AURA: An Application and User Interaction Aware Middleware Framework for Energy Optimization in Mobile Devices

Abstract — Mobile battery-operated devices are becoming an essential instrument for business, communication, and social interaction. In addition to the demand for an acceptable level of performance and a comprehensive set of features, users often desire extended battery lifetime. In fact, limited battery lifetime is one of the biggest obstacles facing the current utility and future growth of increasingly sophisticated “smart” mobile devices. This paper proposes a novel application-aware and user-interaction aware energy optimization middleware framework (AURA) for pervasive mobile devices. AURA optimizes CPU and screen backlight energy consumption while maintaining a minimum acceptable level of performance. The proposed framework employs a novel Bayesian application classifier and management strategies based on Markov Decision Processes to achieve energy savings. Real-world user evaluation studies on a Google Android based HTC Dream smartphone running the AURA framework demonstrate promising results, with up to 24% energy savings compared to the baseline device manager, and up to 5× savings over prior work on CPU and backlight energy co-optimization.

I. INTRODUCTION

Mobile smartphones and other portable battery operated embedded systems (PDAs, tablets) are pervasive computing devices that have emerged in recent years as essential instruments for communication, business, and social interactions. As of 2011, there are more than 5.3 billion mobile subscribers worldwide, with smartphone sales showing the strongest growth. Over 300,000 apps have been developed within the past three years across various mobile platforms. Popular mobile activities include web browsing, multimedia, games, e-mail, and social networking [1]. Overall, these trends suggest that mobile devices are now the new development and computing platform for the 21st century.

Performance, capabilities, and design are all primary considerations when purchasing a mobile device; however, battery lifetime is also a highly desirable attribute. Most portable devices today make use of lithium-ion polymer batteries, which have been used in electronics since the mid 1990’s [2]. Although lithium-ion battery technology and capacity has improved over the years, it still cannot keep pace with the power consumption demands of today’s mobile devices. Until a new battery technology is discovered, this key limiter has led to a strong research emphasis on battery lifetime extension, primarily using software optimizations [3]-[10].

It is important to note that outside of the obvious differences between portable mobile devices and a general PC – weight and size, form factor, computational capabilities, and robustness – a key difference can be found in the user interaction patterns and interfaces. Unlike a desktop or notebook PC in which a user typically interacts with applications using a pointer device or keyboard, applications on mobile devices most often receive user input through a touch screen or keypad events. Many times applications are interacted with for short durations throughout the day (e.g. few seconds or minutes instead of hours) and these

patterns are often unique to each individual user. Significant differences in user interaction patterns make a general-purpose power management strategy unsuitable for mobile devices.

In this work, we present a novel application and user interaction aware energy management framework (AURA) for pervasive mobile devices, which takes advantage of user idle time between interaction events of the foreground application to optimize CPU and backlight energy consumption. In order to balance energy consumption and quality of service (QoS) requirements that are unique to each individual user, AURA makes use of a Bayesian application classifier to dynamically classify applications based on user interaction. Once an application is classified, AURA utilizes Markov Decision Process (MDP) based power management algorithms to adjust processor frequency and screen backlight levels to reduce system energy consumption between user interaction events. Overall, we make the following novel contributions:

• We conduct usage studies with real users and develop a Bayesian application classifier tool to categorize mobile applications based on user interaction activity

• We develop an integrated MDP-based application and user interaction-aware energy management framework that adapts CPU and backlight levels in a mobile device to balance energy consumption and user QoS

• We characterize backlight and CPU power dissipation on an Android OS based HTC Dream mobile architecture

• We implement our framework as middleware running on the HTC Dream smartphone and demonstrate real energy savings on commercial apps running on the device

Real-world user evaluation studies with the Google Android based HTC Dream mobile device running the AURA framework demonstrate promising results, with up to 24% energy savings compared to the baseline device manager; and up to 5× savings over the best known prior work on CPU and backlight energy co-optimization, with negligible impact on user quality of service.

II. AURA ENERGY MANAGEMENT FRAMEWORK

In this section, we present details of the AURA framework. In Section II.A we first describe the fundamental observations that lay the foundation for energy savings in mobile devices. Section II.B presents results of field studies involving users interacting with apps on mobile devices. Section II.C gives a high level overview of the AURA middleware framework. Subsequent sections elaborate on the major components of the framework.

A. Fundamental User-Device Interaction Mechanisms

Here we explain the underlying concepts stemming from the psychology of user-device interactions that drive the CPU and backlight energy optimizations in the AURA framework.

There are three basic processes involved in a user’s response to any interactive system [11], such as a smartphone or personal

Brad K. Donohoo, Chris Ohlsen, and Sudeep Pasricha bdonohoo@rams.colostate.edu, ohlsensc@rams.colostate.edu, sudeep@colostate.edu

Colorado State University, Department of Electrical and Computer Engineering Fort Collins, CO, U.S.A

