scispace - formally typeset
Search or ask a question
Proceedings ArticleDOI

Modeling WiFi Active Power/Energy Consumption in Smartphones

TL;DR: This study builds four versions of a previously proposed linear power-throughput model for WiFi active power/energy consumption based on parameters readily available to smartphone app developers and evaluates its accuracy under a variety of scenarios which have not been considered in previous studies.
Abstract: We conduct the first detailed measurement study of the properties of a class of WiFi active power/energy consumption models based on parameters readily available to smartphone app developers. We first consider a number of parameters used by previous models and show their limitations. We then focus on a recent approach modeling the active power consumption as a function of the application layer throughput. Using a large dataset and an 802.11n-equipped smartphone, we build four versions of a previously proposed linear power-throughput model, which allow us to explore the fundamental trade off between accuracy and simplicity. We study the properties of the model in relation to other parameters such as the packet size and/or the transport layer protocol, and we evaluate its accuracy under a variety of scenarios which have not been considered in previous studies. Our study shows that the model works well in a number of scenarios but its accuracy drops with high throughput values or when tested on different hardware. We further show that a non-linear model can greatly improve the accuracy in these two cases.

Summary (2 min read)

Introduction

  • The Mg–Fe mixing is a complex phenomenon, important to form solid phases that participate in many a natural process over a wide pressure–temperature (P–T) range.
  • In view of the magnesio-wüstite’s wide geological and technological scope, it is important to achieve as full an understanding as possible of the principles underlying the stability of the (Mg,Fe)O solid solution, as a function of those aspects that primarily affect its reactivity in the variety of the transformations in which it participates.
  • Quantum mechanical and semi-empirical calculations, in combination with lattice dynamics and statistical thermodynamics, have proven a powerful tool to model the energetics of solid solutions (for instance: Vinograd et al. 2013, and references therein; De La Pierre et al.
  • In the light of the discussion above, the authors have decided to undertake the present work, in which they model thermochemical properties and T–X phase relations diagram as a function of iron spin configuration (low spin, S = 0: diamagnetic, LS; high spin, S = 2: antiferromagnetic/paramagnetic, HS), composition and temperature of (Mg,Fe)O.

Theoretical

  • The authors remind here some fundamentals for a cluster expansiontype approach to describe the energetics of solid solutions, using the discrete Chebyshev polynomials method (Sanchez et al. 1984).
  • In so doing, the authors have observed that Eq. (10) (i.e. truncation up to 5.5 Å) provides for (Mg,Fe)O’s mixing lattice energy a description numerically comparable to that achievable by more shells and a more complex xFe dependence of the A0 coefficients, which turn out to be highly correlated with each other.

Results and discussion

  • In Fig. 1, the primitive cell volume, V, is displayed as a function of xFe.
  • Figure 4a, b display at 1,373 and 1,573 K their results from computational modelling and experimental data.
  • Using data from Sreҫec only, Fig. 5 shows that the disagreement between theoretical and observed activity decreases upon increasing T. Calculations performed by the HS model 1 3 ΔF(T,xFe)mixing is positive/negative as a function of temperature, it ensures that the pure static energy contribution tends to favour de-mixing but it is offset by vibrational and configuration components of ΔF(T,xFe)mixing.
  • Such a discrepancy, i.e. an excess of Mg in the Fe-rich phase, with respect to observations is ascribable, and the authors deem, either to pressure that is not taken into account in the T–X phase relations diagrams here discussed, or to a model deficiency in reproducing enough asymmetry on the MgO–FeO join.

Conclusions

  • The ΔF(T,xFe)mixing modelled by the HS configuration points to a MgO–FeO solid solution largely controlled by static and configuration contributions, i.e. ΔF(T,xFe)LTmixing, in terms of more than 80 %.
  • The modest vibrationdependent component contributes promoting mixing.
  • The U1(T) function provides a modest contribution to ΔFmixing that has a quasi-symmetric behaviour over the MgO–FeO binary.
  • The sub-solidus solvi exhibits a critical temperature (Tc) of some 950 and 1,200 K, including or neglecting the harmonic part of free Helmholtz energy and using the HS configuration.
  • For the sake of comparison with other minerals hosting Mg–Fe mixing and taking as a reference the former Tc, such a figure is significantly larger than olivine’s and Mg–Fe garnet’s, wherein Mg–Fe species enter octahedral and dodecahedral sites, respectively.

