Available via license: CC BY-NC-ND 4.0
Content may be subject to copyright.
Ultra-Low Power and Green TSCH-based WSNs
with Proactive Reduction of Idle Listening
Stefano Scanzio, Senior Member, IEEE, Federico Quarta, Giacomo Paolini, Member, IEEE,
Gabriele Formis, Student Member, IEEE, and Gianluca Cena, Senior Member, IEEE
Abstract—Wireless sensor networks are characterized by low
power consumption because motes are typically battery-powered.
Time slotted channel hopping (TSCH) relies on a fixed trans-
mission schedule, which enables the receiver module of wireless
motes to be switched off every time it is not needed. Unfor-
tunately, in many practical contexts most of the reserved slots
remain unused, which leads to appreciable energy waste. For
periodic traffic, proactive reduction of idle listening (PRIL)
techniques have been proven able to mitigate this problem.
In this paper, PRIL multi-hop (PRIL-M) is introduced with
the aim to improve existing PRIL techniques, by lowering
energy waste further in large real-world mesh networks. PRIL-M
is advantageous in all those contexts where ultra-low power
consumption is more important than end-to-end latency. Applica-
tions that can benefit from PRIL-M include, e.g., environmental
monitoring, where sensors are deployed over the target area
and must operate for years without maintenance. A thorough
simulation campaign showed that, in these scenarios, energy
consumption of PRIL-M is 75% less than standard TSCH, while
the average latency is about 20 times larger.
Index Terms—IEEE 802.15.4, time slotted channel hopping
(TSCH), wireless sensor networks (WSN), Internet of Things
(IoT), Ultra-low power, green networking, proactive reduction
of idle listening (PRIL), PRIL-F, PRIL-M
I. INTRODUCTION
Wireless communications, and especially wireless sensor
(and actuator) networks (WSN/WSAN), are a key enabling
technology for the Internet of Things (IoT). The contexts
where they can be applied range from unmanned surveillance
[1], natural disaster management [2], industrial networks [3],
and smart and precision agriculture [4], to arrive to the large
and constantly growing home automation sector [5].
Many wireless solutions are exploited nowadays in the
IoT, the most notable being IEEE 802.11 (Wi-Fi), 5G,
IEEE 802.15.4 (WSAN), long range wide area network
(LoRaWAN), and Bluetooth. Every technology has its own
characteristics and peculiarities. Those based on the IEEE
802.15.4 standard [6], which include the time slotted channel
hopping (TSCH) and deterministic and synchronous multi-
channel extension (DSME) MAC-layer profiles, were mainly
This work was partially supported by the European Union under the
Italian National Recovery and Resilience Plan (NRRP) of NextGenerationEU,
partnership on “Telecommunications of the Future” (PE00000001 - program
“RESTART”). (Corresponding author: Stefano Scanzio.)
S. Scanzio and G. Cena are with the National Research Council of Italy
(CNR–IEIIT), 20133 Milano, Italy (e-mail:stefano.scanzio@cnr.it).
F. Quarta is with the Politecnico di Torino, 10129 Torino, Italy.
G. Paolini is with University of Bologna, 40136 Bologna, Italy.
G. Formis is with the Politecnico di Torino, 10129 Torino, Italy, and also with
the National Research Council of Italy (CNR–IEIIT), 20133 Milano, Italy.
conceived for minimizing energy consumption, so that the time
between battery replacements may increase to several years.
As a consequence, they are valuable in all those situations
where maintenance costs must be as low as possible.
Several solutions exist that rely on TSCH concepts, the
most popular being WirelessHART, ISA 100.11a, and 6TiSCH.
DSME is instead a beacon-enabled protocol (that shares some
similarities with WIA-PA) for which open-source implemen-
tations exist like openDSME. This work focuses on TSCH
because of its wider availability in commercial devices. Com-
pared to DSME [7], the main difference is the absence of a
contention access period, which makes TSCH less suitable for
handling sporadic communications. However, its characteris-
tics make it very suitable for periodic traffic.
The solution we propose targets all those real application
contexts where decreasing power consumption is by far more
important than ensuring low end-to-end delays. Examples
include environmental monitoring, which in industrial plants
can be used to support product quality tracking in the brown-
field (that is, without dedicated communication/power-supply
infrastructures), e.g., by logging temperature and humidity
of several shop-floor and warehouse zones. In these cases
there are typically no specific constraints about latency, but
sensors must be able to operate for years with no mainte-
nance (deploy-and-forget). In particular, the following three
aspects must be borne in mind: a) battery replacement may
be problematic/uneconomical when devices are located in
difficult-to-reach places; b) sometimes, the whole device need
to be replaced when batteries are exhausted, e.g., radiator heat
meters; and, c) disposing large amounts of batteries is likely
to increase pollution, with a negative environmental impact.
The basic mechanism TSCH exploits to enable low-power
communication is time slotting, which is a simplified version
of time-division multiple access (TDMA). Time is divided into
slots of fixed duration Tslot and all network traffic (globally
defined in terms of a number of links, every one connecting
one or more senders to one or more receivers) is scheduled
to fit within these intervals. In this way, every node can
switch into a deep-sleep state in all those slots in which no
frames exchanges are scheduled for it, hence saving energy.
The slot schedule repeats periodically over time (the period
is in the order of one or two seconds), and every link is
assigned a specific slot (plus a specific channel offset, but this
is inessential for our work). Every time a slot is scheduled to a
given link but no transmissions takes place, the recipient node
experiences idle listening, i.e., it turns its receiving interface
on unnecessarily, thus wasting energy since there is no frame
This article has been accepted for publication in IEEE Internet of Things Journal. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2024.3406646
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
to be read.
In some cases, the amount of energy spent for idle listening
could be non-negligible. For instance, the charge drawn to
receive a maximal-size 127 B frame (and send back the related
acknowledgment (ACK) frame) in OpenMoteSTM devices is
651.0µJ, while the charge wasted for idle listening is 303.3µJ
[8], which is about one half of what is needed for data
reception. Since WSNs are often used for cyclically sensing
physical quantities characterized by slow dynamics (like tem-
peratures), sampling periods can be quite long (depending on
the application, they can range from less than one minute to
a few hours). Consequently, the amount of energy wasted by
idle listening may be quite high.