978-1-4577-1954-7/11/$26.00 ©2011 IEEE 168

comfrodecthemocanintwhwhthetheresproscrtheslamilredeneari

Aposblinandhumif conof phoin paichadetcerbe exp

B

thevarappintreqburunienehoppot

Wgamtravintfordifon selfornot2 rcom

use

mputer. Duringom the physicalcides on a coure motor procesovements. Thesn be characterizeracting with a

hen a user interhen a response te button changee system resposponse represenocess of the nereen) involves ae input peripheack period, folliseconds. Dur

duced, energy mergy optimizatses whenever a

Another interessibilities for endness [12] as d perception. mans to notice changes occu

nfirmed this limobservers in onotograph graduanother study

inting went undanges such as tected. By gradurtain scenarios, reduced withou

ploited by AUR

B. User-Devic

To understand e real world, wriety of apps ops running on seraction behav

quire constant irsts of user intique interactionergy managemeped to uncovetentially guide sWe selected 18mes, communicvel/local, and eraction loggerr users runningfferent users int

an Android Olected for the sr instance mainlt familiar with relied on her smmfortable whileTable 1 showsers for each of

g the perceptual world. Duringrse of action bas, the user exe

se three proceszed by an interan app on a mracts with an ato that interactioes color) is the onse during whnts the cognitivext interaction (a motor slack. Berals, the systemr hundreds toring this time,may be saved wtions in AURA

user interacts wsting phenomeenergy optimizrevealed by reChange blindnlarge changes

ur in small inmitation of humne study failed

ually disappeare, gradual colodetected by a mthe sudden ad

ually reducing senergy consumut causing user

RA to optimize s

ce Interaction F

how users intewe conducted a on a popular smsmart mobile deviors from usernteraction withteraction follown patterns makent strategies iner trends fromscenario-specifi8 Android appcation, multime

social networ app was develg each of theseteracting with tOS based HTCstudy spanned tly used his smamany of the se

martphone for ae interacting wits a summary othe selected ap

al process, the g the cognitiveased on the inpuecutes the actioses are consec

raction slack. Fmobile device, app (e.g., touchon is perceptiblperceptual slachich the user ve slack; and f(e.g. moving fi

Before the user pm is idle durino possibly ma, if the CPU without affectin

exploit this inwith a mobile denon that canzations is the esearchers in hness refers to in their environ

ncrements. Maman perception

to observe whed over the couror changes ovemajority of subjeddition of an oscreen brightne

mption in mobiler dissatisfactionscreen backligh

Field Study

eract with mobfield study wit

martphone. It isevices can invors – whereas s

h the user, otherwed by long idke conventionaneffective on mm the field stufic energy optims from the foll

edia, shopping, orking. A cusloped to recorde apps. The stuhe apps over th

C Dream smartthe proficiencyartphone for phoelected applicatia variety of tasth the chosen ap

of the interactiopps. The pattern

user senses ine process, the uut. Finally, duron by mechani

cutive in time aFor instance, wh

the time betwehing a button) ale to the user (ek; the period afcomprehends

finally the maninger to touchphysically toucng the cumulatany thousands frequency can

ng user QoS. Cnherent slack tevice. n open up nnotion of chan

human psycholothe inability

nments, especiaany studies ha[13] – a major

hen a building irse of 12 seconer time to an ects but disruptobject were easss over time une devices may thn. This principleht energy.

ile applicationsth users runnins well-known tke vastly differsome applicatiors may experiendle periods. Thal general-purp

mobile devices. Wudies that wo

mizations. lowing categoripersonal/businestom backgrou

d interaction traudy involved fhe course of a dtphone. The usy spectrum: useone calls, and wions, whereas usks, and was qupps. on logged for ns in the table

nput user ring ical and hen een and

e.g., fter the

nual the hes tive

of be

PU that

new nge ogy

of ally ave rity in a nds;

oil tive sily

nder hus e is

s in ng a that rent ons nce

hese ose We

ould

ies: ess, und aces five day sers er 3 was user uite

the are

classifieobviousis that aSudokudifferenthat userspecificalways classifiethen ceanoticeabquite difInteractiWeather

We cspecific would asave enebehavior

C. A

We noptimizaAURA fthe softbackgroapplicatFocus Epreviousstored idatabaseexists inwill makinteractiinvoke optimizeaccessesscreen bexist in is not cevents (classify FollowinappropriAn evenuser inte

D. R

The rmonitor

ed based on inte observation walthough some

and Jewels nt interaction chr-device interac

c. For example, classified as

ed as low-interaase interaction. bly user-speciffferently based ion results for all illustrate than thus concludand applicatio

attempt to exploergy (as discusr uniquely acro

Table 1: A

URA Middlewa

now present a ation frameworframework that tware stack of

ound as an Andrtions, the AURAEvent,” and upds foreground ain memory. The for the currenn the database ake use of the aion pattern statia user-specifi

ed to the aps the mobile dbacklight and Cthe database, o

classified, AUR“Global Touchthe applicationng classificatioiate power mannt emulator weraction events

Runtime Monito

runtime monitorring user interac

eraction frequenwe can immediate apps belong tare both “gamharacteristics. Ictions for someJewels, a fast-phigh-interactio

action because On the other ha

fic, meaning thd on the type ofor Minesweepehis idea. de that users inton-specific ma

oit interaction slssed in the prevoss users and ap

App interaction

are Framewor

high level overk for mobile dis implemented

f the Android roid Service. AA runtime mondate any relevanapplication to he runtime mo

nt foreground apand is classifiedapplication’s leaistics (e.g., meafic power manpplication. Thedevice’s hardwCPU frequency.or the applicatioRA will make h” and “Global Kn using the Bayeon, the MDP pnagement strate

was also integrafor validation p

or

r is an event-drctions with the

ncy (low, mediutely make fromto the same cames”) they poIt is also intere apps are very paced gaming aon, and Musiusers usually

and, some classhat they may f user interactiner, Gallery, an

teract with devianner. Any fralack and changevious section) mpplications.

n classification

rk Overview

erview of the Adevices. Figure d at the middleOS [14], and

As the user switcnitor will receint session inform

an app profionitor will thepplication. If th

d, the MDP powarned classificaan and standard nagement strate power manaware componen. If the applicaton exists in theuse of the useKey” events) toesian applicatipower manageregy on the nexated into AURApurposes.

riven module remobile device.

