CARS: Context Aware Rate Selection for Vehicular Networks Pravin Shankar Tamer Nadeem Justinian Rosca Liviu Iftode

2 Vehicular networks today Ubiquity of WiFi Cheaper, higher peak throughput compared to cellular New applications Traffic Management Urban Sensing (eg. Cartel) In-car Entertainment Social Networking (eg. RoadSpeak, MicroBlog) Requirement: High throughput

3 What is rate selection? PHY: multiple transmission rates 8 bitrates in a/g (6 – 54 Mbps) 8 bitrates in p (3 – 27 Mbps) Different modulation and coding schemes LowHigh Low  High Error Rate HighUnderutilization  Link Quality Bitrate

4 High quality link Low quality link Rate selection problem in vehicular networks 54 Mbps 6 Mbps Rate Selection: Select the best transmission rate based on link quality in real-time to obtain maximum throughput Low quality link 6 Mbps

5 Outline Introduction Existing solutions CARS: Context Aware Rate Selection Evaluation Conclusion

6 Existing rate selection algorithms ARF (1996), RBAR (2001), OAR(2004), AMRR (2004), ONOE (2005), SampleRate (2005), RRAA (2006) (and many more…) Basic scheme in all existing algorithms Estimation: Use physical layer or link layer metrics to estimate the link quality (Re)Action: Switch to lower/higher rate Question: How well do these algorithms work in vehicular environments?

7 Existing schemes + vehicular networks: Experiment Outdoor experiments comparing SampleRate [2005] AMRR [2004] ONOE [2005] 5 runs per rate algorithm 5 runs per fixed rate Slow Mobility: 25 mph Metrics Average goodput Supremum goodput (maximum among all runs for all rates)

8 Existing schemes + vehicular networks: Results Underutilization of link capacity

9 Existing schemes + vehicular networks: Analysis Rapid change in link quality due to distance, speed, density of cars Problems: 1.Estimation delay 2.Sampling requirement 3.Collisions vs. channel errors

10 Problem 1: Estimation delay 6 Mbps 24 Mbps 54 Mbps Link conditions change faster than the estimation window - the rate adaptation lags behind

11 Problem 2: Sampling Requirement When an idle client starts transmitting, there are no recent samples in the estimation window Packet scheduling causes bursty traffic Results in anomalous behavior

12 Problem 3: Collisions vs. errors Hidden-station induced losses should not trigger rate adaptation [CARA06, RRAA06] Lower rate prolongs packet transmission time, aggravating channel collisions Use of RTS/CTS causes additional overhead

13 Outline Introduction Existing solutions CARS: Context Aware Rate Selection Evaluation Conclusion

14 CARS at a glance Rapid change in link quality due to distance, speed (context) Vehicular nodes already have this context information Use this cross-layer information at the link layer to estimate link quality and perform proactive rate selection

15 CARS: reactive + proactive Link Quality: Error Function E H = f(bitrate, len) Reactive Short-term loss statistics from estimation window E C = f(distance, speed, bitrate, len) Proactive Predicted error as a function of context information

16 Proactive rate selection using E c E C = f(distance, speed, bitrate, len) Model link error rate as a function of context information and transmission rate Empirically derived using data from outdoor experiments Simple model is sufficient because of discrete rates in Context recalculation frequency = 100 ms

17 CARS Algorithm

18 CARS Implementation The CARS algorithm was implemented on the open-source MadWifi wireless driver ~ 520 lines of C code Context information obtained from TrafficView [2004] Generic /proc interface: Any other app can be extended to provide a similar interface Extensively tested by means of vehicular field trials and simulations

19 Outline Introduction Existing solutions CARS: Context Aware Rate Selection Evaluation Conclusion

20 CARS Evaluation Effect of Mobility: How does CARS adapt to fast changing link conditions? (Field trial) Effect of Collisions: How robust is CARS to packet losses due to collisions? (Field trial) Effect of Density of Vehicles: How does the throughput improvement scale over large number of vehicles? (Simulation study)

21 Effect of mobility: Setup Scenarios Stationary: Base case Cars are stationary next to each other. SlowMoving: A simple moving scenario Cars are driving around the Rutgers campus: ~25mph speeds FastMoving: A more stressful moving scenario Cars are driving on New Jersey Turnpike: ~70mph speeds in high car/truck traffic conditions Intermittent: A scenario with intermittent connectivity Cars move in and out of each other's range periodically - Hot-spot scenario Workload: UDP traffic from TX to RX using iperf Duration of experiment - 5 minutes

22 Effect of mobility: Results SampleRate CARS StationarySlowMovingFastMovingIntermittent Scenario Goodput (Mbps)

23 Effect of mobility: Analysis Scenario: Intermittent Reactive vs. Proactive

24 Effect of vehicle density - Setup Hotspot scenario: Road of length 5000 m with multiple lanes Base station in the middle of the road Workload: Video stream: 1500 packets of size 1000 bytes each UDP: transmission rate 100 packets per second RTS/CTS disabled Max_retransmits: 4 ns-2 with microscopic traffic generator Compared CARS with AARF and SampleRate