Whenever the effective usage of scheduled cells can be
predicted with good accuracy, proactive reduction of idle
listing (PRIL) can be employed to switch off receivers in
those slots where no frames are expected to be exchanged,
hence enabling ultra-low power communication. If traffic is
periodic with a known period, the PRIL-F technique has been
defined [9] that only considers the first hop in the path between
the source node of a flow and its destination. It can be used
to save a sizable amount of energy in single-hop networks
(star topology), as well as in those setups where part of the
intermediate relays are mains-powered. The major limitation
of PRIL-F is that only a subset of the nodes benefits from a
reduced energy consumption. As a matter of fact, real WSNs
are typically multi-hop, and because of their tree topology
(customarily enforced by routing protocols), nodes closer to
the root consume more energy. Unfortunately, PRIL-F tends
to have little effect on them.
In this paper a new technique is proposed, we named
PRIL-M, which specifically targets multi-hop networks. It has
been conceived to work together with PRIL-F, overcoming its
limitations. The two techniques have been compared and their
joint use (PRIL-F is applied to the first hop and PRIL-M to the
following hops) analyzed by means of an extensive simulation
campaign, to thoroughly investigate the benefits they can
provide in realistic environmental monitoring applications.
Results show that the combined use of PRIL techniques yields
significant energy savings at the expense of latency, which
suits the requirements of the considered scenario.
The next section briefly introduces TSCH and the PRIL-F
technique, while the new PRIL-M technique is presented in
Section III. The experimental setup is described in Section IV,
followed by the presentation of results in Section V. Conclu-
sions are finally drawn in Section VI.
II. PROACTIVE REDUCTION OF IDLE LISTENING
A survey is provided below about the operating principles
of TSCH and PRIL-F, on which PRIL-M grounds.
A. TSCH basics
A scheduled TSCH slot (i.e., assigned to a link) is named
cell.Shared cells can be used by a plurality of transmitting
nodes, whereas every dedicated cell is allocated to a single
transmitter (there is no contention). The latter are by far the
most common option in real setups, and are typically used
Slotframe matrix
1→3
0
15 4→0
1
…
0 1 2
Channel offset
34 5 6
3→0
2→3
3→0
0
1
3
2
4
0Root node N03Relay node N31Leaf node N1
Fig. 1. Example of TSCH operation and slotframe matrix.
for unicast frame transmission (in a given direction) between
a specific pair of nodes, e.g., for propagating data sensed
from the real world toward the intended sink (root node).
The TSCH schedule is defined by a slotframe matrix with
size Nch ×Nslot, an example of which is depicted in Fig. 1.
The schedule spans in two orthogonal directions: time (time
slotting) and frequency (channel hopping).
Regarding time, the slotframe matrix repeats identically
with period Tsfr =Nslot ·Tslot, and defines which commu-
nications are allowed in any specific slot (and between which
nodes). The notion of time in TSCH is uniquely identified
by a counter shared among all the nodes of the network,
known as the absolute slot number (ASN), which is initialized
to zero and is incremented by one on every slot. Nodes are
time synchronized by a specific clock synchronization protocol
[10], precise enough to ensure that boundaries between slots
are adequately obeyed by every node. Typical configuration
parameters of the slotframe matrix in real implementations are
Nslot = 101 and Tslot = 20 ms, which implies Tsfr = 2.02 s.
Regarding frequency, up to Nch transmissions can be carried
out concurrently (between different pairs of nodes) in every
slot (identified by the ASN). In particular, for operations in
the 2.4 GHz band with offset quadrature phase shift keying
(O-QPSK) PHY modulation, the IEEE 802.15.4 standard
defines 16 channels (Nch = 16). Let cbe the logical channel
(or channel offset) of a specific scheduled cell in the slotframe
matrix, which coincides with its row number. The physical
channel ch on which transmission is performed is derived
from cthrough the formula ch = H[(ASN + c) mod Nch],
where H[ ] is the hopping sequence (typically implemented
as a vector of Nch elements) that performs such mapping. If
Nslot and Nch are selected as coprime numbers, in case of
errors subsequent retransmissions of the same packet will be
performed on different channels, which makes link reliability
less susceptible to the behavior of single channels [11].
As thoroughly analyzed in [12], proper selection of network
parameters (e.g., Nslot,Tslot , and retry limit) permits to set a
trade-off among latency, reliability, and power consumption.
PRIL techniques constitute an additional method, orthogonal
to configuration, to decrease energy consumption.
B. Conserving energy in WSNs
Consuming as little energy as possible is one of the main
goals of WSNs, and TSCH is no exception. Whenever no
scheduled receptions or transmissions are foreseen, a TSCH
node can be safely switched into a deep power-saving state,
This article has been accepted for publication in IEEE Internet of Things Journal. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2024.3406646
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
in which energy consumption is negligible. This aspect makes
this solution very attractive.
A number of research activities are available in the litera-
ture that specifically focus on power consumption. Some of
them act on specific aspects of the protocol. The active-scan
procedure proposed in [13] permits to decrease the network
formation time, while in [14] the same goal is obtained
using Q-learning to reduce the congestion of messages used
by the network formation algorithm. Balancing nodes’ load
is another way to reduce battery discharge [15]. In [12] it
is shown that reducing energy consumption is possible by
tuning communication parameters. Other contributions focus
on improving synchronization accuracy among TSCH nodes:
in [16] a guard beacon strategy is proposed to decrease the
guard time, in [17] the guard time is selected by each node
based on its distance from the root, and in [18] a better
synchronization accuracy is obtained by modifying the sending
period of the enhanced beacon (which, besides other purposes,
delivers information about time).
Other techniques optimize instead the routing protocol for
low-power and lossy networks (RPL), which is typically used
to generate the network topology (routes) in TSCH-based
networks. In particular, in [19] RMA-RP is introduced to make
RPL more energy efficient in case of node mobility, in [20]
mobility is managed by improving the way RPL selects the
new parent (e.g., by using more effective metrics), and in [21]
nodes adjust the sending frequency of control messages based
on the mobility level of neighbor nodes.