um, high). An m these results

ategory (e.g., ossess vastly resting to see

application-application, is c is always select a song sifications are be classified

ng with them. nd News &

ices in a user-amework that e blindness to must tailor its

n

AURA energy 1 shows the

eware level in runs in the

ches between ive a “Global mation of the ile database, n search the

he application wer manager ation and user

deviation) to tegy that is ager directly

nts to change tion does not database but er interaction o dynamically ion classifier. r will run an xt invocation. A to simulate

esponsible for The Android

169

OSlistintthealloconappevewhfockeyinvThappstanewsavappprofroforinvBarec

F

E

BinvprospegivhypthewitforexeinvsetfeabettriphigdevappareThof

S makes use oteners, which aeractions (e.g.,

e user interfaceow applicationsnsumed by aplication. In ordents, a set of cuhen any touch ocus, and whenyboard is openvoked whenevehe module keepplication, the c

ate, which are uw application gve session inplication (e.g. cofile database aom the databasereground applicvoke the MDP ayesian classifieceives user inter

Figure 1: AURA

E. Bayesian Ap

Bayesian learnvolves using eviobabilities to cecifically, the pven some obserpothesis), comee hypothesis anth the hypothesrm of Bayesianecuted by Junvolves mappingt of output measatures are the mtween user intepartite categorizgh classes basedviation. We dplication classife low-interactihus the classifie

user and applic

of public callballow applicatitouch and key

e (UI) [14]. Uns to receive “gl

application codder for the runt

ustom event broor key event oc

n the phone coned or closed)er such custom s tracking varia

class of user inupdated on recegains foregroundnformation frclassification, uand retrieve ane for the new focation is classifi

power manageer to dynamicaraction events.

A energy optimi

Application Cla

ning [15] is a fidence or obser

calculate the prprobability thatrved evidence es from a combind the compatisis (the likelihoon learning base

ng et al. [16].g a set of input fsures, Y = {y1, y

mean (x1) and staeraction (e.g., tozation of inputd on correspondefine a singlefication (y). Thion, medium-iner ultimately allcation specific i

back methods, ions to receive

y events) with itnfortunately, thlobal” UI eventde or the cuime monitor toadcasts were crccurs, when anonfiguration ch). The runtime

global UI eveables for each cnteraction eveneiving global Ud focus, the runom the prevuser interaction

ny stored applicoreground applicied, then the runer, otherwise, i

ally classify the

ization middlew

assifier

form of supervrvations along wrobability of at a particular h(the posterior pination of the pibility of the ood of the evidened on the clas This methodfeatures, X = {y2, …, yn}. In oandard deviatioouch) events. Wt features into ding values of m output measu

he pre-defined anteraction, or lows us to get interaction inten

known as eve and handle utems displayed here is no APIts, which are ofurrent foregrou receive global reated that are sn application gahanges (e.g., he monitor is oents are broadccurrent foregrouts, and the pho

UI events. Whentime monitor wvious foregroun stats) to the acation informatcation. If the nntime monitor wit will invoke e application a

ware framework

vised learning twith prior outcoan outcome. Mhypothesis is trprobability of rior probability

observed evidennce). AURA usessification methd of classificat{x1, x2, …, xn}, tour work, the inon (x2) of the timWe create a sim

low, medium, amean and standure: the resultapplication clas

high-interactia characterizat

nsity. Learning

vent user

on I to ften und UI

sent ains hard only ast. und one n a will und app tion new will the s it

k

that ome

More rue, the

y of nce es a hod tion to a nput mes

mple and

dard ting sses ion. tion for

our classize γTS

input feathe class

The probabilikelihooposteriousing th

The deprobabilhypothesame folikelihooapplicatset. probabilapplicat

Thus, compute

Theretogetherdata poprobabil

where differen

M

Consitraining should bvalue, throw be the trainnew inpinteractihigh-int|y=high=1/6. Sprobabil(1/4)×(3

sifier is accompfor each applic

atures to an outsifier to accuratclassification

ilities for each cod, and evidenc

or probability.he following equ

| ,enominatorlity that an ob

eses, and is irreor every class od, is the priotion class is , |lity of seeing tion class is , |

the posterior ped as follows: e may be some r with an outpuints in the trality. To deal wi| is a smoothing

nt attributes of

TabInput Fea

Mean high high med med low med

ider an examplset shown in

be noted that hus a smaller ma new input fe

ning set, that isput is classifiedion, the posteriteraction)=(1/6)-interaction)=1imilarly, for thlity is: Prob(x3/5)=9/160; an

plished throughcation. Findingtput measure ustely predict the

task consists class, based once, then choosinThe posterior

uation: ↓ ,

, , or tbservation X ielevant for decassignment [1

or probability . This is easily

is the pethe input featu. From Jung et |probability of a

, |cases in which

ut measure due aining set, resuith this, we elimg constant

that have been

ble 2: Example T

atures Std. Dev.

low high med low high high

le of dynamic Table 2. Whe

x1=high denotmean gap betweature (x1=med,s presented to d as follows. Fior probability i)×(1/6)×(1/5)=11/6 and Prob(xhe hypothesis yx1=med,x2=highand for hypo

h creation of a tg a consistent msing this traininclass of an appl of calculatin

n the prior probng the class wit

probabilities

| ,the evidence, is is seen under ision making s5]. of the hypoth

y calculated fromer-input-featureure vector X gal. [16] we hav

an application

h input features to an insufficie

ulting in a zerominate them by s,0 , and is t

n observed.

