diversity in smartphone usage - microsoft.com · smartphone usage, user behavior 1. introduction...
TRANSCRIPT
Diversity in Smartphone Usage
Hossein FalakiCENS UCLA
Ratul MahajanMicrosoft Research
Srikanth KandulaMicrosoft Research
Dimitrios LymberopoulosMicrosoft Research
Ramesh GovindanCENS USC
Deborah EstrinCENS UCLA
Abstract ndash Using detailed traces from 255 users we con-duct a comprehensive study of smartphone use We char-acterize intentional user activities ndash interactions with thedevice and the applications used ndash and the impact of thoseactivities on network and energy usage We find immensediversity among users Along all aspects that we study usersdiffer by one or more orders of magnitude For instance theaverage number of interactions per day varies from 10 to 200and the average amount of data received per day varies from1 to 1000 MB This level of diversity suggests that mecha-nisms to improve user experience or energy consumption willbe more effective if they learn and adapt to user behaviorWe find that qualitative similarities exist among users thatfacilitate the task of learning user behavior For instancethe relative application popularity for can be modeled us-ing an exponential distribution with different distributionparameters for different users We demonstrate the value ofadapting to user behavior in the context of a mechanism topredict future energy drain The 90th percentile error withadaptation is less than half compared to predictions basedon average behavior across users
Categories and Subject Descriptors
C4 [Performance of systems] Measurement techniquesGeneral Terms
Measurement human factorsKeywords
Smartphone usage user behavior
1 INTRODUCTIONSmartphones are being adopted at a phenomenal pace but
little is known (publicly) today about how people use thesedevices In 2009 smartphone penetration in the US was 25and 14 of worldwide mobile phone shipments were smart-phones [23 16] By 2011 smartphone sales are projected tosurpass desktop PCs [25] But beyond a few studies that re-port on usersrsquo charging behaviors [2 17] and relative powerconsumption of various components (eg CPU screen) [24]many basic facts on smartphone usage are unknown i) howoften does a user interact with the phone and how long does
Permission to make digital or hard copies of all or part of this work forpersonal or classroom use is granted without fee provided that copies arenot made or distributed for profit or commercial advantage and that copiesbear this notice and the full citation on the first page To copy otherwise torepublish to post on servers or to redistribute to lists requires prior specificpermission andor a feeMobiSysrsquo10 June 15ndash18 2010 San Francisco California USACopyright 2010 ACM 978-1-60558-985-51006 $1000
an interaction last ii) how many applications does a userrun and how is her attention spread across them iii) howmuch network traffic is generated
Answering such questions is not just a matter of academicinterest it is key to understanding which mechanisms caneffectively improve user experience or reduce energy con-sumption For instance if user interactions are frequent andthe sleep-wake overhead is significant putting the phone tosleep aggressively may be counterproductive [8] If the userinteracts regularly with only a few applications applicationresponse time can be improved by retaining those applica-tions in memory [7] Similarly if most transfers are smallbundling multiple transfers [1 22] may reduce per-byte en-ergy cost Smartphone usage will undoubtedly evolve withtime but understanding current usage is important for in-forming the next generation of devices
We analyze detailed usage traces from 255 users of twodifferent smartphone platforms with 7-28 weeks of data peruser Our traces consist of two datasets For the first datasetwe deploy a custom logging utility on the phones of 33 An-droid users Our utility captures a detailed view of userinteractions network traffic and energy drain The seconddataset is from 222 Windows Mobile users across differentdemographics and geographic locations This data was col-lected by a third party
We characterize smartphone usage along four key dimen-sions i) user interactions ii) application use iii) networktraffic and iv) energy drain The first two represent inten-tional user activities and the last two represent the impactof user activities on network and energy resources Insteadof only exploring average case behaviors we are interestedin exploring the range seen across users and time We be-lieve that we are the first to measure and report on manyaspects of smartphone usage of a large population of users
A recurring theme in our findings is the diversity acrossusers Along all dimensions that we study users differ byone or more orders of magnitude For example the meannumber of interactions per day for a user varies from 10 to200 the mean interaction length varies from 10 to 250 sec-onds the number of applications used varies from 10 to 90and the mean amount of traffic per day varies from 1 to1000 MB of which 10 to 90 is exchanged during inter-active use We also find that users are along a continuumbetween the extremes rather than being clustered into asmall number of groups
The diversity among users that we find stems from thefact that users use their smartphones for different purposesand with different frequencies For instance users that use
users Duration Platform Demographic info Information loggedDataset1 33 7-21 weeksuser Android 16 high school students Screen state applications used
17 knowledge workers network traffic battery stateDataset2 222 8-28 weeksuser Windows 61 SC 65 LPU 59 BPU 37 OP Screen state applications used
Mobile Country 116 USA 106 UK
Table 1 An overview of the datasets in our study
games and maps applications more often tend to have longerinteractions Our study also shows that demographic infor-mation can be an unreliable predictor of user behavior andusage diversity exists even when the underlying device isidentical as is the case for one of our datasets
Among the many implications of our findings an over-riding one is that mechanisms to improve user experienceor energy consumption should not follow a one-size-fits-allmindset They should instead adapt by learning relevantuser behaviors otherwise they will likely be only marginallyuseful or benefit only a small proportion of users
We show that despite quantitative differences qualitativesimilarities exist among users which facilitates the task oflearning user behavior For several key aspects of smart-phone usage the same model can describe all users differ-ent users have different model parameters For instance thetime between user interactions can be captured using theWeibull distribution For every user the shape parameterof this model is less than one which implies that the longerit has been since the userrsquos last interaction the less likelyit is for the next interaction to start We also find thatthe relative popularity of applications for each user followsan exponential distribution though the parameters of thedistribution vary widely across users
We demonstrate the value of adapting to user behavior inthe context of a mechanism to predict future energy drainPredicting energy drain is an inherently challenging taskBursty user interactions at short time scales combined withdiurnal patterns at longer time scales lead to an energy con-sumption process that has a very high variance and is seem-ingly unpredictable We show however that reasonably ac-curate predictions can be made by learning the userrsquos energyuse signature in terms of aldquotrend tablerdquo framework For pre-dicting the energy use one hour in the future our predictorrsquos90th percentile error is under 25 Without adaptation andbasing the predictions on average behavior the 90th per-centile error is 60
2 DATA COLLECTIONOur work is based on two sets of data The first is a
high-fidelity data set that we gathered by deploying a cus-tom logger on the phones of 33 Android users The seconddata set consists of 222 Windows Mobile users across dif-ferent demographics Together these data sets provide abroad and detailed view of smartphone usage We leave forthe future the task of studying other smartphone platformssuch as iPhone and BlackBerry The characteristics of ourdatasets are summarized in Table 1
21 Dataset1Our first set of traces is from 33 Android users These
users consisted of 17 knowledge workers and 16 high schoolstudents Knowledge workers were computer science re-searchers and high school students were interns in a single
organization and were recruited by a third person on ourbehalf As stated in our study consent form the usersrsquo iden-tities were not revealed to us The participants were givenHTC Dream smartphones with unlimited voice text anddata plans We encouraged the users to take advantage ofall the features and services of the phones
The data was collected using a custom logging tool thatwe developed and deployed on the smartphones The loggerruns in the background and records a highly detailed viewof smartphone use including the state of the smartphonescreen start and end of incoming and outgoing voice callsthe time the user spends interacting with each applicationthe network traffic sent and received per application andthe battery level The Android OS provides mechanisms toaccess this information The logger keeps data records in alocal SQLite database on the phone and uploads them onlywhen the phone is plugged to the charger to minimize theimpact on the phone battery Our logging utility is availableto other researchers by request
The data was gathered between May and October 2009There is 7-21 weeks of data per user with the average being9 weeks
22 Dataset2Our second data set was collected by an organization that
was interested in investigating smartphone usability and ap-plication popularity This organization provided 222 userswith Windows Mobile smartphones from different hardwarevendors It also paid for their voice and data plans For rep-resentativeness the users were drawn from different demo-graphics as shown in Table 1 The demographic categorieswere defined based on what users stated as the primary moti-vation for using a smartphone Social Communicators (SC)wanted to ldquostay connected via voice and textrdquo Life PowerUsers (LPU) wanted ldquoa multi-function device to help themmanage their liferdquo Business Power Users (BPU) wanted ldquoanadvanced PC-companion to enhance their business produc-tivityrdquo Organizer Practicals (OP) wanted ldquoa simple deviceto manage their liferdquo The subjects were asked about theirintended use of the phone before the study and were catego-rized based on their answers To our knowledge the resultsof this study are not public
Traces were collected using a logger that recorded startand end time of each application This information waslogged using the ActiveApplication API call of the OS whichreports on the executable that currently has the foregroundwindow (with a callback for changes) Other details that ourcustom logger in sect21 records (eg network traffic and bat-tery level) were not logged in this study Thus this datasethas lower fidelity than the first one but it provides a viewof smartphone usage across a broader range of users
The traces were collected between May 2008 and January2009 There is 8-28 weeks of data per user with the averagebeing 16 weeks
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Voi
ce u
sage
rat
io
Dataset1Dataset2
Figure 1 Ratio of voice usage to total usage Thex-axis is user percentile and users are sorted in de-creasing order of ratio
23 Representativeness of conclusionsAn important concern for user studies such as ours is
whether the resulting conclusions represent the entire popu-lation There are two potential sources of bias in our data i)the users are not representative and ii) the measured usageis not representative We believe that our conclusions aregeneral The first concern is alleviated by the fact that asidefrom some quantitative differences we find remarkable con-sistency among users in the two datasets This consistencysuggests generality given that the two datasets are gatheredindependently on different platforms and Dataset2 was pro-fessionally designed to be representative
The second concern stems from the possibility that usersmay not be using the monitored smartphones as their pri-mary devices or that the usage during the monitoring periodmay not be normal All users in Dataset2 used the providedsmartphones as their primary devices We do not know thisaspect with certainty for Dataset1 but we understand fromanecdotal evidence that some users used these devices astheir only phones and others took advantage of the unlim-ited minutes and text plans We study voice usage as in-dicative of the extent to which users relied on the monitoreddevices Higher voice usage suggests use as primary phonesFigure 1 shows the ratio of time users spent in phone callsto total time spent interacting with the phone (sect3) We seethat the overall voice usage in Dataset1 was higher thanthat in Dataset2 in which all users used the phone as theirprimary device
Given that the monitored devices tended to be primaryand the long duration of the monitoring interval we con-jecture that our traces predominantly capture normal us-age Earlier work has pointed to the possibility of an initialadoption process during which usage tends to be differentthan long-term usage [19] To show that our traces are notdominated by the initial excitement of users or other specialevents that cause usage to be appreciably different from thenormal usage Figure 2 shows the average interaction timeper day (sect3) in the first and second halves of the datasets foreach user We see roughly similar usage in the two halvesDetailed investigation shows that the visible differences inthe averages of the two halves especially in Dataset1 arenot statistically significant Other measures of usage (egnetwork activity) look similar We do not claim that in-stances of abnormal usage are absent in the datasets but
110
100
1000
0 20 40 60 80 100User percentile
Act
ive
time
(min
utes
)
Dataset1 1st halfDataset1 2nd halfDataset2 1st halfDataset2 2nd half
Figure 2 Total interaction per day during the firstand second halves of study for each user Userswithin each dataset are sorted based on the interac-tion time in the first half The y-axis is log scale
the monitored period was long enough for our results to notbe impacted by such instances
3 USER INTERACTIONSWe begin our analysis by studying how users interact with
their smartphones independent of the application used Wecharacterize application use in the next section and the im-pact of user actions on network traffic and energy drain inthe following sections
We define an interaction interval also referred to as a ses-sion in this paper differently for each dataset In Dataset1we deem a user to be interacting with the phone wheneverthe screen is on or a voice call is active In Dataset2 aninteraction is defined as the interval that an application isreported to be on the foreground This includes voice callsbecause on Windows Mobile a special program (ldquocprogexerdquo)is reported in the foreground during voice calls
31 Interaction TimeFigure 3(a) shows a basic measure of user interactionmdashthe
number of minutes in a day that a user interacts with thesmartphone The plot shows the mean and the standarddeviation of this number for each user For visual clarityin such graphs we plot only the upper end of the standarddeviation plotting both ends occludes the other curves Theinterested reader can estimate the lower end since standarddeviation is symmetric around the mean
Dataset1 users tend to have more interaction minutes be-cause as we show later they tend to have longer interactionsessions while having a similar number of sessions Withineach dataset however there is an order of magnitude dif-ference among users In Dataset1 the lower end is only 30minutes in a day But the high end is 500 minutes which isroughly eight hours or a third of the day We are surprisedby this extremely high level of usage
Figure 3(a) also shows that users cover the entire rangebetween the two extremes and are not clustered into dif-ferent regions The lack of clusters implies that effectivepersonalization will likely need to learn an individual userrsquosbehavior rather than mapping a user to one or a few pre-defined categories
We examine two factors that can potentially explain theextent to which a user interacts with the phone but find that
110
100
1000
0 20 40 60 80 100User percentile
Act
ive
time
(min
utes
)
Dataset1Dataset2
(a) All users
High schoolKnowledge worker
110
100
1000
0 20 40 60 80 100User percentile
Act
ive
time
(min
utes
)
(b) Dataset1
SCOPBPULPU1
1010
010
00
0 20 40 60 80 100User percentile
Act
ive
time
(min
utes
)
(c) Dataset2
Figure 3 The mean and the upper end of the standard deviation of interaction minutes per day (a) Allusers in each dataset (b)amp(c) Different demographics in the two datasets The y-axes are log scale
neither is effective The first is that heavier users use dif-ferent types of applications (eg games) than lighter usersBut we find that the relative popularity of application typesis similar across classes of users with different interactiontimes (sect42) The second is user demographic But asFigures 3(b) and 3(c) show the interaction times are sim-ilar across the different demographics in the two datasetsWithin each demographic user interaction times span theentire range In sect42 we show that user demographic doesnot predict application popularity either
To understand the reasons behind diversity of user inter-action times we study next how user interaction is spreadacross individual sessions This analysis will show that thereis immense diversity among users in both the number of in-teraction sessions per day and the average session length
32 Interaction SessionsInteraction sessions provide a detailed view of how a user
interacts with the phone Their characteristics are impor-tant also because energy use depends not only on how longthe phone is used in aggregate but also on the usage distribu-tion Many short interactions likely drain more energy thanfew long interactions due to the overheads of awakening thephone and radio Even with negligible overheads batterylifetime depends on how exactly energy is consumed [20]Bursty drain with high current levels during bursts can leadto a lower lifetime than a more consistent drain rate
Figure 4(a) shows the number of sessions per day for dif-ferent users We again see a wide variation Individual usersinteract with their smartphone anywhere between 10 to 200times a day on average
Figure 4(b) shows the mean and standard deviation of in-teraction session lengths Dataset1 users tend to have muchlonger sessions than Dataset2 users Given that they haveroughly similar number of interactions per day as seen inFigure 4(a) their longer sessions explain their higher inter-action time per day as seen in Figure 3(a)
Within each dataset the mean session length varies acrossusers by an order of magnitude Across both datasets therange is 10-250 seconds
Explaining the diversity in session lengths Severalhypothesis might explain the differences in different usersrsquo
Dataset1Dataset21
1010
0
0 20 40 60 80 100User percentile
Ses
sion
cou
nt
(a) Number of sessions
110
100
1000
0 20 40 60 80 100User percentile
Mea
n se
ssio
n le
ngth
(s)
Dataset1Dataset2
(b) Session length
Figure 4 The mean and the upper end of the stan-dard deviation for the number of sessions per dayand the session length The y-axes are log scale
050
100
200
0 20 60 100 140Session length (s)
Ses
sion
cou
nt
Dataset1Dataset2
(a)
050
150
250
Ses
sion
cou
nt
lt10
10minus
20
20minus
30
30minus
40
40minus
60
gt60
Session length (s)
(b)
Figure 5 (a) Scatterplot of session count per dayand mean session length of various users (b) Themean and 95 CI of session count per day for usersin Dataset2 with different mean session lengths
session lengths One hypothesis is that users with longersessions concentrate their smartphone usage in fewer ses-sions Figure 5 shows however that there is little correla-tion between usersrsquo number of sessions and session lengthFigure 5(a) shows a scatterplot of session count versus meanlength for different users There is one data point for eachuser Figure 5(b) shows the dependence of session count onsession length by aggregating data across Dataset2 users Itplots the observed mean and 95 confidence interval (CI) for
Dataset1 example userDataset2 example user
00
02
04
06
08
10
0 200 600 1000Session length (s)
CD
F
(a) Session lengths
Dataset1 example userDataset2 example user
00
02
04
06
08
10
0 1000 2000 3000Time between sessions (s)
CD
F(b) Time between sessions
Figure 6 CDFs of session length and time betweensessions for two example users The x-axis ranges inthe graphs are different
session counts per day for users with different mean sessionlengths The differences in the session counts are not sta-tistically significant In other words it is not the case thatusers who have longer sessions have fewer or more sessions
Our other hypotheses are related to application use Thesecond hypothesis is that users run varying numbers of ap-plications during an interaction and users that tend to usemore applications per session have longer sessions The thirdhypothesis is that users run different applications and someapplications such as maps have longer sessions than othersThe fourth one is that even for the same application usershave different session lengths
Our analysis of application use in sect4 reveals that the sec-ond hypothesis is not explanatory as users overwhelminglyuse only one application per session It also reveals that thethird and fourth hypotheses are likely contributors to diver-sity in session lengths Note that the inability of applicationtypes to explain interaction time per day which we men-tion in the previous section is different from their ability toexplain session lengths
Distribution of a single userrsquos sessions We find thatfor any given user most of the sessions are short but someare very long Figure 6(a) shows the CDF of session lengthsfor two example users The median session length is lessthan a minute but some are longer than an hour (not shownin the graph) A similar skewed distribution can be seen forall users in our datasets albeit with different median andmean session length values This highly skewed distributionalso explains why the standard deviations in Figure 4(b) arehigh relative to the mean In sect81 we show how sessionlengths depend on the screen timeout values
Figure 6(b) shows that the time between sessions whenthe phone is not used also has a skewed distribution Mostare short (relative to the mean) but some are very long Weshow later that these off periods have the property that thelonger a user has been in one of them the greater the chancethat the user will continue in this state
33 Diurnal PatternsWe now study diurnal patterns in interaction The pres-
ence of such patterns has several consequences For instancethe length of time a given level of remaining battery capacitylasts will depend on the time of day
Figure 7 shows for two example users that as expectedstrong diurnal patterns do exist As a function of the hour
Dataset1Dataset2
16
1116
21
0 100 200 300 400Total interaction (minutes)
Diu
rnal
rat
io
(a)
lt30
30minus
60
60minus
90
90minus
120
120minus
150
150minus
200
200minus
300
gt30
0
01
23
45
Diu
rnal
rat
io
Total interaction (minutes)
(b)
Figure 8 (a) Scatterplot of diurnal ratio of interac-tion time per hour and interaction minutes per day(b) The mean and 95 CI of diurnal ratio vs totalinteraction time per day
of the day Figure 7(a) plots the mean number of interactionminutes per hour It also plots the 95 confidence interval(CI) around the mean which can be used to judge if thedifferences in the means are statistically significant We seea clear pattern in which daytime use is much higher thannightime use though the exact pattern for different users isdifferent
Figure 7(a) also shows that usage at hours in the night islow but not completely zero We believe that this non-zerousage stems from a combination of irregular sleeping hoursand users using their devices (eg to check time) when theyget up in the middle of the night
To capture the significance of the diurnal pattern for auser we define the diurnal ratio as the ratio of the meanusage during the peak hour to the mean usage across allhours A diurnal ratio of one implies no diurnal pattern andhigher values reflect stronger patterns Figure 9(a) plots thediurnal ratio in interaction time for all users It shows thatwhile diurnal ratios vary across users roughly 70 of theusers in each dataset have a peak hour usage that is morethan twice their mean usage
Explaining the diversity in diurnal patterns Tohelp explain the variability among usersrsquo diurnal ratios inFigure 8 we study its dependence on interaction time Fig-ure 8(a) shows a scatterplot of the diurnal ratio and themean interaction time per day We see that the diurnal ra-tio tends to be inversely correlated with interaction timeFigure 8(b) shows this negative correlation more clearly byaggregating data across users It plots the mean and 95CI of the diurnal ratio of total interaction time per day forusers with different total interaction times The diurnal ra-tio decreases as interaction time increases This inverse re-lationship suggests that heavy users tend to use their phonemore consistently during the day whereas light users tendto have concentrated use during certain hours of the day
Understanding the source of diurnal patterns Thevariation in interaction time of a user across the day can re-sult from variation in the number of interaction sessions orthe length of individual sessions We find that both factorscontribute Users tend to have different number of sessionsas well as different session lengths at different hours of theday Figures 7(b) and 7(c) illustrate this point for two ex-ample users They plot the mean number of sessions andthe mean session length for each hour of the day
Dataset1 example user0
510
1520
25
0 3 6 9 13 17 21Hour of day
Tot
al in
tera
ctio
n (m
inut
es) Dataset1 example user
05
1015
0 3 6 9 13 17 21Hour of day
Num
ber
of s
essi
ons
Dataset1 example user
050
100
150
0 3 6 9 13 17 21Hour of day
Mea
n se
ssio
n le
ngth
(s)
Dataset2 example user
05
1015
2025
0 3 6 9 13 17 21Hour of day
Tot
al in
tera
ctio
n (m
inut
es)
(a) Interaction time
Dataset2 example user
05
1015
0 3 6 9 13 17 21Hour of day
Num
ber
of s
essi
ons
(b) Number of sessions
Dataset2 example user
050
100
150
0 3 6 9 13 17 21Hour of day
Mea
n se
ssio
n le
ngth
(s)
(c) Session length
Figure 7 The mean and 95 CI of interaction time number of sessions and session length during each hourof the day for an example user from each dataset
Figures 9(b) and 9(c) show the strength of the diurnalpattern for the number of sessions and session length for allthe users Observe that compared to interaction time andsession length the diurnal ratio of the number of sessionstends to be lower
4 APPLICATION USAGEWe now study the applications that users run when they
interact with their smartphones Unlike previous attemptsto understand mobile application usage [5 19 26] that usediaries and interviews we rely on the mobile phonersquos OS toreport application usage We define an application as anyexecutable that the OS reports On Windows Mobile weget timestamped records of start and end times of appli-cation executions in the foreground On Android we logusage counters that are updated by the OS Every time theOS calls the onStart onRestart or onResume method of anAndroid application it starts a timer The timer stops whenthe onPause onStop or onDestroy method is called Werecord periodically the cumulative value of the timer for eachinstalled application This information on the extent of ap-plication use is not as accurate as the equivalent informationon Windows Mobile but it helps us understand relative timespent by the user in each application
41 Number of applicationsFigure 10 shows the number of applications used by users
in each dataset over the length of their trace We see thatthis number varies significantly from 10 to 90 across usersThe median is roughly 50 We are surprised by this highnumber given that the iPhone which is reported to have
020
4060
8010
0
0 20 40 60 80 100User percentile
Num
ber
of a
pplic
atio
ns
Dataset1Dataset2
Figure 10 Number of applications installed andused by users of each dataset
thousands of applications is not part of our study Ourresults show that avid use of smartphone applications is atrait shared by Android and Windows Mobile users as well
42 Application PopularityThe large number of applications installed by the users
does not mean that they use them equally We find thatusers devote the bulk of their attention to a subset of appli-cations of their choice Figure 11 illustrates this popularitybias for example users in each dataset It plots the relativepopularity of each application that is the ratio of the timespent interacting with the application and the total timespent interacting with the smartphone The bars show thepopularity PDF for the top 20 applications and the insetshows the semi-log plot for all applications Because they
15
913
1721
0 20 40 60 80 100User percentile
Diu
rnal
rat
io
Dataset1Dataset2
(a) Interaction time
12
34
5
0 20 40 60 80 100User percentile
Diu
rnal
rat
io
Dataset1Dataset2
(b) Number of sessions
15
913
1721
0 20 40 60 80 100User percentile
Diu
rnal
rat
io
Dataset1Dataset2
(c) Session length
Figure 9 Diurnal ratio of the interaction time the number of sessions and the session length for differentusers The y-axis ranges for the number of sessions is different
laun
cher gm
alar
mcl
ock
talk
cont
acts
map
sm
ms
phon
em
usic
brow
ser
cam
era
syst
emse
nsbe
nchm
ark
wer
tago
vend
ing
cale
ndar
setti
ngs
tippi
ntim
ete
tris
andr
oid
Dataset1 example user
Usa
ge r
atio
000
010
020
030
0 20 401eminus
070
01
iexp
lore
gwes
tmai
lho
me
bubb
lebr
eake
rP
oker
Live
Sea
rch
cale
ndar
Goo
gleM
aps
pout
look
TP
CS
olita
reK
aGlo
mpv
bloa
dcp
rog
serv
ices
tels
hell
pim
gap
pman
YG
oNet
wm
play
er
Dataset2 example user
Usa
ge r
atio
000
005
010
015
0 20 40 601eminus
060
01
Figure 11 Relative time spent running each application for example users in each dataset Inset is thesemi-log plot of application popularity
are binary names even some popular applications may ap-pear unfamiliar For instance in Dataset1 ldquolauncherrdquo is thedefault home screen application on Android in Dataset2ldquogwesrdquo (Graphical Windows and Events Subsystem) is thegraphical shell on Window Mobile
The graphs show that relative application popularity dropsquickly for both users In sect83 we show that for all usersapplication popularity can be modeled by an exponentialdistribution
Diurnal patterns Interestingly application popularityis not stable throughout the day but has a diurnal patternlike the other aspects of smartphone use That is the rel-ative popularity of an application is different for differenttimes of the day Figure 12 illustrates this for an exam-ple user in Dataset2 We see for instance that tmailexewhich is a messaging application on Windows Mobile ismore popular during the day than night Time dependentapplication popularity was recently reported by Trestian etal based on an analysis of the network traffic logs from a3G provider [27] Our analysis based on direct observationof user behavior confirms this effect
0 4 8 12 17 22Hour of day
020
4060
80U
sage
rat
io (
)
servicesexewmplayerexecalendarexepoutlookexeWLMMessengercdialexepimgexecprogexeiexploreexehomeexetmailexe
Figure 12 Relative time spent with each applicationduring each hour of the day for a sample user andher top applications
comm 44
browsing 10
other 11
prod 19media 5
maps 5games 2system 5
Dataset1
comm 49
browsing 12 other 15
prod 2media 9
maps 2
games 10
system 1
Dataset2
Figure 13 Relative popularity of each applicationcategory across all users in each dataset
Aggregate view application popularity To pro-vide an aggregate view of what users use their smartphonesfor we categorize applications into eight distinct categoriesi) communication contains applications for exchanging mes-sages (eg email SMS IM) and voice calls ii) browsingcontains Web browser search and social networking appli-cations iii) media contains applications for consuming orcreating media content (eg pictures music videos) iv)productivity contains applications for calendars alarms andfor viewing and creating text documents (eg Office PDFreader) v) system contains applications for changing userpreferences and viewing system state (eg file explorer)vi) games vii) maps and viii) other contains applicationsthat we could not include in any of the categories aboveeg because we did not know their function
Figure 13 shows the mean relative popularity of each ap-plication category across all users in each dataset Whilethe results are not identical across the two datasets they aresimilar to a first order Communication dominates in bothBrowsing is another major contributor in both datasetsMaps media and games have a comparatively lower butnevertheless substantial share of user attention
Relationship to user demographic To understandthe extent to which user demographic determines applica-tion popularity Figure 14 shows the mean and 95 CI ofrelative popularity of each application category for differ-ent user demographics As for interaction time (sect31) wesee that the impact of user demographics in our datasets isminimal In Dataset2 the relative popularity of various ap-plication types is similar for each of the four demographicsIn Dataset1 there are noticeable differences in the mean forcommunication games and productivity applications Highschool students use communication and games applicationsmore while knowledge workers use productivity applications
Dataset1
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
High schoolKnowledge worker
Dataset2
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
SCOPBPULPU
Figure 14 The mean and 95 CI of relative pop-ularity of application categories among users of dif-ferent demographics
more However considering the overlap of confidence in-tervals these differences in application popularity are notstatistically significant
From this result and the earlier one on the lack of depen-dence between user demographic and interaction time (sect31)we conclude that user demographic at least as defined inour datasets cannot reliably predict how a user will use thephone While demographic information appears to help insome cases (for eg the variation in usage of productivitysoftware in Dataset1) such cases are not the norm and itis hard to guess when demographic information would beuseful Pending development of other ways to classify userssuch that these classes more predictably explain the varia-tion incorporating factors specific to a user appear necessaryThis insensitivity to user demographic has positive as wellas negative implications A negative is that personalizationis more complex we cannot predict a usersrsquo behavior byknowing their demographic A positive implication is thatthe range of user behaviors along many dimensions of in-terest can be found in several common demographics Thissimplifies the task of uncovering the range because recruitingsubjects across multiple demographics tends to be difficult
Relationship to interaction time We also study ifusers that interact more with their phones tend to use differ-ent applications For each dataset we sort users based ontheir average interaction time per day and partition theminto different classes For Dataset1 we use two classesone each for the top and the bottom half of the users ForDataset2 which has more users we use three classes for the
Dataset1
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
HighLow
Dataset2
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
rHighMediumLow
Figure 15 The mean and 95 CI of relative pop-ularity of application categories among differentclasses of users based on interaction time per day
top middle and bottom third of the users Figure 15 showsthe mean and 95 CI for relative time spent with each ap-plication category by each user class We see that users indifferent classes have similar application usage Thus wecannot explain why some users use the phone more simplybased on the applications that they use
43 Application SessionsWe now study the characteristics of application sessions
We conduct this analysis only for Dataset2 based on times-tamps for when an application is started and ended Dataset1does not contain this information precisely Because appli-cations can run in the background start and end refer tothe period when the application is in the foreground
Applications run per interaction We begin by study-ing the number of applications that users run in an interac-tion session Figure 16 shows that an overwhelming major-ity close to 90 of interactions include only one applicationThis graph aggregates data across all users We did not findstatistically significant differences between users A largefraction of sessions of all users have only one application
That interaction sessions very often have only one appli-cation session suggests that users tend to interact with theirsmartphone for one task (eg reading email checking cal-endar etc) at a time and most of these tasks require theuse of only one application
Application session lengths Because interactionsessions are dominated by those with only one application
892
7422 06 02 01
1 2 3 4 5 6
020
4060
8010
0
Apps per session
Per
cent
age
()
Figure 16 Histogram of the number of applica-tions called during each interaction session for allthe users in Dataset2
map
s
gam
es
com
m
med
ia
brow
sing
prod
syst
em
othe
r
110
1000
Ses
sion
leng
th (
s)
Figure 17 The mean and 95 CI of session lengthsof different application categories The y-axis is logscale
the overall properties of application sessions such as theirlengths are similar to those of interaction sessions (sect3)
However studying the session lengths of applications sep-arately reveals interesting insights Different applicationtypes have different session lengths as shown in Figure 17for the categories defined earlier Interactions with mapsand games tend to be the longest and those related to pro-ductivity and system tend to be the shortest
Further given an application different users run themfor different times Figure 18 shows this effect for a mes-saging application tmailexe and a browsing applicationiexploreexe For each application the mean session lengthsof users differ by more than two orders of magnitude
These observations help explain why users have differentsession lengths (sect32) They prefer different applicationsand those application sessions tend to have different lengthsFurther analysis confirms this phenomenon For instance ifwe divide users into two classes based on their mean sessionlengths the popularity of games is twice as high in the classwith high session lengths
5 TRAFFICIn this section we investigate traffic generated by smart-
phones Unlike interaction events and application use net-work traffic is not an intentional user action but a side-effectof those actions Most users are likely oblivious to how much
tmail1
1010
010
00
0 20 40 60 80 100User percentile
Ses
sion
leng
th (
s)iexplore
110
100
1000
0 20 40 60 80 100User percentile
Ses
sion
leng
th (
s)Figure 18 The mean and the upper end of the stan-dard deviation of session lengths of two applications
0 20 40 60 80 100
User percentile
0
1
10
100
1000
10000
Traffic per day (MB) Receive
Send
Figure 19 The mean and the upper end of the stan-dard deviation of the traffic sent and received perday by users in Dataset1
traffic they generate We show that the diversity and diur-nal patterns of this side-effect match those of user actionsthemselves
The analysis in this section includes only Dataset1 we donot have traffic information for Dataset2 In Dataset1 werecord all of the data sent (or received) by the phone exceptfor that exchanged over the USB link ie the traffic hereinincludes data over the 3G radio and the 80211 wireless link
51 Traffic per dayFigure 19 shows that the amount of traffic sent and re-
ceived per day differs across users by almost three orders ofmagnitude The traffic received ranges from 1 to 1000 MBand the traffic sent ranges from 03 to 100 MB The medianvalues are 30 MB sent and 5 MB received
Our results indicate that traffic generated in a day bysmartphone users is comparable to traffic generated by PCsa few years ago A study of a campus WiFi network revealedthat on average users were generating 27 MB of traffic in2001 and 71 MB in 2003 [12] A study of Japanese residen-tial broadband users in 2006 revealed that on average usersgenerate 1000 MB of traffic per day [11] This high level oftraffic has major implications for the provisioning of wirelesscarrier networks as smartphone adoption increases
Relationship to application types To investigate ifcertain types of applications are favored more by users thatgenerate more traffice we divide the users into two equalclasses based on their sum of sent and received traffic perday Figure 20 shows mean and 95 CI of relative popularityof each application category for each user class Expectedly
Dataset1
Per
cent
age
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
HighLow
Figure 20 The mean and 95 CI of relative popu-larity of each application category among high andlow traffic consumers
it shows that communication applications are more popularamong users that consume more traffic
52 ldquoInteractiverdquo trafficWe next estimate what fraction of the total traffic is ldquoin-
teractiverdquo that is has timeliness constraints Approachesto reduce power consumption of data transfers often advo-cate rescheduling network transfers for instance by delay-ing some transfers so that multiple of them may be bun-dled [1 22] Such policies are difficult to implement forinteractive traffic without hurting user response time andthus are likely to be less valuable if the bulk of the traffic isinteractive
We classify traffic as interactive if it was generated whenthe screen is on This classification method might classifysome background traffic as interactive We expect the errorto be low because the screen is on for a small fraction ofthe time for most users Because certain user interactionswith the phone begin right after traffic exchange (eg anew email is received) we also consider traffic received in asmall time window (1 minute) before the screen is turned onas having timeliness constraints The results are robust tothe exact choice of time window Indeed some of the trafficthat we classify as interactive might be delay tolerant ega new email to be sent might be deferred for a little whileHowever the volume of traffic received by the phone whichusers would rather see immediately dominates by one or-der of magnitude the volume that is sent and delay-tolerantmessaging applications such as email contribute only a smallchunk of all traffic
Figure 21 shows the fraction of interactive traffic for eachuser We see that for about 90 of the users over 50of the traffic is interactive but for the rest almost none oftheir traffic is interactive Stated differently for differentsmartphone users almost all to almost none of the traffic isgenerated by applications in the background The extremesrepresent disparate ways in which people use smartphonesand which applications on the smartphone generate mosttraffic Our results imply that the energy savings that can behad by rescheduling network activity will vary across users
53 Diurnal patternsFigure 22(a) shows the diurnal pattern for an example
user with a sharp decline at night Figure 22(b) showsthe strength of the diurnal pattern for individual users by
0 20 40 60 80 100
User percentile
00
02
04
06
08
10
Interactive fraction
Figure 21 The fraction of interactive traffic
0 5 10 15 20
Hour of day
0
5
10
15
Traffic per hour (MB)
Receive
Send
(a)
0 20 40 60 80 100
User percentile
2
4
6
8
10
12
Diurnal ratio
Receive
Send
(b)
Figure 22 (a) The mean and 95 CI for trafficgenerated per hour by an example user (b) Thediurnal ratio of traffic per hour for all users
plotting the diurnal ratio of traffic As defined in sect33 thediurnal ratio reflects how much higher the mean during thepeak hour is to the overall mean
We see that the diurnal ratio varies across users but mosthave a strong diurnal behavior 80 of them generate overtwice their average amount of traffic in their peak hour Thisbehavior is likely a direct result of the high proportion ofinteractive traffic for most users and that user interactionsthemselves have a diurnal pattern
6 ENERGY CONSUMPTIONThe final aspect of smartphone usage that we investigate
is energy consumption Energy drain depends on two fac-tors i) user interactions and applications and ii) platformhardware and software If the second factor dominates theenergy drain of various users with the identical smartphonewill be similar Otherwise the energy drain will be as di-verse as user behaviors
We estimate the amount of energy drain based on the re-maining battery indicator which varies between 0 and 100If the battery indicator has gone down by X in a time pe-riod for a battery with capacity Y mAh we compute theenergy drain in that period to be X middot Y mAh 1 Given thatbatteries are complex electro-chemical devices [14 20] thiscomputation is approximate It assumes that the batterylevel indicator is linear with respect to energy drain
1mAh is technically a unit of charge yet is commonly used toindicate energy drain because battery voltage during normaloperations is typically constant For phones in our datasetthis is 4V so multiplying a mAh reading by 4 would yieldan accurate energy reading in milli-watt-hours
020
4060
8010
0
0 4 8 12 16 20Time (hour)
Bat
tery
leve
l (
) Benchmark1Benchmark2
Figure 23 Timelapse of the remaining battery levelindicator in controlled experiments with two differ-ent workloads at room temperature
110
100
1000
0 20 40 60 80 100User percentile
Bat
tery
Dra
in (
mA
h)
Figure 24 The mean and the upper end of the stan-dard deviation of one hour energy drain for Dataset1users during discharge periods
Controlled experiments suggest that the linearity assump-tion holds to a first order We run a benchmark load thatdrains the battery at a fixed rate in room temperature Un-der this benchmark if the battery level indicator decreaseslinearly with time it must be linear with respect to energydrain Figure 23 shows that the level indicator decreasesroughly linearly for two different benchmarks Benchmark1turns the screen on and off periodically Benchmark2 com-putes and idles periodically We conclude thus that the levelindicator can be used to estimate energy drain
Figure 24 shows the mean and standard deviation of en-ergy that users drain in an hour This graph is computed us-ing only periods in which the battery is not charging becauseenergy drain in those periods are of primary interest Wesee a two orders of magnitude difference among users Whileheaviest users drain close to 250 mAh the lightest of usersdrain only 10 mAh If the battery capacity is 1200 mAhthis leads to a lifetime variation from about 4 to 120 hours
Figure 25(a) shows for an example user that the drain isnot the same throughout the day but has diurnal variationsin which more energy is consumed during the day than dur-ing the night For this user the level of energy consumedchanges by roughly a factor of five Figure 25(b) plots thediurnal ratio of energy use for all users It shows that diurnalvariations occur with different strengths for all users
2060
100
140
0 4 8 12 16 20Hour of the day
Bat
tery
Dra
in (
mA
h)
(a)
0 20 40 60 80 100
12
34
56
7
User percentile
Diu
rnal
rat
io(b)
Figure 25 (a) The mean and 95 CI of energy drainof an example Dataset1 user (b) Diurnal ratio ofall users in Dataset1
Our results show that user activities contribute heavilytowards energy drain users in Dataset1 who have identi-cal smartphones drain energy at different rates and energydrain has diurnal patterns In the future we will developmethods to accurately quantify the energy consumption ofthe platform from that due to user-induced workload
7 IMPLICATIONS OF USER DIVERSITYWe uncover a surprising level of diversity among smart-
phone users For almost every aspect of usage that we studywe find one or more orders of magnitude difference betweenusers Our findings strongly motivate the need for customiz-ing smartphones to their users We believe that this needis greater than that for customizing ordinary cellphones orlaptops Ordinary cellphones do not have as rich an appli-cation environment Laptops are not as portable and aremore resource rich For example many users plug-in theirlaptops while using them
Customization can help at all levels Consider somethingas low-level as the battery Suppose we want batteries to beboth lightweight and last for at least a day with a high proba-bility Meeting the latter goal for all users of a given platformwill require catering to the heaviest users But that will leadto unnecessarily heavy batteries for many users (Highercapacity batteries are heavier) Offering multiple types ofbatteries with different lifetime-weight tradeoffs provides away out of this bind
At levels where intelligent mechanisms to improve user ex-perience or reduce energy consumption reside user diversitymotivates adapting to the smartphone user Driving thesemechanisms based on average case behaviors may not beeffective for a large fraction of the users
The ease and utility of customization depends on twoproperties of user behavior First despite quantitative dif-ferences there must be qualitative similarities among usersFor instance we should be able to describe the behavior ofall users with the same model Different users may havedifferent parameters of this model which will then lead toquantitative differences among them The presence of qual-itative similarities imply that users are not arbitrary pointsin space and it significantly simplifies the task of learninguser behavior Second user behavior in the past must alsobe predictive of the future Otherwise customization basedon past behaviors will be of little value in the future
In the next two sections we present evidence that theseproperties hold for several key aspects of smartphone usage
Dataset1 example user
000
00
010
002
00
030
0 100 200 300 400Session length (s)
Den
sity
Dataset2 example user
000
005
010
015
020
0 20 40 60 80 100Session length (s)
Den
sity
Figure 26 The PDF of session length for sampleusers of each dataset
In sect8 we show that user sessions and relative applicationpopularity can be described using simple models In sect9 weshow that energy drain can be described as well as predictedusing a ldquotrend trablerdquo framework
8 SMARTPHONE USAGE MODELSIn this section we develop simple models that describe
three aspects of smartphone usage ndash session lengths inter-arrival time between sessions and application popularityThe models are common across users but have different pa-rameters for different users While they do not completelydescribe user behavior they capture first order factors andrepresent a first step towards more complete modeling ofsmartphone usage More importantly along with the resultsof the next section they show that qualitative similaritiesdo exist among users
81 Session LengthsWe first consider the statistical properties of session length
distributions of users We find that session length valuestend to stationary With the KPSS test for level stationar-ity [13] 90 of the users have a p-value greater than 01The presence of stationarity is appealing because it suggeststhat past behavior is capable of predicting the future
We also find that session lengths are independent that isthe current value does not have a strong correlation with thevalues seen in the recent past With the Ljung-Box test forindependence [15] 96 of the users have a p-value that isgreater than 01
Stationarity and independence considered together sug-gest that session length values can be modeled as iid sam-ples from a distribution Choosing an appropriate distribu-tion however is complicated by the nature of the sessionlengths Most sessions are very short and the frequencydrops exponentially as the length increases However in-consistent with exponential behavior there are some verylong sessions in the tail for each user
We find that a mixture of exponential and Pareto distri-butions can model both ends of the spectrum The formercaptures short sessions and the latter captures long sessionsThat is session lengths can be described by the followingmixture model
r middot Exp(λ) + (1 minus r) middot Pareto(xm α)
In this equation r is the relative mix of the two distribu-tions λ is the rate of the exponential and xm and α are thelocation and shape parameters of the Pareto distribution
The location for a Pareto distribution represents the min-imum possible value of the random variable The location
010
0020
0030
00
0 1000 2000 3000Observed quantiles
Mod
el q
uant
iles
Figure 27 QQ plot of session lengths model for asample user
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed m
ix
(a) Relative mix (r)
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed r
ate
(b) Exponential rate (λ)
020
4060
0 20 40 60 80 100User percentile
Tim
eout
(s)
(c) Pareto location (xm)
minus1
0minus
04
02
06
10
0 20 40 60 80 100User percentile
Est
imat
ed s
hape
(d) Pareto shape (α)
Figure 28 Distribution of inferred model parame-ters that describe session length values of users inboth datasets
value that offers the best fit is the screen timeout value ofthe user because the session length PDF has a spike at thisvalue The spike corresponds to short sessions that are endedby the timeout (when the user forgets to switch the screenoff) we confirm this using controlled experiments with dif-ferent timeout values Figure 26 shows this spike at 60 and15 seconds for example users from each dataset The time-out provides a natural division between the two componentdistributions We automatically infer its approximate valueusing a simple spike detection algorithm
We use the EM algorithm to infer the maximum likelihoodestimation (MLE) of the remaining three parameters [6]Figure 27 shows the quality of this fit for an example userusing the QQ plot [3] Almost all quantiles are along they = x line indicating a good fit
Figure 28 shows the four inferred parameters for varioususers While users can be modeled using the same mix-
040
0080
00
0 4000 8000Observed quantiles
Mod
el q
uant
iles
Figure 29 QQ plot of session offtime model for asample user
050
010
0015
000 20 40 60 80 100
User percentileE
stim
ated
sca
le
(a) Weibull scale
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed s
hape
(b) Weibull shape
Figure 30 Distribution of inferred model parame-ters that describe the distribution of time betweensessions for users in both datasets
ture model the parameters of this model vary widely acrossusers Because of the way we construct our model the dis-tribution of the parameter r and xm also provide insightinto how frequently usersrsquo screen is switched off by the time-out and the relative popularity of different timeout values60 seconds is the most popular value likely because it is themost common default timeout and many users never changethe default
82 Time between SessionsWe find that the Weibull distribution can explain the
screen off times This distribution has two parameters re-ferred to as its scale and shape We find the MLE for theseparameters for each user From the QQ-plot in Figure 29we notice that the model predicts a greater probability ofseeing some very large offtimes than are observed in thedatasets However the probability of seeing these large off-times is small there are 27 data points that have a y-value greater than 8000 in that graph Hence we believethat Weibull provides a good fit for the length of intervalsbetween interactions
Figure 30 shows the distribution of the estimated shapeand scale of the fitted Weibull distributions Interestinglythe shape is consistently less than one Weibull shape valuesless than one indicate that the longer the screen has been offthe less likely it is for it to be turned on by the user This be-havior has interesting implications for power saving policiesFor instance periodic activities such as checking for emailwhen the screen has been off for a while may be deferred orrescheduled if needed without hurting user experience
020
4060
8010
0
0 20 40 60 80 100User percentile
MS
E (
)
(a) MSE
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed r
ate
(b) Rate parameter
Figure 31 (a) The mean square error (MSE) whenapplication popularity distribution is modeled usingan exponential (b) The inferred rate parameter ofthe exponential distribution for different users
83 Application PopularityWe find that for each user the relative popularity of ap-
plications can be well described by a simple exponential dis-tribution This qualitative invariant is useful for instanceto predict how many applications account for a given frac-tion of user attention For the example users in Figure 11this facet can be seen in the inset plots the semi-log of thepopularity distribution is very close to a straight line
Figure 31(a) shows that this exponential drop in applica-tion popularity is true for almost all users the mean squareerror between modeled exponential and actual popularitydistribution is less than 5 for 95 of the users
Figure 31(b) shows the inferred rate parameter of the ap-plication popularity distribution for various users We seethat the rate varies by an order of magnitude from 01 toalmost 1 The value of the rate essentially captures the paceof the drop in application popularity Lower values describeusers that use more applications on a regular basis Oneimplication of the wide range is that it may be feasible toretain all popular applications in memory for some users andnot for others
9 PREDICTING ENERGY DRAINIn this section we demonstrate the value of adapting to
user behaviors in the context of a mechanism to predict fu-ture energy drain on the smartphone Such a mechanismcan help with scheduling background tasks [9] estimatingbattery lifetime and improving user experience if it is esti-mated that the user will not fully drain the battery until thenext charging cycle [2 21]
Despite its many uses to our knowledge none of thesmartphones today provide a prediction of future energydrain Providing an accurate prediction of energy drain isdifficult2 User diversity of energy use makes any static pre-diction method highly inaccurate Even for the same userthere is a high variance in energy usage Figure 32 shows thisvariance by plotting the ratio of the standard deviation tothe mean energy drain over several time periods for users inDataset 1 The standard deviation of 10-minute windows ishigher than three times the mean for a fifth of the users The
2Energy drain prediction on smartphones is more challeng-ing than what is done for laptops Laptops provide a pre-diction of battery lifetime only for active use Smartphonepredictions on the other hand must cover multiple activeidle periods to be useful
01
23
45
0 20 40 60 80 100User percentile
Err
or r
atio
2 hours1 hour10 minutes
Figure 32 Battery drain predictors have to copewith the high variance in drain
time
window to
predict
n chunks
hellip
Trend Table
hellip
hellip
Figure 33 A userrsquos trend table captures how batterydrain in n consecutive chunks relates to drain in thesubsequent window
bursty nature of user interactions is one cause of this highvariance The variance is high even at longer time scalesThe standard deviation for two hour windows is larger than150 of the mean for half the users The diurnal patternsthat we showed earlier are one cause of high variance oversuch large time scales
Although the variance is high there are patterns in auserrsquos behaviors For instance we showed earlier that aphone that has been idle for a while is likely to remain idlein the near future and it will thus continue to draw energyat a similar rate
We hypothesize that a predictor tuned to user behaviorcan be accurate We present a simple personalized energydrain predictor that incorporates various user-specific fac-tors such as length of idle periods and different types ofbusy periods Instead of explicitly identifying these factorsour predictor captures them in a ldquotrend tablerdquo framework
Our framework is shown in Figure 33 Each entry in thetrend table is indexed by an n-tuple that represents energyusage readings from adjacent chunks of size δ each Thisindex points to energy usage statistics across all time win-dows of size w in the history that follow the n chunk valuesrepresented by that index Thus the trend table captureshow the energy usage in n consecutive chunks is related tothe energy use in the subsequent time windows of size w
In this paper we use n = 3 δ = 10 minutes and maintaindifferent trend tables for each window size w for which theprediction is required To keep the trend table small wequantize the energy readings of chunks We maintain themean and standard deviation for all subsequent windowsthat map to that index
To make a prediction we use the energy readings fromthe n immediately preceding chunks Let these readings be(x1 xn) Then we search for indices in the table thatdiffer from (x1 xn) by no more than a threshold s and
GenericShortminustermTimeminusofminusdayPersonalized
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
(a) 1-hour window
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
GenericShortminustermTimeminusofminusdayPersonalized
(b) 2-hour window
Figure 34 The ratio of error to true battery drainfor different energy drain predictors
predict based on them such that similar indices have higherrelative weight More precisely the prediction is
X
e stforalli=1n|he
iminusxi|lts
ke times ue
P
ke
where e iterates over all tuples in the trend table hei is
the value of the chunk i of the tuple e ue represents thestatistics stored for that index and the relative weight ke =s minus maxi=1n |he
i minus xi| For results shown here we uses = 05
Figures 34 shows how well our Personalized predictor worksfor predicting usage for two different future time windowsThere is one point per user in each graph which correspondsto the median error seen across many predictions for thatuser We see that the 90th percentile error is 25 for 1-hourwindow and 40 for 2-hour window
To place this level of accuracy in context the figure alsoshows the accurary of three alternative predictors TheGeneric predictor uses a trend table but instead of buildinga user-specific table it builds one that combines all usersWe see that its 90th percentile error is roughly twice thatof the Personalized predictor Its predictions are worse forhalf of the users
The Short-term predictor predicts energy drain simply asthe amount drained in the immediately preceding windowof the same size For both time windows its 90th percentileerror is no better than that of the Generic predictor
The Time-of-day predictor predicts energy drain based onvalues observed for the user at the same time of the day inthe past similar to that proposed by Banerjee et al [2]Thus this predictor has some degree of personalization butit learns user behavior in less detail than the Personalizedpredictor We see that its performance for 1-hour windowis similar to the Short-term predictor For 2-hour windowits error is higher than the Personalized predictor across theboard though its worst case error is lower than the Short-term and Generic predictors
Overall these results demonstrate the value of appropri-ately learning user behaviors to design intelligent mecha-nisms on smartphones
10 RELATED WORKIn a range of domains there is a rich history of work that
characterizes user workloads However because smartphoneadoption has gathered pace relatively recently our work rep-resents one of the few to study how people use smartphones
Along with other recent works it helps complete the pic-ture Banerjee et al and Rahmati et al report on batterycharging behaviors [2 18] Like us they find considerablevariation among users Banerjee et al also propose a pre-dictor that estimates the excess energy of the battery atthe time of charging using a histogram of past battery us-age [2] Shye et al study the power consumption character-istics of 20 users [24] They infer properties such as whichcomponents in the phone consume most power and exploreoptimizations based on these characteristics Rahmati andZhong study 14 users of a particular demographic to studywhich applications are popular in that demographic wherethe phone is used and how it is shared among users [19]
In contrast to these works we focus on understandingdifferent aspects of smartphone use (eg interactions andtraffic) and on exposing the diversity of user behaviors in-stead of only the average behavior Our study also entailsan order of magnitude more users than previous efforts
Like us other researchers have developed logging utilitiesMyExperience is one such early utility [10] and the worksabove involve custom utilities as well Our focus in thispaper is not on the development of logging tools but onanalyzing their data to gain insight into user behavior
There is a body of work in modeling the aggregate be-havior of mobile users Using traces from a large cellularoperator some network related aspects of mobile usage havebeen modeled Halepovic et al and Williamson et al re-port that call arrivals are bursty and present diurnal pat-terns [28] Willkomm et al and Brown et al report thatmobile users call duration can be approximately modeled bya lognormal distribution [4 29] We use traces collected onthe mobile device itself and focus on modeling the interac-tions of individual users instead of the aggregate behavior
11 CONCLUSIONSBy studying 255 users of two different smartphone plat-
forms we comprehensively characterized user activities andtheir impact on network and battery We quantify manyhitherto unknown aspects of smartphone usage User diver-sity is an overarching theme in our findings For instancedifferent users interact with their phones 10-200 times a dayon average the mean interaction length of different users
is 10-250 seconds and users receive 1-1000 MB of data perday where 10-90 is received as part of active use
This extent of user diversity implies that mechanisms thatwork for the average case may be ineffective for a large frac-tion of the users Instead learning and adapting to userbehaviors is likely to be more effective as demonstrated byour personalized energy drain predictor We show that qual-itative similarities exist among users to facilitate the devel-opment of such mechanisms For instance the longer theuser has not interacted with the phone the less likely it isfor her to start interacting with it and application popular-ity for a user follows a simple exponential distribution
Our study points to ways in which smartphone platformsshould be enhanced Effective adaptation will require futureplatforms to support light-weight tools that monitor andlearn user behaviors in situ It will also require them toexpose appropriate knobs to control the behavior of lower-level components (eg CPU or radio)
12 ACKNOWLEDGEMENTSWe are grateful to the users in Dataset1 that ran our log-
ger and to the providers of Dataset2 Elizabeth Dawson pro-vided administrative support for data collection Jean BolotKurt Partridge Stefan Saroiu Lin Zhong and the MobiSysreviewers provided many comments that helped improve thepaper We thank them all This work is supported in partby the NSF grants CCR-0120778 and CNS-0627084
13 REFERENCES[1] Balasubramanian N Balasubramanian A
and Venkataramani A Energy consumption inmobile phones A measurement study andimplications for network applications In IMC (2009)
[2] Banerjee N Rahmati A Corner M D
Rollins S and Zhong L Users and batteriesInteractions and adaptive energy management inmobile systems In Ubicomp (2007)
[3] Becker R Chambers J and Wilks A The newS language Chapman amp HallCRC 1988
[4] Brown L Gans N Mandelbaum A Sakov
A Shen H Zeltyn S and Zhao L Statisticalanalysis of a telephone call center Journal of theAmerican Statistical Association 100 469 (2005)
[5] Church K and Smyth B Understanding mobileinformation needs In MobileHCI (2008)
[6] Dempster A Laird N and Rubin D Maximumlikelihood from incomplete data via the EM algorithmJournal of the Royal Statistical Society Series B(Methodological) 39 1 (1977)
[7] Esfahbod B Preload An adaptive prefetchingdaemon Masterrsquos thesis University of Toronto 2006
[8] Falaki H Govindan R and Estrin D Smartscreen management on mobile phones Tech Rep 74Center for Embedded Networked Sesning 2009
[9] Flinn J and Satyanarayanan M Managingbattery lifetime with energy-aware adaptation ACMTOCS 22 2 (2004)
[10] Froehlich J Chen M Consolvo S
Harrison B and Landay J MyExperience Asystem for in situ tracing and capturing of userfeedback on mobile phones In MobiSys (2007)
[11] Fukuda K Cho K and Esaki H The impact ofresidential broadband traffic on Japanese ISPbackbones SIGCOMM CCR 35 1 (2005)
[12] Kotz D and Essien K Analysis of a campus-widewireless network Wireless Networks 11 1ndash2 (2005)
[13] Kwiatkowski D Phillips P Schmidt P and
Shin Y Testing the null hypothesis of stationarityagainst the alternative of a unit root Journal ofEconometrics 54 1-3 (1992)
[14] Lahiri K Dey S Panigrahi D and
Raghunathan A Battery-driven system design Anew frontier in low power design In Asia SouthPacific design automationVLSI Design (2002)
[15] Ljung G and Box G On a measure of lack of fitin time series models Biometrika 65 2 (1978)
[16] Smartphone futures 2009-2014 httpwwwportioresearchcomSmartphone09-14html
[17] Rahmati A Qian A and Zhong L
Understanding human-battery interaction on mobilephones In MobileHCI (2007)
[18] Rahmati A and Zhong L Humanndashbatteryinteraction on mobile phones Pervasive and MobileComputing 5 5 (2009)
[19] Rahmati A and Zhong L A longitudinal study ofnon-voice mobile phone usage by teens from anunderserved urban community Tech Rep 0515-09Rice University 2009
[20] Rao R Vrudhula S and Rakhmatov D
Battery modeling for energy aware system designIEEE Computer 36 12 (2003)
[21] Ravi N Scott J Han L and Iftode L
Context-aware battery management for mobilephones In PerCom (2008)
[22] Sharma A Navda V Ramjee R
Padmanabhan V and Belding E Cool-TetherEnergy efficient on-the-fly WiFi hot-spots usingmobile phones In CoNEXT (2009)
[23] Sharma C US wireless data market update - Q32009 httpwwwchetansharmacomusmarketupdateq309htm
[24] Shye A Sholbrock B and G M Into the wildStudying real user activity patterns to guide poweroptimization for mobile architectures In Micro (2009)
[25] Snol L More smartphones than desktop PCs by2011 httpwwwpcworldcomarticle171380more_smartphones_than_desktop_pcs_by_2011
html
[26] Sohn T Li K Griswold W and Hollan J
A diary study of mobile information needs In SIGCHI(2008)
[27] Trestian I Ranjan S Kuzmanovic A and
Nucci A Measuring serendipity Connecting peoplelocations and interests in a mobile 3G network InIMC (2009)
[28] Williamson C Halepovic E Sun H and Wu
Y Characterization of CDMA2000 Cellular DataNetwork Traffic In Local Computer Networks (2005)
[29] Willkomm D Machiraju S Bolot J and
Wolisz A Primary users in cellular networks Alarge-scale measurement study In DySPAN (2008)
- Introduction
- Data Collection
-
- Dataset1
- Dataset2
- Representativeness of conclusions
-
- User Interactions
-
- Interaction Time
- Interaction Sessions
- Diurnal Patterns
-
- Application Usage
-
- Number of applications
- Application Popularity
- Application Sessions
-
- Traffic
-
- Traffic per day
- ``Interactive traffic
- Diurnal patterns
-
- Energy Consumption
- Implications of User Diversity
- Smartphone Usage Models
-
- Session Lengths
- Time between Sessions
- Application Popularity
-
- Predicting Energy Drain
- Related Work
- Conclusions
- Acknowledgements
- References
-
users Duration Platform Demographic info Information loggedDataset1 33 7-21 weeksuser Android 16 high school students Screen state applications used
17 knowledge workers network traffic battery stateDataset2 222 8-28 weeksuser Windows 61 SC 65 LPU 59 BPU 37 OP Screen state applications used
Mobile Country 116 USA 106 UK
Table 1 An overview of the datasets in our study
games and maps applications more often tend to have longerinteractions Our study also shows that demographic infor-mation can be an unreliable predictor of user behavior andusage diversity exists even when the underlying device isidentical as is the case for one of our datasets
Among the many implications of our findings an over-riding one is that mechanisms to improve user experienceor energy consumption should not follow a one-size-fits-allmindset They should instead adapt by learning relevantuser behaviors otherwise they will likely be only marginallyuseful or benefit only a small proportion of users
We show that despite quantitative differences qualitativesimilarities exist among users which facilitates the task oflearning user behavior For several key aspects of smart-phone usage the same model can describe all users differ-ent users have different model parameters For instance thetime between user interactions can be captured using theWeibull distribution For every user the shape parameterof this model is less than one which implies that the longerit has been since the userrsquos last interaction the less likelyit is for the next interaction to start We also find thatthe relative popularity of applications for each user followsan exponential distribution though the parameters of thedistribution vary widely across users
We demonstrate the value of adapting to user behavior inthe context of a mechanism to predict future energy drainPredicting energy drain is an inherently challenging taskBursty user interactions at short time scales combined withdiurnal patterns at longer time scales lead to an energy con-sumption process that has a very high variance and is seem-ingly unpredictable We show however that reasonably ac-curate predictions can be made by learning the userrsquos energyuse signature in terms of aldquotrend tablerdquo framework For pre-dicting the energy use one hour in the future our predictorrsquos90th percentile error is under 25 Without adaptation andbasing the predictions on average behavior the 90th per-centile error is 60
2 DATA COLLECTIONOur work is based on two sets of data The first is a
high-fidelity data set that we gathered by deploying a cus-tom logger on the phones of 33 Android users The seconddata set consists of 222 Windows Mobile users across dif-ferent demographics Together these data sets provide abroad and detailed view of smartphone usage We leave forthe future the task of studying other smartphone platformssuch as iPhone and BlackBerry The characteristics of ourdatasets are summarized in Table 1
21 Dataset1Our first set of traces is from 33 Android users These
users consisted of 17 knowledge workers and 16 high schoolstudents Knowledge workers were computer science re-searchers and high school students were interns in a single
organization and were recruited by a third person on ourbehalf As stated in our study consent form the usersrsquo iden-tities were not revealed to us The participants were givenHTC Dream smartphones with unlimited voice text anddata plans We encouraged the users to take advantage ofall the features and services of the phones
The data was collected using a custom logging tool thatwe developed and deployed on the smartphones The loggerruns in the background and records a highly detailed viewof smartphone use including the state of the smartphonescreen start and end of incoming and outgoing voice callsthe time the user spends interacting with each applicationthe network traffic sent and received per application andthe battery level The Android OS provides mechanisms toaccess this information The logger keeps data records in alocal SQLite database on the phone and uploads them onlywhen the phone is plugged to the charger to minimize theimpact on the phone battery Our logging utility is availableto other researchers by request
