Transfer Time, Energy, and Quota-Aware Multi-RAT Operation...

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 65, NO. 1, JANUARY 2016 307

Transfer Time, Energy, and Quota-Aware Multi-RATOperation Scheme in Smartphones

Wonbo Lee, Student Member, IEEE, Jonghoe Koo, Student Member, IEEE, Yongseok Park, andSunghyun Choi, Fellow, IEEE

Abstract—Today’s smartphones integrate multiple radio ac-cess technologies (multi-RAT), e.g., third-generation (3G), fourth-generation (4G), Wi-Fi, and Bluetooth, whereas the number ofintegrated RATs is increasing. Therefore, it is becoming moreimportant to select the best RAT set among the available RATsand determine how much data to transfer via each selected RATnetwork. We propose a multi-RAT interface activation algorithmwith supporting system design for smartphones’ file transfer ser-vice (e.g., downloading a movie file). We model a multiattributecost function incorporating file-transfer completion time, energyconsumption, and data usage quota together. The goal of thispaper is to find out the optimal multi-RAT set to be activatedwith the corresponding file segment allocation that minimizes thecost function under given energy and quota constraints. By theproposed algorithm, the suboptimal solution can be found withoutsignificant performance degradation from the optimal solutionwhile making the computational complexity linear in the numberof RATs.

Index Terms—Energy and quota-aware network selection,LTE/WiFi, MADM, multi-attribute, multipath, multi-RAT.

I. INTRODUCTION

A S the number of available radio access technologies(RATs), e.g., third generation (3G), fourth generation

(4G), Wi-Fi, and Bluetooth, in today’s smartphone is increas-ing, users are being given more options for RAT selection for aspecific service. Furthermore, state-of-the-art smartphones arecapable of the parallel transmission by concurrently activat-ing multiple RATs such as Long-Term Evolution (LTE) andWi-Fi to boost up the download speed [1], [2]. Therefore, itis important to develop a strategy to choose the best RAT set tomaximize the user satisfaction or to minimize the cost for dataservice.

Manuscript received August 2, 2014; revised November 6, 2014; acceptedDecember 20, 2014. Date of publication January 21, 2015; date of currentversion January 13, 2016. This work was supported by the project entitled“Development of Context-aware Cross Layer Protocol for 5G Network,”funded by Digital Media & Communications R&D Center, Samsung Elec-tronics, and the Brain Korea 21 Plus Project. An earlier version of this paperwas presented at IEEE WoWMoM 2014, June 2014. The review of this paperwas coordinated by Dr. P. Lin.

W. Lee was with Department of Electrical and Computer Engineering, SeoulNational University, Seoul 151-744, Korea. He is now with Digital Media andCommunication R&D Center, Samsung Electronics, Suwon, 443-742, Korea.

J. Koo and S. Choi are with the Department of Electrical and ComputerEngineering, Seoul National University, Seoul 151-744, Korea.

Y. Park is with Digital Media and Communication R&D Center, SamsungElectronics, Suwon, 443-742, Korea.

Color versions of one or more of the figures in this paper are available onlineat http://ieeexplore.ieee.org.

Digital Object Identifier 10.1109/TVT.2015.2395132

These days, smartphone users are becoming more sensitivenot only to the perceived performance of the service qualitybut to the service charges that they have to pay as well. More-over, the growing appetite for large-volume multimedia dataapplications likely leads to faster battery drainage. Therefore,we propose a multiattribute cost-based multi-RAT interfaceactivation scheme that enables smartphones to automaticallychoose the optimal RAT set among the available n RATs.

We begin by first formulating a multiattribute cost functionconsidering n RAT-based smartphone characteristics. In par-ticular, we incorporate three attributes, namely, file transfercompletion time T , energy consumption E, and data usagequota Q, that are considered most important for typical smart-phone users, into the cost function. The three cost terms T , E,and Q are merged into the multiattribute cost function based onthe simple additive weighting (SAW) method, which is widelyused for multiattribute decision making (MADM) algorithm[3], [4].

With this multiattribute cost function as the primary decisionvehicle, the objective of the proposed scheme is then to de-termine which RAT interface(s) should be activated consid-ering the energy and delay overhead for turning-on/off theseinterface(s) and then calculate how much data should betransmitted through each selected RAT path to minimize theoverall cost. The proposed algorithm enables n RAT-basedsmartphone to adaptively activate and select the optimal RATset to minimize the multiattribute cost according to the file sizeand estimated throughput, while satisfying the energy and costrequirements.

In our previous work [5], we studied an adaptive networkinterface activation scheme considering LTE/Wi-Fi-enabledsmartphones. The multiattribute cost function with regard tothe transfer completion time, energy consumption, and servicecharge are modeled for three modes, namely, LTE-only mode,Wi-Fi-only mode, and LTE/Wi-Fi parallel mode, based on thedelay and energy measurement of commercial smartphones.However, the previous study is only limited to the two-RATcase, and the issues resulting from extension to the case of morethan two RATs are not considered.

For n RAT-based interface activation scheme, a generalizedproblem formulation with n RATs, where n ≥ 2, is needed.Furthermore, a low-complexity algorithm is needed for thedynamic update of the RAT selection during run time becausethe search space to find the optimal RAT set is exponentiallyproportional to the number of available RATs.

In this paper, we formulate an objective function composedof two subproblems: 1) a piecewise linear optimization problem

0018-9545 © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

308 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 65, NO. 1, JANUARY 2016

for the optimal segment allocation to activated RAT pathsand 2) a combinatorial problem to select the optimal RATset considering parallel transmission over multiple RATs. Ourmain contributions are as follows.

• We generalize the multiple network interface activationproblem for n RAT-equipped smartphones. We show thatthe generalized version is the joint combinatorial andpiecewise linear minimization problem that can be stillproperly optimized.

• We propose a heuristic algorithm to find the optimalRAT set and corresponding segment allocation with lowcomputational complexity. Although the number of pos-sible RAT combinations exponentially increases as thenumber of available RATs increases, the proposed linearsearch algorithm reduces the size of the search space fromexponential to linear.

• We also propose a dynamic update algorithm that adaptsthe selected RAT set and segment allocation during thetransfer according to time-varying network conditions.

In the performance evaluation, we verify that the proposedalgorithm shows no significant performance difference from thefull search algorithm. Furthermore, we show that the parallelactivation scheme improves the RAT diversity gain comparedwith the vertical handoff. With time-varying throughput vari-ation of RATs, the performance enhancement through the dy-namic update algorithm is also evaluated.

The remainder of this paper is organized as follows. Wediscuss the related work in Section II. In Section III, the n RAT-based system model in consideration is described. We presentproblem formulation and model the multiattribute (T−E−Q)cost function in Section IV. Then, we analyze the character-istics of the problem in Section V and propose a heuristicalgorithm in Section VI. The performance of the proposedalgorithm is evaluated via simulations in Section VII. Finally,we conclude this paper with the discussion on future work inSection VIII.

