mimo systems

Computer Networks 56 (2012) 2506–2521

Contents lists available at SciVerse ScienceDirect

Computer Networks

journal homepage: www.elsevier .com/locate /comnet

Measurement-based modelling of power consumption at wirelessaccess network gateways q

Karina Gomez a,b, Dejene Boru b, Roberto Riggio a,⇑, Tinku Rasheed a, Daniele Miorandi a,Fabrizio Granelli b

a CREATE-NET, Via Alla Cascata 56/C, 38123 Trento, Italyb DISI – University of Trento, Via Sommarive 14, 38100 Trento, Italy

a r t i c l e i n f o a b s t r a c t

Article history:Received 30 September 2011Received in revised form 23 February 2012Accepted 20 March 2012Available online 19 April 2012

Keywords:Green NetworkingWiMAXWiFiPower consumptionEnergy efficiencyModelling

1389-1286/$ - see front matter � 2012 Elsevier B.Vhttp://dx.doi.org/10.1016/j.comnet.2012.03.028

q Part of the results appeared in Proc. of IEEEresearch leading to these results are supported partCommunity’s Seventh Framework Programme (FPthe framework of the SMART-Net Project (Grant No⇑ Corresponding author.

E-mail addresses: [email protected]@studenti.unitn.it (D. Boru), roberto.riggRiggio), [email protected] (T. [email protected] (D. Miorandi), [email protected]

Improving the energy efficiency of the ICT sector is becoming an ambitious challenge forindustries and research communities alike. Understanding how the energy is consumedin each part of an ICT system becomes fundamental in order to minimize the overall energyconsumed by the system itself. In this paper, we propose an experimentally-drivenapproach to (i) characterize typical wireless access network gateways from an energy con-sumption standpoint and (ii) develop simple and accurate power consumption models forsuch gateways. In this work we focused our attention on the monitoring, measurement andanalysis of the energy consumption patterns of WiFi and WiMAX gateways. Our measure-ments show that the power consumption of such gateways exhibits a linear dependence onthe traffic until a saturation point is reached.

� 2012 Elsevier B.V. All rights reserved.

1. Introduction

In the last decade, ‘‘Green Networking’’ [2–5] has gainedconsiderable importance for both commercial entities andresearchers that aim at understanding and reducing theenergy consumption of computing and communicationinfrastructures [6,7]. Several studies have shown that theICT sector accounts for 2–7% of the global energy consump-tion [8,9,3] and it is also responsible for 2–3% of total emis-sions of CO2 [10–12]. Moreover, it is important to remarkthat about 50% of the total energy used in the ICT sector

. All rights reserved.

GreenCom [1]. Theially by the European7/2007-2013) within. 223937).

(K. Gomez), [email protected] (R.d), daniele.mioran-

.it (F. Granelli).

is consumed at the wireless access part [10,12]. Therefore,network operators and service providers currently com-pete to optimize the energy efficiency of their access infra-structure in order to reduce both CO2 emissions andoperational costs [13].

In such a scenario, two broadband wireless access tech-nologies, WiFi and WiMAX, are witnessing an increasingusage in both metropolitan and rural deployments. Thereason behind their adoption lies in the minimal support-ing infrastructure required for their operations, which intime enables a high level of flexibility in network deploy-ments, allowing connectivity to be provided only whereand when needed. Such fluidity in network deploymentand operations is made possible by the architecturalchoices underpinning the standards. However, while en-ergy efficiency trade-offs have been taken into accountfor the end-users’ terminal, which can be mobile or noma-dic, less attention has been devoted to the networkgateways which, in either the WiFi and the WiMAX stan-dards, are typically connected to the power grid and, thus,

http://dx.doi.org/10.1016/j.comnet.2012.03.028

mailto:[email protected]








http://dx.doi.org/10.1016/j.comnet.2012.03.028

http://www.sciencedirect.com/science/journal/13891286

http://www.elsevier.com/locate/comnet

K. Gomez et al. / Computer Networks 56 (2012) 2506–2521 2507

do not pose energy consumption challenges. As a resultthere is a lack of best practices for designing energy effi-cient network protocols and architectures for broadbandwireless access networks.

The main objective of this work is to experimentallymeasure and analyse the energy consumption patterns ofWiFi and WiMAX gateways1 at both the component andthe network level. In particular, our experiments aim atanswering the following questions:

� Where and how is the power consumed in WiFi andWiMAX access gateways? How much of the power iswasted?� What is the relationship between traffic load and power

consumption in WiFi and WiMAX access gateways?� What are the critical aspects of the IEEE 802.11 and the

IEEE 802.16 standards with respect to powerconsumption?

It is the authors’ standpoint that the answers to thesequestions are very important since they would provideus with an increased insight into the network behaviour,paving the way to the development of (i) realistic modelsfor power consumption in wireless networks and (ii) pro-tocols and algorithms for their operations. The main con-tributions of the paper are the following:

� An empirical approach for understanding the energyconsumption behaviour of WiMAX and WiFi gateways.Thereby, we target the characterization of the powerconsumption of WiFi and WiMAX gateways in termsof (i) the amount of traffic sent/received by the node,(ii) the modulation and coding schemes used, and (iii)the size of the session level data units.� A simple model for the characterization of the power

consumption in WiFi and WiMAX access gateways ispresented.

The remainder of this paper is organized as follows. InSection 2 we present the experimental settings and meth-odology used. Experimental results are reported and dis-cussed in Section 3. In Section 4 we discuss energyefficiency metrics and models. A brief analysis of the re-lated works on measurement driven energy efficiencyanalysis is presented in Section 5. Finally, Section 6 is de-voted to the final conclusions and pointers to promising re-search directions.

2. Evaluation methodology

In this section, we will describe the network setups andthe methodology used in order to evaluate the power effi-ciency of the two wireless networking technologies thathave been considered in this paper, namely an indoor WiFinetwork deployed in a typical office environment and anoutdoor WiMAX network deployed in a rural area. The net-work setups exploited in the WiMAX and the WiFi scenar-ios are sketched respectively in Fig. 1a and b.

1 With a slight abuse of terminology we use the term gateway to refer toboth the WiMAX Base Station (BS) and to the WiFi Access Point (AP).

2.1. Network settings

In the WiMAX case, the network is composed of a BaseStation deployed on the rooftop of a building and a singlestatic Subscriber Station (SS) deployed on the rooftop ofanother building. The testbed is deployed across the Or-ange Labs Campus in Lannion, France. The BS and the SSare about 800 m away from each other and are operatingunder line of sight conditions. The WiMAX equipment iscompliant with the IEEE 802.16-2004 version of thestandard and implements the TDD duplexing scheme.The devices operate between 5.47 and 5.725 GHz usingomni–directional antennas with a gain of 8 dB. With regardto QoS, the devices support the Best Effort (BS), theReal-time Polling Service (rtPS), the Non-real-time PollingService (nrtPS), and the Unsolicited Grant Service (UGS)traffic classes. The BE class has been used throughout theentire measurements campaign reported in this work.

In the WiFi case, the network is composed of a customIEEE 802.11g Access Point and a single DELL LatitudeD620 notebook acting as static wireless client. The testbedis deployed at CREATE-NET premises in Trento, Italy. TheAP is built around a PCEngines ALIX 2C2 (500 MHz �86CPU, 256 MB of RAM) processor board equipped with twoIEEE 802.11a/b/g wireless interfaces (Atheros AR5213Achipset) with RTC/CTS disabled, while the notebook isequipped with an Intel PRO/Wireless 3945AB wirelessadapter. The frequency of operation of the AP is2.412 GHz and the antenna is omni-directional with a gainof 8 dB.

It is important to note that, unless otherwise specified,the rate adaptation algorithm has been set to auto andthe transmission power has been left to its default valueequal to 18 dBm (�63.1 mW) for the WiFi and the WiMAXcases.

2.2. Traffic generation and power consumption monitoring

Traffic is generated using the Iperf traffic generator2 andis injected into the network trough either the Server or theClient. In the former case, we aimed at measuring the powerconsumed by the BS/AP when it is acting as transmitter,while in the latter case we aimed at measuring the powerconsumed by the BS/AP when it is acting as receiver. In bothcases the power consumption figures reported in this workrefer to the BS/AP.