The data was gathered between May and October 2009There is 7-21 weeks of data per user with the average being9 weeks
22 Dataset2Our second data set was collected by an organization that
was interested in investigating smartphone usability and ap-plication popularity This organization provided 222 userswith Windows Mobile smartphones from different hardwarevendors It also paid for their voice and data plans For rep-resentativeness the users were drawn from different demo-graphics as shown in Table 1 The demographic categorieswere defined based on what users stated as the primary moti-vation for using a smartphone Social Communicators (SC)wanted to ldquostay connected via voice and textrdquo Life PowerUsers (LPU) wanted ldquoa multi-function device to help themmanage their liferdquo Business Power Users (BPU) wanted ldquoanadvanced PC-companion to enhance their business produc-tivityrdquo Organizer Practicals (OP) wanted ldquoa simple deviceto manage their liferdquo The subjects were asked about theirintended use of the phone before the study and were catego-rized based on their answers To our knowledge the resultsof this study are not public
Traces were collected using a logger that recorded startand end time of each application This information waslogged using the ActiveApplication API call of the OS whichreports on the executable that currently has the foregroundwindow (with a callback for changes) Other details that ourcustom logger in sect21 records (eg network traffic and bat-tery level) were not logged in this study Thus this datasethas lower fidelity than the first one but it provides a viewof smartphone usage across a broader range of users
The traces were collected between May 2008 and January2009 There is 8-28 weeks of data per user with the averagebeing 16 weeks
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Voi
ce u
sage
rat
io
Dataset1Dataset2
Figure 1 Ratio of voice usage to total usage Thex-axis is user percentile and users are sorted in de-creasing order of ratio
23 Representativeness of conclusionsAn important concern for user studies such as ours is
whether the resulting conclusions represent the entire popu-lation There are two potential sources of bias in our data i)the users are not representative and ii) the measured usageis not representative We believe that our conclusions aregeneral The first concern is alleviated by the fact that asidefrom some quantitative differences we find remarkable con-sistency among users in the two datasets This consistencysuggests generality given that the two datasets are gatheredindependently on different platforms and Dataset2 was pro-fessionally designed to be representative
The second concern stems from the possibility that usersmay not be using the monitored smartphones as their pri-mary devices or that the usage during the monitoring periodmay not be normal All users in Dataset2 used the providedsmartphones as their primary devices We do not know thisaspect with certainty for Dataset1 but we understand fromanecdotal evidence that some users used these devices astheir only phones and others took advantage of the unlim-ited minutes and text plans We study voice usage as in-dicative of the extent to which users relied on the monitoreddevices Higher voice usage suggests use as primary phonesFigure 1 shows the ratio of time users spent in phone callsto total time spent interacting with the phone (sect3) We seethat the overall voice usage in Dataset1 was higher thanthat in Dataset2 in which all users used the phone as theirprimary device
Given that the monitored devices tended to be primaryand the long duration of the monitoring interval we con-jecture that our traces predominantly capture normal us-age Earlier work has pointed to the possibility of an initialadoption process during which usage tends to be differentthan long-term usage [19] To show that our traces are notdominated by the initial excitement of users or other specialevents that cause usage to be appreciably different from thenormal usage Figure 2 shows the average interaction timeper day (sect3) in the first and second halves of the datasets foreach user We see roughly similar usage in the two halvesDetailed investigation shows that the visible differences inthe averages of the two halves especially in Dataset1 arenot statistically significant Other measures of usage (egnetwork activity) look similar We do not claim that in-stances of abnormal usage are absent in the datasets but
110
100
1000
0 20 40 60 80 100User percentile
Act
ive
time
(min
utes
)
Dataset1 1st halfDataset1 2nd halfDataset2 1st halfDataset2 2nd half
Figure 2 Total interaction per day during the firstand second halves of study for each user Userswithin each dataset are sorted based on the interac-tion time in the first half The y-axis is log scale
the monitored period was long enough for our results to notbe impacted by such instances
3 USER INTERACTIONSWe begin our analysis by studying how users interact with
their smartphones independent of the application used Wecharacterize application use in the next section and the im-pact of user actions on network traffic and energy drain inthe following sections
We define an interaction interval also referred to as a ses-sion in this paper differently for each dataset In Dataset1we deem a user to be interacting with the phone wheneverthe screen is on or a voice call is active In Dataset2 aninteraction is defined as the interval that an application isreported to be on the foreground This includes voice callsbecause on Windows Mobile a special program (ldquocprogexerdquo)is reported in the foreground during voice calls
31 Interaction TimeFigure 3(a) shows a basic measure of user interactionmdashthe
number of minutes in a day that a user interacts with thesmartphone The plot shows the mean and the standarddeviation of this number for each user For visual clarityin such graphs we plot only the upper end of the standarddeviation plotting both ends occludes the other curves Theinterested reader can estimate the lower end since standarddeviation is symmetric around the mean
Dataset1 users tend to have more interaction minutes be-cause as we show later they tend to have longer interactionsessions while having a similar number of sessions Withineach dataset however there is an order of magnitude dif-ference among users In Dataset1 the lower end is only 30minutes in a day But the high end is 500 minutes which isroughly eight hours or a third of the day We are surprisedby this extremely high level of usage
Figure 3(a) also shows that users cover the entire rangebetween the two extremes and are not clustered into dif-ferent regions The lack of clusters implies that effectivepersonalization will likely need to learn an individual userrsquosbehavior rather than mapping a user to one or a few pre-defined categories
We examine two factors that can potentially explain theextent to which a user interacts with the phone but find that
110
100
1000
0 20 40 60 80 100User percentile
Act
ive
time
(min
utes
)
Dataset1Dataset2
(a) All users
High schoolKnowledge worker
110
100
1000
0 20 40 60 80 100User percentile
Act
ive
time
(min
utes
)
(b) Dataset1
SCOPBPULPU1
1010
010
00
0 20 40 60 80 100User percentile
Act
ive
time
(min
utes
)
(c) Dataset2
Figure 3 The mean and the upper end of the standard deviation of interaction minutes per day (a) Allusers in each dataset (b)amp(c) Different demographics in the two datasets The y-axes are log scale
neither is effective The first is that heavier users use dif-ferent types of applications (eg games) than lighter usersBut we find that the relative popularity of application typesis similar across classes of users with different interactiontimes (sect42) The second is user demographic But asFigures 3(b) and 3(c) show the interaction times are sim-ilar across the different demographics in the two datasetsWithin each demographic user interaction times span theentire range In sect42 we show that user demographic doesnot predict application popularity either
To understand the reasons behind diversity of user inter-action times we study next how user interaction is spreadacross individual sessions This analysis will show that thereis immense diversity among users in both the number of in-teraction sessions per day and the average session length
32 Interaction SessionsInteraction sessions provide a detailed view of how a user
interacts with the phone Their characteristics are impor-tant also because energy use depends not only on how longthe phone is used in aggregate but also on the usage distribu-tion Many short interactions likely drain more energy thanfew long interactions due to the overheads of awakening thephone and radio Even with negligible overheads batterylifetime depends on how exactly energy is consumed [20]Bursty drain with high current levels during bursts can leadto a lower lifetime than a more consistent drain rate
Figure 4(a) shows the number of sessions per day for dif-ferent users We again see a wide variation Individual usersinteract with their smartphone anywhere between 10 to 200times a day on average
Figure 4(b) shows the mean and standard deviation of in-teraction session lengths Dataset1 users tend to have muchlonger sessions than Dataset2 users Given that they haveroughly similar number of interactions per day as seen inFigure 4(a) their longer sessions explain their higher inter-action time per day as seen in Figure 3(a)
Within each dataset the mean session length varies acrossusers by an order of magnitude Across both datasets therange is 10-250 seconds
Explaining the diversity in session lengths Severalhypothesis might explain the differences in different usersrsquo
Dataset1Dataset21
1010
0
0 20 40 60 80 100User percentile
Ses
sion
cou
nt
(a) Number of sessions
110
100
1000
0 20 40 60 80 100User percentile
Mea
n se
ssio
n le
ngth
(s)
Dataset1Dataset2
(b) Session length
Figure 4 The mean and the upper end of the stan-dard deviation for the number of sessions per dayand the session length The y-axes are log scale
050
100
200
0 20 60 100 140Session length (s)
Ses
sion
cou
nt
Dataset1Dataset2
(a)
050
150
250
Ses
sion
cou
nt
lt10
10minus
20
20minus
30
30minus
40
40minus
60
gt60
Session length (s)
(b)
Figure 5 (a) Scatterplot of session count per dayand mean session length of various users (b) Themean and 95 CI of session count per day for usersin Dataset2 with different mean session lengths
session lengths One hypothesis is that users with longersessions concentrate their smartphone usage in fewer ses-sions Figure 5 shows however that there is little correla-tion between usersrsquo number of sessions and session lengthFigure 5(a) shows a scatterplot of session count versus meanlength for different users There is one data point for eachuser Figure 5(b) shows the dependence of session count onsession length by aggregating data across Dataset2 users Itplots the observed mean and 95 confidence interval (CI) for
Dataset1 example userDataset2 example user
00
02
04
06
08
10
0 200 600 1000Session length (s)
CD
F
(a) Session lengths
Dataset1 example userDataset2 example user
00
02
04
06
08
10
0 1000 2000 3000Time between sessions (s)
CD
F(b) Time between sessions
Figure 6 CDFs of session length and time betweensessions for two example users The x-axis ranges inthe graphs are different
session counts per day for users with different mean sessionlengths The differences in the session counts are not sta-tistically significant In other words it is not the case thatusers who have longer sessions have fewer or more sessions
Our other hypotheses are related to application use Thesecond hypothesis is that users run varying numbers of ap-plications during an interaction and users that tend to usemore applications per session have longer sessions The thirdhypothesis is that users run different applications and someapplications such as maps have longer sessions than othersThe fourth one is that even for the same application usershave different session lengths
Our analysis of application use in sect4 reveals that the sec-ond hypothesis is not explanatory as users overwhelminglyuse only one application per session It also reveals that thethird and fourth hypotheses are likely contributors to diver-sity in session lengths Note that the inability of applicationtypes to explain interaction time per day which we men-tion in the previous section is different from their ability toexplain session lengths
Distribution of a single userrsquos sessions We find thatfor any given user most of the sessions are short but someare very long Figure 6(a) shows the CDF of session lengthsfor two example users The median session length is lessthan a minute but some are longer than an hour (not shownin the graph) A similar skewed distribution can be seen forall users in our datasets albeit with different median andmean session length values This highly skewed distributionalso explains why the standard deviations in Figure 4(b) arehigh relative to the mean In sect81 we show how sessionlengths depend on the screen timeout values
Figure 6(b) shows that the time between sessions whenthe phone is not used also has a skewed distribution Mostare short (relative to the mean) but some are very long Weshow later that these off periods have the property that thelonger a user has been in one of them the greater the chancethat the user will continue in this state
33 Diurnal PatternsWe now study diurnal patterns in interaction The pres-
ence of such patterns has several consequences For instancethe length of time a given level of remaining battery capacitylasts will depend on the time of day
Figure 7 shows for two example users that as expectedstrong diurnal patterns do exist As a function of the hour
Dataset1Dataset2
16
1116
21
0 100 200 300 400Total interaction (minutes)
Diu
rnal
rat
io
(a)
lt30
30minus
60
60minus
90
90minus
120
120minus
150
150minus
200
200minus
300
gt30
0
01
23
45
Diu
rnal
rat
io
Total interaction (minutes)
(b)
Figure 8 (a) Scatterplot of diurnal ratio of interac-tion time per hour and interaction minutes per day(b) The mean and 95 CI of diurnal ratio vs totalinteraction time per day
of the day Figure 7(a) plots the mean number of interactionminutes per hour It also plots the 95 confidence interval(CI) around the mean which can be used to judge if thedifferences in the means are statistically significant We seea clear pattern in which daytime use is much higher thannightime use though the exact pattern for different users isdifferent
Figure 7(a) also shows that usage at hours in the night islow but not completely zero We believe that this non-zerousage stems from a combination of irregular sleeping hoursand users using their devices (eg to check time) when theyget up in the middle of the night
To capture the significance of the diurnal pattern for auser we define the diurnal ratio as the ratio of the meanusage during the peak hour to the mean usage across allhours A diurnal ratio of one implies no diurnal pattern andhigher values reflect stronger patterns Figure 9(a) plots thediurnal ratio in interaction time for all users It shows thatwhile diurnal ratios vary across users roughly 70 of theusers in each dataset have a peak hour usage that is morethan twice their mean usage
Explaining the diversity in diurnal patterns Tohelp explain the variability among usersrsquo diurnal ratios inFigure 8 we study its dependence on interaction time Fig-ure 8(a) shows a scatterplot of the diurnal ratio and themean interaction time per day We see that the diurnal ra-tio tends to be inversely correlated with interaction timeFigure 8(b) shows this negative correlation more clearly byaggregating data across users It plots the mean and 95CI of the diurnal ratio of total interaction time per day forusers with different total interaction times The diurnal ra-tio decreases as interaction time increases This inverse re-lationship suggests that heavy users tend to use their phonemore consistently during the day whereas light users tendto have concentrated use during certain hours of the day
Understanding the source of diurnal patterns Thevariation in interaction time of a user across the day can re-sult from variation in the number of interaction sessions orthe length of individual sessions We find that both factorscontribute Users tend to have different number of sessionsas well as different session lengths at different hours of theday Figures 7(b) and 7(c) illustrate this point for two ex-ample users They plot the mean number of sessions andthe mean session length for each hour of the day
Dataset1 example user0
510
1520
25
0 3 6 9 13 17 21Hour of day
Tot
al in
tera
ctio
n (m
inut
es) Dataset1 example user
05
1015
0 3 6 9 13 17 21Hour of day
Num
ber
of s
essi
ons
Dataset1 example user
050
100
150
0 3 6 9 13 17 21Hour of day
Mea
n se
ssio
n le
ngth
(s)
Dataset2 example user
05
1015
2025
0 3 6 9 13 17 21Hour of day
Tot
al in
tera
ctio
n (m
inut
es)
(a) Interaction time
Dataset2 example user
05
1015
0 3 6 9 13 17 21Hour of day
Num
ber
of s
essi
ons
(b) Number of sessions
Dataset2 example user
050
100
150
0 3 6 9 13 17 21Hour of day
Mea
n se
ssio
n le
ngth
(s)
(c) Session length
Figure 7 The mean and 95 CI of interaction time number of sessions and session length during each hourof the day for an example user from each dataset
Figures 9(b) and 9(c) show the strength of the diurnalpattern for the number of sessions and session length for allthe users Observe that compared to interaction time andsession length the diurnal ratio of the number of sessionstends to be lower
4 APPLICATION USAGEWe now study the applications that users run when they
interact with their smartphones Unlike previous attemptsto understand mobile application usage [5 19 26] that usediaries and interviews we rely on the mobile phonersquos OS toreport application usage We define an application as anyexecutable that the OS reports On Windows Mobile weget timestamped records of start and end times of appli-cation executions in the foreground On Android we logusage counters that are updated by the OS Every time theOS calls the onStart onRestart or onResume method of anAndroid application it starts a timer The timer stops whenthe onPause onStop or onDestroy method is called Werecord periodically the cumulative value of the timer for eachinstalled application This information on the extent of ap-plication use is not as accurate as the equivalent informationon Windows Mobile but it helps us understand relative timespent by the user in each application
41 Number of applicationsFigure 10 shows the number of applications used by users
in each dataset over the length of their trace We see thatthis number varies significantly from 10 to 90 across usersThe median is roughly 50 We are surprised by this highnumber given that the iPhone which is reported to have
020
4060
8010
0
0 20 40 60 80 100User percentile
Num
ber
of a
pplic
atio
ns
Dataset1Dataset2
Figure 10 Number of applications installed andused by users of each dataset
thousands of applications is not part of our study Ourresults show that avid use of smartphone applications is atrait shared by Android and Windows Mobile users as well
42 Application PopularityThe large number of applications installed by the users
does not mean that they use them equally We find thatusers devote the bulk of their attention to a subset of appli-cations of their choice Figure 11 illustrates this popularitybias for example users in each dataset It plots the relativepopularity of each application that is the ratio of the timespent interacting with the application and the total timespent interacting with the smartphone The bars show thepopularity PDF for the top 20 applications and the insetshows the semi-log plot for all applications Because they
15
913
1721
0 20 40 60 80 100User percentile
Diu
rnal
rat
io
Dataset1Dataset2
(a) Interaction time
12
34
5
0 20 40 60 80 100User percentile
Diu
rnal
rat
io
Dataset1Dataset2
(b) Number of sessions
15
913
1721
0 20 40 60 80 100User percentile
Diu
rnal
rat
io
Dataset1Dataset2
(c) Session length
Figure 9 Diurnal ratio of the interaction time the number of sessions and the session length for differentusers The y-axis ranges for the number of sessions is different
laun
cher gm
alar
mcl
ock
talk
cont
acts
map
sm
ms
phon
em
usic
brow
ser
cam
era
syst
emse
nsbe
nchm
ark
wer
tago
vend
ing
cale
ndar
setti
ngs
tippi
ntim
ete
tris
andr
oid
Dataset1 example user
Usa
ge r
atio
000
010
020
030
0 20 401eminus
070
01
iexp
lore
gwes
tmai
lho
me
bubb
lebr
eake
rP
oker
Live
Sea
rch
cale
ndar
Goo
gleM
aps
pout
look
TP
CS
olita
reK
aGlo
mpv
bloa
dcp
rog
serv
ices
tels
hell
pim
gap
pman
YG
oNet
wm
play
er
Dataset2 example user
Usa
ge r
atio
000
005
010
015
0 20 40 601eminus
060
01
Figure 11 Relative time spent running each application for example users in each dataset Inset is thesemi-log plot of application popularity
are binary names even some popular applications may ap-pear unfamiliar For instance in Dataset1 ldquolauncherrdquo is thedefault home screen application on Android in Dataset2ldquogwesrdquo (Graphical Windows and Events Subsystem) is thegraphical shell on Window Mobile
The graphs show that relative application popularity dropsquickly for both users In sect83 we show that for all usersapplication popularity can be modeled by an exponentialdistribution
Diurnal patterns Interestingly application popularityis not stable throughout the day but has a diurnal patternlike the other aspects of smartphone use That is the rel-ative popularity of an application is different for differenttimes of the day Figure 12 illustrates this for an exam-ple user in Dataset2 We see for instance that tmailexewhich is a messaging application on Windows Mobile ismore popular during the day than night Time dependentapplication popularity was recently reported by Trestian etal based on an analysis of the network traffic logs from a3G provider [27] Our analysis based on direct observationof user behavior confirms this effect
0 4 8 12 17 22Hour of day
020
4060
80U
sage
rat
io (
)
servicesexewmplayerexecalendarexepoutlookexeWLMMessengercdialexepimgexecprogexeiexploreexehomeexetmailexe
Figure 12 Relative time spent with each applicationduring each hour of the day for a sample user andher top applications
comm 44
browsing 10
other 11
prod 19media 5
maps 5games 2system 5
Dataset1
comm 49
browsing 12 other 15
prod 2media 9
maps 2
games 10
system 1
Dataset2
Figure 13 Relative popularity of each applicationcategory across all users in each dataset
Aggregate view application popularity To pro-vide an aggregate view of what users use their smartphonesfor we categorize applications into eight distinct categoriesi) communication contains applications for exchanging mes-sages (eg email SMS IM) and voice calls ii) browsingcontains Web browser search and social networking appli-cations iii) media contains applications for consuming orcreating media content (eg pictures music videos) iv)productivity contains applications for calendars alarms andfor viewing and creating text documents (eg Office PDFreader) v) system contains applications for changing userpreferences and viewing system state (eg file explorer)vi) games vii) maps and viii) other contains applicationsthat we could not include in any of the categories aboveeg because we did not know their function
Figure 13 shows the mean relative popularity of each ap-plication category across all users in each dataset Whilethe results are not identical across the two datasets they aresimilar to a first order Communication dominates in bothBrowsing is another major contributor in both datasetsMaps media and games have a comparatively lower butnevertheless substantial share of user attention
Relationship to user demographic To understandthe extent to which user demographic determines applica-tion popularity Figure 14 shows the mean and 95 CI ofrelative popularity of each application category for differ-ent user demographics As for interaction time (sect31) wesee that the impact of user demographics in our datasets isminimal In Dataset2 the relative popularity of various ap-plication types is similar for each of the four demographicsIn Dataset1 there are noticeable differences in the mean forcommunication games and productivity applications Highschool students use communication and games applicationsmore while knowledge workers use productivity applications
Dataset1
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
High schoolKnowledge worker
Dataset2
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
SCOPBPULPU
Figure 14 The mean and 95 CI of relative pop-ularity of application categories among users of dif-ferent demographics
more However considering the overlap of confidence in-tervals these differences in application popularity are notstatistically significant
From this result and the earlier one on the lack of depen-dence between user demographic and interaction time (sect31)we conclude that user demographic at least as defined inour datasets cannot reliably predict how a user will use thephone While demographic information appears to help insome cases (for eg the variation in usage of productivitysoftware in Dataset1) such cases are not the norm and itis hard to guess when demographic information would beuseful Pending development of other ways to classify userssuch that these classes more predictably explain the varia-tion incorporating factors specific to a user appear necessaryThis insensitivity to user demographic has positive as wellas negative implications A negative is that personalizationis more complex we cannot predict a usersrsquo behavior byknowing their demographic A positive implication is thatthe range of user behaviors along many dimensions of in-terest can be found in several common demographics Thissimplifies the task of uncovering the range because recruitingsubjects across multiple demographics tends to be difficult
Relationship to interaction time We also study ifusers that interact more with their phones tend to use differ-ent applications For each dataset we sort users based ontheir average interaction time per day and partition theminto different classes For Dataset1 we use two classesone each for the top and the bottom half of the users ForDataset2 which has more users we use three classes for the
Dataset1
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
HighLow
Dataset2
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
rHighMediumLow
Figure 15 The mean and 95 CI of relative pop-ularity of application categories among differentclasses of users based on interaction time per day
top middle and bottom third of the users Figure 15 showsthe mean and 95 CI for relative time spent with each ap-plication category by each user class We see that users indifferent classes have similar application usage Thus wecannot explain why some users use the phone more simplybased on the applications that they use
43 Application SessionsWe now study the characteristics of application sessions
We conduct this analysis only for Dataset2 based on times-tamps for when an application is started and ended Dataset1does not contain this information precisely Because appli-cations can run in the background start and end refer tothe period when the application is in the foreground
Applications run per interaction We begin by study-ing the number of applications that users run in an interac-tion session Figure 16 shows that an overwhelming major-ity close to 90 of interactions include only one applicationThis graph aggregates data across all users We did not findstatistically significant differences between users A largefraction of sessions of all users have only one application
That interaction sessions very often have only one appli-cation session suggests that users tend to interact with theirsmartphone for one task (eg reading email checking cal-endar etc) at a time and most of these tasks require theuse of only one application
Application session lengths Because interactionsessions are dominated by those with only one application
892
7422 06 02 01
1 2 3 4 5 6
020
4060
8010
0
Apps per session
Per
cent
age
()
Figure 16 Histogram of the number of applica-tions called during each interaction session for allthe users in Dataset2
map
s
gam
es
com
m
med
ia
brow
sing
prod
syst
em
othe
r
110
1000
Ses
sion
leng
th (
s)
Figure 17 The mean and 95 CI of session lengthsof different application categories The y-axis is logscale
the overall properties of application sessions such as theirlengths are similar to those of interaction sessions (sect3)
However studying the session lengths of applications sep-arately reveals interesting insights Different applicationtypes have different session lengths as shown in Figure 17for the categories defined earlier Interactions with mapsand games tend to be the longest and those related to pro-ductivity and system tend to be the shortest
Further given an application different users run themfor different times Figure 18 shows this effect for a mes-saging application tmailexe and a browsing applicationiexploreexe For each application the mean session lengthsof users differ by more than two orders of magnitude
These observations help explain why users have differentsession lengths (sect32) They prefer different applicationsand those application sessions tend to have different lengthsFurther analysis confirms this phenomenon For instance ifwe divide users into two classes based on their mean sessionlengths the popularity of games is twice as high in the classwith high session lengths
5 TRAFFICIn this section we investigate traffic generated by smart-
phones Unlike interaction events and application use net-work traffic is not an intentional user action but a side-effectof those actions Most users are likely oblivious to how much
tmail1
1010
010
00
0 20 40 60 80 100User percentile
Ses
sion
leng
th (
s)iexplore
110
100
1000
0 20 40 60 80 100User percentile
Ses
sion
leng
th (
s)Figure 18 The mean and the upper end of the stan-dard deviation of session lengths of two applications
0 20 40 60 80 100
User percentile
0
1
10
100
1000
10000
Traffic per day (MB) Receive
Send
Figure 19 The mean and the upper end of the stan-dard deviation of the traffic sent and received perday by users in Dataset1
traffic they generate We show that the diversity and diur-nal patterns of this side-effect match those of user actionsthemselves
The analysis in this section includes only Dataset1 we donot have traffic information for Dataset2 In Dataset1 werecord all of the data sent (or received) by the phone exceptfor that exchanged over the USB link ie the traffic hereinincludes data over the 3G radio and the 80211 wireless link
51 Traffic per dayFigure 19 shows that the amount of traffic sent and re-
ceived per day differs across users by almost three orders ofmagnitude The traffic received ranges from 1 to 1000 MBand the traffic sent ranges from 03 to 100 MB The medianvalues are 30 MB sent and 5 MB received
Our results indicate that traffic generated in a day bysmartphone users is comparable to traffic generated by PCsa few years ago A study of a campus WiFi network revealedthat on average users were generating 27 MB of traffic in2001 and 71 MB in 2003 [12] A study of Japanese residen-tial broadband users in 2006 revealed that on average usersgenerate 1000 MB of traffic per day [11] This high level oftraffic has major implications for the provisioning of wirelesscarrier networks as smartphone adoption increases
Relationship to application types To investigate ifcertain types of applications are favored more by users thatgenerate more traffice we divide the users into two equalclasses based on their sum of sent and received traffic perday Figure 20 shows mean and 95 CI of relative popularityof each application category for each user class Expectedly
Dataset1
Per
cent
age
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
HighLow
Figure 20 The mean and 95 CI of relative popu-larity of each application category among high andlow traffic consumers
it shows that communication applications are more popularamong users that consume more traffic
52 ldquoInteractiverdquo trafficWe next estimate what fraction of the total traffic is ldquoin-
teractiverdquo that is has timeliness constraints Approachesto reduce power consumption of data transfers often advo-cate rescheduling network transfers for instance by delay-ing some transfers so that multiple of them may be bun-dled [1 22] Such policies are difficult to implement forinteractive traffic without hurting user response time andthus are likely to be less valuable if the bulk of the traffic isinteractive
We classify traffic as interactive if it was generated whenthe screen is on This classification method might classifysome background traffic as interactive We expect the errorto be low because the screen is on for a small fraction ofthe time for most users Because certain user interactionswith the phone begin right after traffic exchange (eg anew email is received) we also consider traffic received in asmall time window (1 minute) before the screen is turned onas having timeliness constraints The results are robust tothe exact choice of time window Indeed some of the trafficthat we classify as interactive might be delay tolerant ega new email to be sent might be deferred for a little whileHowever the volume of traffic received by the phone whichusers would rather see immediately dominates by one or-der of magnitude the volume that is sent and delay-tolerantmessaging applications such as email contribute only a smallchunk of all traffic
Figure 21 shows the fraction of interactive traffic for eachuser We see that for about 90 of the users over 50of the traffic is interactive but for the rest almost none oftheir traffic is interactive Stated differently for differentsmartphone users almost all to almost none of the traffic isgenerated by applications in the background The extremesrepresent disparate ways in which people use smartphonesand which applications on the smartphone generate mosttraffic Our results imply that the energy savings that can behad by rescheduling network activity will vary across users
53 Diurnal patternsFigure 22(a) shows the diurnal pattern for an example
user with a sharp decline at night Figure 22(b) showsthe strength of the diurnal pattern for individual users by
0 20 40 60 80 100
User percentile
00
02
04
06
08
10
Interactive fraction
Figure 21 The fraction of interactive traffic
0 5 10 15 20
Hour of day
0
5
10
15
Traffic per hour (MB)
Receive
Send
(a)
0 20 40 60 80 100
User percentile
2
4
6
8
10
12
Diurnal ratio
Receive
Send
(b)
Figure 22 (a) The mean and 95 CI for trafficgenerated per hour by an example user (b) Thediurnal ratio of traffic per hour for all users
plotting the diurnal ratio of traffic As defined in sect33 thediurnal ratio reflects how much higher the mean during thepeak hour is to the overall mean
We see that the diurnal ratio varies across users but mosthave a strong diurnal behavior 80 of them generate overtwice their average amount of traffic in their peak hour Thisbehavior is likely a direct result of the high proportion ofinteractive traffic for most users and that user interactionsthemselves have a diurnal pattern
6 ENERGY CONSUMPTIONThe final aspect of smartphone usage that we investigate
is energy consumption Energy drain depends on two fac-tors i) user interactions and applications and ii) platformhardware and software If the second factor dominates theenergy drain of various users with the identical smartphonewill be similar Otherwise the energy drain will be as di-verse as user behaviors
We estimate the amount of energy drain based on the re-maining battery indicator which varies between 0 and 100If the battery indicator has gone down by X in a time pe-riod for a battery with capacity Y mAh we compute theenergy drain in that period to be X middot Y mAh 1 Given thatbatteries are complex electro-chemical devices [14 20] thiscomputation is approximate It assumes that the batterylevel indicator is linear with respect to energy drain
1mAh is technically a unit of charge yet is commonly used toindicate energy drain because battery voltage during normaloperations is typically constant For phones in our datasetthis is 4V so multiplying a mAh reading by 4 would yieldan accurate energy reading in milli-watt-hours
020
4060
8010
0
0 4 8 12 16 20Time (hour)
Bat
tery
leve
l (
) Benchmark1Benchmark2
Figure 23 Timelapse of the remaining battery levelindicator in controlled experiments with two differ-ent workloads at room temperature
110
100
1000
0 20 40 60 80 100User percentile
Bat
tery
Dra
in (
mA
h)
Figure 24 The mean and the upper end of the stan-dard deviation of one hour energy drain for Dataset1users during discharge periods
Controlled experiments suggest that the linearity assump-tion holds to a first order We run a benchmark load thatdrains the battery at a fixed rate in room temperature Un-der this benchmark if the battery level indicator decreaseslinearly with time it must be linear with respect to energydrain Figure 23 shows that the level indicator decreasesroughly linearly for two different benchmarks Benchmark1turns the screen on and off periodically Benchmark2 com-putes and idles periodically We conclude thus that the levelindicator can be used to estimate energy drain
Figure 24 shows the mean and standard deviation of en-ergy that users drain in an hour This graph is computed us-ing only periods in which the battery is not charging becauseenergy drain in those periods are of primary interest Wesee a two orders of magnitude difference among users Whileheaviest users drain close to 250 mAh the lightest of usersdrain only 10 mAh If the battery capacity is 1200 mAhthis leads to a lifetime variation from about 4 to 120 hours
Figure 25(a) shows for an example user that the drain isnot the same throughout the day but has diurnal variationsin which more energy is consumed during the day than dur-ing the night For this user the level of energy consumedchanges by roughly a factor of five Figure 25(b) plots thediurnal ratio of energy use for all users It shows that diurnalvariations occur with different strengths for all users
2060
100
140
0 4 8 12 16 20Hour of the day
Bat
tery
Dra
in (
mA
h)
(a)
0 20 40 60 80 100
12
34
56
7
User percentile
Diu
rnal
rat
io(b)
Figure 25 (a) The mean and 95 CI of energy drainof an example Dataset1 user (b) Diurnal ratio ofall users in Dataset1
Our results show that user activities contribute heavilytowards energy drain users in Dataset1 who have identi-cal smartphones drain energy at different rates and energydrain has diurnal patterns In the future we will developmethods to accurately quantify the energy consumption ofthe platform from that due to user-induced workload
7 IMPLICATIONS OF USER DIVERSITYWe uncover a surprising level of diversity among smart-
phone users For almost every aspect of usage that we studywe find one or more orders of magnitude difference betweenusers Our findings strongly motivate the need for customiz-ing smartphones to their users We believe that this needis greater than that for customizing ordinary cellphones orlaptops Ordinary cellphones do not have as rich an appli-cation environment Laptops are not as portable and aremore resource rich For example many users plug-in theirlaptops while using them
Customization can help at all levels Consider somethingas low-level as the battery Suppose we want batteries to beboth lightweight and last for at least a day with a high proba-bility Meeting the latter goal for all users of a given platformwill require catering to the heaviest users But that will leadto unnecessarily heavy batteries for many users (Highercapacity batteries are heavier) Offering multiple types ofbatteries with different lifetime-weight tradeoffs provides away out of this bind
At levels where intelligent mechanisms to improve user ex-perience or reduce energy consumption reside user diversitymotivates adapting to the smartphone user Driving thesemechanisms based on average case behaviors may not beeffective for a large fraction of the users
The ease and utility of customization depends on twoproperties of user behavior First despite quantitative dif-ferences there must be qualitative similarities among usersFor instance we should be able to describe the behavior ofall users with the same model Different users may havedifferent parameters of this model which will then lead toquantitative differences among them The presence of qual-itative similarities imply that users are not arbitrary pointsin space and it significantly simplifies the task of learninguser behavior Second user behavior in the past must alsobe predictive of the future Otherwise customization basedon past behaviors will be of little value in the future
In the next two sections we present evidence that theseproperties hold for several key aspects of smartphone usage
Dataset1 example user
000
00
010
002
00
030
0 100 200 300 400Session length (s)
Den
sity
Dataset2 example user
000
005
010
015
020
0 20 40 60 80 100Session length (s)
Den
sity
Figure 26 The PDF of session length for sampleusers of each dataset
In sect8 we show that user sessions and relative applicationpopularity can be described using simple models In sect9 weshow that energy drain can be described as well as predictedusing a ldquotrend trablerdquo framework
8 SMARTPHONE USAGE MODELSIn this section we develop simple models that describe
three aspects of smartphone usage ndash session lengths inter-arrival time between sessions and application popularityThe models are common across users but have different pa-rameters for different users While they do not completelydescribe user behavior they capture first order factors andrepresent a first step towards more complete modeling ofsmartphone usage More importantly along with the resultsof the next section they show that qualitative similaritiesdo exist among users
81 Session LengthsWe first consider the statistical properties of session length
distributions of users We find that session length valuestend to stationary With the KPSS test for level stationar-ity [13] 90 of the users have a p-value greater than 01The presence of stationarity is appealing because it suggeststhat past behavior is capable of predicting the future
We also find that session lengths are independent that isthe current value does not have a strong correlation with thevalues seen in the recent past With the Ljung-Box test forindependence [15] 96 of the users have a p-value that isgreater than 01
Stationarity and independence considered together sug-gest that session length values can be modeled as iid sam-ples from a distribution Choosing an appropriate distribu-tion however is complicated by the nature of the sessionlengths Most sessions are very short and the frequencydrops exponentially as the length increases However in-consistent with exponential behavior there are some verylong sessions in the tail for each user
We find that a mixture of exponential and Pareto distri-butions can model both ends of the spectrum The formercaptures short sessions and the latter captures long sessionsThat is session lengths can be described by the followingmixture model
r middot Exp(λ) + (1 minus r) middot Pareto(xm α)
In this equation r is the relative mix of the two distribu-tions λ is the rate of the exponential and xm and α are thelocation and shape parameters of the Pareto distribution
The location for a Pareto distribution represents the min-imum possible value of the random variable The location
010
0020
0030
00
0 1000 2000 3000Observed quantiles
Mod
el q
uant
iles
Figure 27 QQ plot of session lengths model for asample user
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed m
ix
(a) Relative mix (r)
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed r
ate
(b) Exponential rate (λ)
020
4060
0 20 40 60 80 100User percentile
Tim
eout
(s)
(c) Pareto location (xm)
minus1
0minus
04
02
06
10
0 20 40 60 80 100User percentile
Est
imat
ed s
hape
(d) Pareto shape (α)
Figure 28 Distribution of inferred model parame-ters that describe session length values of users inboth datasets
value that offers the best fit is the screen timeout value ofthe user because the session length PDF has a spike at thisvalue The spike corresponds to short sessions that are endedby the timeout (when the user forgets to switch the screenoff) we confirm this using controlled experiments with dif-ferent timeout values Figure 26 shows this spike at 60 and15 seconds for example users from each dataset The time-out provides a natural division between the two componentdistributions We automatically infer its approximate valueusing a simple spike detection algorithm
We use the EM algorithm to infer the maximum likelihoodestimation (MLE) of the remaining three parameters [6]Figure 27 shows the quality of this fit for an example userusing the QQ plot [3] Almost all quantiles are along they = x line indicating a good fit
Figure 28 shows the four inferred parameters for varioususers While users can be modeled using the same mix-
040
0080
00
0 4000 8000Observed quantiles
Mod
el q
uant
iles
Figure 29 QQ plot of session offtime model for asample user
050
010
0015
000 20 40 60 80 100
User percentileE
stim
ated
sca
le
(a) Weibull scale
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed s
hape
(b) Weibull shape
Figure 30 Distribution of inferred model parame-ters that describe the distribution of time betweensessions for users in both datasets
ture model the parameters of this model vary widely acrossusers Because of the way we construct our model the dis-tribution of the parameter r and xm also provide insightinto how frequently usersrsquo screen is switched off by the time-out and the relative popularity of different timeout values60 seconds is the most popular value likely because it is themost common default timeout and many users never changethe default
82 Time between SessionsWe find that the Weibull distribution can explain the
screen off times This distribution has two parameters re-ferred to as its scale and shape We find the MLE for theseparameters for each user From the QQ-plot in Figure 29we notice that the model predicts a greater probability ofseeing some very large offtimes than are observed in thedatasets However the probability of seeing these large off-times is small there are 27 data points that have a y-value greater than 8000 in that graph Hence we believethat Weibull provides a good fit for the length of intervalsbetween interactions
Figure 30 shows the distribution of the estimated shapeand scale of the fitted Weibull distributions Interestinglythe shape is consistently less than one Weibull shape valuesless than one indicate that the longer the screen has been offthe less likely it is for it to be turned on by the user This be-havior has interesting implications for power saving policiesFor instance periodic activities such as checking for emailwhen the screen has been off for a while may be deferred orrescheduled if needed without hurting user experience
020
4060
8010
0
0 20 40 60 80 100User percentile
MS
E (
)
(a) MSE
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed r
ate
(b) Rate parameter
Figure 31 (a) The mean square error (MSE) whenapplication popularity distribution is modeled usingan exponential (b) The inferred rate parameter ofthe exponential distribution for different users
83 Application PopularityWe find that for each user the relative popularity of ap-
plications can be well described by a simple exponential dis-tribution This qualitative invariant is useful for instanceto predict how many applications account for a given frac-tion of user attention For the example users in Figure 11this facet can be seen in the inset plots the semi-log of thepopularity distribution is very close to a straight line
Figure 31(a) shows that this exponential drop in applica-tion popularity is true for almost all users the mean squareerror between modeled exponential and actual popularitydistribution is less than 5 for 95 of the users
Figure 31(b) shows the inferred rate parameter of the ap-plication popularity distribution for various users We seethat the rate varies by an order of magnitude from 01 toalmost 1 The value of the rate essentially captures the paceof the drop in application popularity Lower values describeusers that use more applications on a regular basis Oneimplication of the wide range is that it may be feasible toretain all popular applications in memory for some users andnot for others
9 PREDICTING ENERGY DRAINIn this section we demonstrate the value of adapting to
user behaviors in the context of a mechanism to predict fu-ture energy drain on the smartphone Such a mechanismcan help with scheduling background tasks [9] estimatingbattery lifetime and improving user experience if it is esti-mated that the user will not fully drain the battery until thenext charging cycle [2 21]
Despite its many uses to our knowledge none of thesmartphones today provide a prediction of future energydrain Providing an accurate prediction of energy drain isdifficult2 User diversity of energy use makes any static pre-diction method highly inaccurate Even for the same userthere is a high variance in energy usage Figure 32 shows thisvariance by plotting the ratio of the standard deviation tothe mean energy drain over several time periods for users inDataset 1 The standard deviation of 10-minute windows ishigher than three times the mean for a fifth of the users The
2Energy drain prediction on smartphones is more challeng-ing than what is done for laptops Laptops provide a pre-diction of battery lifetime only for active use Smartphonepredictions on the other hand must cover multiple activeidle periods to be useful
01
23
45
0 20 40 60 80 100User percentile
Err
or r
atio
2 hours1 hour10 minutes
Figure 32 Battery drain predictors have to copewith the high variance in drain
time
window to
predict
n chunks
hellip
Trend Table
hellip
hellip
Figure 33 A userrsquos trend table captures how batterydrain in n consecutive chunks relates to drain in thesubsequent window
bursty nature of user interactions is one cause of this highvariance The variance is high even at longer time scalesThe standard deviation for two hour windows is larger than150 of the mean for half the users The diurnal patternsthat we showed earlier are one cause of high variance oversuch large time scales
Although the variance is high there are patterns in auserrsquos behaviors For instance we showed earlier that aphone that has been idle for a while is likely to remain idlein the near future and it will thus continue to draw energyat a similar rate
We hypothesize that a predictor tuned to user behaviorcan be accurate We present a simple personalized energydrain predictor that incorporates various user-specific fac-tors such as length of idle periods and different types ofbusy periods Instead of explicitly identifying these factorsour predictor captures them in a ldquotrend tablerdquo framework
Our framework is shown in Figure 33 Each entry in thetrend table is indexed by an n-tuple that represents energyusage readings from adjacent chunks of size δ each Thisindex points to energy usage statistics across all time win-dows of size w in the history that follow the n chunk valuesrepresented by that index Thus the trend table captureshow the energy usage in n consecutive chunks is related tothe energy use in the subsequent time windows of size w
In this paper we use n = 3 δ = 10 minutes and maintaindifferent trend tables for each window size w for which theprediction is required To keep the trend table small wequantize the energy readings of chunks We maintain themean and standard deviation for all subsequent windowsthat map to that index
To make a prediction we use the energy readings fromthe n immediately preceding chunks Let these readings be(x1 xn) Then we search for indices in the table thatdiffer from (x1 xn) by no more than a threshold s and
GenericShortminustermTimeminusofminusdayPersonalized
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
(a) 1-hour window
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
GenericShortminustermTimeminusofminusdayPersonalized
(b) 2-hour window
Figure 34 The ratio of error to true battery drainfor different energy drain predictors
predict based on them such that similar indices have higherrelative weight More precisely the prediction is
X
e stforalli=1n|he
iminusxi|lts
ke times ue
P
ke
where e iterates over all tuples in the trend table hei is
the value of the chunk i of the tuple e ue represents thestatistics stored for that index and the relative weight ke =s minus maxi=1n |he
i minus xi| For results shown here we uses = 05
Figures 34 shows how well our Personalized predictor worksfor predicting usage for two different future time windowsThere is one point per user in each graph which correspondsto the median error seen across many predictions for thatuser We see that the 90th percentile error is 25 for 1-hourwindow and 40 for 2-hour window
To place this level of accuracy in context the figure alsoshows the accurary of three alternative predictors TheGeneric predictor uses a trend table but instead of buildinga user-specific table it builds one that combines all usersWe see that its 90th percentile error is roughly twice thatof the Personalized predictor Its predictions are worse forhalf of the users
The Short-term predictor predicts energy drain simply asthe amount drained in the immediately preceding windowof the same size For both time windows its 90th percentileerror is no better than that of the Generic predictor
The Time-of-day predictor predicts energy drain based onvalues observed for the user at the same time of the day inthe past similar to that proposed by Banerjee et al [2]Thus this predictor has some degree of personalization butit learns user behavior in less detail than the Personalizedpredictor We see that its performance for 1-hour windowis similar to the Short-term predictor For 2-hour windowits error is higher than the Personalized predictor across theboard though its worst case error is lower than the Short-term and Generic predictors
Overall these results demonstrate the value of appropri-ately learning user behaviors to design intelligent mecha-nisms on smartphones
10 RELATED WORKIn a range of domains there is a rich history of work that
characterizes user workloads However because smartphoneadoption has gathered pace relatively recently our work rep-resents one of the few to study how people use smartphones
Along with other recent works it helps complete the pic-ture Banerjee et al and Rahmati et al report on batterycharging behaviors [2 18] Like us they find considerablevariation among users Banerjee et al also propose a pre-dictor that estimates the excess energy of the battery atthe time of charging using a histogram of past battery us-age [2] Shye et al study the power consumption character-istics of 20 users [24] They infer properties such as whichcomponents in the phone consume most power and exploreoptimizations based on these characteristics Rahmati andZhong study 14 users of a particular demographic to studywhich applications are popular in that demographic wherethe phone is used and how it is shared among users [19]
In contrast to these works we focus on understandingdifferent aspects of smartphone use (eg interactions andtraffic) and on exposing the diversity of user behaviors in-stead of only the average behavior Our study also entailsan order of magnitude more users than previous efforts
Like us other researchers have developed logging utilitiesMyExperience is one such early utility [10] and the worksabove involve custom utilities as well Our focus in thispaper is not on the development of logging tools but onanalyzing their data to gain insight into user behavior
There is a body of work in modeling the aggregate be-havior of mobile users Using traces from a large cellularoperator some network related aspects of mobile usage havebeen modeled Halepovic et al and Williamson et al re-port that call arrivals are bursty and present diurnal pat-terns [28] Willkomm et al and Brown et al report thatmobile users call duration can be approximately modeled bya lognormal distribution [4 29] We use traces collected onthe mobile device itself and focus on modeling the interac-tions of individual users instead of the aggregate behavior
11 CONCLUSIONSBy studying 255 users of two different smartphone plat-
forms we comprehensively characterized user activities andtheir impact on network and battery We quantify manyhitherto unknown aspects of smartphone usage User diver-sity is an overarching theme in our findings For instancedifferent users interact with their phones 10-200 times a dayon average the mean interaction length of different users
is 10-250 seconds and users receive 1-1000 MB of data perday where 10-90 is received as part of active use
This extent of user diversity implies that mechanisms thatwork for the average case may be ineffective for a large frac-tion of the users Instead learning and adapting to userbehaviors is likely to be more effective as demonstrated byour personalized energy drain predictor We show that qual-itative similarities exist among users to facilitate the devel-opment of such mechanisms For instance the longer theuser has not interacted with the phone the less likely it isfor her to start interacting with it and application popular-ity for a user follows a simple exponential distribution
Our study points to ways in which smartphone platformsshould be enhanced Effective adaptation will require futureplatforms to support light-weight tools that monitor andlearn user behaviors in situ It will also require them toexpose appropriate knobs to control the behavior of lower-level components (eg CPU or radio)
12 ACKNOWLEDGEMENTSWe are grateful to the users in Dataset1 that ran our log-
ger and to the providers of Dataset2 Elizabeth Dawson pro-vided administrative support for data collection Jean BolotKurt Partridge Stefan Saroiu Lin Zhong and the MobiSysreviewers provided many comments that helped improve thepaper We thank them all This work is supported in partby the NSF grants CCR-0120778 and CNS-0627084
13 REFERENCES[1] Balasubramanian N Balasubramanian A
and Venkataramani A Energy consumption inmobile phones A measurement study andimplications for network applications In IMC (2009)
[2] Banerjee N Rahmati A Corner M D
Rollins S and Zhong L Users and batteriesInteractions and adaptive energy management inmobile systems In Ubicomp (2007)
[3] Becker R Chambers J and Wilks A The newS language Chapman amp HallCRC 1988
[4] Brown L Gans N Mandelbaum A Sakov
A Shen H Zeltyn S and Zhao L Statisticalanalysis of a telephone call center Journal of theAmerican Statistical Association 100 469 (2005)
[5] Church K and Smyth B Understanding mobileinformation needs In MobileHCI (2008)
[6] Dempster A Laird N and Rubin D Maximumlikelihood from incomplete data via the EM algorithmJournal of the Royal Statistical Society Series B(Methodological) 39 1 (1977)
[7] Esfahbod B Preload An adaptive prefetchingdaemon Masterrsquos thesis University of Toronto 2006
[8] Falaki H Govindan R and Estrin D Smartscreen management on mobile phones Tech Rep 74Center for Embedded Networked Sesning 2009
[9] Flinn J and Satyanarayanan M Managingbattery lifetime with energy-aware adaptation ACMTOCS 22 2 (2004)
[10] Froehlich J Chen M Consolvo S
Harrison B and Landay J MyExperience Asystem for in situ tracing and capturing of userfeedback on mobile phones In MobiSys (2007)
[11] Fukuda K Cho K and Esaki H The impact ofresidential broadband traffic on Japanese ISPbackbones SIGCOMM CCR 35 1 (2005)
[12] Kotz D and Essien K Analysis of a campus-widewireless network Wireless Networks 11 1ndash2 (2005)
[13] Kwiatkowski D Phillips P Schmidt P and
Shin Y Testing the null hypothesis of stationarityagainst the alternative of a unit root Journal ofEconometrics 54 1-3 (1992)
[14] Lahiri K Dey S Panigrahi D and
Raghunathan A Battery-driven system design Anew frontier in low power design In Asia SouthPacific design automationVLSI Design (2002)
[15] Ljung G and Box G On a measure of lack of fitin time series models Biometrika 65 2 (1978)
[16] Smartphone futures 2009-2014 httpwwwportioresearchcomSmartphone09-14html
[17] Rahmati A Qian A and Zhong L
Understanding human-battery interaction on mobilephones In MobileHCI (2007)
[18] Rahmati A and Zhong L Humanndashbatteryinteraction on mobile phones Pervasive and MobileComputing 5 5 (2009)
[19] Rahmati A and Zhong L A longitudinal study ofnon-voice mobile phone usage by teens from anunderserved urban community Tech Rep 0515-09Rice University 2009
[20] Rao R Vrudhula S and Rakhmatov D
Battery modeling for energy aware system designIEEE Computer 36 12 (2003)
[21] Ravi N Scott J Han L and Iftode L
Context-aware battery management for mobilephones In PerCom (2008)
[22] Sharma A Navda V Ramjee R
Padmanabhan V and Belding E Cool-TetherEnergy efficient on-the-fly WiFi hot-spots usingmobile phones In CoNEXT (2009)
[23] Sharma C US wireless data market update - Q32009 httpwwwchetansharmacomusmarketupdateq309htm
[24] Shye A Sholbrock B and G M Into the wildStudying real user activity patterns to guide poweroptimization for mobile architectures In Micro (2009)
[25] Snol L More smartphones than desktop PCs by2011 httpwwwpcworldcomarticle171380more_smartphones_than_desktop_pcs_by_2011
html
[26] Sohn T Li K Griswold W and Hollan J
A diary study of mobile information needs In SIGCHI(2008)
[27] Trestian I Ranjan S Kuzmanovic A and
Nucci A Measuring serendipity Connecting peoplelocations and interests in a mobile 3G network InIMC (2009)
[28] Williamson C Halepovic E Sun H and Wu
Y Characterization of CDMA2000 Cellular DataNetwork Traffic In Local Computer Networks (2005)
[29] Willkomm D Machiraju S Bolot J and
Wolisz A Primary users in cellular networks Alarge-scale measurement study In DySPAN (2008)
- Introduction
- Data Collection
-
- Dataset1
- Dataset2
- Representativeness of conclusions
-
- User Interactions
-
- Interaction Time
- Interaction Sessions
- Diurnal Patterns
-
- Application Usage
-
- Number of applications
- Application Popularity
- Application Sessions
-
- Traffic
-
- Traffic per day
- ``Interactive traffic
- Diurnal patterns
-
- Energy Consumption
- Implications of User Diversity
- Smartphone Usage Models
-
- Session Lengths
- Time between Sessions
- Application Popularity
-
- Predicting Energy Drain
- Related Work
- Conclusions
- Acknowledgements
- References
-
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Voi
ce u
sage
rat
io
Dataset1Dataset2
Figure 1 Ratio of voice usage to total usage Thex-axis is user percentile and users are sorted in de-creasing order of ratio
23 Representativeness of conclusionsAn important concern for user studies such as ours is
whether the resulting conclusions represent the entire popu-lation There are two potential sources of bias in our data i)the users are not representative and ii) the measured usageis not representative We believe that our conclusions aregeneral The first concern is alleviated by the fact that asidefrom some quantitative differences we find remarkable con-sistency among users in the two datasets This consistencysuggests generality given that the two datasets are gatheredindependently on different platforms and Dataset2 was pro-fessionally designed to be representative
The second concern stems from the possibility that usersmay not be using the monitored smartphones as their pri-mary devices or that the usage during the monitoring periodmay not be normal All users in Dataset2 used the providedsmartphones as their primary devices We do not know thisaspect with certainty for Dataset1 but we understand fromanecdotal evidence that some users used these devices astheir only phones and others took advantage of the unlim-ited minutes and text plans We study voice usage as in-dicative of the extent to which users relied on the monitoreddevices Higher voice usage suggests use as primary phonesFigure 1 shows the ratio of time users spent in phone callsto total time spent interacting with the phone (sect3) We seethat the overall voice usage in Dataset1 was higher thanthat in Dataset2 in which all users used the phone as theirprimary device
Given that the monitored devices tended to be primaryand the long duration of the monitoring interval we con-jecture that our traces predominantly capture normal us-age Earlier work has pointed to the possibility of an initialadoption process during which usage tends to be differentthan long-term usage [19] To show that our traces are notdominated by the initial excitement of users or other specialevents that cause usage to be appreciably different from thenormal usage Figure 2 shows the average interaction timeper day (sect3) in the first and second halves of the datasets foreach user We see roughly similar usage in the two halvesDetailed investigation shows that the visible differences inthe averages of the two halves especially in Dataset1 arenot statistically significant Other measures of usage (egnetwork activity) look similar We do not claim that in-stances of abnormal usage are absent in the datasets but
110
100
1000
0 20 40 60 80 100User percentile
Act
ive
time
(min
utes
)
Dataset1 1st halfDataset1 2nd halfDataset2 1st halfDataset2 2nd half
Figure 2 Total interaction per day during the firstand second halves of study for each user Userswithin each dataset are sorted based on the interac-tion time in the first half The y-axis is log scale
the monitored period was long enough for our results to notbe impacted by such instances
3 USER INTERACTIONSWe begin our analysis by studying how users interact with
their smartphones independent of the application used Wecharacterize application use in the next section and the im-pact of user actions on network traffic and energy drain inthe following sections
We define an interaction interval also referred to as a ses-sion in this paper differently for each dataset In Dataset1we deem a user to be interacting with the phone wheneverthe screen is on or a voice call is active In Dataset2 aninteraction is defined as the interval that an application isreported to be on the foreground This includes voice callsbecause on Windows Mobile a special program (ldquocprogexerdquo)is reported in the foreground during voice calls
31 Interaction TimeFigure 3(a) shows a basic measure of user interactionmdashthe
number of minutes in a day that a user interacts with thesmartphone The plot shows the mean and the standarddeviation of this number for each user For visual clarityin such graphs we plot only the upper end of the standarddeviation plotting both ends occludes the other curves Theinterested reader can estimate the lower end since standarddeviation is symmetric around the mean
Dataset1 users tend to have more interaction minutes be-cause as we show later they tend to have longer interactionsessions while having a similar number of sessions Withineach dataset however there is an order of magnitude dif-ference among users In Dataset1 the lower end is only 30minutes in a day But the high end is 500 minutes which isroughly eight hours or a third of the day We are surprisedby this extremely high level of usage
Figure 3(a) also shows that users cover the entire rangebetween the two extremes and are not clustered into dif-ferent regions The lack of clusters implies that effectivepersonalization will likely need to learn an individual userrsquosbehavior rather than mapping a user to one or a few pre-defined categories
We examine two factors that can potentially explain theextent to which a user interacts with the phone but find that
110
100
1000
0 20 40 60 80 100User percentile
Act
ive
time
(min
utes
)
Dataset1Dataset2
(a) All users
High schoolKnowledge worker
110
100
1000
0 20 40 60 80 100User percentile
Act
ive
time
(min
utes
)
(b) Dataset1
SCOPBPULPU1
1010
010
00
0 20 40 60 80 100User percentile
Act
ive
time
(min
utes
)
(c) Dataset2
Figure 3 The mean and the upper end of the standard deviation of interaction minutes per day (a) Allusers in each dataset (b)amp(c) Different demographics in the two datasets The y-axes are log scale
neither is effective The first is that heavier users use dif-ferent types of applications (eg games) than lighter usersBut we find that the relative popularity of application typesis similar across classes of users with different interactiontimes (sect42) The second is user demographic But asFigures 3(b) and 3(c) show the interaction times are sim-ilar across the different demographics in the two datasetsWithin each demographic user interaction times span theentire range In sect42 we show that user demographic doesnot predict application popularity either
To understand the reasons behind diversity of user inter-action times we study next how user interaction is spreadacross individual sessions This analysis will show that thereis immense diversity among users in both the number of in-teraction sessions per day and the average session length
32 Interaction SessionsInteraction sessions provide a detailed view of how a user
interacts with the phone Their characteristics are impor-tant also because energy use depends not only on how longthe phone is used in aggregate but also on the usage distribu-tion Many short interactions likely drain more energy thanfew long interactions due to the overheads of awakening thephone and radio Even with negligible overheads batterylifetime depends on how exactly energy is consumed [20]Bursty drain with high current levels during bursts can leadto a lower lifetime than a more consistent drain rate
Figure 4(a) shows the number of sessions per day for dif-ferent users We again see a wide variation Individual usersinteract with their smartphone anywhere between 10 to 200times a day on average
Figure 4(b) shows the mean and standard deviation of in-teraction session lengths Dataset1 users tend to have muchlonger sessions than Dataset2 users Given that they haveroughly similar number of interactions per day as seen inFigure 4(a) their longer sessions explain their higher inter-action time per day as seen in Figure 3(a)
Within each dataset the mean session length varies acrossusers by an order of magnitude Across both datasets therange is 10-250 seconds
Explaining the diversity in session lengths Severalhypothesis might explain the differences in different usersrsquo
Dataset1Dataset21
1010
0
0 20 40 60 80 100User percentile
Ses
sion
cou
nt
(a) Number of sessions
110
100
1000
0 20 40 60 80 100User percentile
Mea
n se
ssio
n le
ngth
(s)
Dataset1Dataset2
(b) Session length
Figure 4 The mean and the upper end of the stan-dard deviation for the number of sessions per dayand the session length The y-axes are log scale
050
100
200
0 20 60 100 140Session length (s)
Ses
sion
cou
nt
Dataset1Dataset2
(a)
050
150
250
Ses
sion
cou
nt
lt10
10minus
20
20minus
30
30minus
40
40minus
60
gt60
Session length (s)
(b)
Figure 5 (a) Scatterplot of session count per dayand mean session length of various users (b) Themean and 95 CI of session count per day for usersin Dataset2 with different mean session lengths
session lengths One hypothesis is that users with longersessions concentrate their smartphone usage in fewer ses-sions Figure 5 shows however that there is little correla-tion between usersrsquo number of sessions and session lengthFigure 5(a) shows a scatterplot of session count versus meanlength for different users There is one data point for eachuser Figure 5(b) shows the dependence of session count onsession length by aggregating data across Dataset2 users Itplots the observed mean and 95 confidence interval (CI) for
Dataset1 example userDataset2 example user
00
02
04
06
08
10
0 200 600 1000Session length (s)
CD
F
(a) Session lengths
Dataset1 example userDataset2 example user
00
02
04
06
08
10
0 1000 2000 3000Time between sessions (s)
CD
F(b) Time between sessions
Figure 6 CDFs of session length and time betweensessions for two example users The x-axis ranges inthe graphs are different
session counts per day for users with different mean sessionlengths The differences in the session counts are not sta-tistically significant In other words it is not the case thatusers who have longer sessions have fewer or more sessions
Our other hypotheses are related to application use Thesecond hypothesis is that users run varying numbers of ap-plications during an interaction and users that tend to usemore applications per session have longer sessions The thirdhypothesis is that users run different applications and someapplications such as maps have longer sessions than othersThe fourth one is that even for the same application usershave different session lengths
Our analysis of application use in sect4 reveals that the sec-ond hypothesis is not explanatory as users overwhelminglyuse only one application per session It also reveals that thethird and fourth hypotheses are likely contributors to diver-sity in session lengths Note that the inability of applicationtypes to explain interaction time per day which we men-tion in the previous section is different from their ability toexplain session lengths
Distribution of a single userrsquos sessions We find thatfor any given user most of the sessions are short but someare very long Figure 6(a) shows the CDF of session lengthsfor two example users The median session length is lessthan a minute but some are longer than an hour (not shownin the graph) A similar skewed distribution can be seen forall users in our datasets albeit with different median andmean session length values This highly skewed distributionalso explains why the standard deviations in Figure 4(b) arehigh relative to the mean In sect81 we show how sessionlengths depend on the screen timeout values
Figure 6(b) shows that the time between sessions whenthe phone is not used also has a skewed distribution Mostare short (relative to the mean) but some are very long Weshow later that these off periods have the property that thelonger a user has been in one of them the greater the chancethat the user will continue in this state
33 Diurnal PatternsWe now study diurnal patterns in interaction The pres-
ence of such patterns has several consequences For instancethe length of time a given level of remaining battery capacitylasts will depend on the time of day
Figure 7 shows for two example users that as expectedstrong diurnal patterns do exist As a function of the hour
Dataset1Dataset2
16
1116
21
0 100 200 300 400Total interaction (minutes)
Diu
rnal
rat
io
(a)
lt30
30minus
60
60minus
90
90minus
120
120minus
150
150minus
200
200minus
300
gt30
0
01
23
45
Diu
rnal
rat
io
Total interaction (minutes)
(b)
Figure 8 (a) Scatterplot of diurnal ratio of interac-tion time per hour and interaction minutes per day(b) The mean and 95 CI of diurnal ratio vs totalinteraction time per day
of the day Figure 7(a) plots the mean number of interactionminutes per hour It also plots the 95 confidence interval(CI) around the mean which can be used to judge if thedifferences in the means are statistically significant We seea clear pattern in which daytime use is much higher thannightime use though the exact pattern for different users isdifferent
Figure 7(a) also shows that usage at hours in the night islow but not completely zero We believe that this non-zerousage stems from a combination of irregular sleeping hoursand users using their devices (eg to check time) when theyget up in the middle of the night
