Wireless “ESP”: Using Sensors to Develop Better Network Protocols Lenin Ravindranath Calvin Newport, Hari Balakrishnan, Sam Madden Massachusetts Institute of Technology

Big Changes in Access Devices 172M smartphones sold worldwide in 2009 – 25% of US phone market; 50% in two years Smartphones and tablets will exceed PC sales by 2011 Mobile Internet growing at a tremendous pace

Big Changes in Access Devices Dominant mode of data access in the future

“Truly Mobile” Devices Often switch between static and mobile Exhibit a variety of mobility modes Move through different environments

Protocols need to adapt to different settings – Mobility mode impacts wireless performance The Problem Most protocols optimized for static settings – They perform poorly during mobility Protocols that compensate for mobility are not optimal in static settings

Static vs. Mobile Channel constantly changing – Channel assessments quickly outdated – Protocols should not maintain long histories Channel relatively stable – Protocols can average estimates – Ignore short-term variations

Topology is hardly changing – Probe for links less frequently – Compute routes over long time scales Topology changing rapidly – Probe for links more often – Compute routes over shorter time scales Static vs. Mobile

Current Wireless Protocols Do not differentiate between mobility modes Attempt to adapt to both settings implicitly using measurements of packet loss, SNR, BER Leading to suboptimal performance Lack of explicit knowledge about prevalent mobility mode Can we do better?

Proximity Sensor Camera Ambient Light Sensor Microphone Accelerometer GPS Compass Gyro

Accelerometer Proximity Sensor Camera Ambient Light Sensor Microphone GPS Compass Gyro Many, many, applications…

Accelerometer Proximity Sensor Camera Ambient Light Sensor Microphone GPS Compass Gyro Ignored by Protocols!

Application Transport Network MAC PHY Wireless Radio Wireless Protocol Stack Accelerometer Proximity Sensor Camera Ambient Light Sensor Microphone GPS Compass Gyro Ignored by Protocols!

Application Transport Network MAC PHY Wireless Radio Wireless Protocol Stack Accelerometer GPS Compass Gyro

Application Transport Network MAC PHY Wireless Radio Wireless Protocol Stack GPS Compass Accl Gyro Hints

Application Transport Network MAC PHY Wireless Radio Wireless Protocol Stack GPS Compass Accl Gyro Hints Movement Direction Speed Use hints to adapt to different mobility modes differently Hints Protocol Adapt to hints from neighbors

Application Transport Network MAC PHY Wireless Radio Wireless Protocol Stack GPS Compass Accl Gyro Rate Adaptation Movement Heading AP Association Speed Vehicular Routing Walking

Application Transport Network MAC PHY Wireless Radio Wireless Protocol Stack GPS Compass Accl Gyro Rate Adaptation Movement Heading AP Association Speed Disassociation Packet Scheduling Power Saving Preamble Network Monitoring Speed Walking Location Vehicular Routing

Application Transport Network MAC PHY Wireless Radio Wireless Protocol Stack GPS Compass Accl Gyro Rate Adaptation Movement Heading AP Association Speed Walking Vehicular Routing

Application Transport Network MAC PHY Wireless Radio Wireless Protocol Stack GPS Compass Accl Gyro Rate Adaptation Movement

Accl Movement Reliably detect movement within 100ms

Application Transport Network MAC PHY Wireless Radio Wireless Protocol Stack GPS Compass Accl Gyro Rate Adaptation Movement

Rate Adaptation in Wireless Networks 6 Mbps 9 Mbps 12 Mbps 18 Mbps 24 Mbps 36 Mbps 48 Mbps 54 Mbps a/g bit rates Packet encoded at a particular bit rate Rate Adaptation: Finding the best bit rate to transmit a packet

Static vs. Mobile Performance Static and walking traces – Cycle through bit rates 4 different environments – 80 traces, 20 seconds long Trace-driven simulation – TCP throughput Static Sample Rate85 – 99% RRAA80 – 97% RBAR 70 – 80% CHARM Moving Sample Rate33 – 59% RRAA45 – 63% RBAR 60 – 75% CHARM Compare to optimal throughput

Static vs. Mobile Loss Patterns Probability that packet i+k is lost given packet i is lost 10 ms Losses are more bursty when a node is mobile than when a node is static k

