Context Aware Power Management of Mobile Systems for Sensing Applications

Ralf Herrmann Universität der Bundeswehr München (UniBwM) Werner-Heisenberg-Weg 39; Munich, Germany

ralf.herrmann@unibw.de

Piero Zappi, Tajana Simunic Rosing University of California San Diego

9500 Gilman Dr.; La Jolla, CA, 92093 {pzappi, tajana}@ucsd.edu

ABSTRACT The widespread adoption of mobile, yet powerful smartphones enable the development of distributed sensing applications. For such applications, the smartphone plays a central role by acting as a bridge between user worn sensors and a remote server. Smartphone energy consumption is a critical issue that should be considered. As the user’s daily context continually changes, the application can dynamically adapt its behavior to the user context to reduce the smartphone energy consumption. In this work, we propose to use context knowledge to dynamically adapt the behavior of sensing applications running on smartphones. In our system a context aware application manager starts, suspends and changes the sampling rate of a sensing application according to current user’s context. We tested our technique on a real system and we show that it is can extend smartphone lifetime by 5x.

Categories and Subject Descriptors J.m. [Computer Applications]: Miscellaneous – Power-Aware Computing; C.5.3 [Computer System Implementation]: Microcomputers – Portable Devices; I.5.m [Pattern Recognition] Miscellaneous – Context Recognition.

General Terms Management, Design, Experimentation

Keywords smart phone, energy, context

1. INTRODUCTION Today’s smartphones provide enhanced computational, sensing and user interface capabilities that allow the development of complex applications. Smartphones are an effective solution for a number of applications that need continuous sensing. Examples include health care [1], and participatory sensing [2]. These projects focus primarily on data collection, sharing and dissemination aspects. The typical scenario consists of one or more body worn sensors which continuously collect data from the user or the user surrounding. This data is aggregated on the smartphone with the data from its embedded sensors (i.e. GPS, accelerometer etc.). Furthermore, the smartphone acts both as a gateway to a backend server and as interface for data presentation. For example BikeNet [3] is a mobile sensing system used for mapping the cyclist’s experience. In this project a number of sensors are embedded in a bike and communicate with a Nokia N80 phone to aggregate statistics on the user and his behavior. These statistics are sent to a backend server and shared among the user community. Another example is the CitiSense project [6] which relies on body worn air quality sensors and smartphones to

provide feedback to users on how air quality affects their health. In these examples a sensing application is continuously running on the smartphone collecting and forwarding data to a backend server. Continuous collect and forward of data from sensors to backend impacts on the smartphone's battery life especially when the samples needs to be geo tagged. In fact, just tracking location can reduce smarthphone lifetime to a few hours [4]. Furthermore, wireless interfaces with the body worn sensors, such as Bluetooth, impact the smartphone battery life. However, often data is collected within a context which is not of interest. For example, in an air quality monitoring application, samples collected within a vehicle fail to detect the real pollutant concentration in a certain area due to the filters in the air condition system. A more efficient approach would rely on user context detection to modify the application behavior as a user moves within different situations.

Context-aware computing refers to a general class of mobile systems that can sense their environment and can adapt their behavior accordingly [14]. For example, in [5] a location aware application running on a PDA provides information relevant to caregivers in a hospital. Energy efficient context recognition has been extensively studied in the past. Related work on context aware computing either focus on energy efficient context recognition techniques [7] or propose energy efficient context management frameworks [8][9]. However, to authors best knowledge, the problem to leverage user’s context to optimize how other applications running on the smartphone utilize system resources has received little attention from the research community. Context knowledge to improve system energy efficiency has been proposed in [10]. Here a stochastic model of the user behavior within a smart home is developed to predict the user’s movements. The user proximity to appliances the system context to activate only a subset of devices thus reducing overall system energy consumption. This work targets a scenario significantly different from the one addressed in this paper. First it handles with devices which do not have constraints in their power consumption, and second the application is distributed among several distinct systems while in our scenario the application should be executed always by the same device. Furthermore, in [10] only user location information is used, while several other aspects of the user context can be considered. For example, in this work we include user’s mobility and various types of activities.