To capture the significance of the diurnal pattern for auser we define the diurnal ratio as the ratio of the meanusage during the peak hour to the mean usage across allhours A diurnal ratio of one implies no diurnal pattern andhigher values reflect stronger patterns Figure 9(a) plots thediurnal ratio in interaction time for all users It shows thatwhile diurnal ratios vary across users roughly 70 of theusers in each dataset have a peak hour usage that is morethan twice their mean usage
Explaining the diversity in diurnal patterns Tohelp explain the variability among usersrsquo diurnal ratios inFigure 8 we study its dependence on interaction time Fig-ure 8(a) shows a scatterplot of the diurnal ratio and themean interaction time per day We see that the diurnal ra-tio tends to be inversely correlated with interaction timeFigure 8(b) shows this negative correlation more clearly byaggregating data across users It plots the mean and 95CI of the diurnal ratio of total interaction time per day forusers with different total interaction times The diurnal ra-tio decreases as interaction time increases This inverse re-lationship suggests that heavy users tend to use their phonemore consistently during the day whereas light users tendto have concentrated use during certain hours of the day
Understanding the source of diurnal patterns Thevariation in interaction time of a user across the day can re-sult from variation in the number of interaction sessions orthe length of individual sessions We find that both factorscontribute Users tend to have different number of sessionsas well as different session lengths at different hours of theday Figures 7(b) and 7(c) illustrate this point for two ex-ample users They plot the mean number of sessions andthe mean session length for each hour of the day
Dataset1 example user0
510
1520
25
0 3 6 9 13 17 21Hour of day
Tot
al in
tera
ctio
n (m
inut
es) Dataset1 example user
05
1015
0 3 6 9 13 17 21Hour of day
Num
ber
of s
essi
ons
Dataset1 example user
050
100
150
0 3 6 9 13 17 21Hour of day
Mea
n se
ssio
n le
ngth
(s)
Dataset2 example user
05
1015
2025
0 3 6 9 13 17 21Hour of day
Tot
al in
tera
ctio
n (m
inut
es)
(a) Interaction time
Dataset2 example user
05
1015
0 3 6 9 13 17 21Hour of day
Num
ber
of s
essi
ons
(b) Number of sessions
Dataset2 example user
050
100
150
0 3 6 9 13 17 21Hour of day
Mea
n se
ssio
n le
ngth
(s)
(c) Session length
Figure 7 The mean and 95 CI of interaction time number of sessions and session length during each hourof the day for an example user from each dataset
Figures 9(b) and 9(c) show the strength of the diurnalpattern for the number of sessions and session length for allthe users Observe that compared to interaction time andsession length the diurnal ratio of the number of sessionstends to be lower
4 APPLICATION USAGEWe now study the applications that users run when they
interact with their smartphones Unlike previous attemptsto understand mobile application usage [5 19 26] that usediaries and interviews we rely on the mobile phonersquos OS toreport application usage We define an application as anyexecutable that the OS reports On Windows Mobile weget timestamped records of start and end times of appli-cation executions in the foreground On Android we logusage counters that are updated by the OS Every time theOS calls the onStart onRestart or onResume method of anAndroid application it starts a timer The timer stops whenthe onPause onStop or onDestroy method is called Werecord periodically the cumulative value of the timer for eachinstalled application This information on the extent of ap-plication use is not as accurate as the equivalent informationon Windows Mobile but it helps us understand relative timespent by the user in each application
41 Number of applicationsFigure 10 shows the number of applications used by users
in each dataset over the length of their trace We see thatthis number varies significantly from 10 to 90 across usersThe median is roughly 50 We are surprised by this highnumber given that the iPhone which is reported to have
020
4060
8010
0
0 20 40 60 80 100User percentile
Num
ber
of a
pplic
atio
ns
Dataset1Dataset2
Figure 10 Number of applications installed andused by users of each dataset
thousands of applications is not part of our study Ourresults show that avid use of smartphone applications is atrait shared by Android and Windows Mobile users as well
42 Application PopularityThe large number of applications installed by the users
does not mean that they use them equally We find thatusers devote the bulk of their attention to a subset of appli-cations of their choice Figure 11 illustrates this popularitybias for example users in each dataset It plots the relativepopularity of each application that is the ratio of the timespent interacting with the application and the total timespent interacting with the smartphone The bars show thepopularity PDF for the top 20 applications and the insetshows the semi-log plot for all applications Because they
15
913
1721
0 20 40 60 80 100User percentile
Diu
rnal
rat
io
Dataset1Dataset2
(a) Interaction time
12
34
5
0 20 40 60 80 100User percentile
Diu
rnal
rat
io
Dataset1Dataset2
(b) Number of sessions
15
913
1721
0 20 40 60 80 100User percentile
Diu
rnal
rat
io
Dataset1Dataset2
(c) Session length
Figure 9 Diurnal ratio of the interaction time the number of sessions and the session length for differentusers The y-axis ranges for the number of sessions is different
laun
cher gm
alar
mcl
ock
talk
cont
acts
map
sm
ms
phon
em
usic
brow
ser
cam
era
syst
emse
nsbe
nchm
ark
wer
tago
vend
ing
cale
ndar
setti
ngs
tippi
ntim
ete
tris
andr
oid
Dataset1 example user
Usa
ge r
atio
000
010
020
030
0 20 401eminus
070
01
iexp
lore
gwes
tmai
lho
me
bubb
lebr
eake
rP
oker
Live
Sea
rch
cale
ndar
Goo
gleM
aps
pout
look
TP
CS
olita
reK
aGlo
mpv
bloa
dcp
rog
serv
ices
tels
hell
pim
gap
pman
YG
oNet
wm
play
er
Dataset2 example user
Usa
ge r
atio
000
005
010
015
0 20 40 601eminus
060
01
Figure 11 Relative time spent running each application for example users in each dataset Inset is thesemi-log plot of application popularity
are binary names even some popular applications may ap-pear unfamiliar For instance in Dataset1 ldquolauncherrdquo is thedefault home screen application on Android in Dataset2ldquogwesrdquo (Graphical Windows and Events Subsystem) is thegraphical shell on Window Mobile
The graphs show that relative application popularity dropsquickly for both users In sect83 we show that for all usersapplication popularity can be modeled by an exponentialdistribution
Diurnal patterns Interestingly application popularityis not stable throughout the day but has a diurnal patternlike the other aspects of smartphone use That is the rel-ative popularity of an application is different for differenttimes of the day Figure 12 illustrates this for an exam-ple user in Dataset2 We see for instance that tmailexewhich is a messaging application on Windows Mobile ismore popular during the day than night Time dependentapplication popularity was recently reported by Trestian etal based on an analysis of the network traffic logs from a3G provider [27] Our analysis based on direct observationof user behavior confirms this effect
0 4 8 12 17 22Hour of day
020
4060
80U
sage
rat
io (
)
servicesexewmplayerexecalendarexepoutlookexeWLMMessengercdialexepimgexecprogexeiexploreexehomeexetmailexe
Figure 12 Relative time spent with each applicationduring each hour of the day for a sample user andher top applications
comm 44
browsing 10
other 11
prod 19media 5
maps 5games 2system 5
Dataset1
comm 49
browsing 12 other 15
prod 2media 9
maps 2
games 10
system 1
Dataset2
Figure 13 Relative popularity of each applicationcategory across all users in each dataset
Aggregate view application popularity To pro-vide an aggregate view of what users use their smartphonesfor we categorize applications into eight distinct categoriesi) communication contains applications for exchanging mes-sages (eg email SMS IM) and voice calls ii) browsingcontains Web browser search and social networking appli-cations iii) media contains applications for consuming orcreating media content (eg pictures music videos) iv)productivity contains applications for calendars alarms andfor viewing and creating text documents (eg Office PDFreader) v) system contains applications for changing userpreferences and viewing system state (eg file explorer)vi) games vii) maps and viii) other contains applicationsthat we could not include in any of the categories aboveeg because we did not know their function
Figure 13 shows the mean relative popularity of each ap-plication category across all users in each dataset Whilethe results are not identical across the two datasets they aresimilar to a first order Communication dominates in bothBrowsing is another major contributor in both datasetsMaps media and games have a comparatively lower butnevertheless substantial share of user attention
Relationship to user demographic To understandthe extent to which user demographic determines applica-tion popularity Figure 14 shows the mean and 95 CI ofrelative popularity of each application category for differ-ent user demographics As for interaction time (sect31) wesee that the impact of user demographics in our datasets isminimal In Dataset2 the relative popularity of various ap-plication types is similar for each of the four demographicsIn Dataset1 there are noticeable differences in the mean forcommunication games and productivity applications Highschool students use communication and games applicationsmore while knowledge workers use productivity applications
Dataset1
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
High schoolKnowledge worker
Dataset2
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
SCOPBPULPU
Figure 14 The mean and 95 CI of relative pop-ularity of application categories among users of dif-ferent demographics
more However considering the overlap of confidence in-tervals these differences in application popularity are notstatistically significant
From this result and the earlier one on the lack of depen-dence between user demographic and interaction time (sect31)we conclude that user demographic at least as defined inour datasets cannot reliably predict how a user will use thephone While demographic information appears to help insome cases (for eg the variation in usage of productivitysoftware in Dataset1) such cases are not the norm and itis hard to guess when demographic information would beuseful Pending development of other ways to classify userssuch that these classes more predictably explain the varia-tion incorporating factors specific to a user appear necessaryThis insensitivity to user demographic has positive as wellas negative implications A negative is that personalizationis more complex we cannot predict a usersrsquo behavior byknowing their demographic A positive implication is thatthe range of user behaviors along many dimensions of in-terest can be found in several common demographics Thissimplifies the task of uncovering the range because recruitingsubjects across multiple demographics tends to be difficult
Relationship to interaction time We also study ifusers that interact more with their phones tend to use differ-ent applications For each dataset we sort users based ontheir average interaction time per day and partition theminto different classes For Dataset1 we use two classesone each for the top and the bottom half of the users ForDataset2 which has more users we use three classes for the
Dataset1
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
HighLow
Dataset2
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
rHighMediumLow
Figure 15 The mean and 95 CI of relative pop-ularity of application categories among differentclasses of users based on interaction time per day
top middle and bottom third of the users Figure 15 showsthe mean and 95 CI for relative time spent with each ap-plication category by each user class We see that users indifferent classes have similar application usage Thus wecannot explain why some users use the phone more simplybased on the applications that they use
43 Application SessionsWe now study the characteristics of application sessions
We conduct this analysis only for Dataset2 based on times-tamps for when an application is started and ended Dataset1does not contain this information precisely Because appli-cations can run in the background start and end refer tothe period when the application is in the foreground
Applications run per interaction We begin by study-ing the number of applications that users run in an interac-tion session Figure 16 shows that an overwhelming major-ity close to 90 of interactions include only one applicationThis graph aggregates data across all users We did not findstatistically significant differences between users A largefraction of sessions of all users have only one application
That interaction sessions very often have only one appli-cation session suggests that users tend to interact with theirsmartphone for one task (eg reading email checking cal-endar etc) at a time and most of these tasks require theuse of only one application
Application session lengths Because interactionsessions are dominated by those with only one application
892
7422 06 02 01
1 2 3 4 5 6
020
4060
8010
0
Apps per session
Per
cent
age
()
Figure 16 Histogram of the number of applica-tions called during each interaction session for allthe users in Dataset2
map
s
gam
es
com
m
med
ia
brow
sing
prod
syst
em
othe
r
110
1000
Ses
sion
leng
th (
s)
Figure 17 The mean and 95 CI of session lengthsof different application categories The y-axis is logscale
the overall properties of application sessions such as theirlengths are similar to those of interaction sessions (sect3)
However studying the session lengths of applications sep-arately reveals interesting insights Different applicationtypes have different session lengths as shown in Figure 17for the categories defined earlier Interactions with mapsand games tend to be the longest and those related to pro-ductivity and system tend to be the shortest
Further given an application different users run themfor different times Figure 18 shows this effect for a mes-saging application tmailexe and a browsing applicationiexploreexe For each application the mean session lengthsof users differ by more than two orders of magnitude
These observations help explain why users have differentsession lengths (sect32) They prefer different applicationsand those application sessions tend to have different lengthsFurther analysis confirms this phenomenon For instance ifwe divide users into two classes based on their mean sessionlengths the popularity of games is twice as high in the classwith high session lengths
5 TRAFFICIn this section we investigate traffic generated by smart-
phones Unlike interaction events and application use net-work traffic is not an intentional user action but a side-effectof those actions Most users are likely oblivious to how much
tmail1
1010
010
00
0 20 40 60 80 100User percentile
Ses
sion
leng
th (
s)iexplore
110
100
1000
0 20 40 60 80 100User percentile
Ses
sion
leng
th (
s)Figure 18 The mean and the upper end of the stan-dard deviation of session lengths of two applications
0 20 40 60 80 100
User percentile
0
1
10
100
1000
10000
Traffic per day (MB) Receive
Send
Figure 19 The mean and the upper end of the stan-dard deviation of the traffic sent and received perday by users in Dataset1
traffic they generate We show that the diversity and diur-nal patterns of this side-effect match those of user actionsthemselves
The analysis in this section includes only Dataset1 we donot have traffic information for Dataset2 In Dataset1 werecord all of the data sent (or received) by the phone exceptfor that exchanged over the USB link ie the traffic hereinincludes data over the 3G radio and the 80211 wireless link
51 Traffic per dayFigure 19 shows that the amount of traffic sent and re-
ceived per day differs across users by almost three orders ofmagnitude The traffic received ranges from 1 to 1000 MBand the traffic sent ranges from 03 to 100 MB The medianvalues are 30 MB sent and 5 MB received
Our results indicate that traffic generated in a day bysmartphone users is comparable to traffic generated by PCsa few years ago A study of a campus WiFi network revealedthat on average users were generating 27 MB of traffic in2001 and 71 MB in 2003 [12] A study of Japanese residen-tial broadband users in 2006 revealed that on average usersgenerate 1000 MB of traffic per day [11] This high level oftraffic has major implications for the provisioning of wirelesscarrier networks as smartphone adoption increases
Relationship to application types To investigate ifcertain types of applications are favored more by users thatgenerate more traffice we divide the users into two equalclasses based on their sum of sent and received traffic perday Figure 20 shows mean and 95 CI of relative popularityof each application category for each user class Expectedly
Dataset1
Per
cent
age
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
HighLow
Figure 20 The mean and 95 CI of relative popu-larity of each application category among high andlow traffic consumers
it shows that communication applications are more popularamong users that consume more traffic
52 ldquoInteractiverdquo trafficWe next estimate what fraction of the total traffic is ldquoin-
teractiverdquo that is has timeliness constraints Approachesto reduce power consumption of data transfers often advo-cate rescheduling network transfers for instance by delay-ing some transfers so that multiple of them may be bun-dled [1 22] Such policies are difficult to implement forinteractive traffic without hurting user response time andthus are likely to be less valuable if the bulk of the traffic isinteractive
We classify traffic as interactive if it was generated whenthe screen is on This classification method might classifysome background traffic as interactive We expect the errorto be low because the screen is on for a small fraction ofthe time for most users Because certain user interactionswith the phone begin right after traffic exchange (eg anew email is received) we also consider traffic received in asmall time window (1 minute) before the screen is turned onas having timeliness constraints The results are robust tothe exact choice of time window Indeed some of the trafficthat we classify as interactive might be delay tolerant ega new email to be sent might be deferred for a little whileHowever the volume of traffic received by the phone whichusers would rather see immediately dominates by one or-der of magnitude the volume that is sent and delay-tolerantmessaging applications such as email contribute only a smallchunk of all traffic
Figure 21 shows the fraction of interactive traffic for eachuser We see that for about 90 of the users over 50of the traffic is interactive but for the rest almost none oftheir traffic is interactive Stated differently for differentsmartphone users almost all to almost none of the traffic isgenerated by applications in the background The extremesrepresent disparate ways in which people use smartphonesand which applications on the smartphone generate mosttraffic Our results imply that the energy savings that can behad by rescheduling network activity will vary across users
53 Diurnal patternsFigure 22(a) shows the diurnal pattern for an example
user with a sharp decline at night Figure 22(b) showsthe strength of the diurnal pattern for individual users by
0 20 40 60 80 100
User percentile
00
02
04
06
08
10
Interactive fraction
Figure 21 The fraction of interactive traffic
0 5 10 15 20
Hour of day
0
5
10
15
Traffic per hour (MB)
Receive
Send
(a)
0 20 40 60 80 100
User percentile
2
4
6
8
10
12
Diurnal ratio
Receive
Send
(b)
Figure 22 (a) The mean and 95 CI for trafficgenerated per hour by an example user (b) Thediurnal ratio of traffic per hour for all users
plotting the diurnal ratio of traffic As defined in sect33 thediurnal ratio reflects how much higher the mean during thepeak hour is to the overall mean
We see that the diurnal ratio varies across users but mosthave a strong diurnal behavior 80 of them generate overtwice their average amount of traffic in their peak hour Thisbehavior is likely a direct result of the high proportion ofinteractive traffic for most users and that user interactionsthemselves have a diurnal pattern
6 ENERGY CONSUMPTIONThe final aspect of smartphone usage that we investigate
is energy consumption Energy drain depends on two fac-tors i) user interactions and applications and ii) platformhardware and software If the second factor dominates theenergy drain of various users with the identical smartphonewill be similar Otherwise the energy drain will be as di-verse as user behaviors
We estimate the amount of energy drain based on the re-maining battery indicator which varies between 0 and 100If the battery indicator has gone down by X in a time pe-riod for a battery with capacity Y mAh we compute theenergy drain in that period to be X middot Y mAh 1 Given thatbatteries are complex electro-chemical devices [14 20] thiscomputation is approximate It assumes that the batterylevel indicator is linear with respect to energy drain
1mAh is technically a unit of charge yet is commonly used toindicate energy drain because battery voltage during normaloperations is typically constant For phones in our datasetthis is 4V so multiplying a mAh reading by 4 would yieldan accurate energy reading in milli-watt-hours
020
4060
8010
0
0 4 8 12 16 20Time (hour)
Bat
tery
leve
l (
) Benchmark1Benchmark2
Figure 23 Timelapse of the remaining battery levelindicator in controlled experiments with two differ-ent workloads at room temperature
110
100
1000
0 20 40 60 80 100User percentile
Bat
tery
Dra
in (
mA
h)
Figure 24 The mean and the upper end of the stan-dard deviation of one hour energy drain for Dataset1users during discharge periods
Controlled experiments suggest that the linearity assump-tion holds to a first order We run a benchmark load thatdrains the battery at a fixed rate in room temperature Un-der this benchmark if the battery level indicator decreaseslinearly with time it must be linear with respect to energydrain Figure 23 shows that the level indicator decreasesroughly linearly for two different benchmarks Benchmark1turns the screen on and off periodically Benchmark2 com-putes and idles periodically We conclude thus that the levelindicator can be used to estimate energy drain
Figure 24 shows the mean and standard deviation of en-ergy that users drain in an hour This graph is computed us-ing only periods in which the battery is not charging becauseenergy drain in those periods are of primary interest Wesee a two orders of magnitude difference among users Whileheaviest users drain close to 250 mAh the lightest of usersdrain only 10 mAh If the battery capacity is 1200 mAhthis leads to a lifetime variation from about 4 to 120 hours
Figure 25(a) shows for an example user that the drain isnot the same throughout the day but has diurnal variationsin which more energy is consumed during the day than dur-ing the night For this user the level of energy consumedchanges by roughly a factor of five Figure 25(b) plots thediurnal ratio of energy use for all users It shows that diurnalvariations occur with different strengths for all users
2060
100
140
0 4 8 12 16 20Hour of the day
Bat
tery
Dra
in (
mA
h)
(a)
0 20 40 60 80 100
12
34
56
7
User percentile
Diu
rnal
rat
io(b)
Figure 25 (a) The mean and 95 CI of energy drainof an example Dataset1 user (b) Diurnal ratio ofall users in Dataset1
Our results show that user activities contribute heavilytowards energy drain users in Dataset1 who have identi-cal smartphones drain energy at different rates and energydrain has diurnal patterns In the future we will developmethods to accurately quantify the energy consumption ofthe platform from that due to user-induced workload
7 IMPLICATIONS OF USER DIVERSITYWe uncover a surprising level of diversity among smart-
phone users For almost every aspect of usage that we studywe find one or more orders of magnitude difference betweenusers Our findings strongly motivate the need for customiz-ing smartphones to their users We believe that this needis greater than that for customizing ordinary cellphones orlaptops Ordinary cellphones do not have as rich an appli-cation environment Laptops are not as portable and aremore resource rich For example many users plug-in theirlaptops while using them
Customization can help at all levels Consider somethingas low-level as the battery Suppose we want batteries to beboth lightweight and last for at least a day with a high proba-bility Meeting the latter goal for all users of a given platformwill require catering to the heaviest users But that will leadto unnecessarily heavy batteries for many users (Highercapacity batteries are heavier) Offering multiple types ofbatteries with different lifetime-weight tradeoffs provides away out of this bind
At levels where intelligent mechanisms to improve user ex-perience or reduce energy consumption reside user diversitymotivates adapting to the smartphone user Driving thesemechanisms based on average case behaviors may not beeffective for a large fraction of the users
The ease and utility of customization depends on twoproperties of user behavior First despite quantitative dif-ferences there must be qualitative similarities among usersFor instance we should be able to describe the behavior ofall users with the same model Different users may havedifferent parameters of this model which will then lead toquantitative differences among them The presence of qual-itative similarities imply that users are not arbitrary pointsin space and it significantly simplifies the task of learninguser behavior Second user behavior in the past must alsobe predictive of the future Otherwise customization basedon past behaviors will be of little value in the future
In the next two sections we present evidence that theseproperties hold for several key aspects of smartphone usage
Dataset1 example user
000
00
010
002
00
030
0 100 200 300 400Session length (s)
Den
sity
Dataset2 example user
000
005
010
015
020
0 20 40 60 80 100Session length (s)
Den
sity
Figure 26 The PDF of session length for sampleusers of each dataset
In sect8 we show that user sessions and relative applicationpopularity can be described using simple models In sect9 weshow that energy drain can be described as well as predictedusing a ldquotrend trablerdquo framework
8 SMARTPHONE USAGE MODELSIn this section we develop simple models that describe
three aspects of smartphone usage ndash session lengths inter-arrival time between sessions and application popularityThe models are common across users but have different pa-rameters for different users While they do not completelydescribe user behavior they capture first order factors andrepresent a first step towards more complete modeling ofsmartphone usage More importantly along with the resultsof the next section they show that qualitative similaritiesdo exist among users
81 Session LengthsWe first consider the statistical properties of session length
distributions of users We find that session length valuestend to stationary With the KPSS test for level stationar-ity [13] 90 of the users have a p-value greater than 01The presence of stationarity is appealing because it suggeststhat past behavior is capable of predicting the future
We also find that session lengths are independent that isthe current value does not have a strong correlation with thevalues seen in the recent past With the Ljung-Box test forindependence [15] 96 of the users have a p-value that isgreater than 01
Stationarity and independence considered together sug-gest that session length values can be modeled as iid sam-ples from a distribution Choosing an appropriate distribu-tion however is complicated by the nature of the sessionlengths Most sessions are very short and the frequencydrops exponentially as the length increases However in-consistent with exponential behavior there are some verylong sessions in the tail for each user
We find that a mixture of exponential and Pareto distri-butions can model both ends of the spectrum The formercaptures short sessions and the latter captures long sessionsThat is session lengths can be described by the followingmixture model
r middot Exp(λ) + (1 minus r) middot Pareto(xm α)
In this equation r is the relative mix of the two distribu-tions λ is the rate of the exponential and xm and α are thelocation and shape parameters of the Pareto distribution
The location for a Pareto distribution represents the min-imum possible value of the random variable The location
010
0020
0030
00
0 1000 2000 3000Observed quantiles
Mod
el q
uant
iles
Figure 27 QQ plot of session lengths model for asample user
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed m
ix
(a) Relative mix (r)
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed r
ate
(b) Exponential rate (λ)
020
4060
0 20 40 60 80 100User percentile
Tim
eout
(s)
(c) Pareto location (xm)
minus1
0minus
04
02
06
10
0 20 40 60 80 100User percentile
Est
imat
ed s
hape
(d) Pareto shape (α)
Figure 28 Distribution of inferred model parame-ters that describe session length values of users inboth datasets
value that offers the best fit is the screen timeout value ofthe user because the session length PDF has a spike at thisvalue The spike corresponds to short sessions that are endedby the timeout (when the user forgets to switch the screenoff) we confirm this using controlled experiments with dif-ferent timeout values Figure 26 shows this spike at 60 and15 seconds for example users from each dataset The time-out provides a natural division between the two componentdistributions We automatically infer its approximate valueusing a simple spike detection algorithm
We use the EM algorithm to infer the maximum likelihoodestimation (MLE) of the remaining three parameters [6]Figure 27 shows the quality of this fit for an example userusing the QQ plot [3] Almost all quantiles are along they = x line indicating a good fit
Figure 28 shows the four inferred parameters for varioususers While users can be modeled using the same mix-
040
0080
00
0 4000 8000Observed quantiles
Mod
el q
uant
iles
Figure 29 QQ plot of session offtime model for asample user
050
010
0015
000 20 40 60 80 100
User percentileE
stim
ated
sca
le
(a) Weibull scale
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed s
hape
(b) Weibull shape
Figure 30 Distribution of inferred model parame-ters that describe the distribution of time betweensessions for users in both datasets
ture model the parameters of this model vary widely acrossusers Because of the way we construct our model the dis-tribution of the parameter r and xm also provide insightinto how frequently usersrsquo screen is switched off by the time-out and the relative popularity of different timeout values60 seconds is the most popular value likely because it is themost common default timeout and many users never changethe default
82 Time between SessionsWe find that the Weibull distribution can explain the
screen off times This distribution has two parameters re-ferred to as its scale and shape We find the MLE for theseparameters for each user From the QQ-plot in Figure 29we notice that the model predicts a greater probability ofseeing some very large offtimes than are observed in thedatasets However the probability of seeing these large off-times is small there are 27 data points that have a y-value greater than 8000 in that graph Hence we believethat Weibull provides a good fit for the length of intervalsbetween interactions
Figure 30 shows the distribution of the estimated shapeand scale of the fitted Weibull distributions Interestinglythe shape is consistently less than one Weibull shape valuesless than one indicate that the longer the screen has been offthe less likely it is for it to be turned on by the user This be-havior has interesting implications for power saving policiesFor instance periodic activities such as checking for emailwhen the screen has been off for a while may be deferred orrescheduled if needed without hurting user experience
020
4060
8010
0
0 20 40 60 80 100User percentile
MS
E (
)
(a) MSE
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed r
ate
(b) Rate parameter
Figure 31 (a) The mean square error (MSE) whenapplication popularity distribution is modeled usingan exponential (b) The inferred rate parameter ofthe exponential distribution for different users
83 Application PopularityWe find that for each user the relative popularity of ap-
plications can be well described by a simple exponential dis-tribution This qualitative invariant is useful for instanceto predict how many applications account for a given frac-tion of user attention For the example users in Figure 11this facet can be seen in the inset plots the semi-log of thepopularity distribution is very close to a straight line
Figure 31(a) shows that this exponential drop in applica-tion popularity is true for almost all users the mean squareerror between modeled exponential and actual popularitydistribution is less than 5 for 95 of the users
Figure 31(b) shows the inferred rate parameter of the ap-plication popularity distribution for various users We seethat the rate varies by an order of magnitude from 01 toalmost 1 The value of the rate essentially captures the paceof the drop in application popularity Lower values describeusers that use more applications on a regular basis Oneimplication of the wide range is that it may be feasible toretain all popular applications in memory for some users andnot for others
9 PREDICTING ENERGY DRAINIn this section we demonstrate the value of adapting to
user behaviors in the context of a mechanism to predict fu-ture energy drain on the smartphone Such a mechanismcan help with scheduling background tasks [9] estimatingbattery lifetime and improving user experience if it is esti-mated that the user will not fully drain the battery until thenext charging cycle [2 21]
Despite its many uses to our knowledge none of thesmartphones today provide a prediction of future energydrain Providing an accurate prediction of energy drain isdifficult2 User diversity of energy use makes any static pre-diction method highly inaccurate Even for the same userthere is a high variance in energy usage Figure 32 shows thisvariance by plotting the ratio of the standard deviation tothe mean energy drain over several time periods for users inDataset 1 The standard deviation of 10-minute windows ishigher than three times the mean for a fifth of the users The
2Energy drain prediction on smartphones is more challeng-ing than what is done for laptops Laptops provide a pre-diction of battery lifetime only for active use Smartphonepredictions on the other hand must cover multiple activeidle periods to be useful
01
23
45
0 20 40 60 80 100User percentile
Err
or r
atio
2 hours1 hour10 minutes
Figure 32 Battery drain predictors have to copewith the high variance in drain
time
window to
predict
n chunks
hellip
Trend Table
hellip
hellip
Figure 33 A userrsquos trend table captures how batterydrain in n consecutive chunks relates to drain in thesubsequent window
bursty nature of user interactions is one cause of this highvariance The variance is high even at longer time scalesThe standard deviation for two hour windows is larger than150 of the mean for half the users The diurnal patternsthat we showed earlier are one cause of high variance oversuch large time scales
Although the variance is high there are patterns in auserrsquos behaviors For instance we showed earlier that aphone that has been idle for a while is likely to remain idlein the near future and it will thus continue to draw energyat a similar rate
We hypothesize that a predictor tuned to user behaviorcan be accurate We present a simple personalized energydrain predictor that incorporates various user-specific fac-tors such as length of idle periods and different types ofbusy periods Instead of explicitly identifying these factorsour predictor captures them in a ldquotrend tablerdquo framework
Our framework is shown in Figure 33 Each entry in thetrend table is indexed by an n-tuple that represents energyusage readings from adjacent chunks of size δ each Thisindex points to energy usage statistics across all time win-dows of size w in the history that follow the n chunk valuesrepresented by that index Thus the trend table captureshow the energy usage in n consecutive chunks is related tothe energy use in the subsequent time windows of size w
In this paper we use n = 3 δ = 10 minutes and maintaindifferent trend tables for each window size w for which theprediction is required To keep the trend table small wequantize the energy readings of chunks We maintain themean and standard deviation for all subsequent windowsthat map to that index
To make a prediction we use the energy readings fromthe n immediately preceding chunks Let these readings be(x1 xn) Then we search for indices in the table thatdiffer from (x1 xn) by no more than a threshold s and
GenericShortminustermTimeminusofminusdayPersonalized
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
(a) 1-hour window
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
GenericShortminustermTimeminusofminusdayPersonalized
(b) 2-hour window
Figure 34 The ratio of error to true battery drainfor different energy drain predictors
predict based on them such that similar indices have higherrelative weight More precisely the prediction is
X
e stforalli=1n|he
iminusxi|lts
ke times ue
P
ke
where e iterates over all tuples in the trend table hei is
the value of the chunk i of the tuple e ue represents thestatistics stored for that index and the relative weight ke =s minus maxi=1n |he
i minus xi| For results shown here we uses = 05
Figures 34 shows how well our Personalized predictor worksfor predicting usage for two different future time windowsThere is one point per user in each graph which correspondsto the median error seen across many predictions for thatuser We see that the 90th percentile error is 25 for 1-hourwindow and 40 for 2-hour window
To place this level of accuracy in context the figure alsoshows the accurary of three alternative predictors TheGeneric predictor uses a trend table but instead of buildinga user-specific table it builds one that combines all usersWe see that its 90th percentile error is roughly twice thatof the Personalized predictor Its predictions are worse forhalf of the users
The Short-term predictor predicts energy drain simply asthe amount drained in the immediately preceding windowof the same size For both time windows its 90th percentileerror is no better than that of the Generic predictor
The Time-of-day predictor predicts energy drain based onvalues observed for the user at the same time of the day inthe past similar to that proposed by Banerjee et al [2]Thus this predictor has some degree of personalization butit learns user behavior in less detail than the Personalizedpredictor We see that its performance for 1-hour windowis similar to the Short-term predictor For 2-hour windowits error is higher than the Personalized predictor across theboard though its worst case error is lower than the Short-term and Generic predictors
Overall these results demonstrate the value of appropri-ately learning user behaviors to design intelligent mecha-nisms on smartphones
10 RELATED WORKIn a range of domains there is a rich history of work that
characterizes user workloads However because smartphoneadoption has gathered pace relatively recently our work rep-resents one of the few to study how people use smartphones
Along with other recent works it helps complete the pic-ture Banerjee et al and Rahmati et al report on batterycharging behaviors [2 18] Like us they find considerablevariation among users Banerjee et al also propose a pre-dictor that estimates the excess energy of the battery atthe time of charging using a histogram of past battery us-age [2] Shye et al study the power consumption character-istics of 20 users [24] They infer properties such as whichcomponents in the phone consume most power and exploreoptimizations based on these characteristics Rahmati andZhong study 14 users of a particular demographic to studywhich applications are popular in that demographic wherethe phone is used and how it is shared among users [19]
In contrast to these works we focus on understandingdifferent aspects of smartphone use (eg interactions andtraffic) and on exposing the diversity of user behaviors in-stead of only the average behavior Our study also entailsan order of magnitude more users than previous efforts
Like us other researchers have developed logging utilitiesMyExperience is one such early utility [10] and the worksabove involve custom utilities as well Our focus in thispaper is not on the development of logging tools but onanalyzing their data to gain insight into user behavior
There is a body of work in modeling the aggregate be-havior of mobile users Using traces from a large cellularoperator some network related aspects of mobile usage havebeen modeled Halepovic et al and Williamson et al re-port that call arrivals are bursty and present diurnal pat-terns [28] Willkomm et al and Brown et al report thatmobile users call duration can be approximately modeled bya lognormal distribution [4 29] We use traces collected onthe mobile device itself and focus on modeling the interac-tions of individual users instead of the aggregate behavior
11 CONCLUSIONSBy studying 255 users of two different smartphone plat-
forms we comprehensively characterized user activities andtheir impact on network and battery We quantify manyhitherto unknown aspects of smartphone usage User diver-sity is an overarching theme in our findings For instancedifferent users interact with their phones 10-200 times a dayon average the mean interaction length of different users
is 10-250 seconds and users receive 1-1000 MB of data perday where 10-90 is received as part of active use
This extent of user diversity implies that mechanisms thatwork for the average case may be ineffective for a large frac-tion of the users Instead learning and adapting to userbehaviors is likely to be more effective as demonstrated byour personalized energy drain predictor We show that qual-itative similarities exist among users to facilitate the devel-opment of such mechanisms For instance the longer theuser has not interacted with the phone the less likely it isfor her to start interacting with it and application popular-ity for a user follows a simple exponential distribution
Our study points to ways in which smartphone platformsshould be enhanced Effective adaptation will require futureplatforms to support light-weight tools that monitor andlearn user behaviors in situ It will also require them toexpose appropriate knobs to control the behavior of lower-level components (eg CPU or radio)
12 ACKNOWLEDGEMENTSWe are grateful to the users in Dataset1 that ran our log-
ger and to the providers of Dataset2 Elizabeth Dawson pro-vided administrative support for data collection Jean BolotKurt Partridge Stefan Saroiu Lin Zhong and the MobiSysreviewers provided many comments that helped improve thepaper We thank them all This work is supported in partby the NSF grants CCR-0120778 and CNS-0627084
13 REFERENCES[1] Balasubramanian N Balasubramanian A
and Venkataramani A Energy consumption inmobile phones A measurement study andimplications for network applications In IMC (2009)
[2] Banerjee N Rahmati A Corner M D
Rollins S and Zhong L Users and batteriesInteractions and adaptive energy management inmobile systems In Ubicomp (2007)
[3] Becker R Chambers J and Wilks A The newS language Chapman amp HallCRC 1988
[4] Brown L Gans N Mandelbaum A Sakov
A Shen H Zeltyn S and Zhao L Statisticalanalysis of a telephone call center Journal of theAmerican Statistical Association 100 469 (2005)
[5] Church K and Smyth B Understanding mobileinformation needs In MobileHCI (2008)
[6] Dempster A Laird N and Rubin D Maximumlikelihood from incomplete data via the EM algorithmJournal of the Royal Statistical Society Series B(Methodological) 39 1 (1977)
[7] Esfahbod B Preload An adaptive prefetchingdaemon Masterrsquos thesis University of Toronto 2006
[8] Falaki H Govindan R and Estrin D Smartscreen management on mobile phones Tech Rep 74Center for Embedded Networked Sesning 2009
[9] Flinn J and Satyanarayanan M Managingbattery lifetime with energy-aware adaptation ACMTOCS 22 2 (2004)
[10] Froehlich J Chen M Consolvo S
Harrison B and Landay J MyExperience Asystem for in situ tracing and capturing of userfeedback on mobile phones In MobiSys (2007)
[11] Fukuda K Cho K and Esaki H The impact ofresidential broadband traffic on Japanese ISPbackbones SIGCOMM CCR 35 1 (2005)
[12] Kotz D and Essien K Analysis of a campus-widewireless network Wireless Networks 11 1ndash2 (2005)
[13] Kwiatkowski D Phillips P Schmidt P and
Shin Y Testing the null hypothesis of stationarityagainst the alternative of a unit root Journal ofEconometrics 54 1-3 (1992)
[14] Lahiri K Dey S Panigrahi D and
Raghunathan A Battery-driven system design Anew frontier in low power design In Asia SouthPacific design automationVLSI Design (2002)
[15] Ljung G and Box G On a measure of lack of fitin time series models Biometrika 65 2 (1978)
[16] Smartphone futures 2009-2014 httpwwwportioresearchcomSmartphone09-14html
[17] Rahmati A Qian A and Zhong L
Understanding human-battery interaction on mobilephones In MobileHCI (2007)
[18] Rahmati A and Zhong L Humanndashbatteryinteraction on mobile phones Pervasive and MobileComputing 5 5 (2009)
[19] Rahmati A and Zhong L A longitudinal study ofnon-voice mobile phone usage by teens from anunderserved urban community Tech Rep 0515-09Rice University 2009
[20] Rao R Vrudhula S and Rakhmatov D
Battery modeling for energy aware system designIEEE Computer 36 12 (2003)
[21] Ravi N Scott J Han L and Iftode L
Context-aware battery management for mobilephones In PerCom (2008)
[22] Sharma A Navda V Ramjee R
Padmanabhan V and Belding E Cool-TetherEnergy efficient on-the-fly WiFi hot-spots usingmobile phones In CoNEXT (2009)
[23] Sharma C US wireless data market update - Q32009 httpwwwchetansharmacomusmarketupdateq309htm
[24] Shye A Sholbrock B and G M Into the wildStudying real user activity patterns to guide poweroptimization for mobile architectures In Micro (2009)
[25] Snol L More smartphones than desktop PCs by2011 httpwwwpcworldcomarticle171380more_smartphones_than_desktop_pcs_by_2011
html
[26] Sohn T Li K Griswold W and Hollan J
A diary study of mobile information needs In SIGCHI(2008)
[27] Trestian I Ranjan S Kuzmanovic A and
Nucci A Measuring serendipity Connecting peoplelocations and interests in a mobile 3G network InIMC (2009)
[28] Williamson C Halepovic E Sun H and Wu
Y Characterization of CDMA2000 Cellular DataNetwork Traffic In Local Computer Networks (2005)
[29] Willkomm D Machiraju S Bolot J and
Wolisz A Primary users in cellular networks Alarge-scale measurement study In DySPAN (2008)
- Introduction
- Data Collection
-
- Dataset1
- Dataset2
- Representativeness of conclusions
-
- User Interactions
-
- Interaction Time
- Interaction Sessions
- Diurnal Patterns
-
- Application Usage
-
- Number of applications
- Application Popularity
- Application Sessions
-
- Traffic
-
- Traffic per day
- ``Interactive traffic
- Diurnal patterns
-
- Energy Consumption
- Implications of User Diversity
- Smartphone Usage Models
-
- Session Lengths
- Time between Sessions
- Application Popularity
-
- Predicting Energy Drain
- Related Work
- Conclusions
- Acknowledgements
- References
-
110
100
1000
0 20 40 60 80 100User percentile
Act
ive
time
(min
utes
)
Dataset1Dataset2
(a) All users
High schoolKnowledge worker
110
100
1000
0 20 40 60 80 100User percentile
Act
ive
time
(min
utes
)
(b) Dataset1
SCOPBPULPU1
1010
010
00
0 20 40 60 80 100User percentile
Act
ive
time
(min
utes
)
(c) Dataset2
Figure 3 The mean and the upper end of the standard deviation of interaction minutes per day (a) Allusers in each dataset (b)amp(c) Different demographics in the two datasets The y-axes are log scale
neither is effective The first is that heavier users use dif-ferent types of applications (eg games) than lighter usersBut we find that the relative popularity of application typesis similar across classes of users with different interactiontimes (sect42) The second is user demographic But asFigures 3(b) and 3(c) show the interaction times are sim-ilar across the different demographics in the two datasetsWithin each demographic user interaction times span theentire range In sect42 we show that user demographic doesnot predict application popularity either
To understand the reasons behind diversity of user inter-action times we study next how user interaction is spreadacross individual sessions This analysis will show that thereis immense diversity among users in both the number of in-teraction sessions per day and the average session length
32 Interaction SessionsInteraction sessions provide a detailed view of how a user
interacts with the phone Their characteristics are impor-tant also because energy use depends not only on how longthe phone is used in aggregate but also on the usage distribu-tion Many short interactions likely drain more energy thanfew long interactions due to the overheads of awakening thephone and radio Even with negligible overheads batterylifetime depends on how exactly energy is consumed [20]Bursty drain with high current levels during bursts can leadto a lower lifetime than a more consistent drain rate
Figure 4(a) shows the number of sessions per day for dif-ferent users We again see a wide variation Individual usersinteract with their smartphone anywhere between 10 to 200times a day on average
Figure 4(b) shows the mean and standard deviation of in-teraction session lengths Dataset1 users tend to have muchlonger sessions than Dataset2 users Given that they haveroughly similar number of interactions per day as seen inFigure 4(a) their longer sessions explain their higher inter-action time per day as seen in Figure 3(a)
Within each dataset the mean session length varies acrossusers by an order of magnitude Across both datasets therange is 10-250 seconds
Explaining the diversity in session lengths Severalhypothesis might explain the differences in different usersrsquo
Dataset1Dataset21
1010
0
0 20 40 60 80 100User percentile
Ses
sion
cou
nt
(a) Number of sessions
110
100
1000
0 20 40 60 80 100User percentile
Mea
n se
ssio
n le
ngth
(s)
Dataset1Dataset2
(b) Session length
Figure 4 The mean and the upper end of the stan-dard deviation for the number of sessions per dayand the session length The y-axes are log scale
050
100
200
0 20 60 100 140Session length (s)
Ses
sion
cou
nt
Dataset1Dataset2
(a)
050
150
250
Ses
sion
cou
nt
lt10
10minus
20
20minus
30
30minus
40
40minus
60
gt60
Session length (s)
(b)
Figure 5 (a) Scatterplot of session count per dayand mean session length of various users (b) Themean and 95 CI of session count per day for usersin Dataset2 with different mean session lengths
session lengths One hypothesis is that users with longersessions concentrate their smartphone usage in fewer ses-sions Figure 5 shows however that there is little correla-tion between usersrsquo number of sessions and session lengthFigure 5(a) shows a scatterplot of session count versus meanlength for different users There is one data point for eachuser Figure 5(b) shows the dependence of session count onsession length by aggregating data across Dataset2 users Itplots the observed mean and 95 confidence interval (CI) for
Dataset1 example userDataset2 example user
00
02
04
06
08
10
0 200 600 1000Session length (s)
CD
F
(a) Session lengths
Dataset1 example userDataset2 example user
00
02
04
06
08
10
0 1000 2000 3000Time between sessions (s)
CD
F(b) Time between sessions
Figure 6 CDFs of session length and time betweensessions for two example users The x-axis ranges inthe graphs are different
session counts per day for users with different mean sessionlengths The differences in the session counts are not sta-tistically significant In other words it is not the case thatusers who have longer sessions have fewer or more sessions
Our other hypotheses are related to application use Thesecond hypothesis is that users run varying numbers of ap-plications during an interaction and users that tend to usemore applications per session have longer sessions The thirdhypothesis is that users run different applications and someapplications such as maps have longer sessions than othersThe fourth one is that even for the same application usershave different session lengths
Our analysis of application use in sect4 reveals that the sec-ond hypothesis is not explanatory as users overwhelminglyuse only one application per session It also reveals that thethird and fourth hypotheses are likely contributors to diver-sity in session lengths Note that the inability of applicationtypes to explain interaction time per day which we men-tion in the previous section is different from their ability toexplain session lengths
Distribution of a single userrsquos sessions We find thatfor any given user most of the sessions are short but someare very long Figure 6(a) shows the CDF of session lengthsfor two example users The median session length is lessthan a minute but some are longer than an hour (not shownin the graph) A similar skewed distribution can be seen forall users in our datasets albeit with different median andmean session length values This highly skewed distributionalso explains why the standard deviations in Figure 4(b) arehigh relative to the mean In sect81 we show how sessionlengths depend on the screen timeout values
Figure 6(b) shows that the time between sessions whenthe phone is not used also has a skewed distribution Mostare short (relative to the mean) but some are very long Weshow later that these off periods have the property that thelonger a user has been in one of them the greater the chancethat the user will continue in this state
33 Diurnal PatternsWe now study diurnal patterns in interaction The pres-
ence of such patterns has several consequences For instancethe length of time a given level of remaining battery capacitylasts will depend on the time of day
Figure 7 shows for two example users that as expectedstrong diurnal patterns do exist As a function of the hour
Dataset1Dataset2
16
1116
21
0 100 200 300 400Total interaction (minutes)
Diu
rnal
rat
io
(a)
lt30
30minus
60
60minus
90
90minus
120
120minus
150
150minus
200
200minus
300
gt30
0
01
23
45
Diu
rnal
rat
io
Total interaction (minutes)
(b)
Figure 8 (a) Scatterplot of diurnal ratio of interac-tion time per hour and interaction minutes per day(b) The mean and 95 CI of diurnal ratio vs totalinteraction time per day
of the day Figure 7(a) plots the mean number of interactionminutes per hour It also plots the 95 confidence interval(CI) around the mean which can be used to judge if thedifferences in the means are statistically significant We seea clear pattern in which daytime use is much higher thannightime use though the exact pattern for different users isdifferent
Figure 7(a) also shows that usage at hours in the night islow but not completely zero We believe that this non-zerousage stems from a combination of irregular sleeping hoursand users using their devices (eg to check time) when theyget up in the middle of the night
To capture the significance of the diurnal pattern for auser we define the diurnal ratio as the ratio of the meanusage during the peak hour to the mean usage across allhours A diurnal ratio of one implies no diurnal pattern andhigher values reflect stronger patterns Figure 9(a) plots thediurnal ratio in interaction time for all users It shows thatwhile diurnal ratios vary across users roughly 70 of theusers in each dataset have a peak hour usage that is morethan twice their mean usage
Explaining the diversity in diurnal patterns Tohelp explain the variability among usersrsquo diurnal ratios inFigure 8 we study its dependence on interaction time Fig-ure 8(a) shows a scatterplot of the diurnal ratio and themean interaction time per day We see that the diurnal ra-tio tends to be inversely correlated with interaction timeFigure 8(b) shows this negative correlation more clearly byaggregating data across users It plots the mean and 95CI of the diurnal ratio of total interaction time per day forusers with different total interaction times The diurnal ra-tio decreases as interaction time increases This inverse re-lationship suggests that heavy users tend to use their phonemore consistently during the day whereas light users tendto have concentrated use during certain hours of the day
Understanding the source of diurnal patterns Thevariation in interaction time of a user across the day can re-sult from variation in the number of interaction sessions orthe length of individual sessions We find that both factorscontribute Users tend to have different number of sessionsas well as different session lengths at different hours of theday Figures 7(b) and 7(c) illustrate this point for two ex-ample users They plot the mean number of sessions andthe mean session length for each hour of the day
Dataset1 example user0
510
1520
25
0 3 6 9 13 17 21Hour of day
Tot
al in
tera
ctio
n (m
inut
es) Dataset1 example user
05
1015
0 3 6 9 13 17 21Hour of day
Num
ber
of s
essi
ons
Dataset1 example user
050
100
150
0 3 6 9 13 17 21Hour of day
Mea
n se
ssio
n le
ngth
(s)
Dataset2 example user
05
1015
2025
0 3 6 9 13 17 21Hour of day
Tot
al in
tera
ctio
n (m
inut
es)
(a) Interaction time
Dataset2 example user
05
1015
0 3 6 9 13 17 21Hour of day
Num
ber
of s
essi
ons
(b) Number of sessions
Dataset2 example user
050
100
150
0 3 6 9 13 17 21Hour of day
Mea
n se
ssio
n le
ngth
(s)
(c) Session length
Figure 7 The mean and 95 CI of interaction time number of sessions and session length during each hourof the day for an example user from each dataset
Figures 9(b) and 9(c) show the strength of the diurnalpattern for the number of sessions and session length for allthe users Observe that compared to interaction time andsession length the diurnal ratio of the number of sessionstends to be lower
4 APPLICATION USAGEWe now study the applications that users run when they
interact with their smartphones Unlike previous attemptsto understand mobile application usage [5 19 26] that usediaries and interviews we rely on the mobile phonersquos OS toreport application usage We define an application as anyexecutable that the OS reports On Windows Mobile weget timestamped records of start and end times of appli-cation executions in the foreground On Android we logusage counters that are updated by the OS Every time theOS calls the onStart onRestart or onResume method of anAndroid application it starts a timer The timer stops whenthe onPause onStop or onDestroy method is called Werecord periodically the cumulative value of the timer for eachinstalled application This information on the extent of ap-plication use is not as accurate as the equivalent informationon Windows Mobile but it helps us understand relative timespent by the user in each application
41 Number of applicationsFigure 10 shows the number of applications used by users
in each dataset over the length of their trace We see thatthis number varies significantly from 10 to 90 across usersThe median is roughly 50 We are surprised by this highnumber given that the iPhone which is reported to have
020
4060
8010
0
0 20 40 60 80 100User percentile
Num
ber
of a
pplic
atio
ns
Dataset1Dataset2
Figure 10 Number of applications installed andused by users of each dataset
thousands of applications is not part of our study Ourresults show that avid use of smartphone applications is atrait shared by Android and Windows Mobile users as well
42 Application PopularityThe large number of applications installed by the users
does not mean that they use them equally We find thatusers devote the bulk of their attention to a subset of appli-cations of their choice Figure 11 illustrates this popularitybias for example users in each dataset It plots the relativepopularity of each application that is the ratio of the timespent interacting with the application and the total timespent interacting with the smartphone The bars show thepopularity PDF for the top 20 applications and the insetshows the semi-log plot for all applications Because they
15
913
1721
0 20 40 60 80 100User percentile
Diu
rnal
rat
io
Dataset1Dataset2
(a) Interaction time
12
34
5
0 20 40 60 80 100User percentile
Diu
rnal
rat
io
Dataset1Dataset2
(b) Number of sessions
15
913
1721
0 20 40 60 80 100User percentile
Diu
rnal
rat
io
Dataset1Dataset2
(c) Session length
Figure 9 Diurnal ratio of the interaction time the number of sessions and the session length for differentusers The y-axis ranges for the number of sessions is different
laun
cher gm
alar
mcl
ock
talk
cont
acts
map
sm
ms
phon
em
usic
brow
ser
cam
era
syst
emse
nsbe
nchm
ark
wer
tago
vend
ing
cale
ndar
setti
ngs
tippi
ntim
ete
tris
andr
oid
Dataset1 example user
Usa
ge r
atio
000
010
020
030
0 20 401eminus
070
01
iexp
lore
gwes
tmai
lho
me
bubb
lebr
eake
rP
oker
Live
Sea
rch
cale
ndar
Goo
gleM
aps
pout
look
TP
CS
olita
reK
aGlo
mpv
bloa
dcp
rog
serv
ices
tels
hell
pim
gap
pman
YG
oNet
wm
play
er
Dataset2 example user
Usa
ge r
atio
000
005
010
015
0 20 40 601eminus
060
01
Figure 11 Relative time spent running each application for example users in each dataset Inset is thesemi-log plot of application popularity
are binary names even some popular applications may ap-pear unfamiliar For instance in Dataset1 ldquolauncherrdquo is thedefault home screen application on Android in Dataset2ldquogwesrdquo (Graphical Windows and Events Subsystem) is thegraphical shell on Window Mobile
The graphs show that relative application popularity dropsquickly for both users In sect83 we show that for all usersapplication popularity can be modeled by an exponentialdistribution
Diurnal patterns Interestingly application popularityis not stable throughout the day but has a diurnal patternlike the other aspects of smartphone use That is the rel-ative popularity of an application is different for differenttimes of the day Figure 12 illustrates this for an exam-ple user in Dataset2 We see for instance that tmailexewhich is a messaging application on Windows Mobile ismore popular during the day than night Time dependentapplication popularity was recently reported by Trestian etal based on an analysis of the network traffic logs from a3G provider [27] Our analysis based on direct observationof user behavior confirms this effect
0 4 8 12 17 22Hour of day
020
4060
80U
sage
rat
io (
)
servicesexewmplayerexecalendarexepoutlookexeWLMMessengercdialexepimgexecprogexeiexploreexehomeexetmailexe
Figure 12 Relative time spent with each applicationduring each hour of the day for a sample user andher top applications
comm 44
browsing 10
other 11
prod 19media 5
maps 5games 2system 5
Dataset1
comm 49
browsing 12 other 15
prod 2media 9
maps 2
games 10
system 1
Dataset2
Figure 13 Relative popularity of each applicationcategory across all users in each dataset
Aggregate view application popularity To pro-vide an aggregate view of what users use their smartphonesfor we categorize applications into eight distinct categoriesi) communication contains applications for exchanging mes-sages (eg email SMS IM) and voice calls ii) browsingcontains Web browser search and social networking appli-cations iii) media contains applications for consuming orcreating media content (eg pictures music videos) iv)productivity contains applications for calendars alarms andfor viewing and creating text documents (eg Office PDFreader) v) system contains applications for changing userpreferences and viewing system state (eg file explorer)vi) games vii) maps and viii) other contains applicationsthat we could not include in any of the categories aboveeg because we did not know their function
Figure 13 shows the mean relative popularity of each ap-plication category across all users in each dataset Whilethe results are not identical across the two datasets they aresimilar to a first order Communication dominates in bothBrowsing is another major contributor in both datasetsMaps media and games have a comparatively lower butnevertheless substantial share of user attention
Relationship to user demographic To understandthe extent to which user demographic determines applica-tion popularity Figure 14 shows the mean and 95 CI ofrelative popularity of each application category for differ-ent user demographics As for interaction time (sect31) wesee that the impact of user demographics in our datasets isminimal In Dataset2 the relative popularity of various ap-plication types is similar for each of the four demographicsIn Dataset1 there are noticeable differences in the mean forcommunication games and productivity applications Highschool students use communication and games applicationsmore while knowledge workers use productivity applications
Dataset1
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
High schoolKnowledge worker
Dataset2
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
SCOPBPULPU
Figure 14 The mean and 95 CI of relative pop-ularity of application categories among users of dif-ferent demographics
more However considering the overlap of confidence in-tervals these differences in application popularity are notstatistically significant
From this result and the earlier one on the lack of depen-dence between user demographic and interaction time (sect31)we conclude that user demographic at least as defined inour datasets cannot reliably predict how a user will use thephone While demographic information appears to help insome cases (for eg the variation in usage of productivitysoftware in Dataset1) such cases are not the norm and itis hard to guess when demographic information would beuseful Pending development of other ways to classify userssuch that these classes more predictably explain the varia-tion incorporating factors specific to a user appear necessaryThis insensitivity to user demographic has positive as wellas negative implications A negative is that personalizationis more complex we cannot predict a usersrsquo behavior byknowing their demographic A positive implication is thatthe range of user behaviors along many dimensions of in-terest can be found in several common demographics Thissimplifies the task of uncovering the range because recruitingsubjects across multiple demographics tends to be difficult
Relationship to interaction time We also study ifusers that interact more with their phones tend to use differ-ent applications For each dataset we sort users based ontheir average interaction time per day and partition theminto different classes For Dataset1 we use two classesone each for the top and the bottom half of the users ForDataset2 which has more users we use three classes for the
Dataset1
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
HighLow
Dataset2
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
rHighMediumLow
Figure 15 The mean and 95 CI of relative pop-ularity of application categories among differentclasses of users based on interaction time per day
top middle and bottom third of the users Figure 15 showsthe mean and 95 CI for relative time spent with each ap-plication category by each user class We see that users indifferent classes have similar application usage Thus wecannot explain why some users use the phone more simplybased on the applications that they use
43 Application SessionsWe now study the characteristics of application sessions
We conduct this analysis only for Dataset2 based on times-tamps for when an application is started and ended Dataset1does not contain this information precisely Because appli-cations can run in the background start and end refer tothe period when the application is in the foreground
Applications run per interaction We begin by study-ing the number of applications that users run in an interac-tion session Figure 16 shows that an overwhelming major-ity close to 90 of interactions include only one applicationThis graph aggregates data across all users We did not findstatistically significant differences between users A largefraction of sessions of all users have only one application
That interaction sessions very often have only one appli-cation session suggests that users tend to interact with theirsmartphone for one task (eg reading email checking cal-endar etc) at a time and most of these tasks require theuse of only one application
Application session lengths Because interactionsessions are dominated by those with only one application
892
7422 06 02 01
1 2 3 4 5 6
020
4060
8010
0
Apps per session
Per
cent
age
()
Figure 16 Histogram of the number of applica-tions called during each interaction session for allthe users in Dataset2
map
s
gam
es
com
m
med
ia
brow
sing
prod
syst
em
othe
r
110
1000
Ses
sion
leng
th (
s)
Figure 17 The mean and 95 CI of session lengthsof different application categories The y-axis is logscale
the overall properties of application sessions such as theirlengths are similar to those of interaction sessions (sect3)
However studying the session lengths of applications sep-arately reveals interesting insights Different applicationtypes have different session lengths as shown in Figure 17for the categories defined earlier Interactions with mapsand games tend to be the longest and those related to pro-ductivity and system tend to be the shortest
Further given an application different users run themfor different times Figure 18 shows this effect for a mes-saging application tmailexe and a browsing applicationiexploreexe For each application the mean session lengthsof users differ by more than two orders of magnitude
These observations help explain why users have differentsession lengths (sect32) They prefer different applicationsand those application sessions tend to have different lengthsFurther analysis confirms this phenomenon For instance ifwe divide users into two classes based on their mean sessionlengths the popularity of games is twice as high in the classwith high session lengths
5 TRAFFICIn this section we investigate traffic generated by smart-
phones Unlike interaction events and application use net-work traffic is not an intentional user action but a side-effectof those actions Most users are likely oblivious to how much
tmail1
1010
010
00
0 20 40 60 80 100User percentile
Ses
sion
leng
th (
s)iexplore
110
100
1000
0 20 40 60 80 100User percentile
Ses
sion
leng
th (
s)Figure 18 The mean and the upper end of the stan-dard deviation of session lengths of two applications
0 20 40 60 80 100
User percentile
0
1
10
100
1000
10000
Traffic per day (MB) Receive
Send
Figure 19 The mean and the upper end of the stan-dard deviation of the traffic sent and received perday by users in Dataset1
traffic they generate We show that the diversity and diur-nal patterns of this side-effect match those of user actionsthemselves
The analysis in this section includes only Dataset1 we donot have traffic information for Dataset2 In Dataset1 werecord all of the data sent (or received) by the phone exceptfor that exchanged over the USB link ie the traffic hereinincludes data over the 3G radio and the 80211 wireless link
51 Traffic per dayFigure 19 shows that the amount of traffic sent and re-
ceived per day differs across users by almost three orders ofmagnitude The traffic received ranges from 1 to 1000 MBand the traffic sent ranges from 03 to 100 MB The medianvalues are 30 MB sent and 5 MB received
Our results indicate that traffic generated in a day bysmartphone users is comparable to traffic generated by PCsa few years ago A study of a campus WiFi network revealedthat on average users were generating 27 MB of traffic in2001 and 71 MB in 2003 [12] A study of Japanese residen-tial broadband users in 2006 revealed that on average usersgenerate 1000 MB of traffic per day [11] This high level oftraffic has major implications for the provisioning of wirelesscarrier networks as smartphone adoption increases
Relationship to application types To investigate ifcertain types of applications are favored more by users thatgenerate more traffice we divide the users into two equalclasses based on their sum of sent and received traffic perday Figure 20 shows mean and 95 CI of relative popularityof each application category for each user class Expectedly
Dataset1
Per
cent
age
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
HighLow
Figure 20 The mean and 95 CI of relative popu-larity of each application category among high andlow traffic consumers
it shows that communication applications are more popularamong users that consume more traffic
52 ldquoInteractiverdquo trafficWe next estimate what fraction of the total traffic is ldquoin-
teractiverdquo that is has timeliness constraints Approachesto reduce power consumption of data transfers often advo-cate rescheduling network transfers for instance by delay-ing some transfers so that multiple of them may be bun-dled [1 22] Such policies are difficult to implement forinteractive traffic without hurting user response time andthus are likely to be less valuable if the bulk of the traffic isinteractive
We classify traffic as interactive if it was generated whenthe screen is on This classification method might classifysome background traffic as interactive We expect the errorto be low because the screen is on for a small fraction ofthe time for most users Because certain user interactionswith the phone begin right after traffic exchange (eg anew email is received) we also consider traffic received in asmall time window (1 minute) before the screen is turned onas having timeliness constraints The results are robust tothe exact choice of time window Indeed some of the trafficthat we classify as interactive might be delay tolerant ega new email to be sent might be deferred for a little whileHowever the volume of traffic received by the phone whichusers would rather see immediately dominates by one or-der of magnitude the volume that is sent and delay-tolerantmessaging applications such as email contribute only a smallchunk of all traffic
Figure 21 shows the fraction of interactive traffic for eachuser We see that for about 90 of the users over 50of the traffic is interactive but for the rest almost none oftheir traffic is interactive Stated differently for differentsmartphone users almost all to almost none of the traffic isgenerated by applications in the background The extremesrepresent disparate ways in which people use smartphonesand which applications on the smartphone generate mosttraffic Our results imply that the energy savings that can behad by rescheduling network activity will vary across users
53 Diurnal patternsFigure 22(a) shows the diurnal pattern for an example
user with a sharp decline at night Figure 22(b) showsthe strength of the diurnal pattern for individual users by
0 20 40 60 80 100
User percentile
00
02
04
06
08
10
Interactive fraction
Figure 21 The fraction of interactive traffic
0 5 10 15 20
Hour of day
0
5
10
15
Traffic per hour (MB)
Receive
Send
(a)
0 20 40 60 80 100
User percentile
2
4
6
8
10
12
Diurnal ratio
Receive
Send
(b)
Figure 22 (a) The mean and 95 CI for trafficgenerated per hour by an example user (b) Thediurnal ratio of traffic per hour for all users
plotting the diurnal ratio of traffic As defined in sect33 thediurnal ratio reflects how much higher the mean during thepeak hour is to the overall mean
We see that the diurnal ratio varies across users but mosthave a strong diurnal behavior 80 of them generate overtwice their average amount of traffic in their peak hour Thisbehavior is likely a direct result of the high proportion ofinteractive traffic for most users and that user interactionsthemselves have a diurnal pattern
6 ENERGY CONSUMPTIONThe final aspect of smartphone usage that we investigate
