1. Maximizing the Contact Opportunity for Vehicular Internet Access Authors: Zizhan Zheng †, Zhixue Lu †, Prasun Sinha †, and Santosh Kumar § † The Ohio State University, § University of Memphis INFOCOM 2010, San Diego, CA 1 9/18/2015 TexPoint fonts used in EMF. Read the TexPoint manual before you delete this box.: AAAA A

2. Outline  Motivation  Three Metrics  Contact Opportunity in Distance  Contact Opportunity in Time  Average Throughput  Evaluations  Summary and Future Work 2

3. Motivation: Internet Access for Mobile Vehicles 3  Applications  Infotainment  Cargo tracking  Burglar tracking  Road surface monitoring  Current Approaches  Full Coverage  Opportunistic Service  Sparse Coverage

4. Current Approach I (of III): Full Coverage 4  Wireless Wide-Area Networking  3G Cellular Network  3GPP LTE (Long Term Evolution)  WiMAX Either long range coverage (30 miles) or high data rates (75 Mbps per 20 MHz channel) 3 Mbps downlink bandwidth reported in one of the first deployments in US (Baltimore, MD)  Google WiFi for Mountain View  12 square miles, 5 00 + APs, 95% coverage  1 Mbps upload and download rate  Not very practical for large scale deployment due to the prohibitive cost of deployment and management Google Wifi Coverage Map http://wifi.google.com/city/mv/apmap.html

5. Current Approach II (of III): Opportunistic Service via In-Situ APs 5  Prototype  Drive-Thru Internet (Infocom’04,05)  In-Situ Evaluation  DieselNet (Sigcomm’08, Mobicom’08) Interactive WiFi connectivity (Sigcomm’08) Cost-performance trade-offs of three infrastructure enhancement alternatives (Mobicom’08)  MobiSteer (Mobisys’07) Handoff optimization for a single mobile user in the context of directional antenna and beam steering  Cabernet (Mobicom’08) Fast connection setup (QuickWiFi) and end-to-end throughput improvement (CTP)  Problems  Opportunistic service, no guarantee  Unpredictable interconnection gap Internet AP

6. Current Approach III (of III): Sparse Coverage with Performance Guarantees 6  Basic Idea  Planned deployment  Sparse coverage with performance guarantees  Alpha Coverage (Infocom ’09 mini)  Placing an upper bound on the maximum diameter of coverage holes in a road network  Pure geometric  Does not correspond to the quality of data service directly

7. Contact Opportunity: A More Expressive Sparse Coverage Mode 7  Contact Opportunity – fractional distance/time within range of APs  Closer to user experience  Can be translated to average throughput if all uncertainties resolved  Our Approach  Worst Case perspective  Start with distance measure that involves least uncertainties  Extend to time measure by modeling road traffic  Further extend to average throughput by also modeling data rates, user density, and association

8. Contributions 8  Propose Contact Opportunity, an expressive sparse coverage mode.  Propose efficient solutions with provable performance bounds to maximize the worst-case Contact Opportunity with various uncertainties considered.  Develop the foundations towards providing scalable data service to disconnection-tolerant mobile users with guaranteed performance.

9. Outline 9  Motivation  Three Metrics  Contact Opportunity in Distance  Contact Opportunity in Time  Average Throughput  Evaluations  Summary and Future Work

10. Models and Assumptions 10  Road Network  An undirected graph G  Assumption 1: A set of candidate deployment locations is given, denoted as A.  Mobile Trace  A set of paths on G  Assumption 2: A set of frequently traveled paths is known, denoted as P.  AP Coverage  Geometric model is used  Assumption 3: The covered region for each candidate location is known (but not necessary a disk).

11. Contact Opportunity in Distance 11  For a subset S µ A, a path p 2 P, the Contact Opportunity in Distance of p :  - the cost of S 200m 1000m

12. The Properties of Set Function ´ d 12  The set function ´ d (, p ) : 2 A ! [0,1] is  Normalized: ´ d ( ;, p ) = 0  Nondecreasing: ´ d ( S, p ) · ´ d ( T, p ) if S µ T  Submodular: adding a new AP to a small set helps more than adding it to a large set

13. Submodular Set Function 13  A set function F : 2 A ! R is submodular if for all S µ T µ A and a 2 An T, F ( S [ { a }) – F ( S ) ¸ F ( T [ { a }) – F ( T )  Discrete counterpart of convexity  Example: F ( S ) = ´ d ( S, p ) S T a a

