Design of Non-orthogonal Multi-channel Sensor Networks.

Full text

Page 1

Design of Non-orthogonal Multi-channel Sensor
Networks
Xing Xu
Hong Kong University of
Science and Technology
xingx@cse.ust.hk
Ji Luo
Hong Kong University of
Science and Technology
luoji@cse.ust.hk
Qian Zhang
Hong Kong University of
Science and Technology
qianzh@cse.ust.hk
Abstract—A critical issue in wireless sensor networks (WSNs)
is represented by the network throughput. To meet the through-
put requirement, researchers propose multi-channel design in
802.15.4 networks to better utilize the wireless medium and avoid
the co-channel interference. However, traditional orthogonal
channel design restricts the number of channels and limits the
throughput performance. We argue that the orthogonality is not
necessary for multi-channel design in WSNs.
In this paper, we investigate the feasibility of non-orthogonal
channel design. In our experiment, we observe that with non-
orthogonal transmission, the effect of interference comes from
co-channel and inter-channel is different. More specifically, the
inter-channel interference is tolerable with certain channel center
frequency distance (CFD). According to that, we propose a
novel scheme DCN (Dynamic CCA-threshold for Non-orthogonal
transmission) which adjusts the CCA-threshold to enable the
concurrent transmissions on adjacent non-orthogonal channels
and thus improve the overall network throughput performance.
Through comprehensive experiments on our testbed, we verify
that our DCN achieves about 38.4% ∼ 55.7% throughput
improvement in general network configurations comparing to
the default ZigBee design.
Index Terms—Multi-channel, orthogonal channel, interference,
CCA-threshold, wireless sensor networks.
I. INTRODUCTION
As an emerging technology, wireless sensor networks
(WSNs) have been designed to support a wide range of po-
tential applications, including environment monitoring, event
detection as well as information collection [1][2][14]. In these
applications, large number of sensors are deployed in the
interested region, communicate with each other using wireless
medium. Due to the shared nature of wireless medium, the
interference among multiple transmissions that use the same
channel becomes a fundamental issue [3]. Existing works [20]
show that concurrent transmissions in the same channel is
a great menace to the communication, since it will lead to
collision where the collided packets would not be decoded
successfully. It is well known that the current WSN hard-
ware, such as Micaz and Telos that use the CC2420 radio,
already provide multiple frequencies. Therefore, designing
multi-channel based communication protocols in WSNs to
improve network throughput and provide high performance
communication services is then a nature way [18].
For multi-channel based communication protocol design in
WSNs, one of the most important parameters is the number
of channels which can actually be used. The CC2420 radio
chip that follows ZigBee standard [10] provides 16 channels,
with 5MHz as the center frequency distance (CFD) between
neighboring channels. However, as Wu et al. pointed out
[18], not all channels can be used in a single sensor network
to provide parallel transmissions because of close channel
interferences and interferences caused by other wireless net-
works. To address the channel scarcity issue, several studies
[7][15] have been proposed to utilize overlapping channels,
which were targeted for 802.11 networks only. Similar studies
in 802.15.4 networks are limited. Through our experiment,
we find that 802.15.4 networks show unique interference
characteristics which cannot be captured by existing 802.11
models. More importantly, such uniqueness will benefit multi-
channel design of 802.15.4 networks particularly and plays a
key role in our study.
For a given network with a certain spectrum band, non-
orthogonal channel assignment provides more available chan-
nels. However, comparing to the traditional orthogonal channel
assignment, it introduces a critical challenge, i.e., the inter-
channel interference may affect the transmission significantly.
Thus, to decide a reasonable frequency distance between
neighboring-channels, the trade-off between larger number
of channels and weaker inter-channel interference has to be
carefully considered. Through our following experiments, we
verify that assigning each channel with the channel distance
that guarantees orthogonality (i.e., 9MHz for 802.15.4) cannot
fully utilize the bandwidth medium. Moreover, the default
CFD setting of ZigBee, i.e. 5MHz, is inefficient as well.
Our interesting observation is that the assignment based on
a smaller CFD, e.g., 3MHz, provides better bandwidth uti-
lization. Specifically, if we assign each channel with the CFD
as 3MHz and generate saturated traffic on each channel, the
overall throughput on a given spectrum bandwidth would be
improved significantly comparing to the orthogonal channel
assignment scheme. Comparing with the result obtained with
the ZigBee default setting, more than 40% throughput gain
could be achieved by leveraging smaller CFD, showing that
smaller CFD better exploits the bandwidth.
To further study how to better leverage the non-orthogonal
channels with smaller CFD, e.g., 3MHz, we verify the feasi-
bility of concurrency between assigned non-orthogonal chan-
nels. Different to the collision of co-channel interference that
collided packets cannot be both decoded successfully, packets
transmitted simultaneously from two non-orthogonal channels
2010 International Conference on Distributed Computing Systems 2010 International Conference on Distributed Computing Systems
1063-6927/10 $26.00 © 2010 IEEE
DOI 10.1109/ICDCS.2010.37DOI 10.1109/ICDCS.2010.37
3583581063-6927/10 $26.00 © 2010 IEEE

