Properties and Impact of Vicinity in Mobile Opportunistic Networks - T. Phe-Neau

Properties and Impact of Vicinity in Mobile Opportunistic Networks - PhD Defense Slides

Transcript

1. Tiphaine Phe-Neau Properties and Impact of Vicinity in Mobile Opportunistic Networks André-Luc BEYLOT Reviewer Professeur, ENSEEIHT Emmanuel LOCHIN Reviewer Professeur, ISAE Aline CARNEIRO VIANA Examiner Chargée de Recherche HDR, INRIA Vania CONAN Examiner Directeur de recherche, Thales Anne FLADENMULLER Examiner Maître de Conférence, UPMC Vincent GAUTHIER Examiner Maître de Conférence, Télécom SudParis Marcelo DIAS DE AMORIM Advisor Directeur de recherche, CNRS & UPMC
2. 2   People = Network
3. Disruption-tolerant (DTN) or opportunistic networks 3  
4. A DTN scenario 4   t = t1 t = t2 t = t3 A A A B B B C C C D D D
5. A motivating example 5   A B C D E F G nodes in contact with A Intercontact = Contact
6. A motivating example 6   The binary intercontact vision is NOT enough. A B C D E F G nodes in contact with A A B C D E F G nodes in contact unreachable nodes nodes reachable by an end-to-end path
7. Positionning  MANET-DTN mixed approach –  HYMAD from Whitbeck et al. –  Mixed approach by Ott et al. –  Island-hopping in VANET by Sarafijanovic-Djukic et al.  Similar vision in satellite networks –  « A DTN routing scheme for LEO satellites topology » by Diana et al. 7  
8. Motivation Proximity knowledge is helpful A B C D E F G A B C D E F G t0 t1 8  
9. Contributions I.  Uncovering Vicinity Properties of Intercontacts in DTNs II.  Digging into the Vicinity Dynamics of Mobile Opportunistic Networks III.  Predicting Vicinity Dynamics 9  
10. Datasets 10   Infocom05 Rollernet Unimi Random Trip Community Sigcomm09
11. Uncovering Vicinity Properties of Intercontacts in DTNs 1st contribution Tiphaine Phe-Neau, Marcelo Dias de Amorim, and Vania Conan, “Fine-Grained Intercontact Characterization in Disruption-Tolerant Networks,” IEEE ISCC, Kerkyra, Greece, June 2011. Tiphaine Phe-Neau, Marcelo Dias de Amorim, and Vania Conan, “Vicinity-based DTN Characterization,” ACM MobiOpp, Zurich, Switzerland, March 2012. Tiphaine Phe-Neau, Marcelo Dias de Amorim, and Vania Conan, “The Strength of Vicinity Annexation in Opportunistic Networking,” IEEE NetSciCom, Torino, Italy, April 2013.
12. An interesting observation with RT 12   0 0.2 0.4 0.6 0.8 1 contact intercontact Fraction of time
13. 0 0.2 0.4 0.6 0.8 1 1 2 3 4 5 6 7 8 9 ∞ Fraction of time Number of hops An interesting observation with RT 13   The binary assertion limits the information from the network
14. Sociostructure: Infocom05 14   Most of end-to-end transmission possibilities come from 2-hop paths 0 100 200 300 400 500 75000 85000 95000 105000 115000 Number of connected pairs over 1640 Time t (seconds) 4+ 3 2 contact
15. The notion of k-vicinity   Д-vicinity :  The Д-vicinity of node i is the set of all nodes with shortest paths of length at most Д hops from i.   15   i 1-vicinity
16. The notion of k-vicinity   Д-vicinity :  The Д-vicinity of node i is the set of all nodes with shortest paths of length at most Д hops from i.   16   2-vicinity i
17. k-contact and k-intercontact   Д-contact : Two nodes are in Д-contact when they dwell within each other’s Д-vicinity, with Д ˥ N∗.   Д-intercontact : Two nodes are in Д-intercontact while they do not belong to each other’s Д-vicinity.   17   i j j j leaves i’s κ-vicinity (κ-intercontact starts) κ-vicinity (for κ = 2) j returns to i’s κ-vicinity (end of κ-intercontact time) i
18. k-intercontact characterization: Infocom05 18   2+-intercontact follow the same properties as binary intercontact. Same knee point. 0.001 0.01 0.1 1 1 10 100 1000 10000 P[ κ-intercontact > t] Time t (seconds) Contact 2-interc. 3-interc. 4-interc. 5-interc. 6+-interc. Interc.