Training Set Outpu

Classhigh-inmed-inmed-inmed-inlow-inmed-in

classification ben looking at tes a high-inter

ween events. Le x2=high) whicthe classifier. IFor the hypothis: Prob(x1=me1/180 because Px2=high|y=high

y=med-interactih|y=med-interacothesis y=low

training set of mapping from ng set enables lication. ng posterior bability, class th the highest are obtained

the marginal all possible

since it is the , the class

hesis that the m the training e conditional iven that the e: | class may be

do not occur

ent number of o conditional smoothing:

the number of

t Measure sification nteraction nteraction nteraction nteraction nteraction nteraction

based on the this table, it

raction mean et the bottom ch was not in If 1, this hesis y=high-d,x2=high|y= Prob(x1=med h-interaction) ion, posterior ction)=(3/8)× w-interaction,

170

pos=(1fea

Aevedefclarun

F

MstomaandRS,A

acta inrew

ApowstaproappstaclaleaparadjtheWhthrgivto BtheNoevablinBereaimmutilFigtraneveor fredef

sterior probabil1/6)×(1/3)×(1/5ature is thus medAn application ent. Upon beinfined parameterassified, and an n on its next inv

Figure 2

F. MDP Powe

Markov Decisochastic controlaking models fod partly controA), where S is tions available fn state s at time

ward from transAs part of the Awer manageme

ate and below nobabilities of plication classi

ate flow diagramassified and gaarned user evenrameters (e.g.justment levels e probability of hen the probabreshold (PHLT inven utilization tBelow Nominal

e lowest suppominal, the screealuation intervandness. A tranlow Nominal

aches a low-memediate transitilization exceedgure 2). If procensition back to ent approaches

any user eveequency and scrfaults.

lity is: Prob(x1=5)=1/90. The oud-interaction. is classified in

ng assigned ther) number of timappropriate pow

vocation.

: Control Algor

er Manager

ion Processes l processes thaor systems in wolled. MDPs co

a finite numbfrom each statee t will lead to ssitioning to stateAURA framewoent algorithms nominal state at

user interactfication. All th

m illustrated in ains foregrounnt statistics are, evaluation ΔS, probabilitya user event w

bility of an even Figure 2) andthreshold (UHLT l and the proces

ported frequencen brightness isal, taking adva

nsition to the mstate will also

ed utilization thion back to Nom

ds an even highessor utilizationNominal will oa given low-to-

ent occurs, in reen backlight le

=med,x2=high|yutput measure

this manner oe same classifimes, the applicawer manageme

rithms State Flo

(MDP) [17] at are widely

which outcomes onsist of a fouer of states, A

e, PA is the probstate s* at time e s* from state sork, we developto transition bet discrete time ition, processorhree algorithmsFigure 2. When

nd focus, the e used to set

interval τ, y thresholds PT)ithin the next e

ent is below a d processor util

in Figure 2), thssor frequency lcy. While thes gradually adjuantage of the cmid-level frequeo occur if prohreshold (ULMT minal state will her utilization tn is not a dominoccur as the pro-high threshold

which the nevel are set back

y=low-interactiof the new in

n each interactcation m (a usation is marked

ent strategy will

ow Diagram

are discrete tiused as decisiare partly rand

ur-tuple (S, A, is a finite set

bability that actt + 1, and RS,A i

s. three MDP-baetween a nomiintervals based r demands, a

s make use of n an applicationclassification aalgorithm contscreen backli) and speculate

evaluation intervgiven high-to-lization is belowhe state transitiolevel is adjustede state is Belusted down at eaconcept of chanency while in ocessor utilizatin Figure 2) oroccur if proces

threshold (ULHT

nating factor theobability of a u(PLHT in Figure

nominal procesk to user-specif

on) nput

tion ser-d as l be

ime on-

dom PA,

t of tion is a

sed inal on

and the n is and trol ight

on val. low w a ons d to low ach nge the

tion r an ssor T in en a user e 2) ssor fied

Ourdistributand storwere crduring variant ipredict fof a morecent u

ParaEvaInte

Event PThres

ProUtil

ThreshBac

AdjuLev

Event(

Time (

†COV ≡ C

Tablethat canuser andetermininteractiwhile lointerval.applicata measuFor exawill usesavings.patternstowards design tgrained on appliare againbiased tbackligh40%. Foget a glvalue capattern.which dmoving

Figurmanagemevent retransitioprobabilPoint 3,While inchangesprobabiland the

baseline algorition to predict red user-interacreated to dynameach classifiedis event-driven future interactiooving average user interaction w

Table 3: A

ameter Depealuation erval (τ)

ApClas

Probability shold (PT)

ApClas

ocessor lization hold (UT) cklight ustment

vels (ΔS) t Window (WE)

E-AD

Window (WT)

T-AD

Coefficient of Vari

e 3 above lists n be used to cund application.ned based onion applicationow interaction . Event probation classificatioure of the variabample, low intee high event pro. For applicatio, event probab QoS. Processotime; howevercontrol is requication classificn biased toward

towards energy ht levels are alsor the E-ADAPobal view of than be used for A similar argum

dynamically adaverage windo

re 3 illustrates tment algorithm

eaches the low-ons back to Nomlity of a user ev, and the staten the below nos during this plity of a user eprocessor freq

ithm, NORM, user events basction statistics. mically adapt td application and uses the m

on events. The window of sizwithin a tempor

Algorithm Cont

endencies pplication ssification

Disre-

dpplication ssification

Ttra

Static Consta

COV† Screw

DAPT Only Numto

DAPT Only Sizeu

iance

several key alustomize the al

For example, n the applic

ns will require applications c

ability thresholon and coefficiebility of the leareraction applicobability threshons with high vbility thresholdor utilization th, could also b

uired. Screen bacation, where hds QoS and lowsavings. It is im

so limited by a T variant, a larghe user-interacta more local v

ment can be madapts the user-ow of size WT. the basic workinms. At Point 1-to-high probabminal. After an vent falls belowe transitions toominal state, thepredicted user event rises abovquency is adjust

makes use ofsed on learned Two derivativ

to real-time usinvocation. Th

most recent WE uT-ADAPT varia

ze WT, to dynamral context.

