Ad Hoc Networks 89 (2019) 47–57

Contents lists available at ScienceDirect

Ad Hoc Networks
journal homepage: www.elsevier.com/locate/adhoc

Synchronization of data measurements in wireless sensor networks

for IoT applicationsR
Konstantinos Skiadopoulos a, Athanasios Tsipis a, Konstantinos Giannakis a,∗,
George Koufoudakis a, Eleni Christopoulou a, Konstantinos Oikonomou a,
George Kormentzas c, Ioannis Stavrakakis b
a
Department of Informatics, Ionian University, Corfu, Greece
b
Department of Informatics & Telecommunications, National and Kapodistrian University of Athens, Greece
c
Department of Information and Communication Systems Engineering, University of the Aegean, Samos, Greece

a r t i c l e i n f o a b s t r a c t

Article history: Time synchronization is a challenging problem for wireless sensor networks, as clocks deviate and sensor
Received 28 November 2018 measurements need to be associated with the reference clock time they took place. In Internet of Things
Revised 31 January 2019
(IoT) environments, crucial factors like energy, robustness, limited capabilities and extreme ambient con-
Accepted 5 March 2019
ditions (e.g., when nodes are deployed in soil) lead researchers to consider various schemes of relaxing
Available online 6 March 2019
the synchronization requirements. In this paper, a lightweight synchronization algorithm is proposed for
Keywords: wireless sensor networks, focusing on synchronizing the particular measurements in a per hop basis as
Data measurements synchronization they are transmitted encapsulated in data packets towards the sink node. The aim is to synchronize data
Wireless sensor networks measurements instead of node clocks, and thereby, induce negligible extra overhead, since extra messages
Internet of things and re-synchronization periods are not required. The analysis of the proposed algorithm regarding time
Clock deviation deviation and the corresponding variance shows a dependency on residual time (i.e., the particular time
period a data packet remains within a node), distance (i.e., the number of hops between the sink node
and the sensor node that sensed the data in the first place) and the average skew deviation value. Various
simulations were carried out aiming to exhibit thoroughly the effectiveness of the proposed algorithm,
as well as the validation of the analytical results. Findings from a comparison between the proposed al-
gorithm and a traditional synchronization algorithm along with their simulation results show that the
proposed algorithm can be a suitable solution for demanding environments like IoT systems.
© 2019 Elsevier B.V. All rights reserved.

1. Introduction
Wireless sensor networks have met a tremendous expansion in
the last 20 years, both in theory and applications [1,2]. The emer-
gence of this field, along with related practical achievements, have
an impact on various aspects of everyday life and industry, like
pollution monitoring, precision agriculture, etc. Each sensor node
senses some data of interest (e.g., humidity, temperature) and the
corresponding measurements need to be tagged with the actual
time they took place, i.e., the timestamp, in order to be further an-
R
Part of this work was originally published in the proceedings of the Second
International Balkan Conference on Communications and Networking (BalkanCom
2018), Podgorica, Montenegro, June 6–8, 2018.
∗
Corresponding author.
E-mail addresses: kskiado@ionio.gr (K. Skiadopoulos), atsipis@ionio.gr (A. Tsipis),
kgiann@ionio.gr (K. Giannakis), gkoufoud@ionio.gr (G. Koufoudakis), hristope@
ionio.gr (E. Christopoulou), okon@ionio.gr (K. Oikonomou), gkorm@aegean.gr
(G. Kormentzas), ioannis@di.uoa.gr (I. Stavrakakis).
alyzed (e.g., to exploit time and space dimensions), thus time syn-
chronization is a challenging problem for wireless sensor networks.
Issues related to synchronization begin to attract more and
more interest when it comes to Internet of Things (IoT) systems
and applications [3]. The emergence of IoT has strong bonds with
the notion of “Fog Computing”, as the enormous volume of data
that are generated and exchanged among nodes/devices requires
special treatment and appropriate handling methods [4]. Since
clocks deviate as time passes, e.g., [5,6], the timestamp given by
the node that sensed the data may not be a correct one compared
to a central reference clock, thus prohibiting further processing of
the obtained data measurements.
When particular systems that realize the IoT are considered,
they usually have to be deployed and function in extreme en-
vironments, such as underwater space, areas with wet and ultra
soft/hard soil, and so on. Such systems encounter various issues
that affect the proper operation of a node’s clock mechanism. Some
of these parameters are mechanical vibrations that occur due to

