6
Energy Consumption Model for Mobile Devices in 3G and WLAN Networks Erkki Harjula, Otso Kassinen and Mika Ylianttila Department of Computer Science and Engineering, Erkki Koiso-Kanttilan katu 3, FIN-90014 University of Oulu, Finland [email protected] Abstract—In this paper, we propose an advanced model, called e-Aware, for estimating how application layer protocol properties affect the energy consumption of mobile devices, operating in 3G (WCDMA) and WLAN (802.11) networks. The main motivation for the model is to facilitate designing energy- efficient networking solutions, by reducing the need for time- consuming measurements with real-life networks and devices. The model makes a distinction between signaling and media transfers due to their different energy consumption characteristics, and takes into account the fundamentals of radio interface properties, such as different energy states and timers controlling them. The model is fine-tuned using device- specific coefficients that are defined according to real-world measurements with actual devices. We have implemented the model and simulated it in Matlab environment. The correct functionality is verified by comparing the results with real-life measurements in identical networking scenarios. Keywords—mobile, energy consumption, estimation, simulation I. INTRODUCTION The use of smartphones and other mobile devices is growing rapidly, making them an integral part of today’s Internet paradigm. Battery duration is undoubtedly one of the most essential factors affecting the feasibility of mobile Internet communication. For long, one of the main challenges of mobile networking has been the fact that the battery technology does not improve at the same pace as the power requirements [1]. Thus, energy-efficient solutions are essential for keeping battery life on an acceptable level. With mobile devices, the network interface accounts for a major component of the total power consumption [2], emphasizing the importance of energy-efficient networking. In general, Internet traffic can be categorized into two basic types: signaling and media transfers. Signaling includes network transactions consisting of separately sent and received small messages, transmitted in single packets, such as maintenance messaging, instant messaging, and service lookups. Media transfers include network transactions consisting of series of packets, such as file transfers and media streaming. The above-mentioned traffic types have radically different energy consumption characteristics in wireless networks. Whereas the energy consumption of media transfers in wireless networks can be estimated quite accurately using the amount of moved data, frequent signaling consumes significant amounts of energy regardless of the relatively small amount of transferred data [3]. The latter is especially problematic in today’s peer-to- peer (P2P) networks, where a major part of the traffic consists of frequently received and sent small packets [4][5]. The reason for the energy-inefficiency of signaling in wireless networks is explained in [6] by so called ramp and tail energy components. The ramp and tail energy components are a consequence of the Radio Resource Control (RRC) protocol states in WCDMA 3G networks [7]. RRC includes three states: IDLE, PCH (alternative to IDLE), DCH (dedicated channel), and FACH (forward access channel). IDLE and PCH states provide low energy consumption for idle periods. DCH state gives maximum throughput with minimum delay at the cost of high energy consumption. In FACH, the energy consumption is reduced at the cost of lower throughput, when compared to DCH. Changing from idle state to either of the data transfer states requires a setup time. Ramp energy is the additional energy consumed during the setup time. After the completion of a data transfer, radio link consumes energy by remaining in a data transfer state for a while (defined by an inactivity timer, set by network operator), before moving to idle state. This additional energy is called tail energy. Tail and ramp components are not present with WLAN, but the association cost to access point is high [6]. As a consequence, the traditional energy consumption estimation models, based solely on the amount of moved data, are obsolete for today’s mobile networking. To address this problem, this paper proposes an advanced energy consumption model for mobile devices. The paper is organized as follows. Section 2 presents the related work, Section 3 introduces our e-Aware model, Section 4 evaluates the model, and Section 5 concludes the paper. II. RELATED WORK Several studies have proposed models for estimating the energy consumption of mobile networking. However, to our knowledge, our e-Aware is the first design-phase energy consumption estimation model taking into account the different energy consumption properties of signaling and media transfers. The related work, presented below, clearly proves the need for this kind of advanced energy consumption model. The energy consumption characteristics of mobile networks have been thoroughly studied during the past years. For example, Haverinen et al. [7] have analyzed how keepalive messages, required by e.g. Mobile IP and NAT traversal, affect the battery life of a mobile device in The 9th Annual IEEE Consumer Communications and Networking Conference - Emerging Consumer Technologies 978-1-4577-2071-0/12/$26.00 ©2012 IEEE 532