is energy consumption Energy drain depends on two fac-tors i) user interactions and applications and ii) platformhardware and software If the second factor dominates theenergy drain of various users with the identical smartphonewill be similar Otherwise the energy drain will be as di-verse as user behaviors
We estimate the amount of energy drain based on the re-maining battery indicator which varies between 0 and 100If the battery indicator has gone down by X in a time pe-riod for a battery with capacity Y mAh we compute theenergy drain in that period to be X middot Y mAh 1 Given thatbatteries are complex electro-chemical devices [14 20] thiscomputation is approximate It assumes that the batterylevel indicator is linear with respect to energy drain
1mAh is technically a unit of charge yet is commonly used toindicate energy drain because battery voltage during normaloperations is typically constant For phones in our datasetthis is 4V so multiplying a mAh reading by 4 would yieldan accurate energy reading in milli-watt-hours
020
4060
8010
0
0 4 8 12 16 20Time (hour)
Bat
tery
leve
l (
) Benchmark1Benchmark2
Figure 23 Timelapse of the remaining battery levelindicator in controlled experiments with two differ-ent workloads at room temperature
110
100
1000
0 20 40 60 80 100User percentile
Bat
tery
Dra
in (
mA
h)
Figure 24 The mean and the upper end of the stan-dard deviation of one hour energy drain for Dataset1users during discharge periods
Controlled experiments suggest that the linearity assump-tion holds to a first order We run a benchmark load thatdrains the battery at a fixed rate in room temperature Un-der this benchmark if the battery level indicator decreaseslinearly with time it must be linear with respect to energydrain Figure 23 shows that the level indicator decreasesroughly linearly for two different benchmarks Benchmark1turns the screen on and off periodically Benchmark2 com-putes and idles periodically We conclude thus that the levelindicator can be used to estimate energy drain
Figure 24 shows the mean and standard deviation of en-ergy that users drain in an hour This graph is computed us-ing only periods in which the battery is not charging becauseenergy drain in those periods are of primary interest Wesee a two orders of magnitude difference among users Whileheaviest users drain close to 250 mAh the lightest of usersdrain only 10 mAh If the battery capacity is 1200 mAhthis leads to a lifetime variation from about 4 to 120 hours
Figure 25(a) shows for an example user that the drain isnot the same throughout the day but has diurnal variationsin which more energy is consumed during the day than dur-ing the night For this user the level of energy consumedchanges by roughly a factor of five Figure 25(b) plots thediurnal ratio of energy use for all users It shows that diurnalvariations occur with different strengths for all users
2060
100
140
0 4 8 12 16 20Hour of the day
Bat
tery
Dra
in (
mA
h)
(a)
0 20 40 60 80 100
12
34
56
7
User percentile
Diu
rnal
rat
io(b)
Figure 25 (a) The mean and 95 CI of energy drainof an example Dataset1 user (b) Diurnal ratio ofall users in Dataset1
Our results show that user activities contribute heavilytowards energy drain users in Dataset1 who have identi-cal smartphones drain energy at different rates and energydrain has diurnal patterns In the future we will developmethods to accurately quantify the energy consumption ofthe platform from that due to user-induced workload
7 IMPLICATIONS OF USER DIVERSITYWe uncover a surprising level of diversity among smart-
phone users For almost every aspect of usage that we studywe find one or more orders of magnitude difference betweenusers Our findings strongly motivate the need for customiz-ing smartphones to their users We believe that this needis greater than that for customizing ordinary cellphones orlaptops Ordinary cellphones do not have as rich an appli-cation environment Laptops are not as portable and aremore resource rich For example many users plug-in theirlaptops while using them
Customization can help at all levels Consider somethingas low-level as the battery Suppose we want batteries to beboth lightweight and last for at least a day with a high proba-bility Meeting the latter goal for all users of a given platformwill require catering to the heaviest users But that will leadto unnecessarily heavy batteries for many users (Highercapacity batteries are heavier) Offering multiple types ofbatteries with different lifetime-weight tradeoffs provides away out of this bind
At levels where intelligent mechanisms to improve user ex-perience or reduce energy consumption reside user diversitymotivates adapting to the smartphone user Driving thesemechanisms based on average case behaviors may not beeffective for a large fraction of the users
The ease and utility of customization depends on twoproperties of user behavior First despite quantitative dif-ferences there must be qualitative similarities among usersFor instance we should be able to describe the behavior ofall users with the same model Different users may havedifferent parameters of this model which will then lead toquantitative differences among them The presence of qual-itative similarities imply that users are not arbitrary pointsin space and it significantly simplifies the task of learninguser behavior Second user behavior in the past must alsobe predictive of the future Otherwise customization basedon past behaviors will be of little value in the future
In the next two sections we present evidence that theseproperties hold for several key aspects of smartphone usage
Dataset1 example user
000
00
010
002
00
030
0 100 200 300 400Session length (s)
Den
sity
Dataset2 example user
000
005
010
015
020
0 20 40 60 80 100Session length (s)
Den
sity
Figure 26 The PDF of session length for sampleusers of each dataset
In sect8 we show that user sessions and relative applicationpopularity can be described using simple models In sect9 weshow that energy drain can be described as well as predictedusing a ldquotrend trablerdquo framework
8 SMARTPHONE USAGE MODELSIn this section we develop simple models that describe
three aspects of smartphone usage ndash session lengths inter-arrival time between sessions and application popularityThe models are common across users but have different pa-rameters for different users While they do not completelydescribe user behavior they capture first order factors andrepresent a first step towards more complete modeling ofsmartphone usage More importantly along with the resultsof the next section they show that qualitative similaritiesdo exist among users
81 Session LengthsWe first consider the statistical properties of session length
distributions of users We find that session length valuestend to stationary With the KPSS test for level stationar-ity [13] 90 of the users have a p-value greater than 01The presence of stationarity is appealing because it suggeststhat past behavior is capable of predicting the future
We also find that session lengths are independent that isthe current value does not have a strong correlation with thevalues seen in the recent past With the Ljung-Box test forindependence [15] 96 of the users have a p-value that isgreater than 01
Stationarity and independence considered together sug-gest that session length values can be modeled as iid sam-ples from a distribution Choosing an appropriate distribu-tion however is complicated by the nature of the sessionlengths Most sessions are very short and the frequencydrops exponentially as the length increases However in-consistent with exponential behavior there are some verylong sessions in the tail for each user
We find that a mixture of exponential and Pareto distri-butions can model both ends of the spectrum The formercaptures short sessions and the latter captures long sessionsThat is session lengths can be described by the followingmixture model
r middot Exp(λ) + (1 minus r) middot Pareto(xm α)
In this equation r is the relative mix of the two distribu-tions λ is the rate of the exponential and xm and α are thelocation and shape parameters of the Pareto distribution
The location for a Pareto distribution represents the min-imum possible value of the random variable The location
010
0020
0030
00
0 1000 2000 3000Observed quantiles
Mod
el q
uant
iles
Figure 27 QQ plot of session lengths model for asample user
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed m
ix
(a) Relative mix (r)
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed r
ate
(b) Exponential rate (λ)
020
4060
0 20 40 60 80 100User percentile
Tim
eout
(s)
(c) Pareto location (xm)
minus1
0minus
04
02
06
10
0 20 40 60 80 100User percentile
Est
imat
ed s
hape
(d) Pareto shape (α)
Figure 28 Distribution of inferred model parame-ters that describe session length values of users inboth datasets
value that offers the best fit is the screen timeout value ofthe user because the session length PDF has a spike at thisvalue The spike corresponds to short sessions that are endedby the timeout (when the user forgets to switch the screenoff) we confirm this using controlled experiments with dif-ferent timeout values Figure 26 shows this spike at 60 and15 seconds for example users from each dataset The time-out provides a natural division between the two componentdistributions We automatically infer its approximate valueusing a simple spike detection algorithm
We use the EM algorithm to infer the maximum likelihoodestimation (MLE) of the remaining three parameters [6]Figure 27 shows the quality of this fit for an example userusing the QQ plot [3] Almost all quantiles are along they = x line indicating a good fit
Figure 28 shows the four inferred parameters for varioususers While users can be modeled using the same mix-
040
0080
00
0 4000 8000Observed quantiles
Mod
el q
uant
iles
Figure 29 QQ plot of session offtime model for asample user
050
010
0015
000 20 40 60 80 100
User percentileE
stim
ated
sca
le
(a) Weibull scale
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed s
hape
(b) Weibull shape
Figure 30 Distribution of inferred model parame-ters that describe the distribution of time betweensessions for users in both datasets
ture model the parameters of this model vary widely acrossusers Because of the way we construct our model the dis-tribution of the parameter r and xm also provide insightinto how frequently usersrsquo screen is switched off by the time-out and the relative popularity of different timeout values60 seconds is the most popular value likely because it is themost common default timeout and many users never changethe default
82 Time between SessionsWe find that the Weibull distribution can explain the
screen off times This distribution has two parameters re-ferred to as its scale and shape We find the MLE for theseparameters for each user From the QQ-plot in Figure 29we notice that the model predicts a greater probability ofseeing some very large offtimes than are observed in thedatasets However the probability of seeing these large off-times is small there are 27 data points that have a y-value greater than 8000 in that graph Hence we believethat Weibull provides a good fit for the length of intervalsbetween interactions
Figure 30 shows the distribution of the estimated shapeand scale of the fitted Weibull distributions Interestinglythe shape is consistently less than one Weibull shape valuesless than one indicate that the longer the screen has been offthe less likely it is for it to be turned on by the user This be-havior has interesting implications for power saving policiesFor instance periodic activities such as checking for emailwhen the screen has been off for a while may be deferred orrescheduled if needed without hurting user experience
020
4060
8010
0
0 20 40 60 80 100User percentile
MS
E (
)
(a) MSE
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed r
ate
(b) Rate parameter
Figure 31 (a) The mean square error (MSE) whenapplication popularity distribution is modeled usingan exponential (b) The inferred rate parameter ofthe exponential distribution for different users
83 Application PopularityWe find that for each user the relative popularity of ap-
plications can be well described by a simple exponential dis-tribution This qualitative invariant is useful for instanceto predict how many applications account for a given frac-tion of user attention For the example users in Figure 11this facet can be seen in the inset plots the semi-log of thepopularity distribution is very close to a straight line
Figure 31(a) shows that this exponential drop in applica-tion popularity is true for almost all users the mean squareerror between modeled exponential and actual popularitydistribution is less than 5 for 95 of the users
Figure 31(b) shows the inferred rate parameter of the ap-plication popularity distribution for various users We seethat the rate varies by an order of magnitude from 01 toalmost 1 The value of the rate essentially captures the paceof the drop in application popularity Lower values describeusers that use more applications on a regular basis Oneimplication of the wide range is that it may be feasible toretain all popular applications in memory for some users andnot for others
9 PREDICTING ENERGY DRAINIn this section we demonstrate the value of adapting to
user behaviors in the context of a mechanism to predict fu-ture energy drain on the smartphone Such a mechanismcan help with scheduling background tasks [9] estimatingbattery lifetime and improving user experience if it is esti-mated that the user will not fully drain the battery until thenext charging cycle [2 21]
Despite its many uses to our knowledge none of thesmartphones today provide a prediction of future energydrain Providing an accurate prediction of energy drain isdifficult2 User diversity of energy use makes any static pre-diction method highly inaccurate Even for the same userthere is a high variance in energy usage Figure 32 shows thisvariance by plotting the ratio of the standard deviation tothe mean energy drain over several time periods for users inDataset 1 The standard deviation of 10-minute windows ishigher than three times the mean for a fifth of the users The
2Energy drain prediction on smartphones is more challeng-ing than what is done for laptops Laptops provide a pre-diction of battery lifetime only for active use Smartphonepredictions on the other hand must cover multiple activeidle periods to be useful
01
23
45
0 20 40 60 80 100User percentile
Err
or r
atio
2 hours1 hour10 minutes
Figure 32 Battery drain predictors have to copewith the high variance in drain
time
window to
predict
n chunks
hellip
Trend Table
hellip
hellip
Figure 33 A userrsquos trend table captures how batterydrain in n consecutive chunks relates to drain in thesubsequent window
bursty nature of user interactions is one cause of this highvariance The variance is high even at longer time scalesThe standard deviation for two hour windows is larger than150 of the mean for half the users The diurnal patternsthat we showed earlier are one cause of high variance oversuch large time scales
Although the variance is high there are patterns in auserrsquos behaviors For instance we showed earlier that aphone that has been idle for a while is likely to remain idlein the near future and it will thus continue to draw energyat a similar rate
We hypothesize that a predictor tuned to user behaviorcan be accurate We present a simple personalized energydrain predictor that incorporates various user-specific fac-tors such as length of idle periods and different types ofbusy periods Instead of explicitly identifying these factorsour predictor captures them in a ldquotrend tablerdquo framework
Our framework is shown in Figure 33 Each entry in thetrend table is indexed by an n-tuple that represents energyusage readings from adjacent chunks of size δ each Thisindex points to energy usage statistics across all time win-dows of size w in the history that follow the n chunk valuesrepresented by that index Thus the trend table captureshow the energy usage in n consecutive chunks is related tothe energy use in the subsequent time windows of size w
In this paper we use n = 3 δ = 10 minutes and maintaindifferent trend tables for each window size w for which theprediction is required To keep the trend table small wequantize the energy readings of chunks We maintain themean and standard deviation for all subsequent windowsthat map to that index
To make a prediction we use the energy readings fromthe n immediately preceding chunks Let these readings be(x1 xn) Then we search for indices in the table thatdiffer from (x1 xn) by no more than a threshold s and
GenericShortminustermTimeminusofminusdayPersonalized
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
(a) 1-hour window
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
GenericShortminustermTimeminusofminusdayPersonalized
(b) 2-hour window
Figure 34 The ratio of error to true battery drainfor different energy drain predictors
predict based on them such that similar indices have higherrelative weight More precisely the prediction is
X
e stforalli=1n|he
iminusxi|lts
ke times ue
P
ke
where e iterates over all tuples in the trend table hei is
the value of the chunk i of the tuple e ue represents thestatistics stored for that index and the relative weight ke =s minus maxi=1n |he
i minus xi| For results shown here we uses = 05
Figures 34 shows how well our Personalized predictor worksfor predicting usage for two different future time windowsThere is one point per user in each graph which correspondsto the median error seen across many predictions for thatuser We see that the 90th percentile error is 25 for 1-hourwindow and 40 for 2-hour window
To place this level of accuracy in context the figure alsoshows the accurary of three alternative predictors TheGeneric predictor uses a trend table but instead of buildinga user-specific table it builds one that combines all usersWe see that its 90th percentile error is roughly twice thatof the Personalized predictor Its predictions are worse forhalf of the users
The Short-term predictor predicts energy drain simply asthe amount drained in the immediately preceding windowof the same size For both time windows its 90th percentileerror is no better than that of the Generic predictor
The Time-of-day predictor predicts energy drain based onvalues observed for the user at the same time of the day inthe past similar to that proposed by Banerjee et al [2]Thus this predictor has some degree of personalization butit learns user behavior in less detail than the Personalizedpredictor We see that its performance for 1-hour windowis similar to the Short-term predictor For 2-hour windowits error is higher than the Personalized predictor across theboard though its worst case error is lower than the Short-term and Generic predictors
Overall these results demonstrate the value of appropri-ately learning user behaviors to design intelligent mecha-nisms on smartphones
10 RELATED WORKIn a range of domains there is a rich history of work that
characterizes user workloads However because smartphoneadoption has gathered pace relatively recently our work rep-resents one of the few to study how people use smartphones
Along with other recent works it helps complete the pic-ture Banerjee et al and Rahmati et al report on batterycharging behaviors [2 18] Like us they find considerablevariation among users Banerjee et al also propose a pre-dictor that estimates the excess energy of the battery atthe time of charging using a histogram of past battery us-age [2] Shye et al study the power consumption character-istics of 20 users [24] They infer properties such as whichcomponents in the phone consume most power and exploreoptimizations based on these characteristics Rahmati andZhong study 14 users of a particular demographic to studywhich applications are popular in that demographic wherethe phone is used and how it is shared among users [19]
In contrast to these works we focus on understandingdifferent aspects of smartphone use (eg interactions andtraffic) and on exposing the diversity of user behaviors in-stead of only the average behavior Our study also entailsan order of magnitude more users than previous efforts
Like us other researchers have developed logging utilitiesMyExperience is one such early utility [10] and the worksabove involve custom utilities as well Our focus in thispaper is not on the development of logging tools but onanalyzing their data to gain insight into user behavior
There is a body of work in modeling the aggregate be-havior of mobile users Using traces from a large cellularoperator some network related aspects of mobile usage havebeen modeled Halepovic et al and Williamson et al re-port that call arrivals are bursty and present diurnal pat-terns [28] Willkomm et al and Brown et al report thatmobile users call duration can be approximately modeled bya lognormal distribution [4 29] We use traces collected onthe mobile device itself and focus on modeling the interac-tions of individual users instead of the aggregate behavior
11 CONCLUSIONSBy studying 255 users of two different smartphone plat-
forms we comprehensively characterized user activities andtheir impact on network and battery We quantify manyhitherto unknown aspects of smartphone usage User diver-sity is an overarching theme in our findings For instancedifferent users interact with their phones 10-200 times a dayon average the mean interaction length of different users
is 10-250 seconds and users receive 1-1000 MB of data perday where 10-90 is received as part of active use
This extent of user diversity implies that mechanisms thatwork for the average case may be ineffective for a large frac-tion of the users Instead learning and adapting to userbehaviors is likely to be more effective as demonstrated byour personalized energy drain predictor We show that qual-itative similarities exist among users to facilitate the devel-opment of such mechanisms For instance the longer theuser has not interacted with the phone the less likely it isfor her to start interacting with it and application popular-ity for a user follows a simple exponential distribution
Our study points to ways in which smartphone platformsshould be enhanced Effective adaptation will require futureplatforms to support light-weight tools that monitor andlearn user behaviors in situ It will also require them toexpose appropriate knobs to control the behavior of lower-level components (eg CPU or radio)
12 ACKNOWLEDGEMENTSWe are grateful to the users in Dataset1 that ran our log-
ger and to the providers of Dataset2 Elizabeth Dawson pro-vided administrative support for data collection Jean BolotKurt Partridge Stefan Saroiu Lin Zhong and the MobiSysreviewers provided many comments that helped improve thepaper We thank them all This work is supported in partby the NSF grants CCR-0120778 and CNS-0627084
13 REFERENCES[1] Balasubramanian N Balasubramanian A
and Venkataramani A Energy consumption inmobile phones A measurement study andimplications for network applications In IMC (2009)
[2] Banerjee N Rahmati A Corner M D
Rollins S and Zhong L Users and batteriesInteractions and adaptive energy management inmobile systems In Ubicomp (2007)
[3] Becker R Chambers J and Wilks A The newS language Chapman amp HallCRC 1988
[4] Brown L Gans N Mandelbaum A Sakov
A Shen H Zeltyn S and Zhao L Statisticalanalysis of a telephone call center Journal of theAmerican Statistical Association 100 469 (2005)
[5] Church K and Smyth B Understanding mobileinformation needs In MobileHCI (2008)
[6] Dempster A Laird N and Rubin D Maximumlikelihood from incomplete data via the EM algorithmJournal of the Royal Statistical Society Series B(Methodological) 39 1 (1977)
[7] Esfahbod B Preload An adaptive prefetchingdaemon Masterrsquos thesis University of Toronto 2006
[8] Falaki H Govindan R and Estrin D Smartscreen management on mobile phones Tech Rep 74Center for Embedded Networked Sesning 2009
[9] Flinn J and Satyanarayanan M Managingbattery lifetime with energy-aware adaptation ACMTOCS 22 2 (2004)
[10] Froehlich J Chen M Consolvo S
Harrison B and Landay J MyExperience Asystem for in situ tracing and capturing of userfeedback on mobile phones In MobiSys (2007)
[11] Fukuda K Cho K and Esaki H The impact ofresidential broadband traffic on Japanese ISPbackbones SIGCOMM CCR 35 1 (2005)
[12] Kotz D and Essien K Analysis of a campus-widewireless network Wireless Networks 11 1ndash2 (2005)
[13] Kwiatkowski D Phillips P Schmidt P and
Shin Y Testing the null hypothesis of stationarityagainst the alternative of a unit root Journal ofEconometrics 54 1-3 (1992)
[14] Lahiri K Dey S Panigrahi D and
Raghunathan A Battery-driven system design Anew frontier in low power design In Asia SouthPacific design automationVLSI Design (2002)
[15] Ljung G and Box G On a measure of lack of fitin time series models Biometrika 65 2 (1978)
[16] Smartphone futures 2009-2014 httpwwwportioresearchcomSmartphone09-14html
[17] Rahmati A Qian A and Zhong L
Understanding human-battery interaction on mobilephones In MobileHCI (2007)
[18] Rahmati A and Zhong L Humanndashbatteryinteraction on mobile phones Pervasive and MobileComputing 5 5 (2009)
[19] Rahmati A and Zhong L A longitudinal study ofnon-voice mobile phone usage by teens from anunderserved urban community Tech Rep 0515-09Rice University 2009
[20] Rao R Vrudhula S and Rakhmatov D
Battery modeling for energy aware system designIEEE Computer 36 12 (2003)
[21] Ravi N Scott J Han L and Iftode L
Context-aware battery management for mobilephones In PerCom (2008)
[22] Sharma A Navda V Ramjee R
Padmanabhan V and Belding E Cool-TetherEnergy efficient on-the-fly WiFi hot-spots usingmobile phones In CoNEXT (2009)
[23] Sharma C US wireless data market update - Q32009 httpwwwchetansharmacomusmarketupdateq309htm
[24] Shye A Sholbrock B and G M Into the wildStudying real user activity patterns to guide poweroptimization for mobile architectures In Micro (2009)
[25] Snol L More smartphones than desktop PCs by2011 httpwwwpcworldcomarticle171380more_smartphones_than_desktop_pcs_by_2011
html
[26] Sohn T Li K Griswold W and Hollan J
A diary study of mobile information needs In SIGCHI(2008)
[27] Trestian I Ranjan S Kuzmanovic A and
Nucci A Measuring serendipity Connecting peoplelocations and interests in a mobile 3G network InIMC (2009)
[28] Williamson C Halepovic E Sun H and Wu
Y Characterization of CDMA2000 Cellular DataNetwork Traffic In Local Computer Networks (2005)
[29] Willkomm D Machiraju S Bolot J and
Wolisz A Primary users in cellular networks Alarge-scale measurement study In DySPAN (2008)
- Introduction
- Data Collection
-
- Dataset1
- Dataset2
- Representativeness of conclusions
-
- User Interactions
-
- Interaction Time
- Interaction Sessions
- Diurnal Patterns
-
- Application Usage
-
- Number of applications
- Application Popularity
- Application Sessions
-
- Traffic
-
- Traffic per day
- ``Interactive traffic
- Diurnal patterns
-
- Energy Consumption
- Implications of User Diversity
- Smartphone Usage Models
-
- Session Lengths
- Time between Sessions
- Application Popularity
-
- Predicting Energy Drain
- Related Work
- Conclusions
- Acknowledgements
- References
-
Dataset1 example userDataset2 example user
00
02
04
06
08
10
0 200 600 1000Session length (s)
CD
F
(a) Session lengths
Dataset1 example userDataset2 example user
00
02
04
06
08
10
0 1000 2000 3000Time between sessions (s)
CD
F(b) Time between sessions
Figure 6 CDFs of session length and time betweensessions for two example users The x-axis ranges inthe graphs are different
session counts per day for users with different mean sessionlengths The differences in the session counts are not sta-tistically significant In other words it is not the case thatusers who have longer sessions have fewer or more sessions
Our other hypotheses are related to application use Thesecond hypothesis is that users run varying numbers of ap-plications during an interaction and users that tend to usemore applications per session have longer sessions The thirdhypothesis is that users run different applications and someapplications such as maps have longer sessions than othersThe fourth one is that even for the same application usershave different session lengths
Our analysis of application use in sect4 reveals that the sec-ond hypothesis is not explanatory as users overwhelminglyuse only one application per session It also reveals that thethird and fourth hypotheses are likely contributors to diver-sity in session lengths Note that the inability of applicationtypes to explain interaction time per day which we men-tion in the previous section is different from their ability toexplain session lengths
Distribution of a single userrsquos sessions We find thatfor any given user most of the sessions are short but someare very long Figure 6(a) shows the CDF of session lengthsfor two example users The median session length is lessthan a minute but some are longer than an hour (not shownin the graph) A similar skewed distribution can be seen forall users in our datasets albeit with different median andmean session length values This highly skewed distributionalso explains why the standard deviations in Figure 4(b) arehigh relative to the mean In sect81 we show how sessionlengths depend on the screen timeout values
Figure 6(b) shows that the time between sessions whenthe phone is not used also has a skewed distribution Mostare short (relative to the mean) but some are very long Weshow later that these off periods have the property that thelonger a user has been in one of them the greater the chancethat the user will continue in this state
33 Diurnal PatternsWe now study diurnal patterns in interaction The pres-
ence of such patterns has several consequences For instancethe length of time a given level of remaining battery capacitylasts will depend on the time of day
Figure 7 shows for two example users that as expectedstrong diurnal patterns do exist As a function of the hour
Dataset1Dataset2
16
1116
21
0 100 200 300 400Total interaction (minutes)
Diu
rnal
rat
io
(a)
lt30
30minus
60
60minus
90
90minus
120
120minus
150
150minus
200
200minus
300
gt30
0
01
23
45
Diu
rnal
rat
io
Total interaction (minutes)
(b)
Figure 8 (a) Scatterplot of diurnal ratio of interac-tion time per hour and interaction minutes per day(b) The mean and 95 CI of diurnal ratio vs totalinteraction time per day
of the day Figure 7(a) plots the mean number of interactionminutes per hour It also plots the 95 confidence interval(CI) around the mean which can be used to judge if thedifferences in the means are statistically significant We seea clear pattern in which daytime use is much higher thannightime use though the exact pattern for different users isdifferent
Figure 7(a) also shows that usage at hours in the night islow but not completely zero We believe that this non-zerousage stems from a combination of irregular sleeping hoursand users using their devices (eg to check time) when theyget up in the middle of the night
To capture the significance of the diurnal pattern for auser we define the diurnal ratio as the ratio of the meanusage during the peak hour to the mean usage across allhours A diurnal ratio of one implies no diurnal pattern andhigher values reflect stronger patterns Figure 9(a) plots thediurnal ratio in interaction time for all users It shows thatwhile diurnal ratios vary across users roughly 70 of theusers in each dataset have a peak hour usage that is morethan twice their mean usage
Explaining the diversity in diurnal patterns Tohelp explain the variability among usersrsquo diurnal ratios inFigure 8 we study its dependence on interaction time Fig-ure 8(a) shows a scatterplot of the diurnal ratio and themean interaction time per day We see that the diurnal ra-tio tends to be inversely correlated with interaction timeFigure 8(b) shows this negative correlation more clearly byaggregating data across users It plots the mean and 95CI of the diurnal ratio of total interaction time per day forusers with different total interaction times The diurnal ra-tio decreases as interaction time increases This inverse re-lationship suggests that heavy users tend to use their phonemore consistently during the day whereas light users tendto have concentrated use during certain hours of the day
Understanding the source of diurnal patterns Thevariation in interaction time of a user across the day can re-sult from variation in the number of interaction sessions orthe length of individual sessions We find that both factorscontribute Users tend to have different number of sessionsas well as different session lengths at different hours of theday Figures 7(b) and 7(c) illustrate this point for two ex-ample users They plot the mean number of sessions andthe mean session length for each hour of the day
Dataset1 example user0
510
1520
25
0 3 6 9 13 17 21Hour of day
Tot
al in
tera
ctio
n (m
inut
es) Dataset1 example user
05
1015
0 3 6 9 13 17 21Hour of day
Num
ber
of s
essi
ons
Dataset1 example user
050
100
150
0 3 6 9 13 17 21Hour of day
Mea
n se
ssio
n le
ngth
(s)
Dataset2 example user
05
1015
2025
0 3 6 9 13 17 21Hour of day
Tot
al in
tera
ctio
n (m
inut
es)
(a) Interaction time
Dataset2 example user
05
1015
0 3 6 9 13 17 21Hour of day
Num
ber
of s
essi
ons
(b) Number of sessions
Dataset2 example user
050
100
150
0 3 6 9 13 17 21Hour of day
Mea
n se
ssio
n le
ngth
(s)
(c) Session length
Figure 7 The mean and 95 CI of interaction time number of sessions and session length during each hourof the day for an example user from each dataset
Figures 9(b) and 9(c) show the strength of the diurnalpattern for the number of sessions and session length for allthe users Observe that compared to interaction time andsession length the diurnal ratio of the number of sessionstends to be lower
4 APPLICATION USAGEWe now study the applications that users run when they
interact with their smartphones Unlike previous attemptsto understand mobile application usage [5 19 26] that usediaries and interviews we rely on the mobile phonersquos OS toreport application usage We define an application as anyexecutable that the OS reports On Windows Mobile weget timestamped records of start and end times of appli-cation executions in the foreground On Android we logusage counters that are updated by the OS Every time theOS calls the onStart onRestart or onResume method of anAndroid application it starts a timer The timer stops whenthe onPause onStop or onDestroy method is called Werecord periodically the cumulative value of the timer for eachinstalled application This information on the extent of ap-plication use is not as accurate as the equivalent informationon Windows Mobile but it helps us understand relative timespent by the user in each application
41 Number of applicationsFigure 10 shows the number of applications used by users
in each dataset over the length of their trace We see thatthis number varies significantly from 10 to 90 across usersThe median is roughly 50 We are surprised by this highnumber given that the iPhone which is reported to have
020
4060
8010
0
0 20 40 60 80 100User percentile
Num
ber
of a
pplic
atio
ns
Dataset1Dataset2
Figure 10 Number of applications installed andused by users of each dataset
thousands of applications is not part of our study Ourresults show that avid use of smartphone applications is atrait shared by Android and Windows Mobile users as well
42 Application PopularityThe large number of applications installed by the users
does not mean that they use them equally We find thatusers devote the bulk of their attention to a subset of appli-cations of their choice Figure 11 illustrates this popularitybias for example users in each dataset It plots the relativepopularity of each application that is the ratio of the timespent interacting with the application and the total timespent interacting with the smartphone The bars show thepopularity PDF for the top 20 applications and the insetshows the semi-log plot for all applications Because they
15
913
1721
0 20 40 60 80 100User percentile
Diu
rnal
rat
io
Dataset1Dataset2
(a) Interaction time
12
34
5
0 20 40 60 80 100User percentile
Diu
rnal
rat
io
Dataset1Dataset2
(b) Number of sessions
15
913
1721
0 20 40 60 80 100User percentile
Diu
rnal
rat
io
Dataset1Dataset2
(c) Session length
Figure 9 Diurnal ratio of the interaction time the number of sessions and the session length for differentusers The y-axis ranges for the number of sessions is different
laun
cher gm
alar
mcl
ock
talk
cont
acts
map
sm
ms
phon
em
usic
brow
ser
cam
era
syst
emse
nsbe
nchm
ark
wer
tago
vend
ing
cale
ndar
setti
ngs
tippi
ntim
ete
tris
andr
oid
Dataset1 example user
Usa
ge r
atio
000
010
020
030
0 20 401eminus
070
01
iexp
lore
gwes
tmai
lho
me
bubb
lebr
eake
rP
oker
Live
Sea
rch
cale
ndar
Goo
gleM
aps
pout
look
TP
CS
olita
reK
aGlo
mpv
bloa
dcp
rog
serv
ices
tels
hell
pim
gap
pman
YG
oNet
wm
play
er
Dataset2 example user
Usa
ge r
atio
000
005
010
015
0 20 40 601eminus
060
01
Figure 11 Relative time spent running each application for example users in each dataset Inset is thesemi-log plot of application popularity
are binary names even some popular applications may ap-pear unfamiliar For instance in Dataset1 ldquolauncherrdquo is thedefault home screen application on Android in Dataset2ldquogwesrdquo (Graphical Windows and Events Subsystem) is thegraphical shell on Window Mobile
The graphs show that relative application popularity dropsquickly for both users In sect83 we show that for all usersapplication popularity can be modeled by an exponentialdistribution
Diurnal patterns Interestingly application popularityis not stable throughout the day but has a diurnal patternlike the other aspects of smartphone use That is the rel-ative popularity of an application is different for differenttimes of the day Figure 12 illustrates this for an exam-ple user in Dataset2 We see for instance that tmailexewhich is a messaging application on Windows Mobile ismore popular during the day than night Time dependentapplication popularity was recently reported by Trestian etal based on an analysis of the network traffic logs from a3G provider [27] Our analysis based on direct observationof user behavior confirms this effect
0 4 8 12 17 22Hour of day
020
4060
80U
sage
rat
io (
)
servicesexewmplayerexecalendarexepoutlookexeWLMMessengercdialexepimgexecprogexeiexploreexehomeexetmailexe
Figure 12 Relative time spent with each applicationduring each hour of the day for a sample user andher top applications
comm 44
browsing 10
other 11
prod 19media 5
maps 5games 2system 5
Dataset1
comm 49
browsing 12 other 15
prod 2media 9
maps 2
games 10
system 1
Dataset2
Figure 13 Relative popularity of each applicationcategory across all users in each dataset
Aggregate view application popularity To pro-vide an aggregate view of what users use their smartphonesfor we categorize applications into eight distinct categoriesi) communication contains applications for exchanging mes-sages (eg email SMS IM) and voice calls ii) browsingcontains Web browser search and social networking appli-cations iii) media contains applications for consuming orcreating media content (eg pictures music videos) iv)productivity contains applications for calendars alarms andfor viewing and creating text documents (eg Office PDFreader) v) system contains applications for changing userpreferences and viewing system state (eg file explorer)vi) games vii) maps and viii) other contains applicationsthat we could not include in any of the categories aboveeg because we did not know their function
Figure 13 shows the mean relative popularity of each ap-plication category across all users in each dataset Whilethe results are not identical across the two datasets they aresimilar to a first order Communication dominates in bothBrowsing is another major contributor in both datasetsMaps media and games have a comparatively lower butnevertheless substantial share of user attention
Relationship to user demographic To understandthe extent to which user demographic determines applica-tion popularity Figure 14 shows the mean and 95 CI ofrelative popularity of each application category for differ-ent user demographics As for interaction time (sect31) wesee that the impact of user demographics in our datasets isminimal In Dataset2 the relative popularity of various ap-plication types is similar for each of the four demographicsIn Dataset1 there are noticeable differences in the mean forcommunication games and productivity applications Highschool students use communication and games applicationsmore while knowledge workers use productivity applications
Dataset1
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
High schoolKnowledge worker
Dataset2
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
SCOPBPULPU
Figure 14 The mean and 95 CI of relative pop-ularity of application categories among users of dif-ferent demographics
more However considering the overlap of confidence in-tervals these differences in application popularity are notstatistically significant
From this result and the earlier one on the lack of depen-dence between user demographic and interaction time (sect31)we conclude that user demographic at least as defined inour datasets cannot reliably predict how a user will use thephone While demographic information appears to help insome cases (for eg the variation in usage of productivitysoftware in Dataset1) such cases are not the norm and itis hard to guess when demographic information would beuseful Pending development of other ways to classify userssuch that these classes more predictably explain the varia-tion incorporating factors specific to a user appear necessaryThis insensitivity to user demographic has positive as wellas negative implications A negative is that personalizationis more complex we cannot predict a usersrsquo behavior byknowing their demographic A positive implication is thatthe range of user behaviors along many dimensions of in-terest can be found in several common demographics Thissimplifies the task of uncovering the range because recruitingsubjects across multiple demographics tends to be difficult
Relationship to interaction time We also study ifusers that interact more with their phones tend to use differ-ent applications For each dataset we sort users based ontheir average interaction time per day and partition theminto different classes For Dataset1 we use two classesone each for the top and the bottom half of the users ForDataset2 which has more users we use three classes for the
Dataset1
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
HighLow
Dataset2
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
rHighMediumLow
Figure 15 The mean and 95 CI of relative pop-ularity of application categories among differentclasses of users based on interaction time per day
top middle and bottom third of the users Figure 15 showsthe mean and 95 CI for relative time spent with each ap-plication category by each user class We see that users indifferent classes have similar application usage Thus wecannot explain why some users use the phone more simplybased on the applications that they use
43 Application SessionsWe now study the characteristics of application sessions
We conduct this analysis only for Dataset2 based on times-tamps for when an application is started and ended Dataset1does not contain this information precisely Because appli-cations can run in the background start and end refer tothe period when the application is in the foreground
Applications run per interaction We begin by study-ing the number of applications that users run in an interac-tion session Figure 16 shows that an overwhelming major-ity close to 90 of interactions include only one applicationThis graph aggregates data across all users We did not findstatistically significant differences between users A largefraction of sessions of all users have only one application
That interaction sessions very often have only one appli-cation session suggests that users tend to interact with theirsmartphone for one task (eg reading email checking cal-endar etc) at a time and most of these tasks require theuse of only one application
Application session lengths Because interactionsessions are dominated by those with only one application
892
7422 06 02 01
1 2 3 4 5 6
020
4060
8010
0
Apps per session
Per
cent
age
()
Figure 16 Histogram of the number of applica-tions called during each interaction session for allthe users in Dataset2
map
s
gam
es
com
m
med
ia
brow
sing
prod
syst
em
othe
r
110
1000
Ses
sion
leng
th (
s)
Figure 17 The mean and 95 CI of session lengthsof different application categories The y-axis is logscale
the overall properties of application sessions such as theirlengths are similar to those of interaction sessions (sect3)
However studying the session lengths of applications sep-arately reveals interesting insights Different applicationtypes have different session lengths as shown in Figure 17for the categories defined earlier Interactions with mapsand games tend to be the longest and those related to pro-ductivity and system tend to be the shortest
Further given an application different users run themfor different times Figure 18 shows this effect for a mes-saging application tmailexe and a browsing applicationiexploreexe For each application the mean session lengthsof users differ by more than two orders of magnitude
These observations help explain why users have differentsession lengths (sect32) They prefer different applicationsand those application sessions tend to have different lengthsFurther analysis confirms this phenomenon For instance ifwe divide users into two classes based on their mean sessionlengths the popularity of games is twice as high in the classwith high session lengths
5 TRAFFICIn this section we investigate traffic generated by smart-
phones Unlike interaction events and application use net-work traffic is not an intentional user action but a side-effectof those actions Most users are likely oblivious to how much
tmail1
1010
010
00
0 20 40 60 80 100User percentile
Ses
sion
leng
th (
s)iexplore
110
100
1000
0 20 40 60 80 100User percentile
Ses
sion
leng
th (
s)Figure 18 The mean and the upper end of the stan-dard deviation of session lengths of two applications
0 20 40 60 80 100
User percentile
0
1
10
100
1000
10000
Traffic per day (MB) Receive
Send
Figure 19 The mean and the upper end of the stan-dard deviation of the traffic sent and received perday by users in Dataset1
traffic they generate We show that the diversity and diur-nal patterns of this side-effect match those of user actionsthemselves
The analysis in this section includes only Dataset1 we donot have traffic information for Dataset2 In Dataset1 werecord all of the data sent (or received) by the phone exceptfor that exchanged over the USB link ie the traffic hereinincludes data over the 3G radio and the 80211 wireless link
51 Traffic per dayFigure 19 shows that the amount of traffic sent and re-
ceived per day differs across users by almost three orders ofmagnitude The traffic received ranges from 1 to 1000 MBand the traffic sent ranges from 03 to 100 MB The medianvalues are 30 MB sent and 5 MB received
Our results indicate that traffic generated in a day bysmartphone users is comparable to traffic generated by PCsa few years ago A study of a campus WiFi network revealedthat on average users were generating 27 MB of traffic in2001 and 71 MB in 2003 [12] A study of Japanese residen-tial broadband users in 2006 revealed that on average usersgenerate 1000 MB of traffic per day [11] This high level oftraffic has major implications for the provisioning of wirelesscarrier networks as smartphone adoption increases
Relationship to application types To investigate ifcertain types of applications are favored more by users thatgenerate more traffice we divide the users into two equalclasses based on their sum of sent and received traffic perday Figure 20 shows mean and 95 CI of relative popularityof each application category for each user class Expectedly
Dataset1
Per
cent
age
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
HighLow
Figure 20 The mean and 95 CI of relative popu-larity of each application category among high andlow traffic consumers
it shows that communication applications are more popularamong users that consume more traffic
52 ldquoInteractiverdquo trafficWe next estimate what fraction of the total traffic is ldquoin-
teractiverdquo that is has timeliness constraints Approachesto reduce power consumption of data transfers often advo-cate rescheduling network transfers for instance by delay-ing some transfers so that multiple of them may be bun-dled [1 22] Such policies are difficult to implement forinteractive traffic without hurting user response time andthus are likely to be less valuable if the bulk of the traffic isinteractive
We classify traffic as interactive if it was generated whenthe screen is on This classification method might classifysome background traffic as interactive We expect the errorto be low because the screen is on for a small fraction ofthe time for most users Because certain user interactionswith the phone begin right after traffic exchange (eg anew email is received) we also consider traffic received in asmall time window (1 minute) before the screen is turned onas having timeliness constraints The results are robust tothe exact choice of time window Indeed some of the trafficthat we classify as interactive might be delay tolerant ega new email to be sent might be deferred for a little whileHowever the volume of traffic received by the phone whichusers would rather see immediately dominates by one or-der of magnitude the volume that is sent and delay-tolerantmessaging applications such as email contribute only a smallchunk of all traffic
Figure 21 shows the fraction of interactive traffic for eachuser We see that for about 90 of the users over 50of the traffic is interactive but for the rest almost none oftheir traffic is interactive Stated differently for differentsmartphone users almost all to almost none of the traffic isgenerated by applications in the background The extremesrepresent disparate ways in which people use smartphonesand which applications on the smartphone generate mosttraffic Our results imply that the energy savings that can behad by rescheduling network activity will vary across users
53 Diurnal patternsFigure 22(a) shows the diurnal pattern for an example
user with a sharp decline at night Figure 22(b) showsthe strength of the diurnal pattern for individual users by
0 20 40 60 80 100
User percentile
00
02
04
06
08
10
Interactive fraction
Figure 21 The fraction of interactive traffic
0 5 10 15 20
Hour of day
0
5
10
15
Traffic per hour (MB)
Receive
Send
(a)
0 20 40 60 80 100
User percentile
2
4
6
8
10
12
Diurnal ratio
Receive
Send
(b)
Figure 22 (a) The mean and 95 CI for trafficgenerated per hour by an example user (b) Thediurnal ratio of traffic per hour for all users
plotting the diurnal ratio of traffic As defined in sect33 thediurnal ratio reflects how much higher the mean during thepeak hour is to the overall mean
We see that the diurnal ratio varies across users but mosthave a strong diurnal behavior 80 of them generate overtwice their average amount of traffic in their peak hour Thisbehavior is likely a direct result of the high proportion ofinteractive traffic for most users and that user interactionsthemselves have a diurnal pattern
6 ENERGY CONSUMPTIONThe final aspect of smartphone usage that we investigate
is energy consumption Energy drain depends on two fac-tors i) user interactions and applications and ii) platformhardware and software If the second factor dominates theenergy drain of various users with the identical smartphonewill be similar Otherwise the energy drain will be as di-verse as user behaviors
We estimate the amount of energy drain based on the re-maining battery indicator which varies between 0 and 100If the battery indicator has gone down by X in a time pe-riod for a battery with capacity Y mAh we compute theenergy drain in that period to be X middot Y mAh 1 Given thatbatteries are complex electro-chemical devices [14 20] thiscomputation is approximate It assumes that the batterylevel indicator is linear with respect to energy drain
1mAh is technically a unit of charge yet is commonly used toindicate energy drain because battery voltage during normaloperations is typically constant For phones in our datasetthis is 4V so multiplying a mAh reading by 4 would yieldan accurate energy reading in milli-watt-hours
020
4060
8010
0
0 4 8 12 16 20Time (hour)
Bat
tery
leve
l (
) Benchmark1Benchmark2
Figure 23 Timelapse of the remaining battery levelindicator in controlled experiments with two differ-ent workloads at room temperature
110
100
1000
0 20 40 60 80 100User percentile
Bat
tery
Dra
in (
mA
h)
Figure 24 The mean and the upper end of the stan-dard deviation of one hour energy drain for Dataset1users during discharge periods
Controlled experiments suggest that the linearity assump-tion holds to a first order We run a benchmark load thatdrains the battery at a fixed rate in room temperature Un-der this benchmark if the battery level indicator decreaseslinearly with time it must be linear with respect to energydrain Figure 23 shows that the level indicator decreasesroughly linearly for two different benchmarks Benchmark1turns the screen on and off periodically Benchmark2 com-putes and idles periodically We conclude thus that the levelindicator can be used to estimate energy drain
Figure 24 shows the mean and standard deviation of en-ergy that users drain in an hour This graph is computed us-ing only periods in which the battery is not charging becauseenergy drain in those periods are of primary interest Wesee a two orders of magnitude difference among users Whileheaviest users drain close to 250 mAh the lightest of usersdrain only 10 mAh If the battery capacity is 1200 mAhthis leads to a lifetime variation from about 4 to 120 hours
Figure 25(a) shows for an example user that the drain isnot the same throughout the day but has diurnal variationsin which more energy is consumed during the day than dur-ing the night For this user the level of energy consumedchanges by roughly a factor of five Figure 25(b) plots thediurnal ratio of energy use for all users It shows that diurnalvariations occur with different strengths for all users
2060
100
140
0 4 8 12 16 20Hour of the day
Bat
tery
Dra
in (
mA
h)
(a)
0 20 40 60 80 100
12
34
56
7
User percentile
Diu
rnal
rat
io(b)
Figure 25 (a) The mean and 95 CI of energy drainof an example Dataset1 user (b) Diurnal ratio ofall users in Dataset1
Our results show that user activities contribute heavilytowards energy drain users in Dataset1 who have identi-cal smartphones drain energy at different rates and energydrain has diurnal patterns In the future we will developmethods to accurately quantify the energy consumption ofthe platform from that due to user-induced workload
7 IMPLICATIONS OF USER DIVERSITYWe uncover a surprising level of diversity among smart-
phone users For almost every aspect of usage that we studywe find one or more orders of magnitude difference betweenusers Our findings strongly motivate the need for customiz-ing smartphones to their users We believe that this needis greater than that for customizing ordinary cellphones orlaptops Ordinary cellphones do not have as rich an appli-cation environment Laptops are not as portable and aremore resource rich For example many users plug-in theirlaptops while using them
Customization can help at all levels Consider somethingas low-level as the battery Suppose we want batteries to beboth lightweight and last for at least a day with a high proba-bility Meeting the latter goal for all users of a given platformwill require catering to the heaviest users But that will leadto unnecessarily heavy batteries for many users (Highercapacity batteries are heavier) Offering multiple types ofbatteries with different lifetime-weight tradeoffs provides away out of this bind
At levels where intelligent mechanisms to improve user ex-perience or reduce energy consumption reside user diversitymotivates adapting to the smartphone user Driving thesemechanisms based on average case behaviors may not beeffective for a large fraction of the users
The ease and utility of customization depends on twoproperties of user behavior First despite quantitative dif-ferences there must be qualitative similarities among usersFor instance we should be able to describe the behavior ofall users with the same model Different users may havedifferent parameters of this model which will then lead toquantitative differences among them The presence of qual-itative similarities imply that users are not arbitrary pointsin space and it significantly simplifies the task of learninguser behavior Second user behavior in the past must alsobe predictive of the future Otherwise customization basedon past behaviors will be of little value in the future
In the next two sections we present evidence that theseproperties hold for several key aspects of smartphone usage
Dataset1 example user
000
00
010
002
00
030
0 100 200 300 400Session length (s)
Den
sity
Dataset2 example user
000
005
010
015
020
0 20 40 60 80 100Session length (s)
Den
sity
Figure 26 The PDF of session length for sampleusers of each dataset
In sect8 we show that user sessions and relative applicationpopularity can be described using simple models In sect9 weshow that energy drain can be described as well as predictedusing a ldquotrend trablerdquo framework
8 SMARTPHONE USAGE MODELSIn this section we develop simple models that describe
three aspects of smartphone usage ndash session lengths inter-arrival time between sessions and application popularityThe models are common across users but have different pa-rameters for different users While they do not completelydescribe user behavior they capture first order factors andrepresent a first step towards more complete modeling ofsmartphone usage More importantly along with the resultsof the next section they show that qualitative similaritiesdo exist among users
81 Session LengthsWe first consider the statistical properties of session length
distributions of users We find that session length valuestend to stationary With the KPSS test for level stationar-ity [13] 90 of the users have a p-value greater than 01The presence of stationarity is appealing because it suggeststhat past behavior is capable of predicting the future
We also find that session lengths are independent that isthe current value does not have a strong correlation with thevalues seen in the recent past With the Ljung-Box test forindependence [15] 96 of the users have a p-value that isgreater than 01
Stationarity and independence considered together sug-gest that session length values can be modeled as iid sam-ples from a distribution Choosing an appropriate distribu-tion however is complicated by the nature of the sessionlengths Most sessions are very short and the frequencydrops exponentially as the length increases However in-consistent with exponential behavior there are some verylong sessions in the tail for each user
We find that a mixture of exponential and Pareto distri-butions can model both ends of the spectrum The formercaptures short sessions and the latter captures long sessionsThat is session lengths can be described by the followingmixture model
r middot Exp(λ) + (1 minus r) middot Pareto(xm α)
In this equation r is the relative mix of the two distribu-tions λ is the rate of the exponential and xm and α are thelocation and shape parameters of the Pareto distribution
The location for a Pareto distribution represents the min-imum possible value of the random variable The location
010
0020
0030
00
0 1000 2000 3000Observed quantiles
Mod
el q
uant
iles
Figure 27 QQ plot of session lengths model for asample user
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed m
ix
(a) Relative mix (r)
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed r
ate
(b) Exponential rate (λ)
020
4060
0 20 40 60 80 100User percentile
Tim
eout
(s)
(c) Pareto location (xm)
minus1
0minus
04
02
06
10
0 20 40 60 80 100User percentile
Est
imat
ed s
hape
(d) Pareto shape (α)
Figure 28 Distribution of inferred model parame-ters that describe session length values of users inboth datasets
value that offers the best fit is the screen timeout value ofthe user because the session length PDF has a spike at thisvalue The spike corresponds to short sessions that are endedby the timeout (when the user forgets to switch the screenoff) we confirm this using controlled experiments with dif-ferent timeout values Figure 26 shows this spike at 60 and15 seconds for example users from each dataset The time-out provides a natural division between the two componentdistributions We automatically infer its approximate valueusing a simple spike detection algorithm
We use the EM algorithm to infer the maximum likelihoodestimation (MLE) of the remaining three parameters [6]Figure 27 shows the quality of this fit for an example userusing the QQ plot [3] Almost all quantiles are along they = x line indicating a good fit
Figure 28 shows the four inferred parameters for varioususers While users can be modeled using the same mix-
040
0080
00
0 4000 8000Observed quantiles
Mod
el q
uant
iles
Figure 29 QQ plot of session offtime model for asample user
050
010
0015
000 20 40 60 80 100
User percentileE
stim
ated
sca
le
(a) Weibull scale
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed s
hape
(b) Weibull shape
Figure 30 Distribution of inferred model parame-ters that describe the distribution of time betweensessions for users in both datasets
ture model the parameters of this model vary widely acrossusers Because of the way we construct our model the dis-tribution of the parameter r and xm also provide insightinto how frequently usersrsquo screen is switched off by the time-out and the relative popularity of different timeout values60 seconds is the most popular value likely because it is themost common default timeout and many users never changethe default
82 Time between SessionsWe find that the Weibull distribution can explain the
screen off times This distribution has two parameters re-ferred to as its scale and shape We find the MLE for theseparameters for each user From the QQ-plot in Figure 29we notice that the model predicts a greater probability ofseeing some very large offtimes than are observed in thedatasets However the probability of seeing these large off-times is small there are 27 data points that have a y-value greater than 8000 in that graph Hence we believethat Weibull provides a good fit for the length of intervalsbetween interactions
Figure 30 shows the distribution of the estimated shapeand scale of the fitted Weibull distributions Interestinglythe shape is consistently less than one Weibull shape valuesless than one indicate that the longer the screen has been offthe less likely it is for it to be turned on by the user This be-havior has interesting implications for power saving policiesFor instance periodic activities such as checking for emailwhen the screen has been off for a while may be deferred orrescheduled if needed without hurting user experience
020
4060
8010
0
0 20 40 60 80 100User percentile
MS
E (
)
(a) MSE
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed r
ate
(b) Rate parameter
Figure 31 (a) The mean square error (MSE) whenapplication popularity distribution is modeled usingan exponential (b) The inferred rate parameter ofthe exponential distribution for different users
83 Application PopularityWe find that for each user the relative popularity of ap-
plications can be well described by a simple exponential dis-tribution This qualitative invariant is useful for instanceto predict how many applications account for a given frac-tion of user attention For the example users in Figure 11this facet can be seen in the inset plots the semi-log of thepopularity distribution is very close to a straight line
Figure 31(a) shows that this exponential drop in applica-tion popularity is true for almost all users the mean squareerror between modeled exponential and actual popularitydistribution is less than 5 for 95 of the users
Figure 31(b) shows the inferred rate parameter of the ap-plication popularity distribution for various users We seethat the rate varies by an order of magnitude from 01 toalmost 1 The value of the rate essentially captures the paceof the drop in application popularity Lower values describeusers that use more applications on a regular basis Oneimplication of the wide range is that it may be feasible toretain all popular applications in memory for some users andnot for others
9 PREDICTING ENERGY DRAINIn this section we demonstrate the value of adapting to
user behaviors in the context of a mechanism to predict fu-ture energy drain on the smartphone Such a mechanismcan help with scheduling background tasks [9] estimatingbattery lifetime and improving user experience if it is esti-mated that the user will not fully drain the battery until thenext charging cycle [2 21]
Despite its many uses to our knowledge none of thesmartphones today provide a prediction of future energydrain Providing an accurate prediction of energy drain isdifficult2 User diversity of energy use makes any static pre-diction method highly inaccurate Even for the same userthere is a high variance in energy usage Figure 32 shows thisvariance by plotting the ratio of the standard deviation tothe mean energy drain over several time periods for users inDataset 1 The standard deviation of 10-minute windows ishigher than three times the mean for a fifth of the users The
2Energy drain prediction on smartphones is more challeng-ing than what is done for laptops Laptops provide a pre-diction of battery lifetime only for active use Smartphonepredictions on the other hand must cover multiple activeidle periods to be useful
01
23
45
0 20 40 60 80 100User percentile
Err
or r
atio
2 hours1 hour10 minutes
Figure 32 Battery drain predictors have to copewith the high variance in drain
time
window to
predict
n chunks
hellip
Trend Table
hellip
hellip
Figure 33 A userrsquos trend table captures how batterydrain in n consecutive chunks relates to drain in thesubsequent window
bursty nature of user interactions is one cause of this highvariance The variance is high even at longer time scalesThe standard deviation for two hour windows is larger than150 of the mean for half the users The diurnal patternsthat we showed earlier are one cause of high variance oversuch large time scales
Although the variance is high there are patterns in auserrsquos behaviors For instance we showed earlier that aphone that has been idle for a while is likely to remain idlein the near future and it will thus continue to draw energyat a similar rate
We hypothesize that a predictor tuned to user behaviorcan be accurate We present a simple personalized energydrain predictor that incorporates various user-specific fac-tors such as length of idle periods and different types ofbusy periods Instead of explicitly identifying these factorsour predictor captures them in a ldquotrend tablerdquo framework
Our framework is shown in Figure 33 Each entry in thetrend table is indexed by an n-tuple that represents energyusage readings from adjacent chunks of size δ each Thisindex points to energy usage statistics across all time win-dows of size w in the history that follow the n chunk valuesrepresented by that index Thus the trend table captureshow the energy usage in n consecutive chunks is related tothe energy use in the subsequent time windows of size w
In this paper we use n = 3 δ = 10 minutes and maintaindifferent trend tables for each window size w for which theprediction is required To keep the trend table small wequantize the energy readings of chunks We maintain themean and standard deviation for all subsequent windowsthat map to that index
To make a prediction we use the energy readings fromthe n immediately preceding chunks Let these readings be(x1 xn) Then we search for indices in the table thatdiffer from (x1 xn) by no more than a threshold s and
GenericShortminustermTimeminusofminusdayPersonalized
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
(a) 1-hour window
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
GenericShortminustermTimeminusofminusdayPersonalized
(b) 2-hour window
Figure 34 The ratio of error to true battery drainfor different energy drain predictors
predict based on them such that similar indices have higherrelative weight More precisely the prediction is
X
e stforalli=1n|he
iminusxi|lts
ke times ue
P
ke
where e iterates over all tuples in the trend table hei is
the value of the chunk i of the tuple e ue represents thestatistics stored for that index and the relative weight ke =s minus maxi=1n |he
i minus xi| For results shown here we uses = 05
Figures 34 shows how well our Personalized predictor worksfor predicting usage for two different future time windowsThere is one point per user in each graph which correspondsto the median error seen across many predictions for thatuser We see that the 90th percentile error is 25 for 1-hourwindow and 40 for 2-hour window
To place this level of accuracy in context the figure alsoshows the accurary of three alternative predictors TheGeneric predictor uses a trend table but instead of buildinga user-specific table it builds one that combines all usersWe see that its 90th percentile error is roughly twice thatof the Personalized predictor Its predictions are worse forhalf of the users
The Short-term predictor predicts energy drain simply asthe amount drained in the immediately preceding windowof the same size For both time windows its 90th percentileerror is no better than that of the Generic predictor
The Time-of-day predictor predicts energy drain based onvalues observed for the user at the same time of the day inthe past similar to that proposed by Banerjee et al [2]Thus this predictor has some degree of personalization butit learns user behavior in less detail than the Personalizedpredictor We see that its performance for 1-hour windowis similar to the Short-term predictor For 2-hour windowits error is higher than the Personalized predictor across theboard though its worst case error is lower than the Short-term and Generic predictors
Overall these results demonstrate the value of appropri-ately learning user behaviors to design intelligent mecha-nisms on smartphones
10 RELATED WORKIn a range of domains there is a rich history of work that
characterizes user workloads However because smartphoneadoption has gathered pace relatively recently our work rep-resents one of the few to study how people use smartphones
Along with other recent works it helps complete the pic-ture Banerjee et al and Rahmati et al report on batterycharging behaviors [2 18] Like us they find considerablevariation among users Banerjee et al also propose a pre-dictor that estimates the excess energy of the battery atthe time of charging using a histogram of past battery us-age [2] Shye et al study the power consumption character-istics of 20 users [24] They infer properties such as whichcomponents in the phone consume most power and exploreoptimizations based on these characteristics Rahmati andZhong study 14 users of a particular demographic to studywhich applications are popular in that demographic wherethe phone is used and how it is shared among users [19]
In contrast to these works we focus on understandingdifferent aspects of smartphone use (eg interactions andtraffic) and on exposing the diversity of user behaviors in-stead of only the average behavior Our study also entailsan order of magnitude more users than previous efforts
Like us other researchers have developed logging utilitiesMyExperience is one such early utility [10] and the worksabove involve custom utilities as well Our focus in thispaper is not on the development of logging tools but onanalyzing their data to gain insight into user behavior
There is a body of work in modeling the aggregate be-havior of mobile users Using traces from a large cellularoperator some network related aspects of mobile usage havebeen modeled Halepovic et al and Williamson et al re-port that call arrivals are bursty and present diurnal pat-terns [28] Willkomm et al and Brown et al report thatmobile users call duration can be approximately modeled bya lognormal distribution [4 29] We use traces collected onthe mobile device itself and focus on modeling the interac-tions of individual users instead of the aggregate behavior
11 CONCLUSIONSBy studying 255 users of two different smartphone plat-
forms we comprehensively characterized user activities andtheir impact on network and battery We quantify manyhitherto unknown aspects of smartphone usage User diver-sity is an overarching theme in our findings For instancedifferent users interact with their phones 10-200 times a dayon average the mean interaction length of different users
is 10-250 seconds and users receive 1-1000 MB of data perday where 10-90 is received as part of active use
This extent of user diversity implies that mechanisms thatwork for the average case may be ineffective for a large frac-tion of the users Instead learning and adapting to userbehaviors is likely to be more effective as demonstrated byour personalized energy drain predictor We show that qual-itative similarities exist among users to facilitate the devel-opment of such mechanisms For instance the longer theuser has not interacted with the phone the less likely it isfor her to start interacting with it and application popular-ity for a user follows a simple exponential distribution
Our study points to ways in which smartphone platformsshould be enhanced Effective adaptation will require futureplatforms to support light-weight tools that monitor andlearn user behaviors in situ It will also require them toexpose appropriate knobs to control the behavior of lower-level components (eg CPU or radio)
12 ACKNOWLEDGEMENTSWe are grateful to the users in Dataset1 that ran our log-
ger and to the providers of Dataset2 Elizabeth Dawson pro-vided administrative support for data collection Jean BolotKurt Partridge Stefan Saroiu Lin Zhong and the MobiSysreviewers provided many comments that helped improve thepaper We thank them all This work is supported in partby the NSF grants CCR-0120778 and CNS-0627084
13 REFERENCES[1] Balasubramanian N Balasubramanian A
and Venkataramani A Energy consumption inmobile phones A measurement study andimplications for network applications In IMC (2009)
[2] Banerjee N Rahmati A Corner M D
Rollins S and Zhong L Users and batteriesInteractions and adaptive energy management inmobile systems In Ubicomp (2007)
[3] Becker R Chambers J and Wilks A The newS language Chapman amp HallCRC 1988
[4] Brown L Gans N Mandelbaum A Sakov
A Shen H Zeltyn S and Zhao L Statisticalanalysis of a telephone call center Journal of theAmerican Statistical Association 100 469 (2005)
[5] Church K and Smyth B Understanding mobileinformation needs In MobileHCI (2008)
[6] Dempster A Laird N and Rubin D Maximumlikelihood from incomplete data via the EM algorithmJournal of the Royal Statistical Society Series B(Methodological) 39 1 (1977)
[7] Esfahbod B Preload An adaptive prefetchingdaemon Masterrsquos thesis University of Toronto 2006
[8] Falaki H Govindan R and Estrin D Smartscreen management on mobile phones Tech Rep 74Center for Embedded Networked Sesning 2009
[9] Flinn J and Satyanarayanan M Managingbattery lifetime with energy-aware adaptation ACMTOCS 22 2 (2004)
[10] Froehlich J Chen M Consolvo S
Harrison B and Landay J MyExperience Asystem for in situ tracing and capturing of userfeedback on mobile phones In MobiSys (2007)
[11] Fukuda K Cho K and Esaki H The impact ofresidential broadband traffic on Japanese ISPbackbones SIGCOMM CCR 35 1 (2005)
[12] Kotz D and Essien K Analysis of a campus-widewireless network Wireless Networks 11 1ndash2 (2005)
[13] Kwiatkowski D Phillips P Schmidt P and