In this work we propose a context-aware Dynamic Power Management (DPM) subsystem that leverages context knowledge to adapt at run time the behavior of continuous sensing applications. In contrast to previous work on energy efficient context recognition, here we do not propose a novel framework or technique to detect some aspect of context. Rather, we show the benefit of including context knowledge in the management of other applications running on a smartphone. The DPM subsystem can be combined with other energy efficient techniques for context detection to improve system lifetime and scalability.

Our context-aware Dynamic Power Management (DPM) subsystem consists of three components: a Context Recognition

Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. IPSN'12, April 16–20, 2012, Beijing, China. Copyright 2012 ACM 978-1-4503-1227-1/12/04…$10.00.

Service that detects the current user’s contextavailable in user’s proximity, a Context Managercontext detection and issue configuration commands to the Sensing Application which receives the adapt its data collection according to the context dependent commands sent by the Manager. For example, if the user is interested in collecting data only in an outdoor area, the Context ManagersApplication whenever the user enters a building.is to adjust the sensing rate according to the user mobility to avoid unnecessary sampling over the same area.

We implemented our software architecture on an Android smart phone. We test our context aware DPM participatory sensing system. When comparedno context-aware power management is usedincreases phone battery lifetime by more thanwhile delivering quality results to the users.

2. CONTEXT-AWARE DPM In this work, we assume that each user carries one or more body worn sensors nodes that collect data from the environment. Sensors nodes communicate with a smartphonegeo-tags the data, and forwards it to a server. DPM manages applications running on the smartphone

The general architecture of our context-aware DFigure 1. The Context Manager acts as an middlethe Sensing Applications and a Context Recognition Serviceapplication registers for a set of contexts that are provided by the Context Recognition Service. During this phase each application specify its behavior in each context. For example, an air quality monitoring application is interested in collecting the user is in an outdoor context, or the sampling rate increases when the user is biking with respect to the case when the user is either walking or stationary. The Context Manager a Sensing Application whenever it detects that the user is within or without a context of interest. The Context Recognition Servicecollection of energy efficient context detection algorithms

Figure 1. Context-aware DPM subsystem

Sensing application has three services: sensor samplingand communication. Sensor sampling service communicates with the body area network and retrieves data from the sensors. The storage service saves this data for local processing and enables communication service to forward the data to the backend server. The sensing application behavior is controlled by the Manager according to current context. For example, sensing may be limited to particular hours of the day or location and the sensing rate may change according to user mobility.

Context Recognition Service is a collection of algorithms to detect different context variables from the available sensors. Context manager defines which context should be computed and

context using sensors Manager that manages

issue configuration commands to the adapt its data collection

dependent commands sent by the Context . For example, if the user is interested in collecting data

Context Managers stops the Sensing a building. Another example

sensing rate according to the user mobility to avoid

our software architecture on an Android smart using an air quality

When compared to the case where aware power management is used, our technique

increases phone battery lifetime by more than a factor of five

each user carries one or more body

that collect data from the environment. communicate with a smartphone that timestamps,

forwards it to a server. The context aware DPM manages applications running on the smartphone.

ware DPM is shown in middle layer between

Context Recognition Service. Each that are provided by the

During this phase each application xt. For example, an air quality

monitoring application is interested in collecting data only when context, or the sampling rate increases

when the user is biking with respect to the case when the user is Context Manager start and stops

whenever it detects that the user is within or Context Recognition Service is a

collection of energy efficient context detection algorithms.

DPM subsystem

sensor sampling, storage service communicates with

the body area network and retrieves data from the sensors. The saves this data for local processing and enables

to forward the data to the backend server. behavior is controlled by the Context

according to current context. For example, sensing may to particular hours of the day or location and the

sensing rate may change according to user mobility.

is a collection of algorithms to detect different context variables from the available sensors. The

which context should be computed and

how often according to the application needs. The specific context recognition techniques used by the system are out of the scope of this paper. In this work we implemented our own recognition service, however aframework can be used. Note that the context recognition is an extra cost that the system should pay to obtain context knowledge and adapt accordingly. Thus, if a more energy efficient context recognition framework is used,

Context manager manages the detectionand the behavior of the Sensing applicationevaluated only if an application is registered for it. Whenever an application registers for a context variable it also defines how often that variable should be updated, and the accuracy of the recognition. The Context Managercontext variable to the shortest update period among those requested by each application that registered for that variable. This ensures that the number of context evaluations is minimized. Furthermore, it defines the accuracy in thvariable depending on the needs of eachexample, one application may city where the user is walking and therefore can use cell tower triangulation, while the other estimate (i.e. few meters). As long as the second application is running, the Context Manager requests to the Service to provide high accuracy localization using the GPS module. This information is shared to both sensing applicaWhen the second application endsthe constraints on the localization accuracy and the Recognition Service can rely on other more energy efficient techniques such as cell tower based triangulation.