Energy Consumption Model for Mobile Devices ... - Front · PDF fileEnergy Consumption Model for Mobile Devices in 3G and WLAN Networks ... FACH fo DCH state when the packet size, [bytes]

Embed Size (px)

Citation preview

Energy Consumption Model for Mobile Devices

in 3G and WLAN Networks

Erkki Harjula, Otso Kassinen and Mika Ylianttila

Department of Computer Science and Engineering, Erkki Koiso-Kanttilan katu 3,

FIN-90014 University of Oulu, Finland

[email protected]

Abstract—In this paper, we propose an advanced model, called

e-Aware, for estimating how application layer protocol

properties affect the energy consumption of mobile devices,

operating in 3G (WCDMA) and WLAN (802.11) networks. The

main motivation for the model is to facilitate designing energy-

efficient networking solutions, by reducing the need for time-

consuming measurements with real-life networks and devices.

The model makes a distinction between signaling and media

transfers due to their different energy consumption

characteristics, and takes into account the fundamentals of

radio interface properties, such as different energy states and

timers controlling them. The model is fine-tuned using device-

specific coefficients that are defined according to real-world

measurements with actual devices. We have implemented the

model and simulated it in Matlab environment. The correct

functionality is verified by comparing the results with real-life

measurements in identical networking scenarios.

Keywords—mobile, energy consumption, estimation, simulation

I. INTRODUCTION

The use of smartphones and other mobile devices is growing rapidly, making them an integral part of today’s Internet paradigm. Battery duration is undoubtedly one of the most essential factors affecting the feasibility of mobile Internet communication. For long, one of the main challenges of mobile networking has been the fact that the battery technology does not improve at the same pace as the power requirements [1]. Thus, energy-efficient solutions are essential for keeping battery life on an acceptable level. With mobile devices, the network interface accounts for a major component of the total power consumption [2], emphasizing the importance of energy-efficient networking.

In general, Internet traffic can be categorized into two basic types: signaling and media transfers. Signaling includes network transactions consisting of separately sent and received small messages, transmitted in single packets, such as maintenance messaging, instant messaging, and service lookups. Media transfers include network transactions consisting of series of packets, such as file transfers and media streaming.

The above-mentioned traffic types have radically different energy consumption characteristics in wireless networks. Whereas the energy consumption of media transfers in wireless networks can be estimated quite accurately using the amount of moved data, frequent signaling consumes significant amounts of energy regardless of the relatively small amount of transferred data

[3]. The latter is especially problematic in today’s peer-to-peer (P2P) networks, where a major part of the traffic consists of frequently received and sent small packets [4][5].

The reason for the energy-inefficiency of signaling in wireless networks is explained in [6] by so called ramp and tail energy components. The ramp and tail energy components are a consequence of the Radio Resource Control (RRC) protocol states in WCDMA 3G networks [7]. RRC includes three states: IDLE, PCH (alternative to IDLE), DCH (dedicated channel), and FACH (forward access channel). IDLE and PCH states provide low energy consumption for idle periods. DCH state gives maximum throughput with minimum delay at the cost of high energy consumption. In FACH, the energy consumption is reduced at the cost of lower throughput, when compared to DCH. Changing from idle state to either of the data transfer states requires a setup time. Ramp energy is the additional energy consumed during the setup time. After the completion of a data transfer, radio link consumes energy by remaining in a data transfer state for a while (defined by an inactivity timer, set by network operator), before moving to idle state. This additional energy is called tail energy. Tail and ramp components are not present with WLAN, but the association cost to access point is high [6].

As a consequence, the traditional energy consumption estimation models, based solely on the amount of moved data, are obsolete for today’s mobile networking. To address this problem, this paper proposes an advanced energy consumption model for mobile devices. The paper is organized as follows. Section 2 presents the related work, Section 3 introduces our e-Aware model, Section 4 evaluates the model, and Section 5 concludes the paper.

