This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2018.2808968, IEEE
Transactions on Mobile Computing
1

mSync: Physical Layer Frame

Synchronization Without Preamble Symbols
Bastian Bloessl, Member, IEEE, Falko Dressler, Fellow, IEEE

Abstract—We present a novel physical layer frame format and a corresponding decoding strategy for energy-constraint single-carrier
transceivers, often used in sensor networks and cyber-physical systems. The main advantage of our approach is that decoding does not
rely on dedicated preamble symbols, which usually introduce considerable overhead in terms of energy consumption and utilization of the
wireless channel. We show that omitting the preamble can be achieved by buffering the signal in the receiver and processing the samples
twice; first to synchronize and in a second iteration to decode the actual data. To introduce our approach, we provide a theoretical
description, including a discussion of the implications of synchronizing on data symbols instead of optimized preamble sequences. The
practical feasibility of the algorithm is shown by simulations and experiments using prototype implementations based on software defined
radio. We implemented our algorithm for two technologies, a custom ultra low-power BPSK transceiver and the O-QPSK physical layer of
the IEEE 802.15.4 standard. Finally, we present an extension of the algorithm that allows us to reduce the buffered data to a small
constant number of samples, making our algorithm applicable to physical layers independent from their maximum frame size.

Index Terms—Physical Layer, Frame Synchronization, Preamble, Low-Power Communications, Software Defined Radio

1 I NTRODUCTION In this paper, we propose mSync, a novel orthogonal

Today, people as well as an ever increasing number of things concept that can complement existing approaches. The core
around us are connected every time and everywhere by idea, which we briefly presented in [11], is to decode the
means of wireless communications. To support a diverse frame without relying on dedicated preamble symbols. By
set of applications, a broad spectrum of technologies is in avoiding the overhead of a dedicated preamble, the frames
use. On the one end of the spectrum, there are performance- are shorter, which saves energy and reduces occupancy of
oriented technologies like LTE or WiFi that employ advanced the wireless channel.
concepts such as multi-user MIMO to boost throughput and The improvement can be considerable, especially since
spectral efficiency. On the other end of the spectrum, there are the frame sizes in WSNs are usually small. As an example,
low-power and low-bandwidth single carrier technologies we consider the Offset Quadrature Phase-Shift Keying (O-
that allow for energy efficient operation. Especially with the QPSK) physical layer of the IEEE 802.15.4 standard, which
proliferation of Industry 4.0 and the vision of an Internet of forms the base of the ZigBee stack [12], [13], a popular
Things (IoT) [1], the second kind of devices has shifted into choice for industrial automation and IoT applications. Here,
the focus of research and development. It is well recognized the minimum preamble length is the equivalent of 4 byte
that energy efficient wireless technologies form the central compared to the total size of an Acknowledgement Frame
building block for many applications including Wireless Sen- (ACK) of 11 byte. Admittedly, an ACK presents an extreme
sor Networks (WSNs) [2], Cyber-Physical Systems (CPSs) [3], case where our algorithm can provide a large benefit, but we
industrial automation, wildlife monitoring [4], [5], and medi- also have to consider that in unicast networks they comprise
cal implants. More recently, also car and plane manufacturers 50 % of the frames. Moreover, even a maximum sized frame
are looking into replacing wired system with wireless sensors with 133 byte in total, the overhead of 4 byte is still non-
to save cabling and, thus, weight and fuel. In this paper, we negligible.
will use the more traditional term WSNs when referring to Apart from the O-QPSK physical layer, we apply our new
such energy efficient wireless communication technologies. concept also to a custom ultra low-power transceiver that
In WSNs, the design is mainly affected by the limited we designed for a 2 g sensor mote, used to track bats in their
energy budget available to the device. The challenge is to find natural habitat [4]. In the paper, we will refer to this custom
a good trade-off between functionality and performance, on implementation as B ATS transceiver. Due to architectural
the one hand, and network lifetime [6], on the other hand. To constraints of the sensor mote, we are only able to send very
optimize energy consumption and increase network lifetime, short frames with a total length of 12 byte. In our initial
researchers developed concepts that range from transmit design, we allocated 2 byte, i.e. over 16 %, for the preamble.
power control [7], over duty-cycled MAC protocols [8], [9], This overhead can be completely avoided with mSync.
to wake-up receivers [10]. We introduce our idea by providing a thorough theoretical
description that considers the implications of synchronizing
• B. Bloessl and F. Dressler are with the Heinz Nixdorf Institute and the on data symbols instead of optimized preamble sequences.
Dept. of Computer Science, Paderborn University, Germany, E-mail: To show the practical feasibility and the general applicability
{bloessl, dressler}@ccs-labs.org of the approach, we implement SDR-based prototypes for
• B. Bloessl is also with the Connect Center, Trinity College Dublin, Ireland
two very different technologies: the B ATS transceiver and the

1536-1233 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2018.2808968, IEEE
Transactions on Mobile Computing
2

IEEE 802.15.4 O-QPSK physical layer. Using these prototypes, mechanism saves energy from idle-listening, which would
we assess the performance by means of simulations as well as otherwise consume considerable amounts of energy [15].
experiments. The results highlight that our approach allows Duty-cycling protocols can be divided into synchronized [8]
us to eliminate the overhead of preamble symbols without and unsynchronized [17], [18] approaches. With synchro-
degrading physical layer performance. On the contrary, nized algorithms, the nodes align their duty cycles to agree
omitting the preamble reduces the air-time of the frame, on possible time slots for transmissions. Such approaches
reducing occupancy of the wireless channel. To study this make sending less costly, but introduce signaling overhead
effect in greater detail, we employ a Markov model of the to establish a common time base. Using unsynchronized
IEEE 802.15.4 MAC to show how shorter frames can improve protocols, a sending node does not know about the duty cycle
the saturation goodput. of its neighbors and, therefore, has to extend its transmissions
However, also with mSync, there is no free lunch. The to assert that each node woke up and got a chance to receive
drawback of a naïve implementation is that the receiver the frame, making transmissions costlier in terms of both
has to buffer samples corresponding to a maximum sized energy and channel utilization.
frame. Depending on the technology, this could waste a In that context, a note on terminology is very important:
lot of resources and void the advantage in terms of energy When discussing MAC protocols, many authors refer to a
consumption. To address this issue, we present a slight preamble as a signal that is used to wake up duty-cycled nodes,
variation of the algorithm that allows us to reduce the announcing a transmission. Such mechanism is, for example,
buffered data to a small number of samples independent from required with unsynchronized MAC protocols. This use of
the maximum frame size. With this, we show that mSync is a preamble has to be clearly distinguished from the physical
applicable to many packet-based single-carrier technologies layer preamble, sometimes also called physical layer training
and presents an interesting new option in the design space sequence, which we discuss in this paper. A receiver uses
of low-power wireless communications. the physical layer preamble to synchronize on the signal by
Our main contributions can be summarized as follows: estimating parameters like frequency and clock offsets. Our
approach, which allows saving the overhead of preamble
• We present a novel frame format and a corresponding
symbols, is a pure physical layer concept and, therefore,
decoding strategy that allows to save the overhead of
independent from MAC layer algorithms. In fact, our idea
preamble symbols without degrading physical layer
can be complemented with duty-cycled MAC protocols.
performance. Another strategy to save energy is to use transmit power
• We demonstrate the feasibility and general applica- control [7]. Intuitively, a transmitter can use lower transmit
bility of our approach by implementing SDR-based power when addressing nodes with low channel attenuation,
prototypes for two communication technologies: our such as nodes in close proximity or with unobstructed line
B ATS transceiver and the O-QPSK physical layer of of sight. By decreasing the transmit power, the sender
IEEE 802.15.4. can save energy, which would otherwise be wasted in
• We investigate the reduced energy consumption, the power amplifiers. A potential drawback is that the
discuss the possibility to use more robust receiver sender needs feedback to estimate the channel attenuation,
configurations, and show how shorter frames (i.e., making it a good candidate in relatively static scenarios with
frames without a preamble) can considerably increase bidirectional communication.
system level goodput. A more recent trend is to drop support for complex
• We finally show that, with a small variation of the mesh network topologies, which are, for example, part of
algorithm, the amount of data that has to be buffered IEEE 802.15.4 and stick to simpler star networks. This is
can be reduced to a small fixed number of samples, especially visible with the very successful Bluetooth Low
making our approach applicable also to physical Energy (BLE) standard, whose energy efficiency results
layers with large maximum frame sizes. mainly from changes on the link and network layer [19].
Even though the physical layer was slightly adapted towards
2 R ELATED W ORK larger channel bandwidths, the main improvements stem
from simplified network structures that allow saving energy
Driven by the idea of smart dust [14], many researchers
through less overhead in maintaining connections.
began working towards distributed, decentralized, and self-
organizing networks of small sensor nodes, opening up
2.2 Ultra Low-Power Communication
the field of WSNs [2]. To allow infrastructure-less ad hoc
deployments, sensor motes have to be self-powered, which More recent advancements, sometimes called ultra low-power
shifts energy efficient operation into the focus when reaching communications, allow for even smaller sensor motes that
for long network lifetimes [6]. Given the fact that wireless support network lifetimes of up to several years. Such
communication typically accounts for a large fractions of the motes are based on wake-up receivers that are most of the
overall power consumption [15], the transceiver and espe- time completely switched off (or, to increase sensitivity, in
cially the MAC layer were the subject of many studies [16]. a very low-power mode) and only activated from actual
transmissions [20]. The basic principle is similar to Radio-
Frequency Identification (RFID) [21], where the transmitted
2.1 Energy-Efficient MAC signal induces current in the receiver, which is used to wake
The key concept for low-power MAC protocols is duty-cycling. up the communication module.
With duty-cycling, the transceiver is switched on and off The possibility to activate the radio module only during
following a schedule defined by the MAC protocol. This actual transmissions could be regarded as the ideal duty-

