1. MANET Supernodes March 16, 2005 Barry Demchak Zhong-Yi Jin William Chang

2. Problem How to create a file system on a MANET that is reliable, energy efficient, and has low latency?

3. MANET Mobile Ad-hoc NETwork Characteristics Wireless Energy constrained Transient nodes Nodes are hosts and routers

4. Related Works Ekta DHT substrate on MANET Not a file system Not energy efficient Pangaea Meta-data/data replication Not on MANET

5. Our Approach Group nodes into Supernodes M M M M M D D D M – Meta Data D – Data Hash(D) = SN3 SN3 SN1 SN4 SN3 SN2 SN5

6. Components Supernode – group of nodes sharing a common set of meta-data Meta-data – information about the locations and the name of data Data – shared file residing at one or more nodes Hash() – consistent hashing value of data name

7. Goals Reliability Low latency Energy efficiency Replication of meta-data and data Node group coverage Meta-data propagation optimization

8. Architecture Packet Delivery Route Discovery Join / Leave Split / Merge Meta-Data Save/Retrieve File Save/ Retrieve Picture Viewer Results Grabber File System Replication Meta-Data Replication Application File System Supernode Level Routing Node Level Routing File List File Delete

9. Project Scope Packet Delivery Route Discovery Join / Leave Split / Merge Meta-Data Save/Retrieve File Save/ Retrieve Picture Viewer Results Grabber File System Replication Meta-Data Replication Application File System Supernode Level Routing Node Level Routing File List File Delete

10. Join

11. File Retrieve D M M M M D

12. File Retrieve (Retro) D M M M M D M M M M M

13. File Save M M M M D

14. Forward to every neighbor Poison list optimization Meta-data Propagation M M M M M M M M M M

15. Experiment Simulated on P2PSim Measure reliability Measure energy / latency Measure poison list optimization

16. P2PSim

17. Reliability Sweet spot at 3-5 nodes per supernode

18. Reliability (cont.) With few nodes per supernode, odd are that supernode will die before data node

19. Reliability (cont.) 3 nodes per supernode seem sufficient for protection of up to 5 file copies

20. Energy Larger supernodes have edges closest to any particular node on network

21. Energy (cont.) Latency and energy drop because of spatial locality due to more file copies

22. Scenario A M

23. Scenario B M M M M

24. Latency Assuming 1400 bytes/packet, large files simply involve more packets

25. Poison List Poison list is important energy optimization – definitely worth space in packet

26. Poison List (cont.) Poison list shorter than number of nodes in supernode causes energies and latencies non-linear with respect to supernode size

27. Poison List (cont.) Supernode update energy is linear with respect to supernode size

28. Routing among Groups Apply DHT (Chord) to MANET Characteristics of wireless network Locality: Shared media, limited range Mobility: Fast node join/leave Limited capability: Distribute Load

29. Connect Group to Ring Join Chord Ring Join the group Request super node’s chord info Super node Child Node Super node Child Node

31. Number of JoinsOurChord 128 Nodes102/120/4218 512 Nodes555/1506/1312075 Reduce total Number of Joins

32. Performance Super node Child Node

33. Conclusion Reliability achieved through replication of meta-data and data Low latency & energy efficiency achieved through node grouping Scalability traded for energy efficiency

34. Future Work Routing layer Merge/Split supernodes File delete/modify File listing More realistic experiments Mix node join and crash Realistic routing latency Realistic energy cost Packet loss

35. Q & A

36. Thank You