Shin Y Testing the null hypothesis of stationarityagainst the alternative of a unit root Journal ofEconometrics 54 1-3 (1992)
[14] Lahiri K Dey S Panigrahi D and
Raghunathan A Battery-driven system design Anew frontier in low power design In Asia SouthPacific design automationVLSI Design (2002)
[15] Ljung G and Box G On a measure of lack of fitin time series models Biometrika 65 2 (1978)
[16] Smartphone futures 2009-2014 httpwwwportioresearchcomSmartphone09-14html
[17] Rahmati A Qian A and Zhong L
Understanding human-battery interaction on mobilephones In MobileHCI (2007)
[18] Rahmati A and Zhong L Humanndashbatteryinteraction on mobile phones Pervasive and MobileComputing 5 5 (2009)
[19] Rahmati A and Zhong L A longitudinal study ofnon-voice mobile phone usage by teens from anunderserved urban community Tech Rep 0515-09Rice University 2009
[20] Rao R Vrudhula S and Rakhmatov D
Battery modeling for energy aware system designIEEE Computer 36 12 (2003)
[21] Ravi N Scott J Han L and Iftode L
Context-aware battery management for mobilephones In PerCom (2008)
[22] Sharma A Navda V Ramjee R
Padmanabhan V and Belding E Cool-TetherEnergy efficient on-the-fly WiFi hot-spots usingmobile phones In CoNEXT (2009)
[23] Sharma C US wireless data market update - Q32009 httpwwwchetansharmacomusmarketupdateq309htm
[24] Shye A Sholbrock B and G M Into the wildStudying real user activity patterns to guide poweroptimization for mobile architectures In Micro (2009)
[25] Snol L More smartphones than desktop PCs by2011 httpwwwpcworldcomarticle171380more_smartphones_than_desktop_pcs_by_2011
html
[26] Sohn T Li K Griswold W and Hollan J
A diary study of mobile information needs In SIGCHI(2008)
[27] Trestian I Ranjan S Kuzmanovic A and
Nucci A Measuring serendipity Connecting peoplelocations and interests in a mobile 3G network InIMC (2009)
[28] Williamson C Halepovic E Sun H and Wu
Y Characterization of CDMA2000 Cellular DataNetwork Traffic In Local Computer Networks (2005)
[29] Willkomm D Machiraju S Bolot J and
Wolisz A Primary users in cellular networks Alarge-scale measurement study In DySPAN (2008)
- Introduction
- Data Collection
-
- Dataset1
- Dataset2
- Representativeness of conclusions
-
- User Interactions
-
- Interaction Time
- Interaction Sessions
- Diurnal Patterns
-
- Application Usage
-
- Number of applications
- Application Popularity
- Application Sessions
-
- Traffic
-
- Traffic per day
- ``Interactive traffic
- Diurnal patterns
-
- Energy Consumption
- Implications of User Diversity
- Smartphone Usage Models
-
- Session Lengths
- Time between Sessions
- Application Popularity
-
- Predicting Energy Drain
- Related Work
- Conclusions
- Acknowledgements
- References
-
Dataset1 example user0
510
1520
25
0 3 6 9 13 17 21Hour of day
Tot
al in
tera
ctio
n (m
inut
es) Dataset1 example user
05
1015
0 3 6 9 13 17 21Hour of day
Num
ber
of s
essi
ons
Dataset1 example user
050
100
150
0 3 6 9 13 17 21Hour of day
Mea
n se
ssio
n le
ngth
(s)
Dataset2 example user
05
1015
2025
0 3 6 9 13 17 21Hour of day
Tot
al in
tera
ctio
n (m
inut
es)
(a) Interaction time
Dataset2 example user
05
1015
0 3 6 9 13 17 21Hour of day
Num
ber
of s
essi
ons
(b) Number of sessions
Dataset2 example user
050
100
150
0 3 6 9 13 17 21Hour of day
Mea
n se
ssio
n le
ngth
(s)
(c) Session length
Figure 7 The mean and 95 CI of interaction time number of sessions and session length during each hourof the day for an example user from each dataset
Figures 9(b) and 9(c) show the strength of the diurnalpattern for the number of sessions and session length for allthe users Observe that compared to interaction time andsession length the diurnal ratio of the number of sessionstends to be lower
4 APPLICATION USAGEWe now study the applications that users run when they
interact with their smartphones Unlike previous attemptsto understand mobile application usage [5 19 26] that usediaries and interviews we rely on the mobile phonersquos OS toreport application usage We define an application as anyexecutable that the OS reports On Windows Mobile weget timestamped records of start and end times of appli-cation executions in the foreground On Android we logusage counters that are updated by the OS Every time theOS calls the onStart onRestart or onResume method of anAndroid application it starts a timer The timer stops whenthe onPause onStop or onDestroy method is called Werecord periodically the cumulative value of the timer for eachinstalled application This information on the extent of ap-plication use is not as accurate as the equivalent informationon Windows Mobile but it helps us understand relative timespent by the user in each application
41 Number of applicationsFigure 10 shows the number of applications used by users
in each dataset over the length of their trace We see thatthis number varies significantly from 10 to 90 across usersThe median is roughly 50 We are surprised by this highnumber given that the iPhone which is reported to have
020
4060
8010
0
0 20 40 60 80 100User percentile
Num
ber
of a
pplic
atio
ns
Dataset1Dataset2
Figure 10 Number of applications installed andused by users of each dataset
thousands of applications is not part of our study Ourresults show that avid use of smartphone applications is atrait shared by Android and Windows Mobile users as well
42 Application PopularityThe large number of applications installed by the users
does not mean that they use them equally We find thatusers devote the bulk of their attention to a subset of appli-cations of their choice Figure 11 illustrates this popularitybias for example users in each dataset It plots the relativepopularity of each application that is the ratio of the timespent interacting with the application and the total timespent interacting with the smartphone The bars show thepopularity PDF for the top 20 applications and the insetshows the semi-log plot for all applications Because they
15
913
1721
0 20 40 60 80 100User percentile
Diu
rnal
rat
io
Dataset1Dataset2
(a) Interaction time
12
34
5
0 20 40 60 80 100User percentile
Diu
rnal
rat
io
Dataset1Dataset2
(b) Number of sessions
15
913
1721
0 20 40 60 80 100User percentile
Diu
rnal
rat
io
Dataset1Dataset2
(c) Session length
Figure 9 Diurnal ratio of the interaction time the number of sessions and the session length for differentusers The y-axis ranges for the number of sessions is different
laun
cher gm
alar
mcl
ock
talk
cont
acts
map
sm
ms
phon
em
usic
brow
ser
cam
era
syst
emse
nsbe
nchm
ark
wer
tago
vend
ing
cale
ndar
setti
ngs
tippi
ntim
ete
tris
andr
oid
Dataset1 example user
Usa
ge r
atio
000
010
020
030
0 20 401eminus
070
01
iexp
lore
gwes
tmai
lho
me
bubb
lebr
eake
rP
oker
Live
Sea
rch
cale
ndar
Goo
gleM
aps
pout
look
TP
CS
olita
reK
aGlo
mpv
bloa
dcp
rog
serv
ices
tels
hell
pim
gap
pman
YG
oNet
wm
play
er
Dataset2 example user
Usa
ge r
atio
000
005
010
015
0 20 40 601eminus
060
01
Figure 11 Relative time spent running each application for example users in each dataset Inset is thesemi-log plot of application popularity
are binary names even some popular applications may ap-pear unfamiliar For instance in Dataset1 ldquolauncherrdquo is thedefault home screen application on Android in Dataset2ldquogwesrdquo (Graphical Windows and Events Subsystem) is thegraphical shell on Window Mobile
The graphs show that relative application popularity dropsquickly for both users In sect83 we show that for all usersapplication popularity can be modeled by an exponentialdistribution
Diurnal patterns Interestingly application popularityis not stable throughout the day but has a diurnal patternlike the other aspects of smartphone use That is the rel-ative popularity of an application is different for differenttimes of the day Figure 12 illustrates this for an exam-ple user in Dataset2 We see for instance that tmailexewhich is a messaging application on Windows Mobile ismore popular during the day than night Time dependentapplication popularity was recently reported by Trestian etal based on an analysis of the network traffic logs from a3G provider [27] Our analysis based on direct observationof user behavior confirms this effect
0 4 8 12 17 22Hour of day
020
4060
80U
sage
rat
io (
)
servicesexewmplayerexecalendarexepoutlookexeWLMMessengercdialexepimgexecprogexeiexploreexehomeexetmailexe
Figure 12 Relative time spent with each applicationduring each hour of the day for a sample user andher top applications
comm 44
browsing 10
other 11
prod 19media 5
maps 5games 2system 5
Dataset1
comm 49
browsing 12 other 15
prod 2media 9
maps 2
games 10
system 1
Dataset2
Figure 13 Relative popularity of each applicationcategory across all users in each dataset
Aggregate view application popularity To pro-vide an aggregate view of what users use their smartphonesfor we categorize applications into eight distinct categoriesi) communication contains applications for exchanging mes-sages (eg email SMS IM) and voice calls ii) browsingcontains Web browser search and social networking appli-cations iii) media contains applications for consuming orcreating media content (eg pictures music videos) iv)productivity contains applications for calendars alarms andfor viewing and creating text documents (eg Office PDFreader) v) system contains applications for changing userpreferences and viewing system state (eg file explorer)vi) games vii) maps and viii) other contains applicationsthat we could not include in any of the categories aboveeg because we did not know their function
Figure 13 shows the mean relative popularity of each ap-plication category across all users in each dataset Whilethe results are not identical across the two datasets they aresimilar to a first order Communication dominates in bothBrowsing is another major contributor in both datasetsMaps media and games have a comparatively lower butnevertheless substantial share of user attention
Relationship to user demographic To understandthe extent to which user demographic determines applica-tion popularity Figure 14 shows the mean and 95 CI ofrelative popularity of each application category for differ-ent user demographics As for interaction time (sect31) wesee that the impact of user demographics in our datasets isminimal In Dataset2 the relative popularity of various ap-plication types is similar for each of the four demographicsIn Dataset1 there are noticeable differences in the mean forcommunication games and productivity applications Highschool students use communication and games applicationsmore while knowledge workers use productivity applications
Dataset1
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
High schoolKnowledge worker
Dataset2
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
SCOPBPULPU
Figure 14 The mean and 95 CI of relative pop-ularity of application categories among users of dif-ferent demographics
more However considering the overlap of confidence in-tervals these differences in application popularity are notstatistically significant
From this result and the earlier one on the lack of depen-dence between user demographic and interaction time (sect31)we conclude that user demographic at least as defined inour datasets cannot reliably predict how a user will use thephone While demographic information appears to help insome cases (for eg the variation in usage of productivitysoftware in Dataset1) such cases are not the norm and itis hard to guess when demographic information would beuseful Pending development of other ways to classify userssuch that these classes more predictably explain the varia-tion incorporating factors specific to a user appear necessaryThis insensitivity to user demographic has positive as wellas negative implications A negative is that personalizationis more complex we cannot predict a usersrsquo behavior byknowing their demographic A positive implication is thatthe range of user behaviors along many dimensions of in-terest can be found in several common demographics Thissimplifies the task of uncovering the range because recruitingsubjects across multiple demographics tends to be difficult
Relationship to interaction time We also study ifusers that interact more with their phones tend to use differ-ent applications For each dataset we sort users based ontheir average interaction time per day and partition theminto different classes For Dataset1 we use two classesone each for the top and the bottom half of the users ForDataset2 which has more users we use three classes for the
Dataset1
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
HighLow
Dataset2
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
rHighMediumLow
Figure 15 The mean and 95 CI of relative pop-ularity of application categories among differentclasses of users based on interaction time per day
top middle and bottom third of the users Figure 15 showsthe mean and 95 CI for relative time spent with each ap-plication category by each user class We see that users indifferent classes have similar application usage Thus wecannot explain why some users use the phone more simplybased on the applications that they use
43 Application SessionsWe now study the characteristics of application sessions
We conduct this analysis only for Dataset2 based on times-tamps for when an application is started and ended Dataset1does not contain this information precisely Because appli-cations can run in the background start and end refer tothe period when the application is in the foreground
Applications run per interaction We begin by study-ing the number of applications that users run in an interac-tion session Figure 16 shows that an overwhelming major-ity close to 90 of interactions include only one applicationThis graph aggregates data across all users We did not findstatistically significant differences between users A largefraction of sessions of all users have only one application
That interaction sessions very often have only one appli-cation session suggests that users tend to interact with theirsmartphone for one task (eg reading email checking cal-endar etc) at a time and most of these tasks require theuse of only one application
Application session lengths Because interactionsessions are dominated by those with only one application
892
7422 06 02 01
1 2 3 4 5 6
020
4060
8010
0
Apps per session
Per
cent
age
()
Figure 16 Histogram of the number of applica-tions called during each interaction session for allthe users in Dataset2
map
s
gam
es
com
m
med
ia
brow
sing
prod
syst
em
othe
r
110
1000
Ses
sion
leng
th (
s)
Figure 17 The mean and 95 CI of session lengthsof different application categories The y-axis is logscale
the overall properties of application sessions such as theirlengths are similar to those of interaction sessions (sect3)
However studying the session lengths of applications sep-arately reveals interesting insights Different applicationtypes have different session lengths as shown in Figure 17for the categories defined earlier Interactions with mapsand games tend to be the longest and those related to pro-ductivity and system tend to be the shortest
Further given an application different users run themfor different times Figure 18 shows this effect for a mes-saging application tmailexe and a browsing applicationiexploreexe For each application the mean session lengthsof users differ by more than two orders of magnitude
These observations help explain why users have differentsession lengths (sect32) They prefer different applicationsand those application sessions tend to have different lengthsFurther analysis confirms this phenomenon For instance ifwe divide users into two classes based on their mean sessionlengths the popularity of games is twice as high in the classwith high session lengths
5 TRAFFICIn this section we investigate traffic generated by smart-
phones Unlike interaction events and application use net-work traffic is not an intentional user action but a side-effectof those actions Most users are likely oblivious to how much
tmail1
1010
010
00
0 20 40 60 80 100User percentile
Ses
sion
leng
th (
s)iexplore
110
100
1000
0 20 40 60 80 100User percentile
Ses
sion
leng
th (
s)Figure 18 The mean and the upper end of the stan-dard deviation of session lengths of two applications
0 20 40 60 80 100
User percentile
0
1
10
100
1000
10000
Traffic per day (MB) Receive
Send
Figure 19 The mean and the upper end of the stan-dard deviation of the traffic sent and received perday by users in Dataset1
traffic they generate We show that the diversity and diur-nal patterns of this side-effect match those of user actionsthemselves
The analysis in this section includes only Dataset1 we donot have traffic information for Dataset2 In Dataset1 werecord all of the data sent (or received) by the phone exceptfor that exchanged over the USB link ie the traffic hereinincludes data over the 3G radio and the 80211 wireless link
51 Traffic per dayFigure 19 shows that the amount of traffic sent and re-
ceived per day differs across users by almost three orders ofmagnitude The traffic received ranges from 1 to 1000 MBand the traffic sent ranges from 03 to 100 MB The medianvalues are 30 MB sent and 5 MB received
Our results indicate that traffic generated in a day bysmartphone users is comparable to traffic generated by PCsa few years ago A study of a campus WiFi network revealedthat on average users were generating 27 MB of traffic in2001 and 71 MB in 2003 [12] A study of Japanese residen-tial broadband users in 2006 revealed that on average usersgenerate 1000 MB of traffic per day [11] This high level oftraffic has major implications for the provisioning of wirelesscarrier networks as smartphone adoption increases
Relationship to application types To investigate ifcertain types of applications are favored more by users thatgenerate more traffice we divide the users into two equalclasses based on their sum of sent and received traffic perday Figure 20 shows mean and 95 CI of relative popularityof each application category for each user class Expectedly
Dataset1
Per
cent
age
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
HighLow
Figure 20 The mean and 95 CI of relative popu-larity of each application category among high andlow traffic consumers
it shows that communication applications are more popularamong users that consume more traffic
52 ldquoInteractiverdquo trafficWe next estimate what fraction of the total traffic is ldquoin-
teractiverdquo that is has timeliness constraints Approachesto reduce power consumption of data transfers often advo-cate rescheduling network transfers for instance by delay-ing some transfers so that multiple of them may be bun-dled [1 22] Such policies are difficult to implement forinteractive traffic without hurting user response time andthus are likely to be less valuable if the bulk of the traffic isinteractive
We classify traffic as interactive if it was generated whenthe screen is on This classification method might classifysome background traffic as interactive We expect the errorto be low because the screen is on for a small fraction ofthe time for most users Because certain user interactionswith the phone begin right after traffic exchange (eg anew email is received) we also consider traffic received in asmall time window (1 minute) before the screen is turned onas having timeliness constraints The results are robust tothe exact choice of time window Indeed some of the trafficthat we classify as interactive might be delay tolerant ega new email to be sent might be deferred for a little whileHowever the volume of traffic received by the phone whichusers would rather see immediately dominates by one or-der of magnitude the volume that is sent and delay-tolerantmessaging applications such as email contribute only a smallchunk of all traffic
Figure 21 shows the fraction of interactive traffic for eachuser We see that for about 90 of the users over 50of the traffic is interactive but for the rest almost none oftheir traffic is interactive Stated differently for differentsmartphone users almost all to almost none of the traffic isgenerated by applications in the background The extremesrepresent disparate ways in which people use smartphonesand which applications on the smartphone generate mosttraffic Our results imply that the energy savings that can behad by rescheduling network activity will vary across users
53 Diurnal patternsFigure 22(a) shows the diurnal pattern for an example
user with a sharp decline at night Figure 22(b) showsthe strength of the diurnal pattern for individual users by
0 20 40 60 80 100
User percentile
00
02
04
06
08
10
Interactive fraction
Figure 21 The fraction of interactive traffic
0 5 10 15 20
Hour of day
0
5
10
15
Traffic per hour (MB)
Receive
Send
(a)
0 20 40 60 80 100
User percentile
2
4
6
8
10
12
Diurnal ratio
Receive
Send
(b)
Figure 22 (a) The mean and 95 CI for trafficgenerated per hour by an example user (b) Thediurnal ratio of traffic per hour for all users
plotting the diurnal ratio of traffic As defined in sect33 thediurnal ratio reflects how much higher the mean during thepeak hour is to the overall mean
We see that the diurnal ratio varies across users but mosthave a strong diurnal behavior 80 of them generate overtwice their average amount of traffic in their peak hour Thisbehavior is likely a direct result of the high proportion ofinteractive traffic for most users and that user interactionsthemselves have a diurnal pattern
6 ENERGY CONSUMPTIONThe final aspect of smartphone usage that we investigate
is energy consumption Energy drain depends on two fac-tors i) user interactions and applications and ii) platformhardware and software If the second factor dominates theenergy drain of various users with the identical smartphonewill be similar Otherwise the energy drain will be as di-verse as user behaviors
We estimate the amount of energy drain based on the re-maining battery indicator which varies between 0 and 100If the battery indicator has gone down by X in a time pe-riod for a battery with capacity Y mAh we compute theenergy drain in that period to be X middot Y mAh 1 Given thatbatteries are complex electro-chemical devices [14 20] thiscomputation is approximate It assumes that the batterylevel indicator is linear with respect to energy drain
1mAh is technically a unit of charge yet is commonly used toindicate energy drain because battery voltage during normaloperations is typically constant For phones in our datasetthis is 4V so multiplying a mAh reading by 4 would yieldan accurate energy reading in milli-watt-hours
020
4060
8010
0
0 4 8 12 16 20Time (hour)
Bat
tery
leve
l (
) Benchmark1Benchmark2
Figure 23 Timelapse of the remaining battery levelindicator in controlled experiments with two differ-ent workloads at room temperature
110
100
1000
0 20 40 60 80 100User percentile
Bat
tery
Dra
in (
mA
h)
Figure 24 The mean and the upper end of the stan-dard deviation of one hour energy drain for Dataset1users during discharge periods
Controlled experiments suggest that the linearity assump-tion holds to a first order We run a benchmark load thatdrains the battery at a fixed rate in room temperature Un-der this benchmark if the battery level indicator decreaseslinearly with time it must be linear with respect to energydrain Figure 23 shows that the level indicator decreasesroughly linearly for two different benchmarks Benchmark1turns the screen on and off periodically Benchmark2 com-putes and idles periodically We conclude thus that the levelindicator can be used to estimate energy drain
Figure 24 shows the mean and standard deviation of en-ergy that users drain in an hour This graph is computed us-ing only periods in which the battery is not charging becauseenergy drain in those periods are of primary interest Wesee a two orders of magnitude difference among users Whileheaviest users drain close to 250 mAh the lightest of usersdrain only 10 mAh If the battery capacity is 1200 mAhthis leads to a lifetime variation from about 4 to 120 hours
Figure 25(a) shows for an example user that the drain isnot the same throughout the day but has diurnal variationsin which more energy is consumed during the day than dur-ing the night For this user the level of energy consumedchanges by roughly a factor of five Figure 25(b) plots thediurnal ratio of energy use for all users It shows that diurnalvariations occur with different strengths for all users
2060
100
140
0 4 8 12 16 20Hour of the day
Bat
tery
Dra
in (
mA
h)
(a)
0 20 40 60 80 100
12
34
56
7
User percentile
Diu
rnal
rat
io(b)
Figure 25 (a) The mean and 95 CI of energy drainof an example Dataset1 user (b) Diurnal ratio ofall users in Dataset1
Our results show that user activities contribute heavilytowards energy drain users in Dataset1 who have identi-cal smartphones drain energy at different rates and energydrain has diurnal patterns In the future we will developmethods to accurately quantify the energy consumption ofthe platform from that due to user-induced workload
7 IMPLICATIONS OF USER DIVERSITYWe uncover a surprising level of diversity among smart-
phone users For almost every aspect of usage that we studywe find one or more orders of magnitude difference betweenusers Our findings strongly motivate the need for customiz-ing smartphones to their users We believe that this needis greater than that for customizing ordinary cellphones orlaptops Ordinary cellphones do not have as rich an appli-cation environment Laptops are not as portable and aremore resource rich For example many users plug-in theirlaptops while using them
Customization can help at all levels Consider somethingas low-level as the battery Suppose we want batteries to beboth lightweight and last for at least a day with a high proba-bility Meeting the latter goal for all users of a given platformwill require catering to the heaviest users But that will leadto unnecessarily heavy batteries for many users (Highercapacity batteries are heavier) Offering multiple types ofbatteries with different lifetime-weight tradeoffs provides away out of this bind
At levels where intelligent mechanisms to improve user ex-perience or reduce energy consumption reside user diversitymotivates adapting to the smartphone user Driving thesemechanisms based on average case behaviors may not beeffective for a large fraction of the users
The ease and utility of customization depends on twoproperties of user behavior First despite quantitative dif-ferences there must be qualitative similarities among usersFor instance we should be able to describe the behavior ofall users with the same model Different users may havedifferent parameters of this model which will then lead toquantitative differences among them The presence of qual-itative similarities imply that users are not arbitrary pointsin space and it significantly simplifies the task of learninguser behavior Second user behavior in the past must alsobe predictive of the future Otherwise customization basedon past behaviors will be of little value in the future
In the next two sections we present evidence that theseproperties hold for several key aspects of smartphone usage
Dataset1 example user
000
00
010
002
00
030
0 100 200 300 400Session length (s)
Den
sity
Dataset2 example user
000
005
010
015
020
0 20 40 60 80 100Session length (s)
Den
sity
Figure 26 The PDF of session length for sampleusers of each dataset
In sect8 we show that user sessions and relative applicationpopularity can be described using simple models In sect9 weshow that energy drain can be described as well as predictedusing a ldquotrend trablerdquo framework
8 SMARTPHONE USAGE MODELSIn this section we develop simple models that describe
three aspects of smartphone usage ndash session lengths inter-arrival time between sessions and application popularityThe models are common across users but have different pa-rameters for different users While they do not completelydescribe user behavior they capture first order factors andrepresent a first step towards more complete modeling ofsmartphone usage More importantly along with the resultsof the next section they show that qualitative similaritiesdo exist among users
81 Session LengthsWe first consider the statistical properties of session length
distributions of users We find that session length valuestend to stationary With the KPSS test for level stationar-ity [13] 90 of the users have a p-value greater than 01The presence of stationarity is appealing because it suggeststhat past behavior is capable of predicting the future
We also find that session lengths are independent that isthe current value does not have a strong correlation with thevalues seen in the recent past With the Ljung-Box test forindependence [15] 96 of the users have a p-value that isgreater than 01
Stationarity and independence considered together sug-gest that session length values can be modeled as iid sam-ples from a distribution Choosing an appropriate distribu-tion however is complicated by the nature of the sessionlengths Most sessions are very short and the frequencydrops exponentially as the length increases However in-consistent with exponential behavior there are some verylong sessions in the tail for each user
We find that a mixture of exponential and Pareto distri-butions can model both ends of the spectrum The formercaptures short sessions and the latter captures long sessionsThat is session lengths can be described by the followingmixture model
r middot Exp(λ) + (1 minus r) middot Pareto(xm α)
In this equation r is the relative mix of the two distribu-tions λ is the rate of the exponential and xm and α are thelocation and shape parameters of the Pareto distribution
The location for a Pareto distribution represents the min-imum possible value of the random variable The location
010
0020
0030
00
0 1000 2000 3000Observed quantiles
Mod
el q
uant
iles
Figure 27 QQ plot of session lengths model for asample user
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed m
ix
(a) Relative mix (r)
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed r
ate
(b) Exponential rate (λ)
020
4060
0 20 40 60 80 100User percentile
Tim
eout
(s)
(c) Pareto location (xm)
minus1
0minus
04
02
06
10
0 20 40 60 80 100User percentile
Est
imat
ed s
hape
(d) Pareto shape (α)
Figure 28 Distribution of inferred model parame-ters that describe session length values of users inboth datasets
value that offers the best fit is the screen timeout value ofthe user because the session length PDF has a spike at thisvalue The spike corresponds to short sessions that are endedby the timeout (when the user forgets to switch the screenoff) we confirm this using controlled experiments with dif-ferent timeout values Figure 26 shows this spike at 60 and15 seconds for example users from each dataset The time-out provides a natural division between the two componentdistributions We automatically infer its approximate valueusing a simple spike detection algorithm
We use the EM algorithm to infer the maximum likelihoodestimation (MLE) of the remaining three parameters [6]Figure 27 shows the quality of this fit for an example userusing the QQ plot [3] Almost all quantiles are along they = x line indicating a good fit
Figure 28 shows the four inferred parameters for varioususers While users can be modeled using the same mix-
040
0080
00
0 4000 8000Observed quantiles
Mod
el q
uant
iles
Figure 29 QQ plot of session offtime model for asample user
050
010
0015
000 20 40 60 80 100
User percentileE
stim
ated
sca
le
(a) Weibull scale
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed s
hape
(b) Weibull shape
Figure 30 Distribution of inferred model parame-ters that describe the distribution of time betweensessions for users in both datasets
ture model the parameters of this model vary widely acrossusers Because of the way we construct our model the dis-tribution of the parameter r and xm also provide insightinto how frequently usersrsquo screen is switched off by the time-out and the relative popularity of different timeout values60 seconds is the most popular value likely because it is themost common default timeout and many users never changethe default
82 Time between SessionsWe find that the Weibull distribution can explain the
screen off times This distribution has two parameters re-ferred to as its scale and shape We find the MLE for theseparameters for each user From the QQ-plot in Figure 29we notice that the model predicts a greater probability ofseeing some very large offtimes than are observed in thedatasets However the probability of seeing these large off-times is small there are 27 data points that have a y-value greater than 8000 in that graph Hence we believethat Weibull provides a good fit for the length of intervalsbetween interactions
Figure 30 shows the distribution of the estimated shapeand scale of the fitted Weibull distributions Interestinglythe shape is consistently less than one Weibull shape valuesless than one indicate that the longer the screen has been offthe less likely it is for it to be turned on by the user This be-havior has interesting implications for power saving policiesFor instance periodic activities such as checking for emailwhen the screen has been off for a while may be deferred orrescheduled if needed without hurting user experience
020
4060
8010
0
0 20 40 60 80 100User percentile
MS
E (
)
(a) MSE
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed r
ate
(b) Rate parameter
Figure 31 (a) The mean square error (MSE) whenapplication popularity distribution is modeled usingan exponential (b) The inferred rate parameter ofthe exponential distribution for different users
83 Application PopularityWe find that for each user the relative popularity of ap-
plications can be well described by a simple exponential dis-tribution This qualitative invariant is useful for instanceto predict how many applications account for a given frac-tion of user attention For the example users in Figure 11this facet can be seen in the inset plots the semi-log of thepopularity distribution is very close to a straight line
Figure 31(a) shows that this exponential drop in applica-tion popularity is true for almost all users the mean squareerror between modeled exponential and actual popularitydistribution is less than 5 for 95 of the users
Figure 31(b) shows the inferred rate parameter of the ap-plication popularity distribution for various users We seethat the rate varies by an order of magnitude from 01 toalmost 1 The value of the rate essentially captures the paceof the drop in application popularity Lower values describeusers that use more applications on a regular basis Oneimplication of the wide range is that it may be feasible toretain all popular applications in memory for some users andnot for others
9 PREDICTING ENERGY DRAINIn this section we demonstrate the value of adapting to
user behaviors in the context of a mechanism to predict fu-ture energy drain on the smartphone Such a mechanismcan help with scheduling background tasks [9] estimatingbattery lifetime and improving user experience if it is esti-mated that the user will not fully drain the battery until thenext charging cycle [2 21]
Despite its many uses to our knowledge none of thesmartphones today provide a prediction of future energydrain Providing an accurate prediction of energy drain isdifficult2 User diversity of energy use makes any static pre-diction method highly inaccurate Even for the same userthere is a high variance in energy usage Figure 32 shows thisvariance by plotting the ratio of the standard deviation tothe mean energy drain over several time periods for users inDataset 1 The standard deviation of 10-minute windows ishigher than three times the mean for a fifth of the users The
2Energy drain prediction on smartphones is more challeng-ing than what is done for laptops Laptops provide a pre-diction of battery lifetime only for active use Smartphonepredictions on the other hand must cover multiple activeidle periods to be useful
01
23
45
0 20 40 60 80 100User percentile
Err
or r
atio
2 hours1 hour10 minutes
Figure 32 Battery drain predictors have to copewith the high variance in drain
time
window to
predict
n chunks
hellip
Trend Table
hellip
hellip
Figure 33 A userrsquos trend table captures how batterydrain in n consecutive chunks relates to drain in thesubsequent window
bursty nature of user interactions is one cause of this highvariance The variance is high even at longer time scalesThe standard deviation for two hour windows is larger than150 of the mean for half the users The diurnal patternsthat we showed earlier are one cause of high variance oversuch large time scales
Although the variance is high there are patterns in auserrsquos behaviors For instance we showed earlier that aphone that has been idle for a while is likely to remain idlein the near future and it will thus continue to draw energyat a similar rate
We hypothesize that a predictor tuned to user behaviorcan be accurate We present a simple personalized energydrain predictor that incorporates various user-specific fac-tors such as length of idle periods and different types ofbusy periods Instead of explicitly identifying these factorsour predictor captures them in a ldquotrend tablerdquo framework
Our framework is shown in Figure 33 Each entry in thetrend table is indexed by an n-tuple that represents energyusage readings from adjacent chunks of size δ each Thisindex points to energy usage statistics across all time win-dows of size w in the history that follow the n chunk valuesrepresented by that index Thus the trend table captureshow the energy usage in n consecutive chunks is related tothe energy use in the subsequent time windows of size w
In this paper we use n = 3 δ = 10 minutes and maintaindifferent trend tables for each window size w for which theprediction is required To keep the trend table small wequantize the energy readings of chunks We maintain themean and standard deviation for all subsequent windowsthat map to that index
To make a prediction we use the energy readings fromthe n immediately preceding chunks Let these readings be(x1 xn) Then we search for indices in the table thatdiffer from (x1 xn) by no more than a threshold s and
GenericShortminustermTimeminusofminusdayPersonalized
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
(a) 1-hour window
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
GenericShortminustermTimeminusofminusdayPersonalized
(b) 2-hour window
Figure 34 The ratio of error to true battery drainfor different energy drain predictors
predict based on them such that similar indices have higherrelative weight More precisely the prediction is
X
e stforalli=1n|he
iminusxi|lts
ke times ue
P
ke
where e iterates over all tuples in the trend table hei is
the value of the chunk i of the tuple e ue represents thestatistics stored for that index and the relative weight ke =s minus maxi=1n |he
i minus xi| For results shown here we uses = 05
Figures 34 shows how well our Personalized predictor worksfor predicting usage for two different future time windowsThere is one point per user in each graph which correspondsto the median error seen across many predictions for thatuser We see that the 90th percentile error is 25 for 1-hourwindow and 40 for 2-hour window
To place this level of accuracy in context the figure alsoshows the accurary of three alternative predictors TheGeneric predictor uses a trend table but instead of buildinga user-specific table it builds one that combines all usersWe see that its 90th percentile error is roughly twice thatof the Personalized predictor Its predictions are worse forhalf of the users
The Short-term predictor predicts energy drain simply asthe amount drained in the immediately preceding windowof the same size For both time windows its 90th percentileerror is no better than that of the Generic predictor
The Time-of-day predictor predicts energy drain based onvalues observed for the user at the same time of the day inthe past similar to that proposed by Banerjee et al [2]Thus this predictor has some degree of personalization butit learns user behavior in less detail than the Personalizedpredictor We see that its performance for 1-hour windowis similar to the Short-term predictor For 2-hour windowits error is higher than the Personalized predictor across theboard though its worst case error is lower than the Short-term and Generic predictors
Overall these results demonstrate the value of appropri-ately learning user behaviors to design intelligent mecha-nisms on smartphones
10 RELATED WORKIn a range of domains there is a rich history of work that
characterizes user workloads However because smartphoneadoption has gathered pace relatively recently our work rep-resents one of the few to study how people use smartphones
Along with other recent works it helps complete the pic-ture Banerjee et al and Rahmati et al report on batterycharging behaviors [2 18] Like us they find considerablevariation among users Banerjee et al also propose a pre-dictor that estimates the excess energy of the battery atthe time of charging using a histogram of past battery us-age [2] Shye et al study the power consumption character-istics of 20 users [24] They infer properties such as whichcomponents in the phone consume most power and exploreoptimizations based on these characteristics Rahmati andZhong study 14 users of a particular demographic to studywhich applications are popular in that demographic wherethe phone is used and how it is shared among users [19]
In contrast to these works we focus on understandingdifferent aspects of smartphone use (eg interactions andtraffic) and on exposing the diversity of user behaviors in-stead of only the average behavior Our study also entailsan order of magnitude more users than previous efforts
Like us other researchers have developed logging utilitiesMyExperience is one such early utility [10] and the worksabove involve custom utilities as well Our focus in thispaper is not on the development of logging tools but onanalyzing their data to gain insight into user behavior
There is a body of work in modeling the aggregate be-havior of mobile users Using traces from a large cellularoperator some network related aspects of mobile usage havebeen modeled Halepovic et al and Williamson et al re-port that call arrivals are bursty and present diurnal pat-terns [28] Willkomm et al and Brown et al report thatmobile users call duration can be approximately modeled bya lognormal distribution [4 29] We use traces collected onthe mobile device itself and focus on modeling the interac-tions of individual users instead of the aggregate behavior
11 CONCLUSIONSBy studying 255 users of two different smartphone plat-
forms we comprehensively characterized user activities andtheir impact on network and battery We quantify manyhitherto unknown aspects of smartphone usage User diver-sity is an overarching theme in our findings For instancedifferent users interact with their phones 10-200 times a dayon average the mean interaction length of different users
is 10-250 seconds and users receive 1-1000 MB of data perday where 10-90 is received as part of active use
This extent of user diversity implies that mechanisms thatwork for the average case may be ineffective for a large frac-tion of the users Instead learning and adapting to userbehaviors is likely to be more effective as demonstrated byour personalized energy drain predictor We show that qual-itative similarities exist among users to facilitate the devel-opment of such mechanisms For instance the longer theuser has not interacted with the phone the less likely it isfor her to start interacting with it and application popular-ity for a user follows a simple exponential distribution
Our study points to ways in which smartphone platformsshould be enhanced Effective adaptation will require futureplatforms to support light-weight tools that monitor andlearn user behaviors in situ It will also require them toexpose appropriate knobs to control the behavior of lower-level components (eg CPU or radio)
12 ACKNOWLEDGEMENTSWe are grateful to the users in Dataset1 that ran our log-
ger and to the providers of Dataset2 Elizabeth Dawson pro-vided administrative support for data collection Jean BolotKurt Partridge Stefan Saroiu Lin Zhong and the MobiSysreviewers provided many comments that helped improve thepaper We thank them all This work is supported in partby the NSF grants CCR-0120778 and CNS-0627084
13 REFERENCES[1] Balasubramanian N Balasubramanian A
and Venkataramani A Energy consumption inmobile phones A measurement study andimplications for network applications In IMC (2009)
[2] Banerjee N Rahmati A Corner M D
Rollins S and Zhong L Users and batteriesInteractions and adaptive energy management inmobile systems In Ubicomp (2007)
[3] Becker R Chambers J and Wilks A The newS language Chapman amp HallCRC 1988
[4] Brown L Gans N Mandelbaum A Sakov
A Shen H Zeltyn S and Zhao L Statisticalanalysis of a telephone call center Journal of theAmerican Statistical Association 100 469 (2005)
[5] Church K and Smyth B Understanding mobileinformation needs In MobileHCI (2008)
[6] Dempster A Laird N and Rubin D Maximumlikelihood from incomplete data via the EM algorithmJournal of the Royal Statistical Society Series B(Methodological) 39 1 (1977)
[7] Esfahbod B Preload An adaptive prefetchingdaemon Masterrsquos thesis University of Toronto 2006
[8] Falaki H Govindan R and Estrin D Smartscreen management on mobile phones Tech Rep 74Center for Embedded Networked Sesning 2009
[9] Flinn J and Satyanarayanan M Managingbattery lifetime with energy-aware adaptation ACMTOCS 22 2 (2004)
[10] Froehlich J Chen M Consolvo S
Harrison B and Landay J MyExperience Asystem for in situ tracing and capturing of userfeedback on mobile phones In MobiSys (2007)
[11] Fukuda K Cho K and Esaki H The impact ofresidential broadband traffic on Japanese ISPbackbones SIGCOMM CCR 35 1 (2005)
[12] Kotz D and Essien K Analysis of a campus-widewireless network Wireless Networks 11 1ndash2 (2005)
[13] Kwiatkowski D Phillips P Schmidt P and
Shin Y Testing the null hypothesis of stationarityagainst the alternative of a unit root Journal ofEconometrics 54 1-3 (1992)
[14] Lahiri K Dey S Panigrahi D and
Raghunathan A Battery-driven system design Anew frontier in low power design In Asia SouthPacific design automationVLSI Design (2002)
[15] Ljung G and Box G On a measure of lack of fitin time series models Biometrika 65 2 (1978)
[16] Smartphone futures 2009-2014 httpwwwportioresearchcomSmartphone09-14html
[17] Rahmati A Qian A and Zhong L
Understanding human-battery interaction on mobilephones In MobileHCI (2007)
[18] Rahmati A and Zhong L Humanndashbatteryinteraction on mobile phones Pervasive and MobileComputing 5 5 (2009)
[19] Rahmati A and Zhong L A longitudinal study ofnon-voice mobile phone usage by teens from anunderserved urban community Tech Rep 0515-09Rice University 2009
[20] Rao R Vrudhula S and Rakhmatov D
Battery modeling for energy aware system designIEEE Computer 36 12 (2003)
[21] Ravi N Scott J Han L and Iftode L
Context-aware battery management for mobilephones In PerCom (2008)
[22] Sharma A Navda V Ramjee R
Padmanabhan V and Belding E Cool-TetherEnergy efficient on-the-fly WiFi hot-spots usingmobile phones In CoNEXT (2009)
[23] Sharma C US wireless data market update - Q32009 httpwwwchetansharmacomusmarketupdateq309htm
[24] Shye A Sholbrock B and G M Into the wildStudying real user activity patterns to guide poweroptimization for mobile architectures In Micro (2009)
[25] Snol L More smartphones than desktop PCs by2011 httpwwwpcworldcomarticle171380more_smartphones_than_desktop_pcs_by_2011
html
[26] Sohn T Li K Griswold W and Hollan J
A diary study of mobile information needs In SIGCHI(2008)
[27] Trestian I Ranjan S Kuzmanovic A and
Nucci A Measuring serendipity Connecting peoplelocations and interests in a mobile 3G network InIMC (2009)
[28] Williamson C Halepovic E Sun H and Wu
Y Characterization of CDMA2000 Cellular DataNetwork Traffic In Local Computer Networks (2005)
[29] Willkomm D Machiraju S Bolot J and
Wolisz A Primary users in cellular networks Alarge-scale measurement study In DySPAN (2008)
- Introduction
- Data Collection
-
- Dataset1
- Dataset2
- Representativeness of conclusions
-
- User Interactions
-
- Interaction Time
- Interaction Sessions
- Diurnal Patterns
-
- Application Usage
-
- Number of applications
- Application Popularity
- Application Sessions
-
- Traffic
-
- Traffic per day
- ``Interactive traffic
- Diurnal patterns
-
- Energy Consumption
- Implications of User Diversity
- Smartphone Usage Models
-
- Session Lengths
- Time between Sessions
- Application Popularity
-
- Predicting Energy Drain
- Related Work
- Conclusions
- Acknowledgements
- References
-
15
913
1721
0 20 40 60 80 100User percentile
Diu
rnal
rat
io
Dataset1Dataset2
(a) Interaction time
12
34
5
0 20 40 60 80 100User percentile
Diu
rnal
rat
io
Dataset1Dataset2
(b) Number of sessions
15
913
1721
0 20 40 60 80 100User percentile
Diu
rnal
rat
io
Dataset1Dataset2
(c) Session length
Figure 9 Diurnal ratio of the interaction time the number of sessions and the session length for differentusers The y-axis ranges for the number of sessions is different
laun
cher gm
alar
mcl
ock
talk
cont
acts
map
sm
ms
phon
em
usic
brow
ser
cam
era
syst
emse
nsbe
nchm
ark
wer
tago
vend
ing
cale
ndar
setti
ngs
tippi
ntim
ete
tris
andr
oid
Dataset1 example user
Usa
ge r
atio
000
010
020
030
0 20 401eminus
070
01
iexp
lore
gwes
tmai
lho
me
bubb
lebr
eake
rP
oker
Live
Sea
rch
cale
ndar
Goo
gleM
aps
pout
look
TP
CS
olita
reK
aGlo
mpv
bloa
dcp
rog
serv
ices
tels
hell
pim
gap
pman
YG
oNet
wm
play
er
Dataset2 example user
Usa
ge r
atio
000
005
010
015
0 20 40 601eminus
060
01
Figure 11 Relative time spent running each application for example users in each dataset Inset is thesemi-log plot of application popularity
are binary names even some popular applications may ap-pear unfamiliar For instance in Dataset1 ldquolauncherrdquo is thedefault home screen application on Android in Dataset2ldquogwesrdquo (Graphical Windows and Events Subsystem) is thegraphical shell on Window Mobile
The graphs show that relative application popularity dropsquickly for both users In sect83 we show that for all usersapplication popularity can be modeled by an exponentialdistribution
Diurnal patterns Interestingly application popularityis not stable throughout the day but has a diurnal patternlike the other aspects of smartphone use That is the rel-ative popularity of an application is different for differenttimes of the day Figure 12 illustrates this for an exam-ple user in Dataset2 We see for instance that tmailexewhich is a messaging application on Windows Mobile ismore popular during the day than night Time dependentapplication popularity was recently reported by Trestian etal based on an analysis of the network traffic logs from a3G provider [27] Our analysis based on direct observationof user behavior confirms this effect
0 4 8 12 17 22Hour of day
020
4060
80U
sage
rat
io (
)
servicesexewmplayerexecalendarexepoutlookexeWLMMessengercdialexepimgexecprogexeiexploreexehomeexetmailexe
Figure 12 Relative time spent with each applicationduring each hour of the day for a sample user andher top applications
comm 44
browsing 10
other 11
prod 19media 5
maps 5games 2system 5
Dataset1
comm 49
browsing 12 other 15
prod 2media 9
maps 2
games 10
system 1
Dataset2
Figure 13 Relative popularity of each applicationcategory across all users in each dataset
Aggregate view application popularity To pro-vide an aggregate view of what users use their smartphonesfor we categorize applications into eight distinct categoriesi) communication contains applications for exchanging mes-sages (eg email SMS IM) and voice calls ii) browsingcontains Web browser search and social networking appli-cations iii) media contains applications for consuming orcreating media content (eg pictures music videos) iv)productivity contains applications for calendars alarms andfor viewing and creating text documents (eg Office PDFreader) v) system contains applications for changing userpreferences and viewing system state (eg file explorer)vi) games vii) maps and viii) other contains applicationsthat we could not include in any of the categories aboveeg because we did not know their function
Figure 13 shows the mean relative popularity of each ap-plication category across all users in each dataset Whilethe results are not identical across the two datasets they aresimilar to a first order Communication dominates in bothBrowsing is another major contributor in both datasetsMaps media and games have a comparatively lower butnevertheless substantial share of user attention
Relationship to user demographic To understandthe extent to which user demographic determines applica-tion popularity Figure 14 shows the mean and 95 CI ofrelative popularity of each application category for differ-ent user demographics As for interaction time (sect31) wesee that the impact of user demographics in our datasets isminimal In Dataset2 the relative popularity of various ap-plication types is similar for each of the four demographicsIn Dataset1 there are noticeable differences in the mean forcommunication games and productivity applications Highschool students use communication and games applicationsmore while knowledge workers use productivity applications
Dataset1
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
High schoolKnowledge worker
Dataset2
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
SCOPBPULPU
Figure 14 The mean and 95 CI of relative pop-ularity of application categories among users of dif-ferent demographics
more However considering the overlap of confidence in-tervals these differences in application popularity are notstatistically significant
From this result and the earlier one on the lack of depen-dence between user demographic and interaction time (sect31)we conclude that user demographic at least as defined inour datasets cannot reliably predict how a user will use thephone While demographic information appears to help insome cases (for eg the variation in usage of productivitysoftware in Dataset1) such cases are not the norm and itis hard to guess when demographic information would beuseful Pending development of other ways to classify userssuch that these classes more predictably explain the varia-tion incorporating factors specific to a user appear necessaryThis insensitivity to user demographic has positive as wellas negative implications A negative is that personalizationis more complex we cannot predict a usersrsquo behavior byknowing their demographic A positive implication is thatthe range of user behaviors along many dimensions of in-terest can be found in several common demographics Thissimplifies the task of uncovering the range because recruitingsubjects across multiple demographics tends to be difficult
Relationship to interaction time We also study ifusers that interact more with their phones tend to use differ-ent applications For each dataset we sort users based ontheir average interaction time per day and partition theminto different classes For Dataset1 we use two classesone each for the top and the bottom half of the users ForDataset2 which has more users we use three classes for the
Dataset1
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
HighLow
Dataset2
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
rHighMediumLow
Figure 15 The mean and 95 CI of relative pop-ularity of application categories among differentclasses of users based on interaction time per day
top middle and bottom third of the users Figure 15 showsthe mean and 95 CI for relative time spent with each ap-plication category by each user class We see that users indifferent classes have similar application usage Thus wecannot explain why some users use the phone more simplybased on the applications that they use
43 Application SessionsWe now study the characteristics of application sessions
We conduct this analysis only for Dataset2 based on times-tamps for when an application is started and ended Dataset1does not contain this information precisely Because appli-cations can run in the background start and end refer tothe period when the application is in the foreground
Applications run per interaction We begin by study-ing the number of applications that users run in an interac-tion session Figure 16 shows that an overwhelming major-ity close to 90 of interactions include only one applicationThis graph aggregates data across all users We did not findstatistically significant differences between users A largefraction of sessions of all users have only one application
That interaction sessions very often have only one appli-cation session suggests that users tend to interact with theirsmartphone for one task (eg reading email checking cal-endar etc) at a time and most of these tasks require theuse of only one application
Application session lengths Because interactionsessions are dominated by those with only one application
892
7422 06 02 01
1 2 3 4 5 6
020
4060
8010
0
Apps per session
Per
cent
age
()
Figure 16 Histogram of the number of applica-tions called during each interaction session for allthe users in Dataset2
map
s
gam
es
com
m
med
ia
brow
sing
prod
syst
em
othe
r
110
1000
Ses
sion
leng
th (
s)
Figure 17 The mean and 95 CI of session lengthsof different application categories The y-axis is logscale
the overall properties of application sessions such as theirlengths are similar to those of interaction sessions (sect3)
However studying the session lengths of applications sep-arately reveals interesting insights Different applicationtypes have different session lengths as shown in Figure 17for the categories defined earlier Interactions with mapsand games tend to be the longest and those related to pro-ductivity and system tend to be the shortest
Further given an application different users run themfor different times Figure 18 shows this effect for a mes-saging application tmailexe and a browsing applicationiexploreexe For each application the mean session lengthsof users differ by more than two orders of magnitude
These observations help explain why users have differentsession lengths (sect32) They prefer different applicationsand those application sessions tend to have different lengthsFurther analysis confirms this phenomenon For instance ifwe divide users into two classes based on their mean sessionlengths the popularity of games is twice as high in the classwith high session lengths
5 TRAFFICIn this section we investigate traffic generated by smart-
phones Unlike interaction events and application use net-work traffic is not an intentional user action but a side-effectof those actions Most users are likely oblivious to how much
tmail1
1010
010
00
0 20 40 60 80 100User percentile
Ses
sion
leng
th (
s)iexplore
110
100
1000
0 20 40 60 80 100User percentile
Ses
sion
leng
th (
s)Figure 18 The mean and the upper end of the stan-dard deviation of session lengths of two applications
0 20 40 60 80 100
User percentile
0
1
10
100
1000
10000
Traffic per day (MB) Receive
Send
Figure 19 The mean and the upper end of the stan-dard deviation of the traffic sent and received perday by users in Dataset1
traffic they generate We show that the diversity and diur-nal patterns of this side-effect match those of user actionsthemselves
The analysis in this section includes only Dataset1 we donot have traffic information for Dataset2 In Dataset1 werecord all of the data sent (or received) by the phone exceptfor that exchanged over the USB link ie the traffic hereinincludes data over the 3G radio and the 80211 wireless link
51 Traffic per dayFigure 19 shows that the amount of traffic sent and re-
ceived per day differs across users by almost three orders ofmagnitude The traffic received ranges from 1 to 1000 MBand the traffic sent ranges from 03 to 100 MB The medianvalues are 30 MB sent and 5 MB received
Our results indicate that traffic generated in a day bysmartphone users is comparable to traffic generated by PCsa few years ago A study of a campus WiFi network revealedthat on average users were generating 27 MB of traffic in2001 and 71 MB in 2003 [12] A study of Japanese residen-tial broadband users in 2006 revealed that on average usersgenerate 1000 MB of traffic per day [11] This high level oftraffic has major implications for the provisioning of wirelesscarrier networks as smartphone adoption increases
Relationship to application types To investigate ifcertain types of applications are favored more by users thatgenerate more traffice we divide the users into two equalclasses based on their sum of sent and received traffic perday Figure 20 shows mean and 95 CI of relative popularityof each application category for each user class Expectedly
Dataset1
Per
cent
age
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
HighLow
Figure 20 The mean and 95 CI of relative popu-larity of each application category among high andlow traffic consumers
it shows that communication applications are more popularamong users that consume more traffic
52 ldquoInteractiverdquo trafficWe next estimate what fraction of the total traffic is ldquoin-
teractiverdquo that is has timeliness constraints Approachesto reduce power consumption of data transfers often advo-cate rescheduling network transfers for instance by delay-ing some transfers so that multiple of them may be bun-dled [1 22] Such policies are difficult to implement forinteractive traffic without hurting user response time andthus are likely to be less valuable if the bulk of the traffic isinteractive
We classify traffic as interactive if it was generated whenthe screen is on This classification method might classifysome background traffic as interactive We expect the errorto be low because the screen is on for a small fraction ofthe time for most users Because certain user interactionswith the phone begin right after traffic exchange (eg anew email is received) we also consider traffic received in asmall time window (1 minute) before the screen is turned onas having timeliness constraints The results are robust tothe exact choice of time window Indeed some of the trafficthat we classify as interactive might be delay tolerant ega new email to be sent might be deferred for a little whileHowever the volume of traffic received by the phone whichusers would rather see immediately dominates by one or-der of magnitude the volume that is sent and delay-tolerantmessaging applications such as email contribute only a smallchunk of all traffic
Figure 21 shows the fraction of interactive traffic for eachuser We see that for about 90 of the users over 50of the traffic is interactive but for the rest almost none oftheir traffic is interactive Stated differently for differentsmartphone users almost all to almost none of the traffic isgenerated by applications in the background The extremesrepresent disparate ways in which people use smartphonesand which applications on the smartphone generate mosttraffic Our results imply that the energy savings that can behad by rescheduling network activity will vary across users
53 Diurnal patternsFigure 22(a) shows the diurnal pattern for an example
user with a sharp decline at night Figure 22(b) showsthe strength of the diurnal pattern for individual users by
0 20 40 60 80 100
User percentile
00
02
04
06
08
10
Interactive fraction
Figure 21 The fraction of interactive traffic
0 5 10 15 20
Hour of day
0
5
10
15
Traffic per hour (MB)
Receive
Send
(a)
0 20 40 60 80 100
User percentile
2
4
6
8
10
12
Diurnal ratio
Receive
Send
(b)
Figure 22 (a) The mean and 95 CI for trafficgenerated per hour by an example user (b) Thediurnal ratio of traffic per hour for all users
plotting the diurnal ratio of traffic As defined in sect33 thediurnal ratio reflects how much higher the mean during thepeak hour is to the overall mean
We see that the diurnal ratio varies across users but mosthave a strong diurnal behavior 80 of them generate overtwice their average amount of traffic in their peak hour Thisbehavior is likely a direct result of the high proportion ofinteractive traffic for most users and that user interactionsthemselves have a diurnal pattern
6 ENERGY CONSUMPTIONThe final aspect of smartphone usage that we investigate
is energy consumption Energy drain depends on two fac-tors i) user interactions and applications and ii) platformhardware and software If the second factor dominates theenergy drain of various users with the identical smartphonewill be similar Otherwise the energy drain will be as di-verse as user behaviors
We estimate the amount of energy drain based on the re-maining battery indicator which varies between 0 and 100If the battery indicator has gone down by X in a time pe-riod for a battery with capacity Y mAh we compute theenergy drain in that period to be X middot Y mAh 1 Given thatbatteries are complex electro-chemical devices [14 20] thiscomputation is approximate It assumes that the batterylevel indicator is linear with respect to energy drain
1mAh is technically a unit of charge yet is commonly used toindicate energy drain because battery voltage during normaloperations is typically constant For phones in our datasetthis is 4V so multiplying a mAh reading by 4 would yieldan accurate energy reading in milli-watt-hours
020
4060
8010
0
0 4 8 12 16 20Time (hour)
Bat
tery
leve
l (
) Benchmark1Benchmark2
Figure 23 Timelapse of the remaining battery levelindicator in controlled experiments with two differ-ent workloads at room temperature
110
100
1000
0 20 40 60 80 100User percentile
Bat
tery
Dra
in (
mA
h)
Figure 24 The mean and the upper end of the stan-dard deviation of one hour energy drain for Dataset1users during discharge periods
Controlled experiments suggest that the linearity assump-tion holds to a first order We run a benchmark load thatdrains the battery at a fixed rate in room temperature Un-der this benchmark if the battery level indicator decreaseslinearly with time it must be linear with respect to energydrain Figure 23 shows that the level indicator decreasesroughly linearly for two different benchmarks Benchmark1turns the screen on and off periodically Benchmark2 com-putes and idles periodically We conclude thus that the levelindicator can be used to estimate energy drain
Figure 24 shows the mean and standard deviation of en-ergy that users drain in an hour This graph is computed us-ing only periods in which the battery is not charging becauseenergy drain in those periods are of primary interest Wesee a two orders of magnitude difference among users Whileheaviest users drain close to 250 mAh the lightest of usersdrain only 10 mAh If the battery capacity is 1200 mAhthis leads to a lifetime variation from about 4 to 120 hours
Figure 25(a) shows for an example user that the drain isnot the same throughout the day but has diurnal variationsin which more energy is consumed during the day than dur-ing the night For this user the level of energy consumedchanges by roughly a factor of five Figure 25(b) plots thediurnal ratio of energy use for all users It shows that diurnalvariations occur with different strengths for all users
2060
100
140
0 4 8 12 16 20Hour of the day
Bat
tery
Dra
in (
mA
h)
(a)
0 20 40 60 80 100
12
34
56
7
User percentile
Diu
rnal
rat
io(b)
Figure 25 (a) The mean and 95 CI of energy drainof an example Dataset1 user (b) Diurnal ratio ofall users in Dataset1
Our results show that user activities contribute heavilytowards energy drain users in Dataset1 who have identi-cal smartphones drain energy at different rates and energydrain has diurnal patterns In the future we will developmethods to accurately quantify the energy consumption ofthe platform from that due to user-induced workload
7 IMPLICATIONS OF USER DIVERSITYWe uncover a surprising level of diversity among smart-
phone users For almost every aspect of usage that we studywe find one or more orders of magnitude difference betweenusers Our findings strongly motivate the need for customiz-ing smartphones to their users We believe that this needis greater than that for customizing ordinary cellphones orlaptops Ordinary cellphones do not have as rich an appli-cation environment Laptops are not as portable and aremore resource rich For example many users plug-in theirlaptops while using them
Customization can help at all levels Consider somethingas low-level as the battery Suppose we want batteries to beboth lightweight and last for at least a day with a high proba-bility Meeting the latter goal for all users of a given platformwill require catering to the heaviest users But that will leadto unnecessarily heavy batteries for many users (Highercapacity batteries are heavier) Offering multiple types ofbatteries with different lifetime-weight tradeoffs provides away out of this bind
At levels where intelligent mechanisms to improve user ex-perience or reduce energy consumption reside user diversitymotivates adapting to the smartphone user Driving thesemechanisms based on average case behaviors may not beeffective for a large fraction of the users
The ease and utility of customization depends on twoproperties of user behavior First despite quantitative dif-ferences there must be qualitative similarities among usersFor instance we should be able to describe the behavior ofall users with the same model Different users may havedifferent parameters of this model which will then lead toquantitative differences among them The presence of qual-itative similarities imply that users are not arbitrary pointsin space and it significantly simplifies the task of learninguser behavior Second user behavior in the past must alsobe predictive of the future Otherwise customization basedon past behaviors will be of little value in the future
In the next two sections we present evidence that theseproperties hold for several key aspects of smartphone usage
Dataset1 example user
000
00
010
002
00
030
0 100 200 300 400Session length (s)
Den
sity
Dataset2 example user
000
005
010
015
020
0 20 40 60 80 100Session length (s)
Den
sity
Figure 26 The PDF of session length for sampleusers of each dataset
In sect8 we show that user sessions and relative applicationpopularity can be described using simple models In sect9 weshow that energy drain can be described as well as predictedusing a ldquotrend trablerdquo framework
8 SMARTPHONE USAGE MODELSIn this section we develop simple models that describe
three aspects of smartphone usage ndash session lengths inter-arrival time between sessions and application popularityThe models are common across users but have different pa-rameters for different users While they do not completelydescribe user behavior they capture first order factors andrepresent a first step towards more complete modeling ofsmartphone usage More importantly along with the resultsof the next section they show that qualitative similaritiesdo exist among users
81 Session LengthsWe first consider the statistical properties of session length
distributions of users We find that session length valuestend to stationary With the KPSS test for level stationar-ity [13] 90 of the users have a p-value greater than 01The presence of stationarity is appealing because it suggeststhat past behavior is capable of predicting the future
We also find that session lengths are independent that isthe current value does not have a strong correlation with thevalues seen in the recent past With the Ljung-Box test forindependence [15] 96 of the users have a p-value that isgreater than 01
Stationarity and independence considered together sug-gest that session length values can be modeled as iid sam-ples from a distribution Choosing an appropriate distribu-tion however is complicated by the nature of the sessionlengths Most sessions are very short and the frequencydrops exponentially as the length increases However in-consistent with exponential behavior there are some verylong sessions in the tail for each user
We find that a mixture of exponential and Pareto distri-butions can model both ends of the spectrum The formercaptures short sessions and the latter captures long sessionsThat is session lengths can be described by the followingmixture model
r middot Exp(λ) + (1 minus r) middot Pareto(xm α)
In this equation r is the relative mix of the two distribu-tions λ is the rate of the exponential and xm and α are thelocation and shape parameters of the Pareto distribution
The location for a Pareto distribution represents the min-imum possible value of the random variable The location
010
0020
0030
00
0 1000 2000 3000Observed quantiles
Mod
el q
uant
iles
Figure 27 QQ plot of session lengths model for asample user
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed m
ix
(a) Relative mix (r)
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed r
ate
(b) Exponential rate (λ)
020
4060
0 20 40 60 80 100User percentile
Tim
eout
(s)
(c) Pareto location (xm)
minus1
0minus
04
02
06
10
0 20 40 60 80 100User percentile
Est
imat
ed s
hape
(d) Pareto shape (α)
Figure 28 Distribution of inferred model parame-ters that describe session length values of users inboth datasets
value that offers the best fit is the screen timeout value ofthe user because the session length PDF has a spike at thisvalue The spike corresponds to short sessions that are endedby the timeout (when the user forgets to switch the screenoff) we confirm this using controlled experiments with dif-ferent timeout values Figure 26 shows this spike at 60 and15 seconds for example users from each dataset The time-out provides a natural division between the two componentdistributions We automatically infer its approximate valueusing a simple spike detection algorithm
We use the EM algorithm to infer the maximum likelihoodestimation (MLE) of the remaining three parameters [6]Figure 27 shows the quality of this fit for an example userusing the QQ plot [3] Almost all quantiles are along they = x line indicating a good fit
Figure 28 shows the four inferred parameters for varioususers While users can be modeled using the same mix-
040
0080
00
0 4000 8000Observed quantiles
Mod
el q
uant
iles
Figure 29 QQ plot of session offtime model for asample user
050
010
0015
000 20 40 60 80 100
User percentileE
stim
ated
sca
le
(a) Weibull scale
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed s
hape
(b) Weibull shape
Figure 30 Distribution of inferred model parame-ters that describe the distribution of time betweensessions for users in both datasets
ture model the parameters of this model vary widely acrossusers Because of the way we construct our model the dis-tribution of the parameter r and xm also provide insightinto how frequently usersrsquo screen is switched off by the time-out and the relative popularity of different timeout values60 seconds is the most popular value likely because it is themost common default timeout and many users never changethe default
82 Time between SessionsWe find that the Weibull distribution can explain the
screen off times This distribution has two parameters re-ferred to as its scale and shape We find the MLE for theseparameters for each user From the QQ-plot in Figure 29we notice that the model predicts a greater probability ofseeing some very large offtimes than are observed in thedatasets However the probability of seeing these large off-times is small there are 27 data points that have a y-value greater than 8000 in that graph Hence we believethat Weibull provides a good fit for the length of intervalsbetween interactions
Figure 30 shows the distribution of the estimated shapeand scale of the fitted Weibull distributions Interestinglythe shape is consistently less than one Weibull shape valuesless than one indicate that the longer the screen has been offthe less likely it is for it to be turned on by the user This be-havior has interesting implications for power saving policiesFor instance periodic activities such as checking for emailwhen the screen has been off for a while may be deferred orrescheduled if needed without hurting user experience
020
4060
8010
0
0 20 40 60 80 100User percentile
MS
E (
)
(a) MSE
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed r