Page 2

(e.g., CFD=3MHz) could be decoded successfully in most
power settings or just be slightly corrupted. According to this
unique observation in 802.15.4 networks, we propose a new
scheme DCN (Dynamic CCA-threshold for Non-orthogonal
transmission) that can dynamically modify CCA-threshold1.
By doing so, non-orthogonal channels possess the capability of
concurrency, which is usually supported by orthogonal channel
assignment. We also discuss that the packet recovery scheme
could be integrated with DCN to rescue the slightly corrupted
packets in some special cases. The empirical experiments show
that our scheme DCN can achieve about 38.4% ∼ 55.7%
bandwidth throughput improvement with CFD=3MHz, com-
paring to the default multi-channel settings of ZigBee.
In this paper, we for the first time investigate the fea-
sibility of non-orthogonal multi-channel design for ZigBee-
based WSNs and propose a CCA-threshold adaptation scheme
to maximize the bandwidth throughput. As a summary, we
have the following contributions: 1) we empirically verify
that the default setting of CFD=5MHz in ZigBee is quite
conservative. The assignment with smaller CFD could achieve
better throughput on a given bandwidth; 2) we observe that
different to co-channel interference, the inter-channel interfer-
ence introduced by non-orthogonal assignment is tolerable in
most of the communication settings (e.g., transmission power,
CFD). The collided packets caused by such inter-channel
interference could be decoded successfully or only have a
small portion of bit errors. Thus, we propose a new scheme
DCN to exploit the concurrencies on non-orthogonal channels;
3) with the comprehensive experiments on a real testbed,
we verify that our DCN could obtain significant throughput
improvement comparing to the traditional ZigBee setting.
The rest of the paper is organized as follows. We discuss
the related work in Section II. In Section III, we introduce a
concrete example which motivates our investigation. Section
IV is devoted to the detail analysis of non-orthogonal multi-
channel design, which consists of the study on the co-channel
interference and inter-channel interference; while Section V
presents our scheme DCN. The extensive experiments of our
proposed design is shown in Section VI and in Section VII we
will have a further discussion on the packet recoverability and
the upper bound of bandwidth throughput with DCN. Finally,
a conclusion for the paper will be made in Section VIII.
II. RELATED WORK
It has been well studied that using multiple channels can
significantly enhance the network capacity in wireless net-
works [17]. In recent years, there are also some studies on
efficient delivery in wireless sensor networks by making use
of multi-channel, such as MCMAC [6], TMMAC [21], MMSN
[22], etc. However, all these works assume the channels are
orthogonal and they all evaluated by simulation only.
1CSMA exploits clear channel assessment (CCA) method, which deter-
mines the state of channel by checking whether the in-channel energy is
above a given threshold (CCA-threshold) and transmission can only happen
if the channel appears idle.
Wu et al. proposed TMCP [18], a realistic multichannel
protocol based on empirical experiments. In their conclusions,
they complained that the small number of available channels
limits the multi-channel design of WSNs. Therefore, they find
fully orthogonal and high-quality channels first and partition
the whole network into sub-trees according to the number
of available channels. However, in our experiment, we come
out an interesting observation that though the non-orthogonal
channel interference affects the transmission, such interference
is tolerable. Thus, we use non-orthogonal channels in our
scheme with CFD smaller than the default setting of ZigBee
to provide more available channels.
Although non-orthogonal channel interferences have been
addressed in wireless networks [15] [16], related empirical
works in WSNs are rare. Incel et al. have conducted a measure-
ment study for adjacent channel interference in WSNs using
Ambient uNode [11]. In our work, more popular MicaZ motes
[8] are used. Besides, they disabled the MAC layer behaviors
of sensor node to introduce collisions; but we have studied
a new dimension: setting different CCA-threshold instead of
disabling entire CSMA policy. By varying CCA-threshold, we
can control the power level of introduced interference and
demonstrate the different throughput performance. Recently,
Xing et al. studied the interference of adjacent channel in
WSNs [19], but they use the channel assignment according to
ZigBee [10], while we use even smaller frequency distance to
partition channels for maximizing the utilization on bandwidth.
There are also some works discussed the optimal setting
of CCA-threshold [3] [23]. However, since their works are
based on single channel design where there is no neighboring-
channel interference, they only consider co-channel inter-
ference. Our study is a multi-channel design that exploits
non-orthogonal channels, thus the uniqueness of our CCA-
threshold setting scheme (DCN) is the differentiation of the
interference from co-channel and neighboring-channels, with
an emphasize on dealing with the neighboring-channel inter-
ference. More, Bertocco et al. used White Gaussian noise gen-
erated by signal generator as interference source [3]. However,
in densely deployed WSNs applications, the main interference
is transmitting packets, i.e., valid IEEE 802.15.4 packets.
Obviously, signal generator cannot emulate a real interfering
network: first, White Gaussian noise is different from a valid
packet; and second, constant noise cannot simulate the traffics
of a transmitting network. In this work, we use actual testbed
to generate interference from multiple channels.
III. MOTIVATION: FEASIBLE NON-ORTHOGONAL DESIGN
In this section, we focus on a simple example to demon-
strate the feasibility of non-orthogonal multi-channel design.
Through the experiment, we find that comparing to the
CFD which guarantees the orthogonality, smaller and non-
orthogonal CFD may have better utilization on the given
bandwidth but we also point out that there is a trade-off.
To further explain the advantages of non-orthogonal channel
design, we study the concurrency between non-orthogonal
channels. We observe that the concurrent transmission between
359359

