Directional Antenna Diversity For Mobile Devices: Characterizations and Solutions

Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

Directional Antenna Diversity for Mobile Devices:

Characterizations and Solutions


Ardalan Amiri Sani, Lin Zhong, and Ashutosh Sabharwal
Department of Electrical and Computer Engineering, Rice University

{ardalan, lzhong, ashu}@rice.edu


ABSTRACT
We report a first-of-its-kind realization of directional transmission for smartphone-like mobile devices using multiple passive
directional antennas, supported by only one RF chain. The key is
a multi-antenna system (MiDAS) and its antenna selection methods that judiciously select the right antenna for transmission. It
is grounded by two measurement-driven studies regarding 1) how
smartphones rotate during wireless usage in the field and 2) how
orientation and rotation impact the performance of directional
antennas under various propagation environments.
We implement MiDAS and its antenna selection methods using
the WARP platform. We evaluate the implementation using a
computerized motor to rotate the prototype according to traces
collected from smartphone users in the field. Our evaluation
shows that MiDAS achieves a median of 3dB increase in link
gain. We demonstrate that rate adaptation and power control can
be combined with MiDAS to further improve goodput and power
saving. Real-time experiments with the prototype show that the
link gain translates to 85% goodput improvement for a low SNR
scenario. The same gain translates to 51% transmit power reduction for a high SNR scenario. Compared to other methods in realizing directional communication, MiDAS does not require any
changes to the network infrastructure, and is therefore suitable for
immediate or near-future deployment.

Categories and Subject Descriptors


C.2.1 [Computer-Communication Networks]: Network Architecture and Design - Wireless Communication

General Terms

mobile health monitoring, and video chat, are increasingly employing the uplink. Therefore, the mentioned interference has
become a key bottleneck in the capacity of infrastructure networks with a large number of mobile clients. Similarly, the power
waste also makes wireless transmission one of the most powerhungry activities on mobile devices, contributing a critical barrier
to the usability and wider adoption of mobile Internet. Furthermore, devices that are far away from access points or base stations, also suffer from low uplink data rates since most of their
limited transmission power is radiated to undesirable directions.
Directional transmission can alleviate these problems by focusing
the radiation toward the intended receiver. Beamforming employs
multiple RF chains to achieve directionality. However, its cost
and power overhead has prevented it from being immediately
used on smartphone-like mobile devices. In this work, we study a
much more efficient way of realizing directionality: miniature
passive directional antennas. Many authors use directional antenna to refer to beamforming based on smart antennas. In this
work, we use antenna to refer to the passive antenna without
the RF chain. By placing multiple directional antennas on the
surface of a mobile device, the device can opportunistically select
one for directional transmission without adding RF circuitry, in
contrast to the simultaneous use of multiple RF chains by beamforming. However, because only one antenna can be used at a
time, the device will not be able to find out the best antenna easily, especially when it moves and rotates.
The goal of this work is to realize the benefits of directional antennas for mobile devices. Toward this goal, we experimentally
answer three questions.

Current mobile devices, such as smartphones and laptops, are


omni directional and radiate power to all directions for uplink
transmission. Omni directionality not only introduces interference
between peers, but also leads to power waste. Emerging smartphone applications, such as participatory sensing, media sharing,

First, how do smartphone-like mobile device rotate during wireless access? By using a directional antenna, the device orientation
becomes critical because a mobile device can rotate and the rotation changes device direction much faster than mobility does. We
collect accelerometer and compass readings along with network
usage information from 11 smartphone users, each for one week
in the field. From such field-collected traces, we are able to estimate the orientation and rotation of the smartphones during wireless usage. We show that smartphones rotate relatively slowly;
compared to how fast packets are exchanged during wireless
communication. Moreover, the orientation is quite predictable in
short intervals. We report the characterization in Section 3.

Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
MobiCom10, September 2024, 2010, Chicago, Illinois, USA.
Copyright 2010 ACM 978-1-4503-0181-7/10/09...$10.00.

Second, how do directional antennas behave with indoor and


non-line-of-sight (NLOS) propagations? While recent work has
shown directional communication can work well with stationary
nodes for indoor and NLOS propagation [1, 2], it is still unclear
how orientation and rotation would impact the directional channels. Therefore, using a computerized motor platform, we measure the Received Signal Strength Indicator (RSSI) of directional
antennas for indoor and NLOS environments, with not only con-

Algorithm, Design, Experimentation, Measurement

Keywords
Passive directional antennas, mobile devices, orientation estimation, smartphone rotation

1. INTRODUCTION

trolled orientations, but also rotation according to the fieldcollected traces. We show that directional antennas outperform
omni ones for a considerable range of orientations even in NLOS
indoor environments. More importantly, the directional channels
are highly reciprocal for 802.11-like frequency bands, and their
performance is quite predictable in short intervals even under
realistic rotations. We report the characterization in Section 4.
Finally, how can a device dynamically select the best antenna?
Leveraging the discoveries from the two characterizations, we
design a multi-antenna system (MiDAS) that consists of an omni
directional antenna, one or more directional antennas, and an
antenna switch. MiDAS works with existing mobile devices that
usually have a single RF chain and can use only one antenna at a
time. We further provide two antenna selection methods for MiDAS. The packet-based method uses one packet to assess an antenna without any changes to the network infrastructure. The
symbol-based method uses the PHY training symbols so that all
antennas are assessed with a single packet. It is much more efficient than the packet-based selection but requires changes to the
PHY layer. We report the design of MiDAS in Section 5 and the
two antenna selection methods in Sections 6 and 7, respectively.
We implement MiDAS and its antenna selection methods on the
WARP platform from Rice University and evaluate them with the
computerized motor platform that replays the field-collected rotation traces from smartphone users. Our evaluation shows that
with three 5dBi directional antennas placed on the surfaces of a
device, MiDAS can achieve a median gain close to 3dB, without
any changes to the network infrastructure.
To fully realize the benefits of the link gain of MiDAS, we incorporate rate adaptation and power control mechanisms into our
system. Using real-time experiments, we demonstrate that MiDAS can improve the link goodput by 85% when the SNR is low.
Also, when the SNR is high, it can save 51% of the transmit power, while increasing the goodput by 7%. We report the prototype
and evaluation in Section 8.
To the best of our knowledge, this work represents the first publicly available characterizations of the rotation of smartphone-like
devices and its impact on the performance of directional antennas.
MiDAS and its antenna selection methods are also the first reported directionality implementation for a device that cannot only
move but can also rotate. We will discuss related work in Section
9.
Making smartphone-like devices directional is a radical departure
from existing and emerging wireless technologies. It provides an
inexpensive and immediately deployable solution to improve
network capacity and device efficiency. While we demonstrate
the effectiveness of MiDAS in improving link goodput and device power efficiency, more network support is needed to fully
realize its potential in improving network capacity. From this
perspective, our work complements existing work on directional
MAC protocol design, e.g. [3].

2. DIRECTIONAL ANTENNAS
We are motivated by the commercial availability of miniature
passive directional antennas that can be employed on mobile
devices. Microstrip antennas are good examples. The microstrip
antennas used in our prototype have a patch area of 3.23.2 cm2
[4]. This allows the placement of these antennas on front and
back sides of smartphones. Also theoretically, the width of the
patch can be even smaller while maintaining 5dBi or 8dBi peak

gain. This will enable thin directional microstrip antennas to be


placed on other sides of smartphones.
Moreover, recent advances in multiple reconfigurable antennas
have produced small form factor antennas with a few directional
beams in different directions [5, 6]. Also, sectorized antennas [7]
are another option for having multiple beams on a mobile device.
But, due to their larger size, they can be used for bigger mobile
devices, such as iPad, Kindle, or netbooks.
Unlike digital beamforming, the passive directional antennas
discussed above produce a directional radiation pattern without
extra circuitry or power. With a directional radiation pattern
pointed at the right direction, a mobile client can use reduced
transmit power to deliver a required receiver signal strength
(RSS), or it can increase the RSS with the same transmit power.
In both cases, the clients interference to its peers is reduced.
The key challenge to the use of directional antennas on mobile
devices is that a mobile device can change its orientation through
mobility and rotation. Since one directional antenna can only
provide adequate gain for a limited range of orientations, multiple
antennas should be placed around the device so that they collectively provide a much larger range of orientations in which at
least one of them provides adequate gain.
Because directional antennas have never been studied for smartphone-like mobile devices before, the key questions regarding
their feasibility naturally arises:
Is it possible to track the right antenna when a mobile
device can not only move but can also rotate?
We next answer this question experimentally in three steps.