Generally speaking, all the approaches aimed at increasing
communication quality (latency and reliability) by decreasing
the average number of retransmissions have a positive impact
on power consumption as well. This is the case of black and
white listing techniques, which are meant to dynamically select
the best channel on which transmissions are performed: in [22]
bad channels are blacklisted based on the packet delivery ratio,
[23] introduces the concept of probabilistic blacklisting, where
the probability that a channel is not skipped depends on its
perceived communication quality, in [24] cells are scheduled
based on network metrics including the packet delivery ratio
and end-to-end latency, and finally in [25] blacklisting is
managed as a dynamic multi-armed Bernoulli bandit process
to adapt selected channels to time-varying interference.
Historically, a large part of the techniques aimed at reducing
power consumption in this kind of WSNs relies on scheduling
algorithms [26], by optimally assigning transmission frequen-
cies and slots of the slotframe matrix according to some
specific objective function that minimizes, for instance, power
consumption. In particular, in [27] the VAM-HSA algorithm
was proposed to solve an energy-efficient maximization prob-
lem by reducing its complexity, in [28] the PRCOS algorithm
was proposed to obtain a scheduling aimed at maximizing
system lifetime by taking into account power consumption,
in [29] a MAC layer scheduling algorithm named TREE was
proposed to dynamically optimize allocated cells, instead in
[30] the scheduling algorithm optimizes the use of shared
cells. A suitable scheduling not only permits to decrease power
consumption of the network, but also to improve its real-time
capabilities: in [31] this was achieved with the combination of
spectral clustering unsupervised learning and earliest deadline
first (EDF) scheduling algorithm, while in [32] similar goals
were obtained with an orchestrator (SDN-TSCH) that acts in
a centralized fashion.
Concerning techniques based on proactive reduction of idle
listening, PRIL-F, was preliminary introduced in [33] and
completely defined in [9], whereas PRIL-M is first presented
in this work. Both PRIL-F and PRIL-M are completely in-
dependent from above solutions, hence their joint adoption is
possible to reduce energy waste further.
C. PRIL over the first hop (PRIL-F)
The basic idea behind PRIL to prevent idle listening (or
reduce it as much as possible) is to switch off receivers also
in their scheduled cells when no frames are expected. Several
options exist to determine the occurrence of such conditions.
A quite sophisticated approach is that the nodes analyze
traffic at runtime and automatically determine the actual usage
pattern of their scheduled cells, e.g., by using machine learning
algorithms. This information is then exploited to disable those
cells that, with high likelihood, would remain unused, hence
incurring in idle listening. Simpler techniques exist as well
that foresee that the transmitter on a link directly controls the
receiver based on locally available information, by turning it
off for a given number of subsequent scheduled cells so that
energy can be saved on the receiving side.
In this work, we analyze only techniques of the second type,
where specific sleep commands are added to the frames ex-
changed on a link to explicitly instruct recipients to go to sleep.
We mostly focused on those IoT applications in which sensors
belonging to the perception layer (either leaf or intermediate
relay nodes in the tree network topology) generate cyclic data
that are sent in the upward direction to a sink (root node). This
is the usual case in almost all real WSNs, therefore it does
not constitute a real limitation. However, the techniques we
introduce also work in the reverse (downward) direction. An
efficient way to deliver sleep commands to the receiver is by
means of information elements (IE), which are specific user-
defined elements that can be included in the header of IEEE
802.15.4 frames. Let Ti
app be the generation period of node Ni.
Intuitively, the longer the generation period, the more energy
can be saved by PRIL techniques, because the receiving side
of the link can be turned off for longer periods of time.
The first proposal for reducing idle listening is PRIL-F,
which was conceived to lower the energy waste of receiving
nodes located one hop away from the traffic source. For
example, by looking at the topology of Fig. 1, when the
source node is N1it can disable N3temporarily so as to
make it conserve energy. In the case of traffic flowing in the
upward direction from N1to N0passing through relay N3
(N1→N3→N0), PRIL-F is only effective on N3and can
do nothing for N0. In fact, N3has no clue in advance of
when packets will arrive from N1that it must forward to N0,
and hence it is seemingly unable to generate effective sleep
commands for the latter.
Let notation Ni
x
−→Njrepresent the scheduled cell xof the
link Ni→Njwhen Njleaves its receiving interface on (this
This article has been accepted for publication in IEEE Internet of Things Journal. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2024.3406646
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
TSCH
PRIL-F
RX IDLE IDLE RX
x x+1 x+2 y
RX IDLE
y+1 y+2
ACK
Data frame arrived
(sleep command: 2)
ACK frame arrived
ACK
ACK
RX OFF OFF RX
x x+1 x+2 y
RX OFF
y+1 y+2
2
ACK
IDLE Idle listening
OFF Cell switched off
by PRIL
RX Reception
21
ACK
2
Data frame not arrived
Diff. Diff.
1
2
Fig. 2. Example of PRIL-F operation compared with default TSCH.
corresponds to the default TSCH behavior). When a single
cell of the matrix is allocated to the link, the x-th cell is
found in slotframe x. If a frame is transmitted in the cell, it
is received by Nj(possibly corrupted by errors); otherwise, if
no transmissions took place, Njexperiences idle listening. The
case when reception in cell xhas been (temporarily) disabled
on Njby PRIL is instead represented by the notation Ni
x
−→Nj.
An example that compares TSCH and PRIL-F operations
is shown in the timing diagram of Fig. 2, where a packet is
enqueued in the sender node N1every three slotframes. Let us
assume that the initial transmission attempt is performed for
a certain frame Fin cell N1
x
−→N2(this event is denoted Fx).
Following the successful exchange, the link state of standard
TSCH in the immediate future (before the next packet is
generated) can be described as N1
x+1,x+2
−−−−−→ N2, and hence
cells x+ 1 and x+ 2 suffer from idle listening.
