Anda di halaman 1dari 13

Blue-Fi: Enhancing Wi-Fi Performance using Bluetooth


Ganesh Ananthanarayanan Ion Stoica

UC Berkeley UC Berkeley
Berkeley, California Berkeley, California

ABSTRACT T’put Transfer Idle Scan Range

(J/MB) (W) (W) (m)
Mobile devices are increasingly equipped with multiple net-
Cellular few 100 100 0* 0* 500
work interfaces with complementary characteristics. In par-
ticular, the Wi-Fi interface has high throughput and trans-
Wi-Fi 11-54 5 0.77 1.29 100
fer power efficiency, but its idle power consumption is pro-
hibitive. In this paper we present, Blue-Fi, a sytem that
Bluetooth 700 kbps 0.1 0.01 0.12 10
predicts the availability of the Wi-Fi connectivity by using
a combination of bluetooth contact-patterns and cell-tower ( * - The cellular interface is typically always on.)
information. This allows the device to intelligently switch
Table 1: Various interface characteristics
the Wi-Fi interface on only when there is Wi-Fi connectiv-
ity available, thus avoiding the long periods in idle state and
significantly reducing the the number of scans for discovery.
22] (see Table 1). Ideally, one would like to intelligently
Our prediction results on traces collected from real users
leverage the strengths of these interfaces to optimize the ap-
show an average coverage of 94% and an average accuracy
plication performance and minimize the power consumption.
of 84%, a 47% accuracy improvement over pure cell-tower
Among the typical network interfaces found in today’s mo-
based prediction, and a 57% coverage improvement over the
bile devices, Wi-Fi provides arguably the best combination
pure bluetooth based prediction. For our workload, Blue-Fi
of throughout, range, and power efficiency for data trans-
is up to 62% more energy efficient, which results in increasing
fers. On the downside, Wi-Fi is the least power efficient in
our mobile device’s lifetime by more than a day.
idle state and incurs a high overhead when scanning for new
networks (e.g., anecdotes with iPhone [2, 4]). Thus, ideally,
Categories and Subject Descriptors one should use Wi-Fi whenever it is available, switch it off
C.2.1 [COMPUTER-COMMUNICATION NETWOR when it is not, and avoid scanning whenever possible. This
KS]: [Network Architecture and Design] is particularly useful for applications that use the network
periodically like Microsoft Pocket Outlook [5]. It is easy to
General Terms see that implementing such a strategy requires one to effi-
ciently detect the Wi-Fi availability without switching it on.
Design, Algorithms, Measurement, Performance In this paper, we address this challenge by presenting
Blue-Fi, a simple scheme that predicts the availability of
Keywords Wi-Fi by using bluetooth contact-patterns and cell-tower in-
Bluetooth, Wi-Fi, location, context-awareness, energy-efficiency, formation. The main observation behind our scheme is that
mobile device users tend to repeatedly encounter the same set of bluetooth
devices and cell-towers. Examples of such scenarios include
1. INTRODUCTION a bluetooth mouse or spouse’s mobile device while at home,
colleagues’ devices or printer at work, shop-owner’s mobile
Today’s mobile devices, such as smartphones, are increas-
device at his regular coffee place and even fellow commuters.
ingly equipped with multiple network interfaces, including
Most previous work on context-aware applications has fo-
Wi-Fi, bluetooth, and, of course, cellular interfaces. These
cused on using GPS and cell-tower information. While GPS
interfaces have widely different, often complementary, char-
is highly accurate, it is power hungry (e.g., [24] and [8]),
acteristics in terms of throughput, range, and power [12, 21,
and thus not a good fit for our problem where the focus is
on reducing the power consumption. Furthermore, GPS can-
not be used indoors, where Wi-Fi connectivity is arguably
Permission to make digital or hard copies of all or part of this work for most likely to be available. Cell-tower information has been
personal or classroom use is granted without fee provided that copies are successfully used for inferring contexts [12, 10] but as we
not made or distributed for profit or commercial advantage and that copies demonstrate in this paper, its inaccuracy reduces its appli-
bear this notice and the full citation on the first page. To copy otherwise, to cability to Wi-Fi prediction.
republish, to post on servers or to redistribute to lists, requires prior specific Despite bluetooth’s power efficiency, so far it has been
permission and/or a fee.
MobiSys’09, June 22–25, 2009, Kraków, Poland. largely ignored by the context-aware applications. There
Copyright 2009 ACM 978-1-60558-566-6/09/06 ...$5.00. are several reasons behind this state of affairs. First, blue-