3. CHARACTERIZING SMARTPHONE
ROTATION
We first characterize the orientation and rotation of smartphones
during wireless access with data collected from field usage. It is
important to note that the relative direction of a device with its
access point (base station) is determined by not only device orientation but also location. We do not include location in our characterization because mobile client location and its changes have
been extensively studied as mobility. Moreover, change in relative direction due to mobility is much slower than that due to
rotation thanks to the large distance between a mobile device and
its access point or base station.

3.1 Smartphone Orientation Estimation


We have collected accelerometer and compass readings from 11
smartphone users, each for one week along with both voice and
data usage. The average usage time for a participant is 25.9 hours.
See Appendix A for the trace collection details. The traces will
also be used in the characterization and evaluation reported later
in this submission. We plan to make the traces open-access.
We are able to derive the smartphone orientation, represented by
three Euler angles [8] under the ZYX convention. According to
Euler Theorem, the orientation of a rigid body can be uniquely
defined by three angles, also known as Euler angles. These three
Euler angles are shown as , , and in Figure 1. The X, Y and Z
axes are the coordinates of the device, meaning that they are fixed
to the body of the device and rotate with it. The x, y and z axes
(lower case) represent the axes of the earth. The z axis points to
the sky and is perpendicular to the ground. The y axis is parallel

z Sky
1

Device X axis

Device Z axis

y Magnetic

North

Ground

Figure 1: Euler angles in the ZYX convention used in our


characterization of smartphone orientation and rotation
to the ground and points to the magnetic north. The x axis is also
parallel to the ground and is orthogonal to z and y.
Calculating the three Euler angles using a tri-axis accelerometer
and a tri-axis compass is straightforward in theory. Because the
accelerometer is less subject to external interference, we use the
accelerometer and the gravity to derive two of the angles, and ,
similar to [9]; we only use the compass for , which is impossible
to calculate with the accelerometer only. In Appendix B, we provide the details regarding how we overcome the practical challenges in orientation estimation.

3.2 Rotation Speed


We next examine how a smartphone changes its orientation, or
rotates, when wireless is used. We compute the rotation speed as
a vector including three elements, the rates of change of the Euler
angles: , , and . We calculate the rotation speed for each of
the Euler angles for different time intervals, 0.1s, 1s, and 10s, for
the collected traces. Figure 2 shows the distributions of the rotation speed for each Euler angle. We also find that rotation speeds
are below 120 /second for 90% of time. Such low rotation
speeds indicate that the use of directional antennas can be feasible
because they outperform the omni directional antenna by several
tens of degrees, as we will show in Section 4.
Furthermore, the figures show that the rotation distributions are
not the same for the three Euler angles. , the angle around the z
axis, experiences much faster rotation than the other two angles.
Therefore, we will focus on the impact of rotation in in the next
few sections.

3.3 Predictability of Orientation


We examine how accurate one can predict the device orientation
in the future, given the past orientations. Let sn denotes the nth
sample of the orientation, measured at time tn. Prediction is to
determine sn+1 at tn+1, given ti and si for i n. We study two very
simple prediction methods.

Zero order prediction: sn+1 = sn, and

First order prediction: sn+1 = sn + (sn-sn-1)(tn+1-tn)/(tn-tn-1),


which predicts using linear extrapolation.

We evaluate both of the methods on the three Euler angles in


predicting for 10ms, 100ms, 1s, and 10s ahead, respectively. Figure 3 summarizes the results. It shows that for predictions of
10ms and 100ms, median errors are below 0.2, 0.2, and 3.0 for
, , and , respectively. This demonstrates that a phone orientation is quite predictable in short terms, thanks to the continuous
nature of human movement.

0.5
0

0.1s
1s
10s

0.5

0.5

0
-4

10

-2

10

10

Rotational speed(/s)

0.1s
1s
10s

0
-4

10

10

-2

10

10

-4

10

Rotational speed(/s)

10

-2

10

10

10

Rotational speed(/s)

Figure 2: Cumulative Distribution Function (CDF) of rotation speed for three Euler angles, calculated for 0.1s, 1s,
and 10s intervals

Zero order
First order

()

Device Y axis Y

0.1s
1s
10s

100
1
0.01
10ms 100ms

1s

10s

Prediction Interv.(s)

Zero order
First order

Zero order
First order
100

100
1

0.01

0.01

10ms 100ms

1s

10s

Prediction Interv.(s)

10ms 100ms

1s

10s

Prediction Interv.(s)

Figure 3: Error() of orientation prediction from 10ms to 10s

4. CHARACTERIZING DIRECTIONAL
PROPAGATION
In this section, we report experimental characterization of the
quality of directional links under various propagation environments, orientation, and rotation. While a few recent works have
reported NLOS indoor performance of directional links [1, 2], no
one has characterized NLOS indoor directional propagation under
device rotation.

4.1 Experimental Setup


We employ two WARP nodes for our experiments. Like most
commercial wireless interfaces, the RSSI of the WARP radio card
can be read from software. The RSSI values are then translated to
RSS values according to the datasheet [10]. We report our results
in RSS format in dBm. Considering that the noise floor is approximately constant over time in our hardware, RSS and SNR will
just be an additive constant different. Therefore, all RSS results
can be easily translated to SNR. Also, maximizing received RSS
is equivalent to maximizing received SNR.
We developed a computerized motor platform to rotate one
WARP node and its companion laptop in order to emulate the
orientation and rotation of a mobile client around the z axis, or .
As characterized in Section 3, sees the most rotation among the
three Euler angles. Also, because we mount the directional antennas perpendicular to the z axis, rotation around the z axis is the
most challenging in terms of directional channel changes, as illustrated by Figure 4. Therefore, we believe the evaluation with the
motor platform will reveal key insights into the feasibility of directional antennas.
Two directional antennas of 5dBi [4] and 8dBi [11] with 85 and
75 half-power horizontal beamwidth respectively, and a closeto-omni directional antenna [12] are used and characterized on
this motor-driven WARP node.
A second WARP node with an omni directional antenna is used
as the access point (AP). During the experiments, the motordriven client transmits 802.11-like packets to the AP continuously
and as fast as possible. The AP node records the RSSI for every

NLOS ind. / 5dBi antenna


-30
-40
-50
-60
0

NLOS ind. / 8dBi antenna


-20

Dir-Client
Dir-AP
Omni-Client
Omni-AP

RSS(dBm)

RSS(dBm)

-20

-30
-40
-50
-60
0

60 120 180 240 300 360

LOS ind. / 5dBi antenna

data packet, and the motor-driven node records the RSSI for
every ACK packet sent back from the AP.
All indoor experiments are carried out on the third floor of Duncan Hall, an office building at Rice University. Outdoor experiments are carried out in an open space outside Duncan Hall. During the experiments, the client and AP nodes are separated as
allowed by the WARP nodes, ~15 meters in most cases. For indoor, experiments are carried out at various locations in the building with and without light-of-sight propagation paths, or LOS
indoor and NLOS indoor. The distance between the AP and client
in our indoor experiments is comparable to such distance in an
enterprise 802.11 network. Indoor environments have been
known to be much more challenging for directional communication than outdoor LOS propagation.