If, as depicted in the leftmost part of the figure, the frame
reaches the destination node N2immediately (without any
retry), as notified by the reception of the related ACK frame,
the following two scheduled cells for the link (i.e., x+ 1 and
x+ 2) could be safely disabled to save energy on it. As said
before, PRIL-F does so by making the sender include an IE
in the frame that tells the receiver the number of scheduled
cells for which it must sleep. This can be specified either
in absolute terms (ASN of the scheduled cell when the link
must be re-enabled) or in relative terms (number of scheduled
cells for which the receiver node must stay disabled). These
two solutions are equivalent from the functional point of
view; however, we opted for the latter because it simplifies
descriptions and the information to be stored in the IE is
smaller. Referring again to Fig. 2, if N1knows that the
packet generation of the application executing on it is periodic
with period Tapp = 3 slotframes, it can add a specific sleep
command with value 2. Frame transmission in this case is
symbolically represented by notation F.s=2
x, where “.s = 2”
denotes the sleep IE. If correctly received by N2, the following
two scheduled cells of the link will be disabled, which can be
described as N1
x+1,x+2
−−−−−→N2.
The rightmost part of the figure describes instead what
happens when transmission errors occur that corrupt the frame.
In particular, frame F.s=2
yis sent in slot ybut fails to reach
the destination. Thus, it is retransmitted in slot y+ 1, after
decreasing the value in the sleep command by one, since this
is a retry (this action is synthetically described as F.s=1
y+1 ). In
this case, only cell N1
y+2
−−→N2will be disabled, and the amount
of energy saved on N2is smaller.
If a frame with a sleep IE correctly arrives to destination
but the ACK is lost, the receiver turns the relevant cells off
but the sender is unaware of this. Hence, all further retries of
that frame will no longer reach the destination, until the sleep
command ends and the link is re-enabled on N2.
III. PRIL OVE R MU LTIPLE HOPS (PRIL-M)
The PRIL multi-hop (PRIL-M) technique proposed in this
work is meant to overcome the main limitation of PRIL-F,
i.e., that it can save energy only on nodes located one hop
away from the traffic source. The technique we present is
characterized by modest complexity, which enables direct
and effective implementations on real WSN devices (typically
characterized by low computational power).
In Fig. 3 a sample network setup is shown where three
periodic traffic flows from leaf nodes to the root are defined,
τ1=hN1→N4→N0, T1i,τ2=hN2→N4→N0, T2i, and
τ3=hN3→N4→N0, T3i, where T1,T2,T3represent the
respective generation periods. When applied to this example,
PRIL-F only affects N4, whereas PRIL-M impacts on all the
following nodes in the path to the destination, i.e., those with
a distance from the source greater than one hop. Consequently,
the joint use of PRIL-F and PRIL-M mitigates the idle
listening problem for all the nodes in the WSN that have
receiving cells scheduled for unicast transmissions.
Referring to the above example, PRIL-M counteracts idle
listening in N0by disabling a subset Sof the cells scheduled
to the link N4→N0(this action is denoted N4
S
−→ N0). The
way subset Sis determined is more complex than PRIL-F:
in fact, a link controlled by PRIL-M is typically traversed
by many packet flows, including those not directly generated
by the node that handles transmissions on the link (for them,
it just acts as a relay) but by nodes located at lower levels
in the tree topology (or upper, for downward flows). In other
words, not all frames sent on N4−→N0are generated by N4. In
these conditions, correctly predicting whether or not a specific
scheduled cell will be left unused is hardly possible for relay
nodes like N4. The main problem is due to retransmissions
of frames (when they are not acknowledged), which generate
unpredictability in the arrival time (and order) of packets
N4
N3
N2
N1
(T1= 60 s)
N0
N4
N0
Relay node
(managed by PRIL-F)
Relay node
(managed by PRIL-M)
L0
L1
L2
(T2= 120 s)
(T3=180 s)
Fig. 3. Simple network topology to explain PRIL-M operations.
This article has been accepted for publication in IEEE Internet of Things Journal. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2024.3406646
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
belonging to distinct flows. In large mesh networks, where
paths from sources to sinks may consist of many hops, traffic
flows get mixed in increasingly random way the closer links
are to the root node.
The main idea behind PRIL-M is to have the transmitting
side of a link turning off the receiving side (by using sleep
commands) for a time equal to the minimum among the
periods of the flows traversing the link itself. In fact, the flow
with the smallest period is typically the one with the tightest
constraints in terms of delivery latency (soft deadlines, in our
case). To propagate the knowledge about the period of any flow
τ`to all the following nodes in the related path, the originating
node (source) attaches a timing IE to its frames that specifies
its generation period. Unlike sleep IEs, which act on a single
link, timing IEs attached to the frame are forwarded unchanged
by relay nodes between subsequent hops. For instance, in the
example of Fig. 3, the frames sent by N1(hence belonging to
flow τ1) are encoded as F.T =T1, where the timing IE denoted
“.T =T1” is added by N1and propagates period T1along the
entire path N1→N4→N0.
To operate correctly, every relay that executes PRIL-M
(e.g., N4) must know the minimum period among all the
flows crossing its outgoing link. This information is typically
acquired in the learning phase, and enables PRIL-M to save
energy when the node enters the subsequent runtime phase. A
separate instance of PRIL-M is executed on every relay node
for every outgoing link.
A. Learning phase
At network startup (or upon network topology changes),
when a relay receives a frame F.T=T`that contains a timing
IE with the period T`of the flow τ`it belongs to, it starts a
learning phase whose duration equals T`(or a small multiple
of it). In the learning phase the node collects the periods of
all the frames it forwards (grouped by outgoing link) and
computes the minimum among such values, denoted Tmin. The
node that generates the related “fastest” flow is identified as
Nref . If several flows exist for a given outgoing link whose
periods coincide with Tmin, which are generated by distinct
upstream nodes, one of them is selected as Nref .
