A Novel Easy-to-construct Power Model for Embedded and Mobile

Systems

Using Recursive Neural Nets to Estimate Power Consumption of ARM-based

Embedded Systems and Mobile Devices

Oussama Djedidi, Mohand A. Djeziri, Nacer K. M’Sirdi and Aziz Naamane

Aix-Marseille University, Universit

e de Toulon, CNRS, LIS, SASV, Marseille, France

Keywords:

Data Fitting, Embedded Systems, Modeling, NARX, Neural Nets, Power Consumption, Smartphone.

Abstract:

This paper features a novel modeling scheme for power consumption in embedded and mobile devices. The

model hereafter presented is built thought data ﬁtting techniques using a NARX nonlinear neural net. It

showcases the advantages of using a nonlinear model to estimate power consumption over the widely used

linear regression models, where The NARX neural net is simpler, easier to implement, and more importantly

more suitable as power changes are not always linear. Finally, experimental results validate the model with

one with an accuracy of 97.1% on a smartphone.

1 INTRODUCTION

Power consumption in mobile and embedded devices

occupies a lead role in research since these devices

either come with limited resources (batteries) or have

to be optimized for minimal consumption as is the

case for the internet of things (IoT). The high in-

terest in power consumption and estimation is justi-

ﬁed by a multitude of reasons. First and foremost,

power models allow the developers to estimate the re-

maining battery life (Kim et al., 2016) and estimate

the energy consumption of their applications (Mittal

et al., 2012). They also show developers how user

and application actions and interactions translate into

energy loss (Guo et al., 2017), allow for the said-

developers to debug power consumption in the de-

vice (Hoque et al., 2015), and even identify energy

hogs in the code of their applications (Banerjee et al.,

2014). Moreover, by studying power consumption,

researchers are able to create energy-friendly algo-

rithms (Huang et al., 2017) and even more innova-

tively detect security issues and threats (Caviglione

et al., 2016).

Researchers have been building models to esti-

mate power consumption for over ﬁfteen years, with

each model focusing on either accurate power estima-

tion (Gordon et al., 2011), or granularity (Google Inc.,

2017), or the simplicity of implementation (Kim et al.,

2012; Kim et al., 2014). Most of these models are

repotorted in an extended litterature review (Hoque

et al., 2015), in which Hoque et al. proposed a tax-

onomy for the recorded power proﬁlers and models

according to the measurement mechanisms, different

inputs used to each type of model (utilization, event-

driven and code), modeling philosophies (white-box

vs black-box), proﬁling schemes (On or off-device),

and granularity. The most prominent (and still avail-

able) among the studied are those built by sys-

tem providers like Google’s Android Power Proﬁler

(Google Inc., 2017), Trepn and Snapdragon Proﬁler

(Qualcomm Innovation Center, ), and research pub-

lished in the literature such as PowerBooter (Zhang

et al., 2010), Sesame (Dong and Zhong, 2011), DevS-

cope (Jung et al., 2012), Eprof (Pathak et al., 2012),

and the models presented in (Banerjee et al., 2014)

and (Shye et al., 2009).

Ahmad et al. also established a survey and es-

sentially mentioned the same works. Nevertheless, it

looked at the proﬁlers through the lens of their imple-

mentation (software/hardware-based) (Ahmad et al.,

2015).

Besides the works cited in the surveys, Kim et

al. also proposed a polynomial model for power

consumption modeling, built using linear regression

(Kim et al., 2012). The authors later continued to en-

hance their model with better power estimation for the

GPU (Kim and Chung, 2013), and the display (Kim

et al., 2015). More recently, several new models were

Djedidi, O., Djeziri, M., M’Sirdi, N. and Naamane, A.

A Novel Easy-to-construct Power Model for Embedded and Mobile Systems - Using Recursive Neural Nets to Estimate Power Consumption of ARM-based Embedded Systems and Mobile

Devices.

