1. Catch-Up: A Data Aggregation Scheme for VANETs Bo Yu, Jiayu Gong, Cheng-Zhong Xu Dept. of ECE, Wayne State Univ. ACM VANET08

2. Outline Introduction Related Work Motivation Aggregation Scheme Analysis Simulation Conclusion

3. Introduction Traffic Information Dissemination Each vehicle periodically detects the traffic conditions around it, and then, forwards the information to vehicles following behind it Redundant Data & Limited Bandwidth Multiple redundant copies for the same traffic status Consuming a considerable amount of bandwidth

4. Data Aggregation A useful technique to reduce data redundancy and improve communication efficiency Two aspects Routing-related (our focus) How two reports can meet each other at the same time at the same node Data-related Coding, calculation, and compression of aggregatable data r 1 (30mph) r 3 r 1 +r 2 ((30+35)/2=32.5mph) r 2 (35mph) v1v1 v2v2 v3v3

5. Related Work Structured Aggregation A routing structure, forwarding tree, is maintained to ensure reports can be forwarded to the same node at the same time Widely used in sensor networks, but infeasible in VANETs

6. Related Work (Cont.) Structureless Aggregation Randomized Waiting Wait for a random period before forwarding to the next hop During the waiting period, more reports can be received and aggregated Periodical Waiting TrafficView, SOTIS Wait for a fixed period before forwarding to the next hop An arising question – how long should it wait to achieve better aggregation performance?

7. Motivation Two Properties of VANETs Channel Eavesdropping Every node is able to receive reports being transmitted in the channel and log them into its local database Traffic Information is not delay-sensitive Even a delay of tens of seconds is still acceptable

8. Motivation (Cont.) Determine waiting time based on local observations of individual vehicles Challenge: outdated and incomplete knowledge r1r1 v1v1 r2r2 r 1 is ahead, so r 2 should speed up and catch up with r 1 v2v2 v3v3

9. Distributed MDP Model s – world state o – observation b – internal state a – action SE – State Estimator П – Decision Maker

10. Distributed MDP Model (Cont.) action a: WALK, RUN the propagation speed of a report (how fast shall we propagate a report) can be transformed into two different delays before forwarding to the next hop observation o: eavesdropped reports an observation is a tuple internal state b: estimated position of a report b(r,p t ) : the probability that report r is at position p at time t

11. Expected Future Reward Objective To find a policy which maximizes the expected future reward Policy π a sequence of actions to be performed for a given report in the future Expected Future Reward, γ - a future discount factor w t - the expected reward at time t (the saved communication overhead due to aggregation of the reports)

12. Expected Future Reward (Cont.) Virtual Report – r 0 To encourage some reports to speed up, but the others to slow down Is supposed to be always following the current report Can be configured according to the average frequency of the event source Internal States (r 0,r 1,r 2,r 2,…) Total Expected Reward

13. Decision Tree To find the optimal policy π=(a 0,a 1,a 2,…)

14. Other Issues Report Overtaking r1r1 r1r1 r1r1 r2r2 r 1 +r 2 X v1v1 v2v2 v3v3 v4v4

15. Other Issues (Cont.) How to judge whether a report is contained by another aggregated report r 1 r 3 ? where r 3 =r 1 +r 2 Bloom Filter is a space-efficient probabilistic data structure that is used to test whether an element is a member of a set

16. Property All reports from a given road section and from a given time period can be aggregated into an overview report The convergence time upper bound: The convergence distance upper bound:

17. Simulation Based on NS2 and GrooveNet Compared to Randomized Waiting

18. Results CATCHUP(100,1000) - walking speed at 100m/s, running speed at 1000m/s CATCHUP(200,2000) - walking speed at 200m/s, running speed at 2000m/s For CATCHUP, the aggregation operations mainly reside within the first 5 km. For Randomized Waiting, the aggregation operations are distributed all over the propagation distance.

19. Results (Cont.) Before running, scale them to the same total delay For CATCHUP, the delay mainly reside within the first 4 km CATCHUP trades increased delay for reduced communication overhead For Randomized Waiting, the delay is linear

20. Conclusion We studied the adaptive control of forwarding delay in data aggregation in VANETs Aggregation is a tradeoff between delay and communication overhead We make the delay more controllable in a manner that a report has a better chance to be aggregated with other reports

21. THANKS! Bo Yu, Jiayu Gong, Cheng-Zhong Xu Dept. of ECE, Wayne State Univ. ACM VANET08