After the learning phase is finished, every time a frame
F.T =T`is received for which T`< Tmin, both Tmin and
Nref are updated consequently. Instead, if the relay does not
receive any frames belonging to the flow that originates from
the current Nref for a given time period Tmax (the timeout
can be set, e.g., to Tmax = 10 ·Tmin), it switches back to
the learning phase, looking for another flow with minimum
period. In case of topology changes, in TSCH-based networks
the new routing is typically decided by the RPL protocol. This
operation can take several minutes [34], during which the node
is set to behave according to legacy TSCH. When routing
is re-established, the node enters the learning phase again,
with the aim of re-activating PRIL-M as soon as possible.
Optimizations can be brought to the above procedure, to
lower both the duration of the learning phase and its energy
consumption, but they have not been reported here since
the benefits we obtained were mostly inessential. During the
learning phase, the node obeys standard TSCH rules.
B. Runtime phase
In this phase both Tmin and Nref are correctly initialized.
Let Ni−→ Njbe a link operated according to PRIL-M that is
traversed by frames of flow τ`to reach their destination. With
reference to Fig. 3, Nicorresponds to N4and Njcorresponds
to N0. Node Niunderstands that PRIL-M must be used instead
of PRIL-F because it is not the originator of the flow, i.e., it
is not the first node of the related path. As in PRIL-F, a sleep
command F.s=Tend permits to deactivate the link for a given
amount of time, where Tend represents the number of cells
related to the link Ni−→Njwhere reception will be disabled.
Again, we assume that a relative value is included in the sleep
command, but the same reasoning applies to absolute values.
The basic idea is to deactivate the link for a time period
equal to Tmin, so as to prioritize periodic traffic with the
fastest generation rate. When this approach is implemented in
practice, various aspects must be taken into account, related
to the fact that frames can be lost and retransmitted up to a
maximum number of times. In particular, the link state seen
by Niand Njmay differ due to losses. For this reason, a
reference implementation for PRIL-M is provided by means of
two state machines, we denote RX and TX, reported in Fig. 4,
along with detailed explanations. Every arc is labeled with a
conventional event-action pair. The upper part (event) reports
the conditions that must be verified to enable the transition to
the new state. The leftmost condition (in bold) consists in one
of four possible events: sleep (reception of a sleep command),
ack (correct reception of an ACK frame), no ack (missing or
corrupted ACK frame), and slot end (end of a slot related
to a cell scheduled to link Ni−→ Nj). Instead, the lower part
(action) lists the sequence of operations that are executed on
state change, which are mostly aimed at updating the value of
the variables pertaining to the given state machine instance.
1) Receiving side: The RX state machine of receiver Nj
is depicted in Fig. 4.a. Node Njdefines a separate instance
of this state machine for every link, i.e., for every sender that
communicates with it. In the ON state the link is active and
obeys standard TSCH rules. When a frame F.s=Tend with a
sleep command is received (sleep event), the local counter
sleep_end is initialized to Tend and the state machine
switches to the OFF state, where the link is deactivated so
that Njstarts saving energy. In this state, counter sleep_end
is decreased by one on every new slot scheduled to the link
Ni−→Nj. At the end of every such slots (slot end event) the
counter sleep_end is checked. When it reaches zero the
RX state machine switches back to the ON state, and the link
returns to standard TSCH operation.
2) Transmitting side: The TX state machine for the sender
Niis shown in Fig. 4.b. As can be seen, it is more complex
than RX. When a frame F.T =Tmin sent by Nref is received
by Ni, counter sleep_end is initialized with the number of
cells scheduled to the link Ni−→ Njthat fit in an interval of
duration Tmin. For example, if Tmin = 2 min,Tsfr = 2.02 s,
and one cell is reserved per slotframe, sleep_end is set to
59. The value of sleep_end is then decremented by one on
every slot scheduled to the link. It is worth pointing out that,
at the arrival of F.T=Tmin , the transmission buffer of Nimay
This article has been accepted for publication in IEEE Internet of Things Journal. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2024.3406646
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
a) RX state machine ()
ON
b) TX state machine ()
OFF
ON OFF
RETR
no_ack && queue_size()==1
&& sleep_end>0
slot_end && sleep_end==0
if (new_sleep_end>0)
sleep_end = new_sleep_end
new_sleep_end = 0
slot_end && sleep_end==0
if (new_sleep_end>0)
sleep_end = new_sleep_end
new_sleep_end = 0
ack && queue_size()==1 && sleep_end>0
ack && sleep_end>0
no_ack && sleep_end>0
&& is_last_try(pkt)
slot_end && sleep_end==0
sleep
sleep_end=Tend
Fig. 4. State machine of the receiver side (Nj, on the left) and sender side (Ni, on the right) of an intermediate link in a path.
already contain queued packets. Moreover, additional packets
may arrive after F.T=Tmin is relayed, which will be enqueued.
According to TSCH rules, all packets in the outgoing queue
of link Ni−→ Nj(on Ni) will be orderly sent (one attempt
per scheduled cell). When transmitting the last of them (that
is, when queue_size()==1), Niadds a sleep command to
the frame, that is, it sends F.s=Tend with Tend equal to the
residual value of sleep_end at that time.
If transmission of F.s=Tend succeeds at the first attempt, the
TX state machine switches directly to the OFF state; other-
wise, it moves to the RETR state, whose purpose is managing
retransmissions of frames that embed a sleep command. In
the RETR state, Nican not determine if Njis in the ON or
OFF state, because it does not know whether the data frame or
the ACK frame went lost. Consequently, it keeps on retrying.
If transmission eventually succeeds, i.e., an ACK frame is
received by Ni(ack event), Niswitches to the OFF state
and is aware that also Njis in the OFF state. Instead, if the
retry limit (as specified by Ntries) is exceeded (no ack event
and is_last_try(pkt) returns true), Nistill moves to the
OFF state, but it does not actually know whether Njis in the
ON or OFF state. Doing so prevents the loss of subsequent
frames from Niwhen Njmanaged to switch to the OFF state.