DOI: 10.5220/0006915805410545

In Proceedings of the 15th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2018) - Volume 1, pages 541-545

ISBN: 978-989-758-321-6

541

also published, either for the system as a whole, as

it was the case for (Di Nucci et al., 2017a), (Shukla

et al., 2016) and (Chowdhury and Hindle, 2016), or

just the Central Processing Unit (CPU) like the mod-

els presented in (Yoon et al., 2017) and (Walker et al.,

2017). Some of them even offered better accuracy like

(Yoon et al., 2017) and proved that software proﬁling

is practically as accurate and reliable as hardware pro-

ﬁling (Di Nucci et al., 2017b).

In a previous work, Djedidi et al. proposed a neu-

ral net model for power consumption (Djedidi et al.,

2017). The model is a subsystem part of a general

model to estimate several characteristics of a whole

embedded system on chip (SoC). In their work, the

authors focused on building a model with little over-

head and capable of delivering estimations online

with little lag. Although the model was quite accu-

rate—With an accuracy of 95%—it still presented a

lot of noise that didn’t correspond to estimations (Dje-

didi et al., 2017).

In this work, we propose a new Artiﬁcial Neural

Network (ANN) model to power consumption estima-

tion that offers better results and keeps the estimation

speed and the low overhead. The model is a nonlinear

autoregressive model with exogenous inputs (NARX)

neural net. As it will be showcased by the experimen-

tal results later in this work, the proposed model also

manages to overcome the traditional shortcomings of

the regression-based models which fall short when the

power change is not linear.

The remainder of this work is structured as fol-

lows. In section 2, the model construction process

is described alongside the choice of the type of the

model and the inputs. Section 3 is used to describe

the experimental setup, followed by the experimental

results and a comparative study in section 4. Finally,

the work ends with a conclusion of obtained results

and future works.

2 MODEL CONSTRUCTION

Data-ﬁtting techniques are quite useful when not

all the dynamics are known in a system or are too

complex to be modeled vis-

a-vis the constraints at

hand—computational cost, for instance. These tech-

niques train the model so that its estimations would

match the measured outputs. A lot of the models re-

ported in the literature rely on black-box techniques,

more precisely linear regression (Hoque et al., 2015),

since white-box techniques can become quite com-

plex (Pathak et al., 2012).

Regression-based models, as is the case for

PowerBooter (Zhang et al., 2010), the models by

Dong et al., (Dong and Zhong, 2011), Kim et al. (Kim

et al., 2012), Kim et al. (Kim et al., 2015), Shukla et

al. (Shukla et al., 2016), and Xu et al. (Xu et al., 2013)

amongst others, are quite popular. However, these

models assume—by deﬁnition—that the variations in

power consumption are linear causing an increased

estimation error when it is not the case (Hoque et al.,

2015). In this work, we avoid the pitfalls of using a

linear model by using a nonlinear model capable of

accommodating nonlinear dynamics; The model we

use is a NARX neural network (Xie et al., 2009).

Nonetheless, data-based models—Whether con-

structed through regression techniques or using ma-

chine learning—require no formal relation of inter-

action between the inputs and the estimated outputs

(Zhang et al., 2010). They rely often upon obser-

vations and experimental analyses in various operat-

ing conditions for the choice of model inputs (Hoque

et al., 2015), which—along with the ease of construc-

tion and implementation—explains their popularity.

2.1 Granularity and Input Selection

One of the major differences between the different

power models in embedded and mobile systems is

their granularity, which dictates the lowest level at

which the model can estimate power consumption

(Hoque et al., 2015). It is generally deﬁned by the

purpose behind the building of the model. For in-

stance, in the models built to investigate the inﬂuence

of elements of the code on power consumption, the

granularity should be as ﬁne as possible, at function

level or even more (line of code). The model con-

structed in the work is an evolution and an improve-

ment upon that constructed in (Djedidi et al., 2017).

Hence, it aims to estimate and monitor power con-

sumption of system components mainly those found

in the SoC.