1536-1233 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2018.2808968, IEEE
Transactions on Mobile Computing
3

receiver, the preamble is used to synchronize on the frame

Preamble SFD Data by deriving local estimates for parameters like frequency
p1 ... pk s1 ... sm d1 ... dn offset, sampling clock offset, and symbol timing. This
is typically done with a feedback loop that locks on the
signal by adjusting the parameters according to an error
Data SFD SFD Data signal [25]. With symbol timing recovery, for example, the
dn ... d1 sm ... s1 s1 ... sm d1 ... dn
sampling points are gradually adjusted towards the largest
opening in an eye diagram. This stage usually employs
blind estimators that work without prior knowledge of the
data. For our algorithm, it is irrelevant which exact algorithm
or combination of algorithms is used in the receiver. The
Figure 1. Comparison of normal frames (top) and mSync (bottom). important aspect is that they adjust during reception, which
Instead of a preamble, we use the data to synchronize and calculate local is usually the case, also in state of the art algorithms [26].
estimates for parameters like frequency offset, sampling clock offset, and
symbol timing. This allows omitting the preamble and reduce the frame When designing a new physical layer, it can be chal-
size, as only the components with the solid outline are sent over the air. lenging to find a parameter set that offers a good trade-off
between performance and overhead in terms of preamble
symbols. Ideally the receiver would synchronize exactly at
cycle and allows us to overcome idle listening all together. the end of the preamble and reliably detect the SFD. In
This technology is an enormous step that opens the field practice, it involves a trade-off between the overhead (i.e.,
for completely new applications. In wildlife monitoring, for length of the preamble and, thus, time to lock) and the
example, we can use much smaller batteries and build light- reliability of frame detection. Using a longer preamble, the
weight sensor motes that can even be mounted on small, receiver has more time to synchronize and chances are higher
flying animals like bats [4] or crows [5]. that the SFD can be detected.
Another idea, also adopted from RFID, is the use of
backscattered signals for communication [22], [23]. Here, tags
can be completely passive as they do not have to generate a 3.1 Reversed Frame Structure
signal locally, but merely reflect the signal of an interrogator. While balancing this trade-off for a custom highly-optimized
The actual information is encoded by changing the antenna’s receiver, we thought of using a different frame format that
impedance, which varies the amplitude of the reflected signal. offers important advantages: it relaxes the requirements
While this is an interesting concept, it cannot be directly for the synchronization algorithm, while, at the same time,
applied to WSNs, as it relies on a very capable interrogator. reduces the frame size, improving the energy consumption.
However, with ambient backscatter [22] and bistatic scat- The core idea is to omit the preamble and adapt the frame
ter radio [23], this concept was extended and applied to format shown at the bottom of Figure 1. In contrast to normal
decentralized low-power networks such as WSNs. Instead frames, we send only the data symbols followed by the SFD,
of using an interrogator signal, these works use ubiquitous which might, in this case, better be called an end of frame
signals from broadcast radio, TV stations, or dedicated carrier delimiter. We will, however, stick to the term to emphasize
emitters. Such tags are not completely passive, but the power its correspondence with the SFD in a normal frame. When
consumption is greatly reduced since the signal does not have looking at Figure 1, it is important to note that the dashed
to be generated locally, saving the energy that is otherwise boxes are not sent over the air but added to illustrate the
used in power amplifiers. Very recently, a similar concept concept. Only the solid boxes depict the parts of the signal
was applied to IEEE 802.11b allowing for, what the authors that are actually transmitted. From the figure, we can see
call, Passive WiFi [24]. By modulating a WiFi frame from that mSync allows us to reduce the frame size, while carrying
another device, they are able to produce standard compliant the same amount of data.
IEEE 802.11b frames, while reducing the power consumption The most interesting part of mSync is the decoding
by orders of magnitude. strategy for this modified physical layer frame format. Given
To summarize, we presented a broad spectrum of ideas the fact that the receiver is not synchronized when the first
and concepts to design energy efficient wireless networks. data symbols arrive, it cannot decode the data directly.
With our approach, we introduce another, orthogonal option Instead, it locks on the frame by processing the signal
that can be beneficial for many low-power systems. By using the same blind estimation algorithm as the normal
removing the preamble without degrading physical layer receiver. The process might, however, require more input,
performance, mSync can complement the presented ap- as it uses data symbols instead of a preamble sequence that
proaches and offers a new strategy to further optimize energy is optimized for fast convergence of the synchronization
consumption of low-power transceivers. algorithm.
The crux of mSync is that it buffers samples, for example,
in a ring buffer. This allows us to process the signal twice;
3 P REAMBLE - LESS F RAME D ETECTION AND D E -
once to lock on the frame and another time to decode it.
CODING Using a simple implementation, the buffer size has to be
Before describing mSync, we should first recap the typical set according to the number of samples corresponding to a
physical layer frame format. As illustrated at the top maximum sized frame plus the SFD. We will later discuss
of Figure 1, it consists of a preamble, followed by the Start of ideas on how to reduce the amount of data that has to be
Frame Delimiter (SFD) and the actual data payload. In the buffered.

1536-1233 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2018.2808968, IEEE
Transactions on Mobile Computing
4

Using the data to synchronize, the algorithm has much Actual Symbol Timing
more time, i.e., the whole duration of the data symbols, Ω[n]
1

