1. 1 Routing and Performance Evaluation of Disruption Tolerant Networks Mouhamad IBRAHIM Ph.D. defense Advisor: Philippe Nain INRIA Sophia Antipolis 14 November, 2008

2. 2 Thesis outline Part I: Design and performance evaluation of routing protocols for disruption tolerant networks Part II: Design and performance evaluation of medium access control protocol for IEEE 802.11 standard

3. 3 Routing in mobile ad hoc networks  Mobile Ad Hoc Networks (MANETs)  No fixed infrastructure  Nodes communicate in a peer to peer mode with other nodes  Nodes work as routers: Store-Forward  Routing in MANETs: Main assumption  Existence of end-to-end paths between Source- Destination pairs

4. 4 Routing challenges in MANETs  Instability of wireless paths: node mobility, low node density, interferences,…  Does not help to establish and maintain routes  Appearance of Disruption/Delay Tolerant Networks (DTNs): disconnected mobile networks  Often there is no end-to-end path among Source- Destination pairs  take advantage of node mobility to perform routing Store-Carry-Forward

5. 5 Store-Carry-Forward: how does it work? S V1V1 V3V3 D V2V2 R

6. 6 Routing approaches for DTNs  Classification based on the degree of knowledge that nodes have about their future contact opportunities  Four classes of routing techniques:  Scheduled-contact based routing  Controlled-contact based routing  Predicted-contact based routing  Opportunistic-contact based routing

7. 7 Opportunistic-contact based routing  Flooding mechanism  Epidemic routing protocol  Limit the number of hops  Multicopy Two-hop Relay protocol  Limit the number of copies  Spray-and-Wait protocol  Question: To what extent we can push the performance if we increase number of contact opportunities: Throwboxes

8. 8 Throwboxes (1)  Throwboxes are fixed relays with better storage and energy capabilities  Battery powered for short term use or solar panel for long term use Photos are taken from http://prisms.cs.umass.edu/dome/

9. 9 Throwbox (2)  Operate in Store-Forward paradigm  Promising approach to route messages in DTNs  Adding one throwbox on UMass DieselNet improves packet delivery by 37% and reduces message delivery delay by 10% [1]  Research still in its early stage!! Part I: Evaluate and design routing techniques for opportunistic DTNs augmented by throwboxes [1] N. Banerjee et al. An energy-efficient architecture for DTN throwboxes. Infocom 2007.

10. 10 Opportunistic DTNs: Inter-meeting times  Characteristic of inter-meeting times among nodes  Random mobility:  Inter-meeting times mobile/mobile have shown to follow an exponential distribution [Groenevelt et al.: The message delay in mobile ad hoc networks. Performance Evaluation, 2005]  Human mobility:  Inter-meeting times mobile/mobile have shown to follow power law distribution [Chaintreau et al.: Impact of human mobility on the design of opportunistic forwarding algorithms. Infocom, 2006]

11. 11 Opportunistic-contact: Random mobility X1X1 X2X2 V1V1 V2V2  Directions (α i ) are uniformly distributed (0, 2π) ÄSpeeds (V i ) are uniformly distributed (V min,V max ) ÄTravel times (T i ) are exponentially /generally distributed  Directions (α i ) are uniformly distributed (0, 2π) ÄSpeeds (V i ) are uniformly distributed (V min,V max ) ÄTravel times (T i ) are exponentially /generally distributed R T 1, V 1 T 2, V 2 R Ä Next positions (X i )s are uniformly distributed  Speeds (V i )s are uniformly distributed (V min,V max ) Ä Next positions (X i )s are uniformly distributed  Speeds (V i )s are uniformly distributed (V min,V max ) α1α1 α2α2 Random Waypoint model (RWP) Random Direction model (RD)

12. 12 Mobile/box inter-meeting times CCDF on a linear-log scale: log(Pr(τ > x)) = log(e - μ x )= - μ x Simulation N = 1 Exponential –μx Simulation N = 5 Exponential –5μx Simulation N = 10 Exponential –10μx Simulation N = 1 Exponential –μx Simulation N = 5 Exponential –5μx Simulation N = 10 Exponential –10μx

13. 13 Parameter μ (1)  Stationary probability to find the mobile within neighborhood of a box f(.,.)  stationary spatial pdf of the mobility model  Using Renewal theory, we have Contact time C1C2C3 Time τ1τ2τ3

14. 14 Parameter μ (2)  Unconditioning on throwbox location within the network area L x L  Case of Random Direction model: mobile nodes are uniformly distributed [1]  and hence independent of throwboxes pdf distribution!! pdf of throwboxes distribution Stationary pdf of location for mobility model [1] P. Nain et al. Properties of random direction models. Infocom 2005.