ate
(b) Rate parameter
Figure 31 (a) The mean square error (MSE) whenapplication popularity distribution is modeled usingan exponential (b) The inferred rate parameter ofthe exponential distribution for different users
83 Application PopularityWe find that for each user the relative popularity of ap-
plications can be well described by a simple exponential dis-tribution This qualitative invariant is useful for instanceto predict how many applications account for a given frac-tion of user attention For the example users in Figure 11this facet can be seen in the inset plots the semi-log of thepopularity distribution is very close to a straight line
Figure 31(a) shows that this exponential drop in applica-tion popularity is true for almost all users the mean squareerror between modeled exponential and actual popularitydistribution is less than 5 for 95 of the users
Figure 31(b) shows the inferred rate parameter of the ap-plication popularity distribution for various users We seethat the rate varies by an order of magnitude from 01 toalmost 1 The value of the rate essentially captures the paceof the drop in application popularity Lower values describeusers that use more applications on a regular basis Oneimplication of the wide range is that it may be feasible toretain all popular applications in memory for some users andnot for others
9 PREDICTING ENERGY DRAINIn this section we demonstrate the value of adapting to
user behaviors in the context of a mechanism to predict fu-ture energy drain on the smartphone Such a mechanismcan help with scheduling background tasks [9] estimatingbattery lifetime and improving user experience if it is esti-mated that the user will not fully drain the battery until thenext charging cycle [2 21]
Despite its many uses to our knowledge none of thesmartphones today provide a prediction of future energydrain Providing an accurate prediction of energy drain isdifficult2 User diversity of energy use makes any static pre-diction method highly inaccurate Even for the same userthere is a high variance in energy usage Figure 32 shows thisvariance by plotting the ratio of the standard deviation tothe mean energy drain over several time periods for users inDataset 1 The standard deviation of 10-minute windows ishigher than three times the mean for a fifth of the users The
2Energy drain prediction on smartphones is more challeng-ing than what is done for laptops Laptops provide a pre-diction of battery lifetime only for active use Smartphonepredictions on the other hand must cover multiple activeidle periods to be useful
01
23
45
0 20 40 60 80 100User percentile
Err
or r
atio
2 hours1 hour10 minutes
Figure 32 Battery drain predictors have to copewith the high variance in drain
time
window to
predict
n chunks
hellip
Trend Table
hellip
hellip
Figure 33 A userrsquos trend table captures how batterydrain in n consecutive chunks relates to drain in thesubsequent window
bursty nature of user interactions is one cause of this highvariance The variance is high even at longer time scalesThe standard deviation for two hour windows is larger than150 of the mean for half the users The diurnal patternsthat we showed earlier are one cause of high variance oversuch large time scales
Although the variance is high there are patterns in auserrsquos behaviors For instance we showed earlier that aphone that has been idle for a while is likely to remain idlein the near future and it will thus continue to draw energyat a similar rate
We hypothesize that a predictor tuned to user behaviorcan be accurate We present a simple personalized energydrain predictor that incorporates various user-specific fac-tors such as length of idle periods and different types ofbusy periods Instead of explicitly identifying these factorsour predictor captures them in a ldquotrend tablerdquo framework
Our framework is shown in Figure 33 Each entry in thetrend table is indexed by an n-tuple that represents energyusage readings from adjacent chunks of size δ each Thisindex points to energy usage statistics across all time win-dows of size w in the history that follow the n chunk valuesrepresented by that index Thus the trend table captureshow the energy usage in n consecutive chunks is related tothe energy use in the subsequent time windows of size w
In this paper we use n = 3 δ = 10 minutes and maintaindifferent trend tables for each window size w for which theprediction is required To keep the trend table small wequantize the energy readings of chunks We maintain themean and standard deviation for all subsequent windowsthat map to that index
To make a prediction we use the energy readings fromthe n immediately preceding chunks Let these readings be(x1 xn) Then we search for indices in the table thatdiffer from (x1 xn) by no more than a threshold s and
GenericShortminustermTimeminusofminusdayPersonalized
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
(a) 1-hour window
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
GenericShortminustermTimeminusofminusdayPersonalized
(b) 2-hour window
Figure 34 The ratio of error to true battery drainfor different energy drain predictors
predict based on them such that similar indices have higherrelative weight More precisely the prediction is
X
e stforalli=1n|he
iminusxi|lts
ke times ue
P
ke
where e iterates over all tuples in the trend table hei is
the value of the chunk i of the tuple e ue represents thestatistics stored for that index and the relative weight ke =s minus maxi=1n |he
i minus xi| For results shown here we uses = 05
Figures 34 shows how well our Personalized predictor worksfor predicting usage for two different future time windowsThere is one point per user in each graph which correspondsto the median error seen across many predictions for thatuser We see that the 90th percentile error is 25 for 1-hourwindow and 40 for 2-hour window
To place this level of accuracy in context the figure alsoshows the accurary of three alternative predictors TheGeneric predictor uses a trend table but instead of buildinga user-specific table it builds one that combines all usersWe see that its 90th percentile error is roughly twice thatof the Personalized predictor Its predictions are worse forhalf of the users
The Short-term predictor predicts energy drain simply asthe amount drained in the immediately preceding windowof the same size For both time windows its 90th percentileerror is no better than that of the Generic predictor
The Time-of-day predictor predicts energy drain based onvalues observed for the user at the same time of the day inthe past similar to that proposed by Banerjee et al [2]Thus this predictor has some degree of personalization butit learns user behavior in less detail than the Personalizedpredictor We see that its performance for 1-hour windowis similar to the Short-term predictor For 2-hour windowits error is higher than the Personalized predictor across theboard though its worst case error is lower than the Short-term and Generic predictors
Overall these results demonstrate the value of appropri-ately learning user behaviors to design intelligent mecha-nisms on smartphones
10 RELATED WORKIn a range of domains there is a rich history of work that
characterizes user workloads However because smartphoneadoption has gathered pace relatively recently our work rep-resents one of the few to study how people use smartphones
Along with other recent works it helps complete the pic-ture Banerjee et al and Rahmati et al report on batterycharging behaviors [2 18] Like us they find considerablevariation among users Banerjee et al also propose a pre-dictor that estimates the excess energy of the battery atthe time of charging using a histogram of past battery us-age [2] Shye et al study the power consumption character-istics of 20 users [24] They infer properties such as whichcomponents in the phone consume most power and exploreoptimizations based on these characteristics Rahmati andZhong study 14 users of a particular demographic to studywhich applications are popular in that demographic wherethe phone is used and how it is shared among users [19]
In contrast to these works we focus on understandingdifferent aspects of smartphone use (eg interactions andtraffic) and on exposing the diversity of user behaviors in-stead of only the average behavior Our study also entailsan order of magnitude more users than previous efforts
Like us other researchers have developed logging utilitiesMyExperience is one such early utility [10] and the worksabove involve custom utilities as well Our focus in thispaper is not on the development of logging tools but onanalyzing their data to gain insight into user behavior
There is a body of work in modeling the aggregate be-havior of mobile users Using traces from a large cellularoperator some network related aspects of mobile usage havebeen modeled Halepovic et al and Williamson et al re-port that call arrivals are bursty and present diurnal pat-terns [28] Willkomm et al and Brown et al report thatmobile users call duration can be approximately modeled bya lognormal distribution [4 29] We use traces collected onthe mobile device itself and focus on modeling the interac-tions of individual users instead of the aggregate behavior
11 CONCLUSIONSBy studying 255 users of two different smartphone plat-
forms we comprehensively characterized user activities andtheir impact on network and battery We quantify manyhitherto unknown aspects of smartphone usage User diver-sity is an overarching theme in our findings For instancedifferent users interact with their phones 10-200 times a dayon average the mean interaction length of different users
is 10-250 seconds and users receive 1-1000 MB of data perday where 10-90 is received as part of active use
This extent of user diversity implies that mechanisms thatwork for the average case may be ineffective for a large frac-tion of the users Instead learning and adapting to userbehaviors is likely to be more effective as demonstrated byour personalized energy drain predictor We show that qual-itative similarities exist among users to facilitate the devel-opment of such mechanisms For instance the longer theuser has not interacted with the phone the less likely it isfor her to start interacting with it and application popular-ity for a user follows a simple exponential distribution
Our study points to ways in which smartphone platformsshould be enhanced Effective adaptation will require futureplatforms to support light-weight tools that monitor andlearn user behaviors in situ It will also require them toexpose appropriate knobs to control the behavior of lower-level components (eg CPU or radio)
12 ACKNOWLEDGEMENTSWe are grateful to the users in Dataset1 that ran our log-
ger and to the providers of Dataset2 Elizabeth Dawson pro-vided administrative support for data collection Jean BolotKurt Partridge Stefan Saroiu Lin Zhong and the MobiSysreviewers provided many comments that helped improve thepaper We thank them all This work is supported in partby the NSF grants CCR-0120778 and CNS-0627084
13 REFERENCES[1] Balasubramanian N Balasubramanian A
and Venkataramani A Energy consumption inmobile phones A measurement study andimplications for network applications In IMC (2009)
[2] Banerjee N Rahmati A Corner M D
Rollins S and Zhong L Users and batteriesInteractions and adaptive energy management inmobile systems In Ubicomp (2007)
[3] Becker R Chambers J and Wilks A The newS language Chapman amp HallCRC 1988
[4] Brown L Gans N Mandelbaum A Sakov
A Shen H Zeltyn S and Zhao L Statisticalanalysis of a telephone call center Journal of theAmerican Statistical Association 100 469 (2005)
[5] Church K and Smyth B Understanding mobileinformation needs In MobileHCI (2008)
[6] Dempster A Laird N and Rubin D Maximumlikelihood from incomplete data via the EM algorithmJournal of the Royal Statistical Society Series B(Methodological) 39 1 (1977)
[7] Esfahbod B Preload An adaptive prefetchingdaemon Masterrsquos thesis University of Toronto 2006
[8] Falaki H Govindan R and Estrin D Smartscreen management on mobile phones Tech Rep 74Center for Embedded Networked Sesning 2009
[9] Flinn J and Satyanarayanan M Managingbattery lifetime with energy-aware adaptation ACMTOCS 22 2 (2004)
[10] Froehlich J Chen M Consolvo S
Harrison B and Landay J MyExperience Asystem for in situ tracing and capturing of userfeedback on mobile phones In MobiSys (2007)
[11] Fukuda K Cho K and Esaki H The impact ofresidential broadband traffic on Japanese ISPbackbones SIGCOMM CCR 35 1 (2005)
[12] Kotz D and Essien K Analysis of a campus-widewireless network Wireless Networks 11 1ndash2 (2005)
[13] Kwiatkowski D Phillips P Schmidt P and
Shin Y Testing the null hypothesis of stationarityagainst the alternative of a unit root Journal ofEconometrics 54 1-3 (1992)
[14] Lahiri K Dey S Panigrahi D and
Raghunathan A Battery-driven system design Anew frontier in low power design In Asia SouthPacific design automationVLSI Design (2002)
[15] Ljung G and Box G On a measure of lack of fitin time series models Biometrika 65 2 (1978)
[16] Smartphone futures 2009-2014 httpwwwportioresearchcomSmartphone09-14html
[17] Rahmati A Qian A and Zhong L
Understanding human-battery interaction on mobilephones In MobileHCI (2007)
[18] Rahmati A and Zhong L Humanndashbatteryinteraction on mobile phones Pervasive and MobileComputing 5 5 (2009)
[19] Rahmati A and Zhong L A longitudinal study ofnon-voice mobile phone usage by teens from anunderserved urban community Tech Rep 0515-09Rice University 2009
[20] Rao R Vrudhula S and Rakhmatov D
Battery modeling for energy aware system designIEEE Computer 36 12 (2003)
[21] Ravi N Scott J Han L and Iftode L
Context-aware battery management for mobilephones In PerCom (2008)
[22] Sharma A Navda V Ramjee R
Padmanabhan V and Belding E Cool-TetherEnergy efficient on-the-fly WiFi hot-spots usingmobile phones In CoNEXT (2009)
[23] Sharma C US wireless data market update - Q32009 httpwwwchetansharmacomusmarketupdateq309htm
[24] Shye A Sholbrock B and G M Into the wildStudying real user activity patterns to guide poweroptimization for mobile architectures In Micro (2009)
[25] Snol L More smartphones than desktop PCs by2011 httpwwwpcworldcomarticle171380more_smartphones_than_desktop_pcs_by_2011
html
[26] Sohn T Li K Griswold W and Hollan J
A diary study of mobile information needs In SIGCHI(2008)
[27] Trestian I Ranjan S Kuzmanovic A and
Nucci A Measuring serendipity Connecting peoplelocations and interests in a mobile 3G network InIMC (2009)
[28] Williamson C Halepovic E Sun H and Wu
Y Characterization of CDMA2000 Cellular DataNetwork Traffic In Local Computer Networks (2005)
[29] Willkomm D Machiraju S Bolot J and
Wolisz A Primary users in cellular networks Alarge-scale measurement study In DySPAN (2008)
- Introduction
- Data Collection
-
- Dataset1
- Dataset2
- Representativeness of conclusions
-
- User Interactions
-
- Interaction Time
- Interaction Sessions
- Diurnal Patterns
-
- Application Usage
-
- Number of applications
- Application Popularity
- Application Sessions
-
- Traffic
-
- Traffic per day
- ``Interactive traffic
- Diurnal patterns
-
- Energy Consumption
- Implications of User Diversity
- Smartphone Usage Models
-
- Session Lengths
- Time between Sessions
- Application Popularity
-
- Predicting Energy Drain
- Related Work
- Conclusions
- Acknowledgements
- References
-
comm 44
browsing 10
other 11
prod 19media 5
maps 5games 2system 5
Dataset1
comm 49
browsing 12 other 15
prod 2media 9
maps 2
games 10
system 1
Dataset2
Figure 13 Relative popularity of each applicationcategory across all users in each dataset
Aggregate view application popularity To pro-vide an aggregate view of what users use their smartphonesfor we categorize applications into eight distinct categoriesi) communication contains applications for exchanging mes-sages (eg email SMS IM) and voice calls ii) browsingcontains Web browser search and social networking appli-cations iii) media contains applications for consuming orcreating media content (eg pictures music videos) iv)productivity contains applications for calendars alarms andfor viewing and creating text documents (eg Office PDFreader) v) system contains applications for changing userpreferences and viewing system state (eg file explorer)vi) games vii) maps and viii) other contains applicationsthat we could not include in any of the categories aboveeg because we did not know their function
Figure 13 shows the mean relative popularity of each ap-plication category across all users in each dataset Whilethe results are not identical across the two datasets they aresimilar to a first order Communication dominates in bothBrowsing is another major contributor in both datasetsMaps media and games have a comparatively lower butnevertheless substantial share of user attention
Relationship to user demographic To understandthe extent to which user demographic determines applica-tion popularity Figure 14 shows the mean and 95 CI ofrelative popularity of each application category for differ-ent user demographics As for interaction time (sect31) wesee that the impact of user demographics in our datasets isminimal In Dataset2 the relative popularity of various ap-plication types is similar for each of the four demographicsIn Dataset1 there are noticeable differences in the mean forcommunication games and productivity applications Highschool students use communication and games applicationsmore while knowledge workers use productivity applications
Dataset1
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
High schoolKnowledge worker
Dataset2
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
SCOPBPULPU
Figure 14 The mean and 95 CI of relative pop-ularity of application categories among users of dif-ferent demographics
more However considering the overlap of confidence in-tervals these differences in application popularity are notstatistically significant
From this result and the earlier one on the lack of depen-dence between user demographic and interaction time (sect31)we conclude that user demographic at least as defined inour datasets cannot reliably predict how a user will use thephone While demographic information appears to help insome cases (for eg the variation in usage of productivitysoftware in Dataset1) such cases are not the norm and itis hard to guess when demographic information would beuseful Pending development of other ways to classify userssuch that these classes more predictably explain the varia-tion incorporating factors specific to a user appear necessaryThis insensitivity to user demographic has positive as wellas negative implications A negative is that personalizationis more complex we cannot predict a usersrsquo behavior byknowing their demographic A positive implication is thatthe range of user behaviors along many dimensions of in-terest can be found in several common demographics Thissimplifies the task of uncovering the range because recruitingsubjects across multiple demographics tends to be difficult
Relationship to interaction time We also study ifusers that interact more with their phones tend to use differ-ent applications For each dataset we sort users based ontheir average interaction time per day and partition theminto different classes For Dataset1 we use two classesone each for the top and the bottom half of the users ForDataset2 which has more users we use three classes for the
Dataset1
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
HighLow
Dataset2
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
rHighMediumLow
Figure 15 The mean and 95 CI of relative pop-ularity of application categories among differentclasses of users based on interaction time per day
top middle and bottom third of the users Figure 15 showsthe mean and 95 CI for relative time spent with each ap-plication category by each user class We see that users indifferent classes have similar application usage Thus wecannot explain why some users use the phone more simplybased on the applications that they use
43 Application SessionsWe now study the characteristics of application sessions
We conduct this analysis only for Dataset2 based on times-tamps for when an application is started and ended Dataset1does not contain this information precisely Because appli-cations can run in the background start and end refer tothe period when the application is in the foreground
Applications run per interaction We begin by study-ing the number of applications that users run in an interac-tion session Figure 16 shows that an overwhelming major-ity close to 90 of interactions include only one applicationThis graph aggregates data across all users We did not findstatistically significant differences between users A largefraction of sessions of all users have only one application
That interaction sessions very often have only one appli-cation session suggests that users tend to interact with theirsmartphone for one task (eg reading email checking cal-endar etc) at a time and most of these tasks require theuse of only one application
Application session lengths Because interactionsessions are dominated by those with only one application
892
7422 06 02 01
1 2 3 4 5 6
020
4060
8010
0
Apps per session
Per
cent
age
()
Figure 16 Histogram of the number of applica-tions called during each interaction session for allthe users in Dataset2
map
s
gam
es
com
m
med
ia
brow
sing
prod
syst
em
othe
r
110
1000
Ses
sion
leng
th (
s)
Figure 17 The mean and 95 CI of session lengthsof different application categories The y-axis is logscale
the overall properties of application sessions such as theirlengths are similar to those of interaction sessions (sect3)
However studying the session lengths of applications sep-arately reveals interesting insights Different applicationtypes have different session lengths as shown in Figure 17for the categories defined earlier Interactions with mapsand games tend to be the longest and those related to pro-ductivity and system tend to be the shortest
Further given an application different users run themfor different times Figure 18 shows this effect for a mes-saging application tmailexe and a browsing applicationiexploreexe For each application the mean session lengthsof users differ by more than two orders of magnitude
These observations help explain why users have differentsession lengths (sect32) They prefer different applicationsand those application sessions tend to have different lengthsFurther analysis confirms this phenomenon For instance ifwe divide users into two classes based on their mean sessionlengths the popularity of games is twice as high in the classwith high session lengths
5 TRAFFICIn this section we investigate traffic generated by smart-
phones Unlike interaction events and application use net-work traffic is not an intentional user action but a side-effectof those actions Most users are likely oblivious to how much
tmail1
1010
010
00
0 20 40 60 80 100User percentile
Ses
sion
leng
th (
s)iexplore
110
100
1000
0 20 40 60 80 100User percentile
Ses
sion
leng
th (
s)Figure 18 The mean and the upper end of the stan-dard deviation of session lengths of two applications
0 20 40 60 80 100
User percentile
0
1
10
100
1000
10000
Traffic per day (MB) Receive
Send
Figure 19 The mean and the upper end of the stan-dard deviation of the traffic sent and received perday by users in Dataset1
traffic they generate We show that the diversity and diur-nal patterns of this side-effect match those of user actionsthemselves
The analysis in this section includes only Dataset1 we donot have traffic information for Dataset2 In Dataset1 werecord all of the data sent (or received) by the phone exceptfor that exchanged over the USB link ie the traffic hereinincludes data over the 3G radio and the 80211 wireless link
51 Traffic per dayFigure 19 shows that the amount of traffic sent and re-
ceived per day differs across users by almost three orders ofmagnitude The traffic received ranges from 1 to 1000 MBand the traffic sent ranges from 03 to 100 MB The medianvalues are 30 MB sent and 5 MB received
Our results indicate that traffic generated in a day bysmartphone users is comparable to traffic generated by PCsa few years ago A study of a campus WiFi network revealedthat on average users were generating 27 MB of traffic in2001 and 71 MB in 2003 [12] A study of Japanese residen-tial broadband users in 2006 revealed that on average usersgenerate 1000 MB of traffic per day [11] This high level oftraffic has major implications for the provisioning of wirelesscarrier networks as smartphone adoption increases
Relationship to application types To investigate ifcertain types of applications are favored more by users thatgenerate more traffice we divide the users into two equalclasses based on their sum of sent and received traffic perday Figure 20 shows mean and 95 CI of relative popularityof each application category for each user class Expectedly
Dataset1
Per
cent
age
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
HighLow
Figure 20 The mean and 95 CI of relative popu-larity of each application category among high andlow traffic consumers
it shows that communication applications are more popularamong users that consume more traffic
52 ldquoInteractiverdquo trafficWe next estimate what fraction of the total traffic is ldquoin-
teractiverdquo that is has timeliness constraints Approachesto reduce power consumption of data transfers often advo-cate rescheduling network transfers for instance by delay-ing some transfers so that multiple of them may be bun-dled [1 22] Such policies are difficult to implement forinteractive traffic without hurting user response time andthus are likely to be less valuable if the bulk of the traffic isinteractive
We classify traffic as interactive if it was generated whenthe screen is on This classification method might classifysome background traffic as interactive We expect the errorto be low because the screen is on for a small fraction ofthe time for most users Because certain user interactionswith the phone begin right after traffic exchange (eg anew email is received) we also consider traffic received in asmall time window (1 minute) before the screen is turned onas having timeliness constraints The results are robust tothe exact choice of time window Indeed some of the trafficthat we classify as interactive might be delay tolerant ega new email to be sent might be deferred for a little whileHowever the volume of traffic received by the phone whichusers would rather see immediately dominates by one or-der of magnitude the volume that is sent and delay-tolerantmessaging applications such as email contribute only a smallchunk of all traffic
Figure 21 shows the fraction of interactive traffic for eachuser We see that for about 90 of the users over 50of the traffic is interactive but for the rest almost none oftheir traffic is interactive Stated differently for differentsmartphone users almost all to almost none of the traffic isgenerated by applications in the background The extremesrepresent disparate ways in which people use smartphonesand which applications on the smartphone generate mosttraffic Our results imply that the energy savings that can behad by rescheduling network activity will vary across users
53 Diurnal patternsFigure 22(a) shows the diurnal pattern for an example
user with a sharp decline at night Figure 22(b) showsthe strength of the diurnal pattern for individual users by
0 20 40 60 80 100
User percentile
00
02
04
06
08
10
Interactive fraction
Figure 21 The fraction of interactive traffic
0 5 10 15 20
Hour of day
0
5
10
15
Traffic per hour (MB)
Receive
Send
(a)
0 20 40 60 80 100
User percentile
2
4
6
8
10
12
Diurnal ratio
Receive
Send
(b)
Figure 22 (a) The mean and 95 CI for trafficgenerated per hour by an example user (b) Thediurnal ratio of traffic per hour for all users
plotting the diurnal ratio of traffic As defined in sect33 thediurnal ratio reflects how much higher the mean during thepeak hour is to the overall mean
We see that the diurnal ratio varies across users but mosthave a strong diurnal behavior 80 of them generate overtwice their average amount of traffic in their peak hour Thisbehavior is likely a direct result of the high proportion ofinteractive traffic for most users and that user interactionsthemselves have a diurnal pattern
6 ENERGY CONSUMPTIONThe final aspect of smartphone usage that we investigate
is energy consumption Energy drain depends on two fac-tors i) user interactions and applications and ii) platformhardware and software If the second factor dominates theenergy drain of various users with the identical smartphonewill be similar Otherwise the energy drain will be as di-verse as user behaviors
We estimate the amount of energy drain based on the re-maining battery indicator which varies between 0 and 100If the battery indicator has gone down by X in a time pe-riod for a battery with capacity Y mAh we compute theenergy drain in that period to be X middot Y mAh 1 Given thatbatteries are complex electro-chemical devices [14 20] thiscomputation is approximate It assumes that the batterylevel indicator is linear with respect to energy drain
1mAh is technically a unit of charge yet is commonly used toindicate energy drain because battery voltage during normaloperations is typically constant For phones in our datasetthis is 4V so multiplying a mAh reading by 4 would yieldan accurate energy reading in milli-watt-hours
020
4060
8010
0
0 4 8 12 16 20Time (hour)
Bat
tery
leve
l (
) Benchmark1Benchmark2
Figure 23 Timelapse of the remaining battery levelindicator in controlled experiments with two differ-ent workloads at room temperature
110
100
1000
0 20 40 60 80 100User percentile
Bat
tery
Dra
in (
mA
h)
Figure 24 The mean and the upper end of the stan-dard deviation of one hour energy drain for Dataset1users during discharge periods
Controlled experiments suggest that the linearity assump-tion holds to a first order We run a benchmark load thatdrains the battery at a fixed rate in room temperature Un-der this benchmark if the battery level indicator decreaseslinearly with time it must be linear with respect to energydrain Figure 23 shows that the level indicator decreasesroughly linearly for two different benchmarks Benchmark1turns the screen on and off periodically Benchmark2 com-putes and idles periodically We conclude thus that the levelindicator can be used to estimate energy drain
Figure 24 shows the mean and standard deviation of en-ergy that users drain in an hour This graph is computed us-ing only periods in which the battery is not charging becauseenergy drain in those periods are of primary interest Wesee a two orders of magnitude difference among users Whileheaviest users drain close to 250 mAh the lightest of usersdrain only 10 mAh If the battery capacity is 1200 mAhthis leads to a lifetime variation from about 4 to 120 hours
Figure 25(a) shows for an example user that the drain isnot the same throughout the day but has diurnal variationsin which more energy is consumed during the day than dur-ing the night For this user the level of energy consumedchanges by roughly a factor of five Figure 25(b) plots thediurnal ratio of energy use for all users It shows that diurnalvariations occur with different strengths for all users
2060
100
140
0 4 8 12 16 20Hour of the day
Bat
tery
Dra
in (
mA
h)
(a)
0 20 40 60 80 100
12
34
56
7
User percentile
Diu
rnal
rat
io(b)
Figure 25 (a) The mean and 95 CI of energy drainof an example Dataset1 user (b) Diurnal ratio ofall users in Dataset1
Our results show that user activities contribute heavilytowards energy drain users in Dataset1 who have identi-cal smartphones drain energy at different rates and energydrain has diurnal patterns In the future we will developmethods to accurately quantify the energy consumption ofthe platform from that due to user-induced workload
7 IMPLICATIONS OF USER DIVERSITYWe uncover a surprising level of diversity among smart-
phone users For almost every aspect of usage that we studywe find one or more orders of magnitude difference betweenusers Our findings strongly motivate the need for customiz-ing smartphones to their users We believe that this needis greater than that for customizing ordinary cellphones orlaptops Ordinary cellphones do not have as rich an appli-cation environment Laptops are not as portable and aremore resource rich For example many users plug-in theirlaptops while using them
Customization can help at all levels Consider somethingas low-level as the battery Suppose we want batteries to beboth lightweight and last for at least a day with a high proba-bility Meeting the latter goal for all users of a given platformwill require catering to the heaviest users But that will leadto unnecessarily heavy batteries for many users (Highercapacity batteries are heavier) Offering multiple types ofbatteries with different lifetime-weight tradeoffs provides away out of this bind
At levels where intelligent mechanisms to improve user ex-perience or reduce energy consumption reside user diversitymotivates adapting to the smartphone user Driving thesemechanisms based on average case behaviors may not beeffective for a large fraction of the users
The ease and utility of customization depends on twoproperties of user behavior First despite quantitative dif-ferences there must be qualitative similarities among usersFor instance we should be able to describe the behavior ofall users with the same model Different users may havedifferent parameters of this model which will then lead toquantitative differences among them The presence of qual-itative similarities imply that users are not arbitrary pointsin space and it significantly simplifies the task of learninguser behavior Second user behavior in the past must alsobe predictive of the future Otherwise customization basedon past behaviors will be of little value in the future
In the next two sections we present evidence that theseproperties hold for several key aspects of smartphone usage
Dataset1 example user
000
00
010
002
00
030
0 100 200 300 400Session length (s)
Den
sity
Dataset2 example user
000
005
010
015
020
0 20 40 60 80 100Session length (s)
Den
sity
Figure 26 The PDF of session length for sampleusers of each dataset
In sect8 we show that user sessions and relative applicationpopularity can be described using simple models In sect9 weshow that energy drain can be described as well as predictedusing a ldquotrend trablerdquo framework
8 SMARTPHONE USAGE MODELSIn this section we develop simple models that describe
three aspects of smartphone usage ndash session lengths inter-arrival time between sessions and application popularityThe models are common across users but have different pa-rameters for different users While they do not completelydescribe user behavior they capture first order factors andrepresent a first step towards more complete modeling ofsmartphone usage More importantly along with the resultsof the next section they show that qualitative similaritiesdo exist among users
81 Session LengthsWe first consider the statistical properties of session length
distributions of users We find that session length valuestend to stationary With the KPSS test for level stationar-ity [13] 90 of the users have a p-value greater than 01The presence of stationarity is appealing because it suggeststhat past behavior is capable of predicting the future
We also find that session lengths are independent that isthe current value does not have a strong correlation with thevalues seen in the recent past With the Ljung-Box test forindependence [15] 96 of the users have a p-value that isgreater than 01
Stationarity and independence considered together sug-gest that session length values can be modeled as iid sam-ples from a distribution Choosing an appropriate distribu-tion however is complicated by the nature of the sessionlengths Most sessions are very short and the frequencydrops exponentially as the length increases However in-consistent with exponential behavior there are some verylong sessions in the tail for each user
We find that a mixture of exponential and Pareto distri-butions can model both ends of the spectrum The formercaptures short sessions and the latter captures long sessionsThat is session lengths can be described by the followingmixture model
r middot Exp(λ) + (1 minus r) middot Pareto(xm α)
In this equation r is the relative mix of the two distribu-tions λ is the rate of the exponential and xm and α are thelocation and shape parameters of the Pareto distribution
The location for a Pareto distribution represents the min-imum possible value of the random variable The location
010
0020
0030
00
0 1000 2000 3000Observed quantiles
Mod
el q
uant
iles
Figure 27 QQ plot of session lengths model for asample user
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed m
ix
(a) Relative mix (r)
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed r
ate
(b) Exponential rate (λ)
020
4060
0 20 40 60 80 100User percentile
Tim
eout
(s)
(c) Pareto location (xm)
minus1
0minus
04
02
06
10
0 20 40 60 80 100User percentile
Est
imat
ed s
hape
(d) Pareto shape (α)
Figure 28 Distribution of inferred model parame-ters that describe session length values of users inboth datasets
value that offers the best fit is the screen timeout value ofthe user because the session length PDF has a spike at thisvalue The spike corresponds to short sessions that are endedby the timeout (when the user forgets to switch the screenoff) we confirm this using controlled experiments with dif-ferent timeout values Figure 26 shows this spike at 60 and15 seconds for example users from each dataset The time-out provides a natural division between the two componentdistributions We automatically infer its approximate valueusing a simple spike detection algorithm
We use the EM algorithm to infer the maximum likelihoodestimation (MLE) of the remaining three parameters [6]Figure 27 shows the quality of this fit for an example userusing the QQ plot [3] Almost all quantiles are along they = x line indicating a good fit
Figure 28 shows the four inferred parameters for varioususers While users can be modeled using the same mix-
040
0080
00
0 4000 8000Observed quantiles
Mod
el q
uant
iles
Figure 29 QQ plot of session offtime model for asample user
050
010
0015
000 20 40 60 80 100
User percentileE
stim
ated
sca
le
(a) Weibull scale
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed s
hape
(b) Weibull shape
Figure 30 Distribution of inferred model parame-ters that describe the distribution of time betweensessions for users in both datasets
ture model the parameters of this model vary widely acrossusers Because of the way we construct our model the dis-tribution of the parameter r and xm also provide insightinto how frequently usersrsquo screen is switched off by the time-out and the relative popularity of different timeout values60 seconds is the most popular value likely because it is themost common default timeout and many users never changethe default
82 Time between SessionsWe find that the Weibull distribution can explain the
screen off times This distribution has two parameters re-ferred to as its scale and shape We find the MLE for theseparameters for each user From the QQ-plot in Figure 29we notice that the model predicts a greater probability ofseeing some very large offtimes than are observed in thedatasets However the probability of seeing these large off-times is small there are 27 data points that have a y-value greater than 8000 in that graph Hence we believethat Weibull provides a good fit for the length of intervalsbetween interactions
Figure 30 shows the distribution of the estimated shapeand scale of the fitted Weibull distributions Interestinglythe shape is consistently less than one Weibull shape valuesless than one indicate that the longer the screen has been offthe less likely it is for it to be turned on by the user This be-havior has interesting implications for power saving policiesFor instance periodic activities such as checking for emailwhen the screen has been off for a while may be deferred orrescheduled if needed without hurting user experience
020
4060
8010
0
0 20 40 60 80 100User percentile
MS
E (
)
(a) MSE
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed r
ate
(b) Rate parameter
Figure 31 (a) The mean square error (MSE) whenapplication popularity distribution is modeled usingan exponential (b) The inferred rate parameter ofthe exponential distribution for different users
83 Application PopularityWe find that for each user the relative popularity of ap-
plications can be well described by a simple exponential dis-tribution This qualitative invariant is useful for instanceto predict how many applications account for a given frac-tion of user attention For the example users in Figure 11this facet can be seen in the inset plots the semi-log of thepopularity distribution is very close to a straight line
Figure 31(a) shows that this exponential drop in applica-tion popularity is true for almost all users the mean squareerror between modeled exponential and actual popularitydistribution is less than 5 for 95 of the users
Figure 31(b) shows the inferred rate parameter of the ap-plication popularity distribution for various users We seethat the rate varies by an order of magnitude from 01 toalmost 1 The value of the rate essentially captures the paceof the drop in application popularity Lower values describeusers that use more applications on a regular basis Oneimplication of the wide range is that it may be feasible toretain all popular applications in memory for some users andnot for others
9 PREDICTING ENERGY DRAINIn this section we demonstrate the value of adapting to
user behaviors in the context of a mechanism to predict fu-ture energy drain on the smartphone Such a mechanismcan help with scheduling background tasks [9] estimatingbattery lifetime and improving user experience if it is esti-mated that the user will not fully drain the battery until thenext charging cycle [2 21]
Despite its many uses to our knowledge none of thesmartphones today provide a prediction of future energydrain Providing an accurate prediction of energy drain isdifficult2 User diversity of energy use makes any static pre-diction method highly inaccurate Even for the same userthere is a high variance in energy usage Figure 32 shows thisvariance by plotting the ratio of the standard deviation tothe mean energy drain over several time periods for users inDataset 1 The standard deviation of 10-minute windows ishigher than three times the mean for a fifth of the users The
2Energy drain prediction on smartphones is more challeng-ing than what is done for laptops Laptops provide a pre-diction of battery lifetime only for active use Smartphonepredictions on the other hand must cover multiple activeidle periods to be useful
01
23
45
0 20 40 60 80 100User percentile
Err
or r
atio
2 hours1 hour10 minutes
Figure 32 Battery drain predictors have to copewith the high variance in drain
time
window to
predict
n chunks
hellip
Trend Table
hellip
hellip
Figure 33 A userrsquos trend table captures how batterydrain in n consecutive chunks relates to drain in thesubsequent window
bursty nature of user interactions is one cause of this highvariance The variance is high even at longer time scalesThe standard deviation for two hour windows is larger than150 of the mean for half the users The diurnal patternsthat we showed earlier are one cause of high variance oversuch large time scales
Although the variance is high there are patterns in auserrsquos behaviors For instance we showed earlier that aphone that has been idle for a while is likely to remain idlein the near future and it will thus continue to draw energyat a similar rate
We hypothesize that a predictor tuned to user behaviorcan be accurate We present a simple personalized energydrain predictor that incorporates various user-specific fac-tors such as length of idle periods and different types ofbusy periods Instead of explicitly identifying these factorsour predictor captures them in a ldquotrend tablerdquo framework
Our framework is shown in Figure 33 Each entry in thetrend table is indexed by an n-tuple that represents energyusage readings from adjacent chunks of size δ each Thisindex points to energy usage statistics across all time win-dows of size w in the history that follow the n chunk valuesrepresented by that index Thus the trend table captureshow the energy usage in n consecutive chunks is related tothe energy use in the subsequent time windows of size w
In this paper we use n = 3 δ = 10 minutes and maintaindifferent trend tables for each window size w for which theprediction is required To keep the trend table small wequantize the energy readings of chunks We maintain themean and standard deviation for all subsequent windowsthat map to that index
To make a prediction we use the energy readings fromthe n immediately preceding chunks Let these readings be(x1 xn) Then we search for indices in the table thatdiffer from (x1 xn) by no more than a threshold s and
GenericShortminustermTimeminusofminusdayPersonalized
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
(a) 1-hour window
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
GenericShortminustermTimeminusofminusdayPersonalized
(b) 2-hour window
Figure 34 The ratio of error to true battery drainfor different energy drain predictors
predict based on them such that similar indices have higherrelative weight More precisely the prediction is
X
e stforalli=1n|he
iminusxi|lts
ke times ue
P
ke
where e iterates over all tuples in the trend table hei is
the value of the chunk i of the tuple e ue represents thestatistics stored for that index and the relative weight ke =s minus maxi=1n |he
i minus xi| For results shown here we uses = 05
Figures 34 shows how well our Personalized predictor worksfor predicting usage for two different future time windowsThere is one point per user in each graph which correspondsto the median error seen across many predictions for thatuser We see that the 90th percentile error is 25 for 1-hourwindow and 40 for 2-hour window
To place this level of accuracy in context the figure alsoshows the accurary of three alternative predictors TheGeneric predictor uses a trend table but instead of buildinga user-specific table it builds one that combines all usersWe see that its 90th percentile error is roughly twice thatof the Personalized predictor Its predictions are worse forhalf of the users
The Short-term predictor predicts energy drain simply asthe amount drained in the immediately preceding windowof the same size For both time windows its 90th percentileerror is no better than that of the Generic predictor
The Time-of-day predictor predicts energy drain based onvalues observed for the user at the same time of the day inthe past similar to that proposed by Banerjee et al [2]Thus this predictor has some degree of personalization butit learns user behavior in less detail than the Personalizedpredictor We see that its performance for 1-hour windowis similar to the Short-term predictor For 2-hour windowits error is higher than the Personalized predictor across theboard though its worst case error is lower than the Short-term and Generic predictors
Overall these results demonstrate the value of appropri-ately learning user behaviors to design intelligent mecha-nisms on smartphones
10 RELATED WORKIn a range of domains there is a rich history of work that
characterizes user workloads However because smartphoneadoption has gathered pace relatively recently our work rep-resents one of the few to study how people use smartphones
Along with other recent works it helps complete the pic-ture Banerjee et al and Rahmati et al report on batterycharging behaviors [2 18] Like us they find considerablevariation among users Banerjee et al also propose a pre-dictor that estimates the excess energy of the battery atthe time of charging using a histogram of past battery us-age [2] Shye et al study the power consumption character-istics of 20 users [24] They infer properties such as whichcomponents in the phone consume most power and exploreoptimizations based on these characteristics Rahmati andZhong study 14 users of a particular demographic to studywhich applications are popular in that demographic wherethe phone is used and how it is shared among users [19]
In contrast to these works we focus on understandingdifferent aspects of smartphone use (eg interactions andtraffic) and on exposing the diversity of user behaviors in-stead of only the average behavior Our study also entailsan order of magnitude more users than previous efforts
Like us other researchers have developed logging utilitiesMyExperience is one such early utility [10] and the worksabove involve custom utilities as well Our focus in thispaper is not on the development of logging tools but onanalyzing their data to gain insight into user behavior
There is a body of work in modeling the aggregate be-havior of mobile users Using traces from a large cellularoperator some network related aspects of mobile usage havebeen modeled Halepovic et al and Williamson et al re-port that call arrivals are bursty and present diurnal pat-terns [28] Willkomm et al and Brown et al report thatmobile users call duration can be approximately modeled bya lognormal distribution [4 29] We use traces collected onthe mobile device itself and focus on modeling the interac-tions of individual users instead of the aggregate behavior
11 CONCLUSIONSBy studying 255 users of two different smartphone plat-
forms we comprehensively characterized user activities andtheir impact on network and battery We quantify manyhitherto unknown aspects of smartphone usage User diver-sity is an overarching theme in our findings For instancedifferent users interact with their phones 10-200 times a dayon average the mean interaction length of different users
is 10-250 seconds and users receive 1-1000 MB of data perday where 10-90 is received as part of active use
This extent of user diversity implies that mechanisms thatwork for the average case may be ineffective for a large frac-tion of the users Instead learning and adapting to userbehaviors is likely to be more effective as demonstrated byour personalized energy drain predictor We show that qual-itative similarities exist among users to facilitate the devel-opment of such mechanisms For instance the longer theuser has not interacted with the phone the less likely it isfor her to start interacting with it and application popular-ity for a user follows a simple exponential distribution
Our study points to ways in which smartphone platformsshould be enhanced Effective adaptation will require futureplatforms to support light-weight tools that monitor andlearn user behaviors in situ It will also require them toexpose appropriate knobs to control the behavior of lower-level components (eg CPU or radio)
12 ACKNOWLEDGEMENTSWe are grateful to the users in Dataset1 that ran our log-
ger and to the providers of Dataset2 Elizabeth Dawson pro-vided administrative support for data collection Jean BolotKurt Partridge Stefan Saroiu Lin Zhong and the MobiSysreviewers provided many comments that helped improve thepaper We thank them all This work is supported in partby the NSF grants CCR-0120778 and CNS-0627084
13 REFERENCES[1] Balasubramanian N Balasubramanian A
and Venkataramani A Energy consumption inmobile phones A measurement study andimplications for network applications In IMC (2009)
[2] Banerjee N Rahmati A Corner M D
Rollins S and Zhong L Users and batteriesInteractions and adaptive energy management inmobile systems In Ubicomp (2007)
[3] Becker R Chambers J and Wilks A The newS language Chapman amp HallCRC 1988
[4] Brown L Gans N Mandelbaum A Sakov
A Shen H Zeltyn S and Zhao L Statisticalanalysis of a telephone call center Journal of theAmerican Statistical Association 100 469 (2005)
[5] Church K and Smyth B Understanding mobileinformation needs In MobileHCI (2008)
[6] Dempster A Laird N and Rubin D Maximumlikelihood from incomplete data via the EM algorithmJournal of the Royal Statistical Society Series B(Methodological) 39 1 (1977)
[7] Esfahbod B Preload An adaptive prefetchingdaemon Masterrsquos thesis University of Toronto 2006
[8] Falaki H Govindan R and Estrin D Smartscreen management on mobile phones Tech Rep 74Center for Embedded Networked Sesning 2009
[9] Flinn J and Satyanarayanan M Managingbattery lifetime with energy-aware adaptation ACMTOCS 22 2 (2004)
[10] Froehlich J Chen M Consolvo S
Harrison B and Landay J MyExperience Asystem for in situ tracing and capturing of userfeedback on mobile phones In MobiSys (2007)
[11] Fukuda K Cho K and Esaki H The impact ofresidential broadband traffic on Japanese ISPbackbones SIGCOMM CCR 35 1 (2005)
[12] Kotz D and Essien K Analysis of a campus-widewireless network Wireless Networks 11 1ndash2 (2005)
[13] Kwiatkowski D Phillips P Schmidt P and
Shin Y Testing the null hypothesis of stationarityagainst the alternative of a unit root Journal ofEconometrics 54 1-3 (1992)
[14] Lahiri K Dey S Panigrahi D and
Raghunathan A Battery-driven system design Anew frontier in low power design In Asia SouthPacific design automationVLSI Design (2002)
[15] Ljung G and Box G On a measure of lack of fitin time series models Biometrika 65 2 (1978)
[16] Smartphone futures 2009-2014 httpwwwportioresearchcomSmartphone09-14html
[17] Rahmati A Qian A and Zhong L
Understanding human-battery interaction on mobilephones In MobileHCI (2007)
[18] Rahmati A and Zhong L Humanndashbatteryinteraction on mobile phones Pervasive and MobileComputing 5 5 (2009)
[19] Rahmati A and Zhong L A longitudinal study ofnon-voice mobile phone usage by teens from anunderserved urban community Tech Rep 0515-09Rice University 2009
[20] Rao R Vrudhula S and Rakhmatov D
Battery modeling for energy aware system designIEEE Computer 36 12 (2003)
[21] Ravi N Scott J Han L and Iftode L
Context-aware battery management for mobilephones In PerCom (2008)
[22] Sharma A Navda V Ramjee R
Padmanabhan V and Belding E Cool-TetherEnergy efficient on-the-fly WiFi hot-spots usingmobile phones In CoNEXT (2009)
[23] Sharma C US wireless data market update - Q32009 httpwwwchetansharmacomusmarketupdateq309htm
[24] Shye A Sholbrock B and G M Into the wildStudying real user activity patterns to guide poweroptimization for mobile architectures In Micro (2009)
[25] Snol L More smartphones than desktop PCs by2011 httpwwwpcworldcomarticle171380more_smartphones_than_desktop_pcs_by_2011
html
[26] Sohn T Li K Griswold W and Hollan J
A diary study of mobile information needs In SIGCHI(2008)
[27] Trestian I Ranjan S Kuzmanovic A and
Nucci A Measuring serendipity Connecting peoplelocations and interests in a mobile 3G network InIMC (2009)
[28] Williamson C Halepovic E Sun H and Wu
Y Characterization of CDMA2000 Cellular DataNetwork Traffic In Local Computer Networks (2005)
[29] Willkomm D Machiraju S Bolot J and
Wolisz A Primary users in cellular networks Alarge-scale measurement study In DySPAN (2008)
- Introduction
- Data Collection
-
- Dataset1
- Dataset2
- Representativeness of conclusions
-
- User Interactions
-
- Interaction Time
- Interaction Sessions
- Diurnal Patterns
-
- Application Usage
-
- Number of applications
- Application Popularity
- Application Sessions
-
- Traffic
-
- Traffic per day
- ``Interactive traffic
- Diurnal patterns
-
- Energy Consumption
- Implications of User Diversity
- Smartphone Usage Models
-
- Session Lengths
- Time between Sessions
- Application Popularity
-
- Predicting Energy Drain
- Related Work
- Conclusions
- Acknowledgements
- References
-
Dataset1
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
HighLow
Dataset2
Per
cent
age
()
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
rHighMediumLow
Figure 15 The mean and 95 CI of relative pop-ularity of application categories among differentclasses of users based on interaction time per day
top middle and bottom third of the users Figure 15 showsthe mean and 95 CI for relative time spent with each ap-plication category by each user class We see that users indifferent classes have similar application usage Thus wecannot explain why some users use the phone more simplybased on the applications that they use
43 Application SessionsWe now study the characteristics of application sessions
We conduct this analysis only for Dataset2 based on times-tamps for when an application is started and ended Dataset1does not contain this information precisely Because appli-cations can run in the background start and end refer tothe period when the application is in the foreground
Applications run per interaction We begin by study-ing the number of applications that users run in an interac-tion session Figure 16 shows that an overwhelming major-ity close to 90 of interactions include only one applicationThis graph aggregates data across all users We did not findstatistically significant differences between users A largefraction of sessions of all users have only one application
That interaction sessions very often have only one appli-cation session suggests that users tend to interact with theirsmartphone for one task (eg reading email checking cal-endar etc) at a time and most of these tasks require theuse of only one application
Application session lengths Because interactionsessions are dominated by those with only one application
892
7422 06 02 01
1 2 3 4 5 6
020
4060
8010
0
Apps per session
Per
cent
age
()
Figure 16 Histogram of the number of applica-tions called during each interaction session for allthe users in Dataset2
map
s
gam
es
com
m
med
ia
brow
sing
prod
syst
em
othe
r
110
1000
Ses
sion
leng
th (
s)
Figure 17 The mean and 95 CI of session lengthsof different application categories The y-axis is logscale
the overall properties of application sessions such as theirlengths are similar to those of interaction sessions (sect3)
However studying the session lengths of applications sep-arately reveals interesting insights Different applicationtypes have different session lengths as shown in Figure 17for the categories defined earlier Interactions with mapsand games tend to be the longest and those related to pro-ductivity and system tend to be the shortest
Further given an application different users run themfor different times Figure 18 shows this effect for a mes-saging application tmailexe and a browsing applicationiexploreexe For each application the mean session lengthsof users differ by more than two orders of magnitude
These observations help explain why users have differentsession lengths (sect32) They prefer different applicationsand those application sessions tend to have different lengthsFurther analysis confirms this phenomenon For instance ifwe divide users into two classes based on their mean sessionlengths the popularity of games is twice as high in the classwith high session lengths
5 TRAFFICIn this section we investigate traffic generated by smart-
phones Unlike interaction events and application use net-work traffic is not an intentional user action but a side-effectof those actions Most users are likely oblivious to how much
tmail1
1010
010
00
0 20 40 60 80 100User percentile
Ses
sion
leng
th (
s)iexplore
110
100
1000
0 20 40 60 80 100User percentile
Ses
sion
leng
th (
s)Figure 18 The mean and the upper end of the stan-dard deviation of session lengths of two applications
0 20 40 60 80 100
User percentile
0
1
10
100
1000
10000
Traffic per day (MB) Receive
Send
Figure 19 The mean and the upper end of the stan-dard deviation of the traffic sent and received perday by users in Dataset1
traffic they generate We show that the diversity and diur-nal patterns of this side-effect match those of user actionsthemselves
The analysis in this section includes only Dataset1 we donot have traffic information for Dataset2 In Dataset1 werecord all of the data sent (or received) by the phone exceptfor that exchanged over the USB link ie the traffic hereinincludes data over the 3G radio and the 80211 wireless link
51 Traffic per dayFigure 19 shows that the amount of traffic sent and re-
ceived per day differs across users by almost three orders ofmagnitude The traffic received ranges from 1 to 1000 MBand the traffic sent ranges from 03 to 100 MB The medianvalues are 30 MB sent and 5 MB received
Our results indicate that traffic generated in a day bysmartphone users is comparable to traffic generated by PCsa few years ago A study of a campus WiFi network revealedthat on average users were generating 27 MB of traffic in2001 and 71 MB in 2003 [12] A study of Japanese residen-tial broadband users in 2006 revealed that on average usersgenerate 1000 MB of traffic per day [11] This high level oftraffic has major implications for the provisioning of wirelesscarrier networks as smartphone adoption increases
Relationship to application types To investigate ifcertain types of applications are favored more by users thatgenerate more traffice we divide the users into two equalclasses based on their sum of sent and received traffic perday Figure 20 shows mean and 95 CI of relative popularityof each application category for each user class Expectedly
Dataset1
Per
cent
age
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
HighLow
Figure 20 The mean and 95 CI of relative popu-larity of each application category among high andlow traffic consumers
it shows that communication applications are more popularamong users that consume more traffic
52 ldquoInteractiverdquo trafficWe next estimate what fraction of the total traffic is ldquoin-
teractiverdquo that is has timeliness constraints Approachesto reduce power consumption of data transfers often advo-cate rescheduling network transfers for instance by delay-ing some transfers so that multiple of them may be bun-dled [1 22] Such policies are difficult to implement forinteractive traffic without hurting user response time andthus are likely to be less valuable if the bulk of the traffic isinteractive
We classify traffic as interactive if it was generated whenthe screen is on This classification method might classifysome background traffic as interactive We expect the errorto be low because the screen is on for a small fraction ofthe time for most users Because certain user interactionswith the phone begin right after traffic exchange (eg anew email is received) we also consider traffic received in asmall time window (1 minute) before the screen is turned onas having timeliness constraints The results are robust tothe exact choice of time window Indeed some of the trafficthat we classify as interactive might be delay tolerant ega new email to be sent might be deferred for a little whileHowever the volume of traffic received by the phone whichusers would rather see immediately dominates by one or-der of magnitude the volume that is sent and delay-tolerantmessaging applications such as email contribute only a smallchunk of all traffic
Figure 21 shows the fraction of interactive traffic for eachuser We see that for about 90 of the users over 50of the traffic is interactive but for the rest almost none oftheir traffic is interactive Stated differently for differentsmartphone users almost all to almost none of the traffic isgenerated by applications in the background The extremesrepresent disparate ways in which people use smartphonesand which applications on the smartphone generate mosttraffic Our results imply that the energy savings that can behad by rescheduling network activity will vary across users
53 Diurnal patternsFigure 22(a) shows the diurnal pattern for an example
user with a sharp decline at night Figure 22(b) showsthe strength of the diurnal pattern for individual users by
0 20 40 60 80 100
User percentile
00
02
04
06
08
10
Interactive fraction
Figure 21 The fraction of interactive traffic
0 5 10 15 20
Hour of day
0
5
10
15
Traffic per hour (MB)
Receive
Send
(a)
0 20 40 60 80 100
User percentile
2
4
6
8
10
12
Diurnal ratio
Receive
Send
(b)
Figure 22 (a) The mean and 95 CI for trafficgenerated per hour by an example user (b) Thediurnal ratio of traffic per hour for all users
plotting the diurnal ratio of traffic As defined in sect33 thediurnal ratio reflects how much higher the mean during thepeak hour is to the overall mean
We see that the diurnal ratio varies across users but mosthave a strong diurnal behavior 80 of them generate overtwice their average amount of traffic in their peak hour Thisbehavior is likely a direct result of the high proportion ofinteractive traffic for most users and that user interactionsthemselves have a diurnal pattern
6 ENERGY CONSUMPTIONThe final aspect of smartphone usage that we investigate
is energy consumption Energy drain depends on two fac-tors i) user interactions and applications and ii) platformhardware and software If the second factor dominates theenergy drain of various users with the identical smartphonewill be similar Otherwise the energy drain will be as di-verse as user behaviors
We estimate the amount of energy drain based on the re-maining battery indicator which varies between 0 and 100If the battery indicator has gone down by X in a time pe-riod for a battery with capacity Y mAh we compute theenergy drain in that period to be X middot Y mAh 1 Given thatbatteries are complex electro-chemical devices [14 20] thiscomputation is approximate It assumes that the batterylevel indicator is linear with respect to energy drain
1mAh is technically a unit of charge yet is commonly used toindicate energy drain because battery voltage during normaloperations is typically constant For phones in our datasetthis is 4V so multiplying a mAh reading by 4 would yieldan accurate energy reading in milli-watt-hours
020
4060
8010
0
0 4 8 12 16 20Time (hour)
Bat
tery
leve
l (
) Benchmark1Benchmark2
Figure 23 Timelapse of the remaining battery levelindicator in controlled experiments with two differ-ent workloads at room temperature
110
100
1000
0 20 40 60 80 100User percentile
Bat
tery
Dra
in (
mA
h)
Figure 24 The mean and the upper end of the stan-dard deviation of one hour energy drain for Dataset1users during discharge periods
Controlled experiments suggest that the linearity assump-tion holds to a first order We run a benchmark load thatdrains the battery at a fixed rate in room temperature Un-der this benchmark if the battery level indicator decreaseslinearly with time it must be linear with respect to energydrain Figure 23 shows that the level indicator decreasesroughly linearly for two different benchmarks Benchmark1turns the screen on and off periodically Benchmark2 com-putes and idles periodically We conclude thus that the levelindicator can be used to estimate energy drain
Figure 24 shows the mean and standard deviation of en-ergy that users drain in an hour This graph is computed us-ing only periods in which the battery is not charging becauseenergy drain in those periods are of primary interest Wesee a two orders of magnitude difference among users Whileheaviest users drain close to 250 mAh the lightest of usersdrain only 10 mAh If the battery capacity is 1200 mAhthis leads to a lifetime variation from about 4 to 120 hours
Figure 25(a) shows for an example user that the drain isnot the same throughout the day but has diurnal variationsin which more energy is consumed during the day than dur-ing the night For this user the level of energy consumedchanges by roughly a factor of five Figure 25(b) plots thediurnal ratio of energy use for all users It shows that diurnalvariations occur with different strengths for all users
2060
100
140
0 4 8 12 16 20Hour of the day
Bat
tery
Dra
in (
mA
h)
(a)
0 20 40 60 80 100
12
34
56
7
User percentile
Diu
rnal
rat
io(b)
Figure 25 (a) The mean and 95 CI of energy drainof an example Dataset1 user (b) Diurnal ratio ofall users in Dataset1
Our results show that user activities contribute heavilytowards energy drain users in Dataset1 who have identi-cal smartphones drain energy at different rates and energydrain has diurnal patterns In the future we will developmethods to accurately quantify the energy consumption ofthe platform from that due to user-induced workload
7 IMPLICATIONS OF USER DIVERSITYWe uncover a surprising level of diversity among smart-
phone users For almost every aspect of usage that we studywe find one or more orders of magnitude difference betweenusers Our findings strongly motivate the need for customiz-ing smartphones to their users We believe that this needis greater than that for customizing ordinary cellphones orlaptops Ordinary cellphones do not have as rich an appli-cation environment Laptops are not as portable and aremore resource rich For example many users plug-in theirlaptops while using them
Customization can help at all levels Consider somethingas low-level as the battery Suppose we want batteries to beboth lightweight and last for at least a day with a high proba-bility Meeting the latter goal for all users of a given platformwill require catering to the heaviest users But that will leadto unnecessarily heavy batteries for many users (Highercapacity batteries are heavier) Offering multiple types ofbatteries with different lifetime-weight tradeoffs provides away out of this bind
At levels where intelligent mechanisms to improve user ex-perience or reduce energy consumption reside user diversitymotivates adapting to the smartphone user Driving thesemechanisms based on average case behaviors may not beeffective for a large fraction of the users
The ease and utility of customization depends on twoproperties of user behavior First despite quantitative dif-ferences there must be qualitative similarities among usersFor instance we should be able to describe the behavior ofall users with the same model Different users may havedifferent parameters of this model which will then lead toquantitative differences among them The presence of qual-itative similarities imply that users are not arbitrary pointsin space and it significantly simplifies the task of learninguser behavior Second user behavior in the past must alsobe predictive of the future Otherwise customization basedon past behaviors will be of little value in the future
In the next two sections we present evidence that theseproperties hold for several key aspects of smartphone usage
Dataset1 example user
000
00
010
002
00
030
0 100 200 300 400Session length (s)
Den
sity
Dataset2 example user
000
005
010
015
020
0 20 40 60 80 100Session length (s)
Den
sity
Figure 26 The PDF of session length for sampleusers of each dataset
In sect8 we show that user sessions and relative applicationpopularity can be described using simple models In sect9 weshow that energy drain can be described as well as predictedusing a ldquotrend trablerdquo framework
8 SMARTPHONE USAGE MODELSIn this section we develop simple models that describe
three aspects of smartphone usage ndash session lengths inter-arrival time between sessions and application popularityThe models are common across users but have different pa-rameters for different users While they do not completelydescribe user behavior they capture first order factors andrepresent a first step towards more complete modeling ofsmartphone usage More importantly along with the resultsof the next section they show that qualitative similaritiesdo exist among users
81 Session LengthsWe first consider the statistical properties of session length
distributions of users We find that session length valuestend to stationary With the KPSS test for level stationar-ity [13] 90 of the users have a p-value greater than 01The presence of stationarity is appealing because it suggeststhat past behavior is capable of predicting the future
We also find that session lengths are independent that isthe current value does not have a strong correlation with thevalues seen in the recent past With the Ljung-Box test forindependence [15] 96 of the users have a p-value that isgreater than 01
Stationarity and independence considered together sug-gest that session length values can be modeled as iid sam-ples from a distribution Choosing an appropriate distribu-tion however is complicated by the nature of the sessionlengths Most sessions are very short and the frequencydrops exponentially as the length increases However in-consistent with exponential behavior there are some verylong sessions in the tail for each user
We find that a mixture of exponential and Pareto distri-butions can model both ends of the spectrum The formercaptures short sessions and the latter captures long sessionsThat is session lengths can be described by the followingmixture model
r middot Exp(λ) + (1 minus r) middot Pareto(xm α)
In this equation r is the relative mix of the two distribu-tions λ is the rate of the exponential and xm and α are thelocation and shape parameters of the Pareto distribution
The location for a Pareto distribution represents the min-imum possible value of the random variable The location
010
0020
0030
00
0 1000 2000 3000Observed quantiles
Mod
el q
uant
iles
Figure 27 QQ plot of session lengths model for asample user
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed m
ix
(a) Relative mix (r)
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed r
ate
(b) Exponential rate (λ)
020
4060
0 20 40 60 80 100User percentile
Tim
eout
(s)
(c) Pareto location (xm)
minus1
0minus
04
02
06
10
0 20 40 60 80 100User percentile
Est
imat
ed s
hape
(d) Pareto shape (α)
Figure 28 Distribution of inferred model parame-ters that describe session length values of users inboth datasets
value that offers the best fit is the screen timeout value ofthe user because the session length PDF has a spike at thisvalue The spike corresponds to short sessions that are endedby the timeout (when the user forgets to switch the screenoff) we confirm this using controlled experiments with dif-ferent timeout values Figure 26 shows this spike at 60 and15 seconds for example users from each dataset The time-out provides a natural division between the two componentdistributions We automatically infer its approximate valueusing a simple spike detection algorithm
We use the EM algorithm to infer the maximum likelihoodestimation (MLE) of the remaining three parameters [6]Figure 27 shows the quality of this fit for an example userusing the QQ plot [3] Almost all quantiles are along they = x line indicating a good fit
Figure 28 shows the four inferred parameters for varioususers While users can be modeled using the same mix-
040
0080
00
0 4000 8000Observed quantiles
Mod
el q
uant
iles
Figure 29 QQ plot of session offtime model for asample user
050
010
0015
000 20 40 60 80 100
User percentileE
stim
ated
sca
le
(a) Weibull scale
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed s
hape
(b) Weibull shape
Figure 30 Distribution of inferred model parame-ters that describe the distribution of time betweensessions for users in both datasets
ture model the parameters of this model vary widely acrossusers Because of the way we construct our model the dis-tribution of the parameter r and xm also provide insightinto how frequently usersrsquo screen is switched off by the time-out and the relative popularity of different timeout values60 seconds is the most popular value likely because it is themost common default timeout and many users never changethe default
82 Time between SessionsWe find that the Weibull distribution can explain the
screen off times This distribution has two parameters re-ferred to as its scale and shape We find the MLE for theseparameters for each user From the QQ-plot in Figure 29we notice that the model predicts a greater probability ofseeing some very large offtimes than are observed in thedatasets However the probability of seeing these large off-times is small there are 27 data points that have a y-value greater than 8000 in that graph Hence we believethat Weibull provides a good fit for the length of intervalsbetween interactions