4.2 Impact of Orientation


By orienting the motor-driven node toward different directions,
we measure the RSSI of directional antennas every 10 in a full
circle (360). We repeat the same experiment in LOS and NLOS
indoor environments at various locations. Figure 5 presents traces
of the RSSI measured at the AP for the omni-directional, 5dBi
and 8dBi directional antennas at two representative locations. We
make the following observations.
First, the RSSI pattern of a directional antenna may significantly
deviate from the antenna gain pattern. This is especially true in
NLOS environments. Such deviation is largely due to the rich
multipath effects.
Second, the RSSI pattern is largely continuous. With the continuity in device rotation and orientation predictability, such RSSI
continuity is likely to lead to high predictability in RSSI in real
usage, as will be seen in Section 4.5.
Third, the directional antenna outperforms the omni directional
one over a considerable range of orientation but is significantly
worse over some other range. This highlights the potential benefit
and risk of using a directional antenna.
Finally, the gain of the directional antenna impacts such benefit
and risk. The higher the gain (8dBi vs. 5dBi), the higher the benefit, and the higher the risk. Higher gain antennas have higher peak
gains (benefit), but they have deeper valleys and smaller ranges in
which the directional antenna is better (risk), as can be seen in
Figure 5. In Section 8, we will show how such effects affect the
performance of MiDAS.

-40

Dir-Client
Dir-AP
Omni-Client
Omni-AP

-50
-60
0

LOS ind. / 8dBi antenna


-20

RSS(dBm)

RSS(dBm)

Figure 4: Rotation platform for directional channel characterization. The motor can replay field-collected traces or
orient the WARP board toward any given direction in the
azimuth plane

-30

60 120 180 240 300 360

Direction()

Direction()
-20

Dir-Client
Dir-AP
Omni-Client
Omni-AP

60 120 180 240 300 360

Direction()

-30
-40

Dir-Client
Dir-AP
Omni-Client
Omni-AP

-50
-60
0

60 120 180 240 300 360

Direction()

Figure 5: Directional antenna patterns in NLOS and LOS


indoor environments

4.3 Channel Reciprocity


By comparing the RSSI of a data packet at the AP against the
RSSI of the packets ACK at the motor-driven client, we examine
how reciprocal the channel is for directional antennas. Figure 5
presents both the data packet RSSI and that of its ACK for both
the directional and the omni directional antenna. While there is an
almost constant gap between the RSSI at the mobile client and
that at the AP due to their hardware personalities, their RSSI track
each other closely in their changes. This suggests that the channel
for a directional antenna can be considered reciprocal. Note that
the constant gap between these two measurements is mainly due
to different radio characteristics. In Section 6, we will leverage
this to design a packet-based antenna selection method that selects the antenna for transmission based on RSSI of received
packets. It is worth noting that the reciprocity of 802.11 channels
of omni directional antennas has been shown by others, e.g. [13].

4.4 Superiority of Directional Antennas


An important question about the use of directional antennas on a
mobile device is how long a directional antenna can remain better
than the omni one. To answer this question, we randomly select
~6 minutes of rotation traces from each of the 11 participants
(one hour in total), and replay them on the computerized motor
platform by changing accordingly. During the replay, the mobile client sends packets continuously. In each packet, one directional antenna and one omni antenna take turn and transmit 128
training symbols each. The AP records the RSSI measured for all
packets.
By analyzing the entire RSSI traces, we can calculate the time
intervals in which the directional antenna is always better than the
omni directional one. We call these intervals the superiority intervals of the directional antenna. We can also calculate the average RSS gain over the omni directional antennas during such
intervals. Figure 6 shows the distribution of superiority intervals
in terms of the percentage of the total replay time. It shows that in
a NLOS indoor environment, 53% and 40% of the replay time is
spent in the superiority intervals over 1s for the 5dBi and 8dBi
antenna, respectively. That is, should the directional antenna be
selected, it can provide a better performance over than one second
for close to half of the time. The average gain over the omni di-

5dBi

30

5dBi

100

10

20
15
10

Error(dB)

15

Error(dB)

20

8dBi

100

Zero order
First order

25

total time(%)

total time(%)

25

0.01

Zero order
First order

0.01

5
0

8dBi

30

[0,0.1)

[0.1,1)

[1,10)

[10,inf)

superiority intervals(s)

[0,0.1)

[0.1,1)

[1,10)

[10,inf)

superiority intervals(s)

Figure 6: Distribution of superiority intervals in terms of


the percentage of the total replay time in NLOS indoor

10ms

100ms

5. MULTI-ANTENNA DESIGN (MiDAS)


The characterizations presented in Sections 3 and 4 suggest the
feasibility and the potential benefits of using directional antennas
on mobile devices. Based on the findings, we present a multiantenna design (MiDAS) that can be immediately deployed to
mobile devices with a single RF chain. Figure 8 illustrates the
design. We next describe several design decisions and the rationale underneath them.
First, the design requires only a single RF chain. Therefore, at a
given time, only one antenna can be used or assessed. This makes
the design immediately deployable to power and cost-constrained
mobile devices that only have a single RF chain, such as wireless
handsets.
Second, the design consists of one omni directional antenna and
one or more directional antennas. The omni directional antenna
is included for not only standard-compliance but also as a safety
belt when a good directional antenna is difficult to identify. At a
given time, only one antenna is selected by the transceiver
through an antenna switch. This eliminates much of the requirements on antenna spacing by MIMO and beamforming technologies, allowing MiDAS to fit in small form-factor devices.
Third, the design considers the directional antennas only for data
transmission and receiving their acknowledgments, which are
used for antenna assessment. It uses the omni directional antenna
for other purposes, e.g. for idle listening, control packets, management packets, and data reception, or for data transmission in
case a better directional antenna is not identified. The reasons for
these design decisions are 1) A directional antenna improves the

100ms

1s

Prediction Intervals(s)

Omni-directional antenna

Antenna switch

Transceiver
Directional
antennas

...

By utilizing the RSSI traces collected in the previous section, we


study how predictable the RSS from directional antennas is during real usage. We examine the zero order and first order prediction methods as discussed in Section 3.3. The prediction algorithms are performed on the very lightly smoothed RSSI traces, to
reduce the effect of high frequency noise in RSSI readings. Figure 7 presents the prediction errors. Results show the median RSS
prediction error is no more than 0.1dB for intervals of 10ms and
100ms. Similar to rotation prediction, our results show that RSS
is very predictable in short intervals. Therefore, we will use this
high accuracy of RSS predictions in designing our antenna selection algorithms, which will be discussed in sections 6 and 7.

10ms

Figure 7: RSS prediction error for the 5dBi and 8dBi directional antennas for 10ms to 10s intervals

rectional antenna of these superiority intervals is 2.8dB and 2.5dB


for 5dBi and 8dBi antennas, respectively. This highlights the
potential benefits of using the directional antenna when it is properly selected. It also suggests a more directional antenna may not
necessarily be better.

4.5 Predictability of RSS

1s

Prediction Intervals(s)

RSSI
Antenna selection

Figure 8: Multi-antenna system with one omni directional


antenna and one or more directional antennas that works
with existing wireless interfaces with a single transceiver
device energy efficiency much more in transmission than reception; and 2) listening with directional antennas increases the deafness of the nodes. This is particularly important to CSMA-based
802.11 because the mobile client cannot accurately predict the
arrival of incoming packets.
Key Research Question: With the multi-antenna system described above, the key question to be answered is: which antenna
to use for transmission? In the next two sections, we answer this
question with two antenna selection methods. The first one assesses antennas by packets, and therefore does not require any
changes to the deployed network infrastructure. The second one
assesses antennas by physical layer symbols, which is much more
efficient, but requires changes to the PHY layer.