The power consumption is measured using the WattsUp?3 power meter. Watts Up? is a ‘‘plug load’’ meter thatmeasures the amount of electricity used by whatever elec-trical appliance is plugged into it. The meter incorporatesdigital electronics to perform accurate power consumptionmeasurements. Such measurements are then logged intothe device’s internal memory with a granularity of 0.1 Wattsand a sampling period of 1 s.

The Watts Up? meter is interconnected through its USBinterface to the Server where a custom data logging soft-ware is used in order to extract the power consumption

2 Available at: http://iperf.sourceforge.net/.3 Available at: http://www.wattsupmeters.com/.

http://iperf.sourceforge.net/

http://www.wattsupmeters.com/

Fig. 1. Network scenario used for the measurement campaign.

Table 1IEEE 802.11g and IEEE 802.16 OFDM rates and modulation types.

Modulation type Data rate (Mb/s)

Binary Phase Shift Keying (BPSK) 6/9Quadrature Phase Shift Keying (QPSK) 12/1816-Quadrature Amplitude Modulation (16-QAM) 24/3664-Quadrature Amplitude Modulation (64-QAM) 48/54

2508 K. Gomez et al. / Computer Networks 56 (2012) 2506–2521

samples. It is important to remark that the power con-sumption is monitored for the whole device. Therefore,the results reported in this paper account for both thepower consumed by the processing board for handlingthe incoming and outgoing traffic (e.g. for segmentationand reassembling, protocol overheads, computing check-sums, etc.) as well as for the power consumed to deliverthe actual frame over the wireless link (e.g. power amplifi-ers, modulator/demodulator, etc.).

2.3. Testing methodology

Synthetic traffic is modeled as single UDP flows. Resultsreported in this paper are the average of 4 runs. Results arereported in terms of a 95% confidence interval. Each runwas 500 s long for the WiMAX case and 400 s long for WiFicases. The following scenarios have been considered:

� Constant Traffic Generation Rate, Variable Packet Length.In this test, the traffic generation rate is kept constantwhile the datagram size is progressively increased from32 to 2816 bytes in steps of 256 bytes. Two differentsettings have been considered for the traffic generationrate resulting in application throughput of 1 Mb/s and10 Mb/s for the WiFi case and of 1 Mb/s and 5 Mb/sfor WiMAX case.� Variable Traffic Generation Rate, Fixed Packet Length. In

this test, the datagram size is kept constant at1280 bytes, while the traffic generation rate is progres-sively increased from 0.1 Mb/s up to 54 Mb/s for theWiFi case and from 0.1 Mb/s up to 30 Mb/s for theWiMAX case.� Variable Traffic Generation Rate, Mixed Transmission

Power. In this test, the datagram size is kept constantat 1280 bytes while the traffic generation rate is pro-gressively increased. This test has been repeated usingtwo different transmission power levels. In the WiFicase, the first set of measurements has been performedwith a transmission power set to 10 dBm (�10 mW),while in second round of measurements the transmis-sion power has been increased to 18 dBm (�63.1 mW)for both the AP and the client. Similarly, in the WiMAXthe transmission power level have been set to either10 dBm (�10 mW) or 26 dBm (�398 mW).

� Variable Traffic Generation Rate, Fixed Modulation Type. Inthis test, the datagram size is kept constant at1280 bytes while the message generation rate is pro-gressively increased. The rate control algorithm is dis-abled and the transmission rate is set manually usingthe command line interface. The experiment is repeatedfor each of the transmission rates supported by thewireless adapter (see Table 1).

It is worth noticing that, the frame length of 1280 bytesused in the scenario 2–4 has been chosen in that, accordingto the outcomes of the first scenario, it is the length atwhich the power efficiency of both systems is maximized.

3. Experimental measurements and analysis

In this section, we present the results from the mea-surements campaign using WiMAX and WiFi test environ-ments. We use the following notation for the figuresthroughout the section (a) BS/AP-Receiver is when BS/APis acting as receiver and (b) BS/AP-Transmitter when BS/AP is acting as transmitter.

In Fig. 2a and c, the average power consumption level atthe BS and at the AP as a function of the datagram size for aconstant throughput is presented. This gives us certain in-sights in order to understand ‘‘how’’ the datagram size im-pacts power consumption. Results are plotted when theBS/AP are acting as transmitter and receiver respectively.As it can be seen, the power consumption behaviour forthis experiment is similar for both the gateways (BS andAP). The datagram size has a considerable impact on powerconsumption when the BS/AP are acting as transmitters.However for the WiMAX case, when the BS is acting as re-ceiver, there is no impact on power consumption for anydatagram size.

64 256 512 768 1024 1280 1536 1792 2048 2304 2560 281615.916.116.316.516.716.917.117.317.517.7

Datagram size [bytes]

Pow

er C

onsu

mpt

ion

[W]

BS−Transmitter 1Mp/sBS−Receiver 1Mp/sBS−Transmitter 5Mp/sBS−Receiver 5Mp/s

(a) WiMAX case: Power Consumption

256 512 768 1024 1280 1536 1792 2048 2304 2560 28160

0.51

1.52

2.53

3.54

4.55

Datagram size [byte]

Thro

ughp

ut [M

b/s]

BS−Transmitter 5Mb/sBS−Receiver 5Mb/s

(b) WiMAX case: Throughput for the experi-ment with 5 Mb/s

64 256 512 768 1024 1280 1536 1792 2048 2304 2560 28164.7

4.8

4.9

5.0

5.1

5.2

5.3

Pow

er C

onsu

mpt

ion

[W]


AP−Transmitter 1Mb/sAP−Receiver 1Mb/s AP−Transmitter 10Mb/sAP−Receiver 10Mb/s

(c) WiFi case: Power Consumption

256 512 768 1024 1280 1536 1792 2048 2304 2560 28150123456789

10

Thro

ughp

ut [M

b/s]

Datagram size [byte]

AP−Transmitter 10Mb/sAP−Receiver 10Mb/s

(d) WiFi case: Throughput for the experimentwith 10 Mb/s

Fig. 2. Average power consumption and throughput at either the BS or the AP as a function of the datagram size for a constant traffic generation rate (1, 5and 10 Mb/s). Datagram loss was less than 1% during all measurements.


Instead for the WiFi case, when the AP is acting as recei-ver, the impact on power consumption for any datagramsize follows similar behavior of the AP when acting astransmitter but with a lower power consumption. Such dif-ferences in power consumption while the BS/AP is inreceiving mode can be attributed to the manner in whichthe MAC receives and processes the packets in the caseof the two access technologies. It is also important to notethat the amount of traffic, i.e. 1, 5 and 10 Mb/s for theexperiment, does not affect the behavior of the power con-sumption vs. datagram size. However, for low amounts oftraffic the difference between power consumption whenthe AP is acting as receiver or transmitter is lower. Finally,we can observe that (i) for low datagram sizes, the devicepower is wasted because the BS/AP consumes significantlymore power than for the large datagram sizes under thesame traffic conditions and (ii) when the datagram size be-comes extremely large, fragmentation takes place and con-sequently, the bandwidth utilization decreases. Theseeffects are shown in Fig. 2b and d, the results show thethroughput for the experiments performed by 5 Mb/s forWiMAX case and 10 Mb/s for WiFi case. Therefore, theoptimal datagram length in terms of power consumptionand network performance for a static client is 1280 bytes.The datagram loss for these set of experiments was lowerthan 1%.

In order to study the relationship between traffic loadand power consumption, we performed a set of measure-ments aimed at characterizing the power consumption ofboth the WiFi and the WiMAX networks when synthetic

traffic is injected at an increasing rate. The outcomes of thisset of measurements are reported in Fig. 3. Traffic is in-jected at the BS, respectively at the AP. Each step in thepower consumption corresponds to an increase in the traf-fic rate, starting from no traffic up to 30 Mb/s and 54 Mb/sfor, respectively, the WiMAX and the WiFi networks. As itcan be seen, the power consumption of the BS in idle modeis �16.05 W and monotonically increases with the trafficuntil a saturation point is reached. The same behaviorcan be observed for the AP.