II. RELATED WORK

Network selection schemes in heterogeneous networks canbe classified according to two different points of view: network-centric and user-centric approaches.

The network-centric approaches consider the overall systemperformance in a centralized or distributed manner. Severalstudies [6]–[8] assume a centralized controller that controlseach user’s network selection based on an integrated cost func-tion. A cost-function-based network selection strategy in [6]proposes a strategy to achieve a balanced point between the callblocking probability and the average received signal strength.In [8], a system capacity maximization strategy is proposedbased on resource allocation considering parallel transmissionby activating multiple RATs. Recently, in [7], more diverseattributes have been considered such as spectral efficiency,energy consumption, and fairness. A central global resourcecontroller has been proposed, which manages the resourcesof several heterogeneous wireless networks to balance thesemultiple attributes in a system operator’s perspective.

The decentralized network-centric approach is usually basedon a cooperative [9] or noncooperative game [10]. Thisapproach focuses on the relationship or interaction amongplayers that can be users or networks considering payoffs thatresult from their actions. The game-theoretic algorithms alsoconsider system-wide performance such as call blocking, sumthroughput, fairness, and convergence time.

On the other hand, the user-centric approach is also widelystudied based on an always best connected scheme [11]. In thisapproach, a cost or utility function is also formulated in termsof various attributes such as user’s signal quality, throughput,delay, energy, service charge, and security. While earlier studiesmostly focus on improving user’s throughput or delay per-formance by applying appropriate vertical handoff algorithms[12], [13], the energy consumption and service charge have alsorecently been investigated and considered in [14]–[16].

To consider those multiple attributes for network selection,MADM is widely used [3], [4], [17]. Several MADM methodssuch as SAW, multiplicative exponential weighting, techniquefor order preference by similarity to ideal solution, analytichierarchy process, and Grey relational analysis are popularlyused for MADM. These methods are basically used to selectthe best single network by ranking networks based on weightedmultiple attributes. However, our approach takes into accountthe simultaneous usage of multiple networks, as well as a singlenetwork selection, and hence, we adopt a SAW method foreasy modeling and analysis of the detailed multiattribute costfunction.

Recently, energy-efficient multipath transmission controlprotocol (MPTCP) schemes that exploit multiple networkpaths, e.g., cellular and Wi-Fi, in mobile devices have beenproposed [18]–[20]. They improve the energy efficiency (Mb/J)of MPTCP rather than the aggregate throughput. In this paper,we focus furthermore on balancing the performance, energyconsumption, and data quota usage by finding the optimalRAT set among available n RATs and file segment allocationaccordingly.

III. SYSTEM MODEL

Fig. 1 shows the system model that supports parallel orswitched data transfer through selected RAT path(s) among navailable RATs. To fully utilize each RAT bandwidth, a largedata file is split into a number of small fixed-size segments,e.g., 100 kB, and a set of such segments are allocated toeach selected RAT path and transmitted by the server. Thereceiver combines segments received via multiple RAT pathsto reconstruct the original file in the application layer. Theprocess can be supported by the existing multiple flow-basedapplication layer protocols such as the segmented file transferthat is widely used in peer-to-peer file-sharing systems [21] andthe HTTP range request to download a single file over multiplelinks [22]. These protocols are based on the client-side request;hence, the file can be transferred through end-to-end multiplelinks only with the supporting applications in client and serverterminals.

We only focus on the file download case supported by anend-to-end application layer protocol in the system model since

LEE et al.: MULTI-RAT INTERFACE OPERATION SCHEME IN SMARTPHONES 309

Fig. 1. Multi-RAT file transfer model.

the file download is more dominant than the upload counterpartin real life. To support the upload case, the request-basedend-to-end protocol is not enough, and it would need a newprotocol, e.g., a push-based protocol or a proxy server locatedbetween the client and the server to aggregate file segmentstransmitted through multiple RAT paths. Such a protocol issueis outside the scope of this paper.

On the other hand, for the energy, service charge, andperformance awareness, the quality-of-service (QoS) equalizerobtains weighting and normalization factors based on userpreference and device status for each QoS term, respectively.The device status includes information on the remaining batteryenergy and remaining data quota, and the user preferenceinvolving weighting factors for the transfer completion timeT , energy consumption E, and data quota usage Q, is setthrough the QoS equalizer, as described in Fig. 1. Then, the filesegment allocation manager calculates the multiattribute costfunction based on the weighting and normalization factors anddetermines which RAT set is selected and how many segmentsare requested for each selected RAT path.

IV. PROBLEM FORMULATION

For the optimization problem, we define NA as aset of all the available RATs and SA as a set of allthe possible RAT combinations. For example, a smart-phone with NA = {Wi-Fi, 3G, 4G} will have SA ={{Wi-Fi}, {3G}, {4G}, {Wi-Fi, 3G}, {Wi-Fi, 4G}} if 3G and4G cannot be activated at the same time. Assume that a setincluding the selected RAT(s) is denoted by set S(∈ SA),and x is a (n× 1) file segment allocation vector for theselected n RATs, where each element xi denotes the allocatedsegment size transferred via RAT i. Then, the multiattributecost function FS(x) is represented by the normalized sum ofthe transfer completion time, energy consumption, and datausage quota

FS(x) = αTS(x) + βES(x) + γQS(x) (1)

where TS(x), ES(x), and QS(x) are transfer completion time,energy consumption, and data usage quota, respectively. α, β,and γ are weighted normalization factors to integrate all theseattributes into the cost function, where each one is representedby the multiplication of the user-defined weighting factorsω{t,e,q} and normalization factors as follows: α = ωt/Tmax,β = ωe/Emax, and γ = ωq/Qmax. By these normalization

factors, each cost term is represented by the ratio of the costto the maximum cost, i.e., Tmax, Emax, and Qmax. In [5],we apply the maximum tolerable transfer completion time, theremaining battery energy, and the remaining data quota for themaximum cost of each cost term. In addition to the normaliza-tion factor, a user can require the limitation of energy and datausage, which is denoted by Ereq and Qreq, respectively, for thegiven file transfer service.

Therefore, the optimization problem with the normalizedmultiattribute cost function and required energy/quota con-straints is formulated as follows:

minS∈SA,x

FS(x)

s.t. ES(x) ≤ Ereq, QS(x) ≤ Qreq∑i∈S

xi = Btr xi > 0∀ i ∈ S (2)

where Btr is the size of the file to be transferred. The objectivesof the problem are to find out the best RAT set S∗ with thecorresponding segment allocation vector x∗. In the succeedingsection, we briefly explain the T−E−Q cost model for n RAT-based file transfer services and present the proposed optimiza-tion problem.

