
5
4
6
7
Component ID Vertices
4 4,5
7 7,8
1 1,3
Component ID Vertices
2 2,3
6 6,7
1 1,4

15. Streaming Connected Components (union-find)
15
1
34
5
2
6
87
6
8
1
2
3
5
7
8
1
4
1
3
3
2
5
4
6
7
Component ID Vertices
4 4,5
7 7,8
1 1,3
Component ID Vertices
2 2,3
6 6,7,8
1 1,4

16. Streaming Connected Components (union-find)
16
1
34
5
2
6
87
6
8
1
2
3
5
7
8
1
4
1
3
3
2
5
4
6
7
Component ID Vertices
4 4,5
7 7,8
1 1,2,3
Component ID Vertices
2 2,3
6 6,7,8
1 1,4

17. Streaming Connected Components (union-find)
17
1
34
5
2
6
87
6
8
1
2
3
5
7
8
1
4
1
3
3
2
5
4
6
7
Component ID Vertices
4 4,5
7 7,8
1 1,2,3
Component ID Vertices
2 2,3,5
6 6,7,8
1 1,4

18. Streaming Connected Components (union-find)
18
1
34
5
2
6
87
6
8
1
2
3
5
7
8
1
4
1
3
3
2
5
4
6
7
Component ID Vertices
4 4,5
7 7,8
1 1,2,3
Component ID Vertices
2 2,3,5
6 6,7,8
1 1,4
Component ID Vertices
1 5,4,1,2,3
6 6,7,8

19. Contributions
• A unified comparison framework on Apache Flink
• Classification and experimental comparison
• Performance of partitioning algorithm
• Effect of algorithms on applications
19

20. PARTITIONING ALGORITHMS
20

21. Partitioning Algorithms
21
S
2
W
2
O
2
C
2
N
2
G
2
G
2
SIGOPS,
2007
SIGKDD,
2012
WSDM,
2012
OSDI,
2012
CIKM,
2015
NIPS,
2012
GRADES,
2013
GRADES,
2013

22. Partitioning Algorithms
22
SIGOPS,
2007
SIGKDD,
2012
WSDM,
2012
O
2
C
2
N
2
GRADES,
2013
G
2
SIGOPS,
2007
SIGKDD,
2012
WSDM,
2012
OSDI,
2012
CIKM,
2015
NIPS,
2012
GRADES,
2013
GRADES,
2013

23. Partitioning Algorithms
23
SIGOPS,
2007
SIGKDD,
2012
WSDM,
2012
OSDI,
2012
CIKM,
2015
NIPS,
2012
GRADES,
2013
GRADES,
2013
SIGOPS,
2007
SIGKDD,
2012
WSDM,
2012
OSDI,
2012
CIKM,
2015
NIPS,
2012
GRADES,
2013
GRADES,
2013

24. Partitioning Algorithms
24
O
2
C
2
N
2
G
2
SIGOPS,
2007
SIGKDD,
2012
WSDM,
2012
OSDI,
2012
CIKM,
2015
NIPS,
2012
GRADES,
2013
GRADES,
2013

25. Partitioning Algorithms
25
C
2
G
2
SIGOPS,
2007
SIGKDD,
2012
WSDM,
2012
OSDI,
2012
CIKM,
2015
NIPS,
2012
GRADES,
2013
GRADES,
2013

26. Partitioning Algorithms
26
SIGOPS,
2007
SIGKDD,
2012
WSDM,
2012
OSDI,
2012
CIKM,
2015
NIPS,
2012
GRADES,
2013
GRADES,
2013

27. EXPERIMENTS
27

28. Goals
Our work aims at identifying:
• 1) the benefits of using more complex partitioning methods
• 2) the partitioning overhead for an application
• 3) the effect of partitioning quality on the application performance.
28

29. Datasets and Setup
• On-premises cluster
• A virtualized environment at Amazon consisting of 17x r3.2xlarge
EC2 instances 29

30. RESULTS
30

31. Preview
• Partitioning Performance
• Throughput
• Partitioning Quality
• Cuts and load balance
• Application Performance
31

32. Performance (throughput)
• Vertex partitioning
32

33. Performance (throughput)
Vertex partitioning Edge partitioning
33

34. Performance (throughput)
Vertex partitioning Edge partitioning
34
Hash outperforms others in
terms of throughput

35. Partitioning Quality
• Edge-cuts
35

36. Partitioning Quality
• Edge-cuts
36

37. Partitioning Quality
• Edge-cuts
37
• Vertex-cuts (Replication factor)

38. Partitioning Quality
• Edge-cuts
38
• Vertex-cuts (Replication factor)

39. Partitioning Quality
• Vertex partitioning load balance
39
Dataset Hash Fennel LDG
Twitter 1.0 1.1 1.13
RMAT 1.0 1.1 1.5
Flickr 1.0 1.0 1.0

40. Partitioning Quality
• Vertex partitioning load balance
40
• Edge partitioning load balance
Dataset Hash Fennel LDG
Twitter 1.0 1.1 1.13
RMAT 1.0 1.1 1.5
Flickr 1.0 1.0 1.0
Dataset Hash DBH Greedy Grid HDRF
Friendster 1.0 1.001 1.0 1.0 1.0
Twitter 1.0 1.001 1.0 1.0 1.0
Flickr 1.001 1.002 3.98 1.0 3.98
Hash provides high cuts
but good balance

41. Iterative Application Performance
Input: Twitter graph
41

42. Iterative Application Performance
Input: Twitter graph
42

43. Streaming Application Performance
Bipartiteness Check Connected Components
43

44. Streaming Application Performance
Bipartiteness Check Connected Components
44

45. Conclusion
The trade-off between balancing and reducing
cuts remains
Streaming and iterative applications behave
differently
45
High
Low

46. Open Questions
• Can we design partitioning algorithms with minimal state requirements
for modern stream processing engines?
• Can we adopt existing partitioning algorithms for continuous stream
processing?
• Can we design algorithms with fewer constraints or assumptions about
the input graph?
46

47. THANKS EVERYONE J
Zainab Abbas
zainabab@kth.se
Vasiliki Kalavri
kalavriv@inf.ethz.ch
Paris Carbone
parisc@kth.se
Vladimir Vlassov
vladv@kth.se
47