1

An Application Layer Protocol forEnergy-Efficient Bandwidth Aggregation with

Guaranteed Quality-of-ExperienceZaiyang Tang, Zirui Wang, Peng Li, Member IEEE, Song Guo, Senior Member, IEEE,

Xiaofei Liao, Member, IEEE, and Hai Jin, Senior Member, IEEE

Abstract—WiFi and cellular networks are pervasively provided for mobile Internet access. Although most existing mobile devices areequipped with both WiFi and cellular network interfaces, concurrent data transmissions over these interfaces for improved throughputare not provided. In this paper, a bandwidth aggregation prototype, named ALP-A (Application Layer Protocol based Aggregation),is developed for easy use by simply installing an application in mobile devices without modifing their operating systems or drivers. Itprovides desired quality-of-experience (QoE), i.e., acceptable response delay to users, learned from application characteristics andusers behaviors. Furthermore, we propose an online algorithm of traffic scheduling over WiFi and cellular interfaces with the objectiveof minimizing energy consumption while guaranteeing the QoE. Over the prototype implemented on Andriod-based smartphones, weconduct extensive experiments to show that ALP-A outperforms existing schemes significantly.

Index Terms—Application Layer Protocol, Bandwidth Aggregation, Quality-of-Experience, Energy Efficiency

F

1 INTRODUCTION

The increasing popularity of wireless devices and avail-ability of radio access technologies (e.g., WiFi, Bluetooth,3G and LTE) are stimulating various new mobile services[1]. Modern mobile devices, e.g., smartphones and tablet-s, are usually equipped with multiple network interfaces,such as WiFi and 3G, which imply the opportunities ofmultipath communication with improved throughput ona single device. However, existing mobile devices do notexploit such a great potential because they just simplyactivate a single network interface for data transmission.For example, the 3G-interface of smartphones with An-droid operating system will be automatically turned offwhen WiFi connection is available.

Bandwidth aggregation [2]–[4] is a promising tech-nique for throughput enhancement by allowing concur-rent data transmission over multiple network interfacesas shown in Fig. 1, where a mobile device accesses theInternet via a WiFi access point and a 3G base stationsimultaneously. However, existing bandwidth aggrega-tion solutions are far from providing efficient and prac-tical solutions for mobile users because of the followingweaknesses. First, they are implemented in lower layersof the network protocol stack, such as network or linklayer, relying on the modification on operating systemsor device drivers. To enjoy the benefits of bandwidthaggregation, users need to reinstall the operating system,

Z. Tang, Z. Wang, X. Liao, and H. Jin are with the Services ComputingTechnology and System Lab, Cluster and Grid Computing Lab, School ofComputer Science and Technology, Huazhong University of Science and Tech-nology, China. E-mail: {tangzaiyang, m201172464, xfliao, hjin}@hust.edu.cnP. Li and S. Guo are with the School of Computer Science and Engineering,The University of Aizu, Japan. E-mail: {pengli, sguo}@u-aizu.ac.jp

Fig. 1. Bandwidth aggregation.

which would be time-consuming or even impossible forcommon users without expertise. Second, the variabilityof link quality in both 3G and WiFi networks is ignored.Mobile devices need to contend for network access op-portunities in such networks without any performanceguarantee. The more devices exist in the network, thelower throughput can be achieved by each device. Asmobile users join or leave the network, the achievabletransmission rate of WiFi and 3G interfaces in eachdevice may change, and such influence has not beenstudied.

In this paper, we develop a bandwidth aggregationprototype at the application layer, named ALP-A (Ap-plication Layer Protocol based Aggregation), by takingdynamics of wireless links into consideration. Specifical-ly, we consider mobile devices equipped with both WiFiand 3G network interfaces, and each device has a batch

2

of downloading requests that arrive in an online manner.To guarantee a certain level of quality-of-experience(QoE), ALP-A specifies a deadline for each request ac-cording to the characteristics of applications and userbehaviors. To satisfy these downloading requests, wepropose an online algorithm to dynamically scheduledata transmission on both network interfaces.

In particular, energy efficiency is pursued in our de-sign because mobile devices are usually powered bybatteries with limited capacity. While energy saving formobile device has been intensively studied, the energyconsumption in bandwidth aggregation has not beenaddressed. In our proposed design, data transmissionsare made on both WiFi and 3G interfaces in an energy-efficient manner while the QoE constraints are guaran-teed.

The main contributions of this paper are summarizedas follows.

• We develop an application layer protocol calledALP-A for bandwidth aggregation of WiFi and 3Gwireless links.

• We propose an online algorithm for data transmis-sion over both network interfaces with the objectiveof minimizing energy consumption while guaran-teeing user experiences.

• We implement ALP-A on real devices, and evaluateits performance by conducting extensive experi-ments.

The rest of paper is organized as follows. Section 2 re-views some important related work. Section 3 discussesthe background and motivation. The detailed design ofALP-A is presented in Section 4. Section 5 proposes anonline scheduling algorithm called E-schedule. Section6 shows the effectiveness of our proposal with experi-ments on real devices. Finally, Section 7 concludes thepaper.