Did you find this useful? Give us your feedback

Content maybe subject to copyright    Report

Modeling WiFi Active Power/Energy Consumption in Smartphones
Li Sun, Ramanujan K. Sheshadri, Wei Zheng, Dimitrios Koutsonikolas
Department of Computer Science and Engineering, University at Buffalo, SUNY
Buffalo, NY, 14260-2500
Email: {lsun3, ramanuja, wzheng4, dimitrio}@buffalo.edu
Abstract—We conduct the first detailed measurement study
of the properties of a class of WiFi active power/energy
consumption models based on parameters readily available to
smartphone app developers. We first consider a number of
parameters used by previous models and show their limitations.
We then focus on a recent approach modeling the active
power consumption as a function of the application layer
throughput. Using a large dataset and an 802.11n-equipped
smartphone, we build four versions of a previously proposed
linear power-throughput model, which allow us to explore
the fundamental tradeoff between accuracy and simplicity.
We study the properties of the model in relation to other
parameters such as the packet size and/or the transport layer
protocol, and we evaluate its accuracy under a variety of
scenarios which have not been considered in previous studies.
Our study shows that the model works well in a number of
scenarios but its accuracy drops with high throughput values
or when tested on different hardware. We further show that
a non-linear model can greatly improve the accuracy in these
two cases.
I. INTRODUCTION
Despite the continuously growing popularity of smart-
phones, their utility has been and will always be limited by
their battery life. A major fraction of the energy consumption
in smartphones comes from the WiFi radio, which can
account for more than 50% of the total device power budget
under typical use [1], [2] and can quickly drain the phone’s
battery when transmitting at high peak rates.
The problem becomes particularly pronounced today due
to a combination of reasons. Today’s smartphones run a
variety of network apps which result in a large growth of
network traffic. The discontinuation of unlimited data plans
by most 3G/4G operators forces smartphone users to offload
a continuously growing amount of traffic to WiFi. The
availability of 802.11n/ac in modern smartphones further
exacerbates the situation. Recent studies [3], [4] have shown
that popular 802.11n wireless cards could deplete a typical
smartphone battery in 2-3 hours. Thus, understanding and
optimizing WiFi energy consumption becomes essential for
app developers in order to maximize the battery lifetime of
smartphones.
Since directly measuring WiFi power/energy consump-
tion [5], [1], [6] is a non-trivial task, several recent works
have focused on developing WiFi energy consumption mod-
els [7], [8], [9], [10], [11], [12], [13], [14], [15]. These
models can be broadly classified into two categories.
On one hand, a number of works model the WiFi energy
consumption based on the circuitry or MAC/PHY layer fea-
tures [7], [8], [15], [9], [14]. Such models can offer very high
accuracy; however, they require knowledge only available at
the driver/firmware level and hence, they cannot be used by
app developers. Furthermore, most of these models were
developed for and tested on WiFi cards for laptops/desktops
and they may not be applicable to smartphones [16], [17].
On the other hand, a number of recent works develop
power/energy models for different smartphone components
(disk, CPU, display, cellular, WiFi) that can be easily used
by app developers [11], [10], [13], [18], [19], [20], [21].
In the case of WiFi, one common feature of most of
these models [11], [12], [13], [18], [19] is the assumption
of constant power consumption in each power state.This
assumption is valid for idle/transition power states (e.g.,
sleep, ramp/promotion, and tail power states [10], [13]) and
can also offer satisfactory accuracy for the active power state
(during packet transmissions/receptions) in the case of low
bitrates (e.g., 802.11b) and/or small data transfers.
A major drawback of the constant active power state
assumption is that it fails to capture the characteristics of the
dynamic wireless channel (fading, interference, collisions)
which may trigger retransmissions, exponential backoff, or
rate adaptation during the active power state. Furthermore,
most of the models in this category (with the exception
of [18], [19]) were built for and tested on smartphones
equipped with legacy 802.11b/g WNICs. As recent stud-
ies [3], [14], [16] have shown, the rich set of the new
MAC/PHY features introduced by 802.11n large number
of available modulation and coding schemes (MCS), frame
aggregation/block ACK (FA/BA), channel bonding, MIMO
1
define a large number of active power states. As an
example, the receive power consumption of a Google Nexus
S smartphone at the highest supported 802.11n bitrate (MCS
7) is 23-102% higher than at the lowest bitrate (MCS 0) [16].
Hence, accurate modeling of the WiFi active power/energy
consumption becomes a critical requirement as apps perform
larger data transfers and 802.11 standards move towards
more and higher bitrates.
A small number of recent works model the WiFi active
1
MIMO is not supported by today’s smartphones but it may be supported
in the near future [22].