Signal Amplitude
to lock on the frame. When the receiver is synchronized,
it can recognize the SFD at the end of the frame. If that
B[n]
happens, the receiver keeps its state, i.e., its current estimates
of the signal parameters and, instead of continuing with the B[n+1]
normal sample stream, it processes the ring buffer in reverse Samples
direction and decodes the data. During this reverse operation, e[n]
the receiver traverses the sample stream again, processing -1
the samples in the order indicated by the dashed boxes Time
in Figure 1. This process can also be thought of mirroring
the received signal in time domain at the dashed vertical Figure 2. Illustration of the Mueller and Müller algorithm for timing recovery
and clock offset estimation.
line. The name mSync, for mirror synchronization, is derived
from this central characteristic of the algorithm.
Decoding the signal while traversing the buffer back- points of the SDR are marked as crosses, while the dots
wards also explains why we send the over-the-air signal indicate the points that the receiver considers for decoding
reversed (i.e., why we change the samples of the data from (i.e., the estimated symbol timing). Since the estimated
d1 , · · · dn to dn , · · · d1 ). This is not strictly necessary, but it symbols are not exactly at sample positions, the receiver
asserts that the output of mSync (the boxes with the dashed interpolates the values with a minimum mean squared
outline) correspond to the output of a normal receiver (cf. error FIR interpolator. When using mSync, the receiver
the right hand side of Figure 1). This has the advantage that uses the very same algorithm, but, in addition, stores the
the decoder outputs the exact same bit sequence as a normal samples in a ring buffer and compares the decoded values
receiver, which eases integration of our algorithm. Extending B̂[n − m + 1], · · · , B̂[n] with the reversed SFD sm , · · · , s1
a receiver with mSync is, therefore, straightforward and (cf. Figure 1). Once they match, the receiver continues
merely comprises replacing the synchronization algorithm. processing samples from the buffer. To foster compatibility
The other components can be left unchanged. with the normal receiver, we regenerate the preamble bits in
The part of the algorithm that we did not discuss yet the receiver and prefix them before every frame, just as if it
is how the receiver stays locked while switching directions. would have been received over the air. This way, mSync is
Since this depends on the actual algorithm that is used in the completely transparent for the rest of the receiver.
receiver, we can only discuss exemplary implementations.
In the following, we describe how our algorithm can be
used with the Mueller and Müller (M&M) algorithm [27] for 3.2 Synchronizing on Data
timing recovery. While the M&M algorithm is not state of One important difference of mSync is the use of an unknown
the art, it serves as a good example to describe the concept. data signal instead of optimized preamble sequences for
However, we argue that the general idea can also be applied synchronization. For frame-based systems that need the
to more complex state of the art algorithms like symbol receiver to re-synchronize on every frame, this could, in
timing recovery with polyphase filter banks [26]. The M&M theory, cause problems. Depending on the physical layer
algorithm implements a feedback system that performs and the receive algorithm, not all bit sequences might be
timing recovery and estimates the sampling clock offset, equally suited to derive signal parameters. A Binary Phase
i.e., it calculates estimates for the number of samples per Shift Keying (BPSK) signal that is all ones or all minus
symbol Ω and the position of the n-th symbol in the sample ones, for examples, cannot be used to extract symbol timing
stream B[n]. If we consider a real-valued binary signal that information. Fortunately, this is unlikely to happen, since
encodes its data with {1, −1}, the algorithm calculates the a state-of-the-art physical layer uses a scrambler if the data
error feedback e after decoding the n-th symbol as tends to include long strings of ones or zeros. Also from
an information theoretic perspective, the physical layer will
e[n] = B̂[n − 1] B[n] − B̂[n] B[n − 1], (1) be tuned towards equally probable symbols to maximize
entropy and, thus, self-information.
where B̂ is the, probably corrupt, decoded symbol with To better understand the implications of using mSync,
B̂ ∈ {1, −1}. This error signal is used to adjust the estimates we study how locking on data symbols differentiates from
Ω and B as ideal preamble sequences. Our motivation for this study is
twofold. First, we want to understand how much longer
Ω[n + 1] = Ω[n] + gΩ e[n] (2) it takes to synchronize on a random data pattern. Second,
we want to highlight another potential benefit of mSync.
B[n + 1] = B[n] + Ω[n] + gB e[n]. (3)
Using the data to synchronize, we have much more time to
Here gΩ and gB are gains for the error feedback of the lock on the frame. This allows us to reduce the gain of the
corresponding parameter that can be used to adjust the error feedback in the controller, which slows convergence
sensitivity of the controller. of the synchronization algorithm, but also reduces the noise
An exemplary iteration of the algorithm is illustrated feedback of the controller, making it more stable.
in Figure 2, which shows the noise free analog signal as To show this effect, we set up simulations to compare
a dotted line. The actual symbol timing, unknown to the the ideal locking sequence for the M&M algorithm (i.e.,
receiver, is indicated by the solid vertical lines. The sampling 00110011 · · · ) with a random bit pattern. Similar to our

1536-1233 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2018.2808968, IEEE
Transactions on Mobile Computing
5

normal normal
mSync mSync
20 30 40 50

20 30 40 50
Γ = 10dB, gB = 0.6 Γ = 10dB, gB = 0.3
Phase Error (in %)

Phase Error (in %)

error floor ≈ 16%
error floor ≈ 10.5%
10

10
0

0
20 40 60 80 20 40 60 80
Bit Bit
(a) Faster convergence. (b) Slower convergence.

Figure 3. Impact of the error feedback gain on the time that it takes to lock. With mSync, we have more time to synchronize, allowing us to choose a
setting with slower convergence, but lower error floor.

B ATS transceiver (described in more detail in Section 5), experiments are valid for the M&M algorithm, we expect
we produce a BPSK signal with five samples per bit and similar qualitative behavior also for other synchronization
apply a matched filter. The resulting sample stream is passed algorithms.
through an Additive White Gaussian Noise (AWGN) channel
to produce a signal with a given target Signal to Noise Ratio
(SNR) Γ. To rule out interactions from level controllers and
4 I MPLEMENTATION
to isolate the effect of locking on the data, we scale signal To study our algorithm and to show the feasibility of the
and noise, such that the average signal S plus the average approach, especially that it is possible to stay locked while
noise power N equals unity switching directions, we implemented the algorithm for
GNU Radio, a real-time signal processing framework for use
Γ 1 in SDR systems [28]. In contrast to, for example WARP [29],
S=, N= . (4)
1+Γ 1+Γ GNU Radio implements signal processing on a General
To keep the example simple, we fix Ω to the correct Purpose Processor (GPP), like a normal PC, which lends
value and record the average phase error dependent on the itself well for rapid prototyping [30]. Using GNU Radio,
sample of the frame. The average phase error after the n- signal processing is split in blocks that implement specific
th bit for 100k frame transmissions at an SNR of 10 dB are functions like filters, resamplers, and modulators. To exploit
depicted in Figure 3. With Ω = 5 the maximum phase modern multi-core CPUs, signal processing is parallelized by
error is 2.5, which we set as 100 %. In Figure 3a, we show starting each block in a separate thread. Compared to itera-
a configuration with a rather high error feedback (gB = tive programming environments, like M ATLAB, parallelized
0.6) and fast convergence. Using a normal frame with an processing adds complexity, but it is the central design point
optimized preamble sequence, the phase error stabilizes fast that enables real-time operation.
(after only about 8 bit). Such configuration might be used In GNU Radio, a transceiver is realized with a flow graph.
in a typical physical layer, where the preamble should be as It defines a specific configuration with a set of blocks, their
short as possible to reduce overhead. As expected, mSync parameters, and their connections. To get a better idea of
needs more time to lock on a frame. In this setup, it reaches the concept, Figure 4 shows a screenshot of the relevant
the error floor after about 20 bit. parts of our B ATS receiver in GNU Radio Companion, a
In Figure 3b, we show results for the same configuration, graphical frontend to setup and configure GNU Radio flow
but with reduced error feedback (gB = 0.3). While locking graphs. Integrating our algorithm in the existing receiver was
is slower in that configuration, the general trend is similar. straightforward. We merely had to change the blocks in the
mSync needs more time to lock than an optimized preamble shaded area, which contains the logic to switch between
(about 25 bit for a normal frame compared to about 50 bit the legacy M&M implementation and mSync. With the
for mSync). The advantage of the slower configuration is its Selector blocks, we can pipe the incoming sample stream
higher stability through lower noise feedback. In the slower either through the normal (top) or the mSync implementation
configuration, the phase error reaches an error floor of only (bottom). The other blocks of the receiver can be left
10.5 % compared to 16 % with the faster configuration. Such unchanged. They are used to demodulate the differential
configuration is clearly beneficial, but might not be suitable BPSK signal to a binary stream, search for the preamble
in a normal receiver. The longer convergence time, would ask sequence, and, once found, process the data payload.
for a longer preamble and, therefore, increase the overhead GNU Radio already comes with two very similar imple-
per frame, a problem that we do not face with mSync. mentations of the M&M algorithm; one for complex and one
To summarize, the results highlight mSync’s potential for real signals. We implemented our algorithm for both
to use more stable controller configurations with lower versions, since the B ATS receiver uses the complex variant,
noise feedback. While the quantitative results of these while the IEEE 802.15.4 receiver uses the real one.