2 RELATED WORK

2.1 Bandwidth aggregation solutions over differentlayersBandwidth aggregation can be addressed at differentlayers of the network protocol stack [5]. Early workon application layer bandwidth aggregation is based onmultiple logical channels on the same network interface,such as XFTP [6] and Grid FTP [7]. PSockets [8] createsmultiple parallel sockets to transmit data over multiplelogical TCP connections. However, these approaches useonly a single physical interface, and their performanceis limited by the available bandwidth offered by theinterface. Kaspar et al. [9] [10] [11] have exploitedthe benefits of using multiple Internet connections formultimedia streaming applications on mobile devices.However, a general bandwidth aggregation approach forall applications has not been considered.

Some work exploits bandwidth aggregation at thetransport layer by reducing unnecessary retransmissions[12] [13]. As a widely adopted transport layer protocol,

Stream Control Transmission Protocol (SCTP) provideslittle bandwidth aggregation support for TCP applica-tions. Hsieh et al. [14] [15] have proposed a virtualpipe concept to provide bandwidth aggregation for TCPapplications. The multi-path TCP (MPTCP) has beenproposed by a working group formed from InternetEngineering Task Force (IETF) recently [16], which aimsto enhancing the application performance through PathDiversity mechanism. Chen et al. [17] [18] have shownthe feasibility of using MPTCP for mobile devices inWiFi/Cellular networks. Although they attempt to avoidthe modification of existing Internet infrastructure, theirimplementation still needs to revise operating systemkernel.

Bandwidth aggregation at the network layer has beenstudied in [5], [19]–[21]. Their basic idea is to offerpacket level traffic distribution over multiple networkinterfaces. At the data link layer, multiple links canbe bundled into a single logical communication link,which shows high utilization of the aggregated capacity[22]–[24]. However, above solutions are difficult to beimplemented for mobile devices in wireless networks.

2.2 Energy efficiency of mobile devices

The energy efficiency of mobile devices has been exten-sively investigated in recent years. In [25], the energyconsumption of mobile devices in different environ-ments of 3G and WLAN is considered. The authors haveproposed a mathematical model, e-Aware, consideringtwo energy consumption elements: signaling and mediatransfers, to estimate how application layer protocolproperties affect the energy consumption of mobile de-vices.

Activating multiple network interfaces of a mobiledevice simultaneously incurs more power consumption.Chowdhury et al. [26] present an efficient interface selec-tion scheme. Mahkoum et al. [27] present a frameworkwith power-efficient management from a global pointof view. The basic idea behind is to power off the idleinterface but at the same time to keep it in virtualidle mode in the network by extending IEEE 802.21 onboth mobile node and network side. Chen et al. [18]consider the balance between energy consumption andthroughput of devices using MPTCP.

3 BACKGROUND AND MOTIVATION

3.1 Bandwidth aggregation based on applicationlayer protocol

In this paper, bandwidth aggregation is exploited atthe application layer to support different applicationprotocols such as HTTP and RTP. The major benefit ofsuch approach is no need to modify the protocol stack.Furthermore, the information of upper applications canbe easily obtained and used to adjust the utilization ofmultiple interfaces according to user experience.

3

Fig. 2. Layout of nodes and APs

3.2 Contention and interference in WiFiBecause of the CSMA/CA mechanism adopted by 802.11standard [28], increasing the number of wireless de-vices inevitably aggravates contention and interference,leading to decreased network performance. We showthis phenomenon by measuring the throughput in WiFinetworks with different number of mobile devices.

We consider a wireless network consisting of twowireless routers as access points (APs) and ten mobiledevices, whose locations are shown in Fig. 2. The wire-less routers connect to the Internet via ADSL (Asym-metric Digital Subscriber Line) with 8Mbps downlinkbandwidth. Devices n2, n4, n5 and n7 are laptops, andthe rest are smartphones. All mobile devices connect towireless routers based on the 802.11b/g protocol.

In our experiments, each device runs videowatching through an online video website Youku(http://www.youku.com/) as traffic workload. Theaverage throughput under different number of devicesis shown in Fig. 3. We observe that the averagethroughput decreases sharply as the number of devicesin the network grows. For example, when a singledevice is active in the network, the average throughputachieves nearly 1000KB/s, while it drops to 400KB/swhen two devices are active. Moreover, the interferencebetween two APs is negligible when the channel modeof both AP1 and AP2 is set to auto. Above resultsmotivate us to take the variation of network link qualityinto consideration in our protocol design.

3.3 Energy consumption of Cellular and WiFiWiFi and 3G interfaces have different energy consump-tion characteristics. According to 3GPP regulations [29],a Radio Resource Control (RRC) protocol specifies threestates: IDLE, DCH, and FACH, for 3G network inter-face, which indicate idle channel, dedicated channel and

1c 2c 3c 4c 5c 6c 7c 8c 9c 10c0

200

400

600

800

1000

Ave

rage

thro

ughp

ut p

er d

evic

e (K

B/s

)

Number of clients in AP1

1AP 2APs

Fig. 3. Average per-device throughput v.s. the number ofdevices

forward access channel, respectively. During data trans-mission, the 3G-interface is in the DCH state with highenergy consumption. After data transmission, it remainsin DCH for about 5 seconds, then changes to FACHfor about another 5 second, and eventually switches toIDLE. In contrast, the WiFi-interface switches to idle stateimmediately after the completion of data transmission.

The power consumption modeling and measurement[30], [31] reveal that LTE and 3G are less power efficientthan WiFi. Generally, 3G and LTE are more energy-intensive compared with WiFi.