The average power consumption of the BS as a functionof different traffic generation rates is shown in Fig. 4, for afixed datagram size of 1280 bytes. The power consumptionwhen the BS is acting as transmitter increases according tothe increase in traffic. The power consumption when theBS is acting as receiver remains unchanged, as expected.We can observe in Fig. 4b that when the BS reaches satura-tion, the maximum throughput value remains constant. Inthese set of experiments, the datagram loss starts to in-crease when the BS reaches saturation for both situations,while it is acting as transmitter and receiver, these resultsare shown in Fig. 4c.

In Fig. 5, we report the average power consumption ofthe AP as a function of different traffic generation rates,for a fixed datagram size of 1280 bytes. As it can be intui-tively corroborated, the relationship between powerconsumption and traffic for the AP while acting as trans-mitter follows the same behavior like the BS in WiMAX.In contrast, when the AP is acting as receiver the powerconsumption increases the same way compared to the AP

0 1000 2000 3000 4000 5000 6000 7000 8000 9000 1000015.816.116.416.717.017.317.617.918.218.518.8

Time [s]

Pow

er C

onsu

mpt

ion

[W] BS−Transmitter

idleperiod

(a) WiMAX

0 500 1000 1500 2000 2500 3000 3500 4000 45004.74.84.9

55.15.25.35.4

Time [s]

Pow

er C

onsu

mpt

ion

[W] AP−Transmitter

(b) WiFi

Fig. 3. Power consumption when synthetic traffic is injected at an increasing rate at either the BS (WiMAX) or at the AP (WiFi).

0.11 2 3 4 5 6 7 8 9 10 12 14 16 18 20 22 24 26 28 3015.816.216.617.017.417.818.218.6

Traffic generation rate [Mb/s]

Pow

er C

onsu

mpt

ion

[W] BS−Transmitter

BS−Receiver


0.10.5 1 2 3 4 5 6 7 8 9 10 12 14 16 18 20 22 24 26 28 300

2.5

5.0

7.5

10.0

12.5

15.0


Thro

ughp

ut [M

b/s]

BS−TransmitterBS−Receiver

(b) WiMAX case: Throughput

0.10.5 1 2 3 4 5 6 7 8 9 10 12 14 16 18 20 22 24 26 28 300

10

20

30

40

50

60

70

80


Dat

agra

m L

oss

[%]

BS−TransmitterBS−Receiver

(c) WiMAX case: Datagram loss

Fig. 4. Average power consumption, throughput, and packet loss as a function of different traffic generation rates for a constant datagram size of1280 bytes.


when acting as transmitter but with lower power con-sumption. Fig. 5b reports the throughput while Fig. 5c re-ports the datagram loss for the experiments when AP is

acting as transmitter and receiver. We can also observethe effect of saturation in these figures and the increaseof datagram loss for large amount of traffic when AP is

1 2 3 4 5 6 7 8 910 12 14 16 18 20 24 26 28 32 36 40 44 48 52 544.7

4.8

4.9

5.0

5.1

5.2

5.3

5.4

Pow

er C

onsu

mpt

ion

[W]


AP−Transmitter AP−Receiver

(a) WiFi case: Power Consumption

1 2 3 4 5 6 7 8 9 10 12 14 16 18 20 24 26 28 32 36 40 44 48 52 540

5

10

15

20

25

30

Thro

ughp

ut [M

b/s]


AP−Receiver AP−Transmitter

(b) WiFi case: Throughput

1 2 3 4 5 6 7 8 9 10 12 14 16 18 20 24 26 28 32 36 40 44 48 52 5405

1015202530354045505560

Dat

agra

m L

oss

[%]


AP−ReceiverAP−Transmitter

(c) WiFi case: Datagram loss

Fig. 5. Average power consumption, throughput, and packet loss as a function of different traffic generation rates for a constant datagram size of1280 bytes.


acting as transmitter. This is due to the fact that, due to thelow memory resources of the AP, packets are discarded atthe transmission buffer before they get a chance to betransmitted. Instead, when the AP is acting as receiverand the client is acting as transmitter, the larger memoryresources that are available at the notebook allow packetsto be buffered until a transmission opportunity can be ob-tained. This explains the fact that when the AP is acting asreceiver the packets loss is less that 1% for any amount oftraffic.

In the next set of measurements, we report only thepower consumption statistics relative to the case in whichthe BS is acting as transmitter. The power consumptionwhen the BS is acting as receiver remains always un-changed as we have explained before. We considered twodifferent transmission power levels for the WiMAX case,i.e. 10 dBm and 26 dBm or 10 mW and �398.1 mW,respectively. Results are reported in Fig. 6a. Fig. 6a showsthe average power consumption at the BS as a functionof different traffic generation rates for transmission powerof 10 dBm and 26 dBm with datagram size equal to1280 bytes.

Results show that two different transmitter power lev-els are characterized by the same power consumption insaturation. The reason for this behavior is that the contri-bution of the power amplifier to the overall consumptionis below the sensibility of the meter used in our measure-ments (i.e. lower than 0.1 W). On the other hand whatcame as a surprise is the fact that, when using the hightransmission power level (26 dBm), the saturation isreached for slower data rates than when using the lowtransmission power level (10 dBm). Moreover, increasingthe transmitter power results in an increased datagram

loss, while one would expected that a better SINR wouldgive better performance in terms of packet loss.

In order to investigate this phenomena, we carried outadditional measurements by generating a saturated TCPconnection from the BS to the SS. The test was repeatedusing different transmission power levels. The outcomesare reported in Fig. 7. As it can be seen, for low transmis-sion power levels, the BS uses high rate modulation andcoding schemes (16-QAM 3/4), while when the transmis-sion power level increases the BS switches to less efficientmodulation and coding schemes (16-QAM 1/2 and QPSK 3/4). This choice results in less airtime-efficient modulationand coding schemes being used when the transmitterpower is set to 26 dBm which in time causes the systemto reach the saturation point for lower datarates (seeFig. 6a). Likewise, high datagram loss is experienced bythe system when the traffic generation rate increases inthat less efficient modulation and coding schemes resultis a low bitrate wireless link which in time causes data-grams to be dropped at the wireless interface.

Two different transmission power levels are also con-sidered for the WiFi case, i.e. 10 dBm and 18 dBm or10 mW and �63.1 mW, respectively. Fig. 6c reports theaverage power consumption at the AP as a function of dif-ferent traffic generation rates for transmission power of10 dBm and 18 dBm with datagram size equal to 1280 by-tes. The advantage of decreasing the transmission powercan be clearly observed only when the AP is acting as recei-ver. Instead, when the AP is acting as transmitter with10 dBm and 18 dBm, the results are comparable. Fig. 6d re-ports the datagram loss when the AP is acting as transmit-ter and receiver. The percentage of packet loss when the APis transmitting with 10 dBm is lower than when the AP is

0.11 3 5 7 9 11 13 15 19 2116.016.316.616.917.217.517.818.118.418.719.0


Pow

er C

onsu

mpt

ion

[W]

BS−Transmitter (10dBm)BS−Transmitter (26dBm)


0.1 1 3 5 7 9 11 13 15 19 210

1020304050607080


Dat

agra

m L

oss

[%]

BS−Transmitter (10dBm)BS−Transmitter (26dBm)

(b) WiMAX case: Datagram loss

1 4 7 10 14 18 24 34 44 54

4.74.84.95.05.15.25.35.45.5

Pow

er C

onsu

mpt

ion

[W]


AP−Receiver (10dBm)AP−Receiver (18 dBm)AP−Transmitter (18 dBm)AP−Transmitter (10 dBm)

(c) WiFi case: Power Consumption

1 4 7 10 14 18 24 34 44 5405

1015202530354045505560

Dat

agra

m L

oss

[%]


AP−Transmiter (10 dBm)AP−Transmiter (18 dBm)AP−Receiver (10 dBm)AP−Receiver (18 dBm)

(d) WiFi case: Datagram loss

Fig. 6. Average power consumption at the BS and AP and datagram loss as a function of different traffic generation rates for different transmission powerlevels. Datagram size equal to 1280 bytes.