tooth has a much lower range compared to other ubiquitous (Timestamp, {Bluetooth devices}, {Cell Towers}, {Wi-Fi
network technologies (see Table 1). Second, the discovery networks}). Bluetooth devices are identified by their MAC
process of bluetooth devices is time-consuming and could addresses, cell-towers are identified by the tower identifier
take as long as 10 seconds [14]. Finally, unlike cell-towers and Wi-Fi access points by their SSID/BSSID. The mobile
and access points that are stationary, bluetooth devices are device then uses its log to predict Wi-Fi connectivity.
primarily carried by users and hence mobile. The key question we need to address is: how accurately
Blue-Fi leverages bluetooth’s low range to its advantage can a bluetooth device or a cell-tower predict the Wi-Fi avail-
to achieve high prediction accuracy of Wi-Fi network avail- ability? We first discuss the reliability of bluetooth predic-
ability. However, the high accuracy comes at a price: cov- tion. If all bluetooth devices were fixed, predicting Wi-Fi
erage. The smaller the communication range is, the less availability would be easy. Given a bluetooth device b, just
likely for Blue-Fi to find a nearby bluetooth device. To check whether there is a log entry containing b and a Wi-Fi
increase the prediction coverage, Blue-Fi complements blue- access point: if yes, then we predict the Wi-Fi availability.
tooth with cell-tower information. To alleviate bluetooth’s Unfortunately, in practice many bluetooth devices are mo-
high discovery time, Blue-Fi implements periodic discovery bile (e.g., phones, notebooks), so they cannot straightaway
and uses the latest discovered list of bluetooth devices. Fi- be used to predict fixed Wi-Fi access points.
nally, despite the fact that many bluetooth devices are mo- To account for the mobile bluetooth devices we consider
bile, our results indicate that there is enough repeatability in the correlation between the observations of the bluetooth
contact-patterns of bluetooth devices that can be leveraged devices and Wi-Fi access points. Intuitively, we consider
to provide accurate context information, especially indoors. a bluetooth device b to be a reliable predictor for Wi-Fi
In its basic form, Blue-Fi requires no distributed infras- connectivity, if most of the log entries in which b appears
tructure, or running complex distributed protocols. Each contain at least a Wi-Fi access point.
mobile device periodically logs locally the bluetooth devices,
cell-towers, and Wi-Fi access points in its proximity, and L Log of all network signals over time
later uses this information to predict Wi-Fi connectivity. To τ Threshold for a bluetooth device or
speed up the learning process, we also propose variants of cell-tower to be a reliable predictor
Blue-Fi which employ peer-to-peer protocols or centralized predictcell Cell-towers that are reliable predictors
web services to share the logs. predictBT Bluetooth devices that are reliable predictors
We make the following contributions. First, we advocate SBT Set of all bluetooth devices in the log
the use of bluetooth contact patterns for context inference, Wc Connectable Wi-Fi networks
effectively providing a low-powered location system. In do- EDT Threshold of euclidean distance below
ing so, we address the limitations of the bluetooth devices which a user is considered stationary
with respect to low range, high discovery time and mobil-
ity. Second, we leverage the complementary properties of Table 2: List of notations.
bluetooth and cell-tower to improve the prediction of Wi-Fi
availability. By combining bluetooth and cell tower based More precisely, let SBT be the set of bluetooth devices in
predictions we obtain an average coverage of 94% and an log L and Wc be the set of Wi-Fi networks which provides
average accuracy of 84%, a 47% improvement in accuracy connectivity to the device (we describe mechanisms to ob-
over the pure cell-tower based prediction and a 57% improve- tain Wc in Section 7.1). For each Bluetooth device, BTi 
ment in coverage over the pure bluetooth based prediction SBT , we compute the predictability of BTi as n(L, BTi , Wc )/
scheme. This translates to an energy efficiency of up to 62% n(L, BTi ), where n(L, BTi , Wc ) is the number of entries in
and increase in our mobile device’s lifetime by more than a L when BTi was present and at least one of the networks
day. Finally, we analyze the benefits of collaborative predic- in Wc was present, and n(L, BTi ) is the number of entries
tion including security and privacy concerns. in L when BTi was present. Predictability is a confidence
The rest of the paper is organized as follows. Section 2 measure of how much a device’s presence indicates Wi-Fi
describes the system’s functioning including the prediction connectivity. If predictability is greater than a threshold τ ,
schemes. Section 3 deals with the problem of bluetooth dis- we add BTi to predictBT . In Section 2.2, we present a simple
covery. Algorithms to infer special bluetooth devices - land- algorithm to set this threshold.
mark devices and mobile accessories - are presented in Sec- While a single bluetooth discovery or Wi-Fi scan can po-
tion 4. Collaborative prediction, along with the security and tentially miss some bluetooth devices or access points respec-
privacy implications, is addressed in Section 5. Evaluation tively, we believe the errors due such misses are minimized
of the different aspects of the system is done in Section 6. or negligible when predictBT is calculated over a sufficiently
We present improvements to Blue-Fi - identifying browser- large set of log entries over time.
authenticated Wi-Fi networks, and multihop bluetooth dis- The algorithm to obtain predictcell is exactly the same
covery - in Section 7. We discuss an indoor monitoring sys- as the algorithm described above, with bluetooth devices
tem, and enhanced prediction models in Section 8. Section 9 replaced by cell-towers. predictcell contains the list of cell-
presents related work and contrasts it with Blue-Fi, and we towers at places where the device has Wi-Fi connectivity.
conclude in Section 10.
2.1 Prediction Schemes
2. PREDICTING WI-FI AVAILABILITY In this section, we describe the Wi-Fi prediction schemes
In a nutshell, Blue-Fi predicts the Wi-Fi network avail- based on bluetooth and cell-tower signals. The Wi-Fi pre-
ability by leveraging existing cell-towers and bluetooth de- diction schemes predict Wi-Fi availability using predictBT
vices. Each mobile device periodically logs all the network and predictcell . We evaluate the prediction schemes on two
signals in a log L, locally. The log entries are of the form metrics:

Figure 2: Predict-Signal Matrix

predictcell and predictBT . High values of τ increase accu-

Figure 1: Hybrid Prediction Scheme racy in prediction but lower coverage. On the other hand,
low values of τ result in inaccurate prediction.
We determine the appropriate value of τ using a predict-
• Coverage: The fraction of existing Wi-Fi access points signal matrix (see Figure 2). The matrix models the cases
that are predicted. when our scheme predicts the availability or lack of Wi-Fi
connectivity versus reality. The variables s and s̄ indicate
• Accuracy: The fraction of the Wi-Fi predictions that the presence and absence of Wi-Fi availability respectively,
are correct. and p and p̄ indicate cases when Blue-Fi predicts the avail-
ability of Wi-Fi. Cases 2 and 3 correspond to failures sce-
Bluetooth based Prediction: We predict that Wi-Fi narious. In case 2, Blue-Fi predicts the absence of Wi-Fi
is available if any of the bluetooth devices in predictBT is connectivity even when it is present resulting in the device
currently nearby. If none of the nearby devices are present using the power-inefficient cellular interface. In case 3, the
in predictBT , but there are some present in SBT , then we mobile device wastes power probing for Wi-Fi networks.
predict that there is no Wi-Fi connectivity. If none of the For a data transfer of size F , the expected wastage in
nearby bluetooth devices have been seen before, the predic- energy is Ewaste = p3 ∗ Ep + p2 ∗ (F ∗ ec −
tion scheme offers no prediction (we treat this as absence of (Ep + F ∗ ew )) where p2 and p3 are the probabilities
Wi-Fi connectivity). Bluetooth based prediction is expected of cases 2 and 3, Ep is the energy consumed in probing for
to exhibit high accuracy but low coverage as the range of Wi-Fi networks, and ec and ew are the energy consumed
bluetooth devices is much smaller than Wi-Fi (see Table 1). per data unit for transferring data using the cellular and
For example, a mobile device within Wi-Fi coverage may Wi-Fi interfaces respectively. Ep , ec and ew are constants.
fail to predict Wi-Fi availability, if the bluetooth devices in Automatically measuring and calibrating the values of Ep ,
predictBT are too far to be detected. ec and ew is orthogonal to Blue-Fi’s objectives and has been
Cell-tower based Prediction: Cell-tower based predic- addressed in prior work [17].
tion is based on the presence of currently visible cell-towers For a given transfer size F , the energy wastage Ewaste is
in predictcell . As discussed in Section 1, the ranges of cel- a function of p2 and p3 only. Next, we show that p2 and p3
lular signals is much higher than Wi-Fi networks resulting can be expressed as functions of τ alone. Note that with the
in coarse-grained predictions. As a result, cell-tower based notations in the predict-signal matrix, Accuracy is P (s | p)
prediction achieves high coverage but less accuracy. and Coverage is P (p | s).
Hybrid Prediction: This scheme uses a combination of
bluetooth and cell-tower based predictions. More precisely, p2 = P r(s | p̄)
it uses bluetooth prediction first, and falls back to cell-tower = P r(p̄ | s) (P r(s)/P r(p̄))
= (1 − P r(p | s)) (P r(s)/(1 − P r(p)))
based prediction when the former fails to predict. This way, = (1 − P r(s | p) ∗ P r(p)/P r(s)) (P r(s)/(1 − P r(p)))
hybrid prediction achieves both of best worlds by combining = (1 − Accuracy ∗ P r(p)/P r(s)) (P r(s)/(1 − P r(p)))
the accuracy of bluetooth prediction with the coverage of = (P r(s) − Accuracy ∗ P r(p))/(1 − P r(p))
the cell-tower prediction. Figure 1 shows the flow diagram
of the hybrid prediction scheme. The scheme starts with dis- p3 = P r(s̄ | p)
covering the bluetooth devices that are currently nearby and = 1 − P r(s | p)
= 1 − Accuracy
checks whether any of them is present in either predictBT
or SBT . If yes, it uses bluetooth prediction. Otherwise, it P r(s) is the percentage of times when Wi-Fi networks
resorts to cell-tower based prediction. from Wc was present in the log L and is a constant (not
In the rest of the paper, we assume that if Blue-Fi does dependent on τ ). Accuracy and P r(p) - the percentage of
not make any prediction, then there is no Wi-Fi connectivity, times when we found at least one bluetooth device from the
i.e., “No Prediction” is equivalent to “No Wi-Fi Connectiv- set predictBT or one cell-tower from the set predictcell - are
ity” (see Figure 1). dependent only on τ . Therefore p2 and p3 are essentially
functions of only τ . Note that p2 and p3 can also be ex-
2.2 Prediction Reliability Threshold (τ ) pressed as functions of coverage instead of accuracy.
In this section, we describe an algorithm to select the Hence, for a given transfer size, Blue-Fi finds the value of
appropriate prediction reliability threshold, τ , to populate τ that minimizes Ewaste .