4 SYSTEM DESIGN

4.1 Design objectivesTowards an efficient design, we target several importantobjectives as follows.

(1) Simplicity. ALP-A should be easy to use such thatusers can enjoy the benefits of bandwidth aggregation bysimply installing an application in their mobile devices.For this purpose, the implementation of ALP-A shouldbe independent of operating system and network proto-col stack.

(2) Quality-of-Experience (QoE). ALP-A receives com-munication requests from applications with various QoErequirements. For example, even a bit longer delay inloading CSS and JavaScript for web browsing wouldseriously degrade user-experience, while buffering invideo playing can tolerate a certain level of latency. Toguarantee QoE, ALP-A should schedule the requests ontwo interfaces to guarantee their deadlines specified byapplications.

(3) Energy efficiency. Activating multiple network in-terfaces imposes a great challenge for energy efficiency ofmobile devices powered by batteries with limited capac-ity. When there is no communication request, networkinterfaces should be in low-power states. When com-munication requests arrive, ALP-A needs to determine

4

Application 1

Data recorder

Interface

switch

OS

Interface 1

Application 2 Application 3

Interface 2

Scheduling

Rq1 Rq2 Rq3 ... Rqn Rp1 Rp2 Rp3 ... Rpn

Sub-requests Sub-responses

Buffer pool

Original request Original response

Partitioning Combination

1

2

3

4

5

Scheduler

Fig. 4. Application-layer aggregation framework

which interface should be activated for data commu-nication. The different energy consumption features ofWiFi and 3G interfaces should be exploited in ALP-Ato reduce energy consumption while guaranteeing user-experience.

All above objectives are closely related. In the nextsection, we will address all challenges by designing anapplicant layer protocol for bandwidth aggregation.

4.2 ALP-A: protocol designALP-A receives communication requests from applica-tions and dispatch them to different network interfaces.Meanwhile, ALP-A takes charge of caching responsesfrom lower layers, which will be then delivered toapplications. As illustrated in Fig. 4, ALP-A is composedof three modules: data recorder, scheduler and interfaceswitch.

The interface switch is used to control the activation ofmultiple interfaces. Recall that the Android OS will turnoff 3G-interface automatically when WiFi connection is

Fig. 5. An example of request partition.

available. To implement bandwidth aggregation withoutmodifying OS, ALP-A specifies multiple interfaces withdifferent IP addresses.

To realize the requests partitioning and scheduling,the data recorder module is responsible to acquire thecurrent transmission rate and to predict the rate in thenext time slot.

The scheduler is the core component of ALP-A. Thedetails of its workflow and implementation are present-ed as follows.

1. Original request caching. ALP-A caches data trans-mission requests, such as GET of HTTP requests, fromapplications. For this purpose, ALP-A first enables alistener port, to which all packets are redirected. Then,it listens to this port for any request that will be furtherprocessed by the request partitioning component.

2. Request partition. ALP-A partitions each originalrequest into several sub-requests. An example of requestpartition is shown in Fig. 5, where requests are divid-ed into sub-requests to be scheduled over WiFi or 3Ginterface. To implement request partition, we modifythe “Range” field in the header of HTTP request, whichindicates a range of data to be downloaded. For example,“Range:bytes=0-100” means that the first 100 bytes of thefile are requested. If the request is correctly processed byremote servers providing network services, a responsewith a state code 206 instead of 200 is returned. Forexample, a HTTP response with a header ”Content-Range:byte 0-100/2350” means that this response contains thefirst 100 bytes of the requested file with total 2350 bytes.

3. Sub-requests scheduling. ALP-A uses the APIs(Application Programming Interfaces) provided by theoperating system to schedule sub-requests to availablenetwork interfaces with the objective of reducing energyconsumption while guaranteeing user-experience. Thedetailed design of our scheduling algorithm will bepresented in the next section.

4. Sub-responses buffering. ALP-A buffers all sub-responses returned from remote servers in a pool thatis realized using binary heap. Since the sub-responsesshould be assembled into responses of the correspondingoriginal requests, we use a hash table to facilitate suchprocess. Specifically, the key of the hash table is the orig-inal response, and the associated value is a list of sub-responses sorted in a non-descending order according totheir deadlines.

5. Response assembly. ALP-A combines several sub-

5

responses into a response, and delivers it to applications.To deal with the instability of wireless networks, ALP-A provides a fault-tolerant mechanism for response de-livery. Specifically, each sub-request has a retransmis-sion index that is initialized to be zero. If there is atransmission error, we increase the index, and reset thecorresponding request in the transmission queue. Whenthe retransmission index reaches the threshold (e.g., 10 inour experiments), we set the interface to be unavailable.Accordingly, all requests in the associated queue will bedelivered over another interface.

Note that the request-scheduling algorithm plays animportant role in ALP-A because it determines whichinterface should be active, and how long it be active fordata transmission.

5 E-SCHEDULE:ENERGY-AWARE SCHEDUL-ING ALGORITHM

In this section, we present an online algorithm, which isreferred to as E-Schedule, to schedule requests on WiFiand 3G interfaces with the objective of minimizing ener-gy consumption while guaranteeing QoE. The problemstatement is first given, followed by the online algorithmdesign.