14. Approximation Algorithm (for a relaxed version)  Hard to approximate directly  An instance of budgeted submodular set covering problem  No polynomial time approximation unless P = NP  Relaxing the budget B - a binary search based algorithm  For a given ¸ 2 [0,1], solve the subproblem - find a deployment S of minimum cost that provides worst-case Contact Opportunity of ; An instance of submodular set covering problem A greedy algorithm has a logarithmic factor (L.A. Wolsey 1982)  If w ( S ) > B, a lower ¸ is used; otherwise, a higher ¸ is used;  Repeat until no higher ¸ can be achieved; output ¸  OPT( B ) achieved if ² B is allowed (Andreas Krause 2008)  OPT( B ) - max-min Contact Opportunity of an optimal solution  ² - a logarithmic function of problem parameters 14

15. Contact Opportunity in Time 15  For a subset S µ A, a path p 2 P, the Contact Opportunity in Time of p :  Challenge - uncertain contact time and travel time  Traffic jams, accidents, stop signs, etc.  Solution  Worst-Case perspective  Interval based modeling - for each road segment, an interval of possible travel times is known. 200m 1000m 20s10s 20s

16. Contact Opportunity in Time (Cont.) 16  A traffic scenario k - an assignment of travel time (any value from the interval) to each road segment  k S - the worst traffic scenario  Unfortunately, ´ t ( S, p, k S ) 8 S µ A is not submodular  Approximation by the “mean” scenario  “mean” scenario assigns the average travel time to each road segment   - an upper bound on the ratios of maximum and minimum travel times for all road segments  Factor  achieved by using “mean” scenario

17. From Contact Opportunity to Average Throughput 17  More Assumptions  Each candidate location a 2 A is associated with a worst case data rate r a  The maximum number of users moving on each road segment is known The maximum number of users in the range of an AP at a 2 A can be computed, denoted as v a  A user always selects the AP with the highest normalized rate ( r a / v a ) in range to associate  Handoff time is small enough to be ignored

18. From Contact Opportunity to Average Throughput (Cont.) 18  For a subset S µ A, a path p 2 P, the Average Throughput when moving through p can be estimated as:  Solution similar to “Contact Opportunity in Time”  Limitations  Simplified association protocol  Fairness has been ignored r a = 1 Mbps 200m 1000m 20s10s 20s 223

19. Outline 19  Motivation  Three Metrics  Contact Opportunity in Distance  Contact Opportunity in Time  Average Throughput  Evaluations  Summary and Future Work

20. Simulations 20  Baseline Algorithms  Uniform random sampling  Max-min distance sampling  Road network  A 6x6km 2 region, 1802 intersections,  Obtained from 2008 Tiger/Line Shapefiles  Each edge is associated with an interval of travel speed [  -5,  ] (m/s),  2 [10,20]  Movements: all pair shortest paths ¸ 2km  Each AP has unit cost and a sector based coverage model with radius in [100,200](m)  To evaluate average throughput  Ns-2 based simulation  Restricted random waypoint  1Mbps for each AP  CBR traffic

21. Simulation Results 21  A small controlled experiment in a parking lot at OSU (result in paper) Min Contact Opp in TimeAvg Contact Opp in Time Avg Throughput (2x2km 2, 20 APs, 5 users)

22. Outline 22  Motivation  Three Metrics  Contact Opportunity in Distance  Contact Opportunity in Time  Average Throughput  Evaluations  Summary and Future Work

23. Summary and Future Work 23  We have proposed Contact Opportunity, an expressive sparse coverage mode for providing data service to mobile users, and efficient solutions that maximize the worst-case Contact Opportunity with various uncertainties considered.  Future Work - Expected Contact Opportunity or Throughput  Offline - stochastic modeling of uncertainties on mobility and data flows  Online scheduling to improve fairness

24. Contact Opportunity in Time (Cont.) 24  A traffic scenario k - an assignment of travel time (any value from the interval) to each road segment  K S - the worst traffic scenario that minimizes ´ t ( S, p ) for each p, which assigns the minimum travel time to every segment covered by S and maximum travel time to every segment not covered

25. Contact Opportunity in Time (Cont.) 25  Unfortunately, ´ t ( S, p, k S ) 8 S µ A is normalized, nondecreasing, but not submodular  Approximation by a single scenario independent of S  “mean” scenario assigns the average travel time to each road segment, denoted as k 0  S 0 - optimal deployment with respect to k 0  S * - optimal deployment with respect to k S  If the ratio between the maximum and the minimum travel time is bounded by  for all road segments, then ´ t ( S *, p, k S * ) ·  ´ t ( S 0, p, k S 0 ).