1. Optimisation des DHT à partir des propriétés physiques, logiques et sociologiques des clients Pierre Fraigniaud CNRS LRI, Univ. Paris-Sud http://www.lri.fr/~pierre

2. Plan Distributed Hash Table (DHT) Structural properties Sociological properties Conclusion

3. Principles of DHTs

4. DHT File, data, etc  name Typically: name space = [0,1[ h( file_name ) = 0.10110001101 User  name User name  [0,1[ h( my_IP@ ) = 0.0011010110

5. Correspondence 01 Users = { } user x Data stored by x

6. Overlay network 01 x y z x knows the IP@ of y and z

7. Lookup 01 x h( Andrei Rublev )

8. Node insertion 01 Entry point

9. Examples CAN (D-dimensional meshes) Chord (hypercube) Viceroy (butterfly) D2B, Koorde (de Bruijn) …

10. Structural Properties

11. Desirable properties Small number of hops for lookup: i.e., small diameter and efficient routing Quick updates: i.e., small degree Small congestion: i.e., small probability of contention

12. From the network point of view Taking the inter-node distance in Internet into account! It does not mean that closely related nodes must be close in the Overlay. stretch = max all routes length(Internet route) length(overlay route)

13. Solution Theorem (Abraham & Malkhi) Under some conditions on the physical network,… …there exists an overlay network with strech 1+ε, degree and diameter O(log n).

14. From the user point of view Taking the user interests into account! Closely related users aim at being close in the Overlay. How to measure proximity between users?

15. Requests types Typo: h( André Roublef ) vs. h( Andrei Rublev ) Structure: Prefix search, interval, etc Data-base type requests

16. Sociological Properties

17. Connect users sharing common interets Gnutella enhanced with additional links… Every user keeps links only with users sharing common interest (cf. Maay)

18. Structure of user connections Scale-free structure: Degree distribution = power law Prob( deg(x)=k ) ≈ k -a Guided walk in scale-free graphs Random walk Shortest path Neighbor with largest degree first

19. Rumors and legends Path length Network size Random walk Shortest path Neighbor with highest degree first

20. Using small world properties Milgram’s experiment  six degrees of separation between indivitual Kleinberg’s augmented meshes capture this phenomenon DHT Symphony (!) Why not just doing greedy routing?

21. Conclusion

22. Conclusion: users sociological properties seem to have more impact on DHT’s than network structural properties Unfortunately sociological properties are difficult to model and to measure Warning: this conclusion might be not true in other contexts, e.g., ad hoc, global computing, etc.