1536-1233 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2018.2808968, IEEE
Transactions on Mobile Computing
6

Figure 4. The relevant part of our ultra low-power receiver in GNU Radio Companion. To support both the normal preamble as well as our
preamble-less reversed frame format, we merely had to introduce the possibility to switch the clock recovery algorithm (shaded area).

complexity at a minimum. To assess the performance of this

application-specific physical layer, we developed SDR-based
prototypes that we used for simulations and experiments in
realistic environments [32].

5.1 Physical Layer Performance

Using differential BPSK, the energy in the capacitor is just
Figure 5. The experiments are conducted in an office environment, using enough to send 12 byte frames at a data rate of 200 kbit/s.
GNU Radio together with Ettus Research B210 SDRs. Without mSync, a frame comprises a 2 byte preamble, a
1 byte SFD, and a 2 byte Cyclic Redundancy Check (CRC),
leaving 7 byte for the payload. Given the short frames size,
Apart from rapid prototyping, an important advantage of we stick to error detection using the CRC, but do not employ
GPP-based SDR architecture like GNU Radio is the possibility forward error correction. The SDR implementation uses
to use the same code for simulations as well as experiments. Ω = 5 samples per symbol, which corresponds to a sample
Since signal processing is implemented on the PC, we can rate of 1 Msps. The other parameters of the M&M algorithm
either loop back the samples from the transmitter into the were set to the default values used in GNU Radio version
receiver to perform simulations or connect SDRs for over- 3.7. Finally, we use a moving average over the duration of
the-air experiments. In this paper, we use this feature five bits to normalize the signal to an average power of one,
and evaluate both transceivers by means of simulations before feeding it to the clock recovery algorithm.
over an AWGN channel and real transmissions in an office When conducting simulations using the GNU Radio
environment. implementation of our new algorithm, the first observation
is that it works in the first place. This serves as a proof-
5 C ASE S TUDY 1: BATS T RANSCEIVER of-concept for mSync, but, of course, we wanted to go
The first use-case of our algorithm is the B ATS transceiver, further and provide a quantitative analysis of the impact on
a custom ultra low-power transceiver, that we developed physical layer performance. For this, we conduct simulations
to track mouse-eared bats (Myotis myotis) in their natural where we send frames with a pseudo-random payload over
habitat [4].1 With a weight of about 20 g, these bats can an AWGN channel and measure the frame error rate. In
only be equipped with sensor motes of up to 2 g. In our corresponding runs, i.e., with and without our algorithm, we
project, we used this weight budget for a 1 g mote that use the same channel coefficients and payloads. Furthermore,
is powered by a 1 g battery. This weight constraint has we set the noise to an average power of one and adapt
distinct impact on wireless communications and is asking the amplitude of the signal to reach the desired SNR. The
for a custom application-specific design. The main limitation resulting frame delivery ratio at different SNR levels is
of out platform is that the battery does not supply enough depicted in Figure 6a. The error bars in this and the
current to drive the wireless transceiver directly. Instead, we following plots depict the 95 % confidence intervals. For
have to use the battery to charge a capacitor, which is then better readability, we alter between plotting the confidence
used to power the transceiver and send short bursts of data. intervals for the configuration with normal frames and the
More details and rationale on the design and implementation configuration with mSync.
of this ultra low-power transceiver are available in prior The results show that we start receiving frames at around
work [4], [31]. 0 dB and reach 100 % at about 7 dB. More interestingly,
In the context of this paper, the important aspect is that the graph shows that the two modes behave exactly the
the maximum frame size is limited by the energy stored same, which is a very positive result. It shows that the
in the capacitor. Furthermore, we are restricted to simple improved energy consumption of mSync (through shorter
modulation and coding schemes to keep the sensor mote frames without preamble symbols) does not have to be traded
for physical layer performance. This is the best possible
1. http://www.for-bats.de/ outcome, since the goal was not to improve the frame error

1536-1233 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2018.2808968, IEEE
Transactions on Mobile Computing
7

1
Frame Delivery Ratio
Frame Delivery Ratio
0.75

0.75
0.5

0.5
0.25

0.25
normal normal
mSync mSync
0

0
−2 0 2 4 6 8 10 −2 0 2 4 6 8 10
SNR (in dB) Relative Gain (in dB)
(a) Simulations over an AWGN channel. (b) Experiments in an office environment.

Figure 6. Observed packet delivery ratio of the B ATS PHY in simulations and measurements.

rate, but to benefit from shorter frames without degrading

normal mSync
physical layer performance.
To rule out potential simplifying assumptions in our

Low-Power
simulations, we set up real over-the-air measurements. With

Technology
GNU Radio, switching between simulations and real exper-
iments is straightforward. The possibility to use the same
code in simulations and measurements is a big advantage, as
it allows us to directly compare the results. Figure 5 shows
802.15.4

our measurement setup. In this and the following exper-

iments, we do not employ any unrealistic simplifications
(like using a common clock source for sender and receiver).
The oscillators in sender and receiver run completely in-
0.0 0.2 0.4 0.6 0.8 1.0
dependently, leading to typical hardware impairments like
Relative Frame Size/Energy Consumption
frequency offsets and sampling clock offsets. We use two
B210 SDRs from Ettus Research, which we configured to
Figure 7. Comparison of the relative frame size/the relative energy
send in the 868 MHz band. While we cannot strictly rule consumption between normal frames and mSync. For IEEE 802.15.4,
out interference, we chose a frequency that seemed to be the graph shows 30 byte packets, as used in our simulations and
vacant in our lab. To set different SNR levels, we kept the measurements.
receive gain constant and varied the output power of the
transmitter. Since the B210 is no calibrated measurement
device, this method allows us to change the relative SNR, but shorter frames translate directly into energy savings in the
not to configure an absolute level. Like in the simulations, transmitter, as the sensor mote’s main task is to send periodic
we use 5 samples per symbol, resulting in a sample rate of beacons to the ground network. Apart from this specific
1 Msps. case, also other communication modules could benefit from
The results of the experiment are depicted in Figure 6b, mSync, especially when the analog radio front end (in
where we plot the frame deliver ratio at different SNRs. For particular the power amplifiers) are responsible for a large
better comparison, we shifted the relative SNR on the x- part of the energy consumption. For the receiver, the energy
axis to values corresponding to our simulations. The graph consumption is not straightforward to quantify, since the
shows that the experiments are in perfect compliance with algorithm introduces a slight computational overhead and
our simulations. Again, mSync and normal operation do not requires to buffer samples corresponding to the maximum
show any differences, highlighting that our approach does frame size. A detailed comparative study between an SDR
not degrade the frame error rate. The graph shows two slight implementation of a normal receiver and mSync is, therefore,
dips in the curve, which can be explain by nonlinearities of presented separately in Section 8.
the amplifiers. This is in accordance with the datasheet of While optimized energy consumption per frame is al-
the Analog Devices AD9361 transceiver used with the B210. ready a good argument for our algorithm, wildlife monitor-
ing is a prime application also for other reasons. WSNs for
wildlife monitoring are usually heterogeneous with energy-
5.2 Energy Savings constraint mobile nodes on the animals that send data to
For our custom B ATS transceiver, mSync allows us to shorten more capable stationary nodes. These stationary nodes are
the frame by 17 % (through saving the 2 byte preamble of typically SDR-based, implementing a custom and application-
the 12 byte frame). This improvement is visualized at the specific physical layer. In such scenarios, our algorithm is
top of Figure 7, where we plot the frame size of mSync straightforward to apply: SDRs are easy to extend with
relative to a normal frame. In our particular case, the the required functionality and the mobile nodes rely on a

