Overview Goal: video streaming in vehicular networks via WiFi Compelling usage scenarios –Gas stations and local shops deploy APs to provide video and ads –Taxis/buses provide value-added services to passengers Cellular networks: costly ($60 for 5GB/month  0.1Mbps for <5 days!); limited bandwidth

Enabling High-Bandwidth Vehicular Content Distribution U. Shevade, Y. C. Chen, L. Qiu, Y. Zhang, V. Chandar, M. K. Han, H. H. Song, Y. S. Seung UT Austin

3 Challenges Vehicles move at high speed –WiFi contacts are short and intermittent –70% contacts less than 10 seconds Sparse AP coverage –Dense coverage over large area expensive Internet access links to APs are bottleneck –Naïve solution: Download from Internet during contact –Insufficient b/w if data fetched during contact

Types of Connectivity AP wireline access: persistent connectivity, but insufficient BW –Internet-to-AP throughput is 768Kbps-6Mbps (DSL) –Cannot sustain high data rate if data is fetched only during contact AP wireless access: high BW, but short-lived connectivity –Our measurements: AP-to-car throughput is Mbps using n –High vehicular speed  short contact (70% contacts less than 10s) Wireless mesh network: high BW, but low coverage Vehicle relay traffic between APs: high BW, high delay Q: Can we combine multiple types of connectivity to enable high-bandwidth vehicular content delivery?

5 Synergy among connections High b/w, short-lived High b/w, high delay Low b/w, persistent High b/w, low coverage VCD High b/w, persistent

6 VCD Architecture Controller Content Source Internet Download and upload data Upload GPS location updates, video demands, what car has Download and upload data Upload GPS location updates, video demands, what car has

7 Contributions New techniques to optimize replication –Goal: Fully utilize wireless bandwidth during contact –Optimize wireline replication to Internet-connected APs –Optimize mesh replication and use it for cooperative caching –Replicate using vehicular relays to APs New algorithm for mobility prediction –Predict set of APs that will be visited by vehicles Critical for success of replication techniques

8 Wireline Replication Controller collects vehicle demands for interval (i+1) and what content is present at vehicles and APs Predicts set of APs visited by vehicle in interval (i+1) Computes what content should be replicated to which APs Content servers replicate content to APs At start of interval i Vehicle downloads content from APs During interval (i+1) During interval i

9 Optimize Wireline Replication Interval length, Content present at cars and APs, car demand, AP-to-visit Content to transfer to APs and content to download to cars Total content downloaded to cars weighted by interest, while minimizing the amount of content replicated to APs Total download from AP to car bound by wireless capacity Per-file download to car bound by the difference between file size and what car already has Per-file download to car cannot exceed what AP already has and what is replicated to it from the Internet Per-file replication to AP bound by the difference between file size and what AP already has Total replication to AP does not exceed Internet access link capacity For each interval i, compute replication strategy maximizing user satisfaction for interval (i+1)

10 Contributions New techniques to optimize replication –Goal: Fully utilize wireless bandwidth during contact –Optimize wireline replication to Internet-connected APs –Optimize mesh replication and use it for cooperative caching –Replicate using vehicular relays to APs New algorithm for mobility prediction –Predict set of APs that will be visited by vehicle Critical for success of replication techniques

APs are often close enough to form mesh networks CDF of total contact duration with AP connected components Mesh Networks of APs Substantial contact with APs that can potentially form mesh networks 11 San Francisco, 100m rangeSan Francisco, 200m range Seattle, 100m rangeSeattle, 200m range

Nearby APs can be organized into mesh networks using another wireless card –Replicate content to APs using mesh in addition to Internet link –Fetch missing content from other mesh nodes rather than Internet Changes to linear program –Constraint C3: –Two new constraints: –Objective function: Add Mesh Networks of APs Per-file download to car cannot exceed what AP already has and what is replicated to it from the Internet and from the mesh AP cannot replicate more content over mesh than it has Interference constraint: Total active time of all mesh nodes cannot exceed 100%, assuming all nodes interfere with each other Prefer a replication which uses less mesh traffic among the ones supporting equal traffic demands 12

13 Contributions New techniques to optimize replication –Goal: Fully utilize wireless bandwidth during contact –Optimize wireline replication to Internet-connected APs –Optimize mesh replication and use it for cooperative caching –Replicate using vehicular relays to APs New algorithm for mobility prediction –Predict set of APs that will be visited by vehicle Critical for success of replication techniques