II. RELATED WORK

Several studies have proposed models for estimating the energy consumption of mobile networking. However, to our knowledge, our e-Aware is the first design-phase energy consumption estimation model taking into account the different energy consumption properties of signaling and media transfers. The related work, presented below, clearly proves the need for this kind of advanced energy consumption model.

The energy consumption characteristics of mobile networks have been thoroughly studied during the past years. For example, Haverinen et al. [7] have analyzed how keepalive messages, required by e.g. Mobile IP and NAT traversal, affect the battery life of a mobile device in

The 9th Annual IEEE Consumer Communications and Networking Conference - Emerging Consumer Technologies

978-1-4577-2071-0/12/$26.00 ©2012 IEEE 532

WCDMA networks. The results indicate that the energy consumption is significantly influenced by the RRC parameters and the frequency of keepalive messages.

Balasubramanian et al. [6] present a measurement study of the energy consumption properties of 3G, GSM, and WLAN. They observe that 3G and GSM incur high tail energy overhead due to high power states after completing a transfer, being especially problematic in networks including frequent signaling, such as P2P networks. Kelenyi et al. [4][5] have studied the differences in energy consumption of mobile devices working either as a peers or clients in a structured P2P network, using both 802.11 and WCDMA networks. The studies conclude that the energy consumption is significantly higher in the peer mode when compared to the client mode, due to frequent maintenance signaling. Hence, it is essential that the energy consumption model considers signaling in estimating the energy consumption.

Perälä et al. [8] have studied the WCDMA RRC state transitions in practice. The results suggest that although the 3GPP specifications are followed, reliable prediction of the exact behavior of mobile networks beforehand, based on theories, is difficult. Hence, real-life measurements are essential in tuning the energy consumption model to reflect the real characteristics of a mobile network.

There are also existing models that estimate mobile device energy consumption. For example, Xiao et al. [9] have proposed and evaluated a power consumption model for 802.11g network. However, the study lacks generality since the results are limited to WLAN environment and the model is based solely on the amount of moved data. Zhang et al. [10] have introduced a technique for learning and modeling battery discharge curve of different hardware components of smartphones, in order to estimate the energy consumption of applications. The paper presents a tool for estimating the energy consumption during runtime, but does not provide a detailed off-the shelf model that could be used in design phase before implementing the application.

We have earlier studied the energy consumption of a smartphone in [12]. The article provides energy consumption profiles for Nokia N95 smartphone in 802.11b and WCDMA networks, using different packet transmission intervals and packet sizes as parameters. In this paper we use the result data in defining the parameters for our model.

III. ENERGY CONSUMPTION MODEL

Our energy consumption model, called e-Aware, is a mathematical model for estimating energy consumption originating from network operations, based on two energy consumption elements: signaling and media transfers. The distinction between signaling and media transfers is made due to their different energy consumption characteristics, as justified in the introduction. The model is illustrated in Fig. 1. The network parameters (blue arrows) include the average transmission interval and packet size for transferred data chunks smaller than 1.5kB (often used as the MTU with UDP), and the amount of moved data for data chunks greater than 1.5kB. The device-specific energy profile (green arrow) is based on real-life measurement data gathered by using a real mobile device in a real networking

environment. As output, e-Aware gives power and energy consumption estimates (red arrow) for the scenario. The model works with both static and dynamic (time-varying) parameters. Respectively, the estimates are returned as either static [W or J/s] or dynamic [Ws or J] values.

Figure 1. The operating principle of e-Aware model.

For accurate energy consumption modeling, we need to

understand the factors behind the realized consumption. The

total energy consumption of sending or receiving a data

chunk includes roughly; 1) the ramp energy, consumed

during the state transition from the idle state to the data

transfer state, 2) the transfer energy, consumed during the

transfer of the data chunk through the radio interface and the

associated computing activities, and 3) the tail energy,

consumed during the intermediate power states after the

transmission. With signaling in WCDMA networks, the

transfer energy is only a minor component of the total

energy consumption, as transferring a single packet usually

takes only some milliseconds, while the tail and ramp times

may range from seconds to tens of seconds, depending on

the network parameters. With media, the significance of tail

and ramp components is lower, since subsequential packet

transfers keep the radio continuously in the transfer mode.