Figure 3: Variation in Euclidean Distance when a
device is stationary, walking (5 km/hr) and driving
(30 km/hr).

Bluetooth discovery takes considerable time (over 10 sec-
onds [14]) causing high latency in prediction and subsequent Figure 4: Periodic Bluetooth Discovery combined
application data transfers. We note that when the user is with Euclidean Distance Estimation
not moving at high speeds, Wi-Fi connectivities are likely
to be relatively stable. Based on this observation, Blue-Fi
scans for nearby bluetooth devices periodically, and stores 3.1 Euclidean Distance Threshold (EDT )
and uses the latest discovered list for prediction. Periodic The threshold of euclidean distance for inferring station-
discovery reduces the latency of prediction using Blue-Fi. ary periods, EDT , depends on the fluctuation of cellular
The period of discovery is appropriately set depending on signal in a user’s surroundings and hence fixing a static and
the Wi-Fi scanning frequency and the ratio of power con- uniform value across all users is problematic. We present
sumption of Wi-Fi scanning to Bluetooth discovery. a simple heuristic by which devices can calibrate this value
However, periodic discovery could be wasteful in times themselves.
when device’s Wi-Fi scanning frequency is low or the user The heuristic is based on estimating the times when the
is stationary. In such cases, Wi-Fi characteristics are not device was stationary and calculating the average euclidean
expected to change. For example, a user sitting in his office distance between the cell-tower fingerprints in those periods.
with his office’s Wi-Fi connectivity is unlikely to see a change We consider being in the proximity of a stationary bluetooth
in his Wi-Fi prediction. In cases when the user is stationary, device (we discuss them in Section 4.2) or being connected
the last predicted Wi-Fi connectivity can be used without to the same Wi-Fi access point (BSSID) as an indication
wasting power on updating the visible bluetooth devices. of being stationary. Note that we are only interested in
Observations from prior work [25, 23] and our own exper- inferring times when a user’s Wi-Fi connectivity is relatively
iments indicate that cell tower signal strength values have stable as opposed to a user being completely stationary in
low variance when the device is stationary. A cell-tower fin- the traditional sense. From the log, L, devices can calculate
gerprint is a set of tuples containing the cell-tower identifier the average euclidean distance during stationary periods and
(CT ) and signal strength (SS). Given two cell-tower fin- fix the value of EDT appropriately.
gerprints, Fi = {(CTi1 , SSi1 ), ..., (CTin , SSin )} and Fj =
{(CTj1 , SSj1p), ..., (CTjn , SSjn )}, the euclidean distance is
defined as (SSi1 − SSj1 )2 + ... + (SSin − SSjn )2 ). Cell- 4. SPECIAL BLUETOOTH DEVICES
towers that are present in only one of the fingerprints are There are two classes of bluetooth devices that are espe-
not used in the calculation. High values of euclidean dis- cially interesting - landmark devices and mobile accesories.
tance indicate low similarity between the fingerprints. Landmark devices are devices that are stationary like a blue-
Figure 3 plots the average euclidean distance between con- tooth mouse, keyboard or printer, and mobile accessories
secutive readings in our experiments when a user is station- correspond to devices like bluetooth headsets. This section
ary, walking (5 km/hr) and driving slowly (30 km/hr). Con- describes ways to identify these two categories of devices
secutive measurements were taken a minute apart. As shown from the log, L. Landmark devices can be shared among
in Figure 3, the euclidean distance is a good indicator of users as they are not dependent on any single user’s move-
whether a user is stationary or mobile. ments. Mobile accessories are not reliable indicators of Wi-
Blue-Fi calculates the average euclidean distance since the Fi availability and should be removed from the logs.
last Wi-Fi prediction and if it is below a threshold, EDT , Our identification technique is based on correlating the
periodic bluetooth discovery is not performed and the last Wi-Fi access points and cell-towers, both of which are es-
made prediction of Wi-Fi availability is used. While we note sentially stationary, with the bluetooth devices. In other
that there are alternatives to the average euclidean distance words, we intend to capture the variation in the difference
like Spearman rank correlation coefficient [6] and common locations at which a bluetooth device was sighted. To this
number of cell-towers, our metric is suited for our purpose end, we first introduce the notion of diversity for a device
and produces good results (see Section 6.4). Figure 4 illus- and then use that to identify landmark devices and mobile
trates the steps in Blue-Fi’s periodic bluetooth discovery. accessories.

4.1 Diversity 4.2 Landmark Devices
Diversity captures the variance among the list of loca- Intuitively, a landmark device is one that a user discovers
tions a device is sighted. For every bluetooth device in SBT , only at one location (e.g., home) and always discovers that
we extract the list of Wi-Fi and cell-tower signatures that device when he is at that location. Precisely, a device BT
were co-sighted along with it. A Wi-Fi signature is a list of is classified as landmark if it satisfies two properties: (1) Its
BSSIDs, and a cell signature is a list of cell-tower identifiers. diversity is sufficiently low, and (2) Whenever a signature
Note that the list of signatures can contain repetitions. We that is similar to its median, Wm , occurs in the log, BT
use the list of signatures as an indicator of the locations the should also be present. Property (1) ensures that BT is seen
bluetooth device was present. only at one location. Property (2) ensures that whenever the
We define similarity between two signatures A and B as user is at that location, BT is seen.
|(A ∩ B)| / |(A ∪ B)|. Clearly, the higher the number of
intersecting wireless access points or cell-towers, the greater 4.3 Mobile Accessories
the value of similarity. Mobile accessories are not reliable indicators of Wi-Fi
We use similarity to now calculate diversity for a list of availability and should be removed from the logs. We treat
signatures. Let W = {Ws } be the list of Wi-Fi signatures devices that occur in a high fraction of the log entries to
for a bluetooth device. We identify the median of W us- be mobile personal accessories and unreliable indicators of
ing K-Medians clustering [18]. K-Medians is suited for our a user’s location.
purpose as it can perform clustering even when the “points”
to be clustered are not in Euclidean space, as in our case 5. COLLABORATIVE PREDICTION
where we have to cluster signatures. K-Medians clustering This section explores the scenarios, benefits and issues
exhaustively evaluates the suitability of each Ws in W to if devices were to collaborate and share information about
be the median and picks the best. Table 3 describes the Wi-Fi availability. An example scenario when such sharing