6. SELECTION FOR LEGACY NETWORKS


We first focus on mobile clients of legacy infrastructure networks
such as deployed 802.11 and cellular networks. To work with
legacy networks, the client must select the best antenna without
any support from the infrastructure.

6.1 Packet-based Antenna Assessment


Without any changes to the deployed infrastructure, a mobile
client can only assess one antenna per packet by measuring the
average RSSI of certain bytes of the packet transmitted by that
antenna and leveraging the channel reciprocity to estimate the
receiver RSSI. Fortunately, as observed by us in Section 4.3 and
by others [13], 802.11 channels often see good reciprocity.
To assess an antenna, the client transmits a packet and receives
the ACK packet with the same antenna. To assess N antennas, the
client has to accomplish N Data-ACK exchanges. In 802.11, each
Data-ACK exchange can take several milliseconds, depending on
the data rate and packet size. This ACK-based assessment comes
with a cost in both network capacity and client efficiency. When

Best
Yes

(1) Still the


best?

No

6.3 Heuristic Antenna Selection Algorithm

Safe
(2) Should
we assess?

No

The first order prediction of RSS: avg(tn+1)

The exponential moving average is used mainly to smooth out the


high frequency noise in RSSI readings for better RSS predictions,
as discussed in Section 4.5

Assessantennas
Yes

Figure 9: The packet-based antenna selection with two


modes: Best and Safe. The client transits between the two
modes according to two decisions

The key problem is to decide whether the client should assess the
antennas to find out the best one or simply use the omni directional antenna. Our solution to this problem is based on two intuitive observations:

First, if the client orientation and location have not yet


changed much from last antenna assessment, the client
should continue with the last identified best antenna.

Second, if the channel is likely to change rapidly, it is better


to use the omni directional antenna without assessing the antennas. The reason is that if the channel is likely to change a
lot, the identified best antenna will not bring much benefit
before its optimality expires. The tipping point of the change
rate is actually dependent on the packet transmission rate:
the higher the rate, the more benefit the best antenna can
bring in before its optimality expires.

an inferior antenna is being assessed, the Data-ACK exchange


might be unsuccessful and therefore results in retransmission of
the data packet. Such cost can potentially be reduced by sending
empty packets and using their ACKs for antenna assessment.

6.2 Antenna Bookkeeping


Because finding the best antenna incurs cost and the superiority
interval of the best antenna may be short, it is not always worthwhile to find out the best antenna. Therefore, we make some design decisions in the packet-based assessment discussed below.
The goal of these decisions is to keep the client updated regarding
its antennas in order to avoid unnecessary attempts to find out the
best antenna. We call it antenna bookkeeping.
First, the client continuously assesses the antenna in use, either
omni directional or directional. That is, the client always uses the
same antenna to finish a Data-ACK exchange. This allows the
client to detect changes in the channel of the antenna in use almost immediately.
Second, the client opportunistically assesses the omni antenna by
using it for idle listening and data reception. Therefore, it can
obtain the RSSI of the omni from the overheard packets sent by
the infrastructure to its peers.
Finally, the client leverages the beacons broadcast by the access
point, every 100ms by default in 802.11, to guarantee the omni
antenna and some directional antennas are assessed frequently.
When a directional antenna is used for data transmission and
therefore continuously assessed, the client uses the beacons to
guarantee the omni antenna is assessed every few hundred milliseconds. When the omni antenna is used for data transmission
and therefore continuously assessed, the client uses the beacons to
guarantee that directional antennas are assessed every few hundred milliseconds.
Using these three mechanisms, we can obtain the RSS of the antenna used for data transmission, and the RSS of the antennas
used for receiving other packets (idle listening and beacons) as
time series. We denote the time series of the RSS for an antenna
as S(t) where t is the time of assessment. Let tn denote the time for
the most recent assessment, tn-1 for the one before it and so on.
We can calculate several important measures of the RSS for each
of these antennas:

Exponential moving average of RSS: Savg(tn)

The rate of RSS change R(tn):


R(tn)=[Savg(tn)-Savg(tn-1)]/(tn- tn-1)

We leverage these two observations and devise a simple selection


algorithm as illustrated by Figure 9. The algorithm has the client
working in one of the two modes: Best, and Safe. In the Best
mode, the client uses the antenna that was identified as the best
antenna for data transmission from last antenna assessment. In the
Safe mode, the client uses the omni antenna and does not know
which antenna is the best. While the client may benefit from directional antennas in the Best mode, the Safe mode provides a
safety belt that the client can resort to when the use of directional antennas is considered risky, especially when the client
experiences frequent rotation or rapid mobility. The client transits
between these two modes according to the answers to the following two questions:

In the Best mode, is the last identified best antenna still the
best? The client considers the current antenna no longer the
best if the predicted RSSI of current antenna, avg, is lower
than that of any other antennas assessed in the last 100ms.

In the Safe mode, should the client assess the antennas to


identify the best one, or should it just keep on using omni?
According to the second observation above, the client bases
the decision on both the channel changes and the packet interval. Packet interval, T, is known to the client. The client
gauges the channel change with the rate of RSS change, R,
of the directional antenna that is assessed most recently. The
client bases the decision on the product of |R| and T. If their
product is greater than a threshold, the client will remain in
the Safe mode without assessing antennas. We experimentally choose 2dB for the threshold in the implementation, as it
shows fast and cautious recovery from the Safe mode.

7. SELECTION WITH PHY SUPPORT


We next show that antenna selection can be much more efficient
and effective by assessing all antennas with PHY training symbols. Unlike the packet-based antenna selection, this PHY symbol-based method requires changes to the network infrastructure.
Therefore, we target it for the long-term deployment of directional communication.

Antenna
training packet

Regular packet

SEL

RSS(dBm)

-45

ACK

(a) RTS/CTS-like feedback


Regular packet

Dir3

-50

Omni
-55

Dir1
Dir2

Antenna
training symbols

-60
0

ACK SEL

(b) ACK-based feedback


Figure 10: Two different ways for the receiver to notify
the mobile client of the best antenna for symbol-based antenna selection

(a)

10

15

time(second)

20

(b)

Figure 11: (a) Physical arrangement of the three directional


antennas used in the trace collection and experimentation
(b) A segment of RSSI trace for three 5dBi antennas in
NLOS indoor environment

With PHY support, all the antennas are examined in a single


packet. That is, a mobile client transmits a series of PHY training
periods, each from one of the antennas so that the receiver can
estimate the RSSI of all antennas and notify the mobile client of
the best. This antenna selection strategy is one form of active
probing of the transmit antennas.

8. EVALUATION

7.1 PHY Symbol-based Antenna Assessment

8.1 Hardware Implementation

The method is inspired by the 802.11 Physical Layer Convergence Protocol (PLCP), in particular the 802.11n MIMO PLCP.
The 802.11 PLCP employs training symbols, for the receiver to
estimate the channel and decode the packet. We employ training
symbols similar to the 802.11n MIMO PLCP and existing antenna selection methods for MIMO, e.g. [14]. Each antenna takes
turn to transmit the training period (8s in 802.11). Knowing the
number of training symbols per antenna, the receiver detects the
symbols that belong to each antenna, and estimates the RSSI for
that antenna. The receiver then notifies the transmitter of the antenna with the strongest channel. This feedback from the receiver
to the transmitter makes the transmitter antenna selection different from and more challenging than antenna selection on the receiver side.