The total power consumption at moment t is defined by

equation (1). The details for calculating the power

components are elaborated in the following subsections.

���� = max[��� ���, �������] (1)

A. Signaling

The signaling power consumption is represented by Psig [W]. The equations in this section are based on through analysis of radio channel characteristics [6][7][8][9]. The parameters in the equations are used for fine-tuning the model to correspond with real-life measurements as well as possible. In Fig. 2, we exemplify the power consumption of a mobile device periodically sending and receiving fixed-size packets, as a function of packet transmission interval τ [s]. For clarity, we use τ as the transfer interval symbol, since t is reserved as the simulation time unit symbol.

In [12], we noticed a significant difference between the power consumptions of signaling consisting of packets sized below 200 and above 300 bytes in WCDMA network. The explanation for this is that WCDMA triggers a change from FACH fo DCH state when the packet size, ��� [bytes] exceeds a certain threshold (���������� ), controlled by the network operator [8]. In [10], ���������� was observed to be 151 bytes for uplink and 119 bytes for downlink direction, whereas in [11], it was 250 bytes for both directions. As our

533

measurement data indicates that ���������� is between 200-300 bytes, we use 250 bytes as ���������� for both directions.

With WCDMA, there are three threshold values for packet transmission interval, τ1, τ2 and τ3 [s] (Fig. 2b). When the transmission interval τ is below τ1, the radio remains in sending or receiving mode, and the power consumption is close to the continuous media transfer consumption, and is only slightly affected by the transmission interval. As the power consumption seems to linearly decrease as the transmission interval grows, based on the results of [12], we use a linear function in equation (2) for calculating the power consumption, when � � �1. All the parameters used in equations (2-14) are explained later in Table I.

When ��� ! ���������� , FACH mode is used for

transmission and �" is skipped in the calculations. Fig. 2a illustrates this case. In this case, when �# ! � � �$ , power consumption asymptotically approaches the FACH power consumption (Pfach [W]) following a function of the form %#/� ' %" as τ grows (equation 3). Similarly, when τ gets above τ3, power consumption starts approaching the idle power consumption (Pidle [W]) asymptotically (equation 6).

When ��� ( ����������,DCH is used for transmission,

and the power consumption after �# is calculated using the equation (4). Respectively, equation (5) is used after �", and equation (6) after after �$ , as illustrated in Fig. 2b. The thresholds τ1, τ2 and τ3 are calculated using equations (7-9).

��� _��*��� = + , ��� - . , � ' /, � � �#, (2)

0012 ! 0345604789:

��� _:*;���� =<=<, >��� _��*��#� - �:*;�? ' �:*;�, �# ! � � �$(3)

0012 ( 0345604789:

��� _�;���� =<=<, >��� _��*��#� - ��;�? ' ��;�, �# ! � � �" (4)

��� _:*;� =�@�, >��� _�;���"� - �:*;�? ' �:*;�, �" ! � � �$ (5)

��� _������� =�A�, >��� _:*;���$� - �����? ' ����� , � B �$ (6)

�# = C , ��� ' ����DE (7)

�" = �# ' ��;� (if ��� ! ����������, ��;� = 0, ) (8)

�$ = �" ' �:*;� (9)

With 802.11, we use two threshold values, τ1 and τ2, (see Fig. 2c). The 802.11 standard does not use standby timers, but instead switches the radio to idle mode right after media transfer. However, according to our measurements, there seems to be a period after each networking transaction,

when the power consumption of the radio interface seems to highly fluctuate, being averagely on a higher level than the idle power consumption. We call this the overhead time tovh.