19. k-contact characterization: Infocom05 19   Logical: larger k values, longer average k-contact intervals. 0.001 0.01 0.1 1 1 10 100 1000 10000 P[ κ-contact > t ] Time t (seconds) Contact Contact 2-contact 2-contact 3-contact 4-contact 5+-contact i
20. k-contact characterization: Infocom05 20   Logical: larger k values, longer average k-contact intervals. 0.001 0.01 0.1 1 1 10 100 1000 10000 P[ κ-contact > t ] Time t (seconds) Contact Contact 2-contact 2-contact 3-contact 4-contact 5+-contact i i
21. k-contact characterization: Unimi 21   Unexpected: larger k values, shorter average k-contact intervals Why?
22. Density-related behavior 22    Two different behaviors –  Dense: more longer k-contact intervals (Infocom05) –  Light: more shorter k-contact intervals (Unimi) dense light
23. The power of vicinity annexation  The WAIT protocol –  Delays transmission until direct contact  Metric –  Waiting delay: the time between message creation and destination detection  Performances –  Generation of 10 messages at random times 23  
24. WAIT protocol 24   i!
25. WAIT protocol + 2-vicinity 25   i!
26. The power of vicinity annexation 26   Considering only 2-contact lowers waiting delays up to 80%. Beyond k=4 less gains. 0 100 200 300 400 500 RT Community Infocom05 Rollernet contact κ = 2 κ = 3 κ = 4 κ = 5 κ = 6 κ = 7 1200 1300 1400 1500 1600 1700 1800 Sigcomm09-d1 2500 3000 3500 4000 4500 5000 5500 6000 Shopping 27500 28000 28500 29000 29500 30000 Unimi 57% 80% 40% 40% 16% 48% 5% *
27. 1st contribution wrap up  The binary assertion covers reality  k-vicinity and k-contact show hidden transmission possibilities  k-contact have similar properties to contact  Using a k={3,4} is enough Take this away 27  
28. Digging into the Vicinity Dynamics of Mobile Opportunistic Networks 2nd Contribution T. Phe-Neau, M. E. M. Campista, M. Dias de Amorim, and V. Conan, “Padrões de Mobilidade de Vizinhança em Redes de Contato Intermitente,” In SBRC, Brasilia, DF, Brazil, May 2013 - Best paper candidate! T. Phe-Neau, M. E. M. Campista, M. Dias de Amorim, and V. Conan, “Examining Vicinity Dynamics in Opportunistic Networks,” In ACM MSWiM, Barcelona, Spain, November 2013. T. Phe-Neau, M. E. M. Campista, M. Dias de Amorim, and V. Conan, “Padrões de Mobilidade de Vizinhança em Redes de Contato Intermitente (extended),” In RB-RESD, accepted.