A. T−E−Q Cost Modeling

1) n RAT-Based Transfer Completion Time T Cost: As-suming that a device can access multiple RATs simultaneously,the completion time of the file transfer is determined by thelongest time among the selected RAT path(s). Therefore, thefile transfer completion time TS(x) can be modeled as follows:

TS(x) = maxi∈S

(ttri + tswi

), where ttri = xi/ri (3)

where ttri is the transfer completion time through RAT i for thesegment size xi, which is a part of the Btr-byte file. ri denotesthe achievable throughput of RAT i, and tswi represents theturning-on delay of RAT interface i, including the access delay,until it can actually transfer data. This is a state-dependentvariable, which is equal to zero if the corresponding network isalready connected when the file transfer is about to start. Fromthe measurement study [5], the turning-on delay is different foreach RAT, and we assume that the value is initially given bymanufacturers or obtained by smartphone’s self-training.

2) n RAT-Based Energy Consumption (E) Cost: To transmitor receive packets, the battery energy is consumed by CPUand activated RAT interfaces. When generating packets, theenergy is consumed by CPU to read/write data buffers andto attach/detach TCP, Internet Protocol, and medium accesscontrol headers. Moreover, each RAT interface consumes en-ergy to transmit/receive packets through the wireless channel.Therefore, we take these factors into account for the energyconsumption model written as ES(x) = Enet

S (x) + EcpuS (x),

where EnetS (x) and Ecpu

S (x) represent the energy consumptionat the activated RAT interfaces and CPU, respectively. Thedisplay power is not affected by the change of activated RATinterfaces, and we assume that the display is turned off during


the file transfer service. If needed, we can simply add anapproximated constant power level according to the currentdisplay information, such as the brightness or RGB valuesof pixels.

The energy consumption of each RAT interface is generallycomposed of the base energy, transmission/reception energy,promotion/tail energy,1 and turning-on/off energy. Once thesecomponents are modeled, we can assume that the parameterscan be applied to all the same types of RAT interfaces. The en-ergy consumption of selected RAT interface(s) for data transfercan be expressed as follows:

EnetS (x) =

∑i∈S

{pneti (ri)

xi

ri+ eswi + eohi

}(4)

where pneti (ri) is the average power consumption for datatransfer through RAT i, and it is modeled by the function of theachievable throughput ri. In [23], Huang et al. showed that thepower consumption of 3G, LTE, and Wi-Fi for both of the up-link and downlink transmissions, and their throughput–powercurves are approximately fitted to a linear function of thethroughput variable. Therefore, we apply the linear fit model forthe power consumption according to the achievable throughputfor RAT i as follows: pneti (ri) = airi + bi, where ai and biare the linear fit coefficients for RAT i. eohi is the aggregateenergy overhead including promotion and tail energy to transferfile segments through RAT i. eswi is the state-dependent energyoverhead due to turning on RAT i’s interface, which is set tozero if the RAT interface is already turned on.

On the other hand, it is known that the CPU power islinearly proportional to the CPU usage [24], which is in turnlinearly proportional to the packet generation rate. The packetgeneration rate is proportional to the network throughput;hence, the CPU power can be also approximated by a linearfit model corresponding to the sum data rate. Fig. 2 describesan example of the parallel transmission through n RAT paths,where a smartphone initially connected to RAT-1. As shownin the figure, the horizontal (time) line is divided into 2n− 1sections, where each section is determined by a different set ofactivated RAT paths. Accordingly, each section has a differentsum rate determined by the activated RAT path set. Assumingthat section k of duration tk achieves the sum rate of Rk, theCPU energy consumption model is formulated as follows:

EcpuS (x) =

2n−1∑k=1

(cRk + d)tk = c

2n−1∑k=1

Rktk + d

2n−1∑k=1

tk

= cBtr + dTS(x) (5)

where c and d are the linear fit coefficients—we apply c =3 mW/Mb/s and d = 465 mW referring to [5]. As the equationshows, the CPU energy consumption is neither related to thenumber of sections nor the corresponding sum rate but is onlyrelated to the total file size and transfer time.

1For a power save mode operation, wireless interfaces wake up only whendata transfer is requested. After the data transfer is completed, it waits fora certain duration before going back into sleep mode. The energy overheadfor the former and latter cases are referred to as the promotion and tail energy,respectively.

Fig. 2. Parallel transfer model with n RATs.

3) n RAT-Based Service Quota (Q) Cost: Some RATs, suchas 3G and 4G, incur the service charges, whereas other RATssuch as Wi-Fi and Bluetooth are free of charge. The servicecharge according to the file segment size for each RAT path ismodeled as follows:

QS(x) =∑i∈S

qixi (6)

where qi is the service charge rate per byte for RAT i, whichis zero for the free network. If the service charge is Qp for Bp

bytes per month, we model qi = Qp/Bp.

B. Optimization Problem

Integrating all the T , E, and Q cost terms in (3), (4), (5),and (6) into (2), the normalized cost function can be formulatedas a piecewise linear function of the segment allocation vectorx, i.e., FS(x) = max(uT

i x+ vi). Then, the objective functionwith constraints in (2) is represented as follows:

minS∈SA,x

{maxi∈S

(uTi x+ vi)

}

s.t. maxi∈S

(gTi x+ hi) ≤ Ereq, qTx ≤ Qreq

∑i∈S

xi = Btr, xi > 0 ∀ i ∈ S (7)

where

ui =α+ βd

riIi + β (a+ b) + γq

vi =(α+ βd)tswi + β∑j∈S

(eswj + eohj ) + βcBtr

gi =d

riIi + (a+ b) , hi = dtswi +

∑j∈S

(eswj + eohj ) + cBtr

a =(a1, . . . , an),b = (b1/r1, . . . , bn/rn),q = (q1, . . . , qn)

with n being the number of elements in set S, and Ii beingthe indicator vector that the ith element is one and all theother elements are zero. Then, the objective function formstwo subproblems: 1) a piecewise linear minimization to findthe optimal segment allocation vector x∗ with a given RAT setS and 2) a combinatorial problem to find the optimal RAT setS∗ among all the candidate RAT sets SA. The piecewise linearfunction is known to be convex, and the minimization problemcan be solved transforming it to equivalent linear programming


(LP) by forming the epigraph problem [25] with variable tS asfollows:

min f

s.t. uTi x+ vi ≤ f, gT

i x+ hi ≤ Ereq

qTx ≤ Qreq,∑i∈S

xi = Btr, xi > 0 ∀ i ∈ S. (8)

The constraint xi > 0 can be relaxed as xi ≥ 0 to properlysolve the optimization problem. If a solution of Subproblem1 for RAT set S contains xi = 0, where i ∈ S, this solutioncan be ignored because the optimal solution of the objectivefunction would be surely in the set S\i. The optimal RAT setS∗ is obtained by comparing the normalized costs provided bysolving Subproblem 1 for all the possible RAT set S ∈ SA.