As shown in ﬁg 1, the total amount of power con-

sumed by the device can be broken down into the

Figure 1: Power consumption distribution in a typical mo-

bile or embedded system.

ICINCO 2018 - 15th International Conference on Informatics in Control, Automation and Robotics

542

sum of powers consumed by each individual compo-

nent (i.e. CPU, or Wi-Fi, ...) plus a minimum static

amount of power that is always consumed by the de-

vice (Hoque et al., 2015) :

Total

= P

Static

∑

i=1

(1)

i denotes the component number (CPU, GPU,

RAM...), and n is the total number of these compo-

nents. The power consumed each of these compo-

nents can be correlated with one of its characteris-

tics. These characteristics will be used as inputs to

the power model.

To keep the model as simple as possible—and

consequently as computationally optimized as possi-

ble, we chose to include only the most characterizing

and directly inﬂuential inputs.

2.1.1 The SoC

In most of the embedded system, the SoC is com-

posed of a CPU, the Random Access Memory

(RAM), and a Graphics Processing Unit (GPU) when

needed. CPU power consumption is a function of its

frequency and voltage (Altamimi and Naik, 2015).

Moreover, the voltage in processors with dynamic

voltage and frequency scaling (DVFS) is also a func-

tion of frequency (Djedidi et al., 2017). Thus, the in-

put used to estimate power consumption is the fre-

quency f of each CPU core. Most works in the lit-

erature also use the load, but our experimental results

showed no improvement with its inclusion, and we

chose for optimization purposes to omit it. The same

input is used for the GPU.

To account for the power used by the RAM, in

the ﬁrst trials we opted for the use of the value of the

occupied RAM. However, that value on its own does

not factor the maximum and the minimum possible

values of the RAM on the system. Henceforth, we opt

for a ratio of the occupied RAM over its maximum

value and call it Memory occupation Rate (MOR).

2.1.2 Communication Peripherals

The communication peripherals are the phone’s GSM

chip, Wi-Fi (and Bluetooth), and the GPS. For

the GSM chip, the inputs are the On-call status

(ON/OFF), the signal strength, and the data transfer

rate (download and upload). Similarly, for the Wi-Fi

and Bluetooth, the inputs are the status and the data

rate (download and upload). Finally, for the GPS, we

used the only status.

y(k)

u(k)

TDL

(u(k), u(k -1))

TDL

(y(k), y(k -1))

Figure 2: The general structure of the neural network

model.

2.1.3 Output Peripherals

The output peripherals considered in this study is only

the screen as a case study of the model. For the

touchscreen, the status of the screen (ON/OFF) and

its brightness are used as inputs. We consider that the

power consumed by the touch layer of the screen is

negligible (Kim et al., 2012).

2.2 The NARX Neural Net

NARX neural networkss are discrete-time nonlinear

systems that are used to predict the output of time se-

ries (Xie et al., 2009). Mathematically, they are rep-

resented as follows :

y(k + 1) = f [y(k), ..., y(k − d

+ 1)

u(k − n), u(k − n − 1), ..., u(k − n − d

+ 1)] (2)

y(k) and u(k) are the output to be predicted and the

input a the time step k, respectively. The term n is the

process dead-time. While d

and d

are the orders of

time delays for the input and output, also called input

and output-memory. Since the power changes reacts

pretty quickly with changes in the characteristics (in-

puts), d

, d

, and n were all set to the value of 1. In

the previous paragraph, the input u was set to be:

u = [ f

, ..., f

, f

GPU

, MOR,

GSM

OnCall

, GSM

SignalStrength

, GSM

DataRate

Wi − Fi

Status

, Wi − Fi

DataRate

, GPS

Status

Screen

Status

, Screen

Brightness



(3)

Where j is the total number of CPU cores in the sys-

tem. The neural network is composed of two layers.

The hidden layer contains n neurons whose activation

functions are sigmoids, where n is the length of the

input vector. The output layer contains a single neu-

ron with a linear transfer function. Fig. 2 shows the

general structure of the NARX model with the output

feedback and the time delay line (TDL).

