1. DIMPLE: DynamIc Membership ProtocoL for Epidemic protocols
Jin Sun, Paul Weber, Byung Choi, Roger Kieckhafer
Michigan Technological University
11/15/2018

2. 11/15/2018

3. Epidemic Protocols: Reliable Broadcasting
K. Birman, M. Hayden, O. Ozkasap, Z. Xiao, M. Budiu, Y. Minsky, “Bimodal Multicast,”
ACM Transactions on Computer Systems, 17(2), 41-88, May 1999.
11/15/2018

4. Background Two fundamental assumptions of epidemic protocols
Random selection of next forwarders
Randomly at uniform from the entire membership
Nodes know the network size (N)
Two approaches
Centralized
Distributed
Entire membership at each and every node
Different partial membership at different nodes
11/15/2018

5. Epidemic protocols on P2P systems?
P2P systems can be very large in size, and very dynamic in membership
Difficult to maintain a copy of the entire membership at each node
Inconsistency
Overhead
Partial membership at each node?
11/15/2018

6. Challenges
How to maintain partial membership at each node chosen randomly at uniform from the entire membership?
How to provide different partial membership at different cycles?
How to handle dangling pointers caused by churn?
11/15/2018

7. General Approach
Exchange part of partial membership with part of another partial membership such that, from the node’s perspective, the partial membership is always a random selection of the entire membership at uniform
11/15/2018

8. Shuffle! A well known randomization method in gambling
A practical and easy way of randomization
Is it really random at uniform?
What should happen if not perfectly random?
Network partitioning!
11/15/2018

9. Resilience of Shuffle
Probability of network partitioning is diminishingly small, practically zero!
With reasonable sizes of
Partial membership (O(log(N))
Shuffle length log(N)
Regardless of frequencies
A. Allavena, A. Demers, and J.E. Hopcroft. “Correctness of a gossip based membership protocol” , 24th ACM Symposium on Principles of Distributed Computing (PODC ’05).
11/15/2018

10. Shuffle Properties Convergence
Global randomness, comparable to random networks
Average shortest path length
Clustering coefficient
Regardless of frequencies
S. Voulgaris, D. Gavidia, and M. van Steen. Cyclon: Inexpensive membership management for unstructured p2p overlays. Journal of Network and Systems Management, 13(2):197–217, June 2005.
11/15/2018

11. Shuffle with Churn? Not addressed in the previous work, huh?
Network partition and churn are different!
Measurement study on P2P systems shows:
Average stay time under 10 minutes!
S. Rhea, D. Geels, T. Roscoe, and J. Kubiatowicz, “Handling Churn in a DHT,” USENIX Technical Conference, 2004.
11/15/2018

12. Shuffle with Churn? Found not effective
Produces a large portion of dangling pointers
Would result in poor quality broadcasting
Major reasons:
Time delay to detect dangling pointers
Time delay in join procedures
11/15/2018

13. DIMPLE Improves Shuffle
Reinforcement
At the end of each shuffle
Quick Join
Use a list of visited nodes
11/15/2018

14. DIMPLE-Shuffle works with churn!
11/15/2018

15. DIMPLE-Shuffle works with churn!
11/15/2018

16. DIMPLE-Shuffle works with churn!
11/15/2018

17. DIMPLE-Shuffle works with churn!
11/15/2018

18. DIMPLE: detects dangling pointers fast
11/15/2018

19. In-Degree Distribution
11/15/2018

20. Out-Degree Distribution
11/15/2018

21. DIMPLE: better quality of in-degrees
11/15/2018

22. DIMPLE: better quality of out-degrees
11/15/2018

23. Conclusions Contributions: Future work:
DIMPLE makes shuffle to work with churn
A good practical solution to dynamic membership service (for epidemic protocols)
Future work:
DIMPLE algorithms are improvable
Network size (N) estimation utilizing shuffle is next!
Self-organizing epidemic protocols
11/15/2018