V. NUMERICAL ANALYSIS

Taking a look at the objective function in (7), let us assumethat uT

i x∗ + vi is the maximum for the optimal segment alloca-

tion vector x∗, and uTj x

∗ + vj is the second maximum, wherexi > 0 and xj > 0. Then, the difference between these two val-ues is represented as Δfij = (α+ βd)((xi/ri)− (xj/rj) +tswi − tswj )). If Δfij > 0, it can be reduced to zero by decreas-ing xi and increasing xj maintaining xi + xj as the same size.Because the cost is linearly proportional to xi, decreasing xi

means that the optimal cost is also decreased. This is a con-tradiction with the original value being the optimal. Therefore,we can conclude that Δfij = 0, i.e., uT

i x∗ + vi = uT

j x∗ + vj ,

∀ i, j ∈ S, i �= j, with the optimal segment allocation vector x∗.From this equation and

∑i∈S xi = Btr, the segment size xi is

calculated as follows:

xi =ri∑

k∈S rk

⎛⎝Btr −

∑k∈S\i

(tswi − tswk )rk

⎞⎠ ∀ i ∈ S. (9)

This is the optimal segment allocation if it satisfies the otherconstraints in (7) and the cost for RAT set S is the minimumamong all the sets in SA. When (9) violates any constraints in(7), the algorithm starts to iteratively find the optimal solutionof (7).

We apply Sedumi [26], which is a well-known convex op-timization tool based on the interior point method, to solveSubproblem 1 for all the possible RAT combinations and findthe optimal RAT set that minimizes the cost function. For simu-lations, we apply coefficients of the throughput–RAT interfacepower curve and throughput–CPU power curve for 4G (LTE)and Wi-Fi, respectively, from the case study presented in [5].In addition, the employed coefficients for 3G downlink area3g = 35 mW/Mb/s and b3g = 810 mW, which are obtainedfrom Samsung Galaxy S2 HD LTE, the model used for themeasurement reported in [5]. We set the average throughput for3G, 4G, and Wi-Fi networks as 5, 25, and 15 Mb/s, respectively,for the baseline simulation. The battery energy for Emax is setto 2000 mAh, i.e., 7400 mWh at 3.7 V, and the data quota forQmax is assumed to be 10 GB. In addition, the turning-on delayof each network interface is assumed to be 2, 2, and 7 s for 3G,4G, and Wi-Fi, respectively.

Fig. 3. Segment allocation with initial connection of 3G.

Fig. 4. File size to relative cost with initial connection of 3G.

First, we consider the case that the smartphone is initiallyconnected to 3G. Fig. 3 shows a simulation result of the optimalsegment allocation of each RAT path for varying file sizes.Fig. 4 shows the relative cost of each RAT selection with thecorresponding optimal segment allocation while normalizingthe minimum cost for each file size to one. Each color box inthe bottom of the figure represents the optimal RAT set for thecorresponding file size range.

In the figure, the relative cost of the currently connected net-work, i.e., 3G, for small size files is the least among the relativecosts of all the possible RAT sets. This is because the delayand energy overhead to activate the other RAT interface(s) isrelatively too high, compared with the actual file transferringcost (Area 1). On the other hand, as the file size increases, itbenefits from additionally activating the other RAT interface(s)and boosting up the download speed, meaning to decrease Tcost (Area 2). However, when either the file size is very largeor the transfer energy and/or data quota meets the requiredlimitation, the optimal RAT set converges to RAT(s) that arefree of charge and consume lower energy (Area 3).

In this scenario, we assume that 3G and 4G are compatible,i.e., able to be activated simultaneously. However, if 3G and 4Gare incompatible, i.e., unable to be simultaneously activated,the next best RAT set, i.e., {3G, Wi-Fi} or {4G, Wi-Fi}, will beselected instead of {3G, 4G} or {3G, 4G, Wi-Fi} for that range.Our proposed algorithm, which is explained in the following


Fig. 5. Network selection and segment allocation in various conditions. (a) Initial 4G connection case. (b) Initial Wi-Fi connection case. (c) High 4G throughputcase. (d) Limited energy/data case.

section, takes into account the compatibility test; hence, thisissue can be resolved.

Fig. 5 describes the other cases of the RAT selection andsegment allocation according to the various device conditions.The optimal RAT set for each file size is represented in the colorbox, where 3G, 4G, and Wi-Fi are abbreviated to 3, 4, and W,respectively. Fig. 5(a) presents a case that the initial connectionis 4G instead of 3G, and the other condition is the same as thebaseline simulation. In this case, the portion of the {4G} rangeincreases because of its higher throughput (25 Mb/s) than thatof 3G (5 Mb/s). A case that the initial connection is Wi-Fi isalso presented in Fig. 5(b). In this case, only two kinds of aRAT set, i.e., {Wi-Fi} and {3G, 4G, Wi-Fi}, are chosen.

Fig. 5(c) presents the case that the 4G throughput is veryhigh, i.e., 80 Mb/s, which is 16 times higher than that of3G when the initial connection is 3G. In this case, the {3G,4G} range increases, and the segment size allocated to 4Gis far more increased to make full use of its high throughputperformance. However, the {3G} range remains because it stillhas an advantage for the small size file due to the turning-ondelay and energy overhead of the other interfaces.

On the other hand, if the energy and data budgets are ex-hausted or the Wi-Fi throughput is very high, the optimal RATset changes, as shown in Fig. 5(d). In this case, the remainingenergy is 50%, the remaining data quota is 10%, and theWi-Fi throughput is 50 Mb/s. Then, because Wi-Fi is free ofcharge and consumes less power while performing even better,the {Wi-Fi} range is hence much more extended. The onlyshortcoming of Wi-Fi is the longer turning-on delay, i.e., 7 s;therefore, the {3G} and {3G, 4G} ranges still exist when thefile size is relatively small.

VI. PROPOSED ALGORITHM

First, we propose a heuristic algorithm, called bidirectionallinear search (BLS), which finds the optimal RAT set with

optimal segment allocation vector by solving (7) with low com-putational complexity. It considers the characteristics discussedearlier to reduce the search space of the available RAT set.Second, we propose a dynamic update algorithm that adaptsthe selected RAT set and segment allocation vector over timeaccording to the time-varying network condition.

A. Bidirectional Linear Search Algorithm

The BLS algorithm reduces the search space of candidateRAT sets by selectively visiting possible RAT combinations.It searches candidate RAT sets from the single-RAT set to theset including all the available RATs by adding or subtractinga RAT sequentially. At first, two types of RATs are visited,i.e., the currently connected RAT and another RAT that is freeof charge or consumes the lowest power. If the former one isdifferent from the latter one, the search direction is split intotwo directions, which is the reason why the proposed algorithmis called BLS. If both RATs are the same, then the algorithmsearches only in one direction. The detailed procedure is ex-plained in the following.