algorithm to pick the median and calculate the diversity of is beneficial is when a user goes to a new place with no prior
a given list of signatures. The presence of repetitions in the context (improve coverage). Sharing essentially speeds-up
signature list, W , ensures that the signatures are automati- the learning process. We present two sharing approaches −
cally weighted during the calculation of diversity. Diversity peer-to-peer and global − and also discuss their security and
is calculated for every device and varies between 0 and 1. privacy implications.
Diversity for a list of cell-tower signatures is exactly similar
with Wi-Fi BSSIDs replaced by cell-tower identifiers. 5.1 Peer-to-peer Sharing
Devices query each other over bluetooth for the availabil-
Input: List of signatures, W = {Ws } ity for Wi-Fi networks. Any device that is currently using
Output: Median, Wm and Diversity, D the Wi-Fi network can respond with details including the
bandwidth and authentication mechanisms. In fact, users
Variables: can exchange more useful information like “I performed a 1
Aggregate Similarity, agg sim MB download 30 seconds back and got a 10 Mbps through-
Highest Similarity, best sim ← 0 put”. Such sharing can be used for collaborative selection of
Current Median, Wmc access points. Hence, presence of at least one bluetooth de-
vice can potentially result in a device knowing about Wi-Fi
Algorithm: connectivity.
for each Wsi in W We can extend the hybrid prediction scheme to include
Wmc ← Wsi peer-to-peer querying wherein the device queries its neigh-
agg sim ← 0 bors when bluetooth prediction is inconclusive and before
for each Wsj in W resorting to cell-tower based prediction.
agg sim += Similarity(Wsi , Wsj )
end for 5.2 Global Sharing
if(agg sim > best sim) The alternative to peer-to-peer sharing is global sharing
best sim ← agg sim facilitated through a central service. Devices, when using
Wm ← Wmc the Wi-Fi network, periodically upload entries − (Times-
end for tamp, {Bluetooth devices}, {Cell Towers}, {Wi-Fi network,
{Characteristics}}) − to a centralized server. Characteris-
for each Ws in W tics specify if a network requires authentication along with
D += Similarity(Ws , Wm ) performance parameters like throughput and latency. While
end for information about security of a network (e.g., WEP encryp-
tion) is specified for all Wi-Fi beacons, performance param-
D ← D/|W | eters are provided only for the associated network. Servers
index these entries by bluetooth devices and cell-tower iden-
return (D and Wm ) tifiers for efficient retrieval.
Any device that wants to know about its current Wi-Fi
float Similarity (Signature W1 , Signature W2 ) availability can query the central server by supplying the list
float similarity = |(W1 ∩ W2 )| / |(W1 ∪ W2 )| of currently visible bluetooth devices and cell-towers. The
return similarity server responds by matching the bluetooth devices and cell-
towers in its database and returns the Wi-Fi networks along
Table 3: Diversity of a signature list with their characteristics. Communication with the server

happens over the cellular interface (data channel or simpler 6. EVALUATION
options like SMS). The details of matching application needs We collected logs from twelve volunteers for a period of
with the best suited Wi-Fi network are beyond the scope of two weeks each. In this section, we use the logs to evaluate
this paper. our prediction schemes for accuracy and coverage, appro-
Matching cell-towers need not involve the timestamp field priate threshold selection, effects of periodic discovery and
in the entries in the server because they are stationary. Blue- energy efficiency for our workload. We also evaluate our
tooth devices are mostly dynamic and should be matched algorithms for identifying landmark devices and mobile ac-
only if the time of uploading of the log entry and time of cesories. Finally we quantify the benefits of collaborative
querying is within a bounded interval. Note that the server prediction. We briefly summarize our results here.
can process the logs and obtain the set of landmark devices
as described in Section 4. Matches with landmark devices 1. Hybrid prediction scheme produces high coverage (93.5
need not consider the timestamp. %) and accuracy (84.2%), a 47% improvement in ac-
The ability to leverage landmark devices makes the global curacy over the pure cell-tower based prediction and
sharing option potentially more useful than peer-to-peer shar- a 57% improvement in coverage over pure bluetooth
ing; landmark devices are not expected to do peer-to-peer based prediction scheme.
communication. But on the other hand, peer-to-peer shar-
ing does not need any infrastructure support and hence is 2. Periodic discovery results in negligible reduction in ac-
more readily deployable. curacy and coverage.
3. Energy consumption reduces by up to 62% for our
5.3 Privacy and Security workload using Blue-Fi’s prediction techniques.
In this section, we discuss the security and privacy con-
cerns related to global and peer-to-peer sharing. A trusted 4. Collaborative prediction through sharing improves the
service model, like the emerging location-based services [1], coverage by up to 36.2%.
alleviates these concerns for the case of global sharing. Next,
we discuss the security and privacy issues in peer-to-peer 6.1 Log Collection
sharing, and compare it with the global sharing model. Twelve volunteers were given i-mate PDAs programmed
to log the Wi-Fi, bluetooth and cell-tower signals as men-
1. Intrusion: Peer-to-peer collaboration requires devices tioned in Section 2, every minute. The i-mate runs Win-
talking to other devices around them, and users might dows Mobile 5.0 and is equipped with a Class-2 bluetooth
be wary of such communication as it might potentially interface. We performed Wi-Fi scanning using a library
lead to intrusions. Without any prior knowledge, there that was built using the Windows Driver Development Kit.
is no feasible filter that users can apply. Global sharing Bluetooth scanning happened using the open-source library,
requires the device to just talk to a trusted server. InTheHand [3]. GSM tower information was obtained by
reading a well-known memory location that has been ob-
2. Usage Pattern: If devices were to exchange details tained by the community via reverse-engineering [19]. The
about their network activity (size of download, time of volunteers were a mix of graduate students in Berkeley as
download), a malicious person can continually query well as working professionals in San Francisco Bay Area.
a device in the guise of knowing the Wi-Fi connec- Volunteers carried the PDA along for two weeks in their nor-
tivity characteristics, and end up obtaining the exact mal routine. No instance of the PDA’s battery discharging
access patterns of the device. This can be correlated to or any such incident that stopped the logging was reported.
make decisions about a user’s activity like streaming, Table 4 lists the details about the log. There are two
browsing etc. potentially leading to annoying and tar- noteworthy features from the logs:
geted advertisements. With global sharing, the central 1. The fraction of times when there was Wi-Fi connectiv-
server is expected to hide the identity of the users who ity from the preferred networks varies from 32.1% to
uploaded the data. 68%, indicating it is not ubiquitous and hence the need
for a mechanism for predicting Wi-Fi availability.
3. Industrial Espionage: Assume a scenario where cof-
fee shop A has Wi-Fi connectivity for its customers. A 2. The number of bluetooth devices and the fraction of
competitor, B, is looking to set up its store near A’s times they are visible (49.6% to 77.2%) are encourag-
and expects the same customers to visit B too. Now ing to base a prediction scheme on them.
B can have a person with a device sitting in A’s shop,
querying all the users visiting A continuously and ag- One half of the data was used for training and the other
gregating information about their Wi-Fi usages. B can half for testing.
use this information to appropriately provision its net- 6.2 Coverage and Accuracy
work, thereby obtaining information that A would not
have provided otherwise. The central server in global We measure the accuracy and coverage of using blue-
sharing will provide only representative information in tooth and cell-tower data individually for prediction, high-
response to queries. light their complementary properties and demonstrate the
benefits of using them in conjunction in the hybrid scheme.
While it is true that Wi-Fi frames themselves can be Bluetooth based Prediction: Figures 5 and 6 show the
passively scanned to get the required information, peer- accuracy and coverage (in %) for bluetooth-based prediction
to-peer sharing provides yet another way for this in- for the different users, for varying values of the prediction
formation to be leaked. reliability threshold, τ . We observe that accuracy is directly