When Niis in the OFF state, it waits until sleep_end
reaches 0to return to the ON state. Less likely, but possible,
is the situation where Nireturns to the ON state from the
RETR state. This happens when, during the transmission of
the last packet in the queue, the value of sleep_end is
close to 0and, due to retransmissions, it reaches 0before
Niexceeds the retry limit. The actions performed for the two
arcs from OFF to ON and from RETR to ON make use of the
variable new_sleep_end to deal with the arrival of a new
packet F.T =Tmin at Niwhen it is still in the OFF or RETR
states. This packet may arrive earlier than expected because,
for instance, the previous packet F.T=Tmin was late due to
retransmissions or excessive queuing delays in the preceding
hops of the path (upstream nodes). Since in the OFF and
RETR states the counter sleep_end is already in use by
the above algorithm (and hence, it cannot be overwritten), a
secondary counter new_sleep_end is needed to keep track
of the future reactivation time of the link itself. Both counters
are contextually decreased by one on every scheduled cell of
the link Ni−→ Nj. When the node returns to the ON state, if
new_sleep_end is greater than zero (i.e., it is active), it is
copied in sleep_end.
3) Implementation complexity: PRIL-M requires the addi-
tion of two IEs to the frame, namely the sleep command
F.s and the flow generation period F.T . In addition, the RX
and TX state machines of Fig. 4 have to be implemented:
they are quite simple, and every instance includes just 9
variables, counting both those used to store the link state and
those for managing the learning phase. A separate instance of
PRIL-M must be created for every outgoing link. Summing
up, the variables to store the current state of the RX and
TX state machines require 1 bit and 2 bits, respectively. If
2 bytes are allocated for each one of the remaining seven
variables (encoded as unsigned short integers), every instance
takes less than 15 bytes, which is a modest amount of memory
also for motes. It is worth pointing out that only one copy of
the executable code implementing PRIL-M must be stored in
memory, regardless of the number of instances.
Concerning memory footprint it must be noted that, because
of queuing, larger transmission buffers are seemingly needed.
However, since transmissions are regulated by the flow with
the shortest period, no more than one packet belonging to the
other flows could be in theory enqueued at any given time,
which is the same as TSCH (in other words, buffer size is the
same but its effective usage in PRIL-M is noticeably higher).
Regarding the size of the code segment, it is indicatively less
than 2 kbytes, but the exact amount depends on many aspects,
including the CPU type (x86 or ARM) and the operating
system in use. It is worth pointing out that popular devices
implementing TSCH, such as OpenMote B, are characterized
by 256 kbytes of flash memory and 32 kbytes of RAM,
which is more than enough for implementing PRIL-M. Finally,
computational complexity for relay operations on a link is
O(1), and only marginally affects CPU resource usage.
C. Example
In Fig. 5, three examples are provided to make the details of
PRIL-M operations clearer. In all cases, two frames, denoted
1Fand 2F, are found in the queue of the sender node (TX)
immediately before scheduled cell x. We set Ntries = 2 (only
one retry is allowed). Moreover, 1Fbelongs to the flow with
the shortest generation period (Tmin = 5 slotframes). In all
the cases reported in the figure, the state of the TX node is
described in the upper diagram, while the lower diagram refers
to the RX node. Moreover, let us assume that the first frame
(1F) is successfully delivered at time x.
This article has been accepted for publication in IEEE Internet of Things Journal. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2024.3406646
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
In CASE a, frame 2F.s=3
x+1 (which instructs the receiver to
sleep for 3scheduled cells) fails to reach the destination in
scheduled cell x+ 1 and is retransmitted in slot x+ 2 with a
sleep command decreased to 2, i.e., 2F.s=2
x+2 , which correctly
deactivates the RX node preventing idle listening in two slots.
Conversely, in CASE b also the second transmission of frame
2F.s=2
x+2 in slot x+ 2 fails, and since the maximum number
of allowed transmissions attempts is reached (Ntries = 2) the
frame is discarded and removed from the queue. This means
that the RX node does not reach the OFF state, consequently
behavior remains the same as standard TSCH and idle listening
is experienced. Finally, in CASE c, frame 2F.s=2
x+1 sent in slot
x+ 1 correctly arrives to destination but the corresponding
ACK is lost. The RX node enters the OFF state and starts
saving energy, but the TX node, which is not aware that the
frame has arrived, continues retransmissions until the retry
limit is reached, hence wasting some energy.
Having the state on two sides of the link always coherent
is impossible, since both data and ACK frames may be lost.
Nonetheless, it must be remarked that the state machines were
conceived to never allow the transmitter node to be in the ON
state when the receiver node is OFF. This intended behavior
ensures that frames cannot be lost due to PRIL-M.
D. Overall behavior
Above descriptions refer to the link between a pair of nodes,
managed on its two sides by the TX and RX state machines.
To understand the overall network behavior it must be remem-
bered that the path followed by the packets of every flow is
made up of the ordered sequence of traversed links, where the
state machine in every intermediate relay independently selects
the upstream node Nref with the fastest generation rate (whose
period is denoted Tmin). On the whole, this procedure can be
seen as a network-wide single-elimination tournament, where
the fastest flows autonomously selected by sibling relays (local
minima) compete in their common ancestor, in a process that
stretches from every leaf up to the root (this holds for both
the upward and downward directions).
This has two consequences: first, the values Tend selected
by the different relays (to be included in their respective sleep
commands) are usually not the same. Second, only one flow
CASE a
ON RETR OFF
x x+1 x+2 x+3
OFF ON
x+4 y
ACK
Data frame arrived
(sleep command: 2)
ACK frame arrived
ACK
2
Data frame not arrived
ON ON ON OFF OFF ON
TX
RX
ℱ2
ACK
3
Energy saved
CASE b
ON RETR OFF
x x+1 x+2 x+3
OFF ON
x+4 y
ACK
ON ON ON ON ON ON
TX
RX
2
3
Idle listening
ON RETR OFF
x x+1 x+2 x+3
OFF ON
x+4 y
ACK
ON ON OFF OFF OFF ON
TX
RX
2
3
Energy saved
CASE c
ACK
ACK
ACK frame not arrived
ON
OFF
RETR
States
Frames
queue
ℱℱℱ
ℱ
ℱℱℱ
ℱ
ℱℱℱ
ℱ
ℱ
Fig. 5. Example of PRIL-M operation in three different cases.