At first, n available RATs are sorted in two ways: in ascend-ing orders of the turning-on delay and the E−Q cost, i.e., thenormalized sum of the energy and quota cost, respectively, asshown in the following:

Dπ1< Dπ2

< · · · < Dπn

Cψ1< Cψ2

< · · · < Cψn

(10)

where Dπiand Cψi

are the turning-on delay and E−Q costfor the available RATs πi, ψi ∈ NA ∀ i ∈ {1, 2, . . . , n}, respec-tively. Then, RAT π1 has the minimum turning-on delay, whichis likely to be the currently connected RAT, and RAT ψ1 has theminimum E−Q cost at the current smartphone state. It startsthen to find the minimum cost function with candidate RATsets that are chosen in the order from {π1} to {π1, . . . , πn}


by adding a RAT to the previous RAT set like {Area 1 →Area 2} in Fig. 4, i.e., the search direction is from the shortturning-on delay to the performance-oriented way. A similarprocess is executed with candidate RAT sets chosen from {ψ1}to {ψ1, . . . , ψn} such as (Area 3 → Area 2) in the figure, i.e.,the search direction is from the low energy/quota overhead tothe performance-oriented way. If a newly added RAT is not ableto be concurrently activated with RATs in the previous set andthe turning-on delay or E−Q cost of the new one is smallerthan that of the previous one, the newly added one replaces theprevious one. The details of the proposed algorithm, i.e., theBLS algorithm, are presented in Algorithm 1.

Algorithm 1 Bidirectional Linear Search Algorithm

Initialize:1: Btr ← getFileSize()2: Sπ ← π1, Sψ ← ψ1,Other variables ← ∅3: (F ∗, S∗) ← ( min

πi∈NAFπi

(Btr), argminπi∈NA

Fπi(Btr))

4: Sc ← {{π1}, {π2}, . . . , {πn}}Search direction I: � Area 1 → Area 2 (in Fig. 4)5: for i = 2 to n do6: S ′

π ← Sπ ∪ πi

7: if πi is incompatible with ∃πk ∈ Sπ then8: S ′

π ← S ′π \ πk

9: end if10: if S ′

π �∈ Sc then11: Sπ ← S ′

π, Sc ← Sc ∪ {Sπ}12: (xπ, Fπ) ← OptimalAlloc(Btr, Scur, Sπ, r̂Sπ

)13: if Fπ < F ∗ then14: (x∗, F ∗, S∗) ← (xπ, Fπ , Sπ)15: end if16: end if17: end forSearch direction II: � Area 3 → Area 2 (in Fig. 4)18: for j = 2 to n do19: S ′

ψ ← Sψ ∪ ψj

20: if ψj is incompatible with ∃ψk ∈ Sψ then21: S ′

ψ ← S ′ψ \ ψk

22: end if23: if S ′

ψ �∈ Sc then24: Sψ ← S ′

ψ, Sc ← Sc ∪ {Sψ}25: (xψ, Fψ) ← OptimalAlloc(Btr, Scur, Sψ, r̂Sψ

)26: if Fψ < F ∗ then27: (x∗, F ∗, S∗) ← (xψ , Fψ, Sψ)28: end if29: end if30: end forFinalize:31: (xcur, Fcur, Scur) ← (x∗, F ∗, S∗)32: Build the alternative RAT set Salt for Scur

When a file transfer service is requested, the transferred filesize is stored to Btr. Sπ and Sψ represent currently tested setsin each direction, which are initially set to RAT π1 and ψ1

(Line 2). The normalized cost of a single-RAT connection isdirectly calculated by allocating the whole file size to that RAT

connection, which is denoted by Fπi(Btr). Then, the optimal

(minimum) normalized cost F ∗ and the optimal RAT set S∗

are initialized by the minimum value among the normalizedcosts with single-RAT connections and the corresponding RAT(Line 3). Sc represents a set of already tested RAT sets andis initialized by {{π1}, {π2}, . . . , {πn}} (Line 4) because allthe single-RATs have been tested in Line 3. This set is used tocheck out previously tested RAT sets and to prevent redundantlytesting the same RAT set.

Each loop in the algorithm searches the optimal RAT setwith the optimal segment allocation in the different directionas ordered in (10). In Search direction I, the available RAT setsare tested from {π1} to {π1, . . . , πn}, and the available RATsets are tested from {ψ1} to {ψ1, . . . , ψn} in Search directionII. For each search direction, the union of the previous RAT setSπ/Sψ and RAT πi/ψj is saved to S ′

π/S′ψ (Lines 6 and 19).

If the newly added RAT cannot be concurrently activated withany RAT in the previously tested set, the incompatible RATin the previously tested set is removed from the newly testedRAT set (Lines 7–8 and 20–21). If the RAT set S ′

π/S′ψ is not

in the set of the already tested RAT sets Sc, Sπ/Sψ and Sc areupdated by the new RAT set (Lines 10–11 and 23–24). Withthe updated RAT set Sπ/Sψ, the optimal segment allocation foreach selected RAT is obtained by solving Subproblem 1 in thefunction OptimalAlloc() (Lines 12 and 25). Scur is the activatedRAT set when the algorithm starts, and the RAT activationoverhead would be added when Scur �= Sπ/Sψ. r̂[Sπ/Sψ] is theestimated throughput vector of the RAT set Sπ/Sψ.

The function OptimalAlloc() with these arguments solves (7)and returns the optimal segment allocation vector and its cost(x[π/ψ], F[π/ψ]). If the returned cost is less than the minimumcost F ∗, the optimal segment allocation vector, the optimal cost,and the optimal RAT set are updated by the returned values andthe new RAT set (Lines 14 and 27). After both the loops finish,the RAT set with the corresponding segment allocation vector isfinally updated with the optimal results (Line 31). In addition,the set of the alternative RAT sets Salt is built, where thealternative set replaces any RAT in Scur with the incompatibleRAT(s) (Line 32). This set is used for the proposed dynamicupdate algorithm, and the details are discussed later.

By the BLS algorithm, the number of visited RAT com-binations at most is reduced from 2n to 2n− 1, where n isthe number of available RATs. For performance evaluation,OptimalAlloc() has two versions: 1) to solve Subproblem 1 bythe existing convex tool and 2) to solve it by examining thesegment allocation vector in (9) first and then using the convextool only when (9) is infeasible. We denote the first methodby “BLS w/ cvx_tool” and the second method by “fBLS w/cvx_tool,” where “f” means “fast.”

B. Dynamic Update Algorithm

During the file transfer, the network conditions of the acti-vated RATs, e.g., channel and load, may change dynamically.Accordingly, the throughput of each RAT fluctuates, and theinitially estimated throughput is likely to be different fromthe actual average throughput. The increased throughput isout of our consideration because it reduces the original cost.


However, throughput degradation of the selected RAT causesthe increase in the original cost, and the optimal solution mightbe changed. Therefore, we propose a dynamic update algorithmin Algorithm 2 to adapt the selected RAT set to the dynamicnetwork condition.