User Duration (weeks) Cell-Towers* Wi-Fi SSIDs* Bluetooth Devices*
Preferred All
U1 2 121 (99%) 48.1% 490 (76.1%) 159 (50.1%)
U2 2 337 (99%) 56% 1156 (84.5%) 194 (71%)
U3 2 191 (98%) 32.1% 731 (68.1%) 181 (51.2%)
U4 2 287 (99%) 36% 846 (54.3%) 133 (66.4%)
U5 2 101 (99%) 68% 744 (81.2%) 99 (77.2%)
U6 2 202 (98%) 44.3% 553 (65.2%) 152 (49.6%)
U7 2 221 (99%) 55.8% 701 (77.4%) 201 (55.1%)
U8 2 188 (99%) 67.3% 553 (86.1%) 199 (51.8%)
U9 2 241 (98%) 58% 552 (71.1%) 288 (59.3%)
U10 2 302 (100%) 60.3% 774 (80.8%) 222 (63.7%)
U11 2 254 (99%) 53.2% 691 (63.7%) 241 (66.4%)
U12 2 198 (98%) 54.1% 801 (70.2%) 201 (62.2%)
( *− Fraction of times when at least one signal was observed on that interface)

Table 4: Details of logs collected by users

Figure 5: Accuracy of Bluetooth-based Prediction Figure 6: Coverage of Bluetooth-based Prediction

proportional to τ while coverage is inversely proportional

with τ . Due to the limited range of bluetooth, the average
accuracy of the prediction is high (87.25%), but the aver-
age coverage is low (61%). High accuracy but low coverage
leads to erroneous conclusions of lack of Wi-Fi availability
even in their presence. U4 and U5 have coverage of near
80% indicating that if a user visits very few places routinely,
bluetooth-based prediction can provide good coverage.
Cell-Tower based Prediction: Cell-tower based pre-
diction has complementary properties to bluetooth based
prediction. We see high average coverage of 93.5% but an
average accuracy of only 59.66% (Figures 7 and 8). Note
that U5 has a high accuracy of close to 80%. If a user’s Figure 7: Accuracy of Cell-Tower based Prediction
movements in a given area (when in range of a cell-tower)
is limited, then cell-tower based prediction can be sufficient.
Again, accuracy is directly proportional to τ while coverage of bluetooth contact-patterns and cell-tower information in
is inversely proportional with τ . High coverage and low ac- tandem for inferring contexts.
curacy lead to wastage of energy in unnecessary scanning
for Wi-Fi networks when there are none. 6.3 Prediction Reliability Threshold (τ )
Hybrid Prediction: The Hybrid Scheme combines the Picking the right threshold for a given file transfer influ-
advantage of both the earlier schemes and shows coverage ences the expected energy loss due to erroneous predictions.
of 93.5% and an accuracy of 84.2% for τ = 0.8. Figure 9 We measured how the energy loss, Ewaste , varies with the
plots the accuracy and coverage of the hybrid prediction prediction reliability threshold, τ , for different transfer sizes.
scheme. This is a 47% improvement in accuracy over the We used the values of ec = 100 J/MB, ew = 5 J/MB and Ep
pure cell-tower based prediction and a 57% improvement in = 5 J [12]. Figure 10 plots the results for U3. The optimal
coverage over pure bluetooth based prediction scheme. We threshold for the 100 KB transfer is 0.6 while it is 0.7 for
believe this to be an encouraging validation for the usage higher file transfers. Smaller transfer sizes are more sensi-

Figure 8: Coverage of Cell-Tower based Prediction

Figure 10: Expected Energy Wastage per KB for

different file transfers

User Preferred Networks Landmark Devices

U1 4.1 1.2
U2 5.6 1.6
U3 2.5 2.2
U4 6.7 1.1
U5 6.6 1.9
U6 3.2 0.9
Figure 9: Hybrid Scheme combines the Accuracy U7 2.1 2.2
and Coverage of Bluetooth and Cellular Contexts U8 4.4 1.8
U9 3.1 2.7
U10 4.9 1.2
tive to τ . Note the sharp variation in the value of Ewaste on U11 4.1 2.2
either side of τ = 0.6 for the 100 KB download. We see a U12 2.2 2.4
plateau after the value of τ exceeds 0.7 for transfer sizes of
1 MB and 10 MB. Table 5: Calibrated value of threshold for Euclidean
Across other users, we note that the optimal value of τ Distance, EDT , using BSSIDs of preferred Wi-Fi
varies between 0.5 and 0.7 for the 100 KB download, and networks (Wp ) and landmark bluetooth devices
between 0.7 and 0.8 for the 1 Mb and 10 MB transfer. In-
terestingly, for users U2 and U8, the optimal value of τ is
0.7 for all download sizes. 6.5 Energy Consumption
We evaluated the energy consumption of our schemes and
6.4 Periodic Discovery compared it to other schemes using a workload that models
Recall that Blue-Fi deals with bluetooth’s high discov- commonly used background applications like email synchro-
ery time by periodic discovery and using the latest discov- nizers and RSS feed readers. Background applications use a
ered list. It uses euclidean distance between cell tower fin- “pull” model by periodically polling the server for new data
gerprints to avoid bluetooth scans if a device is stationary. and synchronizing the copy on the mobile device with the
Threshold for variation in euclidean distance, EDT , was cal- server. Our workload consists of periodic synchronization
culated from the readings in L when the user was continu- activities of 100 to 200 KB. Using the full battery capacity
ously in the presence of the same Wi-Fi BSSID provided of the phone, we measure the number of such synchroniza-
by one of his preferred networks, and landmark devices (we tions performed before the phone runs out of power (total
fixed a minimum of five readings). Table 5 lists the EDT capacity of 16200 J).
values calibrated by the different devices using Wp BSSIDs We compare our hybrid prediction scheme with two com-
and landmark devices. True to intuition, we observe that monly used modes of network usage in practice: (a) use the
EDT values obtained using landmark bluetooth devices is cellular interface always (Ecellular ), and (b) scan and check
smaller than those obtained using Wi-Fi BSSIDs as refer- for Wi-Fi availability always − use Wi-Fi if available, cellu-
ence, as the range of bluetooth signals is significantly lower lar connectivity otherwise (EW i−F i ).
than Wi-Fi networks. Also, the diversity in values of EDT For our twelve users, we report encouraging improvements
across different users indicates that individual devices cali- of 19-62% compared to Ecellular and 20-40% compared to
brating themselves is better than fixing a uniform value. EW i−F i . In addition, we make the following observations:
We evaluate change in accuracy and coverage of bluetooth
based prediction when periodic discovery is performed for • Preferred Networks: The overall fraction of times
periods of 10 minutes, 15 minutes and 30 minutes. We see when a user has Wi-Fi connectivity also affects the
negligible reduction in accuracy and coverage because of us- gains compared to Ecellular and EW i−F i . The overall
ing the last discovered list compared to on-demand discov- coverage of Wi-Fi networks vary from 32.1% to 68% of
ery.(see Figures 11 and 12). the time for our twelve users. Figure 14 plots the gains

