1. Multicast Applications: ProbeCast and RelayCast Mario Gerla, Uichin Lee, Soon Oh, SeungHoon Lee CSD, UCLA www.cs.ucla.edu/NRL MURI-DAWN Project review UCSC, Oct 14 2008

2. Progress in 2007-2008 Data dissemination (DTN scenarios) –RelayCast: a scalable DTN multicast protocol (ICNP 2008) –Impact of correlated motion on unicast DTN routing (work in progress) 2 phase inter-contact time distribution: power law head with exponential tail Capacity/delay of DTN unicast routing ProbeCast: multicast admission control (Q2SWINET 2008) –Resource probing + pruning, neighborhood proportional drop (NPROD) for fair share of a channel Network coding configuration/implementation –Communication, disk I/O, encoding overhead analysis (using measurement based models) MobSim: an interactive vehicular motion simulator

3. DTN Multicast Routing Provides reliable data dissemination (e.g., situation awareness data) even in disrupted environments DTN multicast routing strategies –Tree, mesh, ferry/mule, epidemic dissemination –Use mobility-assist routing to deal with disruptions S R2 R3 R1 R4 S R3 R1 R4 R2 S F R1 F S mobility Disrupted node R2 TreeMeshFerry/muleDissemination mobility

4. Scaling Properties of DTN Multicast Questions: –Achievable DTN multicast throughput; average delay –Compare with existing capacity/delay bounds of ad hoc wireless networks (Gupta&Kumar) –Trade-offs: Infinite buffers: throughput/delay trade-offs Finite buffers: throughput/buffer tradeoffs Modeling approach: –Inter-contact time models –Queueing models (for throughput/delay/buffer analysis)

5. DTN Model: Inter-contact Time Common assumption: exponential inter-contact time –Independent and homogeneous Poisson process (e.g., random direction, random waypoint, etc) –Most real world traces also have “ exponential ” tails [Karagiannis07] Inter-contact (meeting) rate ~ speed and radio range product; i.e., λ ~ {speed x radio range} [Groenevelt05] Contact points between two nodes: i and j T ic (1)T ic (2)T ic (3) T ic (n): pair-wise inter-contact time time * Groenevelt05: The message delay in mobile ad hoc networks, Groenevelt et al., Performance’05 § Karagiannis07: Power law and exponential decay of inter contact times between mobile devices, Karagiannis et al., Mobicom’07 T1 T2 T3 T1 T2T3

6. Review: 2 Hop Relay 2-hop relay: 1. Source sends a packet to a relay node 2. Relay node delivers a packet to the corresponding receiver 2-hop Relay by Grossglauser and Tse Source Relay Destination

7. RelayCast: DTN Multicast Routing 2-hop relay based multicast: 1. Source sends a packet to a relay node 2. Relay node delivers the packet to ALL multicast receivers RelayCast: 2-hop relay based multicast Source Relay Destinations D2 D1 D3

8. RelayCast: Throughput Analysis Intuition: average throughput is determined by aggregate encounter rate (src  relay and relay  destinations) –How often does a destination node encounter any of the relay nodes? Answer: n*λ Two-hop relay per node throughput : Θ(nλ) –Aggregate meeting rate at a destination: nλ –Grossglauser and Tse’s results: Θ(nλ)=Θ(1) Recall: λ = 1/n (i.e., speed 1/√n, radio range 1/√n) Source to RelayRelay to Destination Source Destination

9. Two-Hop Relay Review Intuition: average throughput is determined by aggregate encounter rate (src  relay and relay  destinations) –How often does a destination node encounter any of the relay nodes? Answer: n*λ Two-hop relay per node throughput : Θ(nλ) –Aggregate meeting rate at a destination: nλ –Grossglauser and Tse’s results: Θ(nλ)=Θ(1) Recall: λ = 1/n (i.e., speed 1/√n, radio range 1/√n)

10. RelayCast: Throughput Analysis Multicast traffic pattern: –n s sources, each of which is associated with n d random destinations –Different sources may choose the same node as one of their receivers Fraction of sources per receiver : n x = n s n d /n –A source chooses a node as dest with prob. n d /n Fraction of aggregate packets per source = 1/n x RelayCast throughput: Θ(nλ/n x )=Θ(n 2 λ/n s n d ) –i.e. = (#of nodes) x rate x frct of packets per source

11. RelayCast: Delay Analysis One relay node delivers packets to all receivers RelayCast delay: Θ(log n d /λ) –Unlike conventional multicast, delay is proportional to the number of receivers R2 R3 relay node R1 X1 X2 X3 R 3210 3λ2λλ Markov Chain for delivery status: Average delay = average time to absorb = 1/3λ + 1/2λ+1/λ (memoryless!) D=max(X1, X2, X3)