Algorithm 2 Dynamic Update Algorithm

1: while file transfer do2: if update timer expires then3: Br ← Btr −Bdone

4: (xdone, Fdone) ← CalCost(Bdone, Sold, Scur, r̄Scur)

5: (xtodo, Ftodo) ← OptimalAlloc(Br, Scur, Scur, r̂Scur)

6: (xupd, Fupd) ← (Fdone + Ftodo,xdone + xtodo)7: if Fupd > Fcur +mcur then8: Btr ← Br

9: go to BLS algorithm10: else11: while ∀i, Salt

i ∈ Salt do12: (xalt

i , F alti ) ← OptimalAlloc(Br, Scur, S

alti , r̂Salt

i)

13: end while14: o ← arg

iminF alt

i

15: if F alto < Ftodo −malt then

16: Sold ← Scur, Btr ← Br, Bdone ← 017: (xcur, Fcur, Scur) ← (xalt

o , F alto , Salt

o )18: else19: (xcur, Fcur, Scur) ← (xupd, Fupd, Scur)20: end if21: end if22: end if23: end while

For the dynamic update algorithm, the optimality of the net-work selection and segment allocation is periodically checkedby monitoring the achieved throughput during the transfer [5].Br is the remaining file segment size, i.e., Btr subtracted bythe already transferred file segment size Bdone. The cost Fdone

for the segment allocation vector of the already transferred dataxdone is calculated with the cumulative average throughput vec-tor r̄Scur

of the activated RATs (Line 4). The optimal segmentallocation vector xtodo and the cost Ftodo for the remaining filesegment size are recalculated by the OptimalAlloc() with thecurrent RAT set and its estimated throughput vector r̂ (Line 5).We adopt the autoregressive and moving average estimator [27]for throughput estimation as follows:

r̂i = σr̂i−1 +1 − σ

K

K−1∑j=0

ri−j (11)

where we set K = 10 and σ = 0.95 in the simulation.If the updated cost Fupd with the segment allocation vector

xupd, which is the sum of the results in Lines 4 and 5, exceedsthe sum of the current cost and specific margin mcur, “BLSalgorithm” in Algorithm 1 can be revisited to search the betterRAT set (Lines 7–9).

On the other hand, if some RATs cannot be activated simulta-neously with each other, the optimal solution might be the case

that these RATs are sequentially used. For example, assumethat SA={{3G}, {4G}, {Wi-Fi}, {3G,Wi-Fi,}, {4G,Wi-Fi}},but {3G, 4G} and {3G, 4G, Wi-Fi} are not in SA. The optimalactivation strategy could be activating 4G and Wi-Fi for sometime, and then turn off 4G and activating 3G with Wi-Fi.Although this case is not directly included in our problem for-mulation, the proposed dynamic update algorithm in Algorithm2 deals with this issue. At first, RATs that cannot be activatedsimultaneously are defined as the alternative set and stored ina database. When the best RAT set is obtained, the alternativeRAT set Salt is built accordingly (Line 32 in Algorithm 1). Forexample, if SA is given as aforementioned and the optimal RATset is {3G, Wi-Fi}, the alternative RAT set is {4G, Wi-Fi} when3G and 4G are incompatible. Then, the optimal segment alloca-tion vector and the corresponding cost for each alternative RATset is calculated for the remaining file size (Line 12), and theRAT set of the minimum cost among the alternative RAT setsis found out (Line 14). If the minimum alternative cost is malt

less than the cost Ftodo, where malt is a predefined margin,the current RAT set is updated to the alternative RAT set andcorresponding segment allocation (Lines 15–17). Otherwise,the current cost Fcur and segment allocation vector xcur aresubstituted by Fupd and xupd, respectively (Line 19).

VII. PERFORMANCE EVALUATION

In the future smartphones, much more diversified types ofRATs and channels will be available; hence, it will achievelarger RAT diversity gain as the number of RATs increasesat the cost of complexity increase to find the best RAT set.To see the performance of the proposed algorithm, we assumethat a smartphone is equipped with various sets of RATs, i.e.,from three to six RATs, where 3G, 4G, and Wi-Fi interfacesare common, and the other interfaces have arbitrary parametersfor throughput, energy consumption, and data usage quota.We also assume that any combination among the availableRATs can be activated concurrently. However, for the case ofthree RATs with 3G, 4G, and Wi-Fi, we particularly deal withone more case where it is unable to simultaneously activate{3G, 4G} and {3G, 4G, Wi-Fi}, i.e., 3G and 4G can be onlyexclusively used. This case is denoted by “two RATs” instead of“three RATs” in the figure of simulation result. For simulationenvironments, we consider four different conditions, where theaverage achievable throughput of 4G and Wi-Fi, respectively,varies as follows: {25, 15}, {5, 15}, {25, 5}, {5, 5} Mb/s.We assume that the probability density function of the file sizefollows a lognormal distribution where the mean and standarddeviation are 7.17 and 2.41, respectively, referring to [28]. Theexpected optimal costs for the four different conditions areaveraged. With a given RAT set, we use the convex optimizationtools, Sedumi [26] and SDPT3 [29], to solve Subproblem 1 inthe objective function.2 We consider the full-search method,which searches all the possible RAT combinations among theavailable RAT set, to compare with the proposed algorithms.

2LP is typically known to have polynomial time complexity in the worst case.However, Subproblem 1 is always solved in the constant iteration level by thesetools for the scope in consideration (two to six RATs). Therefore, we assumethat the computational complexity is dominated by Subproblem 2.


Fig. 6. Comparison of RAT diversity gain.

Fig. 7. Comparison of complexity.

The full-search method achieves the optimal result, whereas thesearch space exponentially increases as the number of availableRATs increases.

Fig. 6 shows the simulation result that represents normalizedcosts, which are normalized by the cost of the full-searchmethod with Sedumi. The normalized cost of the optimal RATset with the optimal segment allocation vector decreases asthe number of available RATs increases, which shows theRAT diversity gain. Furthermore, we can see no significantperformance gap between the full-search algorithm and fBLSin our simulation environment where the maximum number ofRATs is six.

On the other hand, the performance enhancement in thecomplexity is presented in Fig. 7. Because the computationalcomplexity of the full-search algorithm is O(2n), the number ofiterations to search the optimal RAT set increases exponentially,as shown in the figure, whereas that of BLS just linearlyincreases as the number of available RATs increases. Moreover,fBLS even further reduces the iterations due to the initialexamination with (9). The elapsed time of one iteration with2.4-GHz CPU was about 0.01 s; hence, the total elapsed timeto find the optimal network operation mode with fBLS is under0.5 s even for six RATs, where it is 6–10 s for the full-searchalgorithm.

Fig. 8 presents a performance comparison between the verti-cal handoff and proposed scheme. The vertical handoff impliesthat only a single best RAT is selected for a file transfer, which

Fig. 8. Comparison of network selection strategies.

has the minimum normalized cost among the available RATs. Inaddition, we compare two conditions for the proposed scheme.One is the 2-RAT parallel scheme that can simultaneouslyactivate at most two RATs among the available RATs, and theother is the All-RAT parallel scheme that can exploit all thepossible RAT combinations, including activating all the avail-able RATs. The simulation environment is the same as theprevious evaluation, and the average cost is normalized by thecost of the vertical handoff scheme with two available RATs,which is the maximum among all the results and is set to one.As the number of available RATs increases, the vertical handoffscheme also benefits from the RAT diversity. For example, withsix RATs, the normalized cost is 32% less than that with twoRATs. Utilizing parallel connections enhances the performanceeven further. For example, the normalized costs of the two-RAT parallel and All-RAT parallel schemes with six RATs aredecreased by 26% and 36%, compared with that of the verticalhandoff, respectively.