A WARP node can accommodate up to four antennas using two


radio cards. Each radio card has two antenna ports and can select
from only two antennas in real time. Due to this limitation, we
have to connect four antennas using two radio cards and select the
antenna by first selecting the radio card and then selecting the
right antenna. The unselected radio card is powered off. Therefore, there is only one radio card active at a time. While this implementation is less efficient than the switch-only realization as
discussed in Section 5, it demonstrates the feasibility and gain of
MiDAS and its antenna selection methods in real-time communication. The latency to switch between antennas is around 50ns,
negligible in comparison to the PHY training period duration
(8s). Figure 11 (a) shows the physical placement of the three
directional antennas and two different ways to use two of them
only (two-opp and two-adj).

7.2 Selection Feedback


Because the transmitting mobile client does not need the RSSI
values, the receiver can simply send back the index of the best
antenna with logN bits, N being the number of antennas. There
are two ways that the receiver can notify the transmitting mobile
client, as illustrated in Figure 10. First, the receiver notifies with
the ACK packet so that the transmitter can use the selected antenna for the next packet. We call this ACK-based feedback. The
feedback will incur almost no overhead under 802.11. Second, the
transmitter and receiver can use an exchange similar to RTS/CTS
so the transmitter can use the best antenna for the current data
packet. We call this RTS/CTS-like feedback.
The ACK-based feedback is only effective if the packet intervals
are so short that the best antenna selected during one packet
transmission will remain the best for the next. On the other hand,
it is more efficient than the RTS/CTS-like feedback because the
latter incurs overhead due to the extra packet exchange. Therefore, MiDAS uses the ACK-based method only when the packet
interval is short, below 100ms in our realization according to the
characterization in Section 4. Otherwise, it employs the
RTS/CTS-like feedback.

We have implemented MiDAS with up to three directional antennas on the WARP platform. Using this implementation, we are
able to collect traces for a detailed simulation study and also to
perform real-time experiments to demonstrate the effectiveness of
the packet-based and symbol-based antenna selection methods.

The antennas in the prototype are placed as close to each other as


possible. This is because the antennas need to have highly correlated channels to achieve antenna pattern diversity, which is the
goal of MiDAS. Considering that only one antenna is active at a
time, there will be no constraint on the distance between antennas
and they can be as close as possible. This means that not only
placing the antennas close to each other on the sides of a small
form factor smartphone is not constraint by the distance requirements known in MIMO and beamforming schemes, but also it is
desirable.
We have implemented both packet-based and symbol-based antenna selection on the prototype. However, only the packet-based
implementation is ready for real-time 802.11-like communication
at this moment.

8.2 Trace Collection


To evaluate the antenna selection methods under controlled settings, we first collect RSSI traces of all four antennas of the
hardware prototype. The RSSI traces are collected using the same
experimental setup with the motor platform replaying fieldcollected rotation traces as described in Sections 4.1 and 4.4.

6
5

4
3
2

Upper bound
Symbol-based
Packet-based

6
5

4
3
2

2
1

100ms

1s

10s

NLOS-ind. LOS-ind.

outd.

Environment

Average Packet Interval

1
10ms

Upper bound
Symbol-based
Packet-based

Gain(dB)

Upper bound
Symbol-based
Packet-based

Gain(dB)

Gain(dB)

Gain(dB)

Upper bound
Symbol-based
Packet-based

4
3
2
1

three

two-opp

two-adj

one

Antenna Configuration

5dBi

8dBi

Antenna Gain

(a)
(b)
(c)
(d)
Figure 12: Impact of system settings on the performance of the multi-antenna system. Each bar presents the median, 10th and
90th percentile gain for three antenna selection cases: 1) upper bound assuming the best antenna is always used; 2) symbol-based
antenna selection described in Section 7; and 3) packet-based antenna selection described in Section 6
There is one difference though. That is, four antennas, three directional and one omni directional, are used in the experiments.
Therefore, the prototype repeatedly transmits a training period of
128 symbols from each of the four antennas in turn as fast as
possible. The AP node measures and records the RSSI of all of
the antennas. As a result, we are able to measure the RSSI of each
antenna approximately every 10ms. We collect traces for three
5dBi and three 8dBi directional antennas, each in various propagation environments, including NLOS indoor, LOS indoor, and
outdoor, as described in Section 4.1. The same transmit power of
18dBm is used for all experiments. Figure 11 (b) presents a segment of the RSSI traces of the four antennas for NLOS indoor.

interval >> 100ms) while the symbol-based antenna selection


method always achieves a performance close to the upper bound.
Yet the symbol-based method achieves so with the cost of using
the RTS/CTS-like selection feedback described in Section 7. The
results show that for the traffic with average packet interval of
10ms, the RTS/CTS-like feedback is almost never used and therefore does not introduce any overhead. However, the percentage
grows as the traffic becomes sparser, and we observe 36%, 90%,
and 98% usage of RTS/CTS-like feedback for traffics with average intervals of 100ms, 1s, and 10s, respectively. Therefore, the
multi-antenna system is more effective in more demanding networks that have more packets to transmit.

8.3 Trace-based Simulation

Second, MiDAS and the two antenna selection methods work


well for all propagation environments, as can be observed from
Figure 12 (b). On the other hand, we observe that they provide
more gain for LOS indoor and outdoor environments. This is not
surprising as the same directionality usually leads to lower receiver gain under rich multi-path effects, as we have observed in
Section 4 (Figure 5).

The RSSI traces allow us to simulate the impact of various aspects of the antenna selection methods in a controlled manner. In
the following, the simulation default assumes 1) the traffic is
Poisson with the mean packet interval of 10ms; 2) the configuration of the multi-antenna system is three identical directional
antennas and one omni directional antenna; 3) directional antennas have 5dBi peak gain; and 4) the environment is NLOS indoor.
By changing each of these four aspects, we are able to reveal their
impact on the system performance, reported in terms of RSS gain
at the AP over that of the omni directional antenna case. That is,
we report the difference between the RSS achieved by the proposed multi-antenna system and that by a traditional omni directional system. For each evaluation, we show the system performance for three antenna selection methods: the genie-aided upper
bound that knows the best antenna for every packet, the packetbased and symbol-based antenna selection methods according to
Sections 6 and 7, respectively. Figure 12 summarizes the results.
Each bar in the graphs shows the median gain (in dB) of MiDAS
antenna selection algorithms over an omni system in different
scenarios. We also include the 10th and 90th percentile values of
this gain in each bar. We make the following observations.
First, both symbol-based and packet-based antenna selection methods achieve very good gain (close to the upper bound) when
traffic is intensive, i.e. average packet interval is below 100ms, as
is apparent from Figure 12 (a). The packet-based method is
slightly worse than the symbol-based because the packet-based
method is much coarse-grained and has higher overhead in antenna assessment as discussed in Sections 6 and 7. Moreover, neither
of the two antenna selection methods ever falls below the omni
directional antenna case and therefore the safety belt provided by
the omni directional antenna does work. The packet-based method provides little gain when the traffic become sparser (packet

Third, our results indicate that a design with two directional antennas will provide much of the benefit of the directionality, as is
apparent from Figure 12 (c). It also shows that the opposite pair
(two-opp) provides more gain than the adjacent pair (two-adj) as
illustrated in Figure 11. Because two-opp covers a larger angular
range than two-adj, we do expect that two-opp outperforms twoadj averagely, although it is possible that two-adj outperforms
two-opp under certain propagation environments. Moreover, MiDAS provides decent gain even with only one directional antenna. Note that all these configurations also have one omni directional antenna in addition to their directional antennas.
Finally, our results indicate a more focused antenna may not necessarily provide higher gain. As observed from Figure 12 (d),
MiDAS with three 5dBi antennas achieves better gain than that
with three 8dBi antennas, even for the upper bound case. Comparing Figure 12 (c) and Figure 12 (d), we can see even two 5dBi
antennas outperform three 8dBi antennas. Recall Figure 6 from
Section 4. There we observed that the 5dBi antenna provides
longer superiority intervals than the 8dBi one. As a result, MiDAS with 5dBi antennas requires less frequent antenna assessment and antenna changes.