12. Comparison with Previous Results Assumptions; n fixed; r = √logn/n G&K; r=√1/n for 2-hop relay Throughput scaling comparison with n s = Θ(n) –n d : # receivers, n: total # nodes RelayCast is better than conventional multi-hop multicast (r= √logn/n) Gupta & Kumar, TOIT’00 Shakkottai et al., Mobihoc’07 Li et al., Mobicom’07 Tavli, IEEE Com. Letter’06 Keshavarz-Haddad et al., Mobicom’06 Grossglauser & Tse, INFOCOM’01 Delay Tolerant Apps RelayCast: Delay Tolerant Apps Number of m-cast receivers per source Per node throughput with n s = Θ(n)

13. Simulation Results Comparison with Conventional Multicast Protocol –Connected topology RelayCast is scalable; ODMRP’s throughput decreases significantly, as # sources increases * QualNet v3.9.5 * Mobility: random waypoint (speed = 20, 30m/s) * Network area size: 1000m*1000m * 100 Nodes, 250m TX range 5 destinations

14. Two-phase Inter-contact Time Two-phase distribution: power-law head and exponential tail Chaintreau06 Karagiannis MobiCom 07 Infocom 06 Levy walk: Rhee Infocom 08 Association times with AP (UCSD) or cell tower (MIT cell) Direct contact traces: Infocom, cambridge (imotes), MIT-bt

15. Two-phase Inter-contact Time Why two-phase distribution? –One possible cause: flight distance of each random trip [Cai08] –The shorter the flight distance, the higher the correlation  heavier power tail Examples of correlations: –Manhattan sightseers: In Time Square, sightseers tend to bump into each other; and then depart for other sights Levy flight of human walks [Rhee08]: short flights + occasional long flights –Vehicular mobility: Constrained by road traffic (+traffic jam) High correlation among vehicles in close proximity After leaving locality, vehicles meet like “ships in the night” Power-law head while in the local contention area, vs. exponential tail for future encounters *Cai08: Han Cai and Do Young Eun, Toward Stochastic Anatomy of Inter-meeting Time Distribution under General Mobility Models, MobiHoc’08 *Rhee08: Injong Rhee, Minsu Shin, Seongik Hong, Kyunghan Lee and Song Chong, On the Levy-walk Nature of Human Mobility, INFOCOM’08

16. Two-hop Relay Unicast under Correlated Motion Patterns Impact of correlated motion patterns on throughput/delay performance? –Under the average flight distance of Ω(r); i.e., minimum travel distance ~ one’s radio range –Increase correlation by decreasing flight distance Preliminary analytic results : –Throughput: Independent of node speed and degree of correlation (ie, flight distance) –Average delay is within [1/λ, logn/λ]; i.e., random direction (to wall) and random walk respectively –Delay monotonically increases with the degree of correlation –Buffer requirement also increases Using Little’s results: [Θ(nr/v), Θ(nrlogn/v)] Simulation results: –Correlation increases burstiness of traffic in and out of relays

17. Simulation: Throughput Degree of correlation via average flight distance L –5000m*5000m area –L=R=250m  high correlation  power law head + exponential tail –L=1000m  low correlation  almost exponential Throughput is independent of the degree of correlations Average throughput per node as a function of # relay nodes CCDF of inter-contact time (20m/s) (Log-log plot) L=250m log-linear plot L=1000m L=250m Exp

18. Simulation: Inter-any-contact Time Invariance property (Cai08): average inter-contact time is independent of correlation Residual inter-contact time: –Source sends a packet to a relay node; i.e., like “random probing” the inter-meeting time between a relay and destination  residual time –Average residual inter-contact time is proportional to the second moment of inter-contact time Average Inter-contact time Average residual inter-contact time 20m/s 30m/s 20m/s L=250m 20m/s L=1000m 20m/s L=250m 30m/s L=1000m

19. Simulation: Buffer Utilization Burstiness increases with the degree of correlation Cumulative distribution of the number of consecutive encounters Buffer utilization over time (speed=30m/s)

20. Summary: DTN Routing under Correlated Motion Patterns Per-node throughput is not affected by the degree of correlation However, correlation causes increases in: –Variance in the inter-contact time –Average delay –Buffer requirements –Burstiness of inbound/outbound

21. ProbeCast S. Oh, G. Marfia, M. Gerla, Q2SWINET 2008 The Problem: Resource reservation/allocation schemes are ineffective in inelastic multicast in ad hoc nets –Bookkeeping is very cumbersome (as # of destinations increases); –Also, mobility requires continuous re- adjustments –Without QoS support, quality will collapse Flow 1 has 9 receivers with 200Kbps and flow 2 has 3 receivers with 40Kbps Goal: Achieving reliable QoS support in inelastic multicast flows (e.g., video and audio stream)