trol Parameters

Descriptioscrete time interval-evaluate the systedecide appropriateThreshold to determansitions based on

predictionnditional thresholdate transitions base

demandseen backlight adjus

while in state Below

mber of most receno use for dynamic e of moving averaguse for dynamic ad

lgorithm controgorithms to ea

the evaluationation classific

more frequencan use a largelds are also sent of variancerned user interacations with loholds to bias towvariability in usds will be slihresholds are stbe set dynamicacklight levels high interactionw interaction appmportant to notQoS-driven lowger WE value cation pattern or view of the usade for the T-ADevent statistics

ng of the MDP, the probabili

bility threshold event occurs a

w the high-to-loo the Below Ne screen backliidle time. At

ve the low-to-mted accordingly

f a Gaussian classification

ves of NORM er-interaction he E-ADAPT user events to ant makes use mically track

s

on l in which to

em state and e action(s) mine state user-event n to determine ed on CPU stment levels w Nominal

nt user events adaptation ge window to daptation

ol parameters ch individual n interval is

cation. High nt evaluation, er evaluation set based on , which gives

action pattern. ow variability wards energy er-interaction ightly biased tatically set at ally if finer-are set based

n applications plications are te that screen wer bound of an be used to a smaller WE er-interaction DAPT variant s based on a

-based power ity of a user and the state

at Point 2, the ow threshold,

Nominal state. ght gradually Point 4, the

med threshold y. At Point 5,

171

theandmiNoagathenot

A

IfraHTfor(FrTh2.6MauseinsusiinsandmoSolmewaandtrapowbri

B

DcomfreLinsubproDFdrima

MS528ranmosw

e probability ofd the state trasprediction, in

ominal state. Thain. Mispredictie effectiveness ot diminishing ov

Figure 3:

II

A. Overview

In order to quaamework, poweTC Dream smarr the CPU and broyo) for the H

he Google Andr6.35.9 Linux kachine (VM), aer-space applicstance of the Daing real power strumented the cd measured conitor [18]. Thelutions power t

easurements ovas capable of swd screen backlices from the Mwer measuremightness level.

B. CPU DFS P

Dynamic frequmputing system

equency at runtnux-based systebsystem (cpufreocessor operatioFS drivers, maniver, which is uanually control DThe HTC DreSM7201A proc8 MHz; howevnging from 128odel for the DF

witch frequency

f a user-event ransitions back which a user e

his causes the stions are used asof the algorithmverall QoS.

: Control Algor

II. DEVICE PO

antify the energr analysis was rtphone, with thbacklight. We u

HTC Dream android platform is kernel, and ina variant of Javcations are Dalalvik VM. We b

measurements contact betweenurrent using t

e monitor conneool software th

ver time. We dwitching dynamiight levels overMonsoon Powerments for each

Power Model

uency scaling (ms to conserve time when the ems, such as Aeq) that is a genon. The cpufreq

ny of which are sed in this workDFS levels baseeam smartphocessor. This prver, it support

8 MHz up to 61FS levels, the

y domains. Giv

reaches the lowto Nominal.

event occurs wtate to transitions a coarse QoS ms and ensure e

rithms Timeline

OWER MODELIN

gy-effectivenesperformed on t

he goal of creatuse a variant ofd the Android Scomprised of a

ncorporates theva implementelvik executable

built our power on the HTC D

n the smartphonthe Monsoon

ects to a PC runhat allows real-teveloped a cusic frequency scr time. The corr Tool were thh DFS freque

DFS) is a stanpower by dynasystem is idle Android, provi

neric frameworkq subsystem supconfigurable [1k, allows a usered on a user-dene consists oocessor is nomts 11 different 14 MHz. In ordtest app was

ven the difficu

w-to-high threshPoint 6 shows

while in the Beln back to Nomimetric to evalu

energy savings

e Illustration

NG

ss of the propothe Android bating power modf Android OS 2SDK Revision a slightly modife Dalvik Virt

ed by Google. es that run in estimation mod

Dream device. Wne and the batte

Solutions pownning the Monsoime current/powstom test app taling (DFS) lev

rresponding powhen used to obtency and scre

ndard techniqueamically changor under-utiliz

ide a DFS kerk for managing pports a variety19]. The userspar-space programfined set of rulef a Qualcomm

minally clockedfrequency lev

der to get a powused to manua

ulty of controll

hold s a low inal uate are

sed sed

dels 2.2, 11.

fied tual All an

dels We ery, wer oon wer that vels wer tain een

e in ging zed. rnel the

y of ace

m to es. m® d at vels wer ally ling

CPU ufollowedutilizatiopower cobservedutilizatioLinear coefficieunits of

C. Sc

ScreeMonsoostatic anbacklighrelative MonsoomeasureLinear rwith uni

where ψ

We imsmartphabove wMDP-baany resuAURA fr

• EvetouConStopro(iii)distbur

• Reaappinte

The four imp50; and

utilization expld by a wait on between 0 aconsumption wd measuremenon that lies beregression [20

ents and develomW/% utilizati

PCPU = 0.

creen Backligh

en power conson Power Tool nd dynamic coht levels were

average poweon Power Tool.ed using the toregression was aits of mW:

P

ψ is the backligh

Figure 4: Mo

IV. EXP

mplemented ouhone. The CPUwere integratedased power maulting energy

framework, we o

ent emulator:uch based user nstant – sends chastic – send

obability with a ) Random – sentribution; and (rst of 20 events,

al users: We plications as theeraction events

following algorplementation: fo

for the MDP a

licitly, an intfunction was

and 100% for eawas recorded rnts, linearity etween 0 and 0] was used op our CPU poion:

0024 × freq (M

ht Power Mode

sumption was and our custo

ontrol of screeincremented i

er measuremen. Figure 4 showool as the bacagain used to ob

PSCREEN = 596.

ht level (0.0

onsoon Power M

PERIMENTAL M

ur AURA frameU and backlighd into the AURanagement strasavings. To teobtained stimul

This module interaction ev

events at a useds events usinuser-defined m

nds events usin(iv) High Burst , followed by a

involved five ey would normain the process.

rithm control por the Bayesianlgorithms, WE =

tensive compuused to togg

ach frequency lrespectively. B

was assumed100 % for eac

to determineower estimation

MHz) + 3.0511

el

also measureom test app, wen backlight lein increments

nts were recordws the instantacklight level isbtain a relative

8 × ψ ψ 1.0 .

Monitor Screen

METHODOLOGY

ework on a realht power mode

RA framework, ategies and alsoest the effectivli from two sou

emulates varivent patterns, ier-defined consng a Gaussian

mean and standang a pseudo-ranLong Pause – slong 10 second

users who intally, generating

parameters weren classifier, λ == 6 and WT = 6.

utation loop, gle processor level, and the

Based on the d for CPU ch frequency. e appropriate n model with

1

ed using the which allowed evels. Screen of 10% and

ded with the aneous power s ramped up. power model

Power

HTC Dream els described to guide our

o to quantify veness of the urces:

ious key and including: (i) stant rate; (ii) n distribution ard deviation; ndom uniform sends a quick d pause.

teracted with user-specific

e assumed in = 1 and γTS = . For the state

172

flo= 0= 0

A

IdifourpricomsetbeiThevesucdifcompresatexpmi

Falgaxisavof of eacperdurandis occdueeveadjsligis Nunlan doeevesta

Falgthrthema

w transitions, P0.3/0.4/0.5 (H/M0.5, and ΔS = 0.0

A. Event Emul

In our first studfferent user-inter three MDP-bimarily interestmpared to nomttings on the HTing offered, a

he successful preents that occuccessful predictfficult to quantimmon event ediction rate tisfaction. Thperiments wespredictions, bu

Figure 5: Av

Figure 5 showgorithms for this. It can be seevings (up to 20its dynamic adauser events an

ch evaluation rforms poorly ring the burst ard used to calcunot taken intocurs. On the ote to its time-baent occurrencejusts on each eghtly more comNominal, it wiless CPU utilizevent is very l

es computationent to occur, th

ate transitions toFigure 6 showsgorithm for withree algorithms he robustness anagement algo

PLHT = 0.5/0.6/M/L Classificati05/0.1/0.15 (H/M

V. RE

lation Results

dy, the event emeraction patternbased power mated in the enerminal frequencTC Dream devisuccessful predediction rate me

urred while in tion rate is indicify on a user levery differentlis merely an

he algorithm cere biased ut could be adju

verage Energy

ws the energy e four emulateen that E-ADAP%) due to the raptation. The alnd does not req

interval. Thisin conditions

re stored in the ulate the mean f

account until ther hand, T-ADsed nature. Wh in order to evaluation inter

mputationally inll not drop to B

zation is below low). Because

ns on a regular ihe CPU utilizato Below Nominas average succh the event emuhave high succeof our chos

orithms. A hig

/0.7 (H/M/L Clion), ULHT = 0.8M/L COV).

ESULTS

mulator was uss, to contrast thanagement algorgy savings frocy and default ice. To evaluatediction rate meetric keeps track

Below Nomincative of diminievel – two userly. Thereforen attempt at control parame

towards mausted to allow m

Saved for MDP

savings achieved interaction pPT offers the bresponsive evenlgorithm only aquire constant s also revealslike HBLP - E-ADAPT win

for prediction, bthe event afte

DAPT offers lowhereas E-ADAPTadjust its predrval. This form

ntensive. AdditiBelow NominaUHLT (even if T-ADAPT shifinterval insteadtion is high moal less often. cessful predictiulator patterns. essful predictiosen classificatgher successfu

lassification), P8, ULMT = 0.6, U

sed to imitate fhe effectivenessorithms. We w

om our algorithscreen backli

e the level of Qetric was defink of the numbernal state. A lished QoS. QoSrs may perceive, the success

measuring ueters used in aintaining fewmore.

P Algorithms

ved by the thpatterns on thebest overall enent-triggered natadapts on instan

adaptation durs why E-ADA

the quick evendow as they ocbut the long paer the long pawer energy savinT must wait fordiction, T-ADAm of adaptationonally, if the stl under T-ADAthe probability

fts its window ad of waiting forore often, thus

on rates for eaIt is clear that

on rates, validattion and pow

ul prediction ra

PHLT UHLT

four s of

were hms ight QoS ned. r of low S is

ve a sful user our wer

hree e x-rgy ture nces ring APT ents ccur use use ngs r an APT n is tate

APT y of and r an the

ach t all ting wer ate,

howeverfollowinearlier twill stilBelow Nexceeds until an low agaenergy w(but the reason predictiocharacteADAPTcontrol predictioQoS fopotentiain order