15. 15 Parameter μ (3)  Case of Random Waypoint model: mobile nodes are distributed around the center [3]  μ depends on throwboxes spatial distribution Throwboxes uniformly distributed Throwboxes generally distributed, e.g. [3] J.-Y. Le Boudec and M. Vojnovic. Perfect simulation and stationarity of a class of mobility models, Infocom 2005.

16. 16 Performance evaluation of relaying protocols in DTNs with throwboxes  Epidemic routing protocol (ER)  Multicopy two-hop relay protocol (MTR)

17. 17 S V1V1 V3V3 D V2V2 R B1B2 Epidemic Routing  flooding protocol Epidemic routing protocol

18. 18 S V1V1 V3V3 D V2V2 R B1B2 Copies make at MAX two hops between Source/Destination Multicopy two-hop protocol (MTR)

19. 19 S V1V1 V3V3 D V2V2 R B1B2 Network model Source node Destination node N-1 mobile relay nodes M throwboxes Mobile/mobile: Exponential with λ [4] Mobile/box: Exponential with μ [4] R. Groenevelt, P. Nain, and G. Koole. The message delay in mobile ad hoc networks. Performance Evaluation, 2005.

20. 20 Metrics of interest  Distribution and mean value of  Delivery delay T  user side  Total number of generated copies G when one packet is to be send from source to destination  network operator side

21. 21 Markov analysis  Two-dimensional continuous time absorbing Markov chain I(t) = (R(t),B(t)) as follows:  For t < T:  R(t) {1,2,…,N}  number of mobile nodes holding a copy of the packet (source included)  B(t) {0,1,2,…,M}  number of throwboxes holding a copy of the packet (assumed fully disconnected)  For t > T, I(t)= {a}  absorbing state, i.e. when destination receives the packet

22. 22 MTR protocol: Delivery delay (1) Approach to solve: Stochastic analysis  Delivery delay T MTR is the minimum of N + M mutually independent R.V.s T MTR = (D SD, D r 1, D r 2,…, D r N-1, D B 1,…, D B M )  Hence distribution of T MTR reads as source  destination: exponential with rate λ source  relay  destination: sum of two exponentials with rate λ source  throwbox  destination: sum of two exponentials with rate μ

23. 23 MTR protocol: Delivery delay (2) and mean of T MTR reads as  Using fluid model, we obtained also asymptotic expression for E[T MTR ] when N or M go large

24. 24 MTR protocol: # of generated copies  Define Pr a (n,m) as probability that last visited state before absorption is state (n,m)  Pr a (n,m) is sum of probabilities of different paths joining state (1, 0) to state (n,m)  These probabilities are all equal. Their total number is  The probability distribution of G MTR reads as

25. 25 Epidemic protocol: Delivery delay Approach to solve: Theory of absorbing Markov chain  Delivery delay T ER represents time to absorption Q = infinitesimal generator of Markov chain  M = transition matrix among non-absorbing states m * (i,j) is the (i,j) th entry of M -1

26. 26 Epidemic: # of generated copies  Define Pr a (n,m) as probability that last visited state before absorption is state (n,m)  Case of epidemic protocol: transition rates are state dependent  approach reported by [Gaver et al.: Finite Birth-And-Death Models in Randomly Changing Environments, 1984]  The probability distribution of G ER follows then

27. 27 Case of connected Throwboxes  Underlying assumption: Pass a copy to one throwbox to let all the others infected  Same expressions hold by substituting M  1 μ  M μ

28. 28 Model validation: Delivery delay Epidemic protocol RWP model Epidemic protocol RWP model Throwboxes disconnected and uniformly distributed Throwboxes disconnected and RWP stationary distributed Throwboxes connected and uniformly distributed Throwboxes connected and RWP stationary distributed Throwboxes disconnected and uniformly distributed Throwboxes disconnected and RWP stationary distributed Throwboxes connected and uniformly distributed Throwboxes connected and RWP stationary distributed