3 EXPERIMENTAL SETUPS

For running tests and undertaking experimental val-

idation, we used an Android™ smartphone. These

A Novel Easy-to-construct Power Model for Embedded and Mobile Systems - Using Recursive Neural Nets to Estimate Power Consumption

of ARM-based Embedded Systems and Mobile Devices

543

devices are very popular and developed application

can be effortlessly transferred from one device to an-

other. Moreover, the availability of the source code

for their operating systems makes some of the needed

parameters accessible for reading and even modiﬁca-

tion. Additionally, they are equipped with the all nec-

essary sensors for the measurement of all the variables

mentioned in paragraph 2.1.

The smartphone we used in this study equipped

with an octa-core processor arranged in a big.LITTLE

conﬁguration (ARM Inc, ), running at frequencies up

to 2.3 GHz, a state of the art GPU, and 4 GB of RAM.

4 EXPERIMENTAL RESULTS

In ﬁgure 3, the measured power consumption of the

Android™ phone is drawn against values estimated

by the NARX neural network model. The estimations

made by the model follow the measured values accu-

rately with minimal estimation errors. It has an ac-

curacy of 97.12%, and the recorded Mean Absolute

Error (MAE) 0.0168 W, whereas the Mean Squared

Error (MSE) is 7.04 × 10

−4

. The model is hence val-

idated with encouraginf results.

925 930 935 940 945 950 955 960 965 970

Time (s)

0.5

1.5

Power(W)

meas

estm

Figure 3: Power estimation by the neural network model

against system measurements.

Going a step further, we compared our the accu-

racy of our results with established and recent power

models. In our tests, the measured accuracies of Trepn

and Snapdragon Proﬁler were 94.5% and 95% re-

spectively. PowerBooter has a reported accuracy of

96% (Zhang et al., 2010), and PETrA’s accuracy is

also reported to be the same(Di Nucci et al., 2017a),

as is the case for Eprof (Hoque et al., 2015).

We also compared the power overhead caused by

our modeling and monitoring program to the proﬁlers

we could test; PowerBooter, Trepn, and Snapdragon

Proﬁler. The ﬁrst caused an increase of 9% in power

consumption during a 15 minutes test. While the sec-

ond caused an increase of 8%, in a same scenario test.

The Snapdragon Proﬁler caused an increase of 5%.

Compared to the three aforementioned proﬁlers, our

modeling and monitoring scheme caused an increase

of only 3% in power consumption, during the same

test.

5 CONCLUSION

In this work, we have presented a novel power model

for embedded and mobile devices. The new model

is very simple to construct and train. It causes less

overhead than existing ones and still performs at the

speeds necessary to follow the power changes in the

system. Finally and more importantly, it improves

upon existing models in terms of accuracy with a

mean absolute percentage error of 2.8%, one of the

smallest in the reported literature.

This model will be incorporated in the in the incre-

mental modeling scheme previously built by (Djedidi

et al., 2017) to improve the estimation results. It will

also be used to monitor power consumption in the de-

vice in future studies aiming to study the health status

and operating state of the device.

ACKNOWLEDGEMENTS

This paper is a part of the ”FUI 19” project which

is supported and funded by the Banque Publique

d’Investissement (BPI), to whom we address our

thanks.

REFERENCES

Ahmad, R. W., Gani, A., Hamid, S. H. A., Xia, F., and Shi-

raz, M. (2015). A Review on mobile application en-

ergy proﬁling: Taxonomy, state-of-the-art, and open

research issues. J. Netw. Comput. Appl., 58:42–59.

Altamimi, M. L. and Naik, K. (2015). A Computing Pro-

ﬁling Procedure for Mobile Developers to Estimate

Energy Cost. In Proc. 18th ACM Int. Conf. Model.

Anal. Simul. Wirel. Mob. Syst. - MSWiM ’15, pages

301–305, New York, New York, USA. ACM Press.