https://doi.org/10.1016/j.adhoc.2019.03.002
1570-8705/© 2019 Elsevier B.V. All rights reserved.
48 K. Skiadopoulos, A. Tsipis and K. Giannakis et al. / Ad Hoc Networks 89 (2019) 47–57
the devices’ positions, unpredictable temperature fluctuations, ex-
posure to sunlight, humidity, the unwanted presence of human ac-
tivity, animals, and planting, etc. These conditions are responsi-
ble for clock skew and offset that lead to fast de-synchronization
and unpredictable clock rates [7,8]. Clocks drift even on the order
of seconds, over quite short time periods, which is another factor
pointing out that these clocks are rather unstable. Therefore, al-
ternative approaches and revisiting of traditional synchronization
protocols for IoT environments are needed [3].
Usually, the reference clock in such systems is the clock of the
sink node, which is responsible for collecting all sensed data from
all network nodes in a multihop manner. Most research efforts in
the literature tackle this synchronization problem by synchroniz-
ing all node clocks, e.g., [9,10], the main idea being the exchange
of messages among neighbor nodes e.g., [11,12], or the usage of
GPS, e.g., [13]. Section 2 is devoted on presenting other related ap-
proaches from the literature, which either introduce a certain over-
head in the form of additional transmissions or increase the hard-
ware cost due to the GPS requirement [13].
Perfect synchronization is impossible, and every new synchro-
nization policy has to challenge this problem by tackling one or
more aspects, having in mind the particular needs. In specific IoT
environments, a relaxation of synchronization requirements has to
be considered or multiple messages have to be exchanged in a
regular manner in order to achieve (and consequently maintain) a
near-perfect synchronization [7]. For most wireless sensor network
applications that realize the IoT, a global time reference might be
redundant, and that is why relative time is preferred. In case the
network’s nodes are only sensing information measuring various
quantities, global time reference seems redundant, since only syn-
chronization among measurements is required. Due to cost and en-
ergy consumption, low-resolution clock architectures are preferred,
when the IoT applications can relax their synchronization require-
ments [14–16].
It is apparent that energy and robustness are important factors
for choosing the synchronization method when it comes to IoT sys-
tems, and especially those that are implemented outdoors for in-
dustrial purposes (e.g., [17–20]). Towards the energy-efficient and
robust functioning of actual IoT systems with these characteristics,
appropriate methods have to be considered [3,21].
In this paper, which extends the work in [22], the focus is
on synchronizing the particular measurements in a per hop basis
as they are transmitted encapsulated in data packets towards the
sink, rather than synchronizing the node clocks. Specifically, the
data measurements synchronization approach followed here updates
the timestamp of the obtained measurement in a per hop basis,
thus no extra overhead is introduced.
The system model considered here assumes two types of clock
deviation, i.e., the simple skew model and the general clock model
[23]. A data measurements synchronization algorithm is proposed
for synchronizing the corresponding measurements in a per hop
basis based on MAC [24] layer time stamping as data packets are
transmitted within the network. In particular, each data packet has
a timestamp field that corresponds to an estimation of the current
node’s clock for the particular measurement.
Furthermore, an analysis regarding the proposed algorithm is
carried out, focusing on time deviation, which is defined as the dif-
ference between the timestamp value when the data packet arrives
at the sink and the reference clock value at the time the measure-
ment was initially obtained. Analytical results about the average
time deviation and the corresponding variance show a dependency
on residual time (i.e., the time period a data packet remains within
a node), distance (i.e., the number of hops between the sink node
and the sensor node that sensed the data in the first place) and
the average skew deviation value of the considered model. In par-
ticular, it is shown that as distance increases, the average time de-
viation also increases. The same is valid for residual time, whereas
for average skew, the closer to one (i.e., optimal), the smaller the
average time deviation.
Simulation results demonstrate the effectiveness of the algo-
rithm and validate the analytical results. Various simulation sce-
narios have been conceived regarding different topologies and the
two mentioned types of deviation models (the simple skew model
and the general clock model) that were assumed in the analy-
sis. It is shown that the analytical results are in accordance with
those acquired by the simulations. Finally, the proposed algorithm
is compared against the Timing-Sync Protocol for Sensor Networks
(TPSN) [10] showing that in the particular models that are stud-
ied, it achieves better results. Extensions and improvements over
TPSN have already been proposed, but most of them focus on esti-
mating particular time parameters (like skew and/or offset), which
may not be feasible in extreme environments, like IoT systems for
agricultural monitoring.
Past related work is included in Section 2 and the system model
is described in Section 3. The proposed algorithm is presented in
Section 4 and its analysis is included in Section 5. The simulation
results along with the comparison are presented in Section 6 and
a discussion of the results conclude the study in Section 7. Various
proofs are included in the appendices.
2. Past related work
A series of clock synchronization protocols and schemes for
wireless sensor networks have been proposed and several survey
papers have made an effort to classify them according to various
features and aspects. Sundararaman et al. [5] present a classifica-
tion of clock synchronization protocols based on two kinds of fea-
tures: synchronization issues and application-dependent features.
According to synchronization issues synchronization protocols are
divided into the following: master-slave (e.g., [15]) versus peer-to-
peer (e.g., [25,26]), clock correction (e.g., [15]) versus clock unteth-
ered (e.g., [25,27]), internal versus external (e.g., [25,27,28]), prob-
abilistic versus deterministic (e.g., [29,30]), and sender-to-receiver
versus receiver-to-receiver (e.g., [9,15]). In a master-slave protocol
all nodes try to synchronize to the single master, whereas in peer-
to-peer protocols any node can communicate directly with every
other node in the network. The principle of untethered clocks al-
lows nodes to maintain their own clocks independently without
the need for continually correcting them in order to keep the net-
work synchronized, as is the case with the clock correction model
[25,27]. Internal synchronization, unlike external synchronization,
does not rely on a standard source of time. Both probabilistic and
deterministic synchronization provide a guarantee on the maxi-
mum clock offset [29,30]. In sender-to-receiver model the synchro-
nization is based on the transmissions of sender clock values as
timestamps, whereas in receiver-to-receiver model, synchroniza-
tion exploits the time at which each receiver receives the same
message. Considering application-dependent features, protocols are
divided into the following: single-hop versus multi-hop networks,
stationary versus mobile networks and MAC layer-based versus
standard approach. In single-hop networks, a sensor node can di-
rectly communicate and exchange messages with any other node
[9]. In stationary networks, sensors do not move and the topology
does not change [26]. In the MAC layer approach, the MAC protocol
is used to encapsulate synchronization messages [31].
The survey work by Rhee et al. [6] reviews representative clock
synchronization protocols used in wireless sensor networks and
additionally describes several methods for estimating clock offset
and skew. Specifically, the authors describe the maximum likeli-
hood estimate of clock offset for both RBS (Reference Broadcast-
ing Synchronization) by Elson et al. [9] and TPSN (Timing-sync
Protocol for Sensor Networks) by Ganeriwal et al. [10] protocols,
 K. Skiadopoulos, A. Tsipis and K. Giannakis et al. / Ad Hoc Networks 89 (2019) 47–57
as well as novel methods which use nonparametric bootstrap and
parametric bootstrap techniques and particle filtering techniques,
in the case of TPSN. Wu et al. [32] study clock synchronization of
wireless sensor networks from a signal processing viewpoint, illus-
trating that several clock synchronization protocols can be inter-
preted and their performance is assessed using common statistical
signal processing methods. The presented solutions and estimators
of synchronization parameters are analyzed with respect to Gaus-
sian, exponential, and arbitrary distributions for message exchange
delays.
Another survey paper by Swain and Hansdah [33] considers not
only features that reflect the structure of the networks and the
global objectives that protocols try to achieve, but features that
are associated with different phases of clock synchronization pro-
tocols in wireless sensor networks. They classify these features into
three major categories: structural features, technical features and
global objective features. Structural features characterize the en-
vironment in which the clock synchronization protocols operate
and the main identified features are: internal vs. external clock
synchronization [25,27], stationary vs. mobile nodes [25,34], single
hop [9] vs. multihop wireless sensor networks. Technical features
describe the mechanisms that have been used in various phases
of clock synchronization protocols, like the source from where the
clock value is propagated, how and when the clock value is propa-
gated, when the physical clock is updated, if the synchronization is
a-priori or post-facto [9,30] and if the protocol is based on a proba-
bilistic or a deterministic clock synchronization [29,30]. Global ob-
jective features characterize the overall objectives that need to be