29. 29 Model validation: Delivery delay MTR protocol RWP model MTR protocol RWP model Throwboxes disconnected and uniformly distributed Throwboxes disconnected and RWP stationary distributed Throwboxes connected and uniformly distributed Throwboxes connected and RWP stationary distributed Throwboxes disconnected and uniformly distributed Throwboxes disconnected and RWP stationary distributed Throwboxes connected and uniformly distributed Throwboxes connected and RWP stationary distributed

30. 30 Performance evaluation framework for throwboxes-augmented DTNs Objective: Framework to evaluate and analyze performance of various routing strategies for DTNs extended with throwboxes

31. 31 Proposed five routing strategies (1)  Main idea: define possible message forwarding interactions among the Source, Mobile relays, Throwboxes and the Destination  Ultimate goal: exploit throwboxes presence to minimize copies generations at mobile nodes

32. 32 Proposed five routing strategies (2)  Common forwarding interactions: Source  Relay Relay  Destination Source  Relay Relay  Throwbox Relay  Destination Strategy VStrategy IVStrategy IIIStrategy IIStrategy I Infected throwbox Mobile relay Destination Particular interactions for each strategy Infected mobile relay Mobile relay Source Throwbox Destination

33. 33 Metrics of interest Under a given routing strategy s: 1- Mean delivery delay between a Source/Destination E[T s ]  Mean number of valuable transmissions E[G s ], i.e. those made only by mobile nodes plus the source 2- Mean number of mobile relays infected by the source, I s 3- Mean number of infected throwboxes, K s 4- Proba. Source delivers message to destination, PrS s 5- Proba. Mobile relay delivers message to destination, PrR s

34. 34 Modeling framework (1)  Three-dimensional continuous time absorbing Markov chain A s (t) = (I s (t), J s (t), K s (t)) as follows:  For t < T s, A s (t) = (I s (t), J s (t), K s (t)): I s (t)  Number of mobile nodes infected by the source J s (t)  Number of mobile nodes infected by the throwboxes K s (t)  Number of infected throwboxes  For t > T s : A s (t) = {a}  absorbing state

35. 35 Modeling framework (2) i,j,k i,j+1,k i+1,j,k i,j,k+1 a α (i,j,k) β (i,j,k) θ (i,j,k) γ (i,j,k) S (i,j,k) F s (i,j,k) is mean value of metric F s till absorption starting from (i,j,k) Mean value of metric F s at (i,j,k) 1- Mean sojourn time T s  2- Mean number of mobile relays I s  3- Mean number of throwboxes K s  4- Proba. delivery by source PrS s  5- Proba. delivery by relay PrR s 

36. 36 Modeling framework (3)  Values of F s are known at last states  only one possible transition to state {a}, e.g.  Iterating recursive equation till initial state (1,0,0): N,0,M a θ (N,0,M) Known!

37. 37 Modeling framework (4)  To compute E[T s ] and G[T s ] under a given strategy  Define corresponding state space E s and infinitesimal generator Q s (t)

38. 38 Framework validation N = MMetricAnalyticalSimulationRel. error % 10 T(1,0,0)8.15 10 3 7.95 10 3 2.54 I(1,0,0)3.213.271.75 K(1,0,0)1.481.377.83 PrS(1,0,0)0.270.284.21 PrR(1,0,0)0.550.561.15 100 T(1,0,0)2.30 10 3 2.6 10 3 2.75 I(1,0,0)8.358.461.29 K(1,0,0)4.344.615.81 PrS(1,0,0)0.070.086.48 PrR(1,0,0)0.790.783.0 Strategy II: Analytical versus simulation results

39. 39 Comparing E[T] and E[G] with respect to Epidemic protocol Strategy II Strategy IV Strategy V Strategy II Strategy IV Strategy V

40. 40 Diameter of epidemic protocol Context: Opportunistic DTNs running epidemic protocol WITHOUT throwboxes Objective: Examine the mean length of forwarding path

41. 41 Diameter of epidemic protocol  Instance of epidemic tree:  X S,D denote number of intermediate hops between S and D  Aim is to compute E[X S,D ]: diameter of epidemic protocol R5 S R2R1 D R3 R4

42. 42 Diameter computation (1) Approach to solve: Theory of recursive tree  Recursive tree is like any tree on a graph, however, nodes are labeled with their joining instants to the tree Example: recursive tree of order 4 1 2 34 2 4 3 1 2 3 4 1 234 1 2 3 4 1 3 2 4 1 E[X i,j ] is known for random tree