As with WCDMA, when � � �1 , the 802.11 radio is almost 100% of its time in sending or receiving mode, and the power consumption is close to the consumption of continuous media transfer, and is only slightly affected by the transmission interval. We use equation (10) for calculating the power consumption in this case. Respectively, as illustrated in Fig. 2c, when �# ! � � �",the power consumption approaches asymptotically the overhead power consumption (Povh [W]) as � grows (equation 11). Further, when � B �2 , the power consumption approaches the idle power consumption (Pidle [W]) (equation 12). The thresholds τ1 and τ2 are calculated using equations (13-14).

��� _��*��� = + , ��� - . , �' /, � � �# (10)

��� _�H���� =�=�, >��� _��*��#� - ��H�? ' ��H� , �# ! � � �"(11)

��� _������� =�@�, >��� _�H���"� - �����? ' ����� , � B �" (12)

�# = C , ��� ' ����DE (13)

�" = �# ' ��H� (14)

B. Media transfers

Modeling the power consumption of media transfers (Pmed) [W] is more straightforward. First, the file transfer time (ttra) [s] is calculated using the size of the moved data and the measured data transfer rate for the radio technology in use. The setup time ����DE is included to ttra. During ttra,

the power consumption is averagely Pul [W] for uploads or Pdl [W] for downloads, after which the power consumption returns to Pidle [W] through intermediate power states.

With WCDMA, transferring larger amounts of data is always made in DCH power state. Thus, after the completion of a transfer, the radio interface remains in DCH for tdch seconds, and power consumption is on level Pdch

[W]. Then the power state is switched to FACH, in which the radio interface stays for tfach seconds, while the power consumption is Pfach [W]. The WCDMA power states and triggers are illustrated in Fig. 3a, where a sample file is downloaded using WCDMA. The first 15 seconds should be ignored, as the download was made manually and thus the power consumption is higher due to additional CPU activity and illuminated LCD backlight. The peak of the consumption at the end of the file transfer is caused by CPU activity, which is related to writing the data in the file system. With 802.11, the power consumption is averagely at intermediate level, Povh, [W], for certain overhead time tovh

(a) (b) (c)

Figure 2. Example power consumption of signaling using WCDMA for (a) packets sized < streshold (250B in this case) , (b) packets (streshold and (c) 802.11.

534

(a)

(b)

Figure 3. Power consumption as the function of time during (a) a file

download in WCDMA, (b) a file upload in 802.11b network.

[s] after the transfer completion. The 802.11 power states and triggers are illustrated in Fig. 3b, where a sample file is uploaded. The first 15 seconds should be again ignored due to the extra power consumption caused by manual upload.

e-Aware also enables complicated scenarios with time-varying signaling, with concurrent and non-concurrent transfers of media, etc. As media transfers cause full utilization of the wireless network interface in the direction of the transfer, parallel media transfers (in the same direction) are simulated as sequential transfers (the latter transfer is delayed until the former transfer has completed), and signaling during transfers lengthen the transfer time. We implement concurrent media transfer and signaling by adding the amount of data moved by the signaling during the media transfer to the media chunk size. Let he following example clarify this principle: A transfer of a 10MB file over a 802.11b link with a practical speed of 500kB/s would take 20 seconds. If the signaling frequency were 20 per second (10 signals in one direction) and the packet size were 1.5kB, the file transfer time would be (10MB + 60s × 10signals/s × 1.5kB/signal) / (500kB/s) = 21.8s.

C. Device-specific energy profile

Our earlier work [12] provides energy consumption profiles of N95 smartphone for 802.11b and WCDMA network interfaces, using different packet sending/receiving intervals and packet sizes as parameters. In this paper, we used the result data in defining the parameters for signaling.

For defining the energy consumption profile for media transfers in N95, we had to make new measurements, as previous data was not available. We measured the energy consumption of uploading and downloading files using both WCDMA and 802.11b interfaces. 802.11b allowed data rates of 5Mb/s on average. With WCDMA, the used data

plan allowed the data rate of 384kb/s. For the tests, we needed a tool that can both upload and download files to and from a mobile node. As Fring, an interpersonal communication application, works both on mobile phones and PC computers, and provides a P2P file transfer function, we used it for tests. Fring is based on Mobile VoIP (mVoIP) protocol, using H.323 data channel for media transfers. The test setup consisted of a mobile device (N95) and a laptop PC with Ethernet connection to the university network. By using the PC with a high-speed Internet connection, we eliminated possible bottlenecks at that end. For measuring the power consumptions, we used the Nokia Energy Profiler software. The resulting coefficients are defined by using Matlab’s curve fitting function.