To evaluate the performance of the dynamic update al-gorithm, we artificially generate time-varying throughput, asshown in Fig. 9(a) for each available RAT. Three RATs, namely,3G, 4G, and Wi-Fi, with average throughput of 5, 25, and 15Mb/s, respectively, are considered. The update period of thedynamic update algorithm is set to 10 s. In the simulation, asmartphone downloads several files with the size of {500, 1000,2000, 4000, 8000, 16 000} Mbits and measures the normalizedcost. The simulation is iterated 100 times for each scheme, andthe results are averaged.

The comparison of the proposed algorithms is presentedin Fig. 9(b) and (c) in terms of the remaining battery energyand data quota. “Static” represents the static algorithm,which searches the best RAT set and optimal segmentallocation vector at the start of the file transfer and maintainsthe strategy until the download is completed. “Dynamic”represents the dynamic update algorithm, which monitorsthe achieved throughput and updates the optimal RAT setand segment allocation vector according to Algorithm 2. Weset both the update margins mcur and malt to 0.1 times thecalculated cost. “Genie” searches the best RAT set and segmentallocation vector with the exact average throughput for theentire transfer time, assuming it knows future throughput.The suffix “_I” represents that 3G and 4G are incompatible;hence,SA={{3G}, {4G}, {Wi-Fi}, {3G,Wi-Fi}, {4G,Wi-Fi}},


Fig. 9. Performance comparison of the static algorithm and dynamic update algorithm.

whereas the suffix “_C” represents that all the RATs arecompatible with one another and that all the RAT combinationsare available. The cost of Genie_C, which performs best amongall the compared schemes, is set to the reference value. Then,the cost gaps, i.e., the ratios of the average increased cost ofthe proposed schemes to the cost of Genie_C, are presented inFig. 9(b) and (c).

From the simulation results, we confirm that the perfor-mances of the compatible cases are generally better than thoseof the incompatible cases. In some cases, Dynamic_C performseven better than Genie_I because of more options of the RATselection including {3G, 4G} and {3G, 4G, Wi-Fi}. Second, thecost gap of the static algorithm can be significantly increaseddue to the dynamic network condition. However, the dynamicupdate algorithm can effectively reduce the cost gap. In partic-ular, we observe that the cost gap can be suppressed under 5%by Dynamic_C for all the status of remaining energy and data.

VIII. CONCLUSION AND FUTURE WORK

In this paper, we have proposed and analyzed the multi-RAT selection and segment allocation scheme for smartphonesbased on a multiattribute cost function. The cost function prop-erly normalizes and integrates the transfer completion time,energy consumption, and service charge in n RAT-based smart-phones for file transfer services. Based on the device’s currentstatus-dependent cost model, the proposed scheme adaptivelyactivates n RAT interfaces to minimize the cost function.Considering the increasing number of RATs integrated in thefuture smartphones, the proposed algorithm, namely fast BLS,achieves a nearly optimal RAT selection and segment allocationwith reduced computational complexity. It shows no significantperformance gap from the full-search algorithm, whereas theexponential complexity is reduced to linear. In addition, wehave shown that the performance gain increases as the numberof available RATs increases due to the RAT diversity, and itcan be more improved by exploiting the parallel activation ofmultiple RATs.

On the other hand, the multiattribute cost model might needto be modified according to the performance metric and trafficpattern. Therefore, as future work, we will develop multi-RAT selection algorithms based on multiattribute cost for otherservices such as video streaming and web surfing, in additionto the file transfer service considered in this paper.

REFERENCES

[1] A. Ford, C. Raiciu, M. Handley, S. Barre, and J. Iyengar, Architecturalguidelines for multipath tcp development (RFC 6182), Mar. 2011.

[2] Galaxy S5 Download Booster. (2014). [Online]. Available:http://galaxys5guide.com/samsung-galaxy-s5-features-explained/galaxy-s5-download-booster, (Accessed: May 20, 2014).

[3] Q. Song and A. Jamalipour, “A network selection mechanism for nextgeneration networks,” in Proc. IEEE ICC, May 2005, pp. 1418–1422.

[4] Q.-T. Nguyen-Vuong, N. Agoulmine, and Y. Ghamri-Doudane, “A user-centric and context-aware solution to interface management and accessnetwork selection in heterogeneous wireless environments,” Comput.Netw., vol. 52, no. 18, pp. 3358–3372, Dec. 2008.

[5] W. Lee, J. Koo, S. Choi, and Y. Park, “ESPA: energy, usage ($), andperformance-aware LTE-Wi-Fi adaptive activation scheme for smart-phones,” in Proc. IEEE WoWMoM, Jun. 2014, pp. 1–9.

[6] W. Shen and Q.-A. Zeng, “Cost-function-based network selection strategyin integrated wireless and mobile networks,” IEEE Trans. Veh. Technol.,vol. 57, no. 6, pp. 3778–3788, Nov. 2008.

[7] R. Amin et al., “Balancing spectral efficiency, energy consumption, andfairness in future heterogeneous wireless systems with reconfigurabledevices,” IEEE J. Sel. Areas Commun., vol. 31, no. 5, pp. 969–980,May 2013.

[8] Y. Choi, H. Kim, S. Han, and Y. Han, “Joint resource allocation forparallel multi-radio access in heterogeneous wireless networks,” IEEETrans. Wireless Commun., vol. 9, no. 11, pp. 3324–3329, Nov. 2010.

[9] C.-J. Chang, T.-L. Tsai, and Y.-H. Chen, “Utility and game theory basednetwork selection scheme in heterogeneous wireless networks,” in Proc.IEEE WCNC, Apr. 2009, pp. 1–9.

[10] E. Aryafar, A. Keshavarz-Haddad, M. Wang, and M. Chiang, “RATselection games in HetNets,” in Proc. IEEE INFOCOM, Apr. 2013,pp. 998–1006.

[11] E. Gustafsson and A. Jonsson, “Always best connected,” IEEE WirelessCommun., vol. 10, no. 1, pp. 49–55, Feb. 2003.

[12] M. Kassar, B. Kervella, and G. Pujolle, “An overview of verti-cal handover decision strategies in heterogeneous wireless networks,”Comput. Commun., vol. 31, no. 10, pp. 2607–2620, Jun. 2008.

[13] A. Roy, J. Shin, and N. Saxena, “Multi-objective handover inLTE macro/femto-cell networks,” J. Commun. Netw., vol. 14, no. 5,pp. 578–587, Oct. 2012.

[14] H. Petander, “Energy-aware network selection using traffic estimation,”in Proc. ACM MICNET , Sep. 2009, pp. 55–60.

[15] Y. Choi and S. Choi, “Service charge and energy-aware vertical handoffin integrated ieee 802.16e/802.11 networks,” in Proc. IEEE INFOCOM,May 2007, pp. 589–597.