achieved to meet the requirements of wireless sensor networks
and they include: fault tolerance, scalability, sleep scheduling, en-
ergy efficiency and global/local synchronization.
Following a different approach, Djenouri and Bagaa [35], as well
as Sivrikaya and Yener [13], provide an insight on issues related to
the implementation of synchronization protocols in wireless sensor
networks and analyze real implementations of the synchronization
protocols. The various challenges on implementing and testing syn-
chronization protocols include: the fast drifting of clocks used by
motes, mote limitations due to reduced cost and size for economic
reasons, the delay variation, an inherited feature of the wireless
environment, as well as the faulty nature of motes and lossy chan-
nels, which makes fault tolerance a necessity for a real deployment
and large-scale experimentation.
This is quite common in IoT environments, where IoT nodes
usually synchronize according to an external reference source [3].
On the other hand, NTP-like protocols (where NTP stands for Net-
work Time Protocol) are not fitted for these environments because
of the limited resources and capabilities of these nodes [3]. An-
other issue regarding oscillators that are used for clocks in these
devices is the fact that they are susceptible to changes in supply
voltage, which is common in devices that are used in IoT applica-
tions.
Many clock synchronization protocols are related to the work
presented here, like the Delay Measurement Time Synchroniza-
tion for wireless sensor networks (DMTS) by Ping [15], the Flood-
ing Time Synchronization Protocol (FTSP) by Maróti et al. [30],
the Timing-Sync Protocol for Sensor Networks (TPSN) by Ganeri-
wal et al. [10], the probabilistic clock synchronization service in
sensor networks by PalChaudhuri et al. [26], the Time Diffusion
Synchronization Protocol (TDP) by Su and Akyildiz [25], and the
Asynchronous Diffusion protocol by Li and Rus [12]. More recent
protocols include the Gradient Time Synchronization Protocol
(GTSP) by Sommer and Wattenhofer [36], the Average TimeSync
(ATS) algorithm by Schenato and Fiorentin [34], the PulseSynch by
Lenzen et al. [37], the Round-Robin Timing Exchange (RRTE) pro-
tocol by Huang and Wu [11], the R4 Syn by Djenouri [38], and the
2LTSP by Huang et al. in [39].
49
Similarly to Wu et al. [32], another work regarding clock
synchronization using a reference node in distributed wireless
systems is undertaken by Leng and Wu [40]. In this work, commu-
nication among neighboring nodes suffices to achieve clock syn-
chronization. Simulation results showed that the proposed method
is robust and scalable when compared with other approaches. Each
node holds an estimation of the time which is constantly updated.
Another interesting work is proposed by Du and Wu in [28],
where an effort to estimate the clock skew and offset is made for
distributed networks, dealing with a wide range of network topolo-
gies. Similarly to other works, the achieved result is based on local
information among neighboring nodes. Clock skew and offset, two
factors that are ubiquitous in systems involving oscillators, are also
estimated [41] by Luo and Wu, where a distributedly applied filter
for tracking these parameters is used. The proposed algorithm aims
to calculate the oscillator noise that is to blame for clock skew and
offset. It is also shown, both analytically and experimentally, that
it can cope with scalability and node failing issues.
On the other hand, Luo et al. [8] address the problem of un-
known and unpredictable delays in wireless sensor networks that
affect the synchronization process among the nodes, proposing a
suitable algorithm with low overhead. Similarly to the above ap-
proaches, Noh et al. in [42] aim to efficiently estimate clock skew
and offset in wireless sensor networks. The proposed mechanism
does not require often re-synchronization routines, thus overall
overhead is decreased. An advantage of this method is that it can
be combined with conventional synchronization protocols.
Contrary to other approaches, like FTSP and TPSN, the proposed
algorithm, originally published in [22], can cope with topology
changes and can even be easily transformed to support high mo-
bility, since the timestamping process depends only on routing in-
formation and clock synchronization (that could involve message
exchanges) is not required. On the contrary, TPSN would have to
start over the level discovery phase in cases where changes occur
to the topology or the location of the root node, which yields ex-
tra traffic and complexity. Note that FTSP is able to better address
topology changes since it periodically alters the root node.
3. System definition
After sensing the particular data of interest, each node in the
network creates a data packet, encapsulates its local clock and for-
wards it over a pre-constructed path heading to the sink node. It
is assumed that the first node on this path is node 0, whereas the
kth node is the sink. The distance between node 0 and the sink
node, in terms of number of hops, is k, as it can be seen in Fig. 1.
It is assumed that all packet delays (denoted by τ ) are equal for all
nodes in the network and propagation delay is negligible.
Initially, let node 0 generate a packet at time t0 that will be
forwarded to the 1st node towards the sink at time t1 , to the 2nd
node at time t2 , to the (k − 1 )th at time tk−1 and finally to the kth
(i.e., the sink) at time tk , as depicted in Fig. 1.
p
Let ti denote the time required for processing a packet at node
a
i and ti the time that node i is stalled until the wireless channel
becomes free before transmitting towards node i + 1. Let tγ denote
the time required to transmit the preamble bits and the Start Of
Frame (SOF) field (depicted in Fig. 2). Therefore, τ = ti+1 − ti = ti +
p
tia + t γ and ti = t0 + τ i. It is clear that these delays are hard to be
predicted or calculated, and thus, they are the main factors that
make the synchronization problem of great importance.
Let ci (t) denote the clock of node i at time t and let the sink
node’s clock be the reference clock (i.e., ck (t ) = t). As previously
mentioned, there are two basic models that capture the behavior
(or deviation) of a node’s clock, i.e., the simple skew model and
the general clock model [23]. The general clock model is given by
50 K. Skiadopoulos, A. Tsipis and K. Giannakis et al. / Ad Hoc Networks 89 (2019) 47–57
Fig. 1. Data packet path of k hops from node 0 towards the sink node (kth).
Fig. 2. Data packet structure.
ci (t ) = βi (t )t + θi ,
while the simple skew model is given by
ci (t ) = βi t + θi ,
where θ i is the node’s i clock offset at time t = 0 and β i (t) or β i –
described next – will be referred to hereafter as the clock skew.
Regarding the derivation of the general clock model, node’s
skew is captured by θi (t ) = θi + ai t + ωi (t ), where ωi (t) encap-
sulates the deviations from the simple skew model and obeying
ωi (t ) = 0. Thus,
ci (t ) = θi (t ) + t
= θi + ai t + ωi (t ) + t
= (ai + 1 )t + ωi (t ) + θi .
 fields ci (ti+1 ), TSi (timestamp) and Data, as depicted in Fig. 2.
Let
βi (t ) = ai + 1 + ωit(t ) for t > 0, Then, the above equation
βi (t ) = ai + 1 for t = 0.
can be written as
ci (t ) = βi (t )t + θi .
As it is shown in Eqs. (3) and (2), under the general clock
model, clock skew (i.e., β i (t)) changes in time for node i, while un-
der the simple skew model it remains constant (i.e., β i ) for the
particular node i (but different for each node). The general skew
model is better suited to imitate the stability factor of the clock,
e.g., how much does clock’s frequency drift over time [15]. It has
been shown [3] that even the same kind of IoT devices show vary-
ing clock drifts that are easily affected by factors like ambient
temperature, etc. that emphasize clock skew and offset as time-
dependable parameters. Thus, it is often quite difficult to estimate
skew and offset values.
In general, skew and offset are assumed to be different for each
node and approximately constant for an extended period of time
(from minutes to hours) [31], but there are cases where this is
questionable (e.g., in environments with extreme conditions, such
as underwater etc). For both models, assuming that offset θ i is
known and that the clock skew is equal to one, this corresponds
to the (unrealistic) ideal case where all clocks are synchronized.
Nevertheless, clock skew (either β i (t) or β i , depending on the
model) is not assumed to be ideal and therefore, the aim in the
sequel is to propose and analyze an algorithm for synchronizing
the data measurements rather than the clocks of the nodes. To
(1) keep the notation as simple as possible, the general clock model is
considered in the sequel since it reduces to the simple one when
clock skew does not change in time. The proposed algorithm in the
(2) sequel as well as the subsequent analysis can be applied to both
models.
4. The proposed algorithm
In this part of the work, the proposed algorithm is described
in detail. In particular, Algorithm 1 presents the details of the per
hop basis update of the timestamp field. After the transmission of
the SOF field both nodes (i.e., node i and node i + 1) read their lo-
cal clock values i.e., ci (ti+1 ) and ci+1 (ti+1 ), respectively (lines 7–9).
Node i continues the transmission by giving values to the packet
Upon successful transmission, node i + 1 estimates the dif-
ference between its own clock and that of node i as Ci+1 =
ci+1 (ti+1 ) − ci (ti+1 ) (line 14) and replaces the TSi field by T Si+1 =
T Si + Ci+1 or, T Si+1 = T Si + ci+1 (ti+1 ) − ci (ti+1 ) (line 15). This pro-
(3)
cedure continues until the kth node receives the packet. In a sim-
ilar manner, the kth node calculates the clock difference as Ck =
ck (tk ) − ck−1 (tk ) and replaces the T Sk−1 field by T Sk = T Sk−1 +
ck (tk ) − ck−1 (tk ).
Algorithm 1 is capable of calculating the clock difference be-
tween two nodes assuming no or negligible propagation delay.
Each node increases the TS field by the clock difference between its
clock and the sender,after the reception of a packet. Thus, when a
packet has been received by the kth node, the value of the TS field
is equal to T Sk = T Sk−1 + Ck . As it is proved in Appendix A,
k−1