6 Mbps 9 Mbps 12 Mbps 18 Mbps 24 Mbps 36 Mbps 48 Mbps 54 Mbps 6 Mbps 9 Mbps 12 Mbps 18 Mbps 24 Mbps 36 Mbps 48 Mbps 54 Mbps RapidSample 6 Mbps 9 Mbps 12 Mbps 18 Mbps 24 Mbps 36 Mbps 48 Mbps 54 Mbps 1. After a single loss  Reduce rate 2. History - 10 ms  Don’t retry a failed rate  Or any higher rate 3. Channel not degrading, probably improving  After few successes, sample higher rate not failed  If wrong, come back to the original rate [failed – within last 10ms]

RapidSample, when device is moving Up to 75% better throughput than SampleRate 25% better than other protocols Trace driven (ns3) 30 traces 20 seconds long TCP throughput

But when static… Up to 30% lower throughput than other schemes Trace driven (ns3) 30 traces 20 seconds long TCP throughput

Application Transport Network Rate Adaptation PHY Wireless Radio Wireless Protocol Stack GPS Compass Accl Gyro Movement RapidSample when movement SampleRate when static Movement Hint-Aware Rate Adaptation

40-50% better than other schemes Trace driven (ns3) 10 traces 20 seconds long Static + Moving TCP throughput

Application Transport Network MAC PHY Wireless Radio Wireless Protocol Stack GPS Compass Accl Gyro Rate Adaptation Movement Heading AP Association Speed Walking Vehicular Routing

Application Transport Network MAC PHY Wireless Radio Wireless Protocol Stack GPS Compass Accl Gyro Heading AP Association Walking

AP Association Scan Infrequent scans

AP Association Suboptimal Association Static

Movement-Aware Association 1. Static – Stop Scanning 2. Moving – Scan Periodically 3. Moving to Static – Scan once

Movement-Aware Association On median, 40% more throughput Android implementation 30 traces Static + Moving Throughput

Heading-Aware Association Minimize Handoff Training based approach Heading

Heading-Aware Association 40% median reduction in handoffs Android implementation Training (30 traces) 30 traces # Handoffs

Application Transport Network MAC PHY Wireless Radio Wireless Protocol Stack GPS Compass Accl Gyro Rate Adaptation Movement Heading AP Association Speed Walking Vehicular Routing

Application Transport Network MAC PHY Wireless Radio Wireless Protocol Stack GPS Compass Accl Gyro Heading Speed Vehicular Routing

Routing in Vehicular Mesh Networks “V2V”

Routing in Vehicular Mesh Networks Longevity of links useful – avoids expensive repairs Link between nodes heading in the same direction tend to last longer Connection Time Estimate (CTE) Use heading, speed and position to predict connection duration

Routing in Vehicular Mesh Networks Heading [0, 9)[10, 19)[20, 29)[30, 180]All Links Link Duration (s) Empirical evaluation on taxi traces 15 networks, 100 vehicles each Links with similar heading lasted 4 to 5 times longer than the median duration over all links

Application Transport Network MAC PHY Wireless Radio Wireless Protocol Stack GPS Compass Accl Gyro Rate Adaptation Movement Heading AP Association Speed Walking Vehicular Routing

Application Transport Network MAC PHY Wireless Radio Wireless Protocol Stack GPS Compass Accl Gyro Rate Adaptation Movement Heading AP Association Speed Disassociation Packet Scheduling Power Saving Preamble Network Monitoring Speed Walking Location Vehicular Routing

Related Work Wireless power saving – WakeOnWireless, Cell2Notify, Blue-Fi Vehicular networking – use GPS – AP association Mobisteer, Breadcrumbs – Rate adaptation CARS: Adapt rate based on speed and heading Very recent work – Accelerometer-assisted rate adaptation

Take-Away Message Truly mobile devices will soon be dominant – Variety of mobility modes poses problems for wireless protocols Sensors on these devices give us a new opportunity to develop network protocols Protocol architecture using sensor hints can significantly improve MAC, link, network layers

Backup

Probing How frequently should nodes probe? Delivery Probability ETX, ETT Probes

Infrequent Probing Inaccurate link estimation leads to poor throughput Inaccurate link estimation leads to poor throughput

Frequent Probing Probing wastes bandwidth

Delivery Probability Mobility causes delivery probability to fluctuate with bigger jumps Mobility causes delivery probability to fluctuate with bigger jumps

Static vs. Mobile Mobile case requires 20x more probes to maintain acceptable estimation error Mobile case requires 20x more probes to maintain acceptable estimation error

Adaptive Probing Protocol Adapt probing based on movement hints When a node is static – Probe infrequently (1 probe every 2 seconds) When a node is mobile – Probe frequently (10 probes per second)

Adaptive Probing Tracks the link accurately with fewer probes

Pruning association