1536-1233 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2018.2808968, IEEE
Transactions on Mobile Computing
8

ratio for different SNRs. As in previous figures, the error bars

Length
SFD
Preamble Payload CRC indicate the 95 % confidence intervals. Like with the B ATS
transceiver, the main observation is that the modes behave
exactly the same. That means that also with this transceiver,
4 byte 1 byte 1 byte 2 byte we can benefit from mSync without degrading physical layer
performance.
Figure 8. Structure of an IEEE 802.15.4 frame. The payload size can be Again, we wanted to back up our results with real experi-
up to 125 byte. ments and conducted measurements in an office environment.
We used the same B210 SDRs, this time transmitting in the
2.4 GHz band. Since this band is very crowded, we chose
custom design either way. With these modifications, the
a channel at the upper end, as there are no WiFi networks
more capable stationary node can use the slightly more
allowed in our region. Using this part of the spectrum, we
complex algorithm, while the mobile node can benefit from
were able to avoid most interference sources and had stable
considerable improvements in terms of energy consumption.
experimental conditions. Given the previously discussed
limitations of the B210, we plot the relative SNR and align
6 C ASE S TUDY 2: IEEE 802.15.4 T RANSCEIVER the x-axis for better comparability with the simulation results.
The resulting graph is shown in Figure 9b. Also in this
Motivated by the promising results of the B ATS transceiver,
experiment, the results match very well with simulations,
we were curious to apply the concept also to more complex
proving that mSync is feasible in practice and that we can
transceivers. We chose the IEEE 802.15.4 O-QPSK physical
benefit without any drawbacks in terms of physical layer
layer for the 2.4 GHz band, which forms the base of the Zig-
performance.
Bee stack. This physical layer is designed to provide energy
efficient communication for WSNs and IoT applications and
could, therefore, greatly benefit from the energy savings 6.2 Impact on MAC Layer Goodput
provided by our approach.
While optimizing the energy consumption was the main
motivation to introduce mSync, shorter frames also reduce
6.1 Physical Layer Performance occupancy of the wireless channel, potentially improving
Fortunately, there is already an Open Source implementation network goodput. For our B ATS transceiver, this aspect is not
of IEEE 802.15.4 available for GNU Radio. This implementa- of prime interest, as infrequent transmissions lead to low net-
tion was started by Thomas Schmid [33] and later overhauled work utilization. For IEEE 802.15.4, in contrast, the maximum
by us in [34].2 Based on O-QPSK, the IEEE 802.15.4 physical achievable goodput can be a relevant aspect. Quantifying the
layer is slightly more complex. To encode the data, the impact of shorter frames is, however, not straightforward,
transmitter maps each group of 4 bit to one of 16 pseudo- as the relationship between the overhead per frame and
noise chip sequences. These 32 bit chip sequences are then network goodput is non-trivial. The complexity stems from
O-QPSK modulated to create a signal with a chip rate of the slotted operation of the channel access algorithm and the
2 Mcps. In the SDR transceiver, we process the signal with a fact that we use the channel more efficiently once we access
sample rate of 4 Msps. it.
The IEEE 802.15.4 GNU Radio module already uses the To study the possible improvements of our algorithm, we
M&M algorithm. Integration of our algorithm into the employ the Markov model presented in [35]. This model
receiver is, therefore, straightforward. As shown in Figure 4, considers the stationary throughput of a saturated IEEE
we merely have to replace the M&M block with our modified 802.15.4 network. While the standard defines several network
version. All parameters and the other components of the topologies and modes of operation, we focus on a typical
receiver are left unchanged. The frame format of a normal network, consisting of a Personal Area Network (PAN)
IEEE 802.15.4 frame is shown in Figure 8. Each frame consists coordinator that orchestrates nodes in a star topology. In
of a 4 byte preamble, 4 byte physical layer overhead (for SFD, such networks, the coordinator establishes a superframe cycle
a header, and the CRC), and the data payload of up to that is used to subdivide time into a Contention-Free Period
125 byte. Using our algorithm allows us to save 4 byte of all (CFP) and a Contention Access Period (CAP). The channel
frames, independent from the their total size. For an ACK access during the CAP uses slotted Carrier Sense Multiple
with a total length of 11 byte, this corresponds to 36 %. But Access with Collision Avoidance (CSMA/CA) with a slot
even for a full-sized frame with a total length of 133 byte, the length of aUnitBackoffPeriod that has a duration of 20 symbols.
improvement is still 3 %. For the sake of brevity, we only consider the CAP and
Similar to the previous use-case, we start our evaluations unacknowledged transmissions in our scenario. The model
with simulations over an AWGN channel. We send 30 byte is, however, straightforward to extend to acknowledged
frames with a pseudo random payload and record the frame transmissions and unsaturated conditions, as shown in [35].
delivery ratio. The relative improvement of mSync for a Furthermore, the battery life extension mode, which allows
30 byte frame is depicted in Figure 7. With mSync, we can reducing the size of the initial backoff window, is considered
reduce the air time by 11 % from 1.15 ms to 1.02 ms, while to be disabled.
transmitting the same data. The results of the simulations Similar to Wireless LAN (WLAN), IEEE 802.15.4 uses
are depicted in Figure 9a, where we plot the frame delivery an exponentially increasing backoff algorithm for channel
access. The backoff windows starts at 2macMinBE slots, which
2. https://www.wime-project.net/ are doubled in each round until it reaches a maximum size

1536-1233 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2018.2808968, IEEE
Transactions on Mobile Computing
9

1
Frame Delivery Ratio

Frame Delivery Ratio

0.75

0.75
0.5

0.5
0.25

0.25
normal normal
mSync mSync
0

0
0 2 4 6 8 10 12 0 2 4 6 8 10 12
SNR (in dB) Normalized Gain
(a) Simulations over an AWGN channel. (b) Experiments in an office environment.

Figure 9. Observed packet delivery ratio of the IEEE 802.15.4 PHY in simulations and measurements.

channel twice in two consecutive slots and transmit only if

0.75

10 nodes the channel is sensed idle in both slots. To incorporate this

5 nodes mechanism in the Markov model, Bianchi’s key assumption
Normalized Goodput

of constant and independent collision probabilities had to be

0.5

adapted to a constant and independent probability φ that a

node starts sensing the channel in a randomly selected slot.
At that first channel assessment, the node senses the channel
0.25

busy with a probability α. If the channel is free, the node

15 nodes continues with the second channel assessment during which
normal
it senses the channel busy with a probability β.
mSync The detailed derivation of α, β, and φ is available in [35].
0

In our context, the most interesting result from the paper,

0 20 40 60 80 100 120
is the normalized saturation throughput S that defines the
Payload (in byte)
fraction of time slots spent to transmit non-colliding frames
that contribute to the throughput of the network. It is
Figure 10. Normalized saturation goodput for our approach and normal
frames for typical network sizes and an increasing payload length. calculated as
S = LN φ(1 − φ)N −1 (1 − α)(1 − β) , (5)
of 2macMaxBE slots. Frame transmission fails, if the MAC went where N represents the number of nodes in the network and
through macMaxCSMABackoff rounds without finding the L the length of a frame in time slots. Rounding the size of a
channel idle when trying to send. In that case, the frame is frame to a complete backoff slot, we use the payload size P
dropped, while the algorithm restarts with the next frame in and the overhead O of a frame to calculate its length as
the queue. The most important MAC layer parameters are 
(P + O)RS

also summarized in Table 1. All parameters correspond to L= . (6)
aUnitBackoffPeriod
their default values given in the standard [13].
The Markov model that we use to calculate the saturation Both P and O are given in byte and are converted to
goodput is much inspired by Bianchi’s seminal work on the symbols with a rate of RS = 2 symbols per byte. As payload,
Distributed Coordination Function (DCF) of IEEE 802.11 [36]. we consider the data in the MAC layer, denoted as Payload
Compared to WLAN, the main difference is that sensor in Figure 8, i.e., the MAC frame excluding the CRC. At this
nodes enter a power saving state during their backoff periods point, it is important to differentiate between throughput and
and, therefore, do not recognize if the channel turns busy goodput to understand the advantage of our approach. By
while they are waiting. After the backoff, nodes sense the omitting the preamble, we reduce the overhead from 8 byte
per frame (4 byte preamble, 1 byte SFD, 1 byte length field,
2 byte CRC), to 4 byte. If this difference allows us to save
Table 1 a slot, we use the channel more efficiently as the average
Most relevant MAC layer parameters. goodput per slot G is increased. We calculate G as
P Rs 1
Parameter Value G= · . (7)
aUnitBackoffPeriod L
aUnitBackoffPeriod 20 Symbols
macMinBE 3 Scaling the normalized throughput S with the efficiency
macMaxBE 5 of the channel access G, we calculate the normalized goodput
macMaxCSMABackoff 4 for all payload sizes of up to 125 byte. The results for
networks of 5, 10, and 15 nodes are depicted in Figure 10. In