Page 3

95432
0
500
1000
1500
Channel Frequency Distance (MHz)
Bandwidth Throughput with
Different Frequency Distance
non-orthogonal channels is feasible which implies the potential
bandwidth improvement.
Throughput (packets/s)


N0
N1
N2
N3
N4
N5
Fig. 1.
123456789 1011
0
0.5
1


802.11b
123456789 1011
0
0.5
1
Channel Number
Uniqueness of 802.15.4 Net-
Normalized Throughput (1)
Fig. 2.
works


802.15.4
Fig. 3.
Sender
compliant devices (Crossbow MicaZ motes [8]), we found that
sensor device cannot decode packets from inter-channels, even
the packets are from only 1MHz frequency distance away.
Therefore, inter-channel concurrency is feasible for 802.15.4
networks.
To generate inter-channel concurrencies, we set two links on
non-orthogonal channels sending out packets simultaneously
by disabling their carrier sense module. Since it is hard to
synchronize the collision of two packets, we design the sender
of one link as an attacker which sends out packets at an
extremely fast rate, i.e., 1 packet for each 3ms. With such high
channel occupancy, all the packets from the normal sender on
the other link would be collided as shown in Fig. 3.
We focus on the node’s capability of decoding with interfer-
ence by computing the Collided Packet Receive Rate (CPRR)
for both the attacker and normal sender. By evaluate different
CFD with the same transmission power, the experiment results
shown in Fig. 4 address the feasibility of concurrency on adja-
cent channels: for CFD greater than 4MHz, both CPRR of the
attacker and the normal sender is 100%; CFD=3MHz provides
97% CPRR which indicates that most collided packets could
be decoded successfully; for the case of CFD=2MHz, CPRR
decreases to around 70% due to the severer inter-channel
interference; and for the extreme case of CFD=1MHz, it shows
poor concurrent feature that less than 20% collided packets
could be decoded successfully.
In this section, we observe that simply using smaller CFD
design instead of traditional ZigBee setting or orthogonal
assignment could achieve better bandwidth throughput. Above
experiment results also show the feasibility of concurrent
transmission in non-orthogonal channel which implies the
potential bandwidth improvement that we can obtain.
Collision of Attacker and
54321
0
0.2
0.4
0.6
0.8
1
Channel Frequency Distance (MHz)
CPRR with Different Fre-
quency Distance
Collided Packet Receive Rate (100%)