T Sk = tk − [ci (ti+1 ) − ci (ti )]. (4)
i=0
The timestamp when the packet has arrived at the sink node is
derived by Eq. (4) and it corresponds to the sink node’s estimation
of the time of the data generation at node 0 (i.e., t0 ). It is observed
 K. Skiadopoulos, A. Tsipis and K. Giannakis et al. / Ad Hoc Networks 89 (2019) 47–57 51

Algorithm 1 The Proposed Data Synchronization Algorithm. Considering Eq. (5) and the linearity property of the expected
value [43], it follows that
Var k: Sink node; i: Current node; Data: Local generated data;
T S: Timestamp field; Clock: Local clock; P reamble: Preamble   k−1

bits; E k,0 = τ k − τ [(i + 1 )E[βi ((i + 1 )τ )] − iE[βi (iτ )]].
1: Operate: i=0
2: if Data then Given the previous assumption with respect to the distribution
3: T S := Clock; of β i (t), it follows that
4: if i = k then SendPacket(T S, Data);  If not the sink,
forward the data packet   k−1 
 
E k,0 = τ k − τ (i + 1 )μβ − iμβ
5: if receive < P reamble > then ReceivePacket;
i=0
6: function SendPacket(T S, data) k−1

Var c; SOF : SOF field; = τk − τ μβ
7: send < P reamble, i + 1 >; send < SOF , i + 1 >; i=0
8: c := Clock;
= τ k − τ kμβ
9: send < c, i + 1 >; send < T S, i + 1 >; send < data, i + 1 >;
10: function ReceivePacket which can be written as,
Var c; cR : Clock field; data: Data field; C: Clock difference;    
E k,0 = τ k 1 − μβ . (6)
SOF : SOF field; T S: Timestamp;
11: receive < SOF >; Considering Eq. (5), the variance of time deviation is
12: c := Clock;
receive < cR >; receive < T S >; receive < data >;   k−1

13:
σ 2 k,0 = σ 2 τ k − τ [(i + 1 )βi ((i + 1 )τ ) − iβi (iτ )] .
14: C := c − cR ; i=0
15: T S := T S + C;
16: if i = k then SendPacket(T S, data);  If not the sink, As it is proved in Appendix C,
forward the data packet
 
σ 2 k,0 = kτ 2 σβ2 . (7)
Both Eqs. (6) and (7) help evaluate the performance of the pro-
posed Algorithm 1 with respect to the time deviation. As it is con-
that, in order to achieve packet synchronization, the proposed al- cluded from Eq. (6), the expected value of time deviation E k,0
gorithm requires the instant sending of measurements from every increases linearly with packet delay τ and the number of hops k
node (assuming of course the necessary handling time) towards a between the particular node and the sink node. The mean value of
neighboring node (and then consecutively up to the sink), which the skew function β i (t) also plays an important role. When μβ → 1,
is responsible for the estimation of the measurement’s real gener-
 
this allows for E k,0 → 0. However, μβ → 1 means that β i (t) → 1
ation time. and eventually, the offset by the hardware of each node should be
In summary, the proposed algorithm rather than synchroniz- negligible (i.e., close to zero). Regarding the variance, as given by
ing clocks, it aims to synchronize measurements in a per hop ba- Eq. (7), it is affected by the number of hops k, the packet delay
sis, as they are transmitted encapsulated in data packets. Each τ and variance σ β . As before, when β i (t) → 1, it is expected that
data packet contains a timestamp field that holds an estimation
σ β → 0 and eventually, σ 2 [ k,0 ] → 0.
of the current node’s clock for a particular measurement. Then,
The above analysis focuses on the estimation of the difference
this field is updated in each hop according to the node’s clock,
between the timestamp value when the data packet arrives at the
which does not yield extra overhead (no additional message or re-
sink node and the clock value at the time the measurement was
synchronization periods are required). These packets are sent to-
obtained in an arbitrary node, namely the time deviation of the
wards the sink node, where an estimation of the time of the data
measurement. In particular, the average time deviation is analyt-
sensed by each node is calculated.
ically studied, revealing a dependency on the residual time, the
distance and the average skew deviation value of the considered
model. As distance and residual time increase, the average time
deviation also increases, while the opposite holds for the average
5. Analysis skew (the closer to optimal, the smaller the average time devia-
tion).
The analytical part of this work concentrates on time devi-
ation at the sink node (the kth). Let  k denote the time de- 6. Simulation results
viation of the timestamp at the sink node. Then, k = T Sk − t0
and by replacing TSk from Eq. (4) it follows that, k,0 = tk − t0 − A simulation program in C++ has been developed in order to
k−1
i=0 [ci (ti+1 ) − ci (ti )]. As it is proved in Appendix B, generate geometric random graph topologies with 10 0 0 uniformly
and independently distributed nodes in [0, 1] × [0, 1] plane and a
k−1
 link between any pair exists if their euclidean distance is equal to
k,0 = τ k − τ [(i + 1 )βi ((i + 1 )τ ) − iβi (iτ )]. (5) or smaller than the connectivity radius rc . The considered topolo-
i=0 gies correspond to rc = 0.06, 0.1, 0.2 and 0.3, and are considered
as a simple model suitable for representing wireless sensor net-
Afterwards, the next step is to examine the accuracy of work topologies. Wireless transmissions occur over an error-free
Algorithm 1 by studying the expected value and the variance of channel with a bit rate of 100 kbps and an underlying nonper-
time deviation, i.e., E k,0 and σ 2 [ k,0 ], respectively. Let the dis- sistent CSMA protocol [24]. Note that under CSMA, collisions that
tribution of the skew function β i remain the same across all nodes might put extra delay are present. Each simulation scenario lasts
for the same time step t, and for all time steps t for each node. Let 10 0 0 seconds and the presented simulation results correspond to
μβ and σβ2 denote the corresponding mean and variance. the average of ten independent runs for both skew models. Time
52 K. Skiadopoulos, A. Tsipis and K. Giannakis et al. / Ad Hoc Networks 89 (2019) 47–57

Fig. 3. Simulation results regarding the average time deviation as a function of the number of hops for rc = 0.06, 0.1, 0.2, and 0.3 for the simple skew model. The dotted
lines correspond to the analytical results.

deviation is measured at the sink node that is responsible for the
gathering of measurements.
The description of a simulation scenario is as follows: every
node of the multi-hop network senses some data (which com-
prise the measurements). It is assumed that routing information
is known (or easily obtained through known algorithms) and there
are not frequent topology changes (this does not affect the pro-
posed algorithm’s performance since there is no need for ex-
tra synchronization messages). Then, every node sends its data
(adding the timestamp of the particular measurement), and after
that the next node on the route towards the sink forwards the data
(changing the value of the timestamp accordingly) until they reach
the sink node that is responsible for handling these data. Delay
times are in accordance with the system definition of Section 3.
Immediately after, time deviation is measured in order to test the
algorithm’s performance. Note that this description holds for every
experiment (only the given parameters are changed).
For both models, skew values (β i (t) and β i , respectively) are
uniformly distributed within the range of [0.995, . . . , 1.0 0 0]. For
the particular case that the skew function β i (t) is uniformly dis-
tributed with minimum value β min and maximum value β max ,
βmin +βmax
the mean value and the variance of β i (t) are μβ =
(βmax −βmin )2
σβ2 = 12 , respectively. Thus, Eq. (6) can be written as,
  β + βmax
EU k,0 = τ k 1 − min . (8)
2
6.1. Validation of analytical results
Fig. 3 depicts average values of time deviation (in ms) under
the proposed Algorithm 1 as a function of the number of hops for
the simple skew model for four different topologies. It is obvious
that as the number of hops k increases, the average time deviation
increases linearly. It is interesting to see that the analytical results,
corresponding to Eq. (8) and depicted as a dotted line, capture the
behavior of the system. For each depicted scenario (a, b, c and
d), three different cases with respect to packet delay τ have been
considered, i.e., for (a) τ = 8.32 ms, 12.96 ms and 15.81 ms; for
(b) τ = 8.33 ms, 12.32 ms and 16.11 ms; and for (c) τ = 8.32 ms,
13.98 ms and 17.99 ms. As observed, as τ increases, the average
time deviation of the proposed algorithm increases. Still, the ana-
lytical results capture the system’s behavior.
Fig. 4 shows average values of time deviation (in ms) under the
proposed Algorithm 1 as a function of the number of hops for the
general clock model for four different topologies. As already shown
for the case of the simple skew model, the average time deviation
increases linearly with the number of hops k. For each depicted
scenario (a, b, c and d), three different cases with respect to packet
delay τ have been considered, i.e., for (a) τ = 9.04 ms, 12.61 ms
2 and
and 16.24 ms; for (b) τ = 8.72 ms, 12.29 ms and 16.11 ms; and
for (c) τ = 9.12 ms, 13.09 ms and 16.85 ms. As observed, as τ in-
creases, the average time deviation of the proposed algorithm in-
creases. Still, the analytical results capture the system’s behavior.