5.1 Problem statementWe consider a mobile device equipped with both WiFiand 3G interfaces. The throughput of WiFi and 3G inter-faces is estimated on each time slot t, denoted as V w(t)and V c(t), respectively. The energy consumption of aunit data transmission through WiFi and 3G interfacesis denoted by α and β, respectively. Typically, we haveα < β [25]. The applications on the mobile device issue aset of downloading requests, denoted by {r1, r2, r3, ...},in an online manner. For each request rk that arrivesat time ak, the associated data with size mk shouldbe downloaded before a deadline dk to guarantee acertain level of QoE. Note that ALP-A would divide eachrequest into several sub-requests that will be satisfiedat different time slots. Let mw

k (t) and mck(t) denote

the data downloaded for request rk at time slot t viaWiFi and 3G interfaces, respectively. Given a set of nrequests and the achievable transmission rate of eachnetwork interface within time period [1, T ], the energyminimization problem can be formulated as:

minE, subject to:

E =T∑

t=1

n∑k=1

(α ·mwk (t) + β ·mc

k(t)); (1)

n∑k=1

mck(t) ≤ V c(t), ∀1 ≤ t ≤ T ; (2)

n∑k=1

mwk (t) ≤ V w(t), ∀1 ≤ t ≤ T ; (3)∑

t∈[ak,dk]

mck(t) +mw

k (t) = mk, ∀1 ≤ k ≤ n. (4)

The total energy consumption can be calculated by(1). The capacity constraints of WiFi and 3G interfacesare represented by (2) and (3), respectively. Finally, eachrequest should be satisfied within [ak, dk], which is rep-resented by (4).

Although above linear programming can be easilysolved within polynomial time, it fails to be applied inpractical scenarios where request arrivals and interfacecapacities are unknown in the beginning of each timeslot. However, it provides some insight about the energyminimization problem, i.e., fully exploiting the capacityof low-power WiFi-interface while minimizing usageof high-power 3G-interface to achieve low energy con-sumption. It motivates us to design an online algorithmin the next section.

5.2 Algorithm design

To deal with the challenges of online request arrivaland network dynamic, we propose a heuristic algorith-m called E-Schedule to minimize energy consumptionwhile guaranteeing the QoE of all requests.

The E-Schedule algorithm is executed at the beginningof each time slot. Its basic idea is to give higher priorityto requests with earlier deadlines. The bandwidth ofWiFi-nterface is always fully exploited, while the 3G-interface is activated only if there are requests thatcannot be satisfied expectantly with only WiFi-interface.The pseudo code of E-Schedule algorithm is shown inAlgorithm 1 below.

Algorithm 1 The E-Schedule Algorithm1: sort requests in current time slot according to their

deadlines in a non-descending order, i.e., dσ1 ≤ dσ2 ≤...;

2: measure the transmission rate V w(t) and V c(t);3: Cw(t) = V w(t), Cc(t) = V c(t);4: for each request in the sorted order do5: if mσi(t) ≤ Cw(t) then6: download mσi(t)-unit data of request rσi via

WiFi-interface;7: Cw(t) = Cw(t)−mσi(t);8: mσi(t) = 0;9: else

10: download Cw(t)-unit data of request rσi viaWiFi-interface;

11: Cw(t) = 0;12: mσi(t) = mσi(t)− Cw(t);13: V w(t+ 1) = δV w(t) + (1− δ)V w(t− 1);14: if mσi(t) > (dσi − t)V w(t) then15: m = min{Cc(t),mσi(t)};16: download m-unit data via 3G-interface;17: Cc(t) = Cc(t)−m;18: mσi(t) = mσi(t)−m;19: end if20: end if21: end for

6

Fig. 6. Equipments used in experiments.

In Algorithm 1, all requests at the current time slott are sorted according to their deadlines in a non-descending order, and the results are maintained in σ asshown in line 1. We measure the achievable transmissionrates of WiFi and 3G interfaces, i.e., V w(t) and V c(t). Theresidual capacity of WiFi and 3G interfaces is maintainedin variables Cw(t) and Cc(t), which are initialized toV w(t) and V c(t), respectively. After that, we considerto assign the requests to network interfaces accordingto order σ in the for loop from line 4 to 20. For anyrequest rσi , if its residual data mσi(t) is no greater thanthe capacity of WiFi-interface, we download it usingonly WiFi network, and update Cw(t) and mσi(t) inlines 6 to 8. Otherwise, we fully exploit the residualcapacity of WiFi-interface as shown in lines 10 to 12.Furthermore, we estimate the future transmission rateof WiFi by taking the rates of current and previous timeslots into consideration. Specifically, we use the methodof moving average with parameter δ for estimation asshown in line 13. If the amount of mσi(t)-unit data ofrequest rσi cannot be satisfied by the expected WiFi-rate V w(t + 1) before its deadline, we activate the 3G-interface for data transmission. Finally, we update Cc(t)and mσi(t) in lines 17 and 18.

6 EXPERIMENTS

6.1 Experiments designOur experiments are performed on SAMSUNG GalaxyWonder smartphones with operating system of Android2.3.3. Android supports several APIs to show the operat-ing voltage and remaining battery capacity. To measurethe energy consumption, we use an external DC powersupply on 3.8V, and calculate the power according tothe current. The equipments used in our experimentsare shown in Fig. 6.