3. SMARTPHONE POTo simulate the power consumption of the smartphone we developed a power model for the HTC Aria our experiments. As the user moves iContext Recognition Service and the status of each component. In the next set of figureswith red boxes the transitions driven by context changes.

Figure 2. GPS model with context

GPS operates in three states as shown in transitions from OFF/SLEEP to ACTIVE state are enabled when samples are geo-tagged. Similarly, the frequency of GPS sampling is set as a function of the user’s speed higher. We do not consider GPS signal strength and number of connected satellites to because that does not sufficiently affectoverall energy consumption module has valid almanac and ephemeris data.

Accelerometer model is shown in accelerometer is used to detect user activity. The transition between SLEEP and NORMAL mode is controlled by the application contexts of interest. Since the transitionSLEEP and NORMAL state requires a few ms, the accelerometer is duty cycled between successive context detections.

how often according to the application needs. The specific context recognition techniques used by the system are out of the scope of this paper. In this work we implemented our own Context

, however any other energy efficient framework can be used. Note that the context recognition is an extra cost that the system should pay to obtain context knowledge and adapt accordingly. Thus, if a more energy efficient context

our energy saving improves.

manages the detection of the context variables Sensing application. Context variables are

evaluated only if an application is registered for it. Whenever an application registers for a context variable it also defines how often that variable should be updated, and the accuracy of the

Manager sets the update period of a context variable to the shortest update period among those requested by each application that registered for that variable. This ensures that the number of context evaluations is minimized. Furthermore, it defines the accuracy in the detection of each

needs of each applications. For may need to know only the area of the

city where the user is walking and therefore can use cell tower the other may require a more accurate

. As long as the second application is requests to the Context Recognition

to provide high accuracy localization using the GPS module. This information is shared to both sensing applications.

ends, the Context Manager relaxes the constraints on the localization accuracy and the Context

can rely on other more energy efficient cell tower based triangulation.

SMARTPHONE POWER MODEL To simulate the power consumption of the smartphone we

for the HTC Aria [11] phone we used in . As the user moves into different contexts the

and the Context Manager change the In the next set of figures, we highlight

with red boxes the transitions driven by context changes.

model with context-aware power controls

operates in three states as shown in Figure 2. The power transitions from OFF/SLEEP to ACTIVE state are enabled when

. Similarly, the frequency of GPS sampling is set as a function of the user’s speed – faster when the speed is

not consider GPS signal strength and number of satellites to because that does not sufficiently affect the

overall energy consumption [12]. We assume that the GPS module has valid almanac and ephemeris data.

model is shown in Figure 3. Typically, the accelerometer is used to detect user activity. The transition between SLEEP and NORMAL mode is controlled by the application contexts of interest. Since the transition between SLEEP and NORMAL state requires a few ms, the accelerometer

duty cycled between successive context detections.

Figure 3. Accelerometer with context-aware power controls

Wi-Fi model is shown in Figure 4. It operates in either low or high utilization mode depending on the amount of data that needs to be sent. Several examples to efficiently detect available Wi-Fi access points are available in the literature [15]. Wi-Fi is activated and used for communication within contexts where data is sent to a backend server and Wi-Fi network is available. Wi-Fi module is also activated when the user is moving to help with localization.

Figure 4. Wi-Fi model with context-aware power controls

Figure 5. Cellular radio model with context-aware controls

Cellular module is shown in Figure 5. It has three power states: IDLE, CELL_FACH, and CELL_DCH. The cellular module is used to forward data to a backend server and to perform cell tower based localization only when Wi-Fi connectivity is not available since cellular connectivity is less energy efficient [16].

Figure 6 . Bluetooth radio model with context-aware controls

Bluetooth model is shown in Figure 6. Bluetooth is activated when the smartphone collects samples from the sensor board. The module enters in the PARK state between successive sample transmissions. Whenever the user enters in a context of interest,

the Context Manager activates the Sensing Application which, scans for available sensors and connects to the sensors of interest. Context dependent application sensing rate modulates the transitions between the PARK and CONNECT.