The coefficients and parameters for both signaling and media transfers are presented in Table I. Some of them are also illustrated in Figs. 2 and 3.

TABLE I. ENERGY CONSUMPTION MODEL PARAMETERS FOR N95

Param Description 802.11 WCDMA

A Coefficient for ��� in eq. (2) and (10). 0.025 0.02

B Coefficient for �in eq. (2) and (10). 0.333 0.092

C Constant coefficient in eq. (2) and (10). 1.09 1.235

D Coefficient for ��� in eq. (7) and (13). 0.044 1.15

t setup [s] Delay caused by the transition from idle

to transfer mode. 0.063 4/0.015

(DCH/FACH)

Pdch

[W]

Power consumption of a device when

the WCDMA radio interface is in DCH

mode but not sending or receiving data.

- 0.65

Pfach

[W]

Power consumption of a device when

the WCDMA radio interface is in FACH

mode but not sending or receiving data.

- 0.52

Pidle

[W]

Power consumption of a device when

the radio interface is in IDLE mode. 0.128 0.062

Pul [W] Power consumption of a device when

continuously uploading data. 1.9 1.3

Pdl [W] Power consumption of a device when

continuously downloading data. 1.6 1.25

Povh

[W]

Average power consumption of a device

for tovh seconds after the 802.11 radio has

been switched to the IDLE mode.

0.235/

0.5 (sig/med)

-

t dch [s] DCH release timer for WCDMA. - 5

t fach [s] FACH release timer for WCDMA. - 20

tovh [s]

Time between when the 802.11 interface

switches to IDLE mode and when the

power consumption returns to Pidle.

5 -

IV. EVALUATION

The model was implemented in Matlab environment, and evaluated by comparing the simulation results with the results of real-life measurements. In this section, we first evaluate the accuracy of the signaling and media transfer parts, and then the model’s applicability to a P2P scenario.

A. Signaling

In Fig. 4 the simulation results are compared with real-life measurement results [12] for WCDMA (Fig. 4a) and 802.11b (Fig. 4b) access modes. The solid lines represent the simulated power consumptions of periodical transfers of 10B, 100B and 1000B packets with varying transmission intervals. The vertical bars, colored accordingly, depict the

535

lower and upper limits of the 95% confidence interval for the measured power consumption values for each evaluated interval. The measurements were made for 25ms, 50ms, 150ms, 250ms, 500ms, 2500ms and 5000ms intervals.

From Fig. 4, we can see that the simulated power consumptions are relatively accurate. With WCDMA, 72% of the simulated values in the sample points are inside the 95% confidence intervals of the measured values. With 802.11b, the corresponding percentage is 84%. Further statistical analysis gives 3.2% as the mean error for WCDMA, and 5.3% for 802.11b.

(a)

(b)

Figure 4. Simulated and measured power consumption of signaling as a

function of transmission interval with (a) WCDMA, and (b) 802.11.

B. Media transfers

Fig. 5 illustrates the accuracy of estimating the power consumption of media transfers. As measurement data, we used the data gathered with the setup described in Section 3C. We used a 21.0MB test file with 802.11, and a 3.16MB test file with WCDMA. In Fig. 5a, the file is uploaded using the WCDMA network interface, and in Fig. 5b, a file is downloaded to the mobile device using the 802.11 network interface. Also WCDMA download and 802.11 upload scenarios were measured, although not presented in figures.

As summary, the file transfer simulation seems to estimate the power consumption with adequate accuracy. When ignoring the first 15 seconds (due to the extra power consumption caused by manual transfers), the error in the total energy consumption is in all cases less than 1%. When looking more closely at the idle periods of the 802.11 scenario (Fig. 5b), there seems to be major differences between the power consumptions of real life and simulated scenarios. We have observed in our evaluations that the idle consumption of 802.11 fluctuates significantly over time, probably due to radio channel noise. However, based on our

measurements the long time idle consumption average is around 128mW for N95, which was used in our simulations.