Normal Sender
Attacker
Fig. 4.
A. Channel Distance vs. Overall Throughput
We first evaluate different CFD=9, 5, 4, 3, 2MHz with
ZigBee default protocol on a given bandwidth of 12MHz: for
CFD=9MHz which almost guarantees the fully orthogonality,
we only have 1 channel while for CFD=5MHz which is the
default setting of ZigBee, we have 2 channels. We also try
smaller CFD (e.g., 4, 3, 2MHz) which could assign more
number of channels but introduce severer inter-channel inter-
ference. For each assigned channel, we have 4 MicaZ nodes
with maximum transmission power (i.e., 0dBm). All the nodes
are sending packets at the maximum data rate, generating a
saturated traffic of the channel.
As shown in fig. 1, it demonstrates that the orthogonal
channel assignment will not provide the maximum throughput
(e.g., in Fig. 1 the maximum throughput is achieved at
CFD=3MHz ). The reason is that, though the throughput for a
single channel is maximized by orthogonal CFD setting (i.e.,
CFD=9MHz), but the overall throughput for orthogonal CFD
setting is poor due to the limited number of channels. However,
smaller CFD (e.g., CFD=2MHz) may not always benefit the
overall throughput since severer inter-channel interference may
corrupt more and more packets.
This example indicates that:
CFD=5MHz in ZigBee is quite conservative, smaller CFD for
multi-channel design could obtain better bandwidth through-
put; 2) there is a tradeoff between throughput benefit from
more number of channels and harm from severer inter-channel
interference. To further get the understanding of how to
leverage the non-orthogonal channels to achieve the maximum
bandwidth throughput, we study the concurrent transmission
between non-orthogonal channels in the following subsection.
1)defaultsetting of
B. Concurrency between Non-Orthogonal Channels
In this subsection, we verify the feasibility of concurrent
transmissions in non-orthogonal channels. In 802.11 networks,
concurrent transmissions on non-orthogonal channels is in-
feasible. It is because inter-channel interference acts as valid
packets and force receiver to decode it (even the interference
is from three channels away, i.e., 15MHz away); during
the decoding, the receiver lose desired packet that is sent
simultaneously in the same channel (see experiments result
conducted by Mishra et al. [15] as shown in Fig. 2). As
the comparison, in our experiment on 802.15.4 (ZigBee) [10]
IV. CCA-THRESHOLD RELAXING FOR THROUGHPUT
IMPROVEMENT
In previous section, we have shown that the concurrency is
feasible for non-orthogonal channel design (e.g., CFD=3MHz)
by disabling the carrier sensing module of senders to introduce
the inter-channel collisions. However, there is another respon-
sibility for carrier sensing module, to filter the co-channel
interference. Thus, simply disabling the carrier sensing module
is not a proper way to achieve better throughput. On the other
hand, the CCA-threshold is fixed at -77dBm in the default
setting of ZigBee. It assumes that the interference/noise over
-77dBm would corrupt the packets, and thus makes the sender
backoff when it hears the inter-channel interference but actu-
360 360

Page 4

Fig. 5.
channel Interference
Network Configuration without Co-
−120−100 −80−60−40 −20
0
50
100
150
←−77dBm
CCA−Threshold (−dBm)
Link Throughput with Different CCA-
Threshold (Case without Co-Channel interference)
Throughput (packets/s)


Sent Packets
Received Packets
Fig. 6.
−120−100−80−60 −40−20
800
900
1000
1100
1200
1300
1400
1500
CCA−Threshold (−dBm)
Overall Throughput with Different CCA-
Threshold (Case without Co-Channel interference)
Throughput (packets/s)