(the one for which Tapp is the global minimum) is granted
that its packets never suffer from additional delays due to
the outgoing links found disabled. For all the other flows,
packets receive a privileged treatment only in the lower part
of their path (close to the leaves), while they may suffer
from additional delays in the upper part (near to the root).
In particular, when arriving to a relay where a different Nref
has been selected, forwarding pace is dictated by the packets
of that flow, causing random delays uniformly distributed in
the range [0, Tmin]. In this way, packet forwarding occurs in
bursts, which inherit the treatment of the winner flow. This
means that no further delays will be experienced by them until
a relay is encountered with a different Nref . This situation has
been explicitly shown for the deep topology on the right side
of Fig. 6, by highlighting in green the portion of the paths
where the winner remains the same (and no relevant queuing
delay is experienced due to PRIL-M), whereas discontinuities
(that unavoidably lead to queuing delays) are colored in red.
IV. EXP ER IM EN TAL SETUP
Both PRIL-F and PRIL-M were compared with standard
TSCH using a discrete event simulator we purposely devel-
oped, named TSCH-predictor. It differs from other publicly-
available tools like 6TiSCH simulator [35] and TSCH-Sim [36]
in terms of the drastically lower implementation complexity.
This simulation framework permits new TSCH-based tech-
niques, aimed at improving network behavior, to be easily and
quickly implemented, assessed, and fully understood. In fact,
only those features strictly needed to model basic operations
of the specific techniques and to obtain raw indications about
their performance have to be actually implemented. Results
produced by our simulator were checked against measurements
obtained from a real network setup based on OpenMote B
devices running OpenWSN, and the absolute error concerning
the average latency was found to be negligible (less than
12 ms), this denoting accurate and faithful modeling. This
simulator has already been satisfactorily employed in former
scientific works [9].
Network parameters used for simulation were taken from
the default configuration of real devices, for example, slot
duration (Tslot = 20 ms), number of slots in a slotframe
(Nslot = 101), maximum allowed number of transmission
attempts (Ntries = 16). Instead, the failure probability for
attempts related to data frames (data = 12.6%) and ACK
frames (ACK = 8.0%) was derived from experimental data
obtained from a real setup1.
Regarding the power consumption model, the energy re-
quired to transmit one data frame including 127 B (the data
size used in this work) and to receive the related ACK was
set to 485.7µJ, the energy spent to receive every one of such
frames and to return the related ACK was set to 651.0µJ,
while the energy wasted for idle listening in a single slot was
set to 303.3µJ. These values were obtained from [8], and refer
to an STM32F103RB 32-bit microcontroller and an Atmel
1Experimental data used to compute the loss probability values are included
in the file default-101-16-15days.dat, whose details can be found
in https://dx.doi.org/10.21227/fg62-bp39, while the method for computing the
two probabilities is described in [9].
This article has been accepted for publication in IEEE Internet of Things Journal. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2024.3406646
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
AT86RF231 radiochip mounted on an Open-MoteSTM device.
Other energy models are available in the scientific literature,
which were obtained for different devices, but their adoption
does not substantially change the obtained results in terms of
the reduction of power consumption. The duration of simulated
experiments was set to one year for every campaign.
V. RESULTS
Experimental campaigns were performed for the three net-
work configurations depicted in Fig 6.
A. Simple network configuration
The simple topology (Fig 6.a), already analyzed in Sec-
tion III, is used to compare PRIL-F alone and PRIL-M
(always used in conjunction with PRIL-F), highlighting some
peculiarities. This configuration is characterized by three flows
directed from leaves to the root N0, whose generation periods
were set equal to 3001,6003, and 9005 slots, corresponding to
about 60 s,120 s, and 180 s, respectively. We selected coprime
numbers to models quartz tolerance in real devices. By doing
so, offsets between the generation times of frames belonging
to different traffic flows keep drifting slowly over the entire
duration of experiments (one year), making simulation results
more reliable.
Table I reports the power consumption of every single node,
as well as of all nodes in the network (row labeled “All”). In
particular, Pstands for the total power consumption (transmis-
sions and receptions) while Plisten only considers contributions
of receive modules (RX). As expected, PRIL-F practically
zeroes the Plisten component of the intermediate node N4, but
it has no effect on N0since is located two hops away from
the source. The reason why Plisten is not exactly zero depends
on the learning phase, during which PRIL does not provide
any benefits. When PRIL-M is additionally activated, energy
is saved also on N0. The total power consumption P(all
nodes included) is 663.90 µWfor standard TSCH, 239.22 µW
with PRIL-F (36.0% of TSCH), and 108.46 µWwith PRIL-M
(16.3% of TSCH). Hence, the latter approach doubled the
energy saving capabilities of PRIL-F.
The same configuration is then analyzed in terms of the end-
to-end latency, by taking into account the following perfor-
mance indices: mean value (µd), standard deviation (σd), 99−
to 99.99−percentiles (dp99,dp99.9, and dp99.99 ), and maximum
(dmax). Results, reported in Table II, highlight one of the most
severe drawbacks of PRIL-M: unlike PRIL-F, which always
brings beneficial effects (power consumption improves while
latency remains unchanged), PRIL-M significantly increases
delays, and consequently it should not be used in time-
sensitive applications. However, it is important to remark that,
in many application contexts, energy consumption is the first
and foremost performance index for WSNs.