16 17 18 19 20 21 22 23 24 251

2

3

4

5

6

BS Transmission Power [dBm]

DL

Thro

ughp

ut [M

b/s]

16−QAM 3/4

16−QAM 1/2

QPSK 3/4

Fig. 7. Goodput between BS and SS using different transmission powerlevels. The results refer to a saturated TCP connection.


transmitting with 26 dBm, especially with larger amountsof traffic, as explained earlier.

Lastly, we study the impact of modulation and codingschemes on power consumption patterns. We forced themodulation and coding schemes to be used by wirelessinterfaces and measured the power consumption as afunction of the traffic generation rate. The results for theWiMAX case are the following:

� BPSK 1/2 and QPSK 3/4 modulations: Fig. 8a reports theaverage power consumption at the BS when acting astransmitter, as a function of different traffic generationrates for BPSK 1/2 and QPSK 3/4 modulations. Datagramsize equal to 1280 bytes. Fig. 8b reports the datagramloss.

� QPSK 1/2 and 16-QAM 1/2 modulations: Fig. 8c reportsthe average power consumption at the BS when actingas transmitter, as a function of different traffic genera-tion rates for QPSK 1/2 and 16-QAM 1/2 modulations.Datagram size equal to 1280 bytes. Fig. 8d reports thedatagram loss.� 16-QAM 3/4 modulation and Auto–rate: Fig. 8e reports

the average power consumption at the BS when actingas transmitter, as a function of different traffic genera-tion rates for 16-QAM 3/4 modulation and Auto-modu-lation. Datagram size equal to 1280 bytes. Fig. 8f reportsthe datagram loss.

The figures shows that for the WiMAX case: (i) QPSK 3/4modulation is better than BPSK modulation 1/2 and (ii) 16-QAM 1/2 modulation is better than QPSK 1/2 modulationin terms of power consumption and network performancefor a static client. Therefore higher modulation rates aremore power efficient. This is understood to be due to thefact that higher modulation and coding schemes keep thetransmitter RF interface in the ‘‘on’’ state for a shorteramount of time. Of course this holds in a situation in whichthe channel condition is good. Finally, we can see also thatthe 16-QAM 3/4 modulation is better than auto–modula-tion in terms of power consumption and network perfor-mance for a static client. This is because the rate controlalgorithms are based in parameters related with the chan-nel, such as, RSSI and transition successful probabilities.However, the rate control algorithm cannot quickly adaptto the channel variations and it chooses a lower

0.1 1 3 5 7 9 1118.7

18.8

18.9

19.0

19.1

19.2

19.3

19.4


Pow

er C

onsu

mpt

ion

[W]

BS−Transmitter (BPSK 1/2)BS−Transmitter (QPSK 3/4)


0.1 1 3 5 7 9 110

10203040506070


Dat

agra

m L

oss

[%]

BS−Transmitter (BPKS 1/2) BS−Transmitter (QPSK 3/4)

(b) WiMAX case: Datagram loss

0.1 1 3 5 7 9 1116.016.316.616.917.217.517.818.118.418.719.0


Pow

er C

onsu

mpt

ion

[W]

BS−Transmitter (QPSK 1/2)BS−Transmitter (16−QAM 1/2)

(c) WiMAX case: Power Consumption

0.1 1 3 5 7 9 110

1020304050607080


Dat

agra

m L

oss

[%]

BS−Transmitter (QPSK 1/2)BS−Transmitter (16−QAM 1/2)

(d) WiMAX case: Datagram loss

0.1 1 3 5 7 9 11 1318.618.718.818.919.019.119.219.319.4


Pow

er C

onsu

mpt

ion

[W]

BS−Transmitter (16−QAM 3/4)BS−Transmitter (Auto−modulation)

(e) WiMAX case: Power Consumption

0.1 1 3 5 7 9 11 130

10

20

30

40

50

60


Dat

agra

m L

oss

[%]

BS−Transmitter (16−QAM 3/4)BS−Transmitter (Auto−modulation)

(f) WiMAX case: Datagram loss

Fig. 8. Average of power consumption at the BS (Transmitter) and datagram loss as a function of different traffic generation rate for different modulationtypes. Datagram size equal to 1280 bytes.


modulation scheme also when the channel conditions al-low use a higher modulation scheme.

In the case of WiFi scenario, in Fig. 9a and b, we plot theaverage power consumption at the AP when acting astransmitter and receiver as a function of different trafficgeneration rates for BPSK 3/4, QPSK 3/4, 16-QAM 3/4 and64-QAM 3/4 modulations. Datagram size is kept at1280 bytes. Fig. 9e and f reports the datagram loss whilethe throughput is reported in Fig. 9c and d.

As we can observe from Fig. 9a and c, the higher modu-lation and coding schemes are more efficient at using theavailable power and bandwidth than lower modulationand coding schemes when the AP is acting as transmitter.Fig. 9e and f shows the datagram loss for each modulationscheme. It is worth noticing that the lower modulationscheme (BPSK 3/4) has been tested using traffic generationdatarates up to 24 Mpbs in that higher generation rates re-sulted in either a packet loss of 100% or in misbehavior inthe traffic generator that was not able to maintain the

connection. This is due to buffers overflow at the transmit-ter which in time are caused by the low bitrate link.

In Fig. 9b, we can see that lower modulation and codingschemes are more energy efficient than higher modulationand coding schemes when the AP is acting as receiver.Although, the figure reports the power consumption vs.traffic generation rates at the transmitter side, the trafficreceived by the AP is lower as shown in Fig. 9d. Thus, itis difficult to evaluate the power consumed by each mod-ulation scheme at the receiver side as it depends on thetime that it spends in order to receive all the traffic sentby the transmitter. For this experiment, the datagram losswas lower than 0.2% as show in Fig. 9f.

4. Models

In this section, we aim at introducing a simple, yetaccurate model for estimating the power consumption of

1 4 7 10 14 18 24 34 44 544.7

4.8

4.9

5.0

5.1

5.2

5.3

5.4

Pow

er C

onsu

mpt

ion

[W]


AP−Transmitter (64−QAM 3/4)AP−Transmitter (16−QAM 3/4)AP−Tansmitter (BPSK 3/4)AP−Transmitter (QPSK 3/4)

(a) WiFi case: Power Consumption

1 4 7 10 14 18 24 34 44 544.7

4.8

4.9

5.0

5.1

5.2

Pow

er C

onsu

mpt

ion

[W]


AP−Rx (64−QAM 3/4)AP−Rx (16−QAM 3/4)AP−Rx (QPSK 3/4)AP−Rx (BPSK 3/4)

(b) WiFi case: Power Consumption

1 4 7 10 14 18 24 34 44 540

5

10

15

20

25

Thro

ughp

ut [M

b/s]


AP−Transmitter (BPSK 3/4)AP−Transmitter (QPSK 3/4)AP−Transmitter (16−QAM 3/4)AP−Transmitter (64−QAM 3/4)

(c) WiFi case: Throughput

1 4 7 10 14 18 24 34 44 54

5

10

15

20

25

30

Thro

ughp

ut [M

b/s]


AP−Receiver (BPSK 3/4)AP−Receiver (QPSK 3/4)AP−Receiver (16−QAM 3/4)AP−Receiver (64−QAM 3/4)

(d) WiFi case: Throughput

1 4 7 10 14 18 24 34 44 54

10

20

30

40

50

60

70

80

Dat

agra

m lo

ss [%

]


AP−Transmitter (BPSK 3/4)AP−Transmitter (QPSK 3/4)AP−Transmitter (16−QAM 3/4) AP−Transmitter (64−QAM 3/4)

(e) WiFi case: Datagram loss

1 4 7 10 14 18 24 34 44 540

0.05

0.1

0.15

0.2

Dat

agra

m lo

ss [%

]


AP−Receiver (BPSK 3/4)AP−Receiver (QPSK 3/4)AP−Receiver (16−QAM 3/4)AP−Receiver (64−QAM 3/4)

(f) WiFi case: Datagram loss

Fig. 9. Average of power consumption at the AP and datagram loss and throughput as a function of different traffic generation rate for different modulationtypes. Datagram size equal to 1280 bytes.