29. Vicinity Motion Pairwise Markov chain process –  States: k+1 states representing pairwise distances (∞, 1...k) –  Transitional probabilities: probability of 2 nodes being at distance m during step s knowing that they were at distance n before, m ≠ n –  Asynchronous method 0.10 0.44 0.38 0.24 0.16 0.30 0.31 1 2 3 4 ∞ 0.35 0.30 0.25 0.32 0.35 0.42 0.57 29  
30. Vicinity patterns   Birth –  appearance in the κ-vicinity after a period of k-intercontact   Death –  nodes departure from the κ-vicinity   Sequential –  the process of drifting closer or further from each other, when nodes at a distance m move to distance m-1 or m+1   Erratic 30   0.10 0.44 0.38 0.24 0.16 0.30 0.31 1 2 3 4 ∞ 0.35 0.30 0.25 0.32 0.35 0.42 0.57
31. Vicinity patterns: birth 31   Most births happen at states {3,4} not contact 0 20 40 60 80 100 1 2 3 4 5 6 7 % of total Birth State Rollernet Unimi RT Infocom05 Sigcomm09-d1
32. Vicinity patterns: death 32   Death have quite constant rates 0 20 40 60 80 100 1 2 3 4 5 6 7 % of total Current State Rollernet Unimi Sigcomm09-d1 RT Infocom05
33. Vicinity patterns: sequential 33   Sequential movements have similar patterns 0 20 40 60 80 100 1 2 3 4 5 6 7 % of total Current State Rollernet Unimi RT Infocom05 Sigcomm09-d1
34. Vicinity patterns: repartition Most observed movements come from Death and Sequential patterns 0 20 40 60 80 100 1 2 3 4 5 6 7 % of total Distance Death Sequential Erratic 34  
35. Vicinity patterns: take-aways  Birth –  Monitor k-vicinity up to k = {3,4} hops –  Increases arrival detection to 46% to 72%  Death –  Rates remain quite constant with larger k values  Sequential –  A tendency to come closer –  With death, represents up to 90% of observed movements 35  
36. TiGeR: timeline generator 36    Generator of vicinity behaviors: timelines  Why? To bootstrap vicinity knowledge  Mode I (duration friendly), Mode 2 (transition friendly) 1 2 3 4 5 6 7 50 51 52 53 54 Number of hops Time (×103 seconds)
37. 0 0.2 0.4 0.6 0.8 1 ∞ -1 ∞ -2 ∞ -3 ∞ -4 ∞ -5 ∞ -6 ∞ -7 1-∞ 1-2 1-3 1-4 1-5 1-6 1-7 (From - To) States Average VM Mode I Mode II TiGeR: evaluation (1/2) Average transitions do match 37  
38. TiGeR: evaluation (1/2) Average transitions do match 38   0 0.2 0.4 0.6 0.8 1 2-∞ 2-1 2-3 2-4 2-5 2-6 2-7 3-∞ 3-1 3-2 3-4 3-5 3-6 3-7 4-∞ 4-1 4-2 4-3 4-5 4-6 4-7 (From - To) States Average VM Mode I Mode II
39. 0 0.2 0.4 0.6 0.8 1 5-∞ 5-1 5-2 5-3 5-4 5-6 5-7 6-∞ 6-1 6-2 6-3 6-4 6-5 6-7 7-∞ 7-1 7-2 7-3 7-4 7-5 7-6 (From - To) States Average VM Mode I Mode II TiGeR: evaluation (1/2) Average transitions do match 39  
40. 0 0.2 0.4 0.6 0.8 1 ∞-1 ∞-2 ∞-3 ∞-4 ∞-5 ∞-6 ∞-7 1- ∞ 1-2 1-3 2- ∞ 2-1 2-3 2-4 3- ∞ 3-1 3-2 3-4 3-5 4- ∞ 4-1 4-2 4-3 4-5 Transitions (From - To) States (30,36)-transitions (Infocom05) Original Mode I Mode II TiGeR: evaluation (2/2) 40   Pairwise transitions fit as well
41. TiGeR: evaluation (2/2) 41   Pairwise transitions fit as well 0 0.2 0.4 0.6 0.8 1 5- ∞ 5-4 5-6 6- ∞ 6-3 6-5 7- ∞ 7-6 Transitions (From - To) States (30,36)-transitions (Infocom05) Original Mode I Mode II
42. 2nd contribution wrap up  Vicinity Motion models vicinity behavior under a chain process  3 main patterns: birth, death, and sequential  The TiGeR generator provides synthetic vicinity behaviors Take this away 42  
43. Predicting Vicinity Dynamics 3rd contribution A. Tatar, T. Phe-Neau, M. Dias de Amorim, V. Conan, and S. Fdida, “Beyond Contact Predictions in Mobile Opportunistic Networks,” submitted to IFIP/IEEE WONS.
44. Synchronous Vicinity Motion (SVM) Pairwise Markov chain process –  States: k+1 states representing pairwise distances (∞, 1...k) –  Transitional probabilities: probability of 2 nodes being at distance m during step s knowing that they were at distance n before –  Synchronous method every τ seconds 0.006 0.22 0.18 0.09 0.15 0.27 0.23 1 2 3 4 ∞ 0.01 0.01 0.02 0.13 0.19 0.28 0.36 0.96 0.13 44  
45. Vicinity Motion-based Markovian heuristic 45   Presence vector! (Transition matrix)n ! Result vector ! for the ! nth step!