Figure 11: Reduction in Accuracy because of peri-
odic Bluetooth discovery
Figure 13: Gains compared to Ecellular and EW i−F i
vary with download size for U3

Figure 12: Reduction in Coverage because of peri-

odic Bluetooth discovery

compared to Ecellular and EW i−F i . When the fraction

of Wi-Fi coverage is low, the device ends up using the
cellular interface most times and hence the gains com-
pared to Ecellular are limited. Low fraction of Wi-Fi Figure 14: Savings compared to Ecellular and EW i−F i
coverage potentially leads to many wasteful scans and vary with overall Wi-Fi coverage
hence the utility of prediction is high (compared to
EW i−F i ). For a high fraction of Wi-Fi coverage, the
probability of a scan finding Wi-Fi connectivity is high 6.6 Landmark and Personal Devices
and hence the value of prediction is limited. Now we evaluate the clustering techniques for identifying
landmark and personal devices.
• Download Size: The performance of Wi-Fi predic- Landmark Devices: Users were given a bluetooth mouse
tion schemes vary depending on the amount of data to be kept at their house during the log collection process.
downloaded in every round. Figure 13 plots the gain This is the landmark device we aim to identify. Figure 15
for varying data sizes. The gains compared to Ecellular plots the diversity of the devices in our log for user U2. Even
reduces for small downloads. For small downloads, if a device is sighted only at the same location, the require-
the energy difference between using the cellular inter- ment that it has to be sighted every time the user is at that
face and Wi-Fi (when available) is limited. The gains location ensures that we correctly identify the landmark de-
compared to EW i−F i on the other hand, decreases for vices without false positives or false negatives.
large downloads. Large downloads amortize the energy Note that the majority of the devices have a very low value
wastage due to incorrect Wi-Fi scanning (i.e., scanning of diversity. This implies that users have a spatial corelation
when there is no Wi-Fi connectivity) and the advan- with the bluetooth devices they encounter − users see most
tage of using prediction schemes is limited. bluetooth devices only at select locations. We consider this
• Lifetime: The percentage increase in the device’s bat- as a validation of our technique for using bluetooth contact-
tery lifetime is the same as the percentage increase in patterns for providing context.
the number of synchronizations; the actual increase in Personal Devices: Along with our i-mate PDA, users
lifetime is a function of the synchronization frequency. carried a bluetooth headset with them most of the time and
For downloads of 100 KB, and synchronization fre- also their personal phones with bluetooth on always. Fig-
quencies of 1 minute and 5 minutes, we see that the ure 16 plots the fraction of entries in which a device occurs
lifetime of the device compared to Ecellular increases in the log for user U2. The two devices at the top correspond
by 1.05 to 5.23 hours, and 5.25 to 26.15 hours respec- to the user’s phone and bluetooth headset.
tively for our users. The corresponding numbers com-
pared to EW i−F i are 5.51 to 7.78 hours, and 27.5 to
6.7 Collaborative Prediction
38.9 hours. We make the assumption that the mobile Sharing of Wi-Fi information among devices helps im-
device is not expending power on any other activity. prove the coverage of Blue-Fi. In our logs, we calculated
the fraction of times when at least one bluetooth device

User τ = 0.8 p2p Global
U1 59.6% 68.3% 79.4%
U2 57.8% 72.2% 84.5%
U3 51.1% 78.4% 86.3%
U4 73.4% 78.8% 89.1%
U5 74.2% 81.8% 90.1%
U6 50.2% 66.5% 88.1%
U7 58.8% 71.1% 90.1%
U8 61.2% 69.4% 83.3%
U9 44.6% 65.1% 80.8%
U10 60.2% 70.8% 76.4%
U11 60.1% 72.2% 77.3%
U12 64.5% 70.9% 73.2%

Table 6: Improvement in Coverage because of peer-

Figure 15: Distribution of diversity of bluetooth de-
to-peer and global sharing

aims to learn new access points that provide connectivity,

i.e., open or unsecured wireless networks. However, the
presence of an open network does not automatically imply
connectivity. Often times, web requests are redirected to
authentication pages and this section aims to detect such
access points and discard them from prediction operations.
We choose not to employ an exclusive and dedicated server
for testing redirection as we believe it is more advantageous
in terms of deployability and scalability to use popular web-
servers already available on the Internet. Since majority of
the traffic on the Internet is HTTP based, we test redirec-
tion only for port 80. Note that the problem of picking the
best available access point is orthogonal to our work.
Figure 16: Distribution of the fraction of occurrence Redirection happens after the device has connected and
of bluetooth devices in the log obtained a DHCP assigned address. Web requests are ini-
tially redirected to an authentication webpage. But note
that redirections (HTTP response code 301) in themselves
(not a landmark device) was present in the log and a Wi-Fi do not represent browser-authentication as they are rou-
network from Wc was available. This represents the upper- tinely used in many web servers. We present two techniques
bound on the coverage achievable by peer-to-peer sharing. that detect browser-authentication by essentially making a
The accuracy with peer-to-peer sharing can be 100% if cor- request for a URL with known characteristics and comparing
rect information is provided by the nearby devices. Table 6 if the response is as expected.
lists these numbers and contrasts it with the coverage of
the bluetooth based prediction scheme. We compare it with 1. Response Size: Typically authentication webpages
bluetooth based prediction because it has the highest ac- are a few 10’s to 100’s of kilobytes. Blue-Fi makes a
curacy. Global sharing takes advantage of landmark devices request for a large file (say, a few megabytes) whose
too and hence achieves a higher coverage. Overall, we see an size is known, and check the size of the response.
encouraging result that coverage can reach up to 90% pre-
senting an overall improvement of up to 36.2%. Note that 2. Secure HTTP: Authentication webpages use the se-
the accuracy of the prediction depends on the correctness of cure version of HTTP (https://). A secure response
the information provided by the devices. (https://) for a request for an unsecured webpage (http:
//) indicates browser-level authentication.
7. ENHANCEMENTS TO BLUE-FI The techniques presented above are used to identify open
We now present useful additions to Blue-Fi − detect- networks that require browser-authentication and are dis-
ing open access points that are browser-authenticated, and carded. The rest can be passed to a Wi-Fi profiling system
multi-hop bluetooth discovery. like Virgil [9] for further connectivity tests (like throughput
and latency), the results of which would be stored in Wc .
7.1 Connectable Access Points We evaluate the accuracy of the identification of browser-
Recall that Blue-Fi needs to know the list of Wi-Fi net- authenticated Wi-Fi networks. We collected 19 redirected
works it is allowed to connect, Wc . Most devices store the set webpages from open access points in coffee shops, airports
of Wi-Fi networks to which it has connected in the past [13]. and universities. We used webpages from popular web-
Commonly used operating systems like Windows XP, Linux servers for our testing. Requests for large files were made
or Mac OS, this list is stored under “preferred networks”. Wc to podcast media files on popular news webservers (e.g., on
is initialized with this list. and The popular webservers have
In addition to predicting Wi-Fi availability, Blue-Fi also archiving facilities allowing access to old files. We also made