←−77dBm
Overall Throughput
Fig. 7.
−120−100−80−60−40−20
0
50
100
150
CCA−Threshold (−dBm)
Link Throughput with Different CCA-
Threshold (Case with Co-Channel interference)
ally miss the chance that the sender can still do transmission
since such inter-channel interference is tolerable as observed in
previous section. Therefore, the conservative setting of CCA-
threshold prohibits the concurrency opportunities and draws
the throughput gap between orthogonal and non-orthogonal
channel design as shown in Fig. 1. In this section, we modify
the carrier sensing module, analyze and find a principle to
select proper CCA-threshold for the throughput improvement.
We will focus on the case of CFD=3MHz in the following
experiments since it shows that CFD=3MHz achieves best
performance with default carrier sensing module in Fig. 1.
Throughput (packets/s)


←Min RSS
Sent Packets
Received Packets
Fig. 8.
−120−100−80−60−40−20
0
50
100
150
←−77dBm
CCA−Threshold (−dBm)
Throughput (packets/s)


−8dBm
−11dBm
−15dBm
−22dBm
−33dBm
Fig. 9. Link Throughput with Different Transmis-
sion Power (Case with Co-Channel interference)
−120 −100 −80−60 −40−20
0
0.2
0.4
0.6
0.8
1
CCA−Threshold (−dBm)
Link PRR with Different Transmission
Power (Case with Co-Channel interference)
interference, we could obtain better throughput gain by relax-
ing CCA-threshold. The next step is to answer what CCA-
threshold to select by taking the co-channel interference into
the consideration.
Packet Receive Rate (100%)


−8dBm
−11dBm
−15dBm
−22dBm
−33dBm
Fig. 10.
A. Throughput Improvement without Co-Channel Interference
We start with a simple case without co-channel interference.
For a single link transmitting with 4 neighboring-channel
interference (i.e., CFD=±3, ±6MHz, see Fig. 5) using the
same power (0dBm), we enumerate different CCA-threshold
for this particular link. In Fig. 6, with more relaxed CCA-
threshold, such link sends out more packets since it turns
to consider more interfered channel condition as a clear
channel status. As such neighboring-channel interference is
tolerable, more packets that the link sent out lead to better
throughput. Meanwhile, we also notice that the packet receive
rate (PRR) keeps almost 100%, which indicates the reliability
of neighboring-channel concurrency and matching the result
of our previous concurrency test (Fig. 4) as well.
Furthermore, we check the overall throughput to see
whether such link throughput gain degrades other networks’
throughput. As demonstrated in Fig. 7, the overall throughput
also grows which indicates that the inter-channel concurrency
has been leveraged successfully. Therefore, without co-channel
B. The Case with Co-channel Interference
From the viewpoint of every sensor node, it not only has
neighboring-channel interference but also encounters the co-
channel competitors. Thus, the CCA-threshold should also be
responsible for dealing with the co-channel interference.
Based on previous experiment, we introduce 3 additional
links working together with that particular link at the same
channel. The result in Fig. 8 shows that relaxing CCA-
threshold will not always benefit the throughput.
To further explain it, in the experiment case, the co-channel
interference is stronger than the neighboring-channel interfer-
ence (the vertical line shows the minimum power level of co-
channel interference) and once we relax CCA-threshold more,
the co-channel interference would be introduced which leads
to a disaster. For two collided packets in the same channel,
current common used modulation component of sensor node
could only decode at most one of them; or in the worst case,
both packets are corrupted. The throughput will not benefit
from the co-channel interference but be harmed.
As a short summary, through the experiments above we
observe that relaxing CCA-threshold could exploit more inter-
channel concurrent opportunities to improve the throughput
where the default CCA-threshold setting is quite conservative
for leveraging inter-channel concurrency. On the other side, the
co-channel concurrency should be avoided since decoding co-
channel interfered packets would robber the time for normal
361361