Two typical wireless networks are used for perfor-mance evaluation. One is a WiFi network with 10Mbp-

TABLE 1Details of test sites

Label Web site (url) Requests CSS/JavaScript

1 Sina (www.sina.com.cn) 328 932 163 (www.163.com) 277 653 Yahoo (www.yahoo.com) 215 424 BBC (www.bbc.co.uk) 95 305 CNN (edition.cnn.com) 128 28

s maximum bandwidth, and the other is a 3G net-work using WCDMA technology. The average downlinkthroughput of WiFi and 3G is 500KB/s and 300KB/s,respectively. In addition to our proposed E-schedule (E-S) algorithm, we consider comparison algorithms in ourexperiments, denoted as follows.

• 3G: only 3G network interface is allowed for datatransmission, i.e., V w(t) = 0 under any time slot tin Algorithm 1.

• WiFi: only WiFi network interface is allowed fordata transmission, i.e., V w(t) = 0 under any timeslot t in Algorithm 1.

• GRD: a greedy scheduling algorithm, in which eachrequest is satisfied upon its arrival. Note that the 3G-interface will be activated if the transmission rate ofWiFi network interface is not enough to satisfy therequests.

• OPP: an opportunistic scheduling algorithm pro-posed in [32], which improves the greedy schedul-ing by taking the dynamic channel condition intoconsideration.

In our experiments, we consider two typical appli-cations: web browsing and streaming media playing.In web browsing, we use the default browsing systemto load 5 websites as shown in Table 1. For example,328 requests are issued to load the homepage of Sina(www.sina.com.cn), among which there are 93 requestsbelonging to Cascading Style Sheets (CSS) or JavaScriptelements. In streaming media playing, we watch 5 videosat youku (http://www.youku.com), which is an onlinevideo website similar to YouTube. All results are aver-aged over 20 sets of experiments.

6.2 Results and analysis

We first investigate the energy efficiency of differentalgorithms. As shown in Figs. 7 and 8, WiFi has the min-imum energy consumption under all scenarios, while3G always leads to the worst energy performance. Sinceboth network interfaces are exploited by GRD, OPP andE-S, their performance is bounded by these two extremecases. Specifically, for web browsing as shown in Fig.7, E-S performs closely to WiFi, and better than GRDand OPP by saving 16% and 15.8% energy consumption,respectively. A similar phenomenon can be observedunder the video watching application as shown in Fig.8.

7

1 2 3 4 505

10152025303540455055606570

Ene

rgy

cons

umpt

ion(

J)

Web site

WiFi 3G GRD OPP E-S

Fig. 7. Energy consumption of different methods in webbrowsing

1 2 3 4 50

50

100

150

200

250

300

350

400

450

500

550

600

650

700

Ene

rgy

cons

umpt

ion

(J)

Video

WiFi 3G GRD OPP E-S

Fig. 8. Energy consumption of different methods in videoplaying

We also notice that ALP-A exhibits different perfor-mance in web browsing and streaming media playing.For web browsing, there are various requests (e.g., C-SS, JavaScript, Html, pictures, etc.) with different datasize and QoE requirements. Since CSS and JavaScripthave more influence on pages rendering, these request-s impose close deadlines. On the other hand, high-throughput and continuous connection should be en-sured in streaming media playing. Therefore, both in-terfaces are activated at the beginning to realize thatplaying starts faster.

The real-time power consumption in web browsingand video watching is shown in Figs. 9 and 10, re-spectively. In Fig 9, we observe that the period of highpower state under E-S is shorter than the other twoalgorithms. For example, E-S stays in a low power stateuntil the 7th second while energy consumption of othersincreases from the 5th and 6th second. Furthermore, E-Sswitches back to a low power state a little earlier thanothers. Similarly, as shown in Fig. 10, E-S finishes all data

0 5 10 15 20 25 30 35 40 4520

40

60

80

100GRD

Time (s)

Po

wer

(mW

)

0 5 10 15 20 25 30 35 40 4520

40

60

80

100

OPP

Time (s)

Po

wer

(mW

)

0 5 10 15 20 25 30 35 40 4520

40

60

80

100E−S

Time (s)

Po

wer

(mW

)

Fig. 9. Real-time power consumption in web browsing(www.sina.com.cn)

0 10 20 30 40 50 60 70 80 90 100 110 1201000

2000

3000

4000

Time (s)

Po

wer

(mW

)

GRD

0 10 20 30 40 50 60 70 80 90 100 110 1201000

2000

3000

4000

Time (s)

Po

wer

(mW

)

OPP

0 10 20 30 40 50 60 70 80 90 100 110 1201000

2000

3000

4000E−S

Time (s)

Po

wer

(mW

)

Fig. 10. Real-time power consumption in video playing

transmissions within 100 seconds, while GRD and OPPneed 110 seconds and 105 seconds, respectively.

We also show the energy consumption of differentalgorithms during a long-running test by issuing a seriesof HTTP requests with different sizes for an hour. Asshown in Fig. 11, although the performance gap amongthese algorithms is not obvious in the beginning, E-Ssaves about 30% and 21% battery power of GRD andOPP, respectively, after an hour.