Figure 7 CPU model with context-aware power controls

We added models of CPU (Figure 7) and memory (Figure 8). As the application changes its behavior when using the context aware DPM, so the CPU and memory power consumption is adjusted accordingly. CPU’s frequency of operation depends on the level of utilization. When the smartphone operates in different contexts, the applications generate a different number of jobs. Therefore, the job arrival rate and the percentage of time the CPU operates at high or low frequency is set accordingly in our simulations. For the memory, we model a low power mobile 200MHz DDR SDRAM [13].

Figure 8. Memory model with context-aware power controls

For both components we measure the requests inter-arrival time as the phone runs various jobs. We model request inter-arrival times with an exponential distribution. CPU jobs inter-arrival time has been measured on an Xscale ARM processor running at 525 MHz which has performance similar to the processor in the Aria phone. Memory request inter-arrival rate has been obtained using architecture level simulation and memory power measurements.

4. RESULTS We implemented and evaluated our context aware DPM within the air quality monitoring Citisense [6] project. Within this project, users carry a sensor node that samples air pollutants and a smartphone that acts as a local aggregator for the data from the sensors, provides connectivity to a backend server and presents feedback to the user regarding his/her exposure (see Figure 9).

The CitiSense sensor board is a wearable sensor node able to detect environmental variables (humidity, temperature and barometric pressure) together with a set of pollutants (Nitrogen Dioxide, NO2, and Carbon Monoxide, CO) and to communicate via Bluetooth to the smartphone. The board is able to operate continuously for almost 6 days with the provided 1800mA/h Li-ion battery. The smartphone is an Android HTC Aria [11] where our context-aware power management DPM runs.

Figure 9. Citisense board and HTC Aria smartphone

The Context Recognition Service detects: userlocation, if the user is indoor/outdoor and if he/she is resting.we do not focus on which technique is used to detect the user context. In fact, the energy savings are a consequence of how the application adapts to the current context. The context detection is an extra cost that the system has to pay to gainknowledge. By using more energy efficient techniquesthis extra cost, thus increasing the gain in energy saving.

The smartphone accelerometer is used to recognize mobility states: walking, riding a bike, going by car or bus and standing. 67 hours of recorded data has been used characterize a J48 tree classifier. We achievedand execution time is 2.2 sec.

Location is provided using area detection. In our experimentsindoor and outdoor location is detected based on the network location provider with an accuracy of 10 m to several hundred meters depending on the number of Wi-Fi access points and cell towers in the connection range. Average execution

Indoor/outdoor detection depends on the abilityGPS module a location fix. When the phone is outsidelocalization takes 2 sec. on average. When indoortakes 42ms to execute plus a waiting delay of

We use the magnitude of the acceleration to detect if the user is at rest or moving. Our assumption is that when the user is resting, he/she doesn't carry the smartphone and leaves it on a still surface (e.g. table). The system assumes that the other context variremain unchanged when the user is at rest, therefore we stop the context detection. Detection of this context takes

4.1 Measurement Results Due to HVAC and air filters, the user exposure to pollutants within buildings and vehicles is negligibimplemented a system where sensing and reporting only if the user is outdoors (either resting or moving) and is not a vehicle. We compare two implementations of the collaborative sensing application. One case is where no performed (naïve system), and the other usesDPM. When the application does not leverage userperiodically samples the sensors and forwardbackend. For both cases keep standard phones epolicies (i.e. turn off the screen when idle etc.).

We carried the sensing system for two days. Tstarted at midnight and run until the battery with a subset of measurements is shown in Figure 10. High pollutant locations (in red) occur on main roads. We compare the battery levels with and without context awareness in With our system the increase in battery lifetime is dramatic: from 577 minutes without context-aware power management to minutes when it is, for an increase of 5.3xuser’s exposure estimation is unchanged between the two cases.

A detailed analysis shows that with our consmartphone power consumption is lower at the beginning end of the day, when the user is at home and the phone is at rest.