22. ProbeCast: key insights Insight #1: Resource Probing –No a priori resource allocation –Rather “probe” for resources Insight #2: Pruning via Back-pressure –Back-pressure (“prune”) toward the source when resource is unavailable –Re-route or reject the inelastic flow Insight #3: Neighborhood Proportional Drop (NPROD) –Local rate balancing using proportional dropping –Enforces fair channel sharing  “fair back-pressure” Main Outcome: –Inelastic flows to acquire resources in fair manner without reservation, yet preserving reliable QoS

23. ProbeCast Approach Assumptions: –End-to-End FEC – e.g. erasure coding – always ON –Each flow has packet drop threshold Probing –Each node measures resource overload – e.g. packet drop rate –Broadcast to one hop neighbors own drop rate via piggybacking on packets Proportional Drop (N-PROD) –Overhearing neighbors’ drop rates –Enforcing equal drop rates among flows competing in the same contention domain – packet drop –Nodes in the same contention domain sharing channel fairly

24. ProbeCast Approach (Cont.) Pruning –Drop-Threshold (DT) for flows traffic class and flow age dependent –Piggybacking DT on the packet  Forwarders know Drop Threshold of flows Typically, incoming flow has lower threshold than incumbent When drop rate is > threshold, a flow is back-pressured  no explicit control packets to source –Source action: re-route if there is alternative route; otherwise reject the flow

25. Probe/Prune + N-PROD at Work (A) 3 flows in the same contention domain. Lower graphs shows packet delivery ratios, presented by percentages. (B) Flow 3 starts transmitting and other flows’ rates decrease (N-PROD). (C) Since flow 3 drop rate exceeds the threshold, backpressure starts.

26. Simulation - Fairness Qualnet Simulation: 50 nodes uniformly distributed 1000 mby 1000 m field Flow 1 has 9 receivers with 200Kbps and Flow 2 has 3 receivers with 40Kbps N-PROD eliminates capture problem increasing FAIRNESS

27. Simulation – Pruning Qualnet Simulation: 50 nodes uniformly distributed 1000 mby 1000 m field The number of packet sent by the source in 3 inelastic flows case One flow is rejected at 40s by pruning Two flows can satisfy QoS requirements

28. NC Implementation Guidelines S. Lee, U. Lee, K-W. Lee, M. Gerla SECON 2008 Goal: show that NC can be implemented in military scenarios –Develop configuration guidelines based on measured data We start with Network Coding processing O/H analysis –Linearly proportional to the number of packets in a generation (= generation size) –Generation size must be carefully chosen: max node encoding rate > (available) wireless bandwidth + X1X1 X2X2 X3X3 e1e1 e2e2 e3e3 e 1 X 1 +e 2 X 2 +e 3 X 3 [e 1,e 2, e 3 ] generation

29. NC Throughput Measurement Validation through measurements using portable devices Nokia N800: TI OMAP 2420 (330Mhz) + 802.11b Orinoco 8471 WD IBM Thinkpad R52 Scenario: k contenders in domain (k=1/2/3) N1 N2 N3

30. NC Throughput Measurement large generation => high CPU O/H => low pkt tx As the number of contenders increases, pkt tx rate must decrease  can support a larger generation size For small unit operations, optimal Gen Size < 50 (from experiments) –Well suited for network coding based streaming (i.e., CodeCast) G10 = 10 packets in generation N/A: No network coding Number of contenders in a domain

31. MobSim: An Interactive Simulator for Urban Mobility C. Li, M. Bansal, U. Lee, K.-W. Lee, M. Gerla ACITA Demo Session Limitations of current simulators –Non-realistic urban mobility models –Non-interactive simulations MobSim design goals: –Programmable mobility model –Interactive simulation environment –Built-in appl modules (eg dissemination)

32. MobSim Architecture Mobility Generator: –Tiger map of target urban area –Underlying vehicle motion pattern (eg, commuting, shopping, etc) Application: –E.g., Data Dissemination Processing Module Target selection; Agent vehicles, etc Real-time Visualization Module Interactive Simulation UI Mobility Generator: Tiger map + IDM Data Dissemination Processing Module Applications Real-time Visualization Module Interactive Simulation UI MobSim

33. Road-constrained Motion Model For each car, pick random start/end points and speed Construct the shortest path Travel at variable speed on each segment Start End

34. Agent Tracks Target using Last Encounter Routing Agent moves in direction hinted by cars that last encountered the target While moving, agent continuously looks for fresher encounter information Harvest Move to Last Encounter Point

35. MobSim: Simulation Results Average search time with varying number of agents and number of nodes MobSim Screenshot

36. Future Work Impact of different vehicle and agent motion patterns Impact of density (e.g., intermittent connectivity) Bio-inspired multiple agent collaboration algorithm (i.e., Lévy jump based searching + datataxis) Investigate realistic urban warfare scenario (e.g., hints about enemy movements)