25 Effect of vehicle density - Results

26 Effect of vehicle density - Analysis

27 Outline Introduction Existing solutions CARS: Context Aware Rate Selection Evaluation Conclusion

28 Conclusion Existing rate adaptation algorithms under-utilize vehicular network capacity CARS: uses context information to perform fast rate selection Significant goodput improvement over existing algorithms

29 Backup Slides

30 Limitations of CARS model Other effects (non-modelled) can cause packet loss, eg. multipath, shadowing, environmental effects (rain or snow), background interference Solution: Fall-back mode ( α=0) Enter Fall-back mode if predicted packet loss – measured packet loss > Threshold Future work: Better modeling

31 Signal strength based rate adaptation Stationary Vehicles Moving Vehicles (25 mph) RSSI Spikes (average 5 dB, peaks of upto 14 dB) Moving vehicles: large-scale path loss is more significant than small-scale fading Overhead due to 4-way RTS-CTS-DATA-ACK handshake [Kemp08] frame format (CTS) needs to be extended

32 Estimation window size SampleRate default ew_size = 10 sec We modify SampleRate to ew_size = 1 sec Vehicle with speed 65 mph moves 30m in 1 sec Optimal rate could be different for distances separated by 30m Problem with very small estimation window: Insufficient samples in estimation window [RRAA06] Future work: Estimation window size tuning

33 Capture Effect When there is a collision between the transmitter's frame and a frame sent by a hidden node, the transmitted frame will be successfully demodulated if P t and P j are the received power from transmitter and hidden node α r : threshold ratio at transmission rate r Implications on rate adaptation: α r varies with r Existing collision-aware rate adaptation algorithms do not consider capture effect Future work: model capture effect and use it to guide our rate adaptation scheme

34 Existing Models Existing models in literature Effect of Distance: Free space path loss model Two ray propagation model in LOS environment More complex fading models (Rician, Rayleigh, …) Effect of Mobility: Delay tap model Ray models with Rician delay profiles It is unclear how closely the outdoor VANET environment resembles the existing models Our model is empirically derived using data from extensive outdoor experiments

35 Load and Overhead Comparison LoadOverhead Load: average airtime needed to transmit one packet Overhead: average non- useful airtime needed to transmit one packet

36 Effect of Collisions Scenario: Stationary vehicles located close to hot-spot (to guarantee high-quality links)

37 Evaluation - Mobility - Scenarios Elapsed Time (Sec) Distance (m)Speed (mph)

38 CARS multi-rate retry chain

39 Existing Rate Adaptation Algorithms Auto Rate Fallback [Kamerman et al. ‘97] Drop the transmission rate on successive packet losses and increase it on successive successful packet transmits Adaptive ARF [Lacage et al. ‘04] Uses dynamic instead of fixed frame error thresholds to decrease/increase rate Robust Rate Adaptation Algorithm [Wong et al. ‘06] Uses a short-term loss ratio to opportunistically adapt to dynamic channel variations

40 Existing Rate Adaptation Algorithms SampleRate [Bicket et al. ‘06] Throughput-based scheme Goal is to minimize the mean packet transmission time Sends periodic probe packets at other rates Collision-Aware Rate Adaptation [Kim et al. ‘06] Goal is to distinguish different causes of packet loss Collisions Channel Errors Proposes an adaptive RTS/CTS scheme to prevent hidden-station induced collisions

41 What is context in vehicular networks? Typical vehicular applications make use of location and neighbor information obtained using GPS device Traffic/Safety application Vehicles thus have real-time context information about the environment Examples of context information Distance between transmitter and receiver Relative speed between transmitter and receiver Direct and predictable source of information about link quality

42 Effect of collisions Scenarios: Base: Base case Hidden-Node: Collisions due to hidden node Workload: UDP traffic: iperf Duration: 5 mins TX rate - 3 Mbps IX is out of carrier sensing range of TX

43 Effect of collisions Sequence Number Transmission Rate (Mbps)

44 CARS Evaluation – Field Trial Low Mobility: 25 mph 5 runs per rate algorithm

45 Context Aware Rate Selection (CARS) - Approach Use context information to “learn” the link quality E C = f(distance, speed, bitrate, len) Proactive Predicts large-scale path loss due to mobility Use short-term loss statistics to exploit short- term opportunistic gain E H = f(bitrate, len) Reactive at very small time scale Handles loss due to small-scale fading

46 Putting the two pieces together Issue: When to use E C and when to use E H ? Answer: Weighted decision function PER = α. E C ( ctx,rate,len )+(1-α). E H ( rate,len ) Use context information (vehicle speed) to assign weights α = max(0,min(1,speed/S)) S = 30 m/s (= 65 mph)

47 CARS Algorithm

48 Experiment Trajectory

49 CARS Algorithm

50 Effect of vehicle density