energy consumption of a smartphone as a function of one or
more input parameters [12], [10], [20], [21]. These parame-
ters are either application layer parameters easily measured
by app developers using tools such as tcpdump [23] (e.g.,
transfer size [10], packet transmission/reception rate [12],
throughput [20]) or lower layer parameters available to
app developers through an API (e.g., signal strength [21]).
Instead of directly measuring the energy consumption of
their app (which requires specialized hardware, e.g., a power
monitor or an oscilloscope), app developers can easily obtain
the value of the input parameter while running their app and
estimate the energy consumption using the model. In spite
of their attractiveness due to their simplicity, most of these
models still fail to fully capture the complex characteristics
of the wireless channel as well as those of the 802.11n
WNICs. Consequently, they may work well only under
certain scenarios, e.g., in lossless environments or in the
absence of interference.
Overall, in spite of the large number of models available to
app developers, the true capabilities and weaknesses of these
models remain to a large extent unknown. In certain cases,
the accuracy of the developed models is not evaluated at all
([10] or [20] in the case of WiFi). In other cases, the models
are validated only at the location where the training dataset
was collected, and under ideal conditions, without external
interference [12], [13], [21], or with only one traffic type
(typically small HTTP transfers [20], [21]). Finally, there is
often no detailed information about the methodology used
to collect the training dataset [20], [21], [24].
In this paper, we conduct the first extensive measurement
study of WiFi active energy consumption models based on
parameters easily measured by app developers, trying to
understand their capabilities and limitations. We begin (Sec-
tion IV) by examining various parameters used by recently
proposed models packet loss rate, signal strength, transfer
size, throughput. We show that most of these parameters
fail to accurately capture the dynamics of the wireless
environment and/or the 802.11n MAC/PHY features.
A notable exception, which can uniquely capture both the
wireless channel characteristics and several of the 802.11
MAC/PHY features, is the application layer throughput. A
few recent works have developed linear models of the active
power consumption of a wireless interface as a function
of the throughput [20], [24], [25] or have experimentally
observed such a linear relationship [17]. However, all these
works suffer from at least one of the limitations mentioned
above, i.e., limited or no validation, lack of details about
the data collection methodology, and testing only on older
802.11b/g smartphones. Hence, our second and main contri-
bution of this paper is a detailed evaluation of the capabilities
and properties of throughput-based energy modeling. We
perform the evaluation in four steps.
1) Based on a large dataset which covers different link qual-
ities, transport layer protocols, packet sizes, and MAC/PHY
features, we rebuild the linear model from [20], [25] for a
smartphone equipped with an 802.11n WNIC (Section V).
We explore the fundamental tradeoff between complexity
and accuracy by considering four different options for build-
ing the model.
2) We offer a detailed analysis of the estimation errors with
each of the four options and show that the accuracy can be
improved with the knowledge of the transport layer protocol
and/or the packet size (Section VI).
3) We extensively evaluate the accuracy of the model in
a variety of scenarios, many of which have not been con-
sidered in previous studies (Section VII). Our results show
that the linear model proposed in [20], [25] can accurately
predict the energy consumption in several of these scenarios.
However, its accuracy drops in the case of high throughput
values and when tested on different hardware.
4) We discuss important practical issues, such as how to col-
lect a good training dataset, or how to reduce the complexity
of the most accurate of the four models, and we show that
a non-linear power-throughput model can further improve
the accuracy compared to the linear model from [20], [25]
(Section VIII).
II. W
IFI ENERGY MODELS FOR SMARTPHONES
A number of recent works have developed models of
the power/energy consumption of different smartphone com-
ponents [11], [12], [10], [13], [18], [19], [20], [21], [25],
[24]. In the case of WiFi, one common feature of several
of these models [11], [12], [13], [18], [19], regardless
of their complexity, is the assumption of constant power
consumption at each power state. This assumption does
not hold true for the active power state and can lead to
high inaccuracy as the transfer size, the data rates, or the
complexity of the MAC/PHY layer increase.
In [10], a simple model is developed for the active energy
consumption of an 802.11b WNIC in a smartphone as a
linear function of the data transfer size. Although its sim-
plicity makes it a good candidate for use by app developers,
we note that the data transfer size is not the only factor
that affects energy consumption. The time to download a
file of a given size depends on the channel conditions and
the MAC/PHY protocol characteristics, which determine the
PHY bitrate. Note that the accuracy of the model was not
evaluated in [10].
In [12], the WiFi active energy consumption is modeled
as a function of the data rate (in packets/sec) and the PHY
bitrate. The authors claim that the packet size does not
affect power consumption. However, [16], [24], [15] show a
different result. In Section VI, we also show that the packet
size largely affects the per-bit energy consumption. As far
as the PHY bitrate is concerned, several works have recently
shown that it affects the energy consumption of an 802.11n
WNIC [3], [14], [16]. However, information about the per-
packet bitrate is often not exposed by today’s smartphones,

as rate adaptation in the case of smartphones is typically
implemented at the firmware level. In addition, the bitrate
alone may not always be a good predictor of the energy
consumption, as it cannot capture sender side wireless inter-
ference, which can elongate packet transmissions/receptions
due to carrier sensing.
In [21], a signal strength-aware model is proposed, which
maps different RSSI levels to different values of power con-
sumption. However, [26] showed that signal strength alone
cannot always capture the dynamics of the wireless channel
in 3G/4G networks. Similarly, in a WiFi network, high signal
strength does not always imply low energy consumption due
to hidden terminals and sender-side interference. Note that
the models in [21] were built and evaluated during the night
hours, when interference was limited.
In [20], [25], a linear model of the active power consump-
tion of a wireless interface (3G, 4G, WiFi) is proposed as
a function of the application layer throughput. [24] builds a
similar model of the energy consumption as a linear function
of the packet transmission time (which can be converted to
a linear model of power vs. throughput). The accuracy of
the model in [20] is only evaluated for 4G and short HTTP
transfers. In the case of WiFi, the training dataset includes
only very low data rates (0-2 Mbps) and only TCP traffic. On
the other hand, [24] and [25] used only UDP packets varying
the packet size and the data rate. Note also that only [25]
built the model for an 802.11n-equipped smartphone.
III. E
XPERIMENTAL SETUP
Our experimental setup includes one PC acting as an AP
and one smartphone acting as a client. The PC is part of
a 21-node wireless testbed (UBMesh [28]) deployed on the
3rd floor of UB Davis Hall. Each node has a Ralink RT2860
802.11a/b/g/n mini PCI card, which implements all the
available 802.11n features. The phone is an Android Google
Nexus S (the same model was used in [25]) with a single
core 1000 MHz Cortex A8 processor, and 512 MB RAM.
It was loaded with the CyanogenMod 7 custom ROM based
on Android 2.3. The phone’s 802.11g/n chipset (Broadcom
BCM 4329) is also used in several other smartphones from
manufacturers like Samsung, Apple, HTC, and Motorola. It
supports FA/BA, short guard interval, and PHY bitrates in
the range of 6.5-72.2Mbps (MCS 0-7) [27].
Note that the Android driver does not allow the user to
configure any 802.11 transmission (Tx) parameters (e.g., fix
the MCS, disable FA/BA, etc.). Hence, in this work, we
focus on the receive (Rx) energy consumption on the phone,
similar to in [10], [21]. We also believe that this is of more
interest to app developers, since in most applications WiFi
traffic flows from the AP to the client; the measurement
study in [21] based on a trace from 3785 smartphone users
from 145 countries over a 4-month period shows that the
ratio of downloaded data to uploaded data over WiFi is 20:1.
We measure power consumption on the phone using the
Monsoon Power Monitor [29], following the methodology
in [18], [13], [20], [25], [24]. The measurements were taken
with the screen off, Bluetooth/GSM/3G radios disabled, and
minimal background application activity. This background
activity causes a small base power consumption, which we
subtract from the measured power. The Monsoon Power
Monitor measures the total power consumption and cannot
provide a per-component, per-state, or per-packet break-
down. Hence, our measurements include the idle power con-
sumption between packet receptions (e.g., sender backoff,
carrier sensing, DIFS/SIFS), the transmit power consump-
tion of 802.11 and TCP ACKs, and any CPU processing
power. We consider these components part of the active
power consumption similar to in [18], [13], [20], [25], [24].
We use iperf [30] to generate traffic for our measurement
study. To confirm that iperf does not result in additional
energy consumption, we monitored the CPU usage by the
iperf application using the adb logs [31]. We found that the
CPU usage was negligible and this was true for different
versions of the application. Thus, the total energy measured
with the power monitor (after subtracting the base power)
can be attributed to the energy consumed by the WNIC.
IV. C
ANDIDATE MODEL INPUT PARAMETERS
In this section, through controlled experiments, we expose
the limitations of a set of candidate model input parameters
readily available to app developers. Most of these parame-
ters have been used in previously proposed models. These
limitations motivate the use of application-layer throughput
as the input parameter.
Constant active power The simplest WiFi energy consump-
tion model is one that assumes constant power states [19],
[13], [18]. Figure 1(a) plots the Cumulative Distribution
Function (CDF) of the active Rx power consumption of
a trace of 1640 power values collected over four links of
different quality. The experiment was repeated for all eight
802.11n bitrates, with 1470B TCP and UDP packets, with
and without FA/BA. We observe that the power consumption
varies from 240-795 mW with a median value of 434 mW.
The highest power value is more than 3 times larger than
the lowest one.
Transfer size We selected three links in our testbed, of
high, moderate, and low signal strength, and downloaded
a 10 MB, 50 MB, and 100 MB file over TCP with and
without FA/BA. Each experiment was repeated 15 times.
The use of rate adaptation resulted in different bitrates and
hence in different download times for different experiments.
Figure 1(b) plots the measured energy values for each file
size and the estimated energy (constant per file size) using
the model from [10], which we rebuilt for the Nexus S phone
and our own training dataset, described in Section V-A. We
observe that the estimated energy with the model can be
several times higher or lower than the actual value.