ARM Inc. Technologies — big.LITTLE – Arm Developer.

Banerjee, A., Chong, L. K., Chattopadhyay, S., and Roy-

choudhury, A. (2014). Detecting energy bugs and

hotspots in mobile apps. In Proc. 22nd ACM SIG-

SOFT Int. Symp. Found. Softw. Eng. - FSE 2014, vol-

ume 3, pages 588–598, New York, New York, USA.

ACM Press.

Caviglione, L., Gaggero, M., Lalande, J. F., Mazurczyk, W.,

and Urba

nski, M. (2016). Seeing the unseen: Re-

vealing mobile malware hidden communications via

energy consumption and artiﬁcial intelligence. IEEE

Trans. Inf. Forensics Secur., 11(4):799–810.

ICINCO 2018 - 15th International Conference on Informatics in Control, Automation and Robotics

544

Chowdhury, S. A. and Hindle, A. (2016). GreenOracle: Es-

timating Software Energy Consumption with Energy

Measurement Corpora. In Proc. 13th Int. Work. Min.

Softw. Repos. - MSR ’16, pages 49–60, New York,

New York, USA. ACM Press.

Di Nucci, D., Palomba, F., Prota, A., Panichella, A., Zaid-

man, A., and De Lucia, A. (2017a). PETrA: A

Software-Based Tool for Estimating the Energy Pro-

ﬁle of Android Applications. In 2017 IEEE/ACM 39th

Int. Conf. Softw. Eng. Companion, pages 3–6. IEEE.

Di Nucci, D., Palomba, F., Prota, A., Panichella, A., Zaid-

man, A., and De Lucia, A. (2017b). Software-based

energy proﬁling of Android apps: Simple, efﬁcient

and reliable? In SANER 2017 - 24th IEEE Int. Conf.

Softw. Anal. Evol. Reengineering, volume 15, pages

103–114. IEEE.

Djedidi, O., Djeziri, M. A., M’Sirdi, N. K., and Naa-

mane, A. (2017). Modular Modelling of an Embed-

ded Mobile CPU-GPU Chip for Feature Estimation.

In Proc. 14th Int. Conf. Informatics Control. Autom.

Robot., volume 1, pages 338–345, Mardrid, Spain.

SCITEPRESS - Science and Technology Publications.

Dong, M. and Zhong, L. (2011). Self-constructive high-rate

system energy modeling for battery-powered mobile

systems. In MobiSys, page 335, New York, New York,

USA. ACM Press.

Google Inc. (2017). Measuring Power Values — Android

Open Source Project.

Gordon, M., Zhang, L., and Tiwana, B. (2011). PowerTutor.

Guo, Y., Wang, C., and Chen, X. (2017). Understand-

ing application-battery interactions on smartphones:

A large-scale empirical study. IEEE Access, 5:13387–

13400.

Hoque, M. A., Siekkinen, M., Khan, K. N., Xiao, Y., and

Tarkoma, S. (2015). Modeling, Proﬁling, and De-

bugging the Energy Consumption of Mobile Devices.

ACM Comput. Surv., 48(3):1–40.

Huang, J., Li, R., An, J., Ntalasha, D., Yang, F., and Li,

K. (2017). Energy-Efﬁcient Resource Utilization for

Heterogeneous Embedded Computing Systems. IEEE

Trans. Comput., 66(9):1518–1531.

Jung, W., Kang, C., Yoon, C., Kim, D. D., and Cha, H.

(2012). DevScope: A Nonintrusive and Online Power

Analysis Tool for Smartphone Hardware Compo-

nents. Proc. eighth IEEE/ACM/IFIP Int. Conf. Hard-

ware/software codesign Syst. Synth. - CODES+ISSS

’12, page 353.

Kim, D., Chon, Y., Jung, W., Kim, Y., and Cha, H. (2016).

Accurate Prediction of Available Battery Time for

Mobile Applications. ACM Trans. Embed. Comput.