14 Vehicular Replication Vehicles act as data relays between APs Simple strategy: epidemic dissemination –Vehicle uploads content to AP based on expected future demand at AP AP computes future demand, car notifies what it has AP requests content from the car –Vehicle downloads content from AP First the files it is interested In remaining time, download content randomly

15 Mobility Prediction Predict which APs a car will meet in next interval Challenges: –Vehicles move at high speeds –GPS location updates from vehicles Low frequency Irregular updates –Road and traffic conditions highly dynamic Previous work: 1 st and 2 nd order Markov models –Do not perform well on our dataset

16 Voting among K Nearest Trajectories Exploit history to predict contact: Vehicle’s near history Past trajectories from other vehicles Find K trajectories that most closely match the vehicle’s recent history Obtain future path for K trajectories Report all APs visited by at least T of K trajectories Find K trajectories that most closely match the vehicle’s recent history Obtain future path for K trajectories Report all APs visited by at least T of K trajectories

#Correctly predicted APs #Total predicted APs Setup: Gas stations as APs, radio range = 200m, prediction interval 3min 1200 Seattle city buses Mobility Prediction Results Voting among K nearest trajectories performs best for our dataset #Correctly predicted APs #Total APs actually visited ( 2 ) (1/precision+1/recall) 17 Bus mobility is more predictable 500 San Francisco Yellow Cabs

UDP with congestion control 18 VCD Implementation Controller Coordinator LP Server Content servers b APs n APs Ethernet C++ on Linux TCP for control messages, UDP for data HP iPaq, HTC Tilt C# on Windows Mobile 6.1 Dell, Macbook Pro C# on Windows XP

b Testbed 14 APs deployed in 8 campus buildings –APs are in-building, 20-60ft from the road –802.11b radios with fixed data rate of 11Mbps –3 APs in ACES form a mesh network –Smartphone clients stream H.264 videos at 64Kbps ,8,9,10, 11,12 13, 14

802.11n Testbed n is the new WLAN standard –Considerable throughput increase over b/g –Uses MIMO and 20/40MHz channels Vehicular throughput experiments –Considerable potential throughput increase over 11b Deployed four n APs –Laptops used as clients b4.6Mbps g22.2Mbps n, 2.4GHz39.7Mbps n, 5GHz56.1Mbps

APs: Gas stations, 100m range Results – Simulation Setup: 50 cars, Zipf-like demands, 50% APs not connected to Internet 21 APs: Coffee Shops, 100m range Internet is the bottleneck Benefit from wireline replication Wireless replication helps! Wireline+wireless 5.2X baseline 6.3X better than baseline VCD achieves higher throughput by combining wireline, wireless and mesh replication Mesh adds 3-13%

Results – Simulation Setup: 50 cars, Zipf-like demands, 50% APs not connected to Internet 22 APs: Coffee Shops, 100m range Mesh benefits % Benefits increase with higher range and dense AP deployment APs: Coffee shops, 200m range Low Medium High Video quality over 3G

23 Emulab: Simulator Validation Simulator results within 10% of Emulab results All APs connected to Internet10% APs connected to Internet Setup: 30 APs, 100 cars, 200m range

802.11b testbed: 8 APs, 3 connected by mesh n testbed: 4 APs, all connected by mesh Results - Testbed 24 Download (kB)Play time (sec) No replication Wireline Wireline + Mesh Full replication Download (kB)Play time (sec) No replication Wireline Wireline + Mesh Full replication X 7.8X

Summary: Vehicular Content Distribution KNT: A new mobility prediction algorithm –Based on voting among K nearest trajectories –25-94% more accurate than 1 st and 2 nd order Markov models A series of novel replication schemes –Optimized wireline replication and mesh replication –Opportunistic vehicular relay based replication Extensive evaluation: simulation + testbed + emulation –Simulation using San Francisco taxi and Seattle bus traces 3-6x of no replication, 2-4x of wireline or vehicular alone –Full-fledged prototype deployed on two real testbeds 14-node b testbed and 4-node n testbed x gain over no replication –Emulab emulation with real AP/controller and emulated vehicles Show system works at scale and is efficient Validate our trace-driven simulator