4 The RMSE is a numerical indicator of the differences between valuespredicted by a model and the values actually observed The RMSE is givenby RMSE ¼ 1

N

PNi¼1ðyi � yiÞ2; where yi is the measured value, yi is the

modeled value and N is the total number of samples.


wireless access network gateways as a function of the traf-fic rate and of the datagram size. Throughout the section,we use the following notation:

� x is the amount of traffic transmitted or received by theBS/AP (in Mb/s).� s is the datagram size (expressed in bytes).� f�(x), where � = Tx, Rx is the power consumption model

when BS/AP is acting as transmitter (respectively:receiver).

4.1. Measurements-based modelling

We used a curve fitting approach in order to construct amodel able to match well the power consumption mea-surements presented in the previous section. Different

types of functions were tested, including polynomialsup to the 4th order, mixes of exponential/logarithmicfunctions and piecewise linear functions. The metric usedto assess the goodness of the fit was the standardroot-mean-square error (RMSE).4

For all cases considered, the best fit was given a simplemodel in which the power consumption was linear as afunction of the traffic rate (respectively: datagram size)until a given threshold was reached. After such a threshold,the power consumption was constant regardless of thetraffic rate (respectively: datagram size). The slope of the


curve and the threshold value depend on a number of fac-tors, including technology considered (WiFi/WiMAX) andthe modulation scheme in use.

More formally, the model takes the following form:

f�ðxÞ ¼aðsÞ � xþ b if 0 6 x 6 hðsÞ Mb=s;c if x > hðsÞ Mb=s;

�

f�ðsÞ ¼�dðxÞ � sþ �ðxÞ if p 6 s 6 q byte;gðxÞ if s > q byte;

�8>>><>>>:

ð1Þ

It is worth remarking that the model for the dependencyon x has parameters that in turn depend on s and viceversa.

The parameters of the model have the following physi-cal meaning:

� a(s) (lJ/b) is the amount or energy spent by the wirelessdevice in order to transmit or receive 1 bit of informa-tion from the session layer with a datagram size of sbytes;� b (W) is the amount of power consumed by the wireless

gateway in idle mode;� c (W) is the maximum amount of power consumed by

the wireless gateway and represents the saturationpower consumption;� d(x) (lW/bytes) is the amount of power consumed by

wireless gateway in order to transmit or receive 1 byteof information from the session layer at a rate of x Mb/s;� �(x) (W) is the maximum power consumed by the wire-

less gateway, transmitting at x Mb/s, using extremelysmall packets.� g(x) (W) is the minimum power consumed by the wire-

less gateway to transmit traffic at rate x Mb/s.

4.2. Model validation

We first focused on the power consumption model as afunction of the traffic generation rate. In Fig. 10 we reportthe model for the power consumption at the BS and at theAP vs. different traffic generation rates for a constant data-gram size of 1280 bytes. The model parameters have beenoptimized using Matlab in order to minimize the RMSE.The best fit models obtained are:

WiMAXfTxðxÞ ¼

0:174xþ16:038 if 06 x6 14 Mb=s;18:5900 if 14< x6 30 Mb=s;

�

fRxðxÞ ¼ 0:00077xþ16 if 06 x6 30 Mb=s;fwith s¼ 1280 bytes;

8>>><>>>:

ð2Þ

WiFi

fTxðxÞ¼0:024xþ4:652 if 06 x632 Mb=s;5:396 if 32< x654 Mb=s;

�

fRxðxÞ¼0:016xþ4:677 if 06 x632 Mb=s;5:193 if 32< x654 Mb=s;

�

with s¼1280 bytes;

8>>>>><>>>>>:

ð3Þ

The RMSE figures obtained are 9.3968 � 10�4 (WiMAX, Tx),1.7139 � 10�4 (WiMAX, Rx), 5.7165 � 10�4 (WiFi, Tx),1.234 � 10�4 (WiFi, Tx), which confirm the ability of ourmodel to estimate the actual power consumption in avariety of settings.

Similarly, in Fig. 11 we report the model of the powerconsumption at the BS and at the AP vs. different datagramsizes for a constant traffic generation rate of 1 Mb/s. Themodel parameters have been optimized using Matlab inorder to minimize the RMSE. The best fit models obtainedare:

WiMAXfTxðsÞ¼

�0:0022sþ16:5655 if 326 s6128 byte;16:278 if 128< s62816 byte;

�

fRxðsÞ¼ 16:0469 if 326 s62816 byte;fwith x¼1Mb=s;

8>>><>>>:

ð4Þ

WiFi

fTxðsÞ¼�0:0004sþ4:841 if 326 s6256 byte;4:718 if 256< s62816 byte;

�

fRxðsÞ¼�0:0005sþ4:852 if 326 s6256 byte;4:715 if 256< s62816 byte;

�

with x¼1Mb=s;

8>>>>><>>>>>:

ð5Þ

The RMSE figures obtained are 5.9411 � 10�4 (WiMAX, Tx),1.5548 � 10�4 (WiMAX, Rx), 5.6123 � 10�5 (WiFi, Tx),1.5638 � 10�4 (WiFi, Tx), which again confirm the abilityof our model to estimate the actual power consumption.

We tested the obtained power consumption model withvarious modulation schemes. Results are reported inFig. 12 for a constant datagram size of 1280 bytes. Theparameters used in the model (again derived using Matlab)are reported in Tables 2 and 3 for WiMAX and WiFi accessgateways, respectively.

Finally, in order to validate our power consumptionmodel in the presence of the bi-directional traffic we haveloaded the WiFi AP with the traffic showed in Fig. 13a andmeasured the power consumption of the AP. Fig. 13ashows the amount of traffic that AP is transmitting andreceiving vs. time. It is important to notice that in orderto use the power consumption model described in (1)when the WiFi or the WiMax access gateway is acting astransmitter and receiver at the same time, b must be con-sidered only once. In other words, the total power con-sumption fTx+Rx is given by fTx+Rx = fTx + fRx � b, where fTx

and fRx are given by (1).The model output vs. the AP power consumption is re-

ported in Fig. 13b for a datagram size equal to 1280 bytes.As it can be seen, our model is capable of qualitatively fol-lowing the actual power consumption of the AP in thepresence of bi-directional traffic. However, due to thepower meter’s low resolution (0.1 W) it was not possibleto perform a better validation of our model in this scenario,that, we recall, aims at assessing the capability of our mod-el to predict the instantaneous power consumption of awireless gateway.

5. Related work

Real-world energy consumption measurements of wire-less networking devices have not often been performed inthe past. This, in turn, led to unrealistic and/or over-simpli-fied models being used in simulations.

In Ref. [14], the authors present several measurementsfor an IEEE 802.11a-based wireless network interfaceoperating in idle, sleep, receive and transmit modes. Such

0.11 2 3 4 5 6 7 8 9 10 12 14 16 18 20 22 24 26 28 3015.8

16.2

16.6

17.0

17.4

17.8

18.2

18.6


Pow

er C

onsu

mpt

ion

[W]

fRx(x) = 0.00077x + 16

fTx(x) = 0.174x+16.038 if 0<x<14fTx(x) = 18.5900 if 14<x<30

BS−Transmitter (Tx)Fitting curve fTx(x)

BS−Receiver (Rx)Fitting curve fRx(x)

(a) WiMAX case: Transmitter/Receiver

12345678910 12 14 16 18 20 24 26 28 32 36 40 44 48 52 544.7

4.8

4.9

5

5.1

5.2

5.3

5.4

Pow

er C

onsu

mpt

ion

[W]


fTx(x)=0.024x+4.652 if 0<x<32, fTx(x)=5.396 if 32<x<54

fRx(x)=0.016x+4.677 if 0<x<32, fRx(x)=5.193 if 32<x<54

AP−Transmitter (Tx)Fitting curve fTx(x)

AP−Receiver (Rx)Fitting curve fRx(x)

(b) WiFi case: Transmitter/Receiver

Fig. 10. Fitted curve of the power consumption at the BS and at the AP with a datagram size equal to 1280 bytes


measurements are obtained using an oscilloscope. In thework, the per-packet energy consumption E is approxi-mated using a linear model E = mS + b, where S is thelength of the packet and the values of the linear coeffi-cients m and b must be determined experimentally for var-ious operation modes. The authors conclude that theenergy consumption of an IEEE 802.11a wireless interfacehas a complex range of behaviours according to severalfactors such as relative proportions of broadcast andpoint-to-point traffic, packet size and reliance on promis-cuous mode operations. The behaviour of power consump-tion as a function of packet size is shown to follow a linearbehaviour. However, the model does not consider the caseof link-layer fragmentation and the impact of the differentamount of traffic on power consumption figures.