requests to popular webservers that did not require commu- of resouces like cooling (e.g., conference room A is uncom-
nication over secure HTTP ( and fortably cold) and Wi-Fi performance (e.g., Connectivity in
Checking using the size of the response is most accurate office B is spotty). Users log their sensed data (like Wi-
identifying all the pages in our redirected pages dataset. Fi performance) along with the visible reference bluetooth
The dataset had an average size of 62 KB and a mini- devices. The database of the locations of the reference blue-
mum and maximum of 13 KB and 132 KB respectively. tooth devices can be used to find the location of the point
Secure HTTP identifies only 68.4% of the authentication at which the data was collected and appropriate analysis
pages. This is because splash screens and end-user-license- can be conducted. We plan an extensive deployment and
agreement (EULA) notices are also served over unsecure evaluation of BlueDust.
HTTP. Differentiating between authentication webpages, and
EULA and splash screens is part of future work. 8.2 Prediction Models
Blue-Fi assumes that the sensing of each bluetooth device
7.2 Multi-hop Bluetooth Discovery is an independent event. But clearly, there is a rich opportu-
The discrepancy between the Wi-Fi and the bluetooth nity to incorporate mobility patterns and correlation across
ranges (i.e., 100m vs 10m) is the main reason behind the multiple bluetooth devices in our prediction models. Mo-
high accuracy of bluetooth prediction. On the downside, bility patterns can be used for predicting Wi-Fi availability
this discrepancy leads to low coverage, as a bluetooth de- — e.g., “Wi-Fi connectivity will be available ten minutes af-
vice may not be able to discover other nearby devices, even ter you spot the bluetooth device b”. Such predictions are
though all devices are in the same Wi-Fi coverage. useful for delay-tolerant applications to appropriately plan
One possible solution to reduce the discrepancy between their network activities in future.
the Wi-Fi and bluetooth ranges is to use multi-hop bluetooth Also, predictBT can be augmented to include sets of blue-
discovery. A device establishes a connection with each device tooth devices. This is useful in scenarios when, say, blue-
present in its immediate range, and queries for devices that tooth devices b1 and b2 are not sufficiently reliable on their
are in its range, recursively, for a preset number of hops. own, but whenever sensed together turn out to be reliable
This way, a mobile device will discover bluetooth devices indicators of Wi-Fi availability. Understanding the benefits
beyond its bluetooth range. This technique allows one to of such a correlated predictions on Blue-Fi’s coverage and
trade accuracy for coverage: the higher the number of hops accuracy is part of future work.
the higher the coverage and the lower the accuracy. The
overhead of multi-hop bluetooth discovery can be reduced 9. RELATED WORK
using the technique described in Section 3.
Improving energy efficiency of Wi-Fi networks is a long-
We performed an initial evaluation of the multi-hop dis-
standing problem in wireless networks and has been ap-
covery protocol by placing bluetooth monitoring devices in
proached from different directions. Techniques range from
a large lab area where the bluetooth devices tend to be clus-
protocol optimizations in the various layers of the network-
tered in two groups. We were able to see an increase of up
ing stack for a single Wi-Fi radio to techniques that leverage
to 234% in the number of discovered bluetooth devices by
multiple radios on the same device that often involve spe-
using multi-hop discovery.
cialized infrastructure elements.
The key idea of using a separate low-powered radio to
8. DISCUSSION wake up a high-powered radio was proposed in Wake-on-
In this section, we discuss extensions to Blue-Fi along with Wireless [15]. Wake-on-Wireless proposes the use of a sec-
fugure work: deploy “reference” devices to increase predic- ond special-purpose radio that serves as a wake-up chan-
tion coverage, and use correlated observation of bluetooth nel for a Wi-Fi radio. However the short range custom ra-
devices to improve accuracy. dio necessitates multiple intermediate proxies and presence
servers. Also, it requires significant modifications to exist-
8.1 Reference Bluetooth Devices ing mobile devices. While On-Demand-Paging [27] builds on
Blue-Fi depends fundamentally of the existence of land- this idea to use the widely available Bluetooth radio as the
mark bluetooth devices. However, as the coverage results low-powered channel, it still needs substantial infrastructure
suggest, in a non-trivial number of cases, the mobile device support in the form of specialized access points that have
is not able to identify such bluetooth devices. To address both Wi-Fi and bluetooth interfaces. Cell2Notify [28] uses
this issue, we propose BlueDust, a pervasive deployment of the cellular interface to wake up the Wi-Fi interfaces on an
reference bluetooth devices spatially distributed in a given incoming VOIP call using specialized servers. In contrast,
area. These devices are stationary and serve as means of Blue-Fi does not require any modification to the existing in-
providing contexts to the user. We envisage reference de- frastructure and can be deployed readily on mobile devices.
vices to be simple and inexpensive, for example bluetooth War-driving has been performed in prior work that col-
USB dongles or bluetooth mice. In contrast to Wi-Fi based lects and maps the Wi-Fi access points [7, 10, 24]. War-
indoor monitoring systems [20, 16], BlueDust’s advantage is driving is an expensive operation requiring investment of
power efficiency. Also, a clear benefit of such a system would time and money. While a few major cities have good war-
be increase in coverage for Blue-Fi. driving data, the majority of them have scarce or no map-
A simple and practical deployment mechanism, especially ping. In contrast, Blue-Fi adopts the approach of users auto-
for enterprise environments, is to plug in bluetooth dongles matically learning their own sorroundings and Wi-Fi avail-
into desktops [20]. Reference devices can also be placed us- ability. Also, war-driving data tend to become unreliable
ing systematic techniques. A database contains the locations and out-dated over time.
of the reference devices. Context-for-Wireless [12] aims to provide energy-efficient
BlueDust can be used for monitoring the spatial variation ubiquitous wireless connectivity, and their ideas of using