. Citisense board and HTC Aria smartphone

: user’s mobility and and if he/she is resting. Here

we do not focus on which technique is used to detect the user a consequence of how the

he context detection is m has to pay to gain the context

knowledge. By using more energy efficient techniques, we reduce this extra cost, thus increasing the gain in energy saving.

is used to recognize the following bike, going by car or bus and

has been used to train and achieved 89.4% accuracy

area detection. In our experiments location is detected based on the network

location provider with an accuracy of 10 m to several hundred Fi access points and cell execution time is 1 sec.

Indoor/outdoor detection depends on the ability to obtain from the When the phone is outside, the GPS

When indoor, the algorithm of 10 sec.

the magnitude of the acceleration to detect if the user is at when the user is resting,

he/she doesn't carry the smartphone and leaves it on a still surface The system assumes that the other context variables

remain unchanged when the user is at rest, therefore we stop also Detection of this context takes 1 sec.

Due to HVAC and air filters, the user exposure to pollutants within buildings and vehicles is negligible. Therefore, we implemented a system where sensing and reporting is performed

(either resting or moving) and is not in We compare two implementations of the collaborative

where no context adaptation is uses our context aware

not leverage user’s context, it and forwards the data to the

For both cases keep standard phones energy saving c.).

We carried the sensing system for two days. The application is started at midnight and run until the battery is discharged. A map with a subset of measurements is shown in Figure 10. High

lutant locations (in red) occur on main roads. We compare the battery levels with and without context awareness in Figure 11.

lifetime is dramatic: from aware power management to 3061

minutes when it is, for an increase of 5.3x. The quality of the user’s exposure estimation is unchanged between the two cases.

A detailed analysis shows that with our context aware DPM the smartphone power consumption is lower at the beginning and the

of the day, when the user is at home and the phone is at rest.

Figure 10. NO2 pollution [ppb] samples. Red indicates high pollution locations, while green is for low pollution areas

Figure 11. Smartphone energy consumption aware sensing and the

Table 1. Scenario definitions

Hour Time Activity

0:00 Night Inactive

6:00 Morning Inactive

9:00 Day Inactive

10:45 Day Active

10:50 Day Active

11:00 Day variable

14:00 Day Active

16:10 Day Active

16:25 Day Inactive

18:00 Evening Inactive

4.2 Simulation resultsWe simulate the phone power consumption scenarios through the finite state machinesection. In our experiments the smartphone a backend server, so we did not model the LCD was not changed by our implementation.

Figure 12. Comparison of phone

We compared the energy consumption measured on the HTC Aria with the simulation of the power consumption of the smart phone hardware components using scenariomatches the schedule of the days the phone has been carried by the user. Our simulated results closely match with root mean square error of only 2.08% (

pollution [ppb] samples. Red indicates high

pollution locations, while green is for low pollution areas

nergy consumption when the context-

aware sensing and the naïve systems are used

. Scenario definitions

Contexts Movement In/Out Area

Resting Inside Home

Resting Inside Home

Resting Inside Home

Walking Outside Home

Walking Outside Univ.

variable Inside Univ.

Walking Outside Univ.

Walking Outside Home

Resting Inside Home

Resting Inside Home

Simulation results simulate the phone power consumption within different

finite state machines shown in the previous the smartphone is used as a gateway to

a backend server, so we did not model the LCD power cost, as it was not changed by our implementation.

phone measurements & simulations

compared the energy consumption measured on the HTC Aria with the simulation of the power consumption of the smart phone

scenarios shown in Table 1 which matches the schedule of the days the phone has been carried by

. Our simulated results closely match real measurements with root mean square error of only 2.08% (Figure 12).

Figure 13 shows the simulated power consumption of the smartphone components. As expected, GPS is the main source of power consumption followed by the wireless communication modules (Bluetooth, Wi-Fi and Cellular radio). This pushes the use of low power techniques to reduce the overhead in energy consumption associated to these components. Furthermore, local computation should be preferred to reduce the amount of data sent through the wireless channel.

Figure 13. Smartphone components energy consumption when the context-aware sensing is used

Table 2 compares the power consumption of phone components when using our context aware technique with the naïve system that streams data continuously. As can be seen from this table, our context aware technique is able to dramatically reduce the power consumption of several components. Since the Sensing Application stops when the user is indoors, power consumption of Bluetooth and GPS is 86% lower when using our policy. On the other hand, the cellular and Wi-Fi module power consumption do not change much. The reason is that these components are used also to perform user localization in addition to data transfer, so this affects their energy cost. Finally, note that the accelerometer is used only for the context recognition. This is an extra cost the system has to pay to get context knowledge.