Similar work in terms of methodology is presented inRef. [15]. The paper presents several results for power con-sumption of an IEEE 802.11b wireless network interface.The scenario was built using two laptops with WLAN inter-faces and an oscilloscope in order to monitor powerconsumption on the wireless interface. The authors donot provide a power consumption model but they derivea novel metric from their measurements. The metric isthe energy per bit goodput Ebitgood given by Ebitgood[J/Bit] =AverageCP[W]/Goodput[Bit/s]. In order to determine Ebit good,

the (MAC) goodput must be recorded simultaneously withthe power consumption measurements (Average

CP). The

Ebitgood indicates the amount of energy expenditure in orderto transmit one bit of payload data successfully. Theauthors also conclude that large packets use energy moreefficiently than small ones as well as high modulationschemes are more energy efficient. In contrast to our work,the author’s experimental focus is based only on varyingthe packet size in order to investigate the power consump-tion behaviour while the impact of traffic on power con-sumption figures is not considered.

In Ref. [16], the authors focus their analysis on the novelIEEE 802.11n standard using a wide range of experiments.Each experiment aimed at assessing the impact that a cer-tain feature (e.g. channel width, transmission power, mod-ulation and coding scheme, etc.) has on the global energyconsumption figures. The testbed used for the power con-sumption measurements was composed of two nodesplaced close to each other in order to have good link qual-ity, which in time allowed the authors to effectively exploitall the IEEE 802.11n modulation and coding schemes. Tomeasure the power consumption, the authors placed aresistor (40X) on the 3.3 V power supply to the wirelessinterface. A National Instruments 6218 Data AcquisitionModule (NIDAM) was used in order to measure and record

64 256 512 768 1024 1280 1536 1792 2048 2304 2560 281615.9

16.1

16.3

16.5

16.7


Pow

er C

onsu

mpt

ion

[W] BS−Transmitter 1Mp/s (Tx)

Fitting curve fTx(s)

BS−Receiver 1Mp/s (Rx)Fitting curve fRx(s)

(a) WiMAX case

64 256 512 768 1024 1280 1536 1792 2048 2304 2560 2816

4.7

4.8

4.9

Pow

er C

onsu

mpt

ion

[W]


AP−Transmitter 1Mb/s (Tx)Fitting curve fTx(s)

AP−Receiver 1Mb/s (Rx)Fitting curve fRx(s)

(b) WiFi case

Fig. 11. Fitted curve of the power consumption at the BS and at the AP for a constant traffic rate of 1 Mb/s


the voltage drop across the resistor. Thus, the power con-sumed by the wireless interface was calculated using thedata recorded with NIDAM. The main conclusions of thiswork are that (i) for optimizing energy consumption, it isimperative to use the fastest single-stream rate possible,especially for shorter packets and (ii) the optimal devicesettings will also depend on channel conditions and work-load. The authors also observed that transmit power levelshave very little effect on the power consumed by theinterface.

In Ref. [17], the authors present a power consumptionmodel for IEEE 802.11g WLANs exploiting the power sav-ing mode. The authors also show the power consumptionmodel accuracy w.r.t. physical data measured from threepopular mobile platforms, namely Maemo, Android andSymbian. The model aims at estimating the energy usagebased on the flow characteristics which are easily availableon all the platforms without modifications to low-levelsoftware components or hardware. The authors concludethat energy is wasted by the idle status between packetintervals, in line with our results.

A theoretical energy consumption model is also pre-sented in Ref. [18]. In their work, the authors aims atassessing the amount of energy spent by an IEEE 802.11station in order to transmit 1 MB of data. In contrast toour work, the scenario considered is based on an IEEE802.11 network with N stations rather than 1 station.

The authors assume that the power consumption dependsof five physical radio states which are transmit, receive,listen, sleep, and off. The resulting model estimates thetotal energy (in Joules) consumed by a station in orderto transmit/receive 1 MB of data based on (i) the energyconsumed by the station in a specific state and (ii) the en-ergy consumed by the station for transmitting/receiving1 MB. Based on calculations, the author concludes thatenergy usage for the station grows approximately linearlywith N and as N increases, the energy wastage also in-creases. Such behaviour is mainly due to passive over-hearing of packets intended for another station. Thepower consumption is analysed only for a fixed amountof traffic.

The results reported by the authors in Refs. [14–18] pro-vides us with insights on the power consumption of thesingle wireless interface rather than of the system as awhole, ignoring the energy expenditures related to otherfunctions such as operations for packet forwarding andreception, fragmentation and reassembling etc. On theother hand, in this work, we focus our attention on theoverall energy expenditure in that it is useful to (i) modelthe real power consumption behaviour for WWAN andWLAN gateways and (ii) determine where and how the en-ergy is wasted in wireless access network gateways.

In Ref. [19], the authors focus on the energy consump-tion of a wireless network as a whole. The authors present

0.1 1 3 5 7 9 1116.016.316.616.917.217.517.818.118.418.719.0


Pow

er C

onsu

mpt

ion

[W]

BS−Transmitter (QPSK 1/2)Fitting curve fQPSK 1/2(x)

BS−Transmitter (16−QAM 1/2)Fitting curve f16−QAM 1/2(x)

(a) WiMAX: Transmitter

1 4 7 10 14 18 24 34 44 544.7

4.8

4.9

5.0

5.1

5.2

5.3

5.4

Pow

er C

onsu

mpt

ion

[W]


AP−Transmitter (BPSK 3/4)Fitting curve fBPSK(x)

AP−Transmitter (QPSK 3/4)Fitting curve fQPSK(x)

AP−Transmitter (16−QAM 3/4)Fitting curve f16−QAM(x)

AP−Transmitter (64−QAM 3/4)Fitting curve f64−QAM(x)

(b) WiFi: Transmitter

1 4 7 10 14 18 24 34 44 544.7

4.8

4.9

5.0

5.1

5.2

Pow

er C

onsu

mpt

ion

[W]


AP−Rx (BPSK 3/4)Fitting curve fBPSK(x)

AP−Rx (QPSK 3/4)Fitting curve fQPSK(x)

AP−Rx (16−QAM 3/4)Fitting curve fQAM(x)

AP−Rx (64−QAM 3/4)Fitting curve fQAM(x)

(c) WiFi: Receiver

Fig. 12. Fitted curve of the power consumption at the AP for different modulation/coding schemes. Datagram size equal to 1280 bytes.

Table 2Power consumption model parameters for the WiMAX access gateway using a specific modulation/coding schemes (s = 1280 bytes).

Modulation/coding a (lJi/b) b (W) c (W) h (Mb/s) RMSE (W)

Tx: QPSK 1/2 0.761 16.036 18.748 3 0.002Tx: 16QAM 1/2 0.394 16.132 18.918 7 0.0017

Rx 0 16 16 11 1.7139 � 10�4


a joint experimental evaluation of energy consumptionand performance in a IEEE 802.11-based WLAN using both802.11a and 802.11n operating modes. The testbedconsisted of an AP communicating with a single station.The power consumption measurements are taken using asuitable power meter and traffic is injected using the iperf

traffic generator. The authors have exploited an applica-tion-level approach, varying the packet size and transmis-sion rate and evaluating the energy consumption across awide transmission rates. They also perform a comparisonof the energy consumed by popular Internet applicationssuch as YouTube and Skype. A metric for energy usage

Table 3Power consumption model parameters for the WiFi access gateway using a specific modulation schemes (s = 1280 bytes).