Figure

B. U

A seccommonHTC Drdifferenday, wisession. algorithmwe implalso aimmobile down bo

Figureproposemanagemacross a(but lowcan be eemployeapplicatthe CPUaware aoffer higto the ustime. Fomore enapplicatinteractiexploitesavings

r, does not neceng reason: an than the actual ll be consideredNominal and PLHT, the stateevent occurs an

ain. Therefore,will be saved bprediction willwhy although

on rates dueeristics, it offers

T. It should be nparameters w

on rates, thus mor all three alally be adjusted r to offer even h

e 6: Avg. Succes

User Study Resu

cond study wasn Android appsream device. T

nt applications oith different en

In addition to ms running as lemented the fra

ms to improve Cdevices by us

oth CPU frequee 7 compares thd by Shye et alment techniqu

all users. It can w) energy savinexplained by noed in CHBL is ition type, insteaU frequency anand application-gher energy saser interaction por instance, the nergy comparedtions, respectivion Jewels appled by our algor

for Jewels d

essarily imply hevent may betime of its oc

d successful. Fthe probability

e will switch tond the probabil, if an event decause the Noml still be considh T-ADAPT oe to its sus lower energy

noted that in ourwere biased maintaining a hlgorithms. The

d to accept lowehigher energy sa

ssful Prediction

ults

s conducted, ths and five real uThe users wereover the course nergy saving aour three MDPpart of the AUamework propoCPU and backlising change blency and screenhe average ener. (CHBL) with

ues for the eibe seen that C

ngs across all apoting that the chindependent of ad using constand backlight le-aware algorith

avings because patterns and takE-ADAPT algod to CHBL forely. In some clication, interacrithms. While due to its lac

higher energy sae predicted to currence and thFor example, iy of an eveno Nominal and ity of an event

does not occur minal state will dered successfuloffers the besuperior temposavings than N

r experiments, ttowards high

high minimum e control paramer successful preavings.

n Rate for MDP

his time focususers using these asked to inteof several sess

algorithms enabP-based power

URA middlewareosed by Shye etight energy conlindness to gra

n backlight levelrgy savings of tour three MDPight different

CHBL offers fairpplications. Thehange blindnessuser interaction

ant time triggerevels. In contrahms (specificallthey can dynam

ke full advantagorithm saves up r the Gmail ancases, for instaction slack is toCHBL offers hck of user-aw

avings for the occur much

he prediction if the state is nt occurrence

remain there occurrence is quickly, less be prolonged l). This is the st successful oral locality NORM and E-the algorithm h successful level of user

meters could ediction rates

Algorithms

sing on eight se apps on the ract with the

sions during a bled in each management

e framework, t al. [3] which nsumption for adually ramp ls over time. the technique -based power applications,

rly consistent e consistency

s optimization n patterns and red scaling of ast, our user-ly E-ADAPT) mically adapt ge of user idle

to 4× and 5× nd WordFeud ance the high oo small to be higher energy wareness and

173

appnotpromufreQo5×

Aenefocintawcontran[6]transenGPrecoptthelocimp

enemawoconwitoneautsloHouseapptima uHosysbecdifusepre

Iint

plication-awareticeable QoS doposed AURA fruch as 24% eneequency and scroS, and improv

for some appli

Figure 7: A

A significant aergy optimizatcuses on optimierfaces (e.g., W

ware handoff ansumption of Unsmitting data ], with the goansfers. Other wnsing schemes PS usage. A fcognizes contextimizing the baese works. Whication sensor anplemented alonThere have beeergy consumptiake several imork: (1) the scrnsuming compoth a scheme thae that changes thors implemen

owly reducing sowever, their aper-aware scalinproach tends to

mes can often deuser interactionowever, their apstems, and not tcoming more pfferent in that ies a simple edictor to set CP

In this work weraction aware

eness during degradation iss

framework withergy savings ovreen backlight wes upon prior wcations.

Average Energy

VI. RELA

amount of worion for mobilizing energy coWiFi, 3G/EDGalgorithm is pUMTS (3G) aover different wal of balancing

work [7]-[9] focuaiming to redu

framework for xtual user states acklight and Cile we do not ond wireless int

ngside other straen some effortsion in recent ye

mportant observreen and the Conents, and (2) uat gradually reduabruptly. Base

nt a scheme thscreen brightnespproach does nong like AURA o be overly conetrimentally im

n-aware DFS appproach is diretowards the tourevalent today.t does not consuser-aware bu

PU frequency le

VII. CON

we proposed A energy optimi

scaling, it cosues for the u the E-ADAPT

ver the default dwithout noticeabwork (CHBL [3

y Saved for Rea

ATED WORK

rk has been doe devices. Mu

onsumed by theGE networks). Iproposed basedand WiFi. The wireless interfag energy and duses on energy-uce high battery

mobile sensinis proposed in PU energy is cptimize energyerface, AURA

ategies that do. s to optimize Cears. Shye et a

vations that areCPU are the twusers are generauces screen brigd on the secon

hat utilizes chass and CPU freot consider appl

does, becausenservative at ti

mpacts user QoSpproach in [4] ected towards such-based mobi Their proposedsider backlight ut applicationevels based on d

NCLUSIONS

AURA, an appization framew

omes at a couser. Overall, scheme enablesdevice settings bly degrading u3]) by as much

al Applications

one in the areauch of this we device’s wirelIn [5] an enerd on the ene

energy costs aces is exploreddelay during d-efficient locatiy drain caused

ng that efficien[10]. Our workcomplementary

y consumed by can potentially

CPU and backlil. [3] in particue relevant to two largest powally more satisfghtness rather th

nd observation, ange blindnessequency over timlication-aware ae of which thimes, and at otS. Bi et al. prop

much like AURstationary deskle devices that d approach is aoptimization, a

n-unaware histdemand.

plication and uwork for pervas

ost: our s as for

user h as

a of ork less rgy-rgy for

d in data on-by

ntly k on y to

the be

ight ular this wer fied han the by

me. and heir ther ose RA.

ktop are

also and ory

user sive

mobile dfor clasusing Ba(MDP)framewoand foundependinQoS. Wpower sas 5× thdevelopquantify

[1] Mobcom

[2] MicDevl-12

[3] A. SActiArch

[4] M. Inte

[5] H. Esti

[6] M. RMO

[7] I. CMob

[8] K. Lloca

[9] Z. ZLoc

[10] Y. WAut200

[11] S. KCom

[12] D. Jthe a

[13] D. Jfutu

[14] Goo[15] E.