(a)

(b)

Figure 5. Simulated and measured power consumption of (a) uploading a

large file using WCDMA, and (b) downloading a large file using 802.11b.

C. P2P overlay signaling

The media transfer part of the model is fully applicable to the real-world scenarios, due to its ability to simulate media transfers with varying data chunk sizes and concurrent transfers arbitrarily on time. However, as the signaling part of the model uses average packet sizes, it is expected to give less accurate results in the scenarios with varying packet sizes (even though the average value can be dynamically adjusted). Hence, we evaluated the signaling part in a P2P scenario to find out how the model copes with a scenario with highly varying packet sizes.

In Fig. 6, we compare the measured power consumption of P2P networking [13] with a simulated power consumption estimate for the same scenario. The simulation parameters were based on the result data of [13]. The signaling patterns of each case were known in detail, making it easy to define the average packet size and transmission interval. The blue bars in the figure depict the actual power consumption of the real-life scenario, whereas the red bars depict the simulated power consumption estimates. In total eight cases are presented for both network technologies. The variable overlay parameters for each case are presented below the x-axis. The parameters are the number of peers in the overlay (N), the lookup request interval (tlookup), and the average node online time (tchurn) .

With WCDMA, the average estimation error is 14.9%. The results are relatively accurate except the first few bars on the left. In theory, when the average transmission interval exceeds the threshold value τ1 (illustrated in Figs. 2a and 2b), the power consumption starts to dramatically drop. In the measurements the average packet sizes ranged between 108 and 114 bytes, leading to the situation, where τ1 ranges

536

between 146ms (108 byte packets) and 155ms (114 byte packets). Hence, τ1 is clearly inside the lower (114ms) and upper (250ms) bounds of the average transmission intervals of the measurements. The effect of τ1 can be seen in the first two bars of the simulated values, but it is not clearly visible in the measured values. A probable reason for this is that whereas in reality the packet sizes vary between 59 and 1000 bytes, simulations use the average packet sizes (108-114 bytes). Since the packets sized above 250 bytes trigger the transfer mode transition from FACH to DCH, it is clear that even scattered packets over 250 bytes significantly increase the power consumption in the real-life scenario (i.e. instead of Fig. 2a, the power consumption follows the curve of Fig. 2b). Since the average packet size remains below 250 bytes, the transfer mode is never triggered to DCH in the simulation, which explains the difference.

Figure 6. Measured and simulated power consumptions of P2P signaling.

With 802.11, the average estimation error is 20.5%. The

simulation accuracy is on an acceptable level, but is slightly

lower than with WCDMA. The signaling intervals (114-

250ms) are clearly between the threshold values τ1 and τ2

(see Fig. 2c) of the 802.11 interface, as can be seen in Fig.

4b. Thus, threshold phenomena related to transmission

interval are not present. Moreover, with 802.11, the

variation in packet sizes does not cause as significant

differences in power consumption as with WCDMA.

Instead, the most probable reason for the >20% estimation

error is unstable radio conditions. Even during idle periods

the 802.11 interface regularly kept sending and receiving

data that was not related to the scenario, and the amount of

this data significantly fluctuated over the time.

V. CONCLUSION

In this paper, we proposed a model, called e-Aware, for

estimating the effect of application-layer protocol solutions

on the energy consumption of mobile devices, operating in

WCDMA and 802.11 networks. The model takes into

account the fundamentals of radio interface properties, and

makes a distinction between signaling and media transfers

for maximizing the accuracy. The model is fine-tuned for

different devices using device-specific coefficients.

According to the evaluation results, the model achieves

high accuracy (3-6% estimation error) in signaling scenarios

with fixed packet sizes, as well as in file transfer scenarios

(<1% error). In a signaling scenario with high variance in

packet sizes, the accuracy is lower (14-21% error), as

expected. Since the model is primarily aimed as a design-

phase tool, in most cases only rough estimates of the

messaging patterns and message sizes can be given as

parameters. In such scenario, the measured estimation errors

can be considered acceptable.