Modulation/coding a (lJi/b) b (W) c (W) h (Mb/s) RMSE (W)

Tx: BPSK 3/4 0.089 4.667 5.292 7 0.0012Tx: QPSK 3/4 0.045 4.705 5.360 14 0.0011Tx: 16-QAM 3/4 0.025 4.697 5.329 24 0.0014Tx64-QAM 3/4 0.019 4.695 5.374 34 0.0012

Rx: BPSK 3/4 0.0035 4.711 4.748 10 5.009 � 10�6

Rx: QPSK 3/4 0.0065 4.702 4.796 14 4.538 � 10�5

Rx: 16-QAM 3/4 0.0154 4.6933 5.0389 24 9.7329 � 10�5

Rx: 64-QAM 3/4 0.015 4.700 5.1886 34 1.2910 � 10�4

0 50 100 150 200 250 300 350 400 450 500 550 600 650 7000

2

4

6

8

10

12

14

Time [s]

Traf

fic G

ener

atio

n ra

te [M

b/s]

TransmissionReception

(a) Traffic

0 50 100 150 200 250 300 350 400 450 500 550 600 650 7004.6

4.7

4.8

4.9

5.0

5.1

Time [s]

Pow

er C

onsu

mpt

ion

[W] real measurement

model

(b) Power Consumption

Fig. 13. Comparison between real and modeled values of the power consumption at the AP vs. time in the presence of the bi-directional traffic. Datagramsize equal to 1280 bytes


namely Effective Application-specific energy-usage (EA)was defined, as follows: EA = (mean power used duringtransmission of flow)[J]/(mean throughput of flow)[Mb].The authors also observed that both the application’stransmission rate and the packet size have an impact onpower consumption when the device is acting as transmit-ter. In contrast to our work, the case when the device isacting as receiver in not considered, and no power con-sumption model is provided.

In Ref. [20], the authors present a power consumptionmodel for wireless access networks and, in particular, formobile WiMAX, HSPA and LTE networks. The scenario isa suburban area and a physical bit rate of 10 Mb/s is used.The authors compare wireless technologies for one SISOsystem and three MIMO systems considering a ranking ofthe wireless technologies as a function of their power con-sumption, range and energy efficiency. To compare the dif-ferent technologies, the authors define models in order to

calculate the total power consumption per user and thepower consumption of the base station. In contrast to ourwork, the model provided by the authors is based in theamount of power consumed by each part of the systemseparately rather than the power consumption of the sys-tem as a whole. The authors also present relevant analysisto (i) determine which technology is the best solution forthe specified area and (ii) compare the power consumptionof the wireless access networks with the power consump-tion of the wired access networks. The most relevant con-clusion of the paper is that with a pre-defined bit rate of10 Mb/s, the mobile WiMAX is the most energy-efficientsolution compared with HSPA and LTE. However, theinvestigation is limited to a fixed amount of traffic conse-quently the scenarios of low traffic load and very high traf-fic load are not considered leaving the question ‘‘how thetraffic affects the power consumption of the WiMAX, HSPAand LTE devices’’ unanswered.


6. Conclusions and future work

In this work, we proposed a measurement-based meth-odology for characterizing the energy consumption behav-iour of networked wireless devices. In particular, wefocused our attention on WiFi and WiMAX access gate-ways and we derived their energy consumption figuresas a function of (i) the traffic load, (ii) the modulationand coding schemes and (iii) the size of the datagramsused.

It is the authors’ standpoint that a simple and accuratepower consumption model, that can be easily plugged intotypical network simulations tools such as ns3 or Omnet++,is essential to drive the design and development of energyaware network protocols and algorithms. Such an ap-proach will pave the way to an energy proportional net-working paradigm where wireless networks are designedin order to provide coverage and capacity but only whenand where needed. Consequently, it is imperative to refor-mulate wireless communication system design in terms ofenergy efficiency by taking into consideration that ‘‘alwaysavailable’’ does not mean ‘‘always on’’.

The main observations of this work are:

� The energy consumption of the wireless access networkgateways depends on several factors, such as theamount of traffic, packet size, transmission power level,modulation and coding schemes and channelconditions.� Large packets use energy more efficiently than small

ones as well as high modulation and coding schemesare more energy efficient. Additionally, the transmitpower levels have very little effect on the power con-sumed by the wireless gateways.� The power consumption follows a linear behaviour until

the gateway reaches a saturation point.� The energy consumed when the gateways are in idle

mode is approximately 80% of the maximum powerconsumption in saturation regime, in both the WiFiand WiMAX cases.

As it can be seen from the results, the challenges inwireless networking in terms of energy efficiency are (i)improving the energy efficiency of the wireless access net-work gateway when it is transmitting data as well as whenit is in idle mode and (ii) introduce novel indicator of en-ergy efficiency as part of the standard of each wirelesstechnologies.

We are currently analysing the effects of traffic onpower consumption in multi-hop wireless networks withmultiple clients considering real application scenariosand traffic classes. We are also working toward a general-ized model capable of taking into account traffic load anddatagram size. Such a model shall be able to provide amore precise prediction of the power consumed by a wire-less gateway in realistic settings where multiple clients aretransmitting and receiving different types of traffic.

Finally, we are currently developing a custom hardwareand software solution for power consumption monitoringcapable of addressing the limitations of the approach used

in this work, namely low sampling rate (1 sample per sec-ond) and low resolution (0.1 W). The new solution is ex-pected to deliver increased insight into the behavior insaturation (higher resolution) and in the transition be-tween linear and saturation (higher sampling rate).

References

[1] K.M. Gomez, R. Riggio, T. Rasheed, D. Miorandi, I. Chlamtac, F.Granelli, Analysing the energy consumption behaviour of WiFinetworks, in: Proc. of the IEEE GreenCom, 2011.

[2] A.P. Jardosh, K. Papagiannaki, E.M. Belding, K.C. Almeroth, G.Iannaccone, B. Vinnakota, Green WLANs: on-demand WLANinfrastructures, Mob. Netw. Appl. 14 (2009) 798–814.

[3] Y. Al-Hazmi, H. De Meer, K.A. Hummel, H. Meyer, M. Meo, D.Remondo, Energy-efficient wireless mesh infrastructures, IEEENetwork Mag. 25 (2011) 32–36.

[4] S. Nedevschi, L. Popa, G. Iannaccone, S. Ratnasamy, D. Wetherall,Reducing network energy consumption via sleeping and rate-adaptation, in: Proc. of the NSDI, Berkeley, CA, USA, 2008.

[5] K. Pentikousis, In search of energy-efficient mobile networking, IEEECommun. Mag. 48 (1) (2010) 95–103.

[6] Green Touch Initiative. <http://www.greentouch.org>.[7] European Union Funded Projects’ Information in Green Networks.

<http://cordis.europa.eu/fp7/ict/sustainable-growth/>.[8] M. Marsan, L. Chiaraviglio, D. Ciullo, M. Meo, Optimal energy savings

in cellular access networks, in: IEEE International Conference onCommunications Workshops, 2009, pp. 1–5.

[9] K. Dufkova and, M. Bjelica, B. Moon, L. Kencl, J.-Y. Le Boudec, Energysavings for cellular network with evaluation of impact on data trafficperformance, in: Wireless Conference (EW), 2010 European, 2010,pp. 916–923.

[10] H.-O. Scheck, ICT & wireless networks and their impact on globalwarming, in: Proc. of the European Wireless Conference, Lucca, Italy,2010, pp. 911–915.

[11] O. Blume, D. Zeller, U. Barth, Approaches to energy efficient wirelessaccess networks, in: Proc. of the 4th International Symposium onCommunications, ISCCSP 2010, Control and Signal Processing,Limasso1, Cyprus, 2010, pp. 1–5.

[12] J. Lorincz, A. Capone, M. Bogarelli, Energy savings in wireless accessnetworks through optimized network management, in: Proc. of the5th IEEE International Conference on Wireless Pervasive Computing,ISWPC’10, IEEE Press, Mondena, Italy, 2010, pp. 449–454.

[13] D. Kilper, G. Atkinson, S. Korotky, S. Goyal, P. Vetter, D. Suvakovic, O.Blume, Power trends in communication networks, IEEE J. Sel. TopicsQuant. Electron. 17 (2) (2011) 275–284.

[14] L.M. Feeney, M. Nilsson, Investigating the energy consumption of awireless network interface in an ad hoc networking environment,in: Proc. of the IEEE INFOCOM, Anchorage, USA, 2001, pp. 1548–1557.

[15] J. Ebert, B. Burns, A. Wolisz, A trace-based approach for determiningthe energy consumption of a WLAN network interface, in: Proc. ofthe European Wireless Conference, 2002, pp. 230–236.