[16] Y.-L. Chung, Z. Tsai, and C.-H. Yang, “A Study of quota-based dynamicnetwork selection for multimode terminal users,” IEEE Syst. J., vol. 8,no. 3, pp. 759–768, Sep. 2014.

[17] Y. Yu, B. Yong, and C. Lan, “Utility-dependent network selectionusing madm in heterogeneous wireless networks,” in Proc. IEEE PIMRC,Sep. 2007, pp. 1–5.

[18] Q. Peng, M. Chen, A. Walid, and S. H. Low, “Energy efficient mul-tipath TCP for mobile devices,” in Proc. ACM MobiHoc, Aug. 2014,pp. 257–266.

[19] Y.-S. Lim, Y.-C. Chen, E. M. Nahum, D. Towsley, and R. J. Gibbens,“How green is multipath TCP for mobile devices?” in Proc. ACMAllThingsCellular, Aug. 2014, pp. 3–8.


[20] S. Chen, Z. Yuan, and G.-M. Muntean, “An energy-aware multipath-TCP-based content delivery scheme in heterogeneous wireless networks,” inProc. IEEE WCNC, Apr. 2013, pp. 1291–1296.

[21] J. Pouwelse, P. Garbacki, D. Epema, and H. Sips, “The BitTorrent P2Pfile-sharing system: Measurements and analysis,” in Proc. Int. WorkshopPeer-to-Peer Syst., Feb. 2005, pp. 205–216.

[22] D. Kaspar et al., “Enhancing video-on-demand playout over multi-ple heterogeneous access networks,” in Proc. IEEE CCNC, Jan. 2010,pp. 1–5.

[23] J. Huang et al., “A close examination of performance and power char-acteristics of 4G LTE networks,” in Proc. ACM MobiSys, Jun. 2012,pp. 225–238.

[24] C. Yoon, D. Kim, W. Jung, C. Kang, and H. Cha, “AppScope: Applicationenergy metering framework for android smartphones using kernel activitymonitoring,” in Proc. USENIX ATC, Jun. 2012, pp. 36.

[25] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, U.K.:Cambridge Univ. Press, 2003.

[26] J. F. Sturm, “Using SeDuMi 1.02, a MATLAB toolbox for optimiza-tion over symmetric cones,” Optim. Methods Softw., vol. 11, no. 1–4,pp. 625–633, 1999.

[27] W. A. Gardner, Introduction to Random Processes: with Applications toSignals and Systems, 2nd ed. New York, NY, USA: McGraw-Hill, Inc.,1990.

[28] R. Pries, F. Wamser, D. Staehle, K. Heck, and P. Tran-Gia, “On trafficcharacteristics of a broadband wireless internet access,” in Proc. IEEENGI, Jul. 2009, pp. 1–7.

[29] R. H. Tütüncü, K. Toh, and M. Todd, “Solving semidefinite-quadratic-linear programs using SDPT3,” Math. Programm., vol. 95, no. 2,pp. 189–217, 2003.

Wonbo Lee (S’12) received the B.S., M.S., andPh.D. degrees in electrical and computer engineeringfrom Seoul National University, Seoul, Korea, in2003, 2006, and 2015, respectively.

From 2006 to 2009, he was a Researcherwith the Institute of Advanced Technology, Dae-woo Electronics, Seoul. Beginning in March 2015,he started working as a Senior Engineer withSamsung Electronics, Suwon, Korea. His researchinterests include multiattribute operation policiesfor multiple-radio-access-technology smartphones,

context-aware cross-layer protocols, and self-organizing networks for fifth-generation systems.

Jonghoe Koo (S’12) received the B.S. degree fromSeoul National University, Seoul, Korea, in 2011.He is currently working toward the Ph.D. degreewith the Department of Electrical and ComputerEngineering, Seoul National University.

His research interests include reliable video mul-ticast streaming over wireless local area networksand context-aware cross-layer protocol for fifth-generation networks.

Yongseok Park received the B.S. and M.S.degrees in electronics from Seoul National Univer-sity, Seoul, Korea, and the Ph.D. degree in electricaland computer engineering from Purdue University,West Lafayette, IN, USA.

He is currently a Principal Engineer with theDigital Media and Communications R&D Center,Samsung Electronics, Suwon, Korea, working on thedevelopment of network solutions for smartphonesand Internet of Things. Before joining SamsungElectronics, he was with AT&T, working in the areas

of network management and data communication service development.

Sunghyun Choi (S’96–M’00–SM’05–F’14) re-ceived the B.S. (summa cum laude) and M.S.degrees from Korea Advanced Institute of Scienceand Technology, Daejeon, Korea, in 1992 and 1994,respectively, and the Ph.D. degree from The Univer-sity of Michigan, Ann Arbor, MI, USA, in 1999.

Since 2002, he has been with the Departmentof Electrical and Computer Engineering, Seoul Na-tional University (SNU), Seoul, Korea, where he iscurrently a Professor. Before working with SNU,he was with Philips Research USA, and from June

2009 to June 2010, he was a Visiting Associate Professor with StanfordUniversity, Stanford, CA, USA. He is the author or coauthor of over 190technical papers and book chapters in the areas of wireless/mobile networksand communications. He is a coauthor (with B. G. Lee) of the book BroadbandWireless Access and Local Networks: Mobile WiMAX and Wi-Fi (Artech House,2008). He is a holder of about 120 patents and has many patents pending.

Dr. Choi has served as a General Cochair for the Third International Con-ference on Communication System Software and Middleware (COMSWARE)in 2008 and as a Technical Program Committee Cochair for the Association forComputing Machinery (ACM) Multimedia, the IEEE International Symposiumon World of Wireless Mobile and Multimedia Networks (IEEE WoWMoM),and COMSWARE in 2007. He has also served on program and organiza-tion committees of numerous leading wireless and networking conferences,including the ACM Annual International Conference on Mobile Computingand Networking, the IEEE Conference on Computer Communications, theIEEE Communications Society Conference on Sensor and Ad Hoc Commu-nications, and IEEE WoWMoM. He is also currently serving as an Editorfor the IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS and IEEEWIRELESS COMMUNICATIONS.He served as an Associate Editor for the IEEETRANSACTIONS ON MOBILE COMPUTING and as a Guest Editor for theIEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS and ACMWireless Networks (WINET). He was an active voting member of the IEEE802.11 WLAN Working Group from 2000 to 2007. He has received numerousawards, including the Best Teaching Award from the College of Engineering,SNU, in 2006; the IEEK/IEEE Joint Award for Young IT Engineer in 2007;the Best Paper Award from IEEE WoWMoM, the Presidential Young ScientistAward, and the Outstanding Research Award in 2008; the Shinyang ScholarshipAward in 2011; and the KICS Dr. Irwin Jacobs Award in 2013.

Transfer Time, Energy, and Quota-Aware Multi-RAT Operation...

Documents

Transcript of Transfer Time, Energy, and Quota-Aware Multi-RAT Operation...