(a) Constant active power. (b) Transfer size. (c) Loss rate. (d) RSSI.
Figure 1. Limitations of candidate input parameters.
Loss Rate Packet loss rate can easily be measured at the
application layer and it is a good measure of channel quality
at a given MCS. However, a lower loss rate does not
always lead to lower energy consumption [14]. Furthermore,
a loss-based energy model ignores sender-side wireless
interference, which can increase energy consumption even
under zero packet loss. As an example, Figure 1(c) plots
the energy consumption against the loss rate, for a dataset
including measurements with 1470B TCP and UDP packets
over four links of varying quality, with rate adaptation,
with and without FA/BA. We observe that there is no clear
relationship between the energy consumption and the loss
rate; different measured energy values for the same loss rate
may vary by up to 10x.
Signal strength Similar to a packet loss-aware model, a
signal strength-aware model [21] cannot capture sender-
side interference. In addition, it fails to capture receiver-
side interference, i.e., collisions due to hidden terminals. In
Figure 1(d), we plot the power consumption against RSSI for
a dataset including measurements with 1470B UDP packets
over 20 links, with rate adaptation. We observe that there is
no clear relationship between RSSI and power consumption.
Different measured power values for the same RSSI value
may differ by more than 2x.
V. T
HROUGHPUT-BASED ENERGY MODELS
In the remaining of the paper we focus on the application
layer throughput as the input parameter. We conduct an
extensive study of the properties and the accuracy of the
linear model of power vs. throughput proposed in [20], [25],
[24]. None of these works provides detailed information on
the training dataset (e.g., number of measurement samples,
link conditions, WiFi parameters, etc.). In addition, [20],
[25] used different phones and we do not know whether
a model built for a given device offers the same accuracy
when tested with different devices (we briefly examine this
issue in Section VII-F). Hence, we rebuild the linear model
for the Nexus S phone using our own training dataset.
A. Training dataset
We selected 8 links of varying signal strength levels
by keeping the smartphone at a fixed location and using
testbed nodes located in different offices as senders. All
our measurements were conducted at night to avoid any
interference. We conducted measurements of the application
layer throughput and the Rx power consumption for all
the supported 802.11 g/n bitrates. We built separate models
for the two WiFi standards and we found that they exhibit
similar properties. In the remaining of the paper, we focus
on 802.11n, and we omit the discussion on 802.11g due to
space limitation.
Each measurement involves a 10-second iperf session
during which the sender sends TCP or UDP traffic to
the phone at full speed. We repeated each UDP exper-
iment for 4 different packet sizes: 100B, 700B, 1470B,
and 1470B with FA/BA
2
. For the TCP experiments, we
only used a packet size of 1470B with/without FA/BA,
since TCP transfers typically use large packets. For each
𝑡𝑟𝑎𝑛𝑠𝑝𝑜𝑟𝑡 𝑝𝑟𝑜𝑡𝑜𝑐𝑜𝑙, 𝑝𝑎𝑐𝑘𝑒𝑡 𝑠𝑖𝑧𝑒, 𝑀𝐶𝑆 setting, we took
15 measurements. In total, we collected around 6,000
throughput/energy samples.
Although [20], [25] build a linear model of the power (𝑃 )
as a function of throughput (𝑇ℎ) of the form 𝑃 = 𝛽 𝑇ℎ+𝛼
(1), we preferred an equivalent model of the per bit energy
consumption (𝐸
𝑏
)oftheform𝐸
𝑏
= 𝑎𝑇ℎ
1
+𝛽 (2). The per
bit energy consumption (in nJ/bit) is calculated as the power
consumption divided by throughput. This model can be
directly used to calculate the total energy consumption for a
given data transfer size without considering the downloading
time.
Figure 2(a) plots the 802.11n per bit Rx energy con-
sumption against the Rx application layer throughput for
all 6 𝑡𝑟𝑎𝑛𝑠𝑝𝑜𝑟𝑡 𝑝𝑟𝑜𝑡𝑜𝑐𝑜𝑙, 𝑝𝑎𝑐𝑘𝑒𝑡 𝑠𝑖𝑧𝑒 settings. In contrast
to [20], which built the model for a very small range of
throughputs (0-2 Mbps), the throughput values in our train-
ing dataset span the whole range of achievable throughputs
for the MCS set supported by the phone (0.12-44 Mbps) with
a median value of 6.7 Mbps and an average value of 7.79
Mbps. The corresponding energy per bit values range from
13.97-1902.11 nJ/bit with a median value of 60.15 nJ/bit
and an average value of 131.94 nJ/bit.
From Figure 2(a), as well as from Figures 3(a), 3(c), 4(a),
5(a), 5(c), 5(e), which plot parts of the training dataset corre-
2
Although packet-size distribution in the Internet is bimodal with packets
of around 1500B or smaller than 100B [32], for completeness we also
consider an intermediate size of 700B.