6.2. Comparison against TPSN

These values are selected in order to imitate an unstable envi-
ronment with frequent drift changes (for the general skew model),
assuming also that any ambient noise is embedded in this param-
eter. In other words, stability [15] is assumed to be quite low in
order to simulate actual paradigms of IoT environments that reside
in unfriendly areas. That is why both skew models are used.
Besides the demonstration of the proposed algorithm in the
previous part, a comparison with another well-established ap-
proach is described in this section. In particular, the Timing-Sync
Protocol for Sensor Networks (TPSN) protocol by Ganeriwal et al.
[10] is considered for the evaluation of the proposed algorithm.
 K. Skiadopoulos, A. Tsipis and K. Giannakis et al. / Ad Hoc Networks 89 (2019) 47–57
Fig. 4. Simulation results regarding the average time deviation as a function of the number of hops for rc = 0.06, 0.1, 0.2, and 0.3 for the general clock model. The dotted
lines correspond to the analytical results.
The choice of TPSN is dictated by some special attributes re-
garding functionality and its wide adoption. Note that, as described
previously in Section 2, several variations of TPSN have already
been proposed that yield improvements for particular instances of
networks, when, for example, clock skew and offset can be esti-
mated [20,42]. All of them focus on synchronizing clocks through
re-adjustments after proper estimation of clock skew and/or off-
set. Nevertheless, since the proposed algorithm assumes general
instances, it should be compared against TPSN (described in [10]).
It is important, though, to recall that TPSN aims at synchronizing
nodes’ clocks, while the proposed algorithm operates in a per-hop
basis and ensures packet and not clock synchronization.
TPSN is a two-step algorithm. At the beginning, a hierarchi-
cal structure is established in the network and then a pair-wise
synchronization is performed along the edges of this structure to
establish a global timescale throughout the network. This obvi-
ously yields exchange of messages and hence extra overhead is in-
troduced. Eventually, all nodes in the network synchronize their
clocks to a predefined reference node. Unlike the proposed algo-
rithm, the limitation of instantaneous sending of measurements
from every node to the sink is leveraged, since TPSN permits each
node to maintain measurements inside themselves and send them
at will.
It is observed that, in order to achieve packet measurements’
synchronization, the proposed algorithm requires instantaneous
sending of measurements from every node (assuming of course the
necessary handling time) towards the sink, which is responsible for
the estimation of the measurements’ actual time of generation.
For the simulation purposes four distinct runs are examined for
three different time periods, i.e., synchronization was occured ev-
ery 50, 100, and 500 ms, for each run. This synchronization cycle
is fixed for each case and takes place by following the root’s clock.
The rest of the simulation parameters use the same setting as be-
fore.
The first observation from Figs. 5 and 6 is that the average time
deviation increases linearly with the number of hops k (which is
53
also true for the case of the proposed algorithm). This is not sur-
prising given the intrinsic character of the synchronization prob-
lem. Moreover, time deviation is also affected by the synchroniza-
tion period, the more frequent it is, the less the deviation.
Fig. 5 shows average values of time deviation for the TPSN as
a function of the number of hops for the simple skew model and
for four different values of the connectivity radius. Again, for each
scenario, three different cases with respect to synchronization pe-
riod have been considered. It is easily seen that as the value of ts
increases, the average time deviation of the proposed algorithm in-
creases, as well. When compared with the respective results of the
proposed algorithm (as seen in Fig. 3), it is observed that TPSN’s
time deviation is increased, since the maximum time deviation is
10 times higher than that of the proposed algorithm.
In Fig. 6 the results for the general skew model are depicted.
Again, average values of time deviation under the TPSN algorithm
as a function of the number of hops for four different topologies
are depicted. Similarly to the simple skew model, the average time
deviation increases linearly with the number of hops k, whereas,
as ts increases, the average time deviation of the TPSN algorithm
increases. Similar to the case of simple skew model (as seen in
Fig. 4), it is observed that TPSN’s time deviation is increased, since
the maximum time deviation is 10 times higher that that of the
proposed algorithm.
6.3. Synchronizing clocks vs. synchronizing measurements
The above results support the claim that in environments with
rapidly varying values for clock drifting, the proposed algorithm
outperforms traditional synchronization algorithms that focus on
synchronizing nodes’ clocks, rather than measurements. This is at-
tributed to the fact that since each timestep the clock drift is
changing, any effort to establish synchronization is of limited ef-
fectiveness, not to mention the extra overhead induced by the
exchange of messages needed for this establishment. Clocks can
drift even on the order of seconds, which might take place over
54 K. Skiadopoulos, A. Tsipis and K. Giannakis et al. / Ad Hoc Networks 89 (2019) 47–57

Fig. 5. Simulation results of TSPN regarding the average time deviation as a function of the number of hops for rc = 0.06, 0.1, 0.2, and 0.3 for the simple skew model and
various synchronization intervals.

Fig. 6. Simulation results of TSPN regarding the average time deviation as a function of the number of hops for rc = 0.06, 0.1, 0.2, and 0.3 for the general clock model and
various synchronization intervals.

relatively short time periods, emphasizing the fact that these
clocks are rather unstable. It is true that in most cases. skew and
offset are different for each node and are considered constant for
an extended period of time [31], but extreme environmental condi-
tions can severely shorten this period, compromising the process of
clock synchronization. This underlines the need for alternative ap-
proaches and revisiting of conventional synchronization algorithms
for IoT environments.
On the other hand, the advantage of using TPSN over the
proposed algorithm is that nodes’ clocks remain (not perfectly)
synchronized, whereas in the proposed approach clocks are al-
lowed to run on their own pace. However, having in mind that
 K. Skiadopoulos, A. Tsipis and K. Giannakis et al. / Ad Hoc Networks 89 (2019) 47–57 55

unstable clocks operating under extreme circumstances (where Appendix A. Derivation of Eq. (4)
skew and offset are different for each node and not constant for
an extended period of time [31]), with low energy capabilities, it The update is:
is concluded that the proposed approach is the better one and
it is more energy-efficient, since no synchronization messages are
T Sk = T Sk−1 + Ck
sent. Also, unlike the proposed algorithm, nodes under TPSN can = T Sk−1 + [ck (tk ) − ck−1 (tk )]
hold the sensed measurements and send them at will, whereas the = T Sk−2 + [ck−1 (tk−1 ) − ck−2 (tk−1 )] + [ck (tk ) − ck−1 (tk )]
proposed algorithm strictly dictates each node to send them right
= T S0 + [c1 (t1 ) − c0 (t1 )] + [c2 (t2 ) − c1 (t2 )] . . .
away.
+ [ck−1 (tk−1 ) − ck−2 (tk−1 )] + [ck (tk ) − ck−1 (tk )]
= T S0 − c0 (t1 ) − [c1 (t2 ) − c1 (t1 )] − [c2 (t3 ) − c2 (t2 )] . . .
7. Summary and conclusions
− [ck−1 (ks ) − ck−1 (tk−1 )] + ck (tk )
The problem of time synchronization is constantly under ex- k−1

amination, since the emergence of wireless sensor networks and = ck (tk ) − [ci (ti+1 ) − ci (ti )] − [c0 (t1 ) − T S0 ].
the broad spread of Internet of Things (IoT) systems. Such envi- i=1
ronments are usually distributed, multi-hop, and at the same time,
Note that TS0 is node’s 0 timestamp of the generated
the involved devices have particular needs and capabilities. For ex-
data at time t = t0 (i.e., T S0 = c0 (t0 )) and also kth node’s
ample, ambient noise and node malfunctions are present, whereas
clock is the reference clock (i.e., ck (t ) = t since the kth node
these systems are also characterized by non-deterministic delays
is the sink). Thus, the previous expression can be written
(e.g., due to jitter [44]) that are not easy to predict, with noise lev-  −1
as T Sk = tk − ki=1 [ci (ti+1 ) − ci (ti )] − [c0 (t1 ) − c0 (t0 )] or, T Sk = tk −
els being usually higher than those of other conventional systems k−1
[5]. i=0 [c i (t i +1 ) − c i (t i )].