cell-tower information for predicting Wi-Fi availability are Fourteenth Annual International Conference on Mobile
closely related to our work. But as we demonstrate cell- Computing and Networking (Mobicom ’08), Sep 2008.
tower information is coarse-grained and can be beneficially [9] A. J. Nicholson, Y. Chawathe, M. Y. Chen, B. D.
combined with bluetooth contact-patterns for fine-grained Noble, and D. Wetherall. Improved access point
context localization. Intel Place Lab [19, 10] collected ex- selection. In Proceedings of The Fourth International
tensive network data on GSM cellular networks for the sake Conference on Mobile Systems, Applications, and
of positioning Wi-Fi hotspots. To the best of our knowl- Services (MobiSys ’06), June 2006.
edge, ours is the first work that demonstrates the benefits [10] A. LaMarca, Y. Chawathe, S. Consolvo, J. Hightower,
of using bluetooth contact-patterns (including mobile blue- I. Smith, J. Scott . Place lab: Device positioning using
tooth devices) for context localization. We are not aware of radio beacons in the wild. In Proceedings of The Third
prior work that demonstrates a beneficial convergence of the International Conference on Pervasive Computing
bluetooth, cellular and Wi-Fi interfaces. (Pervasive ’05), May 2005.
Our work has been supported by recent studies that show [11] A. Natarajan, M. Motani, and V. Srinivasan.
the existence of bluetooth contact-patterns in devices [11, Understanding Urban Interactions from Bluetooth
26]. CoolSpots [22] proposes the installation of “bluetooth Phone Contact Traces. In Proceedings of The Eighth
access-points”, which if deployed widely, strengthens our Passive and Active Measurement Conference (PAM
idea of using bluetooth contact-patterns for providing con- ’07), Apr 2007.
text information to mobile devices. [12] A. Rahmati and L. Zhong. Context-for-Wireless:
Context-Sensitive Energy-Efficient Wireless Data
10. CONCLUSIONS Transfer. In Proceedings of The Fifth International
Blue-Fi proposed using bluetooth contact-patterns as con- Conference on Mobile Systems, Applications, and
text identifiers for predicting Wi-Fi availability. The low Services (MobiSys ’07), Jun 2007.
range of bluetooth devices make them accurate predictors [13] B. Greenstein, R. Gummadi, J. Pang, M. Y. Chen, T.
of Wi-Fi availability; we compensated for their lack of cov- Kohno, S. Seshan, and D. Wetherall. Can ferris bueller
erage by using cell-tower identifiers. Our evaluation with still have his day off? protecting privacy in the
data collected from real users’ contact-patterns shows en- wireless era. In Proceedings of The Eleventh Workshop
couraging results in providing coverage as well as accuracy. on Hot Topics in Operating Systems (HotOS-XI), May
In addition, we also presented techniques to overcome 2007.
bluetooth’s high discovery time essentially using periodic [14] E. S. Hall, D. K. Vawdrey, and C. D. Knutson. RF
scanning. To speed up the learning process, Blue-Fi uses Rendez-Blue: reducing power and inquiry costs in
collaborative prediction through sharing of logs and Wi-Fi Bluetooth-enabled mobile systems. In Proceedings of
connectivity details. The Eleventh International Conf. on Computer
Communications and Networks (ICCCN ’02), Oct
[15] E. Shih, P. Bahl, and M. J. Sinclair. Wake on wireless:
We would like to thank Joe Hellerstein , Dilip Joseph, An-
An event driven energy saving strategy for battery
thony Joseph, Randy Katz, Venkat Padmanabhan, Lucian
operated devices. In Proceedings of The Eighth Annual
Popa and Matei Zaharia for their advice and guidance re-
International Conference on Mobile Computing and
garding various technical aspects as well as writing of this
Networking (MobiCom ’02), 2002.
paper. We would also like to thank our shepherd, Albrecht
Schmidt, and the anonymous reviewers for their comments [16] F. Dabek, R. Cox, F. Kaashoek, and R. Morris.
and suggestions. Vivaldi: A decentralized network coordinate system.
In Proceedings of the 2004 conference on Applications,
technologies, architectures, and protocols for computer
12. REFERENCES communications (SIGCOMM ’04), Aug 2004.
[1] Companies Eye Location-Services Market. [17] G. Ananthanarayanan, V. N. Padmanabhan, L. Ravindranath, and C. A. Thekkath. COMBINE:
[2] IMAP IDLE monitor/notifier. Leveraging the Power of Wireless Peers through Collaborative Downloading. In Proceedings of The
[3] In The Hand. Fifth International Conference on Mobile Systems,
[4] Is WiFi draining when not connected. Applications, and Services (MobiSys ’07), Jun 2007. [18] Ke Chen. On k-Median Clustering in High
330279.html. Dimensions. In Proceedings of The Seventeenth
[5] Microsoft Pocket Outlook. ACM-SIAM Symposium on Discrete Algorithms (SODA ’06), Jan 2006.
us/downloads/microsoft/office-outlook-mobile.mspx. [19] M. Chen, T. Sohn, D. Chmelev, D. Haehnel, J.
[6] Spearman Rank Correlation Coefficient. Hightower, J. Hughes, A. LaMarca, F. Potter, I. Smith, and A. Varshavsky. Practical
onCoefficient.html. Metropolitan-Scale Positioning for GSM Phones. In
[7] WIGLE: Wireless Geographic Logging Engine. Proceedings of The Eighth International Conference on Ubiquitous Computing (UbiComp ’06), Sep 2006.
[8] A. J. Nicholson and B. D. Noble. BreadCrumbs: [20] P. Bahl, R. Chandra, J. Padhye, L. Ravindranath, M.
Forecasting mobile connectivity. In Proceedings of The Singh, A. Wolman and B. Zill. Enhancing the security

of corporate wi-fi networks using dair. In Proceedings [25] V. Otsason, A. Varshavsky, A. LaMarca, and E. de
of The Fourth International Conference on Mobile Lara. Accurate GSM Indoor Localization. In The
Systems, Applications, and Services (MobiSys ’06), Seventh International Conference on Ubiquitous
Jun 2006. Computing (Ubicomp ’05), Sep 2005.
[21] P. Mohan, V. N. Padmanabhan, and R. Ramjee. [26] V. S. W. Wang and M. Motani. Adaptive contact
TrafficSense: Rich Monitoring of Road and Traffic probing mechanisms for delay tolerant applications. In
Conditions using Mobile Smartphones. In Microsoft Proceedings of The Thirteenth Annual International
Technical Report, MSR-TR-2008-59, Apr 2008. Conference on Mobile Computing and Networking
[22] T. Pering, Y. Agarwal, R. Gupta, and R. Want. (MobiCom ’07), Sep 2007.
Coolspots: Reducing the power consumption of [27] Y. Agarwal, C. Schurgers, and R. Gupta. Dynamic
wireless mobile devices with multiple radio interfaces. Power Management Using On Demand Paging for
In Proceedings of The Fourth International Conference Networked Embedded Systems. In Proceedings of The
on Mobile Systems, Applications, and Services 2005 Conference on Asia South Pacific Design
(MobiSys ’06), Jun 2006. Automation (ASP-DAC ’05), Jan 2005.
[23] T. Sohn, A. Varshavsky, A. Lamarca, M. Chen, T. [28] Y. Agarwal, R. Chandra, A. Wolman, P. Bahl, and R.
Choudhury, I. Smith, S. Consolvo, J. Hightower, W. Gupta. Wireless wakeups revisited: Energy
Griswold, and E. de Lara. Mobility detection using management for voip over wi-fi smartphones. In
everyday gsm traces. In Proceedings of The Eighth Proceedings of The Fifth International Conference on
International Conference on Ubiquitous Computing Mobile Systems, Applications, and Services (MobiSys
(UbiComp ’06), Sep 2006. ’07), Jun 2007.
[24] V. Bychkovsky, B. Hull, A. Miu, H. Balakrishnan, and
S. Madden. A measurement study of vehicular
internet access using in situ wi-fi networks. In
Proceedings of The Twelfth Annual International
Conference on Mobile Computing and Networking
(MobiCom ’06), 2006.