46. Vicinity Motion-based Markovian heuristic 46   [0 1 0] 0,5 0,1 0,4 0,7 0,2 0,1 0,6 0,3 0,1 x! 1 = [0,7 0,2 0,1] Sf! Ss!   Sf = ∞ : first highest probability guess   Ss = 1 : second highest probability guess Nodes in state 1!
47. Evaluation  Is it right? –  nth future steps : n = {1...10} –  Exact Distance (ED): if Sf or Ss = observed value  OK! –  Upper bound Distance (UbD): if Sf or Ss ≥ observed value  OK!  Full vs partial knowledge –  SVM over full experiment –  SVM over half experiment, performance test on the second half 47  
48. Complete knowledge heuristic: Sigcomm09 It is highly efficient because DTN scenarios are mainly disconnected 0.8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 1 1 2 3 4 5 6 7 8 9 10 Proportion of correct guesses nth forward step SVM-full ED UbD 48  
49. Partial knowledge heuristic: Sigcomm09 Light decrease prediction accuracy. Shows how behavior is repetitive. 0.8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 1 1 2 3 4 5 6 7 8 9 10 Proportion of correct guesses nth forward step SVM-half ED UbD 49  
50. 3rd contribution wrap up  Vicinity Motion embeds a prediction scheme  It is highly efficient in DTN because of their disconnected behavior  Useable in the close future Take this away 50  
51. The Vicinity Package for DTN Transversal contribution
52. Vicinity package details  The package embeds: –  Vicinity Motion, TiGeR, heuristics, and various tools  Python, NumPy, and NetworkX  http://vicinity.lip6.fr 52  
53. 53   Take this away
54. 54   Take this away The Vicinity package http://vicinity.lip6.fr Vicinity is important in DTN. k-vicinity up to {3,4} hops is enough. Vicinity Motion. Birth, death, and sequential. TiGeR generator. Predictions with the Vicinity Motion-based Heuristic. Simple with high efficiency.
55. Research directions  Vicinity usage in routing protocols and vehicular networks  Opportunistic networks interoperability  Predictions in DTNs 55  
56. Publications (1/2)   T. Phe-Neau, M. Dias de Amorim and V. Conan, “Caractérisation en diptyque de l’intercontact pour les réseaux à connectivité intermittente,” In Algotel, La Grande Motte, France, May 2012.   T. Phe-Neau, M. E. M. Campista, M. Dias de Amorim, and V. Conan, “Padrões de Mobilidade de Vizinhança em Redes de Contato Intermitente,” In SBRC, Brasilia, DF, Brazil, May 2013. Best paper candidate!   T. Phe-Neau, M. E. M. Campista, M. Dias de Amorim, and V. Conan, “Padrões de Mobilidade de Vizinhança em Redes de Contato Intermitente,” In RB-RESD, accepted. 56  
57. Publications (2/2)   T. Phe-Neau, M. Dias de Amorim and V. Conan, “Fine-Grained Intercontact Characterization in Disruption-Tolerant Networks,” In IEEE ISCC, Kerkyra, Greece, June 2011.   T. Phe-Neau, M. Dias de Amorim and V. Conan, “Vicinity-based DTN Characterization,” In ACM MobiOpp, Zurich, Switzerland, March 2012.   T. Phe-Neau, M. Dias de Amorim and V. Conan, “The Strength of Vicinity Annexation in Opportunistic Networking,” In IEEE NetSciCom, Torino, Italy, April 2013.   T. Phe-Neau, M. E. M. Campista, M. Dias de Amorim, and V. Conan, “Examining Vicinity Dynamics in Opportunistic Networks (poster),” In ACM MSWiM, Barcelona, Spain, November 2013. 57  
58. Under review   T. Phe-Neau, M. E. M. Campista, M. Dias de Amorim, and V. Conan, “Analyzing and Generating Vicinity Traces,” submitted to Elsevier Ad Hoc Networks.   T. Phe-Neau, M. Dias de Amorim, and V. Conan, “Uncovering Vicinity Properties in Disruption-Tolerant Networks,” submitted to Elsevier Computer Networks.   A. Tatar, T. Phe-Neau, M. Dias de Amorim, V. Conan, and S. Fdida, “Beyond Contact Predictions in Mobile Opportunistic Networks,” submitted to IFIP/IEEE WONS. 58  
59. Questions