8.4 Number of Directional Antennas:


Our trace-based simulation results showed that with three 5dBi
directional antennas, MiDAS provides ~3dB gain compared to a
conventional omni system. But due to the hardware limitation in
WARP, we were unable to investigate the performance of more

Gain(dB)

Algorithm 1: Rate adaptation and power control in MiDAS


with packet-based antenna selection

Upper bound
Symbol-based
Packet-based

4
3
2
1
0

one

two-opp

three

four

five

six

seven

eight

Number of Directional Antennas

Figure 13: Impact of number of directional antennas on the


performance of upper bound, MiDAS symbol-based, and
MiDAS packet-based
than three directional antennas. Using the NLOS channel characterization of directional antennas in Figure 5, we try to answer the
following question in this section: How much will more directional antennas help?
We use the same one hour mobility trace that we used in our previous experiments along with the results in Figure 5 to emulate
the performance of one to eight 5dBi directional antennas in
NLOS indoor. All scenarios use a Poisson distribution with the
average time interval of 10ms. Figure 13 shows the results for the
genie-aided upper bound gain and the MiDAS antenna selection
algorithms. We make three observations: First, the increase in the
upper bound gain is marginal for more than 4 directional antennas. Second, MiDAS symbol-based always performs close to the
upper bound. But it is important to note that more directional
antennas introduce additional training overhead to MiDAS symbol-based, which increases linearly with the number of antennas.
Third, as also observed in Figure 12 (c), the performance gap
between the upper bound and the MiDAS packet-based increases
as the number of antennas increases due to the additional assessment overhead, and as a result, the gain of MiDAS packet-based
does not increase anymore for more than 4 directional antennas.

8.5 Moving Scenario


In all of the experiments reported above, the AP and client did not
change locations while the client was rotating on the motor platform. In this section, we try to answer the following question:
What is the possible gain of MiDAS when the client moves?
Instead of moving the client being rotated by the motor, we
mount the AP on a cart and move the cart around at a walking
speed until the connection is lost. During the mobility, the setup
experiences both NLOS and LOS indoor environments. The average packet interval is set to be 10ms.
We observe that the upper bound gain, the symbol-based and
packet-based antenna selections achieve average gains of 4.5dB,
4.4dB, and 2.6dB, respectively, compared to 3.0dB, 3.0dB, and
2.7dB, when the AP is not moving. Therefore, MiDAS maintains
its gain even in moving scenarios. We believe that the reason for
high upper bound gain of this scenario is that the antennas on the
client, although mounted as close to each other as possible, are
not fully correlated. Therefore, the system also benefits from
spatial diversity at some spots.

9. MIDAS WITH RATE ADAPTATION AND


POWER CONTROL:
In this section, we apply rate adaption and transmit power control
to MiDAS, MiDAS+RA/PC, to fully realize its gain in terms of

1
2
3
4
5
6
7

Input: Predicted SNR value for the next packet in dBm,


the goodput-SNR table
for all possible rates
1, 2, , goodput loss threshold
Output: optimal rate
, and maximum transmit power reduction
in dB
If in best mode
pick

s.t.
pick
else
pick the base rate (
1)
pick
0
end

goodput and power saving. Goodput is defined as the ratio of the


delivered data payload to the transmission time. We show that
MiDAS can increase the link goodput significantly in low SNR
scenarios and reduce the transmit power noticeably in high SNR
scenarios.
We adopt the SNR-triggered rate adaptation because it has been
shown to outperform other mechanisms if trained well [15] and
MiDAS already uses highly accurate link RSS prediction that can
be translated to SNR predictions. SNR-triggered rate adaptation
mechanisms use the goodput-SNR table, which maps the link
SNR to its expected effective goodput for each physical layer data
rate in the hardware. Deriving the goodput-SNR table is a onetime effort given a wireless interface card.
We apply rate adaptation and power control to MiDAS with
packet-based antenna selection as follows. If MiDAS is in the
Best mode, for every packet, MiDAS predicts the link SNR using
the methods discussed before. Then it uses the goodput-SNR
table and chooses the rate that has the highest expected goodput
in the predicted link SNR. Using the goodput-SNR table, MiDAS
also reduces its transmit power as much as the expected goodput
is not reduced by more than a defined percentage e.g. 1%, also
known as the goodput loss threshold. The threshold can be chosen to be 0% if the device is unwilling to sacrifice any goodput. If
MiDAS is not in the Best mode, we choose the base rate and no
power control is applied, in order to maintain the safety belt.
Algorithm 1 presents a detailed description of the algorithm used
in MiDAS+RA/PC for packet-based antenna selection. The same
algorithm can also be applied to MiDAS with symbol-based antenna selection.

9.1 Simulation Results


We simulate MiDAS+RA/PC with the channel traces. Our simulation uses the rates supported by 802.11a. We use the effective
goodput-SNR table reported by the authors of [16] for eight
802.11a rates using 2000 byte packets. The goodput loss threshold is chosen to be 1%.
Figure 14 shows the results. We observed that the performances
of packet-based antenna selection and symbol-based antenna
selection are very close to each other in terms of goodput gain
and power saving. Therefore, we only show the results for packetbased antenna selection, which has a slightly worse performance
than symbol-based antenna selection. Also the baseline for the
performance comparison is an omni system that also uses the
same rate adaptation algorithm. We have not included power
control in the baseline omni algorithm because the effect of omni

25

150

MiDAS-5dBi
MiDAS-8dBi
omni

GG-Upper bound
GG-MiDAS
PR-Upper bound
PR-MiDAS

100

20
15

50

10
5
0
-10

10

20

Omni SNR(dB)

30

0
0

10

20

30

Omni SNR(dB)

(a)
(b)
Figure 14: (a) Goodput of MiDAS+RA/PC (three 5dBi and
three 8dBi) and Omni+RA in NLOS indoor; (b) Goodput
Gain (GG) and transmit Power Reduction (PR) of MiDAS+RA/PC (three 5dBi) versus omni+RA in NLOS indoor
transmit power is already included in the corresponding omni link
SNR in the x axis of both graphs in Figure 14.
Figure 14 (a) shows the goodput of MiDAS+RA/PC with three
5dBi and three 8dBi antennas. The same figure also shows the
goodput for the baseline omni+RA system. We make the following observations: First, three 5dBi antennas outperform three
8dBi antennas. Second, there are some SNR values for which
omni system has no connection or its goodput is low (below 3.3
Mbps), but MiDAS+RA/PC is capable of maintaining a much
higher goodput. We refer to these SNR values as Weak or no
connection region. Third, MiDAS+RA/PC always outperforms
the omni system. Figure 14 (b) shows the performance of MiDAS+RA/PC in terms of goodput improvement and power reduction for three 5dBi antenna compared to the omni system. We
have not plotted these graphs for Weak or no connection region, because the goodput gain can be infinite in this regime. The
first point in Figure 14 (b) corresponds to the SNR value in which
omni system has 3.3Mbps goodput.
We have also computed the genie-aided upper bound for the two
evaluated metrics with the assumption that the node has perfect
knowledge of channel conditions in the future. The results show
that MiDAS+RA/PC has a performance very close to the upper
bound. The results show that, for omni-SNR = 3 dB, where omni
system has a goodput of 3.3Mbps, MiDAS+RA/PC increases the
goodput by 132% while reducing the transmit power by 7%. Also, for omni-SNR = 30 dB, where omni system has a goodput of
34.4 Mbps, MiDAS+RA/PC maintains the same goodput while
reducing the transmit power by 48%