As future work, the accuracy of the model could be

improved by developing the model further to consider the

variance of packet sizes in estimating the energy

consumption of signaling. The future work also includes

more comprehensive evaluations with different devices in

order to confirm the generality of the model.

ACKNOWLEDGMENTS

The financial support from TEKES, Infotech Oulu, and

foundations of HPY, Nokia, TeliaSonera, Walter Ahlström,

Seppo Säynäjäkangas, KAUTE, and Riitta & Jorma J.

Takanen is gratefully acknowledged.

REFERENCES

[1] N. Ravi, J. Scott, L. Han, and L. Iftode, “Context-aware Battery Management for Mobile Phones,” in, IEEE Conference on Pervasive Computing and Communications, Hong Kong, 2008, pp.224-233.

[2] T. Pering, Y. Agarwal, R. Gupta, and R. Want, "CoolSpots: Reducing the Power Consumption of Wireless Mobile Devices with Multiple Radio Interfaces,” in, International Conference on Mobile Systems, Uppsala, Sweden, 2006, pp. 220-232.

[3] P. Eronen, “TCP Wake-Up: Reducing Keep-Alive Traffic in Mobile IPv4 and IPsec NAT Traversal,” Nokia Research Center, Helsinki, Finland, Rep. NRC-TR-2008-002, 2008.

[4] I. Kelenyi and J.K. Nurminen, ”Energy Aspects of Peer Cooperation - Measurements with a Mobile DHT System,” in, IEEE International Conference on Communications, Beijing, China, 2008, pp. 164-168.

[5] J. Nurminen and J.K. Noyranen, “Energy-consumption in mobile peer-to-peer - quantitative results from file sharing,” in, Consumer Communications and Networking Conference, Las Vegas, USA,2008.

[6] N. Balasubramanian, A. Balasubramanian, and A. Venkataramani, “Energy Consumption in Mobile Phones: A Measurement Study and Implications for Network Applications”, in, ACM Internet Measurement Conference, Chicago, USA, 2009, pp. 280-293.

[7] H. Haverinen, J. Siren, and P. Eronen, ”Energy Consumption of Always-On Applications in WCDMA Networks,” in, IEEE Vehicular Technology Conference, Dublin, Ireland, 2007, pp. 964-968.

[8] P.H.J. Perälä, A. Barbuzzi, G. Boggia, K. Pentikousis, “Theory and Practice of RRC State Transitions in UMTS Networks,” in, IEEE Broadband Wireless Access Workshop, Hawaii, USA, 2009, pp.1-6.

[9] Y. Xiao, P. Savolainen, A. Karppanen, M. Siekkinen, and A. Ylä-Jääski, “Practical power modeling of data transmission over 802.11g for wireless applications” in, International Conference on Energy-Efficient Computing and Networking,Passau, Germany, 2010.

[10] L. Zhang, B. Tiwana, Z. Qian, Z. Wang, R.P. Dick, Z.M. Mao, and L. Yang, “Accurate Online Power Estimation and Automatic Battery Behavior Based Power Model Generation for Smartphones,” in, International Conference on Hardware-Software Codesign and System Synthesis, Scottsdale, USA, 2010, pp. 105–114.

[11] Holma, H., Toskala, A. “WCDMA for UMTS: HSPA Evolution and LTE,” 4th edition, John. Wiley & Sons, 2007.

[12] Z. Ou, E. Harjula, O. Kassinen, and M. Ylianttila, “Performance Evaluation of a Kademlia-based Communication-oriented P2P System under Churn,” Elsevier Journal of Computer Networks, vol. 54, no. 5, pp. 689-705.

[13] O. Kassinen, E. Harjula, J. Korhonen, M. Ylianttila, “Battery Life of Mobile Peers with UMTS and WLAN in a Kademlia-based P2P Overlay,” in, Personal, Indoor and Mobile Radio Communications Symposium, Tokyo, Japan, 2009, pp. 662-665.

537