(a) Energy per bit vs. throughput: Training
dataset and model.
(b) Absolute error vs. throughput. (c) Relative error vs. throughput.
Figure 2. Error analysis of the Universal model.
Table I
ENERGY-THROUGHPUT MODEL PARAMETERS
Model Types Parameter 𝑎 Parameter 𝑏
Universal 305.3 13.1
Protocol
TCP 229.4 23.5
UDP 311.2 10.1
Packet
1470B FA/BA 228.0 19.1
1470B 214.7 23.2
700B 199.0 38.5
100B 197.3 258.7
Packet/Protocol
TCP-1470B FA/BA 258.8 20.3
TCP-1470B 210.0 28.0
UDP-1470B FA/BA 216.8 15.4
UDP-1470B 207.4 20.0
UDP-700B 199.0 38.5
UDP-100B 197.3 258.7
sponding to different settings, we observe a clear monotonic
relationship between the energy per bit and the throughput.
This result confirms the superiority of throughput as an input
parameter to an energy/power model compared to the other
candidate parameters we examined in Section IV.
B. Energy models
For an in-depth study of the properties and potential
limitations of throughput-based energy modeling under dif-
ferent settings, we explore four different options for building
a WiFi active energy model based on equation (2) by
considering the fundamental tradeoff between complexity
and accuracy. By complexity here we mean the number of
different equations of the form (2) required to describe the
relationship between energy and throughput.
Universal model: This is the simplest and most generic
model requiring only knowledge of the throughput. It uses
only one equation regardless of the setting.
Protocol model: We use two equations, one for each of the
two transport protocols, i.e., TCP and UDP. The motivation
for this comes from the inherent differences between the two
protocols which can result in different energy consumption
for the same transfer size. For example, TCP’s reaction
to loss may create more idle intervals between packet
receptions. In addition, TCP ACKs increase the total energy
consumption.
Packet model: Another option is to use different equations
for different packet sizes. The primary motivation for this
comes from a recent measurement study [16] showing that
the Rx energy per bit in the Nexus S phone can differ by
up to an order of magnitude for different packet sizes.
Packet/Protocol model: We also explore the case of devel-
oping a per-packet and per-protocol model, i.e., using a dif-
ferent equation for each 𝑡𝑟𝑎𝑛𝑠𝑝𝑜𝑟𝑡 𝑝𝑟𝑜𝑡𝑜𝑐𝑜𝑙, 𝑝𝑎𝑐𝑘𝑒𝑡 𝑠𝑖𝑧𝑒
setting. The accuracy of such a model is expected to be
the highest among the considered models at the cost of the
highest complexity.
We build the models with MATLAB’s Curve Fitting
Tool [33] using the Trust-Region algorithm. The parameters
of the four models for 802.11n are listed in Table I.
VI. E
RROR ANALYSIS
In this section, we evaluate the accuracy of the four
models using the training dataset.
A. Error metrics
We use two metrics to evaluate the accuracy of the models.
Our primary metric is the relative estimation error defined
as
𝐸𝑟𝑟
𝑟𝑒𝑙
=
𝐸𝑟𝑟
𝑎𝑏𝑠
𝐸
𝑏
=
𝐸
𝑏
𝐸
𝑏𝑒
𝐸
𝑏
where 𝐸
𝑏
is the measured energy per bit value and 𝐸
𝑏𝑒
is
the estimated value from the model. Occasionally, we also
use the absolute estimation error defined as
𝐸𝑟𝑟
𝑎𝑏𝑠
= 𝐸
𝑏
𝐸
𝑏𝑒
since the same relative error may be of varying significance
depending on the absolute energy values. E.g., a 50% error
may correspond to a measured value of 2 nJ/bit and an
estimated value of 1 nJ/bit or to a measured value of 2000
nJ/bit and an estimated value of 1000 nJ/bit; we consider
the latter case to be much worse.
For each model, we use the throughput values of the
training dataset as input to the set of equations describing the
model (Table I) and we estimate the Rx energy consumption
corresponding to each throughput value. We then compare
the estimated energy values with the measured values from
the training data set. We analyze the errors for each model