Page 5

Fig. 11.Architecture of DCN
transmission. This observation would be a basis to propose
our dynamic CCA-threshold scheme in next section.
C. Effect of Transmission Power
In all the experiments above, we discuss feasibility of
concurrency and relaxing CCA-threshold with the setting of
the same transmission power. To further verify whether our
observation holds for different power setting, we will evaluate
the effect of transmission power in this subsection.
Fig. 9 illustrates that with different transmission power,
there would always be some throughput improvement by
relaxing CCA-threshold to introduce the inter-channel inter-
ference. Without surprise, the gain is different with different
transmission power due to the node’s capability of decoding
with interference. However, we found that for most of the
cases, i.e., transmission power is greater than -22dBm, the
PRRs are all 100% (in Fig. 10). Even for the case that link’s
transmission power is -22dBm vs. 0dBm of the interferers, the
PRR is higher than 80%. Such experiment results provide a
support for our observation so that we could design our scheme
for most general cases.
V. DESIGN AND IMPLEMENTATION
As a conclusion in previous section, in terms of bandwidth
throughput the CCA-threshold should be relaxed to intro-
duce concurrencies with neighboring-channel transmissions.
However, the CCA-threshold could not be relaxed in an
arbitrary way since there is also a responsibility for it to
prohibit the co-channel collisions. Thus, the CCA-threshold
must be able to introduce the inter-channel interference and
filter the co-channel interference at the same time. Towards this
end, we design and implement DCN, a scheme of Dynamic
CCA-Threshold for Non-orthogonal transmission. In DCN, we
modify sensor MAC protocol by adding an new component,
CCA-Adjustor. In the following subsections, we first show the
architecture of our DCN, followed by the design of CCA-
Adjustor and finally introduce the implementation in detail.
A. Architecture
In the traditional CSMA/CA policy (see Fig. 11), the upper
layer wishing to send packets has to first sense the power
of using channel to check whether there is any activity on the
channel by comparing the sensing power with a predetermined
−100−80−60−40 −20
0
0.5
1
(1) Overlapped Interference


←Updating Phase
←Initializing Phase
Co−Channel
Inter−Channel
−100 −80−60−40 −20
0
0.5
1
(2) Separated Interference
Probability Fraction (100%)


←Updating Phase
←Initializing Phase
Co−Channel
Inter−Channel
Fig. 12.Setting of CCA-Threshold
CCA-threshold. If the sensing power is stronger than the CCA-
threshold, it regards the channel as currently occupied and
postpones the transmission request; otherwise, the channel is
available and the packet will be transmitted immediately.
Based on the CSMA/CA policy above, we add a new
component, CCA-Adjustor, in our design to adjust the CCA-
threshold dynamically, instead of using the fixed value. The
purpose of such CCA-Adjustor is to relax the CCA-threshold
according to different transmitting power of interference, for
leveraging neighboring-channel concurrencies while avoiding
the co-channel collisions.
B. Design of CCA-Adjustor
There are two types of interference information we could
obtain from the sensor node: 1) RSSI of co-channel interfer-
ence packets; 2) in-channel sensing power which means the
signal power of current using channel, including not only co-
channel packets but also inter-channel interference. To avoid
the co-channel collisions, a safe principle for CCA-threshold
setting is to be smaller than the power level of any co-
channel interference packets and with such constraint as high
as possible to introduce the concurrencies on non-orthogonal
channels.
Concretely, CCA-Adjustor modify the CCA-threshold in
two phases: Initializing Phase which focuses on getting a con-
servative initial CCA-threshold setting without any relaxing;
and Updating Phase which is devoted to update the CCA-
threshold according to most recent interference traffics.
1) Initializing Phase: In the initial stage when sensor
nodes have just started, aggressive CCA-threshold setting
may introduce unexpected co-channel interference. Thus, the
CCA-threshold should be determined cautiously to avoid any
potential co-channel interference in the initializing phase.
More specifically, assume we obtain the RSSI SI
channel interference packet MSGI
the in-channel sensing power sequence of {PI
threshold CCAIshould satisfy:
∀i, CCAI < SI
and be even lower as shown in Fig. 12 (2) to avoid the
potential co-channel interference occurring in the gap between
current co-channel and inter-channel interference records:
CCAI = min{SI
iof co-
iin the initial stage and
j}. The CCA-
i
(1)
1,SI
2,...,max{PI
1,PI
2,...}}
(2)
362362