1536-1233 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2018.2808968, IEEE
Transactions on Mobile Computing
10

mSync
Relative Goodput Improvement 5 nodes
0.05 0.1 0.15 0.2 10 nodes Data SFD
dn ... d1 sm ... s1
15 nodes

mSync++
Data SFD Data
dk ... d1 sm ... s1 dk+1 ... dn

(1)
(2)
0

0 20 40 60 80 100 120 Figure 12. Comparison of a normal mSync frame (top) and an optimized
Payload (in byte) version that requires less buffering in the receiver (bottom). With the
optimized version, the SFD is inserted after a fixed number of byte k. The
resulting frame is first decoded in backwards direction (1) and then like a
Figure 11. Relative goodput improvement of our approach over normal normal frame in forward direction (2).
frames for typical network sizes and an increasing payload length.

format is depicted in Figure 12. To decode the frame, we

accordance with [35], we can see that the goodput decreases
start like with normal mSync. We use the data to synchronize
with the number of nodes as well as with smaller frame
and search for the SFD. Once found, we store the state of the
sizes. Most importantly, however, we can see that our
receiver, and traverse the signal in reverse direction, while
approach performs better or at least equally good, with a
decoding the data. The difference is that we only have to
rather constant difference to normal frames.
go backwards for k byte. After that, we restore the state
The regular zigzag pattern stems from the slotted oper-
of the receiver to the state when we detected the SFD and
ation of the MAC layer, which uses slots with a duration
continue in forward direction. The frame is designed so that
corresponding to 10 byte. Since a frame blocks the channel
the decoded bits match with a normal frame, allowing us to
always for full slots, and since we save 4 byte from the
leave the rest of the receiver unchanged.
preamble, we benefit only in 4 of 10 cases.
To quantify the gain, we provide another view on the Since k is a fixed system parameter, the number of sam-
data in Figure 11, where we plot the relative improvement ples that are stored locally do not depend on the maximum
over normal IEEE 802.15.4 frames. The graph shows that, es- frame size, allowing the algorithm to scale. To optimize the
pecially for smaller payload sizes, the gain is significant with receiver, k could be used to balance the trade-off between the
increases of over 20 %. Even for large frames an improvement amount of data that has to be buffered and the probability
of about 5 % can be achieved. Furthermore, we can see a that the receiver is synchronized at the SFD. In other
higher improvement for larger networks, which, however, words, it balances the performance of the receiver against its
stems from their lower network goodput as opposed to complexity. With lower k, the SFD is placed earlier in the
higher absolute gains. frame, decreasing the data that has to be buffered. However,
To summarize, these analytical evaluations highlight that lowering k also gives the receiver less time to lock the signal,
our approach, which was initially motivated by optimizing potentially degrading performance.
the energy consumption of sensor motes, can also provide To demonstrate the practical feasibility of the approach,
considerable improvements in terms of network goodput. we also implement it on SDR. As discussed earlier, the
platform of our B ATS transceiver supports only short frame
sizes and, therefore, does not face the problem that mSync++
7 M INIMIZING S IGNAL B UFFERING solves. For that reason, we decided to implement our
We are aware that the need to buffer samples in the receiver improved algorithm for the IEEE 802.15.4 transceiver, where
is the main drawback of our approach. With the presented we can significantly reduce the amount of buffered data.
version of mSync, the amount of data that has to be buffered Apart from that, IEEE 802.15.4 is the more complex physical
would grow linearly with the signal bandwidth and the layer, highlighting the general applicability of the idea.
maximum frame size. This could easily render our approach In our proof-of-concept implementation, we varied k to
unfeasible in practice, especially when implementing it on the number of samples corresponding to 1 byte, 3 byte and
an integrated platform. 5 byte. Even with the largest k, we are able to reduce the
To address this issue, we introduce mSync++, a straight- amount of buffered samples by over 96 % for a full sized
forward extension of our algorithm. It is motivated by the frame (from 129 byte to 5 byte). However, the important
observation that the main problem with long frames is that aspect is that the buffered data is constant and independent
the receiver has to buffer the whole frame, even though from the maximum frame size. This makes mSync++ an
it is likely synchronized already after a few data symbols. interesting option also for other physical layers that support
In Section 3.2, we have seen that even with more stable larger frame sizes.
and slower converging configurations, the synchronization We start our evaluation with simulations over an AWGN
algorithm locks after a few byte. Therefore, with mSync++, channel. To ease comparison, we use 30 byte frames with
we do not put the SFD at the very end of the frame, but place pseudo-random payload, like in prior experiments. The
it latest after a fixed number of byte k. The resulting frame resulting packet delivery ratio for different SNR levels is

1536-1233 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2018.2808968, IEEE
Transactions on Mobile Computing
11

1
Frame Delivery Ratio
Frame Delivery Ratio

0.75
0.75

0.5
0.5

normal normal
mSync++, k=5 mSync++, k=5

0.25
0.25

mSync++, k=3 mSync++, k=3

mSync++, k=1 mSync++, k=1

0
0

0 2 4 6 8 10 12 0 2 4 6 8 10 12
SNR (in dB) Normalized Gain
(a) Simulations over an AWGN channel. (b) Experiments in an office environment.

Figure 13. Observed packet delivery ratio of the IEEE 802.15.4 PHY in simulations and measurements.

depicted in Figure 13a. For the sake of readability, we normal algorithm. Thus, the computational overhead of
did not plot the confidence intervals as the lines are very mSync is twofold: First, we have to save and restore the
close. The figure is, however, based on the same number of internal state of the algorithm when switching directions
measurements as the previous plots and showed a similar (with mSync++, the algorithm has to restore its internal state
confidence level. The results indicate that already low values before it can continue in forward direction). We believe that
for k (i.e., k = 3 and k = 5) provide very similar performance this overhead is negligible since it comprises only saving
as the normal receiver, which means we can benefit from and restoring of a few floating point numbers. Second,
shorter frames of mSync++ without suffering from frame parts of the sample stream have to be processed twice by
loss. Furthermore, the plot shows that a k of only 3 was the synchronization algorithm. For mSync, the number of
large enough to have the receiver synchronized with a high samples depends on the frame size, while for mSync++ the
probability. Otherwise, some SFDs would have been missed, number of samples depends on the placement of the SFD,
leading to worse performance compared to normal frames. i.e., the parameter k.
To validate the simulations and to rule out any unrealistic To present exemplary results, we prepared a sample
simplifications, we also conducted real over-the-air measure- stream with 30 byte IEEE 802.15.4 frames, as used in the
ments. We ran the same configuration with B210 SDRs in previous experiments. We used a sample rate of 4 Msps and
an office environment and varied the transmit gain to set an inter-frame space of 100 ms, corresponding to ten frames
different SNR levels. Using this method, we only know per second. The SNR was set to 30 dB to make sure that all
the relative change of the SNR, but not the absolute level. frames are received, i.e., that all frames go through the whole
In Figure 13b, we shifted the x-axis to a similar level as in the decoding process. For mSync++, we used k = 5, i.e., placed
simulations to ease comparison of the results. Overall, we the SFD after 5 byte.
see that mSync++ offers very similar performance as normal The resulting sample stream was loaded into memory
frames, proving the practical feasibility of our approach. and piped into the SDR receiver with its real-time sample rate
of 4 Msps. Using GNU Radio’s performance counters [37],
8 C OMPUTATIONAL C OMPLEXITY we monitored the CPU time of each block when running
We have already looked into most aspects of mSync. In the receiver on an Intel i7-7560U processor. To do this, we
particular, we showed through simulations and experiments developed a custom application that connects to the running
that it allows us to reduce the frame size without degrading flow graph, resets all performance counters, waits for 60 s,
physical layer performance. For the transmitter, this directly and writes CPU times of all block into a file. With this
results in energy savings since we are able to send shorter approach, we can perform precise measurements, which are
frames without introducing additional complexity. Therefore, not impacted by the start-up time of the flow graph.
mSync allows to trade-off energy consumption at the trans- The results of these measurements are depicted in Fig-
mitter against a more complex receiver. In the following, we ure 14, where we plot the CPU time of individual receiver
try to quantify this complexity. components during the 60 s measurement period. The Demod-
For that reason, we compare mSync and mSync++ with a ulator, Filter, and Subtract components are for demodulation
baseline transceiver that employs blind estimation for symbol and normalization of the signal level before feeding it to the
timing recovery and performs frame detection through synchronization algorithm. These three components are not
correlating with the SFD in subsequent stages. Such receiver affected by mSync and, therefore, show very similar CPU
was, for example, used for the SDR-based ground nodes of times in all modes.
the B ATS project and the GNU Radio IEEE 802.15.4 physical The most interesting component is the synchronization
layer. When evaluating the complexity, an important insight algorithm, which we labeled Sync in the figure. Already
is that mSync does not introduce a new signal processing in the normal configuration, it is the most demanding
algorithm, but merely changes the input that is fed to the component. When switching to mSync, we have to process