[16] D. Halperin, B. Greenstein, A. Sheth, D. Wetherall, Demystifying802.11n power consumption, in: Proc. of the 2010 InternationalConference on Power Aware Computing and Systems, HotPower’10,USENIX Association, Berkeley, CA, USA, 2010, pp. 1–5. <http://dl.acm.org/citation.cfm?id=1924920.1924928>.

[17] Y. Xiao, P. Savolainen, A. Karppanen, M. Siekkinen, A. Ylä-Jääski,Practical power modeling of data transmission over 802.11g forwireless applications, in: Proc. of the 1st International Conference onEnergy-Efficient Computing and Networking, e-Energy ’10, ACM,New York, NY, USA, 2010, pp. 75–84.

[18] M. Ergen, P. Varaiya, Decomposition of energy consumption inIEEE 802.11, in: Proc. of IEEE ICC, Glasgow, Scotland, 2007, pp. 403–408.

[19] Y.Y.M. Tauber, S.N. Bhatti, Application level energy and performancemeasurements in a wireless LAN, in: Proc. of the GreenCom2011 –IEEE/ACM 2011 International Conference on Green Computing andCommunications, Chengdu, Sichuan, China, 2011.

[20] M. Deruyck, W. Vereecken, E. Tanghe, W. Joseph, M. Pickavet, L.Martens, P. Demeester, Comparison of power consumption of mobileWiMAX, HSPA and LTE access networks, in: 9th Conference on Proc.of the Telecommunications Internet and Media Techno Economics(CTTE 2010), 2010, pp. 1 –7.

.

http://www.greentouch.org

http://cordis.europa.eu/fp7/ict/sustainable-growth/

http://dl.acm.org/citation.cfm?id=1924920.1924928

http://dl.acm.org/citation.cfm?id=1924920.1924928


Karina Gomez received the engineeringdegree (cum laude) in Electronics and Tele-communication Engineering from theNational Polytechnic School in Ecuador, in2006. She received her Masters degree inWireless Systems and Related Technologiesfrom the Turin Polytechnic, Italy, during 2007.In 2007, she joined FIAT Research Center,becoming part of the Infomobility-Communi-cation and location Technologies. Since July2008, she is part of the iNSPIRE Area at Create-Net, working on the several research projects

in the wireless communication and networking area. She is also a PhDcandidate at University of Trento, in Italy. Her current research activity ismainly focused on Green Networking. Other research interests include:
wireless mesh networks, sensor and ad hoc networks, Performanceevaluation (Omnet++, Matlab).
Dejene Boru was born in Mida Kegn, Ethiopia.He received his B.Sc degree in Electrical andComputer Engineering from Addis AbabaInstitute of Technology in 2008. In year 2008he joined ZTE Corporation as Fixed LineEngineer. From November 2008 to October2009 he was Graduate Assistant in ElectricalEngineering Department at Dire Dawa Uni-versity, Ethiopia. Currently he is workingtoward his M.Sc in TelecommunicationsEngineering at University of Trento, Italy. Hisresearch interests are in the areas of Green

networking and Cloud Computing.

Roberto Riggio is currently Senior Researcherstaff member in the iNSPIRE team at CREATE-NET, Italy. He received the B.S. degree and theM.S. degrees (summa cum laude) inTelecommunication Engineering from theUniversity of Reggio Calabria, Italy, in,respectively, 2001 and 2004. He received thePh.D. degree in Telecommunication sciencefrom the University of Trento, Italy, in 2008,with a thesis entitled ‘‘Cross-layer Design inWireless Mesh Networks’’. From May 2007 toNovember 2007 he was a visiting researcher

at the Wireless Networks Laboratory (WINET) at the University of Florida.His main research interests lie in field of wireless mesh networks with aparticular design and implementation of wireless mesh networking
solutions, autonomous network monitoring, and network virtualization.He co-authored several peer reviewed research papers in internationalconferences and journals. He served as TPC member for the MediaWiNconferences, and as a peer reviewer for the Elsevier Computer Networks,Springer Wireless Networks, Springer Mobile Networks and Applications,and IEEE Transaction on Vehicular Technology journals.
Tinku Rasheed received the B.E. degree inelectronics and communications from KeralaUniversity, India, in 2002, the M.S. degree intelecommunications from Aston University,Birmingham, UK, in 2004, and the Ph.D.degree in computer science from the Univer-sity of Paris-Sud XI, Orsay, France, in 2007.From May 2003 until December 2006, he waswith Orange Labs, Lannion, France, where hewas extensively involved in the developmentof protocols and quality-of-service provision-ing schemes for mobile access networks and

contributed to several national and internal projects and IEEE standard-ization groups. Currently, he is a Senior research staff member with

Create-Net Research Center, Trento, Italy, where he is involved in industryand European-funded projects focusing on next generation light infra-structure wireless access solutions. He has served as the Co-Editor of aSpecial Issue of the Springer Mobile Networks and Applications Journaland as Guest Editor for IEEE Networks Magazine. He is the holder ofseveral patents.

Daniele Miorandi is the head of the iNSPIREArea at CREATE-NET, Italy. He received a Ph.D.in Communications Engineering from Univ. ofPadova, Italy, in 2005, and a Laurea degree(summa cum lauda) in CommunicationsEngineering from Univ. of Padova, Italy, in2001. He joined CREATE-NET in January 2005,where he is leading the iNSPIRE (Networkingand Security Solutions for Pervasive Comput-ing Systems: Research & Experimentation).His research interests include bio-inspiredapproaches to networking and service provi-

sioning in large-scale computing systems, modelling and performanceevaluation of wireless networks, prototyping of wireless mesh solutions.He has co-authored more than 100 papers in internationally refereed
journals and conferences. He serves on the Steering Committee of variousinternational events (WiOpt, Autonomics, ValueTools), for some of whichhe was a co-founder (Autonomics and ValueTools). He also serves on theTPC of leading conferences in the networking field, including, e.g. IEEEINFOCOM, IEEE ICC, IEEE Globecom. He is a member of IEEE, ACM andICST.
Fabrizio Granelli received the ‘‘Laurea’’(M.Sc.) degree in Electronic Engineering fromthe University of Genoa, Italy, in 1997, with athesis on video coding, awarded with theTELECOM Italy prize, and the Ph.D. in Tele-communications from the same university, in2001. Since 2000 he is carrying on hisactivity as Assistant Professor in Telecom-munications at the Dept. of InformationEngineering and Computer Science – Uni-versity of Trento (Italy) and coordinator ofthe Networking Laboratory. In August 2004

and August 2010, he was visiting professor at the State University ofCampinas (Brasil). He is author or co-author of more than 100 paperspublished in international journals, books and conferences, and he is
member of the Technical Committee of the International Conference onCommunications (from 2003 to 2007) and Global TelecommunicationsConference (GLOBECOM2003 and GLOBECOM2004). He is guest-editorof ACM Journal on Mobile Networks and Applications, special issues on‘‘WLAN Optimization at the MAC and Network Levels’’ and ‘‘Ultra-WideBand for Sensor Networks’’, and Co-Chair of 10th IEEE Workshop onComputer-Aided Modeling, Analysis, and Design of CommunicationLinks and Networks (CAMAD’04). He is General Vice-Chair of the FirstInternational Conference on Wireless Internet (WICON’05) and GeneralChair of the 11th and 15th IEEE Workshop on Computer-Aided Model-ing, Analysis, and Design of Communication Links and Networks(CAMAD’06 and CAMAD’10). He is Co-Chair of GLOBECOM 2007–2009Symposia on ‘‘Communications QoS, Reliability and Performance Mod-eling’’. His main research activities are in the field of networking andsignal processing, with particular reference to network performancemodeling, cross-layering, medium access control, wireless networks,cognitive radio systems, and video transmission over packet networks.He is Senior Member of IEEE, Chair of IEEE ComSoc Technical Committeeon Communication Systems Integration and Modeling (CSIM) andAssociate Editor of IEEE Communications Letters, IEEE CommunicationsSurveys and Tutorials, International Journal on Communication Systems,Journal of Wireless Communications and Networking.

mimo systems

Documents

wimax gateways

total energy

network operators

access infrastructure

global energy consumption

ict system

power grid

wimax standards