Citations
More filters
Proceedings ArticleDOI
24 Aug 2015
TL;DR: The power consumption in various states of the wireless interface, the impact of various features of 802.11n/ac (PHY bitrate, frame aggregation, channel bonding, MIMO), and the tradeoffs between these two metrics are investigated.
Abstract: This paper presents the first, to the best of our knowledge, detailed experimental study of 802.11n/ac throughput and power consumption in modern smartphones. We experiment with a variety of smartphones, supporting different subsets of 802.11n/ac features. We investigate the power consumption in various states of the wireless interface (sleep, idle, active), the impact of various features of 802.11n/ac (PHY bitrate, frame aggregation, channel bonding, MIMO) on both throughput and power consumption, and the tradeoffs between these two metrics. Some of our findings are significantly different from the findings of previous studies using 802.11n/ac wireless cards for laptop/desktop computers. We believe that these findings will help in understanding various performance and power consumption issues in today's smartphones and will guide the design of power optimization algorithms for the next generation of mobile devices.

59 citations

Journal ArticleDOI
25 Jun 2018
TL;DR: This paper proposes a new paradigm in designing and realizing energy efficient wireless indoor access networks, namely, a hybrid system enabled by traditional RF access, such as WiFi, as well as the emerging visible light communication (VLC), and proposes a randomized online algorithm that achieves a competitive ratio.
Abstract: In this paper, we propose a new paradigm in designing and realizing energy efficient wireless indoor access networks, namely, a hybrid system enabled by traditional RF access, such as WiFi, as well as the emerging visible light communication (VLC). VLC facilitates the great advantage of being able to jointly perform illumination and communications, and little extra power beyond illumination is required to empower communications, thus rendering wireless access with almost zero power consumption. On the other hand, when illumination is not required from the light source, the energy consumed by VLC could be more than that consumed by the RF. By capitalizing on the above properties, the proposed hybrid RF-VLC system is more energy efficient and more adaptive to the illumination conditions than the individual VLC or RF systems. To demonstrate the viability of the proposed system, we first formulate the problem of minimizing the power consumption of the hybrid RF-VLC system while satisfying the users requests and maintaining acceptable level of illumination, which is NP-complete. Therefore, we divide the problem into two subproblems. In the first subproblems, we determine the set of VLC access points (AP) that needs to be turned on to satisfy the illumination requirements. Given this set, we turn our attention to satisfying the users’ requests for real-time communications, and we propose a randomized online algorithm that, against an oblivious adversary, achieves a competitive ratio of ${\log (N)\log (M)}$ with probability of success $ {(1-({1}/{N}))}$ , where ${N}$ is the number of users and ${M}$ is the number of VLC and RF APs. We also show that the best online algorithm to solve this problem can achieve a competitive ratio of $ {\log } {M}$ . Simulation results further demonstrate the advantages of the hybrid system.

54 citations


Cites methods or result from "Modeling WiFi Active Power/Energy C..."

  • ...Other experimental studies for WiFi also validate this model [45]–[48]....

    [...]

  • ...Here, the power additive model is utilized, which is validated by preliminary experimental results and by [45]–[48]....

    [...]

Proceedings ArticleDOI
12 Jun 2017
TL;DR: This work evaluates the power and energy consumption of two standard-compliant 60 GHz wireless adapters in different operating states and under a number of different configurations, and compares their results against 802.11ac and discusses power-performance tradeoffs for the two technologies.
Abstract: The millimeter-wave technology is emerging as an alternative to legacy 2.4/5 GHz WiFi, offering multi-Gigabit throughput. While a lot of attention has been paid recently to analyzing the performance of the 60 GHz technology and adapting it for indoor WLAN usage, the power consumption aspect has largely been neglected. Given that mobile devices are the next target for 60 GHz, any discussion about this technology is incomplete without considering power consumption. In this work, we present the first, to our best knowledge, detailed study of the power consumption of 60 GHz commodity devices. We evaluate the power and energy consumption of two standard-compliant 60 GHz wireless adapters in different operating states and under a number of different configurations. We also compare our results against 802.11ac and discuss power-performance tradeoffs for the two technologies.

38 citations


Cites background from "Modeling WiFi Active Power/Energy C..."

  • ...11n/ac chipsets and smartphones over the past few years have shown that power increases with the PHY data rates [10], [11], [12] and channel width [13], [12], as well as with the application layer throughput [14], [15], [16], [17]....

    [...]

01 Jan 2012
TL;DR: The Curve Fitting Toolbox as discussed by the authors is a curve fitting toolbox that can be used to fit a variety of tools in a MATLAB MATLAB classifier, such as:
Abstract: 介绍了使用MATLAB曲线拟合工具箱(Curve Fitting Toolbox)对光电效应实验数据进行曲线拟合,并利用曲线拟合的结果计算得到了普朗克常数.结果显示,利用曲线拟合工具箱能方便、直观的对实验数据进行拟合,使得实验的数据处理过程方便、准确.

32 citations

Journal ArticleDOI
TL;DR: A categorization of fundamental aspects regarding computation offloading in heterogeneous cloud computing from the perspective of smartphone applications and state-of-the-art solutions for the identified categories is developed.

28 citations


Cites background from "Modeling WiFi Active Power/Energy C..."

  • ...[74] studied the properties of WiFi based on parameters that are readily available to smartphone applications, such as packet loss rate, signal strength, transfer...

    [...]

  • ...The models estimate the power consumption based on different network variables like transfer size [44], data rate [38], signal strength [69], throughput [41, 74], transmission time [75]....

    [...]