1536-1233 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2018.2808968, IEEE
Transactions on Mobile Computing
12

the contrary, through analytical evaluations, we were able

normal

Sync.
25 to show that the shorter frames can improve the goodput of
mSync networks considerably.
mSync++ Finally, we addressed the drawback of our algorithm, i.e.,
CPU Time (in sec)
20

Demodulator
the need to buffer samples in the receiver. With a simple
variation of the algorithm, we can reduce the buffered data
15

to a small constant number of samples, which makes our

approach applicable to physical layers independent from
10

their maximum frame size. We believe that the algorithm

Filter

Subtract

Decoder
occupies a sweet spot between performance and computa-
5

tional complexity, making it an attractive option for a broad

range of single carrier communication systems.
0

Receiver Component
ACKNOWLEDGEMENTS
Figure 14. Computational complexity of the IEEE 802.15.4 transceiver This work has been supported in part by the German
running in different modes. Research Foundation (DFG) under grant no. FOR 1508.

R EFERENCES
the 30 byte frame twice, which increases the overall CPU
[1] L. Atzori, A. Iera, and G. Morabito, “The Internet of Things: A
time from 17.8 s to 21.1 s. With mSync++, the part that has Survey,” Elsevier Computer Networks, vol. 54, no. 15, pp. 2787–2805,
to be processed twice is reduced to 5 byte, leading to a CPU Oct. 2010.
time of 20.4 s. [2] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci,
Another difference between both mSync variants and “A Survey on Sensor Networks,” IEEE Communications Magazine,
vol. 40, no. 8, pp. 102–114, Aug. 2002.
the normal receiver is that searching for the SFD becomes [3] F.-J. Wu, Y.-F. Kao, and Y.-C. Tseng, “From wireless sensor
part of the synchronization algorithm. Normally, the Decoder networks towards cyber physical systems,” Elsevier Pervasive and
component processes the continuous bit stream that is output Mobile Computing, vol. 7, no. 4, pp. 397–413, Aug. 2011.
[4] F. Dressler, B. Bloessl, M. Hierold, C.-Y. Hsieh, T. Nowak, R. Weigel,
by the synchronization algorithm. In this stream, it searches and A. Koelpin, “Protocol Design for Ultra-Low Power Wake-
for the SFD and, once found, decodes the data by demapping Up Systems for Tracking Bats in the Wild,” in IEEE International
the spreading sequences of the IEEE 802.15.4 physical layer Conference on Communications (ICC 2015). London, UK: IEEE, Jun.
to the data bits. With mSync, searching for the SFD becomes 2015, pp. 6345–6350.
[5] C. Rutz, Z. T. Burns, R. James, S. M. Ismar, J. Burt, B. Otis, J. Bowen,
part of the synchronization algorithm, since it has to know and J. J. S. Clair, “Automated mapping of social networks in wild
when to switch directions. That means that the increased birds,” Current Biology, vol. 22, no. 17, pp. R669–R671, 2012.
CPU times of the mSync variants also stem from the fact [6] I. Dietrich and F. Dressler, “On the Lifetime of Wireless Sensor
Networks,” ACM Transactions on Sensor Networks, vol. 5, no. 1, pp.
that this functionality is shifted from the decoder to the
1–39, Feb. 2009.
synchronization algorithm. In fact, when using mSync or [7] M. Kubisch, H. Karl, and A. Wolisz, “Distributed Algorithms for
mSync++, the CPU times of the decoder drop from 1.5 s to Transmission Power Control in Wireless Sensor Networks,” in IEEE
below 60 ms, making the values hardly visible in Figure 14. Wireless Communications and Networking Conference (WCNC 2003),
New Orleans, LA, Mar. 2003.
Overall, these measurements underline the practical [8] W. Ye, J. Heidemann, and D. Estrin, “An Energy-Efficient MAC
feasibility of our approach. While the absolute values Protocol for Wireless Sensor Networks,” in 21st IEEE Conference on
might vary depending on the platform, our experiment show Computer Communications (INFOCOM 2002). New York, NY: IEEE,
that the computational overhead of the mSync variants is Jun. 2002, pp. 1567–1576.
[9] O. Landsiedel, E. Ghadimi, S. Duquennoy, and M. Johansson, “Low
manageable. power, low delay: Opportunistic routing meets duty cycling,” in
11th ACM/IEEE International Symposium on Information Processing in
Sensor Networks (IPSN 2012). Beijing, China: IEEE, Apr. 2012, pp.
9 C ONCLUSION 185–196.
[10] I. Demirkol, C. Ersoy, and F. Alagoz, “MAC Protocols for Wireless
We presented a novel physical layer frame format and a Sensor Networks: a Survey,” IEEE Communications Magazine, vol. 44,
corresponding decoding strategy for single carrier wireless no. 4, pp. 115–121, Apr. 2006.
communication systems, as often used in Wireless Sensor [11] B. Bloessl and F. Dressler, “mSync - Frames without Preambles,”
in 21st ACM International Conference on Mobile Computing and
Networks (WSNs), Cyber-Physical Systems (CPSs), Internet Networking (MobiCom 2015), 4th ACM Software Radio Implementation
of Things (IoT) devices, and industrial automation systems. Forum (SRIF 2015), Poster Session. Paris, France: ACM, Sep. 2015,
Our approach works without dedicated preamble symbols, pp. 11–11.
[12] P. Baronti, P. Pillai, V. W. Chook, S. Chessa, A. Gotta, and Y. F. Hu,
which results in shorter frames with less physical layer
“Wireless Sensor Networks: a Survey on the State of the Art and the
overhead, saving energy and decreasing occupancy of the 802.15.4 and ZigBee Standards,” Elsevier Computer Communications,
wireless channel. To assess the performance of our approach vol. 30, no. 7, pp. 1655–1695, May 2007.
and to prove its feasibility, we incorporated it in two different [13] “Low-Rate Wireless Personal Area Networks (LR-WPANs),” IEEE,
Std 802.15.4-2011, Jun. 2011.
Software Defined Radio (SDR)-based prototypes: a custom Bi- [14] J. M. Kahn, R. Katz, and K. Pister, “Emerging Challenges:
nary Phase Shift Keying (BPSK) ultra low-power transceiver Mobile Networking for ’Smart Dust’,” Journal of Communications
and the Offset Quadrature Phase-Shift Keying (O-QPSK) and Networking, vol. 2, no. 3, Sep. 2000.
physical layer of the IEEE 802.15.4 standard. Both simulations [15] M. Stemm, R. H. Katz, and Y. H. Katz, “Measuring and Reducing
Energy Consumption of Network Interfaces in Hand-Held Devices,”
and over-the-air measurements showed that omitting the IEICE Transactions on Communications, vol. E80-B, no. 8, pp. 1125–
preamble did not degrade physical layer performance. On 1131, Aug. 1997.

