BreadCrumbsForecasting Mobile

Connectivity

Presented by Tao HUANGLingzhi XU

Context• Mobile devices need exploit

variety of connectivity options as they travel.  • Operating systems manage wireless networks in the moment,

reactively choosing connections only when circumstances change.

Problem• Applications cannot make reliable assumptions about the

quality of connectivity.

Observation• People are creatures of habit; they take similar paths every

day.

Background

Determining AP quality • Virgil

Estimating client location• GPS• Place lab

Determining AP QualityCurrently• Select the unencrypted AP with the strongest signal

VirgilConnects to reference servers to estimate connection quality:• Downstream bandwidth• Whether AP blocks certain services• Estimated latency

BreadCrumbs uses Adapted Virgil, estimates three values:• Downstream bandwidth, • Upstream bandwidth, • Estimated latency

Estimating Client Location

GPS• Provide latitude and longitude (0.001◦×0.001◦, 110m*80m)

Place Lab• For devices without GPS• Works well indoors and in urban canyons• Rely on public Wardriving database

Contribution

• Authors introduce the concept of connectivity forecasts for mobile devices.

• Authors demonstrate that such forecasts can be accurate over regular, day-to-day use, without requiring GPS hardware or extensive centralized infrastructure.

• Authors illustrate the potential benefits of the system through three example applications

Connectivity Forecasting

It is an estimate of the quality of a given facet of network connectivity at some future time.

• Predicting future mobility

• Forecasting future conditions

Predicting Future Mobility

• Each state consists of two sets of coordinates: the location where the device was during the last state, and its current location.

• The frequency with which BreadCrumbs estimates the device’s GPS location bounds the resolution of the mobility model.

Forecasting Future Conditions (1)

• Build an AP quality database to estimate the “quality” of a connection to the Internet

• When BreadCrumbs first encounters an unencrypted AP, it attempts to estimate (1) downstream bandwidth, (2) upstream bandwidth, and (3) latency to remote Internet hosts.

• The test database tracks access points both by ESSID and by MAC address, and tags all AP test results with GPS coordinates.

Forecasting Future Conditions (2)

BreadCrumbs combines the custom user mobility modeland the AP quality database to provide connectivity forecasts.

One step is τ seconds in the future.

Example

Implementation

Scanning thread• Scans for access points and fixes the device’s GPS

every10 seconds• Updating the transition probability• All data are stored in local database

Application interface• Handle application requests for connectivity forecasts.• Requests consist of two value: criterion of interest and

number of seconds in the future

Evaluation: Methodology

Track movements for two weeks• Mobile: Familiar Linux on iPAQ + WiFi• Mixture of walking and bus• First week as training set, while second week for evaluation

Authors use three sample applications to examine how both the operating system and different mobile applications could benefit from connectivity forecasts.

Forecast Accuracy Location Prediction

Accuracy Bandwidth Prediction

Sample Applications: Map View

• BreadCrumbs’ forecasts could avoid wasteful pre-fetching.

Sample Applications: Streaming Media

• Connectivity forecasts provide the same playback experience while using significantly less of the mobile device’s battery.

Sample Applications: Opportunistic Writeback

• Although BreadCrumbs completes slower than no prediction algorithm, it makes more efficient use of the wireless radio.

Overhead

Related work

• Context-for-wireless: Choose between WiFi and cellular data networks, no prediction

• MobiSteer: Improve wireless network connectivity in motion by using a directional antenna.

• Predictability of WLAN mobility and its effects on bandwidth extensive:

Using different prediction methods to improve bandwidth provisioning and handoff for VoIP telephony

• Building realistic mobility models from coarse-grained traces: Build model from different client traces, more close real movements

Future Work

• Deploy BreadCrumbs on the mobile device to investigate how sharing of these databases among co-located users can reduce this scanning overhead further.

• Add encrypted WiFis which device has authority to access to database. Those encrypted WiFis own higher weight than those unencrypted. Test whether it will improve performance.

Conclusion

• Applications cannot make reliable assumptions about the quality of connectivity. So it cannot provide relative stable connectivity performance.

• BreadCrumbs tracks device trend of connectivity quality as its owner moves around. The predictions of the mobility model and the AP quality database yield connectivity forecasts.

• BreadCrumbs can provide improved performance while reducing power consumption only with one week training time.

• It can works on devices without GPS.

Questions