Due to bandwidth dynamic of both network inter-faces, some requests may not be satisfied within their

8

0 500 1000 1500 2000 2500 3000 350020

30

40

50

60

70

80

90

100

Time (s)

Ba

tte

ry l

ev

el

(%)

WiFi

GRD

OPP

E−S

Fig. 11. Energy consumption in a one-hour test

1 2 3 4 50

2

4

6

8

10

12

14

16

18

Req

uest

s m

iss

dead

line

(%)

Web site

WiFi GRD OPP E-S

Fig. 12. Number of requests completed upon deadlineover different methods

deadlines. We show the percentage of requests missingtheir deadlines under different algorithms in Fig. 12.Although WiFi always has the minimum energy con-sumption as we observe in Fig. 7, there are over 10%requests that cannot be finished within their deadlines.On the other hand, the corresponding percentages ofother algorithms are all under 5%.

We compare the loading time and rendering timefor web browsing (www.sina.com.cn) under differentalgorithms. The loading time is defined as the totaltime of satisfying all requests (328 requests as shownin Table 1) belonging to the web page. The renderingtime is the time of delivering only CSS and JavaScriptelements (93 elements as shown in Table 1). Comparedwith loading time, rendering time has greater effecton QoE because mobile users are sensitive to CSS andJavaScript elements as the main contents of web pages.As shown in Fig. 13, the GRD, OPP and E-S algorithmsexploiting both network interfaces have shorter loadingtime and rendering time. Furthermore, although E-S haslonger loading timer than GRD and OPP, it has similarrendering time with them, demonstrating that mobile

WiFi 3G GRD OPP E-S0

5

10

15

20

25

Tim

e (s

)

Rendering time Loading time

Fig. 13. Completion time and rendering time over differentmethods in web browsing (www.sina.com.cn)

100090080070060050040030020010000

5

10

15

20

25

30

WiFi throughput (KB/s)

3G

uti

liza

tio

n p

er

slo

t (K

B)

GRD

OPP

E−S

Fig. 14. 3G utilization over different WiFi throughput.

users can hardly be aware of QoE differences under thesealgorithms.

We investigate the 3G utilization of GRD, OPP andE-S under different WiFi transmission rate in Fig. 14.As WiFi rate decreases, 3G utilization grows under allalgorithms because more data need to be delivered over3G-interface to guarantee QoE. However, 3G utilizationof E-S is always lower than GRD and OPP due to itsawareness of energy efficiency.

After request partition in ALP-A, the generated sub-requests will be sent to lower layers (e.g., network layerand data link layer), where additional headers will beadded. To evaluate the overhead of this process, we mea-sure the ratio of additional header size and original datasize. We conduct five sets of experiments using randomlygenerated requests with different sizes. Specifically, weuse large request size in the first two sets of experiments,and small size in the rest. As shown in Table 2, the ratiois less than 0.14% in all experiments.

9

TABLE 2Overhead of ALP-A

No. of requests Additional (Byte) Original (Byte) Ratio

104 6656 48712096 0.014%51 1920 23542334 0.008%

1175 3392 3745042 0.09%380 2112 1537925 0.137%2043 4608 4396977 0.105%

7 CONCLUSION

In this paper, we develop a bandwidth aggregation pro-totype, named ALP-A, at the application layer for mobiledevices. An energy-aware scheduling algorithm calledE-Schedule is proposed to reduce energy consumptionwhile guaranteeing desired QoE. We implement ALP-A in real mobile devices and conduct extensive experi-ments to evaluate its performance. Compared with twostart-of-the-art solutions, our proposed algorithm cansave about 16% energy under web browsing and videostreaming applications while guaranteeing QoE.

ACKNOWLEDGEMENT

This paper was partially supported by China NationalNatural Science Foundation under grant No. 61272408,61322210, National High-tech Research and Develop-ment Program of China (863 Program) under grantNo.2012AA010905, and Doctoral Fund of Ministry ofEducation of China under grant No. 20130142110048.

REFERENCES[1] P. Bellavista, A. Corradi, and S. A. Malik, ”A survey of context

data distribution for mobile ubiquitous systems,” ACM ComputingSurveys (CSUR), vol. 44, no. 4, p. 24, 2012.

[2] M. Yousaf, A. Qayyum, and S. A. Malik, ”An architecture forexploiting multihoming in mobile devices for vertical handovers &bandwidth aggregation,” Wireless Personal Communications, vol.66, no. 1, pp. 57-79, 2012.

[3] A. Bazzi, G. Pasolini, and O. Andrisano, ”Multiradio resource man-agement: parallel transmission for higher throughput,” EURASIPjournal on advances in signal processing, vol. 2008, no. 1, p. 763264,2008.

[4] A. Miu, J. G. Apostolopoulos, W.-t. Tan and M. Trott, ”Low-latency wireless video over 802.11 networks using path diversity,”Proc. IEEE. International Conference on Multimedia and Expo(ICME’03), pp. II-441, 2003.

[5] A. L. Ramaboli, O. E. Falowo, and A. H. Chan, ”Bandwidthaggregation in heterogeneous wireless networks: A survey ofcurrent approaches and issues,” Journal of Network and ComputerApplications, vol. 35, no. 6, pp. 1674-1690, 2012.

[6] M. Allman, H. Kruse, and S. Ostermann, ”An application-level so-lution to TCPs satellite inefficiencies,” Proc. The First InternationalWorkshop on Satellite-based Information Services (WOSBIS), 1996.