Table 2. Average power consumption per component

Component Context Aware (mW)

Naïve system (mW)

Power savings (%)

CPU 2.8 37.5 92.4 Memory 8.5 40.6 79.0

Accelerometer 0.02 0 0 Wifi 16.8 20.4 17.5 GPS 32.3 242.2 86.7

Bluetooth 14.0 103.1 86.4 Cellular 13.8 17.8 22.6 Total 88.2 461.7 80.9

5. CONCLUSION In this paper, we propose the use of context knowledge to improve energy efficiency of sensing applications running on smartphones. Our context aware DPM is able to recognize and adapt to changing context. We show that with our approach we are able to extend the smartphone lifetime by more than 5x. We also analyze per HW component contributions to energy cost when context awareness is enabled. As expected, highest contributors are associated to GPS and wireless communication modules.

6. ACKNOWLEDGMENTS This research was supported by NSF Grant CNS-0932403, Center for Networked Systems (http://cns.ucsd.edu) grants CNS08TR and CNS10TR, and Qualcomm.

7. REFERENCES [1] Denning, T., Andrew, A., Chaudhri, R., Hartung, C., Lester,

J., Borriello, G., and Duncan, G. 2009. BALANCE: towards a usable pervasive wellness application with accurate activity inference. HotMobile '10.

[2] Perez, A.J., Labrador, M.A., and Barbeau, S.J. 2010. G-Sense: a scalable architecture for global sensing and monitoring. IEEE Network 24,4. 57-64.

[3] Eisenman, S., Miluzzo, E., Lane, N., Peterson, R., Ahn, G-S., and Campbell, A. 2007. The BikeNet mobile sensing system for cyclist experience mapping. SenSys'07. 87-101.

[4] Anand, A, Manikopoulos, C., Jones, Q., and Borcea, C. 2007. A Quantitative Analysis of Power Consumption for Location-Aware Applications on Smart Phones. ISIE'07. 1986-1991.

[5] Rodriguez, M.D., Favela, J., Martinez, E.A., and Munoz, M.A. 2004. Location-aware access to hospital information and services. IEEE Trans. on Information Technology in Biomedicine, 8, 4. 448-455.

[6] UCSD, "The CitiSense Project," https://sosa.ucsd.edu/ confluence/display/CitiSensePublic/CitiSense

[7] Zappi, P., Farella, E., Benini L., Lombriser, C., Stiefmeier, T., Roggen, D., and Tröster, G. 2008. Activity recognition from on-body sensors: accuracy-power trade-off by dynamic sensor selection. EWSN'08. 17-33.

[8] Petre, L. 2008. Energy-Aware Middleware. ECBS'08. 326-334.

[9] Kang, S., Lee, J., Jang, H., Lee, Y. Park, S., Song, J., 2010. A Scalable and Energy-Efficient Context Monitoring Framework for Mobile Personal Sensor Networks. In IEEE Transaction on Mobile Computing, vol.9, no.5, pp.686-702.

[10] Roy, A., Das Bhaumik, S.K., Bhattacharya, A., Basu, K., Cook, D.J., and Das, S.K. 2003. Location aware resource management in smart homes. PerCom'03. 481- 488.

[11] HTC, "HTC Aria", http://www.htc.com/us/products/aria-att

[12] Carroll A. and Heiser, G. 2010. An Analysis of Power Consumption in a Smartphone. USENIX'10. 271-284.

[13] Fan, X., Ellis, C. S. and Lebeck, A. R. 2003. Modeling of DRAM Power Control Policies Using Deterministic and Stochastic Petri Nets. PACS'03. 130-140.

[14] Beach, A., Gartrell, M., Xing, X., Han, R., Lv, Q., Mishra, S., and Seada, K. 2010. Fusing mobile, sensor, and social data to fully enable context-aware computing. HotMobile'10. 60-65.

[15] Rahmati,A., and Zhong, L. 2007. Context for Wireless: Context-sensitive energy-efficient wireless data transfer. In MobiSys'07. 165-178.

[16] Ananthanarayanan, G., and Stoica, I. 2008. Predicting Wi-Fi Availability using Bluetooth and Cellular Signals. In Special Interest Group on Data Communication. SIGCOMM 2008.