Syst., 15(3):1–17.

Kim, K., Shin, D., Xie, Q., Wang, Y., Pedram, M., and

Chang, N. (2014). FEPMA: Fine-Grained Event-

Driven Power Meter for Android Smartphones Based

on Device Driver Layer Event Monitoring. In Des.

Autom. Test Eur. Conf. Exhib., pages 1–6, New Jersey.

IEEE Conference Publications.

Kim, M. and Chung, S. W. (2013). Accurate GPU power es-

timation for mobile device power proﬁling. Dig. Tech.

Pap. - IEEE Int. Conf. Consum. Electron., pages 183–

184.

Kim, M., Kong, J., and Chung, S. W. (2012). Enhancing

online power estimation accuracy for smartphones.

IEEE Trans. Consum. Electron., 58(2):333–339.

Kim, Y. G., Kim, M., Kim, J. M., Sung, M., and Chung,

S. W. (2015). A novel GPU power model for accurate

smartphone power breakdown. ETRI J., 37(1):157–

164.

Mittal, R., Kansal, A., and Chandra, R. (2012). Empow-

ering developers to estimate app energy consumption.

In Proc. 18th Annu. Int. Conf. Mob. Comput. Netw. -

Mobicom ’12, page 317, New York, New York, USA.

ACM Press.

Pathak, A., Hu, Y. C., and Zhang, M. (2012). Where is the

Energy Spent Inside My App?: Fine Grained Energy

Accounting on Smartphones with Eprof. In Proc. 7th

ACM Eur. Conf. Comput. Syst., EuroSys ’12, pages

29–42, New York, NY, USA. ACM.

Qualcomm Innovation Center, I. Trepn Proﬁler - Android

Apps on Google Play.

Shukla, N. K., Pila, R., and Rawat, S. (2016). Utilization-

based power consumption proﬁling in smartphones. In

Proc. 2016 2nd Int. Conf. Contemp. Comput. Infor-

matics, IC3I 2016, pages 881–886. IEEE.

Shye, A., Scholbrock, B., and Memik, G. (2009). Into

the wild: Studying real user activity patterns to guide

power optimizations for mobile architectures. Micro,

pages 168–178.

Walker, M. J., Diestelhorst, S., Hansson, A., Das, A. K.,

Yang, S., Al-Hashimi, B. M., and Merrett, G. V.

(2017). Accurate and Stable Run-Time Power Mod-

eling for Mobile and Embedded CPUs. IEEE Trans.

Comput. Des. Integr. Circuits Syst., 36(1):106–119.

Xie, H., Tang, H., and Liao, Y. H. (2009). Time series pre-

diction based on narx neural networks: An advanced

approach. In Proc. 2009 Int. Conf. Mach. Learn. Cy-

bern., volume 3, pages 1275–1279. IEEE.

Xu, F., Liu, Y., Li, Q., and Zhang, Y. (2013). V-edge:

fast self-constructive power modeling of smartphones

based on battery voltage dynamics. In nsdi’13 Proc.

10th USENIX Conf. Networked Syst. Des. Implement.,

pages 43–55. USENIX Association.

Yoon, C., Lee, S., Choi, Y., Ha, R., and Cha, H. (2017). Ac-

curate power modeling of modern mobile application

processors. J. Syst. Archit., 81:17–31.

Zhang, L., Tiwana, B., Qian, Z., Wang, Z., Dick, R. P., Mao,

Z. M., and Yang, L. (2010). Accurate online power es-

timation and automatic battery behavior based power

model generation for smartphones. In Proc. eighth

IEEE/ACM/IFIP Int. Conf. Hardware/software code-

sign Syst. Synth. - CODES/ISSS ’10, page 105, New

York, New York, USA. ACM Press.

A Novel Easy-to-construct Power Model for Embedded and Mobile Systems - Using Recursive Neural Nets to Estimate Power Consumption

of ARM-based Embedded Systems and Mobile Devices

545