References
More filters
Proceedings ArticleDOI
22 Apr 2001
TL;DR: A series of experiments are described which obtained detailed measurements of the energy consumption of an IEEE 802.11 wireless network interface operating in an ad hoc networking environment, and some implications for protocol design and evaluation in ad hoc networks are discussed.
Abstract: Energy-aware design and evaluation of network protocols requires knowledge of the energy consumption behavior of actual wireless interfaces. But little practical information is available about the energy consumption behavior of well-known wireless network interfaces and device specifications do not provide information in a form that is helpful to protocol developers. This paper describes a series of experiments which obtained detailed measurements of the energy consumption of an IEEE 802.11 wireless network interface operating in an ad hoc networking environment. The data is presented as a collection of linear equations for calculating the energy consumed in sending, receiving and discarding broadcast and point-to-point data packets of various sizes. Some implications for protocol design and evaluation in ad hoc networks are discussed.

1,810 citations

Proceedings Article
23 Jun 2010
TL;DR: A detailed analysis of the power consumption of a recent mobile phone, the Openmoko Neo Freerunner, measuring not only overall system power, but the exact breakdown of power consumption by the device's main hardware components.
Abstract: Mobile consumer-electronics devices, especially phones, are powered from batteries which are limited in size and therefore capacity. This implies that managing energy well is paramount in such devices. Good energy management requires a good understanding of where and how the energy is used. To this end we present a detailed analysis of the power consumption of a recent mobile phone, the Openmoko Neo Freerunner. We measure not only overall system power, but the exact breakdown of power consumption by the device's main hardware components. We present this power breakdown for micro-benchmarks as well as for a number of realistic usage scenarios. These results are validated by overall power measurements of two other devices: the HTC Dream and Google Nexus One. We develop a power model of the Freerunner device and analyse the energy usage and battery lifetime under a number of usage patterns. We discuss the significance of the power drawn by various components, and identify the most promising areas to focus on for further improvements of power management. We also analyse the energy impact of dynamic voltage and frequency scaling of the device's application processor.

1,579 citations


"Modeling WiFi Active Power/Energy C..." refers background in this paper

  • ...A major fraction of the energy consumption in smartphones comes from the WiFi radio, which can account for more than 50% of the total device power budget under typical use [1], [2] and can quickly drain the phone’s battery when transmitting at high peak rates....

    [...]

Proceedings ArticleDOI
04 Nov 2009
TL;DR: TailEnder is developed, a protocol that reduces energy consumption of common mobile applications and aggressively prefetches several times more data and improves user-specified response times while consuming less energy.
Abstract: In this paper, we present a measurement study of the energy consumption characteristics of three widespread mobile networking technologies: 3G, GSM, and WiFi. We find that 3G and GSM incur a high tail energy overhead because of lingering in high power states after completing a transfer. Based on these measurements, we develop a model for the energy consumed by network activity for each technology.Using this model, we develop TailEnder, a protocol that reduces energy consumption of common mobile applications. For applications that can tolerate a small delay such as e-mail, TailEnder schedules transfers so as to minimize the cumulative energy consumed meeting user-specified deadlines. We show that the TailEnder scheduling algorithm is within a factor 2x of the optimal and show that any online algorithm can at best be within a factor 1.62x of the optimal. For applications like web search that can benefit from prefetching, TailEnder aggressively prefetches several times more data and improves user-specified response times while consuming less energy. We evaluate the benefits of TailEnder for three different case study applications - email, news feeds, and web search - based on real user logs and show significant reduction in energy consumption in each case. Experiments conducted on the mobile phone show that TailEnder can download 60% more news feed updates and download search results for more than 50% of web queries, compared to using the default policy.

1,239 citations


"Modeling WiFi Active Power/Energy C..." refers background or methods in this paper

  • ...The time to download a file of a given size depends on the channel conditions and the MAC/PHY protocol characteristics, which determine the PHY bitrate....

    [...]

  • ...On the other hand, a number of recent works develop power/energy models for different smartphone components (disk, CPU, display, cellular, WiFi) that can be easily used by app developers [11], [10], [13], [18], [19], [20], [21]....

    [...]

  • ...…important practical issues, such as how to collect a good training dataset, or how to reduce the complexity of the most accurate of the four models, and we show that a non-linear power-throughput model can further improve the accuracy compared to the linear model from [20], [25] (Section VIII)....

    [...]

  • ...This assumption is valid for idle/transition power states (e.g., sleep, ramp/promotion, and tail power states [10], [13]) and can also offer satisfactory accuracy for the active power state (during packet transmissions/receptions) in the case of low bitrates (e.g., 802.11b) and/or small data…...

    [...]

  • ...Since directly measuring WiFi power/energy consumption [5], [1], [6] is a non-trivial task, several recent works have focused on developing WiFi energy consumption models [7], [8], [9], [10], [11], [12], [13], [14], [15]....

    [...]

Proceedings ArticleDOI
24 Oct 2010
TL;DR: PowerBooter is an automated power model construction technique that uses built-in battery voltage sensors and knowledge of battery discharge behavior to monitor power consumption while explicitly controlling the power management and activity states of individual components.
Abstract: This paper describes PowerBooter, an automated power model construction technique that uses built-in battery voltage sensors and knowledge of battery discharge behavior to monitor power consumption while explicitly controlling the power management and activity states of individual components. It requires no external measurement equipment. We also describe PowerTutor, a component power management and activity state introspection based tool that uses the model generated by PowerBooter for online power estimation. PowerBooter is intended to make it quick and easy for application developers and end users to generate power models for new smartphone variants, which each have different power consumption properties and therefore require different power models. PowerTutor is intended to ease the design and selection of power efficient software for embedded systems. Combined, PowerBooter and PowerTutor have the goal of opening power modeling and analysis for more smartphone variants and their users.