9.2 Real-Time Experiment Results


We implemented MiDAS+RA/PC in WARP for real-time experimentation. The WARP current physical layer only supports
three different modulations, being BPSK, QPSK, and QAM16
which correspond to 6, 12, 24 Mbps physical layer rates. The
physical layer in WARP is uncoded; hence supported rates differ
in their modulations only. We choose to implement an SNRtriggered rate adaptation with equal air time to benefit more from
higher data rates. That is, the data size of the packet is bigger for
higher rates in order to have equal transmission periods over the
air for all rates.
First, we derive the goodput-SNR table for different rates in
WARP with equal air time packets. To do so, we connect two
WARP boards over wire with 60 dB attenuation in between. The
choice of wire is to emulate a constant SNR medium. Therefore,
the goodput values in this table are the expected value of goodput
given some predicted SNR. In our experiment, we change the

14

Goodput (Mbps)

30

Goodput (Mbps)

35

12
10

BPSK
QPSK
QAM16

8
6
4
2
0
10

20

30

40

SNR (dB)

Figure 15: Effective goodput of different modulations in


WARP with equal-air-time packets.

transmit power over all possible values. For each transmit power,
we send packets as fast as possible for 30 seconds. The size of the
data part of the packet is 1400 bytes for QAM16 modulation, and
is modified accordingly for other modulations to have equal airtime. Figure 15 shows the results. We observe that QPSK modulation always outperforms BPSK over all possible SNR values
in terms of link goodput. Therefore, an SNR-triggered rate adaptation algorithm will not choose BPSK at any input SNR. This
means that we can only benefit from the other two modulations in
our experiments.
We performed two experiments using three 5dBi directional antennas. The client is placed on the motor platform which rotates
according to the traces, and the AP is stationary, and each experiment lasts for 20 minutes. The packets are transmitted as fast as
possible using the selected modulations. In the first experiment,
the omni antenna experiences a low SNR which is not enough to
use the QAM16 modulation. In the second experiment, the omni
link is strong enough to switch to the higher modulation. The
results show that, in the first experiment, the gain from MiDAS is
enough to switch to the higher modulation for most of the packets, and therefore the goodput is increased by 85%, using the
same transmit power as the omni transmitter. In the second experiment, this gain from MiDAS is used to save transmit power by
51% while increasing the goodput by 7%.

10. RELATED WORK


We next discuss related work in smartphone orientation estimation, directional communication, directional antenna, and antenna
diversity. In summary, all existing work considers smartphonelike mobile devices as omni directional.
Smartphone Orientation Estimation: Accelerometers and compasses have been used for different purposes. These applications
include localization [17], inferring the orientation of the device in
motion [18], and monitoring the road and traffic conditions in a
city [19], which includes estimating the tilt angles of the phone
(, ). However, no previous work has characterized the rotational patterns of real-life usage of smartphones, and to the best of
our knowledge, our work is the first to do so.
Directionality based on Smart Antenna System: Beamforming
using smart antenna systems has been widely studied and deployed for cellular base stations (e.g. [20]), WLAN access points
(e.g. [2]), and even vehicles (e.g. [21]), but not for batterypowered mobile devices like smartphones. Compared with MiDAS, beamforming requires multiple RF chains and therefore
comes with much higher power and hardware cost. Therefore,
compared to beamforming based directional communication,

MiDAS is more amenable to immediate deployments due to its


small changes to the hardware and infrastructure.

time under device mobility and rotation, in comparison to


ones with a lower gain.

Directionality from Directional Antenna: Directional antennas


have also been studied for 802.11 mesh nodes (e.g. [7]) and sensor nodes (e.g. [22]). Because existing work only applies directional antennas to fixed platforms that do not move or rotate, it
does not address the antenna selection challenge that must be
solved for use on mobile devices. Our previous work, BeamSwitch [23], is the most related work. It requires the directional
antennas to be placed in a way to form an omni directional coverage. It employs some form of packet-based antenna assessment to
identify the best directional antenna all the time. In contrast, we
do not have any constraints on the number or placement of antennas. Moreover, we show that it is not always worthwhile to find
out the best antenna and therefore answer a different question:
should the antennas be assessed or not? Also, we study the much
more efficient symbol-based antenna selection method so that the
order of antenna assessment is no longer important.

Finally, by experimentally demonstrating the link gain of passive


directional antennas on mobile devices, this work is an important
step toward an efficient network of directional mobile clients It
invites further research into many interesting networking problems when mobile clients employ directional communication
solutions like MiDAS.

Antenna Diversity on Mobile Devices: Mobile devices such as


wireless handsets have used multiple (often two) omni-directional
antennas for antenna diversity, e.g. see [24, 25]. Compared to
MiDAS that uses directional antennas, antenna pattern diversity
from omni directional antennas is lower, and the benefit is mainly
due to spatial diversity. Some mobile devices actually use two
directional antennas to form an omni directional antenna and treat
them as a single antenna for communication. Therefore, they do
not gain from the directionality of the directional antennas. Also,
all the antenna selection algorithms for receiver diversity systems
are intended for the receiver side, which makes them different
from MiDAS.

11. CONCLUSIONS
We make the following conclusions following the characterization, simulation, and experimentation reported in this work.

Through MiDAS and its antenna selection methods, directional antennas can be effectively employed to improve the
transmission gain of smartphone-like mobile devices by almost 3dB under various propagation environments and realistic rotation. Such gain can be achieved without any change
to the deployed network infrastructure.
Three 5dBi antennas, placed on the adjacent sides of a mobile device, can provide 3dB gain of MiDAS.

ACKNOWLEDGEMENTS
The work was supported in part by NSF awards ECCS/IHCS
0925942, CNS/NeTS-WN 0721894, CNS/CRI 0751173, CNS0551692, CNS-0551692, and support from the TI Leadership
University program. The authors would like to thank the anonymous reviewers for their constructive comments and thank Dr.
Dina Papagiannaki for shepherding the camera ready.

REFERENCES
[1]

[2]

[3]

[4]

[5]

[6]

[7]

[8]

Such gain is possible because of the following characteristics we


identified from field-collected traces and trace-driven experimentation with directional antennas:

[9]

[10]

Smartphone-like devices rotate relatively slow, below


120/second for 90% of the time for each Euler angle, during
wireless usage, compared to how fast modern wireless
clients exchange packets with their access point or base stations.

[11]

The channel of directional antennas is quite reciprocal at


least under various 802.11 propagation environments.

[12]

The channel of directional antennas is quite predictable in


short terms (100ms or shorter) when the device is undergoing human-inflicted rotations.

[13]

Directional antennas with a higher gain may not necessarily


be better because they will be good for a shorter period of

[14]