When IoT systems are considered, they usually have to be

deployed and operate in environments with extreme conditions, Appendix B. Derivation of Eq. (5)
such as areas with wet and ultra soft/hard soil, etc. Such systems
Given, τ = ti+1 − ti = ti + tia + t γ and ti = t0 + iτ , the time de-
p
come up against plenty of issues that affect the proper operation
 −1 
of a node’s clock mechanism leading to fast de-synchronization viation of TSk can be written as k,0 = τ k − ki=0 ci (t0 + (i + 1 )τ )
and unpredictable clock frequencies. This emphasizes the fact that 
− ci (t0 + iτ ) and by replacing ci (t) from Eq. (3) it follows that
such clocks are rather unstable, and consequently, alternative ap-
proaches and revisiting of traditional solutions are needed in order k−1

to deal with these issues, like the one proposed in the current pa- k,0 = τ k − [(t0 + (i + 1 )τ )βi (t0 + (i + 1 )τ ) + θi,0
per. i=0
The problem of synchronization in wireless sensor networks −(t0 + iτ )βi (t0 + iτ ) − θi,0 ]
was addressed here under a different perspective. Specifically, the
k−1

main idea here was to synchronize data measurements as they
= τk − [(t0 + (i + 1 )τ )βi (t0 + (i + 1 )τ )
are transmitted towards the sink node in a per hop basis instead
i=0
of synchronizing the node clocks, thus inducing no extra over-
head. This is due to the no need for additional messages and re- −(t0 + iτ )βi (t0 + iτ )].
synchronization periods, contrary to other approaches that syn- Without loss of generality it is assumed that t0 = 0 thus,
chronize and adjust over and over the nodes’ clocks. Thereby, a
k−1

lightweight synchronization algorithm was proposed and subse-
quently analyzed, whereas indicative simulation results demon- k,0 = τ k − τ [(i + 1 )βi ((i + 1 )τ ) − iβi (iτ )].
i=0
strated the algorithm’s effectiveness in the considered environ-
ment. Furthermore, it was shown that the analytical results are in
accordance with the simulations. Appendix C. Derivation of Eq. (7)
Contrary to other approaches, the proposed algorithm can cope
with topology changes and can even be easily transformed to sup- Considering Eq. (5), the variance of the time deviation is
port high mobility, since the timestamping process depends only
on routing information and no extra effort for clock synchroniza-   k−1

σ 2 k,0 = σ 2 τ k − τ [(i + 1 )βi ((i + 1 )τ ) − iβi (iτ )]
tion (that could require message exchanges) is required. For future
i=0
work, the implementation of the proposed algorithm in a large
k−1

scale setting would be useful in order to test the method in ac-
tual extreme conditions. = τ 2σ 2 [(i + 1 )βi ((i + 1 )τ ) − iβi (iτ )] .
i=0

Since β i (t) and β j = i (t) are uncorrelated it follows that their co-
Acknowledgments  
variance equals to zero (i.e., ρ βi (t ), β j=i (t ) = 0), where ρ ( · ) is
 
This work was supported in part by project “A Pilot Wireless the covariance function. Also ρ βi (tm ), βi (tn=m ) = σβ2 thus,
Sensor Networks System for Synchronized Monitoring of Climate
and Soil Parameters in Olive Groves,” (MIS 5007309) which is par-   k−1

tially funded by European and National Greek Funds (ESPA) under σ 2 k,0 = τ 2 σ 2 [(i + 1 )βi ((i + 1 )τ ) − iβi (iτ )]
the Regional Operational Programme “Ionian Islands 2014-2020.” i=0

In addition, this work was supported in part by the European Com- k−1 

mission as part of the SAURON project (Horizon H2020 Frame- = τ2 (i + 1 )2 σ 2 [βi ((i + 1 )τ )] + i2 σ 2 [βi (iτ )]
work Programme of the European Union) and by the National and i=0

Kapodistrian University of Athens (S.A.R.G.). −2(i + 1 )iρ (βi ((i + 1 )τ ), βi (iτ ) )]