Mas[16] H. J

Tech183

[17] T. LDecMob

[18] MonLab

[19] D. BDocMar

[20] J.J. [21] HTC

spec[22] W. L

DFS6, S

[23] S. PMan

[24] N. Batthand

devices. As parssifying applicaayesian classifibased power mork on both emnd that it can ang on the appli

When comparedsavings in mobitheir energy saing better pred

ying lifetime im

VbiThinking, “Glo

m, April 2011. cro Power White Pvices,” http://www274743697.pdf [OnShye, B. Scholbroivity Patterns thitectures,” In MIBi, I. Crk, C.

eractive ApplicatioPetander, “Ene

imation,” In MICNRa, et al., “Energy

OBISYS ’10, pp. 25Constandache, et bile Phones,” In INLin, et al., “Energation,” In MOBISYZhuang, K. Kim,

cation Sensing on SWang, et al., “A Ftomatic User State9.

K. Card, T. P. Momputer InteractionJ. Simons, S. L. Frabsence of a visuaJ. Simons, and R. ure”, Trends in Cogogle Android, offic

Alpaydin, “Intrssachusetts: The MJung, M. Pedram, hniques by Using, Mar. 2008. L. Cheung, K. Ocision Process (Mbile Phones,” In Ensoon Solutions

bEquipment/PowerBrodowski et al., cumentation,” hrch 2011. Faraway, “Linear C, “HTC G1 Ocification.html Liang, P. Lai, “DeS Mechanism for t

Sept. 2010. P. Tarzia, P. A. nagement PoliciesVallina-Rodriguetery Statistics: Udsets,” In MOBIH

rt of AURA, weations based oication, and thremanagement algmulated and reachieve averageication, without

d with prior wile devices [3], avings. Our ondiction techniqu

mprovement.

VIII. REFERE

obal mobile statis

Paper, “Using Lithw.micro-power.conline].

ock, G. Memik, “Into Guide Powe

ICRO-42 ’09, pp. 1Gniady, “IADVS

ons,” In HPCA ’10ergy-Aware NetwNET ’09, pp. 55-60y-Delay Tradeoffs5-270, Jun. 2010. al., “EnLoc: En

NFOCOM ’09, ppgy-accuracy trade-YS, pp. 285-298. Ju, J. P. Singh, “ISmartphones,” In MFramework of Enee Recognition,” In

oran, and A. New,” Hillsdale, NJ: Lranconeri, and R.

al disruption”, In PA. Rensink, “Chagnitive Sciences, Vcial website, http:/troduction to MMIT Press, 2010.

“Improving the Eg Bayesian Classif

Okomoto, F. MakeMDP) Framework

EMSOFT ’09, pp. 1Inc., official w

rMonitor, 2008. “Linux CPUFreq

http://mjmwired.ne

Models with R”, COverview,” http:

esign and Implemethe Android Opera

Dinda, R. P. Dis in Practice,” In ICz, P. Hui, J. Cr

Understanding thHELD ’10, pp. 9-14

e introduced a non user interacee Markov Decgorithms. We eal user interacte energy savingt perceptible im

work on CPU aAURA can ach

ngoing work is ues and battery

ENCES stics 2011,” http:

hium Polymer Battom/userfiles/file/wp

nto the Wild: Studer Optimizations168-178, Jan. 2009S: On-Demand P0, pp. 1-10, Apr. 20work Selection 0, Sept. 2009. s in Smartphone A

nergy-Efficient L. 19-25, Jun. 2009-off for continuousun 2010. Improving EnergyMOBISYS, pp. 315ergy Efficient MobMOBISYS ’09, pp

well, “The PsycholLawrence Erlbaum

L. Reimer, “ChanPerception, 29:114ange blindness: PaVol.9 No.1 Januar//www.android.comMachine Learnin

Efficiency of Powfication,” In ISQED

er, X. Liu, V. Ak for Optimizing11-20, Oct. 2009.website, http://ww

q – CPUFreq Govet/kernel/Documen

CRC Press, 2004//www.htc.com/w

entation of a Criticating System,” In E

ick, G. Memik, “CAC ’10, pp. 51-6rowcroft, A. Ric

he energy deman4, Aug. 2010.

novel method ction patterns cision Process evaluated the tion patterns, gs up to 24% mpact on user and backlight hieve as much

focusing on y models for

://mobithinking.

teries in Mobile p_polymer_fina

dying Real User s for Mobile 9.

Performance for 010. Using Traffic

Applications,” In

Localization for 9. s mobile device

y Efficiency of 5-330,. 2010. bile Sensing for p. 179-192, Jun.

logy of Human-m Assoc., 1983.

nge blindness in 43–1154, 2000. ast, present, and ry 2005. m ng,” 2nd ed.

wer Management ED ’08, pp. 178-

Akella, “Markov g Software on

ww.msoon.com/

vernors. Linux ntation/cpu-freq,

www/product/g1/

cal Speed-based EMS ’10, pp. 1-

“Display Power 0, Jun. 2010.

ce, “Exhausting nds on mobile

174