1,225 citations


"Modeling WiFi Active Power/Energy C..." refers background or methods in this paper

  • ...The improvement is more prominent in the two high throughput scenarios (Sections VII-B, VII-E)....

    [...]

  • ...Our experiments in Sections VII-B, VII-E further confirm this finding for high throughput scenarios....

    [...]

  • ...In the case of WiFi, one common feature of most of these models [11], [12], [13], [18], [19] is the assumption of constant power consumption in each power state....

    [...]

  • ...We discuss some of these issues further in Sections VIII-B, VIII-C....

    [...]

  • ...…important practical issues, such as how to collect a good training dataset, or how to reduce the complexity of the most accurate of the four models, and we show that a non-linear power-throughput model can further improve the accuracy compared to the linear model from [20], [25] (Section VIII)....

    [...]

Proceedings ArticleDOI
25 Jun 2012
TL;DR: This paper develops the first empirically derived comprehensive power model of a commercial LTE network with less than 6% error rate and state transitions matching the specifications, and identifies that the performance bottleneck for web-based applications lies less in the network, compared to the previous study in 3G.
Abstract: With the recent advent of 4G LTE networks, there has been increasing interest to better understand the performance and power characteristics, compared with 3G/WiFi networks. In this paper, we take one of the first steps in this direction.Using a publicly deployed tool we designed for Android called 4GTest attracting more than 3000 users within 2 months and extensive local experiments, we study the network performance of LTE networks and compare with other types of mobile networks. We observe LTE generally has significantly higher downlink and uplink throughput than 3G and even WiFi, with a median value of 13Mbps and 6Mbps, respectively. We develop the first empirically derived comprehensive power model of a commercial LTE network with less than 6% error rate and state transitions matching the specifications. Using a comprehensive data set consisting of 5-month traces of 20 smartphone users, we carefully investigate the energy usage in 3G, LTE, and WiFi networks and evaluate the impact of configuring LTE-related parameters. Despite several new power saving improvements, we find that LTE is as much as 23 times less power efficient compared with WiFi, and even less power efficient than 3G, based on the user traces and the long high power tail is found to be a key contributor. In addition, we perform case studies of several popular applications on Android in LTE and identify that the performance bottleneck for web-based applications lies less in the network, compared to our previous study in 3G [24]. Instead, the device's processing power, despite the significant improvement compared to our analysis two years ago, becomes more of a bottleneck.

1,029 citations


"Modeling WiFi Active Power/Energy C..." refers background or methods in this paper

  • ...On the other hand, a number of recent works develop power/energy models for different smartphone components (disk, CPU, display, cellular, WiFi) that can be easily used by app developers [11], [10], [13], [18], [19], [20], [21]....

    [...]

  • ...As far as the PHY bitrate is concerned, several works have recently shown that it affects the energy consumption of an 802.11n WNIC [3], [14], [16]....

    [...]

  • ...…important practical issues, such as how to collect a good training dataset, or how to reduce the complexity of the most accurate of the four models, and we show that a non-linear power-throughput model can further improve the accuracy compared to the linear model from [20], [25] (Section VIII)....

    [...]

  • ...The motivation for this comes from the inherent differences between the two protocols which can result in different energy consumption for the same transfer size....

    [...]

  • ...A small number of recent works model the WiFi active 1MIMO is not supported by today’s smartphones but it may be supported in the near future [22]. energy consumption of a smartphone as a function of one or more input parameters [12], [10], [20], [21]....

    [...]

Frequently Asked Questions (9)
Q1. What have the authors contributed in "Modeling wifi active power/energy consumption in smartphones" ?

The authors conduct the first detailed measurement study of the properties of a class of WiFi active power/energy consumption models based on parameters readily available to smartphone app developers. The authors first consider a number of parameters used by previous models and show their limitations. 11n-equipped smartphone, the authors build four versions of a previously proposed linear power-throughput model, which allow us to explore the fundamental tradeoff between accuracy and simplicity. The authors study the properties of the model in relation to other parameters such as the packet size and/or the transport layer protocol, and they evaluate its accuracy under a variety of scenarios which have not been considered in previous studies. The authors further show that a non-linear model can greatly improve the accuracy in these two cases. 

a loss-based energy model ignores sender-side wireless interference, which can increase energy consumption even under zero packet loss. 

The time to download a file of a given size depends on the channel conditions and the MAC/PHY protocol characteristics, which determine the PHY bitrate. 

In the case of WiFi, one common feature of several of these models [11], [12], [13], [18], [19], regardless of their complexity, is the assumption of constant power consumption at each power state. 

Loss Rate Packet loss rate can easily be measured at the application layer and it is a good measure of channel quality at a given MCS. 

In [20], [25], a linear model of the active power consumption of a wireless interface (3G, 4G, WiFi) is proposed as a function of the application layer throughput. [24] builds a similar model of the energy consumption as a linear function of the packet transmission time (which can be converted to a linear model of power vs. throughput). 

Assume an app uses two packet sizes 𝑝1, 𝑝2 which give throughputs 𝑇ℎ1, 𝑇ℎ2, respectively, and total throughput 𝑇ℎ. The authors consider two methods. 

In Table II, 82-98% of the relative errors of the models for the settings without FA/BA have absolute values lower than 10% and 96-100% of the relative errors have absolute values lower than 20%. 

The authors observe that the model for 1470B packets can predict the energy consumption of both 1000B packets and 1470B with FA/BA with very high accuracy (86%/79% of the relative errors haveFigure 6.