56 K. Skiadopoulos, A. Tsipis and K. Giannakis et al. / Ad Hoc Networks 89 (2019) 47–57
k−1 
 
= τ2 (i + 1 )2 σβ2 + i2 σβ2 − 2(i + 1 )iσβ2
i=0
k−1 
 
= τ2 (i + 1 − i )2 σβ2
i=0
= kτ 2 σβ2 .
References
[1] I. Akyildiz, W. Su, Y. Sankarasubramaniam, E. Cayirci, Wireless sensor net-
works: a survey, Comput. Netw. 38 (4) (2002) 393–422.
[2] I. Akyildiz, T. Melodia, K. Chowdhury, A survey on wireless multimedia sensor
networks, Comput. Netw. 51 (4) (2007) 921–960.
[3] S.K. Mani, R. Durairajan, P. Barford, J. Sommers, An architecture for IoT clock
synchronization, in: Proceedings of the 8th International Conference on the
Internet of Things (IOT’18), ACM, 2018 Article 17, 8 pages, doi:10.1145/3277593.
3277606.
[4] F. Bonomi, R. Milito, J. Zhu, S. Addepalli, Fog computing and its role in the
internet of things, in: Proceedings of the First Edition of the MCC Workshop
on Mobile Cloud Computing, ACM, 2012, pp. 13–16.
[5] B. Sundararaman, U. Buy, A.D. Kshemkalyani, Clock synchronization for wire-
less sensor networks: a survey, Ad Hoc Netw. 3 (3) (2005) 281–323, doi:10.
1016/j.adhoc.2005.01.002.
[6] I.-K. Rhee, J. Lee, J. Kim, E. Serpedin, Y.-C. Wu, Clock synchronization in wire-
less sensor networks: an overview, Sensors 9 (1) (2009) 56–85, doi:10.3390/
s9010 0 056.
[7] A. Elsts, X. Fafoutis, S. Duquennoy, G. Oikonomou, R. Piechocki, I. Craddock,
Temperature-resilient time synchronization for the internet of things, IEEE
Trans. Ind. Inf. 14 (5) (2018) 2241–2250.
[8] B. Luo, L. Cheng, Y.-C. Wu, Fully distributed clock synchronization in wireless
sensor networks under exponential delays, Signal Process. 125 (2016) 261–273,
doi:10.1016/j.sigpro.2016.02.007.
[9] J. Elson, L. Girod, D. Estrin, Fine-grained network time synchronization using
reference broadcasts, ACM SIGOPS Operat. Syst. Rev. 36 (SI) (2002) 147–163.
[10] S. Ganeriwal, R. Kumar, M.B. Srivastava, Timing-sync protocol for sensor net-
works, in: Proceedings of the 1st International Conference on Embedded Net-
worked Sensor Systems, ACM, 2003, pp. 138–149.
[11] Y.-H. Huang, S.-H. Wu, Time synchronization protocol for small-scale wire-
less sensor networks, in: Wireless Communications and Networking Confer-
ence (WCNC), 2010 IEEE, IEEE, 2010, pp. 1–5.
[12] Q. Li, D. Rus, Global clock synchronization in sensor networks, IEEE Trans.
Comput. 55 (2) (2006) 214–226.
[13] F. Sivrikaya, B. Yener, Time synchronization in sensor networks: a survey, IEEE
Netw. 18 (4) (2004) 45–50, doi:10.1109/MNET.2004.1316761.
[14] F. Terraneo, A.V. Papadopoulos, A. Leva, M. Prandini, FLOPSYNC-QACS: quanti-
zation-aware clock synchronization for wireless sensor networks, ACM SIGBED
Rev. 14 (4) (2018) 33–38.
[15] S. Ping, Delay measurement time synchronization for wireless sensor net-
works, Intel Res. Berkeley Lab 6 (2003) 1–12.
[16] J. Polastre, R. Szewczyk, D. Culler, Telos: enabling ultra-low power wireless
research, in: Information Processing in Sensor Networks, 20 05. IPSN 20 05.
Fourth International Symposium on, IEEE, 2005, pp. 364–369.
[17] N. Aakvaag, M. Mathiesen, G. Thonet, Timing and power issues in wireless sen-
sor networks-an industrial test case, in: Parallel Processing, 2005. ICPP 2005
Workshops. International Conference Workshops on, IEEE, 2005, pp. 419–426.
[18] J. Höller, D. Boyle, S. Karnouskos, S. Avesand, C. Mulligan, V. Tsiatsis, From Ma-
chine-to-Machine to the Internet of Things, Elsevier, 2014.
[19] G. Han, C. Zhang, L. Shu, J.J. Rodrigues, Impacts of deployment strategies on
localization performance in underwater acoustic sensor networks, IEEE Trans.
Ind. Electron. 62 (3) (2015) 1725–1733.
[20] H. Wang, L. Shao, M. Li, B. Wang, P. Wang, Estimation of clock skew for time
synchronization based on two-way message exchange mechanism in indus-
trial wireless sensor networks, IEEE Trans. Ind. Inf. 14 (11) (2018) 4755–4765,
doi:10.1109/TII.2018.2799595.
[21] T. Qiu, Y. Zhang, D. Qiao, X. Zhang, M.L. Wymore, A.K. Sangaiah, A robust time
synchronization scheme for industrial internet of things, IEEE Trans. Ind. Inf.
14 (8) (2018) 3570–3580.
[22] S. Fanarioti, A. Tsipis, K. Giannakis, G. Koufoudakis, E. Christopoulou,
K. Oikonomou, I. Stavrakakis, A proposed algorithm for data measurements
synchronization in wireless sensor networks, in: Second International Balkan
Conference on Communications and Networking 2018 (BalkanCom’18), Podgor-
ica, Montenegro, 2018, pp. 1–5.
[23] D. Veitch, S. Babu, A. Pàsztor, Robust synchronization of software clocks across
the internet, in: Proceedings of the 4th ACM SIGCOMM conference on Internet
measurement, ACM, 2004, pp. 219–232.
[24] L. Kleinrock, F. Tobagi, Packet switching in radio channels: Part I–carrier sense
multiple-access modes and their throughput-delay characteristics, IEEE Trans.
Commun. 23 (12) (1975) 1400–1416.
[25] W. Su, I.F. Akyildiz, Time-diffusion synchronization protocol for wireless sen-
sor networks, IEEE/ACM Trans. Netw. 13 (2) (2005) 384–397, doi:10.1109/TNET.
2004.842228.
[26] S. PalChaudhuri, A. Saha, D.B. Johnson, Probabilistic clock synchronization ser-
vice in sensor networks, IEEE Trans. Netw. 2 (2) (2003) 177–189.
[27] R. Fan, I. Chakraborty, N. Lynch, Clock synchronization for wireless networks,
in: T. Higashino (Ed.), Principles of Distributed Systems, Springer Berlin Hei-
delberg, Berlin, Heidelberg, 2005, pp. 400–414.
[28] J. Du, Y. Wu, Distributed clock skew and offset estimation in wireless sen-
sor networks: asynchronous algorithm and convergence analysis, IEEE Trans.
Wirel. Commun. 12 (11) (2013) 5908–5917, doi:10.1109/TWC.2013.100213.
130553.
[29] S. PalChaudhuri, A.K. Saha, D.B. Johnson, Adaptive clock synchronization in
sensor networks, in: Proceedings of the 3rd International Symposium on In-
formation Processing in Sensor Networks, ACM, 2004, pp. 340–348.
[30] M. Maróti, B. Kusy, G. Simon, Á. Lédeczi, The flooding time synchronization
protocol, in: Proceedings of the 2nd International Conference on Embedded
Networked Sensor Systems, ACM, 2004, pp. 39–49.
[31] M.L. Sichitiu, C. Veerarittiphan, Simple, accurate time synchronization for wire-
less sensor networks, in: Wireless Communications and Networking, 2003.
WCNC 2003. 2003 IEEE, vol. 2, IEEE, 2003, pp. 1266–1273.
[32] Y.-C. Wu, Q. Chaudhari, E. Serpedin, Clock synchronization of wireless sensor
networks, IEEE Signal Process. Mag. 28 (1) (2011) 124–138.
[33] A.R. Swain, R. Hansdah, A model for the classification and survey of clock syn-
chronization protocols in WSNs, Ad Hoc Netw. 27 (2015) 219–241, doi:10.1016/
j.adhoc.2014.11.021.
[34] L. Schenato, F. Fiorentin, Average timesynch: a consensus-based protocol for
clock synchronization in wireless sensor networks, Automatica 47 (9) (2011)
1878–1886.
[35] D. Djenouri, M. Bagaa, Synchronization protocols and implementation issues in
wireless sensor networks: a review, IEEE Syst. J. 10 (2016) 617–627.
[36] P. Sommer, R. Wattenhofer, Gradient clock synchronization in wireless sensor
networks, 2009 International Conference on Information Processing in Sensor
Networks (2009) 37–48.
[37] C. Lenzen, P. Sommer, R. Wattenhofer, Optimal clock synchronization in net-
works, in: Proceedings of the 7th ACM Conference on Embedded Networked
Sensor Systems, ACM, 2009, pp. 225–238.
[38] D. Djenouri, R4 Syn: relative referenceless receiver/receiver time synchroniza-
tion in wireless sensor networks, IEEE Signal Process. Lett. 19 (2012) 175–178.
[39] G. Huang, A.Y. Zomaya, F.C. Delicato, P.F. Pires, Long term and large scale time
synchronization in wireless sensor networks, Comput. Commun. 37 (2014)
77–91.
[40] M. Leng, Y. Wu, Distributed clock synchronization for wireless sensor networks
using belief propagation, IEEE Trans. Signal Process. 59 (11) (2011) 5404–5414,
doi:10.1109/TSP.2011.2162832.
[41] B. Luo, Y.C. Wu, Distributed clock parameters tracking in wireless sensor net-
work, IEEE Trans. Wirel. Commun. 12 (12) (2013) 6464–6475, doi:10.1109/TWC.
2013.103013.130811.
[42] K. Noh, Q.M. Chaudhari, E. Serpedin, B.W. Suter, Novel clock phase offset and
skew estimation using two-way timing message exchanges for wireless sensor
networks, IEEE Trans. Commun. 55 (4) (2007) 766–777, doi:10.1109/TCOMM.
2007.894102.
[43] A. Papoulis, Probability, Random Variables, and Stochastic Processes, third ed.,
Mcgraw-Hill College, 1991.
[44] R. Landry, I. Stavrakakis, Study of delay jitter with and without peak rate
enforcement, IEEE/ACM Trans. Netw. 5 (4) (1997) 543–553, doi:10.1109/90.
649512.
Konstantinos Skiadopoulos obtained his BSc in physics
from the University of Athens, Greece in 1985 and his
MSc in Informatics from the Department of Informatics
at the Ionian University in Corfu, Greece in 2013. He is
now a PhD Candidate at the Department of Informatics
at the Ionian University and a member of the NMS lab in
the same department. He is currently a Secondary Edu-
cation teacher (1993-today) in Corfu, Greece. His research
interests focus on algorithms, graph theory and informa-
tion dissemination in networks.
Athanasios Tsipis was born in Corfu, Greece in 1989.
In February 2015, he received his B.Sc. degree in Infor-
matics with specialization in the field of “Information
Systems” from the Department of Informatics, Ionian Uni-
versity, Greece. Two years later, in February 2017, af-
ter receiving a full tuition-free scholarship, he received
his M.Sc. diploma in Informatics with excellence in the
subject area of “Computer Technologies and Information
Systems” from the same Department. Since March 2015,
Athanasios Tsipis is a member of the “Networks, Multime-
dia and Security Systems Laboratory” (NMSLab) of Ionian
University. During this time he has participated in EU and
national funded development and research projects and
has worked as an independent web and database programmer. Athanasios Tsipis
is currently a Ph.D. student at the aforementioned Department in the area of net-
working. His research interests concentrate on cloud gaming, cloud networking, 5G
computing, wireless systems, facility location issues, etc.
 K. Skiadopoulos, A. Tsipis and K. Giannakis et al. / Ad Hoc Networks 89 (2019) 47–57 57