Figure 30 shows the distribution of the estimated shapeand scale of the fitted Weibull distributions Interestinglythe shape is consistently less than one Weibull shape valuesless than one indicate that the longer the screen has been offthe less likely it is for it to be turned on by the user This be-havior has interesting implications for power saving policiesFor instance periodic activities such as checking for emailwhen the screen has been off for a while may be deferred orrescheduled if needed without hurting user experience
020
4060
8010
0
0 20 40 60 80 100User percentile
MS
E (
)
(a) MSE
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed r
ate
(b) Rate parameter
Figure 31 (a) The mean square error (MSE) whenapplication popularity distribution is modeled usingan exponential (b) The inferred rate parameter ofthe exponential distribution for different users
83 Application PopularityWe find that for each user the relative popularity of ap-
plications can be well described by a simple exponential dis-tribution This qualitative invariant is useful for instanceto predict how many applications account for a given frac-tion of user attention For the example users in Figure 11this facet can be seen in the inset plots the semi-log of thepopularity distribution is very close to a straight line
Figure 31(a) shows that this exponential drop in applica-tion popularity is true for almost all users the mean squareerror between modeled exponential and actual popularitydistribution is less than 5 for 95 of the users
Figure 31(b) shows the inferred rate parameter of the ap-plication popularity distribution for various users We seethat the rate varies by an order of magnitude from 01 toalmost 1 The value of the rate essentially captures the paceof the drop in application popularity Lower values describeusers that use more applications on a regular basis Oneimplication of the wide range is that it may be feasible toretain all popular applications in memory for some users andnot for others
9 PREDICTING ENERGY DRAINIn this section we demonstrate the value of adapting to
user behaviors in the context of a mechanism to predict fu-ture energy drain on the smartphone Such a mechanismcan help with scheduling background tasks [9] estimatingbattery lifetime and improving user experience if it is esti-mated that the user will not fully drain the battery until thenext charging cycle [2 21]
Despite its many uses to our knowledge none of thesmartphones today provide a prediction of future energydrain Providing an accurate prediction of energy drain isdifficult2 User diversity of energy use makes any static pre-diction method highly inaccurate Even for the same userthere is a high variance in energy usage Figure 32 shows thisvariance by plotting the ratio of the standard deviation tothe mean energy drain over several time periods for users inDataset 1 The standard deviation of 10-minute windows ishigher than three times the mean for a fifth of the users The
2Energy drain prediction on smartphones is more challeng-ing than what is done for laptops Laptops provide a pre-diction of battery lifetime only for active use Smartphonepredictions on the other hand must cover multiple activeidle periods to be useful
01
23
45
0 20 40 60 80 100User percentile
Err
or r
atio
2 hours1 hour10 minutes
Figure 32 Battery drain predictors have to copewith the high variance in drain
time
window to
predict
n chunks
hellip
Trend Table
hellip
hellip
Figure 33 A userrsquos trend table captures how batterydrain in n consecutive chunks relates to drain in thesubsequent window
bursty nature of user interactions is one cause of this highvariance The variance is high even at longer time scalesThe standard deviation for two hour windows is larger than150 of the mean for half the users The diurnal patternsthat we showed earlier are one cause of high variance oversuch large time scales
Although the variance is high there are patterns in auserrsquos behaviors For instance we showed earlier that aphone that has been idle for a while is likely to remain idlein the near future and it will thus continue to draw energyat a similar rate
We hypothesize that a predictor tuned to user behaviorcan be accurate We present a simple personalized energydrain predictor that incorporates various user-specific fac-tors such as length of idle periods and different types ofbusy periods Instead of explicitly identifying these factorsour predictor captures them in a ldquotrend tablerdquo framework
Our framework is shown in Figure 33 Each entry in thetrend table is indexed by an n-tuple that represents energyusage readings from adjacent chunks of size δ each Thisindex points to energy usage statistics across all time win-dows of size w in the history that follow the n chunk valuesrepresented by that index Thus the trend table captureshow the energy usage in n consecutive chunks is related tothe energy use in the subsequent time windows of size w
In this paper we use n = 3 δ = 10 minutes and maintaindifferent trend tables for each window size w for which theprediction is required To keep the trend table small wequantize the energy readings of chunks We maintain themean and standard deviation for all subsequent windowsthat map to that index
To make a prediction we use the energy readings fromthe n immediately preceding chunks Let these readings be(x1 xn) Then we search for indices in the table thatdiffer from (x1 xn) by no more than a threshold s and
GenericShortminustermTimeminusofminusdayPersonalized
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
(a) 1-hour window
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
GenericShortminustermTimeminusofminusdayPersonalized
(b) 2-hour window
Figure 34 The ratio of error to true battery drainfor different energy drain predictors
predict based on them such that similar indices have higherrelative weight More precisely the prediction is
X
e stforalli=1n|he
iminusxi|lts
ke times ue
P
ke
where e iterates over all tuples in the trend table hei is
the value of the chunk i of the tuple e ue represents thestatistics stored for that index and the relative weight ke =s minus maxi=1n |he
i minus xi| For results shown here we uses = 05
Figures 34 shows how well our Personalized predictor worksfor predicting usage for two different future time windowsThere is one point per user in each graph which correspondsto the median error seen across many predictions for thatuser We see that the 90th percentile error is 25 for 1-hourwindow and 40 for 2-hour window
To place this level of accuracy in context the figure alsoshows the accurary of three alternative predictors TheGeneric predictor uses a trend table but instead of buildinga user-specific table it builds one that combines all usersWe see that its 90th percentile error is roughly twice thatof the Personalized predictor Its predictions are worse forhalf of the users
The Short-term predictor predicts energy drain simply asthe amount drained in the immediately preceding windowof the same size For both time windows its 90th percentileerror is no better than that of the Generic predictor
The Time-of-day predictor predicts energy drain based onvalues observed for the user at the same time of the day inthe past similar to that proposed by Banerjee et al [2]Thus this predictor has some degree of personalization butit learns user behavior in less detail than the Personalizedpredictor We see that its performance for 1-hour windowis similar to the Short-term predictor For 2-hour windowits error is higher than the Personalized predictor across theboard though its worst case error is lower than the Short-term and Generic predictors
Overall these results demonstrate the value of appropri-ately learning user behaviors to design intelligent mecha-nisms on smartphones
10 RELATED WORKIn a range of domains there is a rich history of work that
characterizes user workloads However because smartphoneadoption has gathered pace relatively recently our work rep-resents one of the few to study how people use smartphones
Along with other recent works it helps complete the pic-ture Banerjee et al and Rahmati et al report on batterycharging behaviors [2 18] Like us they find considerablevariation among users Banerjee et al also propose a pre-dictor that estimates the excess energy of the battery atthe time of charging using a histogram of past battery us-age [2] Shye et al study the power consumption character-istics of 20 users [24] They infer properties such as whichcomponents in the phone consume most power and exploreoptimizations based on these characteristics Rahmati andZhong study 14 users of a particular demographic to studywhich applications are popular in that demographic wherethe phone is used and how it is shared among users [19]
In contrast to these works we focus on understandingdifferent aspects of smartphone use (eg interactions andtraffic) and on exposing the diversity of user behaviors in-stead of only the average behavior Our study also entailsan order of magnitude more users than previous efforts
Like us other researchers have developed logging utilitiesMyExperience is one such early utility [10] and the worksabove involve custom utilities as well Our focus in thispaper is not on the development of logging tools but onanalyzing their data to gain insight into user behavior
There is a body of work in modeling the aggregate be-havior of mobile users Using traces from a large cellularoperator some network related aspects of mobile usage havebeen modeled Halepovic et al and Williamson et al re-port that call arrivals are bursty and present diurnal pat-terns [28] Willkomm et al and Brown et al report thatmobile users call duration can be approximately modeled bya lognormal distribution [4 29] We use traces collected onthe mobile device itself and focus on modeling the interac-tions of individual users instead of the aggregate behavior
11 CONCLUSIONSBy studying 255 users of two different smartphone plat-
forms we comprehensively characterized user activities andtheir impact on network and battery We quantify manyhitherto unknown aspects of smartphone usage User diver-sity is an overarching theme in our findings For instancedifferent users interact with their phones 10-200 times a dayon average the mean interaction length of different users
is 10-250 seconds and users receive 1-1000 MB of data perday where 10-90 is received as part of active use
This extent of user diversity implies that mechanisms thatwork for the average case may be ineffective for a large frac-tion of the users Instead learning and adapting to userbehaviors is likely to be more effective as demonstrated byour personalized energy drain predictor We show that qual-itative similarities exist among users to facilitate the devel-opment of such mechanisms For instance the longer theuser has not interacted with the phone the less likely it isfor her to start interacting with it and application popular-ity for a user follows a simple exponential distribution
Our study points to ways in which smartphone platformsshould be enhanced Effective adaptation will require futureplatforms to support light-weight tools that monitor andlearn user behaviors in situ It will also require them toexpose appropriate knobs to control the behavior of lower-level components (eg CPU or radio)
12 ACKNOWLEDGEMENTSWe are grateful to the users in Dataset1 that ran our log-
ger and to the providers of Dataset2 Elizabeth Dawson pro-vided administrative support for data collection Jean BolotKurt Partridge Stefan Saroiu Lin Zhong and the MobiSysreviewers provided many comments that helped improve thepaper We thank them all This work is supported in partby the NSF grants CCR-0120778 and CNS-0627084
13 REFERENCES[1] Balasubramanian N Balasubramanian A
and Venkataramani A Energy consumption inmobile phones A measurement study andimplications for network applications In IMC (2009)
[2] Banerjee N Rahmati A Corner M D
Rollins S and Zhong L Users and batteriesInteractions and adaptive energy management inmobile systems In Ubicomp (2007)
[3] Becker R Chambers J and Wilks A The newS language Chapman amp HallCRC 1988
[4] Brown L Gans N Mandelbaum A Sakov
A Shen H Zeltyn S and Zhao L Statisticalanalysis of a telephone call center Journal of theAmerican Statistical Association 100 469 (2005)
[5] Church K and Smyth B Understanding mobileinformation needs In MobileHCI (2008)
[6] Dempster A Laird N and Rubin D Maximumlikelihood from incomplete data via the EM algorithmJournal of the Royal Statistical Society Series B(Methodological) 39 1 (1977)
[7] Esfahbod B Preload An adaptive prefetchingdaemon Masterrsquos thesis University of Toronto 2006
[8] Falaki H Govindan R and Estrin D Smartscreen management on mobile phones Tech Rep 74Center for Embedded Networked Sesning 2009
[9] Flinn J and Satyanarayanan M Managingbattery lifetime with energy-aware adaptation ACMTOCS 22 2 (2004)
[10] Froehlich J Chen M Consolvo S
Harrison B and Landay J MyExperience Asystem for in situ tracing and capturing of userfeedback on mobile phones In MobiSys (2007)
[11] Fukuda K Cho K and Esaki H The impact ofresidential broadband traffic on Japanese ISPbackbones SIGCOMM CCR 35 1 (2005)
[12] Kotz D and Essien K Analysis of a campus-widewireless network Wireless Networks 11 1ndash2 (2005)
[13] Kwiatkowski D Phillips P Schmidt P and
Shin Y Testing the null hypothesis of stationarityagainst the alternative of a unit root Journal ofEconometrics 54 1-3 (1992)
[14] Lahiri K Dey S Panigrahi D and
Raghunathan A Battery-driven system design Anew frontier in low power design In Asia SouthPacific design automationVLSI Design (2002)
[15] Ljung G and Box G On a measure of lack of fitin time series models Biometrika 65 2 (1978)
[16] Smartphone futures 2009-2014 httpwwwportioresearchcomSmartphone09-14html
[17] Rahmati A Qian A and Zhong L
Understanding human-battery interaction on mobilephones In MobileHCI (2007)
[18] Rahmati A and Zhong L Humanndashbatteryinteraction on mobile phones Pervasive and MobileComputing 5 5 (2009)
[19] Rahmati A and Zhong L A longitudinal study ofnon-voice mobile phone usage by teens from anunderserved urban community Tech Rep 0515-09Rice University 2009
[20] Rao R Vrudhula S and Rakhmatov D
Battery modeling for energy aware system designIEEE Computer 36 12 (2003)
[21] Ravi N Scott J Han L and Iftode L
Context-aware battery management for mobilephones In PerCom (2008)
[22] Sharma A Navda V Ramjee R
Padmanabhan V and Belding E Cool-TetherEnergy efficient on-the-fly WiFi hot-spots usingmobile phones In CoNEXT (2009)
[23] Sharma C US wireless data market update - Q32009 httpwwwchetansharmacomusmarketupdateq309htm
[24] Shye A Sholbrock B and G M Into the wildStudying real user activity patterns to guide poweroptimization for mobile architectures In Micro (2009)
[25] Snol L More smartphones than desktop PCs by2011 httpwwwpcworldcomarticle171380more_smartphones_than_desktop_pcs_by_2011
html
[26] Sohn T Li K Griswold W and Hollan J
A diary study of mobile information needs In SIGCHI(2008)
[27] Trestian I Ranjan S Kuzmanovic A and
Nucci A Measuring serendipity Connecting peoplelocations and interests in a mobile 3G network InIMC (2009)
[28] Williamson C Halepovic E Sun H and Wu
Y Characterization of CDMA2000 Cellular DataNetwork Traffic In Local Computer Networks (2005)
[29] Willkomm D Machiraju S Bolot J and
Wolisz A Primary users in cellular networks Alarge-scale measurement study In DySPAN (2008)
- Introduction
- Data Collection
-
- Dataset1
- Dataset2
- Representativeness of conclusions
-
- User Interactions
-
- Interaction Time
- Interaction Sessions
- Diurnal Patterns
-
- Application Usage
-
- Number of applications
- Application Popularity
- Application Sessions
-
- Traffic
-
- Traffic per day
- ``Interactive traffic
- Diurnal patterns
-
- Energy Consumption
- Implications of User Diversity
- Smartphone Usage Models
-
- Session Lengths
- Time between Sessions
- Application Popularity
-
- Predicting Energy Drain
- Related Work
- Conclusions
- Acknowledgements
- References
-
tmail1
1010
010
00
0 20 40 60 80 100User percentile
Ses
sion
leng
th (
s)iexplore
110
100
1000
0 20 40 60 80 100User percentile
Ses
sion
leng
th (
s)Figure 18 The mean and the upper end of the stan-dard deviation of session lengths of two applications
0 20 40 60 80 100
User percentile
0
1
10
100
1000
10000
Traffic per day (MB) Receive
Send
Figure 19 The mean and the upper end of the stan-dard deviation of the traffic sent and received perday by users in Dataset1
traffic they generate We show that the diversity and diur-nal patterns of this side-effect match those of user actionsthemselves
The analysis in this section includes only Dataset1 we donot have traffic information for Dataset2 In Dataset1 werecord all of the data sent (or received) by the phone exceptfor that exchanged over the USB link ie the traffic hereinincludes data over the 3G radio and the 80211 wireless link
51 Traffic per dayFigure 19 shows that the amount of traffic sent and re-
ceived per day differs across users by almost three orders ofmagnitude The traffic received ranges from 1 to 1000 MBand the traffic sent ranges from 03 to 100 MB The medianvalues are 30 MB sent and 5 MB received
Our results indicate that traffic generated in a day bysmartphone users is comparable to traffic generated by PCsa few years ago A study of a campus WiFi network revealedthat on average users were generating 27 MB of traffic in2001 and 71 MB in 2003 [12] A study of Japanese residen-tial broadband users in 2006 revealed that on average usersgenerate 1000 MB of traffic per day [11] This high level oftraffic has major implications for the provisioning of wirelesscarrier networks as smartphone adoption increases
Relationship to application types To investigate ifcertain types of applications are favored more by users thatgenerate more traffice we divide the users into two equalclasses based on their sum of sent and received traffic perday Figure 20 shows mean and 95 CI of relative popularityof each application category for each user class Expectedly
Dataset1
Per
cent
age
0
20
40
60
80
com
m
brow
sing
gam
es
med
ia
map
s
prod
syst
em
othe
r
HighLow
Figure 20 The mean and 95 CI of relative popu-larity of each application category among high andlow traffic consumers
it shows that communication applications are more popularamong users that consume more traffic
52 ldquoInteractiverdquo trafficWe next estimate what fraction of the total traffic is ldquoin-
teractiverdquo that is has timeliness constraints Approachesto reduce power consumption of data transfers often advo-cate rescheduling network transfers for instance by delay-ing some transfers so that multiple of them may be bun-dled [1 22] Such policies are difficult to implement forinteractive traffic without hurting user response time andthus are likely to be less valuable if the bulk of the traffic isinteractive
We classify traffic as interactive if it was generated whenthe screen is on This classification method might classifysome background traffic as interactive We expect the errorto be low because the screen is on for a small fraction ofthe time for most users Because certain user interactionswith the phone begin right after traffic exchange (eg anew email is received) we also consider traffic received in asmall time window (1 minute) before the screen is turned onas having timeliness constraints The results are robust tothe exact choice of time window Indeed some of the trafficthat we classify as interactive might be delay tolerant ega new email to be sent might be deferred for a little whileHowever the volume of traffic received by the phone whichusers would rather see immediately dominates by one or-der of magnitude the volume that is sent and delay-tolerantmessaging applications such as email contribute only a smallchunk of all traffic
Figure 21 shows the fraction of interactive traffic for eachuser We see that for about 90 of the users over 50of the traffic is interactive but for the rest almost none oftheir traffic is interactive Stated differently for differentsmartphone users almost all to almost none of the traffic isgenerated by applications in the background The extremesrepresent disparate ways in which people use smartphonesand which applications on the smartphone generate mosttraffic Our results imply that the energy savings that can behad by rescheduling network activity will vary across users
53 Diurnal patternsFigure 22(a) shows the diurnal pattern for an example
user with a sharp decline at night Figure 22(b) showsthe strength of the diurnal pattern for individual users by
0 20 40 60 80 100
User percentile
00
02
04
06
08
10
Interactive fraction
Figure 21 The fraction of interactive traffic
0 5 10 15 20
Hour of day
0
5
10
15
Traffic per hour (MB)
Receive
Send
(a)
0 20 40 60 80 100
User percentile
2
4
6
8
10
12
Diurnal ratio
Receive
Send
(b)
Figure 22 (a) The mean and 95 CI for trafficgenerated per hour by an example user (b) Thediurnal ratio of traffic per hour for all users
plotting the diurnal ratio of traffic As defined in sect33 thediurnal ratio reflects how much higher the mean during thepeak hour is to the overall mean
We see that the diurnal ratio varies across users but mosthave a strong diurnal behavior 80 of them generate overtwice their average amount of traffic in their peak hour Thisbehavior is likely a direct result of the high proportion ofinteractive traffic for most users and that user interactionsthemselves have a diurnal pattern
6 ENERGY CONSUMPTIONThe final aspect of smartphone usage that we investigate
is energy consumption Energy drain depends on two fac-tors i) user interactions and applications and ii) platformhardware and software If the second factor dominates theenergy drain of various users with the identical smartphonewill be similar Otherwise the energy drain will be as di-verse as user behaviors
We estimate the amount of energy drain based on the re-maining battery indicator which varies between 0 and 100If the battery indicator has gone down by X in a time pe-riod for a battery with capacity Y mAh we compute theenergy drain in that period to be X middot Y mAh 1 Given thatbatteries are complex electro-chemical devices [14 20] thiscomputation is approximate It assumes that the batterylevel indicator is linear with respect to energy drain
1mAh is technically a unit of charge yet is commonly used toindicate energy drain because battery voltage during normaloperations is typically constant For phones in our datasetthis is 4V so multiplying a mAh reading by 4 would yieldan accurate energy reading in milli-watt-hours
020
4060
8010
0
0 4 8 12 16 20Time (hour)
Bat
tery
leve
l (
) Benchmark1Benchmark2
Figure 23 Timelapse of the remaining battery levelindicator in controlled experiments with two differ-ent workloads at room temperature
110
100
1000
0 20 40 60 80 100User percentile
Bat
tery
Dra
in (
mA
h)
Figure 24 The mean and the upper end of the stan-dard deviation of one hour energy drain for Dataset1users during discharge periods
Controlled experiments suggest that the linearity assump-tion holds to a first order We run a benchmark load thatdrains the battery at a fixed rate in room temperature Un-der this benchmark if the battery level indicator decreaseslinearly with time it must be linear with respect to energydrain Figure 23 shows that the level indicator decreasesroughly linearly for two different benchmarks Benchmark1turns the screen on and off periodically Benchmark2 com-putes and idles periodically We conclude thus that the levelindicator can be used to estimate energy drain
Figure 24 shows the mean and standard deviation of en-ergy that users drain in an hour This graph is computed us-ing only periods in which the battery is not charging becauseenergy drain in those periods are of primary interest Wesee a two orders of magnitude difference among users Whileheaviest users drain close to 250 mAh the lightest of usersdrain only 10 mAh If the battery capacity is 1200 mAhthis leads to a lifetime variation from about 4 to 120 hours
Figure 25(a) shows for an example user that the drain isnot the same throughout the day but has diurnal variationsin which more energy is consumed during the day than dur-ing the night For this user the level of energy consumedchanges by roughly a factor of five Figure 25(b) plots thediurnal ratio of energy use for all users It shows that diurnalvariations occur with different strengths for all users
2060
100
140
0 4 8 12 16 20Hour of the day
Bat
tery
Dra
in (
mA
h)
(a)
0 20 40 60 80 100
12
34
56
7
User percentile
Diu
rnal
rat
io(b)
Figure 25 (a) The mean and 95 CI of energy drainof an example Dataset1 user (b) Diurnal ratio ofall users in Dataset1
Our results show that user activities contribute heavilytowards energy drain users in Dataset1 who have identi-cal smartphones drain energy at different rates and energydrain has diurnal patterns In the future we will developmethods to accurately quantify the energy consumption ofthe platform from that due to user-induced workload
7 IMPLICATIONS OF USER DIVERSITYWe uncover a surprising level of diversity among smart-
phone users For almost every aspect of usage that we studywe find one or more orders of magnitude difference betweenusers Our findings strongly motivate the need for customiz-ing smartphones to their users We believe that this needis greater than that for customizing ordinary cellphones orlaptops Ordinary cellphones do not have as rich an appli-cation environment Laptops are not as portable and aremore resource rich For example many users plug-in theirlaptops while using them
Customization can help at all levels Consider somethingas low-level as the battery Suppose we want batteries to beboth lightweight and last for at least a day with a high proba-bility Meeting the latter goal for all users of a given platformwill require catering to the heaviest users But that will leadto unnecessarily heavy batteries for many users (Highercapacity batteries are heavier) Offering multiple types ofbatteries with different lifetime-weight tradeoffs provides away out of this bind
At levels where intelligent mechanisms to improve user ex-perience or reduce energy consumption reside user diversitymotivates adapting to the smartphone user Driving thesemechanisms based on average case behaviors may not beeffective for a large fraction of the users
The ease and utility of customization depends on twoproperties of user behavior First despite quantitative dif-ferences there must be qualitative similarities among usersFor instance we should be able to describe the behavior ofall users with the same model Different users may havedifferent parameters of this model which will then lead toquantitative differences among them The presence of qual-itative similarities imply that users are not arbitrary pointsin space and it significantly simplifies the task of learninguser behavior Second user behavior in the past must alsobe predictive of the future Otherwise customization basedon past behaviors will be of little value in the future
In the next two sections we present evidence that theseproperties hold for several key aspects of smartphone usage
Dataset1 example user
000
00
010
002
00
030
0 100 200 300 400Session length (s)
Den
sity
Dataset2 example user
000
005
010
015
020
0 20 40 60 80 100Session length (s)
Den
sity
Figure 26 The PDF of session length for sampleusers of each dataset
In sect8 we show that user sessions and relative applicationpopularity can be described using simple models In sect9 weshow that energy drain can be described as well as predictedusing a ldquotrend trablerdquo framework
8 SMARTPHONE USAGE MODELSIn this section we develop simple models that describe
three aspects of smartphone usage ndash session lengths inter-arrival time between sessions and application popularityThe models are common across users but have different pa-rameters for different users While they do not completelydescribe user behavior they capture first order factors andrepresent a first step towards more complete modeling ofsmartphone usage More importantly along with the resultsof the next section they show that qualitative similaritiesdo exist among users
81 Session LengthsWe first consider the statistical properties of session length
distributions of users We find that session length valuestend to stationary With the KPSS test for level stationar-ity [13] 90 of the users have a p-value greater than 01The presence of stationarity is appealing because it suggeststhat past behavior is capable of predicting the future
We also find that session lengths are independent that isthe current value does not have a strong correlation with thevalues seen in the recent past With the Ljung-Box test forindependence [15] 96 of the users have a p-value that isgreater than 01
Stationarity and independence considered together sug-gest that session length values can be modeled as iid sam-ples from a distribution Choosing an appropriate distribu-tion however is complicated by the nature of the sessionlengths Most sessions are very short and the frequencydrops exponentially as the length increases However in-consistent with exponential behavior there are some verylong sessions in the tail for each user
We find that a mixture of exponential and Pareto distri-butions can model both ends of the spectrum The formercaptures short sessions and the latter captures long sessionsThat is session lengths can be described by the followingmixture model
r middot Exp(λ) + (1 minus r) middot Pareto(xm α)
In this equation r is the relative mix of the two distribu-tions λ is the rate of the exponential and xm and α are thelocation and shape parameters of the Pareto distribution
The location for a Pareto distribution represents the min-imum possible value of the random variable The location
010
0020
0030
00
0 1000 2000 3000Observed quantiles
Mod
el q
uant
iles
Figure 27 QQ plot of session lengths model for asample user
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed m
ix
(a) Relative mix (r)
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed r
ate
(b) Exponential rate (λ)
020
4060
0 20 40 60 80 100User percentile
Tim
eout
(s)
(c) Pareto location (xm)
minus1
0minus
04
02
06
10
0 20 40 60 80 100User percentile
Est
imat
ed s
hape
(d) Pareto shape (α)
Figure 28 Distribution of inferred model parame-ters that describe session length values of users inboth datasets
value that offers the best fit is the screen timeout value ofthe user because the session length PDF has a spike at thisvalue The spike corresponds to short sessions that are endedby the timeout (when the user forgets to switch the screenoff) we confirm this using controlled experiments with dif-ferent timeout values Figure 26 shows this spike at 60 and15 seconds for example users from each dataset The time-out provides a natural division between the two componentdistributions We automatically infer its approximate valueusing a simple spike detection algorithm
We use the EM algorithm to infer the maximum likelihoodestimation (MLE) of the remaining three parameters [6]Figure 27 shows the quality of this fit for an example userusing the QQ plot [3] Almost all quantiles are along they = x line indicating a good fit
Figure 28 shows the four inferred parameters for varioususers While users can be modeled using the same mix-
040
0080
00
0 4000 8000Observed quantiles
Mod
el q
uant
iles
Figure 29 QQ plot of session offtime model for asample user
050
010
0015
000 20 40 60 80 100
User percentileE
stim
ated
sca
le
(a) Weibull scale
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed s
hape
(b) Weibull shape
Figure 30 Distribution of inferred model parame-ters that describe the distribution of time betweensessions for users in both datasets
ture model the parameters of this model vary widely acrossusers Because of the way we construct our model the dis-tribution of the parameter r and xm also provide insightinto how frequently usersrsquo screen is switched off by the time-out and the relative popularity of different timeout values60 seconds is the most popular value likely because it is themost common default timeout and many users never changethe default
82 Time between SessionsWe find that the Weibull distribution can explain the
screen off times This distribution has two parameters re-ferred to as its scale and shape We find the MLE for theseparameters for each user From the QQ-plot in Figure 29we notice that the model predicts a greater probability ofseeing some very large offtimes than are observed in thedatasets However the probability of seeing these large off-times is small there are 27 data points that have a y-value greater than 8000 in that graph Hence we believethat Weibull provides a good fit for the length of intervalsbetween interactions
Figure 30 shows the distribution of the estimated shapeand scale of the fitted Weibull distributions Interestinglythe shape is consistently less than one Weibull shape valuesless than one indicate that the longer the screen has been offthe less likely it is for it to be turned on by the user This be-havior has interesting implications for power saving policiesFor instance periodic activities such as checking for emailwhen the screen has been off for a while may be deferred orrescheduled if needed without hurting user experience
020
4060
8010
0
0 20 40 60 80 100User percentile
MS
E (
)
(a) MSE
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed r
ate
(b) Rate parameter
Figure 31 (a) The mean square error (MSE) whenapplication popularity distribution is modeled usingan exponential (b) The inferred rate parameter ofthe exponential distribution for different users
83 Application PopularityWe find that for each user the relative popularity of ap-
plications can be well described by a simple exponential dis-tribution This qualitative invariant is useful for instanceto predict how many applications account for a given frac-tion of user attention For the example users in Figure 11this facet can be seen in the inset plots the semi-log of thepopularity distribution is very close to a straight line
Figure 31(a) shows that this exponential drop in applica-tion popularity is true for almost all users the mean squareerror between modeled exponential and actual popularitydistribution is less than 5 for 95 of the users
Figure 31(b) shows the inferred rate parameter of the ap-plication popularity distribution for various users We seethat the rate varies by an order of magnitude from 01 toalmost 1 The value of the rate essentially captures the paceof the drop in application popularity Lower values describeusers that use more applications on a regular basis Oneimplication of the wide range is that it may be feasible toretain all popular applications in memory for some users andnot for others
9 PREDICTING ENERGY DRAINIn this section we demonstrate the value of adapting to
user behaviors in the context of a mechanism to predict fu-ture energy drain on the smartphone Such a mechanismcan help with scheduling background tasks [9] estimatingbattery lifetime and improving user experience if it is esti-mated that the user will not fully drain the battery until thenext charging cycle [2 21]
Despite its many uses to our knowledge none of thesmartphones today provide a prediction of future energydrain Providing an accurate prediction of energy drain isdifficult2 User diversity of energy use makes any static pre-diction method highly inaccurate Even for the same userthere is a high variance in energy usage Figure 32 shows thisvariance by plotting the ratio of the standard deviation tothe mean energy drain over several time periods for users inDataset 1 The standard deviation of 10-minute windows ishigher than three times the mean for a fifth of the users The
2Energy drain prediction on smartphones is more challeng-ing than what is done for laptops Laptops provide a pre-diction of battery lifetime only for active use Smartphonepredictions on the other hand must cover multiple activeidle periods to be useful
01
23
45
0 20 40 60 80 100User percentile
Err
or r
atio
2 hours1 hour10 minutes
Figure 32 Battery drain predictors have to copewith the high variance in drain
time
window to
predict
n chunks
hellip
Trend Table
hellip
hellip
Figure 33 A userrsquos trend table captures how batterydrain in n consecutive chunks relates to drain in thesubsequent window
bursty nature of user interactions is one cause of this highvariance The variance is high even at longer time scalesThe standard deviation for two hour windows is larger than150 of the mean for half the users The diurnal patternsthat we showed earlier are one cause of high variance oversuch large time scales
Although the variance is high there are patterns in auserrsquos behaviors For instance we showed earlier that aphone that has been idle for a while is likely to remain idlein the near future and it will thus continue to draw energyat a similar rate
We hypothesize that a predictor tuned to user behaviorcan be accurate We present a simple personalized energydrain predictor that incorporates various user-specific fac-tors such as length of idle periods and different types ofbusy periods Instead of explicitly identifying these factorsour predictor captures them in a ldquotrend tablerdquo framework
Our framework is shown in Figure 33 Each entry in thetrend table is indexed by an n-tuple that represents energyusage readings from adjacent chunks of size δ each Thisindex points to energy usage statistics across all time win-dows of size w in the history that follow the n chunk valuesrepresented by that index Thus the trend table captureshow the energy usage in n consecutive chunks is related tothe energy use in the subsequent time windows of size w
In this paper we use n = 3 δ = 10 minutes and maintaindifferent trend tables for each window size w for which theprediction is required To keep the trend table small wequantize the energy readings of chunks We maintain themean and standard deviation for all subsequent windowsthat map to that index
To make a prediction we use the energy readings fromthe n immediately preceding chunks Let these readings be(x1 xn) Then we search for indices in the table thatdiffer from (x1 xn) by no more than a threshold s and
GenericShortminustermTimeminusofminusdayPersonalized
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
(a) 1-hour window
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
GenericShortminustermTimeminusofminusdayPersonalized
(b) 2-hour window
Figure 34 The ratio of error to true battery drainfor different energy drain predictors
predict based on them such that similar indices have higherrelative weight More precisely the prediction is
X
e stforalli=1n|he
iminusxi|lts
ke times ue
P
ke
where e iterates over all tuples in the trend table hei is
the value of the chunk i of the tuple e ue represents thestatistics stored for that index and the relative weight ke =s minus maxi=1n |he
i minus xi| For results shown here we uses = 05
Figures 34 shows how well our Personalized predictor worksfor predicting usage for two different future time windowsThere is one point per user in each graph which correspondsto the median error seen across many predictions for thatuser We see that the 90th percentile error is 25 for 1-hourwindow and 40 for 2-hour window
To place this level of accuracy in context the figure alsoshows the accurary of three alternative predictors TheGeneric predictor uses a trend table but instead of buildinga user-specific table it builds one that combines all usersWe see that its 90th percentile error is roughly twice thatof the Personalized predictor Its predictions are worse forhalf of the users
The Short-term predictor predicts energy drain simply asthe amount drained in the immediately preceding windowof the same size For both time windows its 90th percentileerror is no better than that of the Generic predictor
The Time-of-day predictor predicts energy drain based onvalues observed for the user at the same time of the day inthe past similar to that proposed by Banerjee et al [2]Thus this predictor has some degree of personalization butit learns user behavior in less detail than the Personalizedpredictor We see that its performance for 1-hour windowis similar to the Short-term predictor For 2-hour windowits error is higher than the Personalized predictor across theboard though its worst case error is lower than the Short-term and Generic predictors
Overall these results demonstrate the value of appropri-ately learning user behaviors to design intelligent mecha-nisms on smartphones
10 RELATED WORKIn a range of domains there is a rich history of work that
characterizes user workloads However because smartphoneadoption has gathered pace relatively recently our work rep-resents one of the few to study how people use smartphones
Along with other recent works it helps complete the pic-ture Banerjee et al and Rahmati et al report on batterycharging behaviors [2 18] Like us they find considerablevariation among users Banerjee et al also propose a pre-dictor that estimates the excess energy of the battery atthe time of charging using a histogram of past battery us-age [2] Shye et al study the power consumption character-istics of 20 users [24] They infer properties such as whichcomponents in the phone consume most power and exploreoptimizations based on these characteristics Rahmati andZhong study 14 users of a particular demographic to studywhich applications are popular in that demographic wherethe phone is used and how it is shared among users [19]
In contrast to these works we focus on understandingdifferent aspects of smartphone use (eg interactions andtraffic) and on exposing the diversity of user behaviors in-stead of only the average behavior Our study also entailsan order of magnitude more users than previous efforts
Like us other researchers have developed logging utilitiesMyExperience is one such early utility [10] and the worksabove involve custom utilities as well Our focus in thispaper is not on the development of logging tools but onanalyzing their data to gain insight into user behavior
There is a body of work in modeling the aggregate be-havior of mobile users Using traces from a large cellularoperator some network related aspects of mobile usage havebeen modeled Halepovic et al and Williamson et al re-port that call arrivals are bursty and present diurnal pat-terns [28] Willkomm et al and Brown et al report thatmobile users call duration can be approximately modeled bya lognormal distribution [4 29] We use traces collected onthe mobile device itself and focus on modeling the interac-tions of individual users instead of the aggregate behavior
11 CONCLUSIONSBy studying 255 users of two different smartphone plat-
forms we comprehensively characterized user activities andtheir impact on network and battery We quantify manyhitherto unknown aspects of smartphone usage User diver-sity is an overarching theme in our findings For instancedifferent users interact with their phones 10-200 times a dayon average the mean interaction length of different users
is 10-250 seconds and users receive 1-1000 MB of data perday where 10-90 is received as part of active use
This extent of user diversity implies that mechanisms thatwork for the average case may be ineffective for a large frac-tion of the users Instead learning and adapting to userbehaviors is likely to be more effective as demonstrated byour personalized energy drain predictor We show that qual-itative similarities exist among users to facilitate the devel-opment of such mechanisms For instance the longer theuser has not interacted with the phone the less likely it isfor her to start interacting with it and application popular-ity for a user follows a simple exponential distribution
Our study points to ways in which smartphone platformsshould be enhanced Effective adaptation will require futureplatforms to support light-weight tools that monitor andlearn user behaviors in situ It will also require them toexpose appropriate knobs to control the behavior of lower-level components (eg CPU or radio)
12 ACKNOWLEDGEMENTSWe are grateful to the users in Dataset1 that ran our log-
ger and to the providers of Dataset2 Elizabeth Dawson pro-vided administrative support for data collection Jean BolotKurt Partridge Stefan Saroiu Lin Zhong and the MobiSysreviewers provided many comments that helped improve thepaper We thank them all This work is supported in partby the NSF grants CCR-0120778 and CNS-0627084
13 REFERENCES[1] Balasubramanian N Balasubramanian A
and Venkataramani A Energy consumption inmobile phones A measurement study andimplications for network applications In IMC (2009)
[2] Banerjee N Rahmati A Corner M D
Rollins S and Zhong L Users and batteriesInteractions and adaptive energy management inmobile systems In Ubicomp (2007)
[3] Becker R Chambers J and Wilks A The newS language Chapman amp HallCRC 1988
[4] Brown L Gans N Mandelbaum A Sakov
A Shen H Zeltyn S and Zhao L Statisticalanalysis of a telephone call center Journal of theAmerican Statistical Association 100 469 (2005)
[5] Church K and Smyth B Understanding mobileinformation needs In MobileHCI (2008)
[6] Dempster A Laird N and Rubin D Maximumlikelihood from incomplete data via the EM algorithmJournal of the Royal Statistical Society Series B(Methodological) 39 1 (1977)
[7] Esfahbod B Preload An adaptive prefetchingdaemon Masterrsquos thesis University of Toronto 2006
[8] Falaki H Govindan R and Estrin D Smartscreen management on mobile phones Tech Rep 74Center for Embedded Networked Sesning 2009
[9] Flinn J and Satyanarayanan M Managingbattery lifetime with energy-aware adaptation ACMTOCS 22 2 (2004)
[10] Froehlich J Chen M Consolvo S
Harrison B and Landay J MyExperience Asystem for in situ tracing and capturing of userfeedback on mobile phones In MobiSys (2007)
[11] Fukuda K Cho K and Esaki H The impact ofresidential broadband traffic on Japanese ISPbackbones SIGCOMM CCR 35 1 (2005)
[12] Kotz D and Essien K Analysis of a campus-widewireless network Wireless Networks 11 1ndash2 (2005)
[13] Kwiatkowski D Phillips P Schmidt P and
Shin Y Testing the null hypothesis of stationarityagainst the alternative of a unit root Journal ofEconometrics 54 1-3 (1992)
[14] Lahiri K Dey S Panigrahi D and
Raghunathan A Battery-driven system design Anew frontier in low power design In Asia SouthPacific design automationVLSI Design (2002)
[15] Ljung G and Box G On a measure of lack of fitin time series models Biometrika 65 2 (1978)
[16] Smartphone futures 2009-2014 httpwwwportioresearchcomSmartphone09-14html
[17] Rahmati A Qian A and Zhong L
Understanding human-battery interaction on mobilephones In MobileHCI (2007)
[18] Rahmati A and Zhong L Humanndashbatteryinteraction on mobile phones Pervasive and MobileComputing 5 5 (2009)
[19] Rahmati A and Zhong L A longitudinal study ofnon-voice mobile phone usage by teens from anunderserved urban community Tech Rep 0515-09Rice University 2009
[20] Rao R Vrudhula S and Rakhmatov D
Battery modeling for energy aware system designIEEE Computer 36 12 (2003)
[21] Ravi N Scott J Han L and Iftode L
Context-aware battery management for mobilephones In PerCom (2008)
[22] Sharma A Navda V Ramjee R
Padmanabhan V and Belding E Cool-TetherEnergy efficient on-the-fly WiFi hot-spots usingmobile phones In CoNEXT (2009)
[23] Sharma C US wireless data market update - Q32009 httpwwwchetansharmacomusmarketupdateq309htm
[24] Shye A Sholbrock B and G M Into the wildStudying real user activity patterns to guide poweroptimization for mobile architectures In Micro (2009)
[25] Snol L More smartphones than desktop PCs by2011 httpwwwpcworldcomarticle171380more_smartphones_than_desktop_pcs_by_2011
html
[26] Sohn T Li K Griswold W and Hollan J
A diary study of mobile information needs In SIGCHI(2008)
[27] Trestian I Ranjan S Kuzmanovic A and
Nucci A Measuring serendipity Connecting peoplelocations and interests in a mobile 3G network InIMC (2009)
[28] Williamson C Halepovic E Sun H and Wu
Y Characterization of CDMA2000 Cellular DataNetwork Traffic In Local Computer Networks (2005)
[29] Willkomm D Machiraju S Bolot J and
Wolisz A Primary users in cellular networks Alarge-scale measurement study In DySPAN (2008)
- Introduction
- Data Collection
-
- Dataset1
- Dataset2
- Representativeness of conclusions
-
- User Interactions
-
- Interaction Time
- Interaction Sessions
- Diurnal Patterns
-
- Application Usage
-
- Number of applications
- Application Popularity
- Application Sessions
-
- Traffic
-
- Traffic per day
- ``Interactive traffic
- Diurnal patterns
-
- Energy Consumption
- Implications of User Diversity
- Smartphone Usage Models
-
- Session Lengths
- Time between Sessions
- Application Popularity
-
- Predicting Energy Drain
- Related Work
- Conclusions
- Acknowledgements
- References
-
0 20 40 60 80 100
User percentile
00
02
04
06
08
10
Interactive fraction
Figure 21 The fraction of interactive traffic
0 5 10 15 20
Hour of day
0
5
10
15
Traffic per hour (MB)
Receive
Send
(a)
0 20 40 60 80 100
User percentile
2
4
6
8
10
12
Diurnal ratio
Receive
Send
(b)
Figure 22 (a) The mean and 95 CI for trafficgenerated per hour by an example user (b) Thediurnal ratio of traffic per hour for all users
plotting the diurnal ratio of traffic As defined in sect33 thediurnal ratio reflects how much higher the mean during thepeak hour is to the overall mean
We see that the diurnal ratio varies across users but mosthave a strong diurnal behavior 80 of them generate overtwice their average amount of traffic in their peak hour Thisbehavior is likely a direct result of the high proportion ofinteractive traffic for most users and that user interactionsthemselves have a diurnal pattern
6 ENERGY CONSUMPTIONThe final aspect of smartphone usage that we investigate
is energy consumption Energy drain depends on two fac-tors i) user interactions and applications and ii) platformhardware and software If the second factor dominates theenergy drain of various users with the identical smartphonewill be similar Otherwise the energy drain will be as di-verse as user behaviors
We estimate the amount of energy drain based on the re-maining battery indicator which varies between 0 and 100If the battery indicator has gone down by X in a time pe-riod for a battery with capacity Y mAh we compute theenergy drain in that period to be X middot Y mAh 1 Given thatbatteries are complex electro-chemical devices [14 20] thiscomputation is approximate It assumes that the batterylevel indicator is linear with respect to energy drain
1mAh is technically a unit of charge yet is commonly used toindicate energy drain because battery voltage during normaloperations is typically constant For phones in our datasetthis is 4V so multiplying a mAh reading by 4 would yieldan accurate energy reading in milli-watt-hours
020
4060
8010
0
0 4 8 12 16 20Time (hour)
Bat
tery
leve
l (
) Benchmark1Benchmark2
Figure 23 Timelapse of the remaining battery levelindicator in controlled experiments with two differ-ent workloads at room temperature
110
100
1000
0 20 40 60 80 100User percentile
Bat
tery
Dra
in (
mA
h)
Figure 24 The mean and the upper end of the stan-dard deviation of one hour energy drain for Dataset1users during discharge periods
Controlled experiments suggest that the linearity assump-tion holds to a first order We run a benchmark load thatdrains the battery at a fixed rate in room temperature Un-der this benchmark if the battery level indicator decreaseslinearly with time it must be linear with respect to energydrain Figure 23 shows that the level indicator decreasesroughly linearly for two different benchmarks Benchmark1turns the screen on and off periodically Benchmark2 com-putes and idles periodically We conclude thus that the levelindicator can be used to estimate energy drain
Figure 24 shows the mean and standard deviation of en-ergy that users drain in an hour This graph is computed us-ing only periods in which the battery is not charging becauseenergy drain in those periods are of primary interest Wesee a two orders of magnitude difference among users Whileheaviest users drain close to 250 mAh the lightest of usersdrain only 10 mAh If the battery capacity is 1200 mAhthis leads to a lifetime variation from about 4 to 120 hours
Figure 25(a) shows for an example user that the drain isnot the same throughout the day but has diurnal variationsin which more energy is consumed during the day than dur-ing the night For this user the level of energy consumedchanges by roughly a factor of five Figure 25(b) plots thediurnal ratio of energy use for all users It shows that diurnalvariations occur with different strengths for all users
2060
100
140
0 4 8 12 16 20Hour of the day
Bat
tery
Dra
in (
mA
h)
(a)
0 20 40 60 80 100
12
34
56
7
User percentile
Diu
rnal
rat
io(b)
Figure 25 (a) The mean and 95 CI of energy drainof an example Dataset1 user (b) Diurnal ratio ofall users in Dataset1
Our results show that user activities contribute heavilytowards energy drain users in Dataset1 who have identi-cal smartphones drain energy at different rates and energydrain has diurnal patterns In the future we will developmethods to accurately quantify the energy consumption ofthe platform from that due to user-induced workload
7 IMPLICATIONS OF USER DIVERSITYWe uncover a surprising level of diversity among smart-
phone users For almost every aspect of usage that we studywe find one or more orders of magnitude difference betweenusers Our findings strongly motivate the need for customiz-ing smartphones to their users We believe that this needis greater than that for customizing ordinary cellphones orlaptops Ordinary cellphones do not have as rich an appli-cation environment Laptops are not as portable and aremore resource rich For example many users plug-in theirlaptops while using them
Customization can help at all levels Consider somethingas low-level as the battery Suppose we want batteries to beboth lightweight and last for at least a day with a high proba-bility Meeting the latter goal for all users of a given platformwill require catering to the heaviest users But that will leadto unnecessarily heavy batteries for many users (Highercapacity batteries are heavier) Offering multiple types ofbatteries with different lifetime-weight tradeoffs provides away out of this bind
At levels where intelligent mechanisms to improve user ex-perience or reduce energy consumption reside user diversitymotivates adapting to the smartphone user Driving thesemechanisms based on average case behaviors may not beeffective for a large fraction of the users
The ease and utility of customization depends on twoproperties of user behavior First despite quantitative dif-ferences there must be qualitative similarities among usersFor instance we should be able to describe the behavior ofall users with the same model Different users may havedifferent parameters of this model which will then lead toquantitative differences among them The presence of qual-itative similarities imply that users are not arbitrary pointsin space and it significantly simplifies the task of learninguser behavior Second user behavior in the past must alsobe predictive of the future Otherwise customization basedon past behaviors will be of little value in the future
In the next two sections we present evidence that theseproperties hold for several key aspects of smartphone usage
Dataset1 example user
000
00
010
002
00
030
0 100 200 300 400Session length (s)
Den
sity
Dataset2 example user
000
005
010
015
020
0 20 40 60 80 100Session length (s)
Den
sity
Figure 26 The PDF of session length for sampleusers of each dataset
In sect8 we show that user sessions and relative applicationpopularity can be described using simple models In sect9 weshow that energy drain can be described as well as predictedusing a ldquotrend trablerdquo framework
8 SMARTPHONE USAGE MODELSIn this section we develop simple models that describe
three aspects of smartphone usage ndash session lengths inter-arrival time between sessions and application popularityThe models are common across users but have different pa-rameters for different users While they do not completelydescribe user behavior they capture first order factors andrepresent a first step towards more complete modeling ofsmartphone usage More importantly along with the resultsof the next section they show that qualitative similaritiesdo exist among users
81 Session LengthsWe first consider the statistical properties of session length
distributions of users We find that session length valuestend to stationary With the KPSS test for level stationar-ity [13] 90 of the users have a p-value greater than 01The presence of stationarity is appealing because it suggeststhat past behavior is capable of predicting the future
We also find that session lengths are independent that isthe current value does not have a strong correlation with thevalues seen in the recent past With the Ljung-Box test forindependence [15] 96 of the users have a p-value that isgreater than 01
Stationarity and independence considered together sug-gest that session length values can be modeled as iid sam-ples from a distribution Choosing an appropriate distribu-tion however is complicated by the nature of the sessionlengths Most sessions are very short and the frequencydrops exponentially as the length increases However in-consistent with exponential behavior there are some verylong sessions in the tail for each user
We find that a mixture of exponential and Pareto distri-butions can model both ends of the spectrum The formercaptures short sessions and the latter captures long sessionsThat is session lengths can be described by the followingmixture model
r middot Exp(λ) + (1 minus r) middot Pareto(xm α)
In this equation r is the relative mix of the two distribu-tions λ is the rate of the exponential and xm and α are thelocation and shape parameters of the Pareto distribution
The location for a Pareto distribution represents the min-imum possible value of the random variable The location
010
0020
0030
00
0 1000 2000 3000Observed quantiles
Mod
el q
uant
iles
Figure 27 QQ plot of session lengths model for asample user
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed m
ix
(a) Relative mix (r)
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed r
ate
(b) Exponential rate (λ)
020
4060
0 20 40 60 80 100User percentile
Tim
eout
(s)
(c) Pareto location (xm)
minus1
0minus
04
02
06
10
0 20 40 60 80 100User percentile
Est
imat
ed s
hape
(d) Pareto shape (α)
Figure 28 Distribution of inferred model parame-ters that describe session length values of users inboth datasets
value that offers the best fit is the screen timeout value ofthe user because the session length PDF has a spike at thisvalue The spike corresponds to short sessions that are endedby the timeout (when the user forgets to switch the screenoff) we confirm this using controlled experiments with dif-ferent timeout values Figure 26 shows this spike at 60 and15 seconds for example users from each dataset The time-out provides a natural division between the two componentdistributions We automatically infer its approximate valueusing a simple spike detection algorithm
We use the EM algorithm to infer the maximum likelihoodestimation (MLE) of the remaining three parameters [6]Figure 27 shows the quality of this fit for an example userusing the QQ plot [3] Almost all quantiles are along they = x line indicating a good fit
Figure 28 shows the four inferred parameters for varioususers While users can be modeled using the same mix-
040
0080
00
0 4000 8000Observed quantiles
Mod
el q
uant
iles
Figure 29 QQ plot of session offtime model for asample user
050
010
0015
000 20 40 60 80 100
User percentileE
stim
ated
sca
le
(a) Weibull scale
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed s
hape
(b) Weibull shape
Figure 30 Distribution of inferred model parame-ters that describe the distribution of time betweensessions for users in both datasets
ture model the parameters of this model vary widely acrossusers Because of the way we construct our model the dis-tribution of the parameter r and xm also provide insightinto how frequently usersrsquo screen is switched off by the time-out and the relative popularity of different timeout values60 seconds is the most popular value likely because it is themost common default timeout and many users never changethe default
82 Time between SessionsWe find that the Weibull distribution can explain the
screen off times This distribution has two parameters re-ferred to as its scale and shape We find the MLE for theseparameters for each user From the QQ-plot in Figure 29we notice that the model predicts a greater probability ofseeing some very large offtimes than are observed in thedatasets However the probability of seeing these large off-times is small there are 27 data points that have a y-value greater than 8000 in that graph Hence we believethat Weibull provides a good fit for the length of intervalsbetween interactions
Figure 30 shows the distribution of the estimated shapeand scale of the fitted Weibull distributions Interestinglythe shape is consistently less than one Weibull shape valuesless than one indicate that the longer the screen has been offthe less likely it is for it to be turned on by the user This be-havior has interesting implications for power saving policiesFor instance periodic activities such as checking for emailwhen the screen has been off for a while may be deferred orrescheduled if needed without hurting user experience
020
4060
8010
0
0 20 40 60 80 100User percentile
MS
E (
)
(a) MSE
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed r
ate
(b) Rate parameter
Figure 31 (a) The mean square error (MSE) whenapplication popularity distribution is modeled usingan exponential (b) The inferred rate parameter ofthe exponential distribution for different users
83 Application PopularityWe find that for each user the relative popularity of ap-
plications can be well described by a simple exponential dis-tribution This qualitative invariant is useful for instanceto predict how many applications account for a given frac-tion of user attention For the example users in Figure 11this facet can be seen in the inset plots the semi-log of thepopularity distribution is very close to a straight line
Figure 31(a) shows that this exponential drop in applica-tion popularity is true for almost all users the mean squareerror between modeled exponential and actual popularitydistribution is less than 5 for 95 of the users
Figure 31(b) shows the inferred rate parameter of the ap-plication popularity distribution for various users We seethat the rate varies by an order of magnitude from 01 toalmost 1 The value of the rate essentially captures the paceof the drop in application popularity Lower values describeusers that use more applications on a regular basis Oneimplication of the wide range is that it may be feasible toretain all popular applications in memory for some users andnot for others
9 PREDICTING ENERGY DRAINIn this section we demonstrate the value of adapting to
user behaviors in the context of a mechanism to predict fu-ture energy drain on the smartphone Such a mechanismcan help with scheduling background tasks [9] estimatingbattery lifetime and improving user experience if it is esti-mated that the user will not fully drain the battery until thenext charging cycle [2 21]
Despite its many uses to our knowledge none of thesmartphones today provide a prediction of future energydrain Providing an accurate prediction of energy drain isdifficult2 User diversity of energy use makes any static pre-diction method highly inaccurate Even for the same userthere is a high variance in energy usage Figure 32 shows thisvariance by plotting the ratio of the standard deviation tothe mean energy drain over several time periods for users inDataset 1 The standard deviation of 10-minute windows ishigher than three times the mean for a fifth of the users The
2Energy drain prediction on smartphones is more challeng-ing than what is done for laptops Laptops provide a pre-diction of battery lifetime only for active use Smartphonepredictions on the other hand must cover multiple activeidle periods to be useful
01
23
45
0 20 40 60 80 100User percentile
Err
or r
atio
2 hours1 hour10 minutes
Figure 32 Battery drain predictors have to copewith the high variance in drain
time
window to
predict
n chunks
hellip
Trend Table
hellip
hellip
Figure 33 A userrsquos trend table captures how batterydrain in n consecutive chunks relates to drain in thesubsequent window
bursty nature of user interactions is one cause of this highvariance The variance is high even at longer time scalesThe standard deviation for two hour windows is larger than150 of the mean for half the users The diurnal patternsthat we showed earlier are one cause of high variance oversuch large time scales
Although the variance is high there are patterns in auserrsquos behaviors For instance we showed earlier that aphone that has been idle for a while is likely to remain idlein the near future and it will thus continue to draw energyat a similar rate
We hypothesize that a predictor tuned to user behaviorcan be accurate We present a simple personalized energydrain predictor that incorporates various user-specific fac-tors such as length of idle periods and different types ofbusy periods Instead of explicitly identifying these factorsour predictor captures them in a ldquotrend tablerdquo framework
Our framework is shown in Figure 33 Each entry in thetrend table is indexed by an n-tuple that represents energyusage readings from adjacent chunks of size δ each Thisindex points to energy usage statistics across all time win-dows of size w in the history that follow the n chunk valuesrepresented by that index Thus the trend table captureshow the energy usage in n consecutive chunks is related tothe energy use in the subsequent time windows of size w
In this paper we use n = 3 δ = 10 minutes and maintaindifferent trend tables for each window size w for which theprediction is required To keep the trend table small wequantize the energy readings of chunks We maintain themean and standard deviation for all subsequent windowsthat map to that index
To make a prediction we use the energy readings fromthe n immediately preceding chunks Let these readings be(x1 xn) Then we search for indices in the table thatdiffer from (x1 xn) by no more than a threshold s and
GenericShortminustermTimeminusofminusdayPersonalized
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
(a) 1-hour window
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
GenericShortminustermTimeminusofminusdayPersonalized
(b) 2-hour window
Figure 34 The ratio of error to true battery drainfor different energy drain predictors
predict based on them such that similar indices have higherrelative weight More precisely the prediction is
X
e stforalli=1n|he
iminusxi|lts
ke times ue
P
ke
where e iterates over all tuples in the trend table hei is
the value of the chunk i of the tuple e ue represents thestatistics stored for that index and the relative weight ke =s minus maxi=1n |he
i minus xi| For results shown here we uses = 05
Figures 34 shows how well our Personalized predictor worksfor predicting usage for two different future time windowsThere is one point per user in each graph which correspondsto the median error seen across many predictions for thatuser We see that the 90th percentile error is 25 for 1-hourwindow and 40 for 2-hour window
To place this level of accuracy in context the figure alsoshows the accurary of three alternative predictors TheGeneric predictor uses a trend table but instead of buildinga user-specific table it builds one that combines all usersWe see that its 90th percentile error is roughly twice thatof the Personalized predictor Its predictions are worse forhalf of the users
The Short-term predictor predicts energy drain simply asthe amount drained in the immediately preceding windowof the same size For both time windows its 90th percentileerror is no better than that of the Generic predictor
The Time-of-day predictor predicts energy drain based onvalues observed for the user at the same time of the day inthe past similar to that proposed by Banerjee et al [2]Thus this predictor has some degree of personalization butit learns user behavior in less detail than the Personalizedpredictor We see that its performance for 1-hour windowis similar to the Short-term predictor For 2-hour windowits error is higher than the Personalized predictor across theboard though its worst case error is lower than the Short-term and Generic predictors
Overall these results demonstrate the value of appropri-ately learning user behaviors to design intelligent mecha-nisms on smartphones
10 RELATED WORKIn a range of domains there is a rich history of work that
characterizes user workloads However because smartphoneadoption has gathered pace relatively recently our work rep-resents one of the few to study how people use smartphones
Along with other recent works it helps complete the pic-ture Banerjee et al and Rahmati et al report on batterycharging behaviors [2 18] Like us they find considerablevariation among users Banerjee et al also propose a pre-dictor that estimates the excess energy of the battery atthe time of charging using a histogram of past battery us-age [2] Shye et al study the power consumption character-istics of 20 users [24] They infer properties such as whichcomponents in the phone consume most power and exploreoptimizations based on these characteristics Rahmati andZhong study 14 users of a particular demographic to studywhich applications are popular in that demographic wherethe phone is used and how it is shared among users [19]
In contrast to these works we focus on understandingdifferent aspects of smartphone use (eg interactions andtraffic) and on exposing the diversity of user behaviors in-stead of only the average behavior Our study also entailsan order of magnitude more users than previous efforts
Like us other researchers have developed logging utilitiesMyExperience is one such early utility [10] and the worksabove involve custom utilities as well Our focus in thispaper is not on the development of logging tools but onanalyzing their data to gain insight into user behavior
There is a body of work in modeling the aggregate be-havior of mobile users Using traces from a large cellularoperator some network related aspects of mobile usage havebeen modeled Halepovic et al and Williamson et al re-port that call arrivals are bursty and present diurnal pat-terns [28] Willkomm et al and Brown et al report thatmobile users call duration can be approximately modeled bya lognormal distribution [4 29] We use traces collected onthe mobile device itself and focus on modeling the interac-tions of individual users instead of the aggregate behavior
11 CONCLUSIONSBy studying 255 users of two different smartphone plat-
forms we comprehensively characterized user activities andtheir impact on network and battery We quantify manyhitherto unknown aspects of smartphone usage User diver-sity is an overarching theme in our findings For instancedifferent users interact with their phones 10-200 times a dayon average the mean interaction length of different users
is 10-250 seconds and users receive 1-1000 MB of data perday where 10-90 is received as part of active use
This extent of user diversity implies that mechanisms thatwork for the average case may be ineffective for a large frac-tion of the users Instead learning and adapting to userbehaviors is likely to be more effective as demonstrated byour personalized energy drain predictor We show that qual-itative similarities exist among users to facilitate the devel-opment of such mechanisms For instance the longer theuser has not interacted with the phone the less likely it isfor her to start interacting with it and application popular-ity for a user follows a simple exponential distribution
Our study points to ways in which smartphone platformsshould be enhanced Effective adaptation will require futureplatforms to support light-weight tools that monitor andlearn user behaviors in situ It will also require them toexpose appropriate knobs to control the behavior of lower-level components (eg CPU or radio)
12 ACKNOWLEDGEMENTSWe are grateful to the users in Dataset1 that ran our log-
ger and to the providers of Dataset2 Elizabeth Dawson pro-vided administrative support for data collection Jean BolotKurt Partridge Stefan Saroiu Lin Zhong and the MobiSysreviewers provided many comments that helped improve thepaper We thank them all This work is supported in partby the NSF grants CCR-0120778 and CNS-0627084
13 REFERENCES[1] Balasubramanian N Balasubramanian A
and Venkataramani A Energy consumption inmobile phones A measurement study andimplications for network applications In IMC (2009)
[2] Banerjee N Rahmati A Corner M D
Rollins S and Zhong L Users and batteriesInteractions and adaptive energy management inmobile systems In Ubicomp (2007)
[3] Becker R Chambers J and Wilks A The newS language Chapman amp HallCRC 1988
[4] Brown L Gans N Mandelbaum A Sakov
A Shen H Zeltyn S and Zhao L Statisticalanalysis of a telephone call center Journal of theAmerican Statistical Association 100 469 (2005)
[5] Church K and Smyth B Understanding mobileinformation needs In MobileHCI (2008)
[6] Dempster A Laird N and Rubin D Maximumlikelihood from incomplete data via the EM algorithmJournal of the Royal Statistical Society Series B(Methodological) 39 1 (1977)
[7] Esfahbod B Preload An adaptive prefetchingdaemon Masterrsquos thesis University of Toronto 2006
[8] Falaki H Govindan R and Estrin D Smartscreen management on mobile phones Tech Rep 74Center for Embedded Networked Sesning 2009
[9] Flinn J and Satyanarayanan M Managingbattery lifetime with energy-aware adaptation ACMTOCS 22 2 (2004)
[10] Froehlich J Chen M Consolvo S
Harrison B and Landay J MyExperience Asystem for in situ tracing and capturing of userfeedback on mobile phones In MobiSys (2007)
[11] Fukuda K Cho K and Esaki H The impact ofresidential broadband traffic on Japanese ISPbackbones SIGCOMM CCR 35 1 (2005)
[12] Kotz D and Essien K Analysis of a campus-widewireless network Wireless Networks 11 1ndash2 (2005)
[13] Kwiatkowski D Phillips P Schmidt P and
Shin Y Testing the null hypothesis of stationarityagainst the alternative of a unit root Journal ofEconometrics 54 1-3 (1992)
[14] Lahiri K Dey S Panigrahi D and
Raghunathan A Battery-driven system design Anew frontier in low power design In Asia SouthPacific design automationVLSI Design (2002)
[15] Ljung G and Box G On a measure of lack of fitin time series models Biometrika 65 2 (1978)
[16] Smartphone futures 2009-2014 httpwwwportioresearchcomSmartphone09-14html
[17] Rahmati A Qian A and Zhong L
Understanding human-battery interaction on mobilephones In MobileHCI (2007)
[18] Rahmati A and Zhong L Humanndashbatteryinteraction on mobile phones Pervasive and MobileComputing 5 5 (2009)
[19] Rahmati A and Zhong L A longitudinal study ofnon-voice mobile phone usage by teens from anunderserved urban community Tech Rep 0515-09Rice University 2009
[20] Rao R Vrudhula S and Rakhmatov D
Battery modeling for energy aware system designIEEE Computer 36 12 (2003)
[21] Ravi N Scott J Han L and Iftode L
Context-aware battery management for mobilephones In PerCom (2008)
[22] Sharma A Navda V Ramjee R
Padmanabhan V and Belding E Cool-TetherEnergy efficient on-the-fly WiFi hot-spots usingmobile phones In CoNEXT (2009)
[23] Sharma C US wireless data market update - Q32009 httpwwwchetansharmacomusmarketupdateq309htm
[24] Shye A Sholbrock B and G M Into the wildStudying real user activity patterns to guide poweroptimization for mobile architectures In Micro (2009)
[25] Snol L More smartphones than desktop PCs by2011 httpwwwpcworldcomarticle171380more_smartphones_than_desktop_pcs_by_2011
html
[26] Sohn T Li K Griswold W and Hollan J
A diary study of mobile information needs In SIGCHI(2008)
[27] Trestian I Ranjan S Kuzmanovic A and
Nucci A Measuring serendipity Connecting peoplelocations and interests in a mobile 3G network InIMC (2009)