1536-1233 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2018.2808968, IEEE
Transactions on Mobile Computing
13

[16] P. Huang, L. Xiao, S. Soltani, M. W. Mutka, and N. Xi, “The [36] G. Bianchi, “Performance Analysis of the IEEE 802.11 Distributed
Evolution of MAC Protocols in Wireless Sensor Networks: A Coordination Function,” IEEE Journal on Selected Areas in Communi-
Survey,” IEEE Communications Surveys and Tutorials, vol. 15, no. 1, cations, vol. 18, no. 3, pp. 535–547, Mar. 2000.
pp. 101–120, Feb. 2013. [37] T. W. Rondeau, T. O’Shea, and N. Goergen, “Inspecting GNU
[17] J. Polastre, J. Hill, and D. Culler, “Versatile Low Power Media Radio Applications with ControlPort and Performance Counters,”
Access for Wireless Sensor Networks,” in 2nd ACM Conference on in ACM SIGCOMM 2013, 2nd ACM SIGCOMM Workshop of Software
Embedded Networked Sensor Systems (SenSys 2004). Baltimore, MD: Radio Implementation Forum (SRIF 2013). Hong Kong, China: ACM,
ACM, 2004, pp. 95–107. Aug 2013, pp. 65–70.
[18] A. Dunkels, “The ContikiMAC Radio Duty Cycling Protocol,”
Swedish Institute of Computer Science, Tech. Rep. T2011:13, Dec.
2011.
[19] C. Gomez, J. Oller, and J. Paradells, “Overview and Evaluation
of Bluetooth Low Energy: An Emerging Low-Power Wireless
Technology,” Sensors, vol. 12, no. 9, pp. 11 734–11 753, Sep. 2012.
[20] I. Demirkol, C. Ersoy, and E. Onur, “Wake-up receivers for
wireless sensor networks: benefits and challenges,” IEEE Wireless
Communications, vol. 16, no. 4, pp. 88–96, Aug. 2009.
[21] E. Welbourne, L. Battle, G. Cole, K. Gould, K. Rector, S. Raymer, Bastian Bloessl received the diploma degree
M. Balazinska, and G. Borriello, “Building the Internet of Things in Computer Science from the University of
Using RFID,” IEEE Internet Computing, vol. 33, no. 3, pp. 48–55, Würzburg, Germany, in 2011. He is working
May/June 2009. toward the PhD degree at the chair for Distributed
[22] V. Liu, A. Parks, V. Talla, S. Gollakota, D. Wetherall, and J. R. Smith, Embedded Systems at the Heinz Nixdorf Institute
“Ambient Backscatter: Wireless Communication Out of Thin Air,” and the Dept. of Computer Science, Paderborn
in ACM SIGCOMM 2013. Hong Kong, China: ACM, Aug. 2013, University. In 2015, he won a FitWeltweit schol-
pp. 39–50. arship from the German Academic Exchange
[23] J. Kimionis, A. Bletsas, and J. N. Sahalos, “Increased Range Bistatic Service (DAAD), which funded a six-month stay
Scatter Radio,” IEEE Transactions on Communications, vol. 62, no. 3, in the research group of Prof. Mario Gerla at the
pp. 1091–1104, Mar. 2014. Computer Science Department of the University
[24] B. Kellogg, V. Talla, S. Gollakota, and J. R. Smith, “Passive Wi-Fi: of California, Los Angeles (UCLA). Since 2017, he has been a researcher
Bringing Low Power to Wi-Fi Transmissions,” in 13th USENIX at the CONNECT Center, Trinity College Dublin, Ireland’s Research
Symposium on Networked Systems Design and Implementation (NSDI Center for Future Networks and Communications, where he is funded
2016). Santa Clara, CA: USENIX, Mar. 2016, pp. 151–164. through a Marie Skłodowska-Curie fellowship. His research is focused on
[25] F. Harris, “Let’s Assume the System Is Synchronized,” in Glob- using software defined radio-based prototypes to assess the performance
alization of Mobile and Wireless Communications, ser. Signals and and robustness of vehicular and sensor networks. He is a member of the
Communication Technology, R. Prasad, S. Dixit, R. van Nee, and IEEE.
T. Ojanpera, Eds. Springer, 2011, pp. 311–325.
[26] F. J. Harris and M. Rice, “Multirate Digital Filters for Symbol
Timing Synchronization in Software Defined Radios,” IEEE Journal
on Selected Areas in Communications, vol. 19, no. 12, pp. 2346–2357,
Dec. 2001.
[27] K. Mueller and M. Müller, “Timing Recovery in Digital Syn-
chronous Data Receivers,” IEEE Transactions on Communications,
vol. 24, no. 5, pp. 516–531, May 1976.
[28] T. W. Rondeau, “On the GNU Radio Ecosystem,” in Opportunistic Falko Dressler received the MSc and PhD de-
Spectrum Sharing and White Space Access: The Practical Reality, grees from the Department of Computer Science,
O. Holland, H. Bogucka, and A. Medeisis, Eds. Wiley, May University of Erlangen, in 1998 and 2003, respec-
2015, pp. 25–48. tively. He is full professor of computer science
[29] A. Khattab, J. Camp, C. Hunter, P. Murphy, A. Sabharwal, and and chair for Distributed Embedded Systems at
E. W. Knightly, “WARP: A Flexible Platform for Clean-Slate the Heinz Nixdorf Institute and the Dept. of
Wireless Medium Access Protocol Design,” ACM SIGMOBILE Computer Science, Paderborn University, where
Mobile Computing and Communications Review, vol. 12, no. 1, pp. he is also a member of the University Senate.
56–58, Jan. 2008. He is associate editor-in-chief for Elsevier Com-
[30] G. Sklivanitis, A. Gannon, S. N. Batalama, and D. A. Pados, “Ad- puter Communications as well as an editor for
dressing Next-Generation Wireless Challenges with Commercial journals such as the IEEE Transaction on Mobile
Software-Defined Radio Platforms,” IEEE Communications Magazine, Computing, the IEEE Transaction on Network Science and Engineering,
vol. 54, no. 1, pp. 59–67, Jan. 2016. the Elsevier Ad Hoc Networks, and the Elsevier Nano Communication
[31] F. Dressler, S. Ripperger, M. Hierold, T. Nowak, C. Eibel, B. Cassens, Networks. He has been guest editor of special issues in the IEEE Journal
F. Mayer, K. Meyer-Wegener, and A. Koelpin, “From Radio on Selected Areas in Communications, the IEEE Communications
Telemetry to Ultra-Low-Power Sensor Networks: Tracking Bats Magazine, the Elsevier Ad Hoc Networks, and many others. He has been
in the Wild,” IEEE Communications Magazine, vol. 54, no. 1, pp. chairing conferences such as IEEE INFOCOM, ACM MobiSys, ACM
129–135, Jan. 2016. MobiHoc, IEEE VNC, IEEE GLOBECOM, and many others. He authored
[32] M. Nabeel, B. Bloessl, and F. Dressler, “On Using BOC Modulation the textbooks Self-Organization in Sensor and Actor Networks published
in Ultra-Low Power Sensor Networks for Wildlife Tracking,” in by Wiley & Sons and Vehicular Networking published by Cambridge
IEEE Wireless Communications and Networking Conference (WCNC University Press. He has been an IEEE Distinguished Lecturer as
2016). Doha, Qatar: IEEE, Apr. 2016, pp. 848–853. well as an ACM Distinguished Speaker. He is a fellow the IEEE as
[33] T. Schmid, “GNU Radio 802.15.4 En-and Decoding,” Networked well as a senior member of the ACM, and member of the GI (German
& Embedded Systems Laboratory, UCLA, Technical Report TR- Computer Science Society). He also serves on the IEEE COMSOC
UCLA-NESL-200609-06, Jun. 2006. Conference Council and the ACM SIGMOBILE Executive Committee.
[34] B. Bloessl, C. Leitner, F. Dressler, and C. Sommer, “A GNU Radio- His research objectives include adaptive wireless networking, self-
based IEEE 802.15.4 Testbed,” in 12. GI/ITG KuVS Fachgespräch organization techniques, and embedded system design with applications
Drahtlose Sensornetze (FGSN 2013), Cottbus, Germany, Sep. 2013, in ad hoc and sensor networks, vehicular networks, industrial wireless
pp. 37–40. networks, and nano-networking.
[35] S. Pollin, M. Ergen, S. C. Ergen, B. Bougard, L. Van der Perre,
I. Moerman, A. Bahai, P. Varaiya, and F. Catthoor, “Performance
Analysis of Slotted Carrier Sense IEEE 802.15.4 Medium Access
Layer,” IEEE Transactions on Wireless Communications, vol. 7, no. 9,
pp. 3359–3371, Sep. 2008.

1536-1233 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.