[7] W. Allcock, J. Bester, J. Bresnahan, A. Chervenak, L. Liming, and S.Tuecke, ”GridFTP: Protocol extensions to FTP for the Grid,” GlobalGrid ForumGFD-RP, vol. 20, 2003.

[8] H. Sivakumar, S. Bailey, and R. L. Grossman, ”PSockets: The casefor application-level network striping for data intensive applica-tions using high speed wide area networks,” Proc. IEEE. The 2000ACM/IEEE conference on Supercomputing (CDROM), p. 37, 2000.

[9] D. Kaspar, K. Evensen, P. Engelstad, A. F. Hansen, P. Halvorsen,and C. Griwodz, ”Enhancing video-on-demand playout over mul-tiple heterogeneous access networks,” Proc. IEEE. 7th ConsumerCommunications and Networking Conference (CCNC), pp. 1-5,2010

[10] D. Kaspar, K. Evensen, P. Engelstad, and A. F. Hansen, ”UsingHTTP pipelining to improve progressive download over multipleheterogeneous interfaces,” Proc. IEEE. International Conference onCommunications (ICC), pp. 1-5, 2010.

[11] K. Evensen, D. Kaspar, C. Griwodz, P. Halvorsen, A. Hansen,and P. Engelstad, ”Improving the performance of quality-adaptivevideo streaming over multiple heterogeneous access networks,”Proc. ACM. The second annual ACM conference on MultimediaSystems, pp. 57-68, 2011.

[12] J. R. Iyengar, P. D. Amer, and R. Stewart, ”Concurrent multipathtransfer using SCTP multihoming over independent end-to-endpaths,” IEEE/ACM Trans. Networking, vol.14, no.5, pp. 951-964,2006.

[13] H. Adhari, T. Dreibholz, and M. Becke, ”SCTP socket API ex-tensions for concurrent multipath transfer,” Draft-dreibholz-tsvwg-sctpsocket-multipath-01, 2012.

[14] H.-Y. Hsieh, K.-H. Kim, and R. Sivakumar, ”An end-to-endapproach for transparent mobility across heterogeneous wirelessnetworks,” Mobile Networks and Applications, vol. 9, no. 4, pp.363-378, 2004.

[15] H.-Y. Hsieh and R. Sivakumar, ”A transport layer approach forachieving aggregate bandwidths on multi-homed mobile hosts,”Wireless Networks, vol. 11, no. 1-2, pp. 99-114, 2005.

[16] A. Ford, C. Raiciu, M. Handley, S. Barre, and J. Iyengar, ”Archi-tectural guidelines for multipath TCP development,” IETF, Infor-mational RFC, vol. 6182, pp. 2070-1721, 2011.

[17] C. Paasch, G. Detal, F. Duchene, C. Raiciu, and O. Bonaventure,”Exploring mobile/WiFi handover with multipath TCP,” Proc.ACM. The 2012 ACM SIGCOMM workshop on Celluar networks:operations, challenges, and future design, pp. 31-36, 2012.

[18] S. Chen, Z. Yuan, and G.-M. Muntean, ”An energy-awaremultipath-TCP-based content delivery scheme in heterogeneouswireless networks,” Proc. IEEE. Wireless Communications andNetworking Conference (WCNC), pp. 1291-1296, 2013.

[19] Y. Zhang, C. Wang, and Y. Gao, ”Weighted size-aware packetdistribution for multipath live streaming,” Proc. IEEE. Commu-nications (ICC’09), pp. 1-5, 2009.

[20] K. Chebrolu, B. Raman, and R. R. Rao, ”A network layer ap-proach to enable TCP over multiple interfaces,” Springer. WirelessNetworks, vol. 11, no. 5, pp. 637-650, 2005.

[21] K. Chebrolu and R. R. Rao, ”Bandwidth aggregation for real-time applications in heterogeneous wireless networks,” IEEE Trans.Mobile Computing, vol. 5, no. 4, pp. 388-403, 2006.

[22] J. Sachs, ”A generic link layer for future generation wirelessnetworking,” Proc. IEEE. International Conference on Communi-cations (ICC’03), vol. 2, pp. 834-838, 2003.

[23] G. P. Koudouridis, H. R. Karimi, and K. Dimou, ”Switched multi-radio transmission diversity in future access networks,” Proc. IEEE.Vehicular Technology Conference (VTC’05), pp. 235-239, 2005.

[24] G. P. Koudouridis, R. Aguero, E. Alexandri, J. Choque, K. Dimou,H. Karimi, H. Lederer, J. Sachs, and R. Sigle, ”Generic link layerfunctionality for multi-radio access networks,” IST Mobile andWireless Communications Summit, 2005.

[25] E. Harjula, O. Kassinen, and M. Ylianttila, ”Energy consumptionmodel for mobile devices in 3G and WLAN networks,” Proc.IEEE. Consumer Communications and Networking Conference(CCNC’12), pp. 280-293, 2009.

[26] M. Z. Chowdhury, Y. M. Jang, C. S. Ji, S. Choi, H. Jeon, J.Jee, and C. Park, ”Interface selection for power management inUMTS/WLAN overlaying network,” Proc. IEEE. Advanced Com-munication Technology (ICACT’09), pp. 795-799, 2009.