[28] Williamson C Halepovic E Sun H and Wu
Y Characterization of CDMA2000 Cellular DataNetwork Traffic In Local Computer Networks (2005)
[29] Willkomm D Machiraju S Bolot J and
Wolisz A Primary users in cellular networks Alarge-scale measurement study In DySPAN (2008)
- Introduction
- Data Collection
-
- Dataset1
- Dataset2
- Representativeness of conclusions
-
- User Interactions
-
- Interaction Time
- Interaction Sessions
- Diurnal Patterns
-
- Application Usage
-
- Number of applications
- Application Popularity
- Application Sessions
-
- Traffic
-
- Traffic per day
- ``Interactive traffic
- Diurnal patterns
-
- Energy Consumption
- Implications of User Diversity
- Smartphone Usage Models
-
- Session Lengths
- Time between Sessions
- Application Popularity
-
- Predicting Energy Drain
- Related Work
- Conclusions
- Acknowledgements
- References
-
2060
100
140
0 4 8 12 16 20Hour of the day
Bat
tery
Dra
in (
mA
h)
(a)
0 20 40 60 80 100
12
34
56
7
User percentile
Diu
rnal
rat
io(b)
Figure 25 (a) The mean and 95 CI of energy drainof an example Dataset1 user (b) Diurnal ratio ofall users in Dataset1
Our results show that user activities contribute heavilytowards energy drain users in Dataset1 who have identi-cal smartphones drain energy at different rates and energydrain has diurnal patterns In the future we will developmethods to accurately quantify the energy consumption ofthe platform from that due to user-induced workload
7 IMPLICATIONS OF USER DIVERSITYWe uncover a surprising level of diversity among smart-
phone users For almost every aspect of usage that we studywe find one or more orders of magnitude difference betweenusers Our findings strongly motivate the need for customiz-ing smartphones to their users We believe that this needis greater than that for customizing ordinary cellphones orlaptops Ordinary cellphones do not have as rich an appli-cation environment Laptops are not as portable and aremore resource rich For example many users plug-in theirlaptops while using them
Customization can help at all levels Consider somethingas low-level as the battery Suppose we want batteries to beboth lightweight and last for at least a day with a high proba-bility Meeting the latter goal for all users of a given platformwill require catering to the heaviest users But that will leadto unnecessarily heavy batteries for many users (Highercapacity batteries are heavier) Offering multiple types ofbatteries with different lifetime-weight tradeoffs provides away out of this bind
At levels where intelligent mechanisms to improve user ex-perience or reduce energy consumption reside user diversitymotivates adapting to the smartphone user Driving thesemechanisms based on average case behaviors may not beeffective for a large fraction of the users
The ease and utility of customization depends on twoproperties of user behavior First despite quantitative dif-ferences there must be qualitative similarities among usersFor instance we should be able to describe the behavior ofall users with the same model Different users may havedifferent parameters of this model which will then lead toquantitative differences among them The presence of qual-itative similarities imply that users are not arbitrary pointsin space and it significantly simplifies the task of learninguser behavior Second user behavior in the past must alsobe predictive of the future Otherwise customization basedon past behaviors will be of little value in the future
In the next two sections we present evidence that theseproperties hold for several key aspects of smartphone usage
Dataset1 example user
000
00
010
002
00
030
0 100 200 300 400Session length (s)
Den
sity
Dataset2 example user
000
005
010
015
020
0 20 40 60 80 100Session length (s)
Den
sity
Figure 26 The PDF of session length for sampleusers of each dataset
In sect8 we show that user sessions and relative applicationpopularity can be described using simple models In sect9 weshow that energy drain can be described as well as predictedusing a ldquotrend trablerdquo framework
8 SMARTPHONE USAGE MODELSIn this section we develop simple models that describe
three aspects of smartphone usage ndash session lengths inter-arrival time between sessions and application popularityThe models are common across users but have different pa-rameters for different users While they do not completelydescribe user behavior they capture first order factors andrepresent a first step towards more complete modeling ofsmartphone usage More importantly along with the resultsof the next section they show that qualitative similaritiesdo exist among users
81 Session LengthsWe first consider the statistical properties of session length
distributions of users We find that session length valuestend to stationary With the KPSS test for level stationar-ity [13] 90 of the users have a p-value greater than 01The presence of stationarity is appealing because it suggeststhat past behavior is capable of predicting the future
We also find that session lengths are independent that isthe current value does not have a strong correlation with thevalues seen in the recent past With the Ljung-Box test forindependence [15] 96 of the users have a p-value that isgreater than 01
Stationarity and independence considered together sug-gest that session length values can be modeled as iid sam-ples from a distribution Choosing an appropriate distribu-tion however is complicated by the nature of the sessionlengths Most sessions are very short and the frequencydrops exponentially as the length increases However in-consistent with exponential behavior there are some verylong sessions in the tail for each user
We find that a mixture of exponential and Pareto distri-butions can model both ends of the spectrum The formercaptures short sessions and the latter captures long sessionsThat is session lengths can be described by the followingmixture model
r middot Exp(λ) + (1 minus r) middot Pareto(xm α)
In this equation r is the relative mix of the two distribu-tions λ is the rate of the exponential and xm and α are thelocation and shape parameters of the Pareto distribution
The location for a Pareto distribution represents the min-imum possible value of the random variable The location
010
0020
0030
00
0 1000 2000 3000Observed quantiles
Mod
el q
uant
iles
Figure 27 QQ plot of session lengths model for asample user
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed m
ix
(a) Relative mix (r)
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed r
ate
(b) Exponential rate (λ)
020
4060
0 20 40 60 80 100User percentile
Tim
eout
(s)
(c) Pareto location (xm)
minus1
0minus
04
02
06
10
0 20 40 60 80 100User percentile
Est
imat
ed s
hape
(d) Pareto shape (α)
Figure 28 Distribution of inferred model parame-ters that describe session length values of users inboth datasets
value that offers the best fit is the screen timeout value ofthe user because the session length PDF has a spike at thisvalue The spike corresponds to short sessions that are endedby the timeout (when the user forgets to switch the screenoff) we confirm this using controlled experiments with dif-ferent timeout values Figure 26 shows this spike at 60 and15 seconds for example users from each dataset The time-out provides a natural division between the two componentdistributions We automatically infer its approximate valueusing a simple spike detection algorithm
We use the EM algorithm to infer the maximum likelihoodestimation (MLE) of the remaining three parameters [6]Figure 27 shows the quality of this fit for an example userusing the QQ plot [3] Almost all quantiles are along they = x line indicating a good fit
Figure 28 shows the four inferred parameters for varioususers While users can be modeled using the same mix-
040
0080
00
0 4000 8000Observed quantiles
Mod
el q
uant
iles
Figure 29 QQ plot of session offtime model for asample user
050
010
0015
000 20 40 60 80 100
User percentileE
stim
ated
sca
le
(a) Weibull scale
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed s
hape
(b) Weibull shape
Figure 30 Distribution of inferred model parame-ters that describe the distribution of time betweensessions for users in both datasets
ture model the parameters of this model vary widely acrossusers Because of the way we construct our model the dis-tribution of the parameter r and xm also provide insightinto how frequently usersrsquo screen is switched off by the time-out and the relative popularity of different timeout values60 seconds is the most popular value likely because it is themost common default timeout and many users never changethe default
82 Time between SessionsWe find that the Weibull distribution can explain the
screen off times This distribution has two parameters re-ferred to as its scale and shape We find the MLE for theseparameters for each user From the QQ-plot in Figure 29we notice that the model predicts a greater probability ofseeing some very large offtimes than are observed in thedatasets However the probability of seeing these large off-times is small there are 27 data points that have a y-value greater than 8000 in that graph Hence we believethat Weibull provides a good fit for the length of intervalsbetween interactions
Figure 30 shows the distribution of the estimated shapeand scale of the fitted Weibull distributions Interestinglythe shape is consistently less than one Weibull shape valuesless than one indicate that the longer the screen has been offthe less likely it is for it to be turned on by the user This be-havior has interesting implications for power saving policiesFor instance periodic activities such as checking for emailwhen the screen has been off for a while may be deferred orrescheduled if needed without hurting user experience
020
4060
8010
0
0 20 40 60 80 100User percentile
MS
E (
)
(a) MSE
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed r
ate
(b) Rate parameter
Figure 31 (a) The mean square error (MSE) whenapplication popularity distribution is modeled usingan exponential (b) The inferred rate parameter ofthe exponential distribution for different users
83 Application PopularityWe find that for each user the relative popularity of ap-
plications can be well described by a simple exponential dis-tribution This qualitative invariant is useful for instanceto predict how many applications account for a given frac-tion of user attention For the example users in Figure 11this facet can be seen in the inset plots the semi-log of thepopularity distribution is very close to a straight line
Figure 31(a) shows that this exponential drop in applica-tion popularity is true for almost all users the mean squareerror between modeled exponential and actual popularitydistribution is less than 5 for 95 of the users
Figure 31(b) shows the inferred rate parameter of the ap-plication popularity distribution for various users We seethat the rate varies by an order of magnitude from 01 toalmost 1 The value of the rate essentially captures the paceof the drop in application popularity Lower values describeusers that use more applications on a regular basis Oneimplication of the wide range is that it may be feasible toretain all popular applications in memory for some users andnot for others
9 PREDICTING ENERGY DRAINIn this section we demonstrate the value of adapting to
user behaviors in the context of a mechanism to predict fu-ture energy drain on the smartphone Such a mechanismcan help with scheduling background tasks [9] estimatingbattery lifetime and improving user experience if it is esti-mated that the user will not fully drain the battery until thenext charging cycle [2 21]
Despite its many uses to our knowledge none of thesmartphones today provide a prediction of future energydrain Providing an accurate prediction of energy drain isdifficult2 User diversity of energy use makes any static pre-diction method highly inaccurate Even for the same userthere is a high variance in energy usage Figure 32 shows thisvariance by plotting the ratio of the standard deviation tothe mean energy drain over several time periods for users inDataset 1 The standard deviation of 10-minute windows ishigher than three times the mean for a fifth of the users The
2Energy drain prediction on smartphones is more challeng-ing than what is done for laptops Laptops provide a pre-diction of battery lifetime only for active use Smartphonepredictions on the other hand must cover multiple activeidle periods to be useful
01
23
45
0 20 40 60 80 100User percentile
Err
or r
atio
2 hours1 hour10 minutes
Figure 32 Battery drain predictors have to copewith the high variance in drain
time
window to
predict
n chunks
hellip
Trend Table
hellip
hellip
Figure 33 A userrsquos trend table captures how batterydrain in n consecutive chunks relates to drain in thesubsequent window
bursty nature of user interactions is one cause of this highvariance The variance is high even at longer time scalesThe standard deviation for two hour windows is larger than150 of the mean for half the users The diurnal patternsthat we showed earlier are one cause of high variance oversuch large time scales
Although the variance is high there are patterns in auserrsquos behaviors For instance we showed earlier that aphone that has been idle for a while is likely to remain idlein the near future and it will thus continue to draw energyat a similar rate
We hypothesize that a predictor tuned to user behaviorcan be accurate We present a simple personalized energydrain predictor that incorporates various user-specific fac-tors such as length of idle periods and different types ofbusy periods Instead of explicitly identifying these factorsour predictor captures them in a ldquotrend tablerdquo framework
Our framework is shown in Figure 33 Each entry in thetrend table is indexed by an n-tuple that represents energyusage readings from adjacent chunks of size δ each Thisindex points to energy usage statistics across all time win-dows of size w in the history that follow the n chunk valuesrepresented by that index Thus the trend table captureshow the energy usage in n consecutive chunks is related tothe energy use in the subsequent time windows of size w
In this paper we use n = 3 δ = 10 minutes and maintaindifferent trend tables for each window size w for which theprediction is required To keep the trend table small wequantize the energy readings of chunks We maintain themean and standard deviation for all subsequent windowsthat map to that index
To make a prediction we use the energy readings fromthe n immediately preceding chunks Let these readings be(x1 xn) Then we search for indices in the table thatdiffer from (x1 xn) by no more than a threshold s and
GenericShortminustermTimeminusofminusdayPersonalized
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
(a) 1-hour window
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
GenericShortminustermTimeminusofminusdayPersonalized
(b) 2-hour window
Figure 34 The ratio of error to true battery drainfor different energy drain predictors
predict based on them such that similar indices have higherrelative weight More precisely the prediction is
X
e stforalli=1n|he
iminusxi|lts
ke times ue
P
ke
where e iterates over all tuples in the trend table hei is
the value of the chunk i of the tuple e ue represents thestatistics stored for that index and the relative weight ke =s minus maxi=1n |he
i minus xi| For results shown here we uses = 05
Figures 34 shows how well our Personalized predictor worksfor predicting usage for two different future time windowsThere is one point per user in each graph which correspondsto the median error seen across many predictions for thatuser We see that the 90th percentile error is 25 for 1-hourwindow and 40 for 2-hour window
To place this level of accuracy in context the figure alsoshows the accurary of three alternative predictors TheGeneric predictor uses a trend table but instead of buildinga user-specific table it builds one that combines all usersWe see that its 90th percentile error is roughly twice thatof the Personalized predictor Its predictions are worse forhalf of the users
The Short-term predictor predicts energy drain simply asthe amount drained in the immediately preceding windowof the same size For both time windows its 90th percentileerror is no better than that of the Generic predictor
The Time-of-day predictor predicts energy drain based onvalues observed for the user at the same time of the day inthe past similar to that proposed by Banerjee et al [2]Thus this predictor has some degree of personalization butit learns user behavior in less detail than the Personalizedpredictor We see that its performance for 1-hour windowis similar to the Short-term predictor For 2-hour windowits error is higher than the Personalized predictor across theboard though its worst case error is lower than the Short-term and Generic predictors
Overall these results demonstrate the value of appropri-ately learning user behaviors to design intelligent mecha-nisms on smartphones
10 RELATED WORKIn a range of domains there is a rich history of work that
characterizes user workloads However because smartphoneadoption has gathered pace relatively recently our work rep-resents one of the few to study how people use smartphones
Along with other recent works it helps complete the pic-ture Banerjee et al and Rahmati et al report on batterycharging behaviors [2 18] Like us they find considerablevariation among users Banerjee et al also propose a pre-dictor that estimates the excess energy of the battery atthe time of charging using a histogram of past battery us-age [2] Shye et al study the power consumption character-istics of 20 users [24] They infer properties such as whichcomponents in the phone consume most power and exploreoptimizations based on these characteristics Rahmati andZhong study 14 users of a particular demographic to studywhich applications are popular in that demographic wherethe phone is used and how it is shared among users [19]
In contrast to these works we focus on understandingdifferent aspects of smartphone use (eg interactions andtraffic) and on exposing the diversity of user behaviors in-stead of only the average behavior Our study also entailsan order of magnitude more users than previous efforts
Like us other researchers have developed logging utilitiesMyExperience is one such early utility [10] and the worksabove involve custom utilities as well Our focus in thispaper is not on the development of logging tools but onanalyzing their data to gain insight into user behavior
There is a body of work in modeling the aggregate be-havior of mobile users Using traces from a large cellularoperator some network related aspects of mobile usage havebeen modeled Halepovic et al and Williamson et al re-port that call arrivals are bursty and present diurnal pat-terns [28] Willkomm et al and Brown et al report thatmobile users call duration can be approximately modeled bya lognormal distribution [4 29] We use traces collected onthe mobile device itself and focus on modeling the interac-tions of individual users instead of the aggregate behavior
11 CONCLUSIONSBy studying 255 users of two different smartphone plat-
forms we comprehensively characterized user activities andtheir impact on network and battery We quantify manyhitherto unknown aspects of smartphone usage User diver-sity is an overarching theme in our findings For instancedifferent users interact with their phones 10-200 times a dayon average the mean interaction length of different users
is 10-250 seconds and users receive 1-1000 MB of data perday where 10-90 is received as part of active use
This extent of user diversity implies that mechanisms thatwork for the average case may be ineffective for a large frac-tion of the users Instead learning and adapting to userbehaviors is likely to be more effective as demonstrated byour personalized energy drain predictor We show that qual-itative similarities exist among users to facilitate the devel-opment of such mechanisms For instance the longer theuser has not interacted with the phone the less likely it isfor her to start interacting with it and application popular-ity for a user follows a simple exponential distribution
Our study points to ways in which smartphone platformsshould be enhanced Effective adaptation will require futureplatforms to support light-weight tools that monitor andlearn user behaviors in situ It will also require them toexpose appropriate knobs to control the behavior of lower-level components (eg CPU or radio)
12 ACKNOWLEDGEMENTSWe are grateful to the users in Dataset1 that ran our log-
ger and to the providers of Dataset2 Elizabeth Dawson pro-vided administrative support for data collection Jean BolotKurt Partridge Stefan Saroiu Lin Zhong and the MobiSysreviewers provided many comments that helped improve thepaper We thank them all This work is supported in partby the NSF grants CCR-0120778 and CNS-0627084
13 REFERENCES[1] Balasubramanian N Balasubramanian A
and Venkataramani A Energy consumption inmobile phones A measurement study andimplications for network applications In IMC (2009)
[2] Banerjee N Rahmati A Corner M D
Rollins S and Zhong L Users and batteriesInteractions and adaptive energy management inmobile systems In Ubicomp (2007)
[3] Becker R Chambers J and Wilks A The newS language Chapman amp HallCRC 1988
[4] Brown L Gans N Mandelbaum A Sakov
A Shen H Zeltyn S and Zhao L Statisticalanalysis of a telephone call center Journal of theAmerican Statistical Association 100 469 (2005)
[5] Church K and Smyth B Understanding mobileinformation needs In MobileHCI (2008)
[6] Dempster A Laird N and Rubin D Maximumlikelihood from incomplete data via the EM algorithmJournal of the Royal Statistical Society Series B(Methodological) 39 1 (1977)
[7] Esfahbod B Preload An adaptive prefetchingdaemon Masterrsquos thesis University of Toronto 2006
[8] Falaki H Govindan R and Estrin D Smartscreen management on mobile phones Tech Rep 74Center for Embedded Networked Sesning 2009
[9] Flinn J and Satyanarayanan M Managingbattery lifetime with energy-aware adaptation ACMTOCS 22 2 (2004)
[10] Froehlich J Chen M Consolvo S
Harrison B and Landay J MyExperience Asystem for in situ tracing and capturing of userfeedback on mobile phones In MobiSys (2007)
[11] Fukuda K Cho K and Esaki H The impact ofresidential broadband traffic on Japanese ISPbackbones SIGCOMM CCR 35 1 (2005)
[12] Kotz D and Essien K Analysis of a campus-widewireless network Wireless Networks 11 1ndash2 (2005)
[13] Kwiatkowski D Phillips P Schmidt P and
Shin Y Testing the null hypothesis of stationarityagainst the alternative of a unit root Journal ofEconometrics 54 1-3 (1992)
[14] Lahiri K Dey S Panigrahi D and
Raghunathan A Battery-driven system design Anew frontier in low power design In Asia SouthPacific design automationVLSI Design (2002)
[15] Ljung G and Box G On a measure of lack of fitin time series models Biometrika 65 2 (1978)
[16] Smartphone futures 2009-2014 httpwwwportioresearchcomSmartphone09-14html
[17] Rahmati A Qian A and Zhong L
Understanding human-battery interaction on mobilephones In MobileHCI (2007)
[18] Rahmati A and Zhong L Humanndashbatteryinteraction on mobile phones Pervasive and MobileComputing 5 5 (2009)
[19] Rahmati A and Zhong L A longitudinal study ofnon-voice mobile phone usage by teens from anunderserved urban community Tech Rep 0515-09Rice University 2009
[20] Rao R Vrudhula S and Rakhmatov D
Battery modeling for energy aware system designIEEE Computer 36 12 (2003)
[21] Ravi N Scott J Han L and Iftode L
Context-aware battery management for mobilephones In PerCom (2008)
[22] Sharma A Navda V Ramjee R
Padmanabhan V and Belding E Cool-TetherEnergy efficient on-the-fly WiFi hot-spots usingmobile phones In CoNEXT (2009)
[23] Sharma C US wireless data market update - Q32009 httpwwwchetansharmacomusmarketupdateq309htm
[24] Shye A Sholbrock B and G M Into the wildStudying real user activity patterns to guide poweroptimization for mobile architectures In Micro (2009)
[25] Snol L More smartphones than desktop PCs by2011 httpwwwpcworldcomarticle171380more_smartphones_than_desktop_pcs_by_2011
html
[26] Sohn T Li K Griswold W and Hollan J
A diary study of mobile information needs In SIGCHI(2008)
[27] Trestian I Ranjan S Kuzmanovic A and
Nucci A Measuring serendipity Connecting peoplelocations and interests in a mobile 3G network InIMC (2009)
[28] Williamson C Halepovic E Sun H and Wu
Y Characterization of CDMA2000 Cellular DataNetwork Traffic In Local Computer Networks (2005)
[29] Willkomm D Machiraju S Bolot J and
Wolisz A Primary users in cellular networks Alarge-scale measurement study In DySPAN (2008)
- Introduction
- Data Collection
-
- Dataset1
- Dataset2
- Representativeness of conclusions
-
- User Interactions
-
- Interaction Time
- Interaction Sessions
- Diurnal Patterns
-
- Application Usage
-
- Number of applications
- Application Popularity
- Application Sessions
-
- Traffic
-
- Traffic per day
- ``Interactive traffic
- Diurnal patterns
-
- Energy Consumption
- Implications of User Diversity
- Smartphone Usage Models
-
- Session Lengths
- Time between Sessions
- Application Popularity
-
- Predicting Energy Drain
- Related Work
- Conclusions
- Acknowledgements
- References
-
010
0020
0030
00
0 1000 2000 3000Observed quantiles
Mod
el q
uant
iles
Figure 27 QQ plot of session lengths model for asample user
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed m
ix
(a) Relative mix (r)
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed r
ate
(b) Exponential rate (λ)
020
4060
0 20 40 60 80 100User percentile
Tim
eout
(s)
(c) Pareto location (xm)
minus1
0minus
04
02
06
10
0 20 40 60 80 100User percentile
Est
imat
ed s
hape
(d) Pareto shape (α)
Figure 28 Distribution of inferred model parame-ters that describe session length values of users inboth datasets
value that offers the best fit is the screen timeout value ofthe user because the session length PDF has a spike at thisvalue The spike corresponds to short sessions that are endedby the timeout (when the user forgets to switch the screenoff) we confirm this using controlled experiments with dif-ferent timeout values Figure 26 shows this spike at 60 and15 seconds for example users from each dataset The time-out provides a natural division between the two componentdistributions We automatically infer its approximate valueusing a simple spike detection algorithm
We use the EM algorithm to infer the maximum likelihoodestimation (MLE) of the remaining three parameters [6]Figure 27 shows the quality of this fit for an example userusing the QQ plot [3] Almost all quantiles are along they = x line indicating a good fit
Figure 28 shows the four inferred parameters for varioususers While users can be modeled using the same mix-
040
0080
00
0 4000 8000Observed quantiles
Mod
el q
uant
iles
Figure 29 QQ plot of session offtime model for asample user
050
010
0015
000 20 40 60 80 100
User percentileE
stim
ated
sca
le
(a) Weibull scale
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed s
hape
(b) Weibull shape
Figure 30 Distribution of inferred model parame-ters that describe the distribution of time betweensessions for users in both datasets
ture model the parameters of this model vary widely acrossusers Because of the way we construct our model the dis-tribution of the parameter r and xm also provide insightinto how frequently usersrsquo screen is switched off by the time-out and the relative popularity of different timeout values60 seconds is the most popular value likely because it is themost common default timeout and many users never changethe default
82 Time between SessionsWe find that the Weibull distribution can explain the
screen off times This distribution has two parameters re-ferred to as its scale and shape We find the MLE for theseparameters for each user From the QQ-plot in Figure 29we notice that the model predicts a greater probability ofseeing some very large offtimes than are observed in thedatasets However the probability of seeing these large off-times is small there are 27 data points that have a y-value greater than 8000 in that graph Hence we believethat Weibull provides a good fit for the length of intervalsbetween interactions
Figure 30 shows the distribution of the estimated shapeand scale of the fitted Weibull distributions Interestinglythe shape is consistently less than one Weibull shape valuesless than one indicate that the longer the screen has been offthe less likely it is for it to be turned on by the user This be-havior has interesting implications for power saving policiesFor instance periodic activities such as checking for emailwhen the screen has been off for a while may be deferred orrescheduled if needed without hurting user experience
020
4060
8010
0
0 20 40 60 80 100User percentile
MS
E (
)
(a) MSE
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed r
ate
(b) Rate parameter
Figure 31 (a) The mean square error (MSE) whenapplication popularity distribution is modeled usingan exponential (b) The inferred rate parameter ofthe exponential distribution for different users
83 Application PopularityWe find that for each user the relative popularity of ap-
plications can be well described by a simple exponential dis-tribution This qualitative invariant is useful for instanceto predict how many applications account for a given frac-tion of user attention For the example users in Figure 11this facet can be seen in the inset plots the semi-log of thepopularity distribution is very close to a straight line
Figure 31(a) shows that this exponential drop in applica-tion popularity is true for almost all users the mean squareerror between modeled exponential and actual popularitydistribution is less than 5 for 95 of the users
Figure 31(b) shows the inferred rate parameter of the ap-plication popularity distribution for various users We seethat the rate varies by an order of magnitude from 01 toalmost 1 The value of the rate essentially captures the paceof the drop in application popularity Lower values describeusers that use more applications on a regular basis Oneimplication of the wide range is that it may be feasible toretain all popular applications in memory for some users andnot for others
9 PREDICTING ENERGY DRAINIn this section we demonstrate the value of adapting to
user behaviors in the context of a mechanism to predict fu-ture energy drain on the smartphone Such a mechanismcan help with scheduling background tasks [9] estimatingbattery lifetime and improving user experience if it is esti-mated that the user will not fully drain the battery until thenext charging cycle [2 21]
Despite its many uses to our knowledge none of thesmartphones today provide a prediction of future energydrain Providing an accurate prediction of energy drain isdifficult2 User diversity of energy use makes any static pre-diction method highly inaccurate Even for the same userthere is a high variance in energy usage Figure 32 shows thisvariance by plotting the ratio of the standard deviation tothe mean energy drain over several time periods for users inDataset 1 The standard deviation of 10-minute windows ishigher than three times the mean for a fifth of the users The
2Energy drain prediction on smartphones is more challeng-ing than what is done for laptops Laptops provide a pre-diction of battery lifetime only for active use Smartphonepredictions on the other hand must cover multiple activeidle periods to be useful
01
23
45
0 20 40 60 80 100User percentile
Err
or r
atio
2 hours1 hour10 minutes
Figure 32 Battery drain predictors have to copewith the high variance in drain
time
window to
predict
n chunks
hellip
Trend Table
hellip
hellip
Figure 33 A userrsquos trend table captures how batterydrain in n consecutive chunks relates to drain in thesubsequent window
bursty nature of user interactions is one cause of this highvariance The variance is high even at longer time scalesThe standard deviation for two hour windows is larger than150 of the mean for half the users The diurnal patternsthat we showed earlier are one cause of high variance oversuch large time scales
Although the variance is high there are patterns in auserrsquos behaviors For instance we showed earlier that aphone that has been idle for a while is likely to remain idlein the near future and it will thus continue to draw energyat a similar rate
We hypothesize that a predictor tuned to user behaviorcan be accurate We present a simple personalized energydrain predictor that incorporates various user-specific fac-tors such as length of idle periods and different types ofbusy periods Instead of explicitly identifying these factorsour predictor captures them in a ldquotrend tablerdquo framework
Our framework is shown in Figure 33 Each entry in thetrend table is indexed by an n-tuple that represents energyusage readings from adjacent chunks of size δ each Thisindex points to energy usage statistics across all time win-dows of size w in the history that follow the n chunk valuesrepresented by that index Thus the trend table captureshow the energy usage in n consecutive chunks is related tothe energy use in the subsequent time windows of size w
In this paper we use n = 3 δ = 10 minutes and maintaindifferent trend tables for each window size w for which theprediction is required To keep the trend table small wequantize the energy readings of chunks We maintain themean and standard deviation for all subsequent windowsthat map to that index
To make a prediction we use the energy readings fromthe n immediately preceding chunks Let these readings be(x1 xn) Then we search for indices in the table thatdiffer from (x1 xn) by no more than a threshold s and
GenericShortminustermTimeminusofminusdayPersonalized
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
(a) 1-hour window
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
GenericShortminustermTimeminusofminusdayPersonalized
(b) 2-hour window
Figure 34 The ratio of error to true battery drainfor different energy drain predictors
predict based on them such that similar indices have higherrelative weight More precisely the prediction is
X
e stforalli=1n|he
iminusxi|lts
ke times ue
P
ke
where e iterates over all tuples in the trend table hei is
the value of the chunk i of the tuple e ue represents thestatistics stored for that index and the relative weight ke =s minus maxi=1n |he
i minus xi| For results shown here we uses = 05
Figures 34 shows how well our Personalized predictor worksfor predicting usage for two different future time windowsThere is one point per user in each graph which correspondsto the median error seen across many predictions for thatuser We see that the 90th percentile error is 25 for 1-hourwindow and 40 for 2-hour window
To place this level of accuracy in context the figure alsoshows the accurary of three alternative predictors TheGeneric predictor uses a trend table but instead of buildinga user-specific table it builds one that combines all usersWe see that its 90th percentile error is roughly twice thatof the Personalized predictor Its predictions are worse forhalf of the users
The Short-term predictor predicts energy drain simply asthe amount drained in the immediately preceding windowof the same size For both time windows its 90th percentileerror is no better than that of the Generic predictor
The Time-of-day predictor predicts energy drain based onvalues observed for the user at the same time of the day inthe past similar to that proposed by Banerjee et al [2]Thus this predictor has some degree of personalization butit learns user behavior in less detail than the Personalizedpredictor We see that its performance for 1-hour windowis similar to the Short-term predictor For 2-hour windowits error is higher than the Personalized predictor across theboard though its worst case error is lower than the Short-term and Generic predictors
Overall these results demonstrate the value of appropri-ately learning user behaviors to design intelligent mecha-nisms on smartphones
10 RELATED WORKIn a range of domains there is a rich history of work that
characterizes user workloads However because smartphoneadoption has gathered pace relatively recently our work rep-resents one of the few to study how people use smartphones
Along with other recent works it helps complete the pic-ture Banerjee et al and Rahmati et al report on batterycharging behaviors [2 18] Like us they find considerablevariation among users Banerjee et al also propose a pre-dictor that estimates the excess energy of the battery atthe time of charging using a histogram of past battery us-age [2] Shye et al study the power consumption character-istics of 20 users [24] They infer properties such as whichcomponents in the phone consume most power and exploreoptimizations based on these characteristics Rahmati andZhong study 14 users of a particular demographic to studywhich applications are popular in that demographic wherethe phone is used and how it is shared among users [19]
In contrast to these works we focus on understandingdifferent aspects of smartphone use (eg interactions andtraffic) and on exposing the diversity of user behaviors in-stead of only the average behavior Our study also entailsan order of magnitude more users than previous efforts
Like us other researchers have developed logging utilitiesMyExperience is one such early utility [10] and the worksabove involve custom utilities as well Our focus in thispaper is not on the development of logging tools but onanalyzing their data to gain insight into user behavior
There is a body of work in modeling the aggregate be-havior of mobile users Using traces from a large cellularoperator some network related aspects of mobile usage havebeen modeled Halepovic et al and Williamson et al re-port that call arrivals are bursty and present diurnal pat-terns [28] Willkomm et al and Brown et al report thatmobile users call duration can be approximately modeled bya lognormal distribution [4 29] We use traces collected onthe mobile device itself and focus on modeling the interac-tions of individual users instead of the aggregate behavior
11 CONCLUSIONSBy studying 255 users of two different smartphone plat-
forms we comprehensively characterized user activities andtheir impact on network and battery We quantify manyhitherto unknown aspects of smartphone usage User diver-sity is an overarching theme in our findings For instancedifferent users interact with their phones 10-200 times a dayon average the mean interaction length of different users
is 10-250 seconds and users receive 1-1000 MB of data perday where 10-90 is received as part of active use
This extent of user diversity implies that mechanisms thatwork for the average case may be ineffective for a large frac-tion of the users Instead learning and adapting to userbehaviors is likely to be more effective as demonstrated byour personalized energy drain predictor We show that qual-itative similarities exist among users to facilitate the devel-opment of such mechanisms For instance the longer theuser has not interacted with the phone the less likely it isfor her to start interacting with it and application popular-ity for a user follows a simple exponential distribution
Our study points to ways in which smartphone platformsshould be enhanced Effective adaptation will require futureplatforms to support light-weight tools that monitor andlearn user behaviors in situ It will also require them toexpose appropriate knobs to control the behavior of lower-level components (eg CPU or radio)
12 ACKNOWLEDGEMENTSWe are grateful to the users in Dataset1 that ran our log-
ger and to the providers of Dataset2 Elizabeth Dawson pro-vided administrative support for data collection Jean BolotKurt Partridge Stefan Saroiu Lin Zhong and the MobiSysreviewers provided many comments that helped improve thepaper We thank them all This work is supported in partby the NSF grants CCR-0120778 and CNS-0627084
13 REFERENCES[1] Balasubramanian N Balasubramanian A
and Venkataramani A Energy consumption inmobile phones A measurement study andimplications for network applications In IMC (2009)
[2] Banerjee N Rahmati A Corner M D
Rollins S and Zhong L Users and batteriesInteractions and adaptive energy management inmobile systems In Ubicomp (2007)
[3] Becker R Chambers J and Wilks A The newS language Chapman amp HallCRC 1988
[4] Brown L Gans N Mandelbaum A Sakov
A Shen H Zeltyn S and Zhao L Statisticalanalysis of a telephone call center Journal of theAmerican Statistical Association 100 469 (2005)
[5] Church K and Smyth B Understanding mobileinformation needs In MobileHCI (2008)
[6] Dempster A Laird N and Rubin D Maximumlikelihood from incomplete data via the EM algorithmJournal of the Royal Statistical Society Series B(Methodological) 39 1 (1977)
[7] Esfahbod B Preload An adaptive prefetchingdaemon Masterrsquos thesis University of Toronto 2006
[8] Falaki H Govindan R and Estrin D Smartscreen management on mobile phones Tech Rep 74Center for Embedded Networked Sesning 2009
[9] Flinn J and Satyanarayanan M Managingbattery lifetime with energy-aware adaptation ACMTOCS 22 2 (2004)
[10] Froehlich J Chen M Consolvo S
Harrison B and Landay J MyExperience Asystem for in situ tracing and capturing of userfeedback on mobile phones In MobiSys (2007)
[11] Fukuda K Cho K and Esaki H The impact ofresidential broadband traffic on Japanese ISPbackbones SIGCOMM CCR 35 1 (2005)
[12] Kotz D and Essien K Analysis of a campus-widewireless network Wireless Networks 11 1ndash2 (2005)
[13] Kwiatkowski D Phillips P Schmidt P and
Shin Y Testing the null hypothesis of stationarityagainst the alternative of a unit root Journal ofEconometrics 54 1-3 (1992)
[14] Lahiri K Dey S Panigrahi D and
Raghunathan A Battery-driven system design Anew frontier in low power design In Asia SouthPacific design automationVLSI Design (2002)
[15] Ljung G and Box G On a measure of lack of fitin time series models Biometrika 65 2 (1978)
[16] Smartphone futures 2009-2014 httpwwwportioresearchcomSmartphone09-14html
[17] Rahmati A Qian A and Zhong L
Understanding human-battery interaction on mobilephones In MobileHCI (2007)
[18] Rahmati A and Zhong L Humanndashbatteryinteraction on mobile phones Pervasive and MobileComputing 5 5 (2009)
[19] Rahmati A and Zhong L A longitudinal study ofnon-voice mobile phone usage by teens from anunderserved urban community Tech Rep 0515-09Rice University 2009
[20] Rao R Vrudhula S and Rakhmatov D
Battery modeling for energy aware system designIEEE Computer 36 12 (2003)
[21] Ravi N Scott J Han L and Iftode L
Context-aware battery management for mobilephones In PerCom (2008)
[22] Sharma A Navda V Ramjee R
Padmanabhan V and Belding E Cool-TetherEnergy efficient on-the-fly WiFi hot-spots usingmobile phones In CoNEXT (2009)
[23] Sharma C US wireless data market update - Q32009 httpwwwchetansharmacomusmarketupdateq309htm
[24] Shye A Sholbrock B and G M Into the wildStudying real user activity patterns to guide poweroptimization for mobile architectures In Micro (2009)
[25] Snol L More smartphones than desktop PCs by2011 httpwwwpcworldcomarticle171380more_smartphones_than_desktop_pcs_by_2011
html
[26] Sohn T Li K Griswold W and Hollan J
A diary study of mobile information needs In SIGCHI(2008)
[27] Trestian I Ranjan S Kuzmanovic A and
Nucci A Measuring serendipity Connecting peoplelocations and interests in a mobile 3G network InIMC (2009)
[28] Williamson C Halepovic E Sun H and Wu
Y Characterization of CDMA2000 Cellular DataNetwork Traffic In Local Computer Networks (2005)
[29] Willkomm D Machiraju S Bolot J and
Wolisz A Primary users in cellular networks Alarge-scale measurement study In DySPAN (2008)
- Introduction
- Data Collection
-
- Dataset1
- Dataset2
- Representativeness of conclusions
-
- User Interactions
-
- Interaction Time
- Interaction Sessions
- Diurnal Patterns
-
- Application Usage
-
- Number of applications
- Application Popularity
- Application Sessions
-
- Traffic
-
- Traffic per day
- ``Interactive traffic
- Diurnal patterns
-
- Energy Consumption
- Implications of User Diversity
- Smartphone Usage Models
-
- Session Lengths
- Time between Sessions
- Application Popularity
-
- Predicting Energy Drain
- Related Work
- Conclusions
- Acknowledgements
- References
-
020
4060
8010
0
0 20 40 60 80 100User percentile
MS
E (
)
(a) MSE
00
02
04
06
08
10
0 20 40 60 80 100User percentile
Est
imat
ed r
ate
(b) Rate parameter
Figure 31 (a) The mean square error (MSE) whenapplication popularity distribution is modeled usingan exponential (b) The inferred rate parameter ofthe exponential distribution for different users
83 Application PopularityWe find that for each user the relative popularity of ap-
plications can be well described by a simple exponential dis-tribution This qualitative invariant is useful for instanceto predict how many applications account for a given frac-tion of user attention For the example users in Figure 11this facet can be seen in the inset plots the semi-log of thepopularity distribution is very close to a straight line
Figure 31(a) shows that this exponential drop in applica-tion popularity is true for almost all users the mean squareerror between modeled exponential and actual popularitydistribution is less than 5 for 95 of the users
Figure 31(b) shows the inferred rate parameter of the ap-plication popularity distribution for various users We seethat the rate varies by an order of magnitude from 01 toalmost 1 The value of the rate essentially captures the paceof the drop in application popularity Lower values describeusers that use more applications on a regular basis Oneimplication of the wide range is that it may be feasible toretain all popular applications in memory for some users andnot for others
9 PREDICTING ENERGY DRAINIn this section we demonstrate the value of adapting to
user behaviors in the context of a mechanism to predict fu-ture energy drain on the smartphone Such a mechanismcan help with scheduling background tasks [9] estimatingbattery lifetime and improving user experience if it is esti-mated that the user will not fully drain the battery until thenext charging cycle [2 21]
Despite its many uses to our knowledge none of thesmartphones today provide a prediction of future energydrain Providing an accurate prediction of energy drain isdifficult2 User diversity of energy use makes any static pre-diction method highly inaccurate Even for the same userthere is a high variance in energy usage Figure 32 shows thisvariance by plotting the ratio of the standard deviation tothe mean energy drain over several time periods for users inDataset 1 The standard deviation of 10-minute windows ishigher than three times the mean for a fifth of the users The
2Energy drain prediction on smartphones is more challeng-ing than what is done for laptops Laptops provide a pre-diction of battery lifetime only for active use Smartphonepredictions on the other hand must cover multiple activeidle periods to be useful
01
23
45
0 20 40 60 80 100User percentile
Err
or r
atio
2 hours1 hour10 minutes
Figure 32 Battery drain predictors have to copewith the high variance in drain
time
window to
predict
n chunks
hellip
Trend Table
hellip
hellip
Figure 33 A userrsquos trend table captures how batterydrain in n consecutive chunks relates to drain in thesubsequent window
bursty nature of user interactions is one cause of this highvariance The variance is high even at longer time scalesThe standard deviation for two hour windows is larger than150 of the mean for half the users The diurnal patternsthat we showed earlier are one cause of high variance oversuch large time scales
Although the variance is high there are patterns in auserrsquos behaviors For instance we showed earlier that aphone that has been idle for a while is likely to remain idlein the near future and it will thus continue to draw energyat a similar rate
We hypothesize that a predictor tuned to user behaviorcan be accurate We present a simple personalized energydrain predictor that incorporates various user-specific fac-tors such as length of idle periods and different types ofbusy periods Instead of explicitly identifying these factorsour predictor captures them in a ldquotrend tablerdquo framework
Our framework is shown in Figure 33 Each entry in thetrend table is indexed by an n-tuple that represents energyusage readings from adjacent chunks of size δ each Thisindex points to energy usage statistics across all time win-dows of size w in the history that follow the n chunk valuesrepresented by that index Thus the trend table captureshow the energy usage in n consecutive chunks is related tothe energy use in the subsequent time windows of size w
In this paper we use n = 3 δ = 10 minutes and maintaindifferent trend tables for each window size w for which theprediction is required To keep the trend table small wequantize the energy readings of chunks We maintain themean and standard deviation for all subsequent windowsthat map to that index
To make a prediction we use the energy readings fromthe n immediately preceding chunks Let these readings be(x1 xn) Then we search for indices in the table thatdiffer from (x1 xn) by no more than a threshold s and
GenericShortminustermTimeminusofminusdayPersonalized
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
(a) 1-hour window
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
GenericShortminustermTimeminusofminusdayPersonalized
(b) 2-hour window
Figure 34 The ratio of error to true battery drainfor different energy drain predictors
predict based on them such that similar indices have higherrelative weight More precisely the prediction is
X
e stforalli=1n|he
iminusxi|lts
ke times ue
P
ke
where e iterates over all tuples in the trend table hei is
the value of the chunk i of the tuple e ue represents thestatistics stored for that index and the relative weight ke =s minus maxi=1n |he
i minus xi| For results shown here we uses = 05
Figures 34 shows how well our Personalized predictor worksfor predicting usage for two different future time windowsThere is one point per user in each graph which correspondsto the median error seen across many predictions for thatuser We see that the 90th percentile error is 25 for 1-hourwindow and 40 for 2-hour window
To place this level of accuracy in context the figure alsoshows the accurary of three alternative predictors TheGeneric predictor uses a trend table but instead of buildinga user-specific table it builds one that combines all usersWe see that its 90th percentile error is roughly twice thatof the Personalized predictor Its predictions are worse forhalf of the users
The Short-term predictor predicts energy drain simply asthe amount drained in the immediately preceding windowof the same size For both time windows its 90th percentileerror is no better than that of the Generic predictor
The Time-of-day predictor predicts energy drain based onvalues observed for the user at the same time of the day inthe past similar to that proposed by Banerjee et al [2]Thus this predictor has some degree of personalization butit learns user behavior in less detail than the Personalizedpredictor We see that its performance for 1-hour windowis similar to the Short-term predictor For 2-hour windowits error is higher than the Personalized predictor across theboard though its worst case error is lower than the Short-term and Generic predictors
Overall these results demonstrate the value of appropri-ately learning user behaviors to design intelligent mecha-nisms on smartphones
10 RELATED WORKIn a range of domains there is a rich history of work that
characterizes user workloads However because smartphoneadoption has gathered pace relatively recently our work rep-resents one of the few to study how people use smartphones
Along with other recent works it helps complete the pic-ture Banerjee et al and Rahmati et al report on batterycharging behaviors [2 18] Like us they find considerablevariation among users Banerjee et al also propose a pre-dictor that estimates the excess energy of the battery atthe time of charging using a histogram of past battery us-age [2] Shye et al study the power consumption character-istics of 20 users [24] They infer properties such as whichcomponents in the phone consume most power and exploreoptimizations based on these characteristics Rahmati andZhong study 14 users of a particular demographic to studywhich applications are popular in that demographic wherethe phone is used and how it is shared among users [19]
In contrast to these works we focus on understandingdifferent aspects of smartphone use (eg interactions andtraffic) and on exposing the diversity of user behaviors in-stead of only the average behavior Our study also entailsan order of magnitude more users than previous efforts
Like us other researchers have developed logging utilitiesMyExperience is one such early utility [10] and the worksabove involve custom utilities as well Our focus in thispaper is not on the development of logging tools but onanalyzing their data to gain insight into user behavior
There is a body of work in modeling the aggregate be-havior of mobile users Using traces from a large cellularoperator some network related aspects of mobile usage havebeen modeled Halepovic et al and Williamson et al re-port that call arrivals are bursty and present diurnal pat-terns [28] Willkomm et al and Brown et al report thatmobile users call duration can be approximately modeled bya lognormal distribution [4 29] We use traces collected onthe mobile device itself and focus on modeling the interac-tions of individual users instead of the aggregate behavior
11 CONCLUSIONSBy studying 255 users of two different smartphone plat-
forms we comprehensively characterized user activities andtheir impact on network and battery We quantify manyhitherto unknown aspects of smartphone usage User diver-sity is an overarching theme in our findings For instancedifferent users interact with their phones 10-200 times a dayon average the mean interaction length of different users
is 10-250 seconds and users receive 1-1000 MB of data perday where 10-90 is received as part of active use
This extent of user diversity implies that mechanisms thatwork for the average case may be ineffective for a large frac-tion of the users Instead learning and adapting to userbehaviors is likely to be more effective as demonstrated byour personalized energy drain predictor We show that qual-itative similarities exist among users to facilitate the devel-opment of such mechanisms For instance the longer theuser has not interacted with the phone the less likely it isfor her to start interacting with it and application popular-ity for a user follows a simple exponential distribution
Our study points to ways in which smartphone platformsshould be enhanced Effective adaptation will require futureplatforms to support light-weight tools that monitor andlearn user behaviors in situ It will also require them toexpose appropriate knobs to control the behavior of lower-level components (eg CPU or radio)
12 ACKNOWLEDGEMENTSWe are grateful to the users in Dataset1 that ran our log-
ger and to the providers of Dataset2 Elizabeth Dawson pro-vided administrative support for data collection Jean BolotKurt Partridge Stefan Saroiu Lin Zhong and the MobiSysreviewers provided many comments that helped improve thepaper We thank them all This work is supported in partby the NSF grants CCR-0120778 and CNS-0627084
13 REFERENCES[1] Balasubramanian N Balasubramanian A
and Venkataramani A Energy consumption inmobile phones A measurement study andimplications for network applications In IMC (2009)
[2] Banerjee N Rahmati A Corner M D
Rollins S and Zhong L Users and batteriesInteractions and adaptive energy management inmobile systems In Ubicomp (2007)
[3] Becker R Chambers J and Wilks A The newS language Chapman amp HallCRC 1988
[4] Brown L Gans N Mandelbaum A Sakov
A Shen H Zeltyn S and Zhao L Statisticalanalysis of a telephone call center Journal of theAmerican Statistical Association 100 469 (2005)
[5] Church K and Smyth B Understanding mobileinformation needs In MobileHCI (2008)
[6] Dempster A Laird N and Rubin D Maximumlikelihood from incomplete data via the EM algorithmJournal of the Royal Statistical Society Series B(Methodological) 39 1 (1977)
[7] Esfahbod B Preload An adaptive prefetchingdaemon Masterrsquos thesis University of Toronto 2006
[8] Falaki H Govindan R and Estrin D Smartscreen management on mobile phones Tech Rep 74Center for Embedded Networked Sesning 2009
[9] Flinn J and Satyanarayanan M Managingbattery lifetime with energy-aware adaptation ACMTOCS 22 2 (2004)
[10] Froehlich J Chen M Consolvo S
Harrison B and Landay J MyExperience Asystem for in situ tracing and capturing of userfeedback on mobile phones In MobiSys (2007)
[11] Fukuda K Cho K and Esaki H The impact ofresidential broadband traffic on Japanese ISPbackbones SIGCOMM CCR 35 1 (2005)
[12] Kotz D and Essien K Analysis of a campus-widewireless network Wireless Networks 11 1ndash2 (2005)
[13] Kwiatkowski D Phillips P Schmidt P and
Shin Y Testing the null hypothesis of stationarityagainst the alternative of a unit root Journal ofEconometrics 54 1-3 (1992)
[14] Lahiri K Dey S Panigrahi D and
Raghunathan A Battery-driven system design Anew frontier in low power design In Asia SouthPacific design automationVLSI Design (2002)
[15] Ljung G and Box G On a measure of lack of fitin time series models Biometrika 65 2 (1978)
[16] Smartphone futures 2009-2014 httpwwwportioresearchcomSmartphone09-14html
[17] Rahmati A Qian A and Zhong L
Understanding human-battery interaction on mobilephones In MobileHCI (2007)
[18] Rahmati A and Zhong L Humanndashbatteryinteraction on mobile phones Pervasive and MobileComputing 5 5 (2009)
[19] Rahmati A and Zhong L A longitudinal study ofnon-voice mobile phone usage by teens from anunderserved urban community Tech Rep 0515-09Rice University 2009
[20] Rao R Vrudhula S and Rakhmatov D
Battery modeling for energy aware system designIEEE Computer 36 12 (2003)
[21] Ravi N Scott J Han L and Iftode L
Context-aware battery management for mobilephones In PerCom (2008)
[22] Sharma A Navda V Ramjee R
Padmanabhan V and Belding E Cool-TetherEnergy efficient on-the-fly WiFi hot-spots usingmobile phones In CoNEXT (2009)
[23] Sharma C US wireless data market update - Q32009 httpwwwchetansharmacomusmarketupdateq309htm
[24] Shye A Sholbrock B and G M Into the wildStudying real user activity patterns to guide poweroptimization for mobile architectures In Micro (2009)
[25] Snol L More smartphones than desktop PCs by2011 httpwwwpcworldcomarticle171380more_smartphones_than_desktop_pcs_by_2011
html
[26] Sohn T Li K Griswold W and Hollan J
A diary study of mobile information needs In SIGCHI(2008)
[27] Trestian I Ranjan S Kuzmanovic A and
Nucci A Measuring serendipity Connecting peoplelocations and interests in a mobile 3G network InIMC (2009)
[28] Williamson C Halepovic E Sun H and Wu
Y Characterization of CDMA2000 Cellular DataNetwork Traffic In Local Computer Networks (2005)
[29] Willkomm D Machiraju S Bolot J and
Wolisz A Primary users in cellular networks Alarge-scale measurement study In DySPAN (2008)
- Introduction
- Data Collection
-
- Dataset1
- Dataset2
- Representativeness of conclusions
-
- User Interactions
-
- Interaction Time
- Interaction Sessions
- Diurnal Patterns
-
- Application Usage
-
- Number of applications
- Application Popularity
- Application Sessions
-
- Traffic
-
- Traffic per day
- ``Interactive traffic
- Diurnal patterns
-
- Energy Consumption
- Implications of User Diversity
- Smartphone Usage Models
-
- Session Lengths
- Time between Sessions
- Application Popularity
-
- Predicting Energy Drain
- Related Work
- Conclusions
- Acknowledgements
- References
-
GenericShortminustermTimeminusofminusdayPersonalized
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
(a) 1-hour window
00
05
10
15
0 20 40 60 80 100User percentile
Err
or r
atio
GenericShortminustermTimeminusofminusdayPersonalized
(b) 2-hour window
Figure 34 The ratio of error to true battery drainfor different energy drain predictors
predict based on them such that similar indices have higherrelative weight More precisely the prediction is
X
e stforalli=1n|he
iminusxi|lts
ke times ue
P
ke
where e iterates over all tuples in the trend table hei is
the value of the chunk i of the tuple e ue represents thestatistics stored for that index and the relative weight ke =s minus maxi=1n |he
i minus xi| For results shown here we uses = 05
Figures 34 shows how well our Personalized predictor worksfor predicting usage for two different future time windowsThere is one point per user in each graph which correspondsto the median error seen across many predictions for thatuser We see that the 90th percentile error is 25 for 1-hourwindow and 40 for 2-hour window
To place this level of accuracy in context the figure alsoshows the accurary of three alternative predictors TheGeneric predictor uses a trend table but instead of buildinga user-specific table it builds one that combines all usersWe see that its 90th percentile error is roughly twice thatof the Personalized predictor Its predictions are worse forhalf of the users
The Short-term predictor predicts energy drain simply asthe amount drained in the immediately preceding windowof the same size For both time windows its 90th percentileerror is no better than that of the Generic predictor
The Time-of-day predictor predicts energy drain based onvalues observed for the user at the same time of the day inthe past similar to that proposed by Banerjee et al [2]Thus this predictor has some degree of personalization butit learns user behavior in less detail than the Personalizedpredictor We see that its performance for 1-hour windowis similar to the Short-term predictor For 2-hour windowits error is higher than the Personalized predictor across theboard though its worst case error is lower than the Short-term and Generic predictors
Overall these results demonstrate the value of appropri-ately learning user behaviors to design intelligent mecha-nisms on smartphones
10 RELATED WORKIn a range of domains there is a rich history of work that
characterizes user workloads However because smartphoneadoption has gathered pace relatively recently our work rep-resents one of the few to study how people use smartphones
Along with other recent works it helps complete the pic-ture Banerjee et al and Rahmati et al report on batterycharging behaviors [2 18] Like us they find considerablevariation among users Banerjee et al also propose a pre-dictor that estimates the excess energy of the battery atthe time of charging using a histogram of past battery us-age [2] Shye et al study the power consumption character-istics of 20 users [24] They infer properties such as whichcomponents in the phone consume most power and exploreoptimizations based on these characteristics Rahmati andZhong study 14 users of a particular demographic to studywhich applications are popular in that demographic wherethe phone is used and how it is shared among users [19]
In contrast to these works we focus on understandingdifferent aspects of smartphone use (eg interactions andtraffic) and on exposing the diversity of user behaviors in-stead of only the average behavior Our study also entailsan order of magnitude more users than previous efforts
Like us other researchers have developed logging utilitiesMyExperience is one such early utility [10] and the worksabove involve custom utilities as well Our focus in thispaper is not on the development of logging tools but onanalyzing their data to gain insight into user behavior
There is a body of work in modeling the aggregate be-havior of mobile users Using traces from a large cellularoperator some network related aspects of mobile usage havebeen modeled Halepovic et al and Williamson et al re-port that call arrivals are bursty and present diurnal pat-terns [28] Willkomm et al and Brown et al report thatmobile users call duration can be approximately modeled bya lognormal distribution [4 29] We use traces collected onthe mobile device itself and focus on modeling the interac-tions of individual users instead of the aggregate behavior
11 CONCLUSIONSBy studying 255 users of two different smartphone plat-
forms we comprehensively characterized user activities andtheir impact on network and battery We quantify manyhitherto unknown aspects of smartphone usage User diver-sity is an overarching theme in our findings For instancedifferent users interact with their phones 10-200 times a dayon average the mean interaction length of different users
is 10-250 seconds and users receive 1-1000 MB of data perday where 10-90 is received as part of active use
This extent of user diversity implies that mechanisms thatwork for the average case may be ineffective for a large frac-tion of the users Instead learning and adapting to userbehaviors is likely to be more effective as demonstrated byour personalized energy drain predictor We show that qual-itative similarities exist among users to facilitate the devel-opment of such mechanisms For instance the longer theuser has not interacted with the phone the less likely it isfor her to start interacting with it and application popular-ity for a user follows a simple exponential distribution
Our study points to ways in which smartphone platformsshould be enhanced Effective adaptation will require futureplatforms to support light-weight tools that monitor andlearn user behaviors in situ It will also require them toexpose appropriate knobs to control the behavior of lower-level components (eg CPU or radio)
12 ACKNOWLEDGEMENTSWe are grateful to the users in Dataset1 that ran our log-
ger and to the providers of Dataset2 Elizabeth Dawson pro-vided administrative support for data collection Jean BolotKurt Partridge Stefan Saroiu Lin Zhong and the MobiSysreviewers provided many comments that helped improve thepaper We thank them all This work is supported in partby the NSF grants CCR-0120778 and CNS-0627084
13 REFERENCES[1] Balasubramanian N Balasubramanian A
and Venkataramani A Energy consumption inmobile phones A measurement study andimplications for network applications In IMC (2009)
[2] Banerjee N Rahmati A Corner M D
Rollins S and Zhong L Users and batteriesInteractions and adaptive energy management inmobile systems In Ubicomp (2007)
[3] Becker R Chambers J and Wilks A The newS language Chapman amp HallCRC 1988
[4] Brown L Gans N Mandelbaum A Sakov
A Shen H Zeltyn S and Zhao L Statisticalanalysis of a telephone call center Journal of theAmerican Statistical Association 100 469 (2005)
[5] Church K and Smyth B Understanding mobileinformation needs In MobileHCI (2008)
[6] Dempster A Laird N and Rubin D Maximumlikelihood from incomplete data via the EM algorithmJournal of the Royal Statistical Society Series B(Methodological) 39 1 (1977)
[7] Esfahbod B Preload An adaptive prefetchingdaemon Masterrsquos thesis University of Toronto 2006
[8] Falaki H Govindan R and Estrin D Smartscreen management on mobile phones Tech Rep 74Center for Embedded Networked Sesning 2009
[9] Flinn J and Satyanarayanan M Managingbattery lifetime with energy-aware adaptation ACMTOCS 22 2 (2004)
[10] Froehlich J Chen M Consolvo S
Harrison B and Landay J MyExperience Asystem for in situ tracing and capturing of userfeedback on mobile phones In MobiSys (2007)
[11] Fukuda K Cho K and Esaki H The impact ofresidential broadband traffic on Japanese ISPbackbones SIGCOMM CCR 35 1 (2005)
[12] Kotz D and Essien K Analysis of a campus-widewireless network Wireless Networks 11 1ndash2 (2005)
[13] Kwiatkowski D Phillips P Schmidt P and
Shin Y Testing the null hypothesis of stationarityagainst the alternative of a unit root Journal ofEconometrics 54 1-3 (1992)
[14] Lahiri K Dey S Panigrahi D and
Raghunathan A Battery-driven system design Anew frontier in low power design In Asia SouthPacific design automationVLSI Design (2002)
[15] Ljung G and Box G On a measure of lack of fitin time series models Biometrika 65 2 (1978)
[16] Smartphone futures 2009-2014 httpwwwportioresearchcomSmartphone09-14html
[17] Rahmati A Qian A and Zhong L
Understanding human-battery interaction on mobilephones In MobileHCI (2007)
[18] Rahmati A and Zhong L Humanndashbatteryinteraction on mobile phones Pervasive and MobileComputing 5 5 (2009)
[19] Rahmati A and Zhong L A longitudinal study ofnon-voice mobile phone usage by teens from anunderserved urban community Tech Rep 0515-09Rice University 2009
[20] Rao R Vrudhula S and Rakhmatov D
Battery modeling for energy aware system designIEEE Computer 36 12 (2003)
[21] Ravi N Scott J Han L and Iftode L
Context-aware battery management for mobilephones In PerCom (2008)
[22] Sharma A Navda V Ramjee R
Padmanabhan V and Belding E Cool-TetherEnergy efficient on-the-fly WiFi hot-spots usingmobile phones In CoNEXT (2009)
[23] Sharma C US wireless data market update - Q32009 httpwwwchetansharmacomusmarketupdateq309htm
[24] Shye A Sholbrock B and G M Into the wildStudying real user activity patterns to guide poweroptimization for mobile architectures In Micro (2009)
[25] Snol L More smartphones than desktop PCs by2011 httpwwwpcworldcomarticle171380more_smartphones_than_desktop_pcs_by_2011
html
[26] Sohn T Li K Griswold W and Hollan J
A diary study of mobile information needs In SIGCHI(2008)
[27] Trestian I Ranjan S Kuzmanovic A and
Nucci A Measuring serendipity Connecting peoplelocations and interests in a mobile 3G network InIMC (2009)
[28] Williamson C Halepovic E Sun H and Wu
Y Characterization of CDMA2000 Cellular DataNetwork Traffic In Local Computer Networks (2005)
[29] Willkomm D Machiraju S Bolot J and
Wolisz A Primary users in cellular networks Alarge-scale measurement study In DySPAN (2008)
- Introduction
- Data Collection
-
- Dataset1
- Dataset2
- Representativeness of conclusions
-
- User Interactions
-
- Interaction Time
- Interaction Sessions
- Diurnal Patterns
-
- Application Usage
-
- Number of applications
- Application Popularity
- Application Sessions
-
- Traffic
-
- Traffic per day
- ``Interactive traffic
- Diurnal patterns
-
- Energy Consumption
- Implications of User Diversity
- Smartphone Usage Models
-
- Session Lengths
- Time between Sessions
- Application Popularity
-
- Predicting Energy Drain
- Related Work
- Conclusions
- Acknowledgements
- References
-
is 10-250 seconds and users receive 1-1000 MB of data perday where 10-90 is received as part of active use
This extent of user diversity implies that mechanisms thatwork for the average case may be ineffective for a large frac-tion of the users Instead learning and adapting to userbehaviors is likely to be more effective as demonstrated byour personalized energy drain predictor We show that qual-itative similarities exist among users to facilitate the devel-opment of such mechanisms For instance the longer theuser has not interacted with the phone the less likely it isfor her to start interacting with it and application popular-ity for a user follows a simple exponential distribution
Our study points to ways in which smartphone platformsshould be enhanced Effective adaptation will require futureplatforms to support light-weight tools that monitor andlearn user behaviors in situ It will also require them toexpose appropriate knobs to control the behavior of lower-level components (eg CPU or radio)
12 ACKNOWLEDGEMENTSWe are grateful to the users in Dataset1 that ran our log-
ger and to the providers of Dataset2 Elizabeth Dawson pro-vided administrative support for data collection Jean BolotKurt Partridge Stefan Saroiu Lin Zhong and the MobiSysreviewers provided many comments that helped improve thepaper We thank them all This work is supported in partby the NSF grants CCR-0120778 and CNS-0627084
13 REFERENCES[1] Balasubramanian N Balasubramanian A
and Venkataramani A Energy consumption inmobile phones A measurement study andimplications for network applications In IMC (2009)
[2] Banerjee N Rahmati A Corner M D
Rollins S and Zhong L Users and batteriesInteractions and adaptive energy management inmobile systems In Ubicomp (2007)
[3] Becker R Chambers J and Wilks A The newS language Chapman amp HallCRC 1988
[4] Brown L Gans N Mandelbaum A Sakov
A Shen H Zeltyn S and Zhao L Statisticalanalysis of a telephone call center Journal of theAmerican Statistical Association 100 469 (2005)
[5] Church K and Smyth B Understanding mobileinformation needs In MobileHCI (2008)
[6] Dempster A Laird N and Rubin D Maximumlikelihood from incomplete data via the EM algorithmJournal of the Royal Statistical Society Series B(Methodological) 39 1 (1977)
[7] Esfahbod B Preload An adaptive prefetchingdaemon Masterrsquos thesis University of Toronto 2006
[8] Falaki H Govindan R and Estrin D Smartscreen management on mobile phones Tech Rep 74Center for Embedded Networked Sesning 2009
[9] Flinn J and Satyanarayanan M Managingbattery lifetime with energy-aware adaptation ACMTOCS 22 2 (2004)
[10] Froehlich J Chen M Consolvo S
Harrison B and Landay J MyExperience Asystem for in situ tracing and capturing of userfeedback on mobile phones In MobiSys (2007)
[11] Fukuda K Cho K and Esaki H The impact ofresidential broadband traffic on Japanese ISPbackbones SIGCOMM CCR 35 1 (2005)
[12] Kotz D and Essien K Analysis of a campus-widewireless network Wireless Networks 11 1ndash2 (2005)
[13] Kwiatkowski D Phillips P Schmidt P and
Shin Y Testing the null hypothesis of stationarityagainst the alternative of a unit root Journal ofEconometrics 54 1-3 (1992)
[14] Lahiri K Dey S Panigrahi D and
Raghunathan A Battery-driven system design Anew frontier in low power design In Asia SouthPacific design automationVLSI Design (2002)
[15] Ljung G and Box G On a measure of lack of fitin time series models Biometrika 65 2 (1978)
[16] Smartphone futures 2009-2014 httpwwwportioresearchcomSmartphone09-14html
[17] Rahmati A Qian A and Zhong L
Understanding human-battery interaction on mobilephones In MobileHCI (2007)
[18] Rahmati A and Zhong L Humanndashbatteryinteraction on mobile phones Pervasive and MobileComputing 5 5 (2009)
[19] Rahmati A and Zhong L A longitudinal study ofnon-voice mobile phone usage by teens from anunderserved urban community Tech Rep 0515-09Rice University 2009
[20] Rao R Vrudhula S and Rakhmatov D
Battery modeling for energy aware system designIEEE Computer 36 12 (2003)
[21] Ravi N Scott J Han L and Iftode L
Context-aware battery management for mobilephones In PerCom (2008)
[22] Sharma A Navda V Ramjee R
Padmanabhan V and Belding E Cool-TetherEnergy efficient on-the-fly WiFi hot-spots usingmobile phones In CoNEXT (2009)
[23] Sharma C US wireless data market update - Q32009 httpwwwchetansharmacomusmarketupdateq309htm
[24] Shye A Sholbrock B and G M Into the wildStudying real user activity patterns to guide poweroptimization for mobile architectures In Micro (2009)
[25] Snol L More smartphones than desktop PCs by2011 httpwwwpcworldcomarticle171380more_smartphones_than_desktop_pcs_by_2011
html
[26] Sohn T Li K Griswold W and Hollan J
A diary study of mobile information needs In SIGCHI(2008)
[27] Trestian I Ranjan S Kuzmanovic A and
Nucci A Measuring serendipity Connecting peoplelocations and interests in a mobile 3G network InIMC (2009)
[28] Williamson C Halepovic E Sun H and Wu
Y Characterization of CDMA2000 Cellular DataNetwork Traffic In Local Computer Networks (2005)
[29] Willkomm D Machiraju S Bolot J and
Wolisz A Primary users in cellular networks Alarge-scale measurement study In DySPAN (2008)
- Introduction
- Data Collection
-
- Dataset1
- Dataset2
- Representativeness of conclusions
-
- User Interactions
-
- Interaction Time
- Interaction Sessions
- Diurnal Patterns
-
- Application Usage
-
- Number of applications
- Application Popularity
- Application Sessions
-
- Traffic
-
- Traffic per day
- ``Interactive traffic
- Diurnal patterns
-
- Energy Consumption
- Implications of User Diversity
- Smartphone Usage Models
-
- Session Lengths
- Time between Sessions
- Application Popularity
-
- Predicting Energy Drain
- Related Work
- Conclusions
- Acknowledgements
- References
-