Konstantinos Giannakis obtained his BSc and his MSc in
informatics from the Department of Informatics at the Io-
nian University in Corfu, Greece (in 2011 and 2013, re-
spectively). He received his PhD in Theoretical Computer
Science from the Dept. of Informatics of the Ionian Uni-
versity in 2016. He is a member of the NMS lab in the
same Department, as well as a member of the QUIT
research group. He has participated in various funded
projects and he has also served as a reviewer and TPC
member for several conferences and journals with high
impact. His research interests focus on the theory of com-
putation with emphasis on automata theory, algorithms,
quantum and unconventional computation, the founda-
tions of the semantic web, and graph theory. Currently, he is a postdoctoral as-
sociate of the Dept. of Informatics and the NMS lab at the Ionian University.
Georgios Koufoudakis received his diploma in electrical
and computer engineering from National Technical Uni-
versity of Athens in 2005. In 2012 he received his post-
graduate degree: M.Sc. in Informatics from Ionian Univer-
sity. He is currently a PhD Candidate at Ionian University,
Department of Informatics. His current interests lie in the
area of algebraic graph theory for computer networks.
Dr. George Kormentzas is currently associate professor
in the University of the Aegean, Department of Infor-
mation and Communication Systems Engineering. George
Kormentzas was born in Athens, Greece on 1973. He re-
ceived the Diploma in Electrical and Computer Engineer-
ing and the Ph.D. in Computer Engineering both from the
National Technical University of Athens (NTUA), Greece, in
1995 and 20 0 0, respectively. He has been actively work-
ing for many years on the area of network management
and quality of service of computer and communication
systems where he has introduced the concept of abstract
information model, an ancestor of next generation net-
working middleware, which constitutes his main current
research interest along with event-based distributed systems. He has published ex-
tensively in international scientific journals, edited books and conference proceed-
ings. He is a member of pronounced professional societies, an active reviewer for
several journals and conferences and EU-Independent Expert (being both Evalua-
tor and Rapporteur) for various EC framework Programmes. Furthermore, George
Kormentzas acts within Technical Program and Organizing Committees for numer-
ous conferences and workshops and as a Guest Editor for scientific journals. After
invitation, George Kormentzas is member of the IFIP WG6.6 (Management of Net-
works and Distributed Systems) of the TC6-Communication Systems. At 2008, he
served as General Chair of 4th IEEE/IFIP International Week on Management of Net-
works and Services and TPC co-chair of 19th IFIP/IEEE International Workshop on
Distributed Systems: Operations and Management (DSOM). For three years (2010 -
2012), George Kormentzas acted as Vice President of OPEKEPE, the Greek Paying
Agency for Community Aids according to the EC Common Agriculture Policy (CAP).
The last four years, G. Kormentzas is member of COPA-COGECA praesidium repre-
senting the Greek Farmers’ cooperatives.

Eleni Christopoulou is a lecturer at the Department of

Prof. Ioannis Stavrakakis (IEEE Fellow) (http://cgi.di.uoa.
Informatics of the Ionian University, Corfu, Greece and
gr/∼ioannis/), is professor in the Dept of Informatics
a tutor at the Undergraduate program on Computer Sci-
and Telecommunications of the National and Kapodistrian
ence of the Hellenic Open University, Greece. She received
University of Athens. He received his Diploma in Electri-
her Diploma in Computer Engineering and Informatics
cal Engineering from the Aristotelian University of Thes-
from the University of Patras in 2001 and her MSc in
saloniki and his PhD in the same field from University of
Computer Science and Engineering from the University of
Virginia, USA. He served as Assistant Professor in CSEE,
Patras in 2004. She holds a PhD on Computer Science
University of Vermont (USA), 1988–1994; Associate Pro-
with focus on ontologies in context-aware mobile and
fessor of Electrical and Computer Engineering, Northeast-
ubiquitous computing systems from the Department of
ern University, Boston (USA), 1994–1999; Associate Pro-
Computer Engineering and Informatics of the University
fessor of Informatics and Telecommunications, National
of Patras, Greece since 2013. She has taught as an ad-
and Kapodistrian University of Athens (Greece), 1999–
junct lecturer at the Department of Business Administra-
2002; and Professor since 2002. Teaching and research
tion of the Technological Educational Institute of Epirus, Greece and the Depart-
interests are focused on Analysis and Design aspects of networking technolo-
ment of Business Administration of the Technological Educational Institute of West-
gies ranging from link to application layers: Social, mobile, ad hoc, autonomic,
ern Greece and she has worked as an R&D engineer at the Computer Technology
information-centric, delay tolerant and future Internet networking; network re-
Institute in Patras, Greece. Her research interests include knowledge representation
source allocation algorithms & protocols, traffic management and performance
and management in mobile and ubiquitous computing systems, context-awareness
evaluation; content dissemination, placement and (cooperative) replication in un-
in smart systems and smart cities as well as big data in IoT environments. She has
structured P2P and social networks; (human-driven) decision making in competi-
contributed to a number of Greek and European research programs mainly in the
tive environments. His research has been published in over 220 scientific journals
area of ubiquitous computing, context-aware systems and smart cities and she has
and conference proceedings and was funded by USA-NSF, DARPA, GTE, BBN and Mo-
been a reviewer and TPC member of numerous conferences and journals in the
torola (USA) as well as Greek and European Union (IST, FET, FIRE) funding agencies.
area. She has received a scholarship from the John S. Latsis Public Benefit Founda-
He has received 2 Marie-Curie grants for training post and has supervised about
tion, Greece, and a visiting researcher scholarship from the University of Oulu, Fin-
20 Ph.D. graduates. He has served repeatedly in NSF and EU-IST research proposal
land. She holds a best student paper award from the 2nd Workshop on Pervasive
review panels and involved in the TPC and organization of numerous conferences
Urban Applications and a best paper award from the 16th Panhellenic Conference
sponsored by IEEE, ACM, ITC and IFIP societies. He has served as chairman of IFIP
on Informatics.
WG6.3 and elected officer for IEEE Technical Committee on Computer Communica-
tions (TCCC). He has been in the editorial board of Proceedings of IEEE (2015-),
Konstantinos Oikonomou has received his MEng in com- Computer Communications (2008-), IEEE/ACM transactions on Networking, ACM/
puter engineering and informatics from University of Springer Wireless Networks and Computer Networks journals. He has served as
Patras in 1998. In September 1999 he received his post- head of the Communications and Signal Processing Division, Director of Graduate
graduate degree: M.Sc. in Communication and Signal Pro- Studies, Dept Vice-Chair and Dept Chair.
cessing from the Electrical and Electronic Engineering
Department, Imperial College (London). He received his
Ph.D. degree in 2004 from the Informatics and Telecom-
munications Department, University of Athens, Greece.
His Ph.D. thesis focuses on medium access control poli-
cies in ad hoc networks. He is currently the dean of the
faculty of Information Science and Informatics of the Io-
nian University, Corfu, Greece. Since May 2018 he is the
director of the M.Sc. Course “Research Directions in Infor-
matics” of the Department of Informatics of the same university. Since 2006 he is a
faculty member in Computer Networks at the same department. He is an Associate
Professor and he has also served as the Head of the Department. Between Decem-
ber 1999 and January 2005 he was employed at Intracom S.A, as a research and
development engineer. His current interests involve medium access control in ad
hoc networks, performance issues in wireless networks, information dissemination,
service discovery, facility location, energy consumption in wireless sensor networks
and network cost reduction in cloud computing environments. Dr. Oikonomou has a
long experience regarding wireless systems and he has been co-ordinating a num-
ber of EC research projects in the computer networks research area and in various
local development projects (e.g., virtual worlds). He has been a reviewer and TPC
member of numerous conferences and journals in the area. He holds an award for
the best paper from the Hawaii International Conference on System Service and an-
other one from the Hellenic Mathematical Society.