M. Blanco, R. Kokku, K. Ramachandran, S. Rangarajan, and


K. Sundaresan, "On the Effectiveness of Switched Beam Antennas in Indoor Environments," in Proc. Passive and Active
Measurement (PAM), April 2008.
X. Liu, A. Sheth, M. Kaminsky, K. Papagiannaki, S. Seshan,
and P. Steenkiste, "DIRC: increasing indoor wireless capacity
using directional antennas," in Proc. ACM SIGCOMM, August
2009.
R. R. Choudhury, X. Yang, R. Ramanathan, and N. H. Vaidya,
"Using directional antennas for medium access control in ad
hoc networks," in Proc. ACM MobiCom, September 2002.
Telex, 5dBi Patch Antenna:
http://www.telex.com/Wireless/product.aspx?MarketID=1&Cat
egoryID=3&ProductID=49, 2008.
B. Cetiner, H. Jafarkhani, J. Qian, H. Yoo, A. Grau, and F. De
Flaviis, "Multifunctional reconfigurable MEMS integrated antennas for adaptive MIMO systems," in IEEE Communications
Magazine, vol. 42, issue 12, 2004.
S. Zhang, G. Huff, J. Feng, and J. Bernhard, "A pattern reconfigurable microstrip parasitic array," in IEEE Transactions on
Antennas and Propagation, vol. 52, issue 10, 2004.
A. P. Subramanian, H. Lundgren, and T. Salonidis, "Experimental characterization of sectorized antennas in dense 802.11
wireless mesh networks," in Proc. ACM MobiHoc, May 2009.
H. Goldstein, C. P. Poole, and J. L. Safko, Classical Mechanics, 2nd ed.: Addison-Wesley, 1980.
Kionix, Application Note: screen rotation and device orientation: http://www.kionix.com/App-Notes/AN011 Screen Rotation.pdf.
Maxim Integrated Products, Datasheet for MAX2828MAX2829, Single-/Dual-Band 802.11a/b/g World-Band Transceiver ICs, 2008.
L-COM Global Connectivity, RE08P:
HyperGain 2.4 GHz 8 dBi Round Patch Antenna: http://www.lcom.com/multimedia/datasheets/DS_RE08P-RTP.PDF.
Telex, Omni-directional 2406 Antenna:
http://www.telex.com/Wireless/Product.aspx?MarketID=1&Cat
egoryID=3&ProductID=21, 2008.
G. Judd, X. Wang, and P. Steenkiste, "Efficient channel-aware
rate adaptation in dynamic environments," in Proc. ACM MobiSys, June 2008.
T. Onizawa, A. Ohta, and Y. Asai, "Experiments on FPGAImplemented Eigenbeam MIMO-OFDM With Transmit An-

[15]

[16]

[17]

[18]

[19]

[20]

[21]

[22]

[23]

[24]

[25]

[26]

tenna Selection," in IEEE Transactions on Vehicular Technology, vol. 58, issue 3, 2009.
J. Camp and E. Knightly, "Modulation rate adaptation in urban
and vehicular environments: cross-layer implementation and
experimental evaluation," in Proc. ACM MobiCom, September
2008.
D. Qiao, S. Choi, and K. Shin, "Goodput analysis and link
adaptation for IEEE 802.11 a wireless LANs," in IEEE Transactions on Mobile Computing, vol. 1, issue 4, 2002.
M. Youssef, M. Yosef, and M. El-Derini, "GAC: EnergyEfficient Hybrid GPS-Accelerometer-Compass GSM Localization," in Arxiv preprint arXiv:1004.3174.
K. Kunze, P. Lukowicz, K. Partridge, and B. Begole, "Which
way am I facing: Inferring horizontal device orientation from
an accelerometer signal," in Proc. IEEE Int. Symp. on Wearable Computers (ISWC), Septemeber 2009.
P. Mohan, V. Padmanabhan, and R. Ramjee, "Nericell: Rich
monitoring of road and traffic conditions using mobile smartphones," in Proc. ACM SenSys, November 2008.
J. C. Liberti and T. S. Rappaport, Smart Antennas for Wireless
Communications: IS-95 and Third Generation CDMA Applications: Prentice Hall PTR Upper Saddle River, 1999.
V. Navda, A. P. Subramanian, K. Dhanasekaran, A. TimmGiel, and S. Das, "MobiSteer: using steerable beam directional
antenna for vehicular network access," in Proc. ACM MobiSys,
June 2007.
C.-L. Yang and W. J. Chappell, "Angular Diversity Prominent for Compact Devices of Wireless Sensor Network in Indoor Environments," in Proc. IEEE Antennas and Propagation
Society Int. Symp. (APS), June 2007.
A. Amiri Sani, H. Dumanli, L. Zhong, and A. Sabharwal,
"Power-efficient directional wireless communication on small
form-factor mobile devices," in Proc. ACM/IEEE Int. Symp.
Low Power Electronics and Design (ISLPED), August 2010.
O. Norklit, P. D. Teal, and R. G. Vaughan, "Measurement and
evaluation of multi-antenna handsets in indoor mobile communication," in IEEE Transactions on Antennas and Propagation,
vol. 49, issue 3, 2001.
J. S. Colburn, Y. Rahmat-Samii, M. A. Jensen, and G. J. Pottie,
"Evaluation of personal communications dual-antenna handset
diversity performance," in IEEE Transactions on Vehicular
Technology, vol. 47, issue 3, 1998.
M. Rogers, K. Hrovat, K. McPherson, M. Moskowitz, and T.
Reckart, "Accelerometer data analysis and presentation techniques," Technical Report, National Aeronautics and Space
Administration, Lewis Research Center, 1997.

Appendix A: Trace Collection


We employ a commercial smartphone with a built-in tri-axis accelerometer and compass, HTC G1, to collect data regarding
device orientation from real mobile device usage. We have developed a software logger to collect accelerometer and compass
readings along with information regarding wireless network status, and applications being used for HTC G1. The logger runs in
the background and collects data about every 10ms, including
both voice and Internet usage. We employ a larger battery and an
extended battery compartment cover for the HTC G1 so that the
smartphone has 10 to 15 hours of battery lifetime with normal
usage, long enough not to significantly affect participants usage.

Table 1: Orientation estimation error ()


Euler
Angle

Static

Accelerated

Avg.

Max.

Avg.

Max.

0.9

1.5

1.1

1.8

3.8

9.5

1.5

3.3

11.7

21.1

4.6

9.5

We recruited 11 participants through flyers distributed on Rice


university campus to use the HTC G1 with their own SIM cards
inside. We collected data from each for one week. Each participant was rewarded with a $50 gift card. We plan to make the
collected traces open-access.

Appendix B: Euler Angle Estimation


We next provide details regarding how we estimate the three
Euler angles from accelerometer and compass readings.

B.1. Overcoming Hardware Limitation


In practice, there are a few challenges to the estimation algorithm
from the hardware limitations of accelerometer and compass in
smartphones. First, both the accelerometer and the compass are
prone to hardware noise. In particular, the compass can be easily
interfered by electromagnetic activities nearby, including those by
the host device. From controlled measurement conducted in the
lab, we observe that such interferences and hardware noise tend
to be much higher frequency than what could be possibly introduced by device rotation or mobility. Therefore, we suppress the
noise and interference with a moving average window of 25 samples. Moreover, the numerical readings of a sensor can be different from the physical value. For example, when there is no acceleration, the accelerometer reading can be non-zero. We remove
this offset by data demeaning [26]. Finally, while we leverage the
impact of gravity on the accelerometer reading to estimate the
orientation, the accelerometer reading also reflects the mobility of
a device through acceleration. We note that the gravity is almost
constant while the external acceleration due to mobility changes
much faster. Therefore, the impact from mobility is suppressed
with the mentioned low pass filter.

B.2. Validation of Orientation Estimation


We validate our orientation estimation algorithm with two experiments. In the first experiment (static), we place the HTC G1
smartphone on a plastic supporter in a series of orientations that
are measured accurately as the ground truth. In the second experiment (accelerated), we fix the orientation of the phone and
hand-move the phone with high acceleration toward various directions. In both experiments, we use our algorithm to calculate
the orientation and compare it with the ground truth.
Table 1 provides the average and maximum error for each of the
Euler angles for both the experiments. As the results show, the
estimation of is the most accurate. The estimation of has very
small error in most of its range except for when is equal to 90 or
-90 degrees, which is known as the Gimbal Lock [8]. is the
most error-prone angle. This is mainly due to the noisy compass
reading. However, our interest is in the device rotation, instead of
orientation. Because the random nature of the errors, we can consider the rotation derived from the orientation estimation acceptable for our characterization purpose.

You might also like

pFad - Phonifier reborn

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.


Alternative Proxies:

Alternative Proxy

pFad Proxy

pFad v3 Proxy

pFad v4 Proxy