43. 43 Diameter computation (2)  Conditioning on possible labels of the destination among the N nodes  Look to the impact of limiting number of forwarding hops on relaying performance  Using the framework, we analyze different dissemination algorithm with limited number of hops

44. 44 Epidemic protocol: Limiting # of hops Max. hop = 2 Max. hop = 3 Max. hop = 4 Max. hop = 5 Max. hop = 2 Max. hop = 3 Max. hop = 4 Max. hop = 5

45. 45 Part II

46. 46 Adaptive Backoff Algorithm for IEEE 802.11  Motivation: IEEE 802.11 performs poorly in congested network  Following a successful transmission, source station chooses backoff duration randomly in {0,…,CW 0 }  Objectives:  Adaptive algorithm aware of active stations  Maximize system throughput and minimize end-to- end delay Inadequate for large networks

47. 47 How to transmit at optimal transmission probability τ*  Bianchi model [5] :  Transmission probability  Our idea: m= log(CW max /CW 0 ) [5] G. Bianchi. Performance analysis of the IEEE 802.11 distributed coordination function. JSAC 2000.

48. 48 Estimating # of active stations  Active stations are decoding all transmitted packets on the channel  identify emitting stations  Stations counts signs of life coming from others stations  signs of life: error free data and RTS packets  Measured during virtual transmission times  Samples used as input to a corrected WMA filter Ň k : sample at k th period CW k : window at k th period α, β : correcting factors

49. 49 Algorithm performance Adaptive Standard Adaptive Standard Group entrance Group departure Group entrance Group departure

50. 50 Conclusions (1)  Accurate approximation for meeting rate between a mobile/throwbox:  For two common mobility model  For general throwboxes spatial distribution  Explicit expressions for the distribution and the mean of delivery delay and number of generated copies  Under epidemic and MTR protocols  Asymptotic expressions for these means under MTR

51. 51 Conclusions (2)  Proposed various routing strategies for DTNs augmented with throwboxes  Markovian framework to evaluate performance of various routing strategies  Can be extended to evaluate other performance metrics and routing techniques  Explicit expression for the diameter of forwarding path under epidemic protocol

52. 52 Conclusions (3)  Proposed an efficient MAC protocol for IEEE 802.11  Adapt starting value of contention window to network size  Original mechanism to estimate number of active stations

53. 53 Future research direction  Analyze correlation and heterogeneous movement patterns in real mobility traces  Elaborate corresponding mobility models and evaluate proposed routing strategies over them  e.g. markovian model for community based mobility, bus mobility  Analyze impact of different buffer management techniques on routing under heterogeneous mobility model

54. 54 The end … Thank you!

55. 55 Publications  M. Ibrahim, A. Al Hanbali, P. Nain, "Delay and Resource Analysis in MANETs in Presence of Throwboxes", Performance Evaluation, Vol. 64, Issues 9-12, P. 933-947, October 2007.  Al Hanbali, M. Ibrahim, V. Simon, E. Varga, I. Carreras "A Survey of Message Diffusion Protocols in Mobile Ad Hoc Networks", Inter-Perf 2008, Athens, Greece, Octobre 2008.  M. Ibrahim, S. Alouf, "Design and Analysis of an Adaptive Backoff Algorithm for IEEE 802.11 DCF mechanism", Networking 2006, Coimbra, Portugal, Mai 2006.  Under submission: M. Ibrahim, P. Nain, I. Carreras. "On routing trade-offs in throwbox-embedded DTN networks".

56. 56

57. 57 ZebraNet: mobility based routing  Objective  track zebras in wildlife  Collars attached to zebras  Base stations move sporadically to collect data

58. 58 Model validation: Delivery delay Throwboxes disconnected and uniformly distributed Throwboxes disconnected and RWP stationary distributed Throwboxes connected and uniformly distributed Throwboxes connected and RWP stationary distributed Throwboxes disconnected and uniformly distributed Throwboxes disconnected and RWP stationary distributed Throwboxes connected and uniformly distributed Throwboxes connected and RWP stationary distributed MTR Epidemic

59. 59 Virtual transmission time  Virtual transmission time = time separating two successful random transmissions

60. 60 Mean wasted time  E[wasted_time] = E[collision_time] + E[idle_time]  E[collision_time] = f( τ,N ) w ith τ for fixed N  E[idle_time] = g( τ,N ) with τ for fixed N N = 50 N = 20 N = 10 N = 50 N = 20 N = 10 Bianchi model