[27] H. Mahkoum, B. Sarikaya, and A. Hafid, ”A framework forpower management of handheld devices with multiple radios,”Proc. IEEE. Wireless Communications and Networking Conference(WCNC’09), pp. 1-6, 2009.

[28] E. Ziouva and T. Antonakopoulos, ”CSMA/CA performanceunder high traffic conditions: throughput and delay analysis,”Elsevier. Computer Communications, vol. 25, no. 3, pp. 313-321,2002.

[29] 3GPP TR 37.804, http://www.qtc.jp/3GPP/Specs/37804-b00.pdf,2006

[30] Balasubramanian N, Balasubramanian A, Venkataramani A. ”En-ergy consumption in mobile phones: a measurement study andimplications for network applications,” Proc. ACM. The 9th ACMSIGCOMM conference on Internet measurement conference, pp.280-293, 2009.

10

[31] J. Huang, F. Qian, A. Gerber, Z. M. Mao, S. Sen, and O. Spatscheck,”A close examination of performance and power characteristics of4G LTE networks” Proc. ACM. Mobisys, pp. 225-238, 2012.

[32] X. Liu, E. K. Chong, and N. B. Shroff, ”A framework for op-portunistic scheduling in wireless networks,” Elsevier. ComputerNetworks, vol. 41, no. 4, pp. 451-474, 2003.

Zaiyang Tang received his BS degree fromHuazhong University of Science and Technolo-gy, China, in 2011. He is a PhD candidate inHuazhong University of Science and Technol-ogy. His research interests include networkingmodeling, network virtualization, machine learn-ing and recommendation system.

Zirui Wang received his BS degree from Cheng-du University of Technology, China, in 2011, theMS degrees from the Huazhong University ofScience and Technology. His research interestsinclude mobile network and network virtualiza-tion.

Peng Li received his BS degree from HuazhongUniversity of Science and Technology, China, in2007, the MS and PhD degrees from the Univer-sity of Aizu, Japan, in 2009 and 2012, respec-tively. He is currently an Associate Professor atSchool of Computer Science and Engineering,the University of Aizu, Japan. His research in-terests include networking modeling, cross-layeroptimization, wireless sensor networks, cloudcomputing, smart grid, performance evaluationof wireless and mobile networks for reliable,

energy-efficient, and cost-effective communications. He is a member ofthe IEEE.

Song Guo (M’02-SM’11) received the PhD de-gree in computer science from University of Ot-tawa, Canada. He is currently a Full Professorat School of Computer Science and Engineer-ing, the University of Aizu, Japan. His researchinterests are mainly in the areas of protocoldesign and performance analysis for computerand telecommunication networks. He receivedthe Best Paper Awards at ACM IMCOM 2014,IEEE CSE 2011, and IEEE HPCC 2008. Dr. Guocurrently serves as Associate Editor of the IEEE

Transactions on Parallel and Distributed Systems. He is in the editorialboards of ACM/Springer Wireless Networks, Wireless Communicationsand Mobile Computing, and many others. He has also been in orga-nizing and technical committees of numerous international conferences,including serving as a General Co-Chair of MobiQuitous 2013. Dr. Guois a senior member of the IEEE and the ACM.

Xiaofei Liao received his Ph.D. degree in com-puter science and engineering from HuazhongUniversity of Science and Technology (HUST),China, in 2005. He is now a professor in theschool of Computer Science and Engineering atHUST. He has served as a reviewer for manyconferences and journal papers. His researchinterests are in the areas of system software,P2P system, cluster computing and streamingservices. He is a member of the IEEE and theIEEE Computer society.

Hai Jin is a Cheung Kung Scholars Chair Pro-fessor of computer science and engineering atthe Huazhong University of Science and Tech-nology (HUST) in China. He is now Dean ofthe School of Computer Science and Technolo-gy at HUST. Jin received his PhD in computerengineering from HUST in 1994. In 1996, hewas awarded a German Academic ExchangeService fellowship to visit the Technical Uni-versity of Chemnitz in Germany. Jin worked atThe University of Hong Kong between 1998 and

2000, and as a visiting scholar at the University of Southern Californiabetween 1999 and 2000. He was awarded Excellent Youth Award fromthe National Science Foundation of China in 2001. Jin is the chiefscientist of ChinaGrid, the largest grid computing project in China, andthe chief scientist of National 973 Basic Research Program Project ofVirtualization Technology of Computing System. Jin is a senior memberof the IEEE and a member of the ACM. Jin is the member of GridForum Steering Group (GFSG). He has co-authored 15 books andpublished over 400 research papers. His research interests includecomputer architecture, virtualization technology, cluster computing andgrid computing, peer-to-peer computing, network storage, and networksecurity. Jin is the steering committee chair of International Confer-ence on Grid and Pervasive Computing (GPC), Asia-Pacific ServicesComputing Conference (APSCC), International Conference on Frontierof Computer Science and Technology (FCST), and Annual ChinaGridConference. Jin is a member of the steering committee of the IEEE/ACMInternational Symposium on Cluster Computing and the Grid (CCGrid),the IFIP International Conference on Network and Parallel Computing(NPC), and the International Conference on Grid and CooperativeComputing (GCC), International Conference on Autonomic and TrustedComputing (ATC), International Conference on Ubiquitous Intelligenceand Computing (UIC).