Latency increase is more evident for flows other than the
one with the minimum generation period Tmin, as their packets
may remain stuck for a while in intermediate nodes when
the related outgoing links are temporarily disabled. As can be
seen, the mean latency µdexperienced by τ2and τ3is about
30 s. In fact, since the link returns to the ON state every ∼60 s
TABLE I
COMPARISON AMONG TSCH AND TH E PRO POS ED PRIL S TRATE GI ES
(SIMPLE TOPOLOGY): POWER CONSUMPTION.
a) Simple topology (2 layers)
TSCH PRIL-F PRIL-M
Node Hops Plisten P Plisten P Plisten P
[µW] [µW] [µW]
N02 138.64 163.34 138.62 163.36 0.19 23.83
N41 438.92 482.09 .0017 41.20 .0017 50.11
N30 0 3.36 0 6.34 0 6.25
N20 0 5.04 0 9.46 0 9.42
N10 0 10.07 0 18.85 0 18.87
All 577.56 663.90 138.63 239.22 0.20 108.46
TABLE II
COMPARISON AMONG TSCH AND TH E PRO POS ED PRIL S TRATE GI ES
(SI MPL E TOP OL OGY ): E ND-T O-EN D LATE NCY.
a) Simple topology (2 layers)
Flow µdσddp99 dp99.9dp99.99 dmax
(Source) [s]
TSCH All 1.720 1.389 6.220 9.400 12.000 17.260
PRIL-F All 1.722 1.383 6.240 9.400 12.120 17.580
PRIL-M
τ3(N3)30.229 16.879 60.340 63.820 67.200 69.940
τ2(N2)30.446 16.930 60.200 62.640 66.080 68.600
τ1(N1)4.282 2.337 11.160 14.620 17.860 21.660
All 16.134 17.365 59.000 62.280 65.280 69.940
(i.e., Tmin) due to packets from N1, and the generation times
on N2and N3are not synchronized with N1, the time they
have to wait in the transmission queue before the link switches
to ON is, on average, Tmin
2.
Overall power consumption (for the whole network) and
mean latency (experienced by packets of source nodes N1,
N2, and N3) are graphically sketched in the two diagrams of
Fig. 7, which permit to quickly compare the performance of
TSCH, PRIL-F, and PRIL-M.
To set a trade-off between power consumption and latency
in PRIL-M, an upper limit for the time a cell can remain
disabled could be possibly enforced. Because this aspect is
outside the scope of this paper, it is left as future work.
B. Realistic network configurations
Two larger and more realistic network configurations were
additionally analyzed, we term star (Fig. 6.b) and deep
(Fig. 6.c) topologies, whose complexity resembles real WSNs.
Again, flows are directed from leaves to the root node N0.
The star topology consists of 3layers (root excluded) and
interconnects 16 leaves (sensors), 6intermediate nodes (relay
only) at a distance of one hop from leaves and 3nodes at two
hops, plus the root N0(sink), for a total of 26 nodes. Two
different periods were defined for flows, namely, ∼10 min
and ∼30 min. Again, actual periods were defined in slot times
as coprime numbers. Mapping between leaves and periods
is shown in the figure. This topology represents a typical
example of a WSN in which nodes are deployed in such a
way to provide complete coverage of a certain area and the
sink (gateway) is positioned near the center of this area.
This article has been accepted for publication in IEEE Internet of Things Journal. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2024.3406646
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
a) Simple topology b) Star topology c) Deep topology
Fig. 6. Simple, star (3 layers), and deep (6 layers) network configurations used for simulations: the generation period (in slot times) is reported between
parentheses.
The deep topology of Fig. 6.c is also quite common in
the real world, especially when the area to be covered has
an elongated or irregular shape. In this case, the maximum
number of hops from leaves to the sink may be sensibly higher
than the star topology, if networks with the same number of
nodes are considered. In the example we considered, which
consists of 6layers, the maximum number of children for
every node is two and the total number of nodes is 29. The
generation period was set equal to about 1 min for all the 8
leaf nodes (actual periods of flows are similar but coprime).
Only for the experiment described in Section V-C, an
additional node NSwas also inserted in the deep network
topology to analyze the behavior of PRIL-M in the presence
of sporadic traffic flows.
Table III about power consumption shows both aggregate
results (distinguishing among leaves at the lowest network
layer and the two layers of relays located one and two hops
away from leaves, respectively) and results related to some
specific individual nodes belonging to the different layers (N0,
N23,N21 ,N17, and N1). Regarding the entire network (row
in the table labeled “All”), the effect of PRIL-M is clear:
compared to PRIL-F the total power consumption decreases to
46.4% (from 2140.2µWto 993.71 µW) for the star topology
and to 34.3% (from 3941.5µWto 1350.2µW) for the deep
topology. When using PRIL-M, the contribution Plisten to
power consumption (which is related to the idle listening
phenomenon) is practically canceled. This is not true for
PRIL-F: in this case, when the distance of the considered node
from the source of the packet flow (number of hops) is greater
0
100
200
300
400
500
600
700
TSCH PRIL-F PRIL-M
0
2
4
6
8
10
12
14
16
18
TSCH PRIL-F PRIL-M
[µW]
P
Plisten
(a) Power consumption
[s]
P
(b) Latency µd
Fig. 7. Comparison among the proposed PRIL strategies (simple topology).
TABLE III
COMPARISON AMONG TSCH AND TH E PRO POS ED PRIL S TRATE GI ES
(STAR A ND D EEP T OPO LO GIE S): POW ER CO NSU MP TIO N.
b) Star topology (3 layers)
TSCH PRIL-F PRIL-M
Node Hops Plisten P Plisten P Plisten P
[µW] [µW] [µW]
N03 375.15 536.77 375.15 536.76 .13 159.07
N23 2 275.18 369.32 275.18 369.28 .13 101.01
N21 1 575.48 669.60 .0012 89.83 .0012 98.60
N17 1 287.73 334.82 .00058 44.93 .00058 53.66
N10 0 10.07 0 18.90 0 18.84
Relays 2 825.57 1107.8 825.57 1107.8 .19 299.71
Relays 1 2327.1 2609.3 .0069 269.36 .0069 308.79
Leaves 0 0 120.63 0 226.2 0 226.14
All 3527.8 4374.6 1200.7 2140.2 0.33 993.71
c) Deep topology (6 layers)
TSCH PRIL-F PRIL-M
Node Hops Plisten P Plisten P Plisten P
[µW] [µW] [µW]
N06 250.17 357.77 250.16 357.78 .53 105.97
N27 5 125.04 219.16 125.05 219.15 .064 101.87
N25 4 275.19 369.30 275.18 369.29 0.78 101.52
N21 3 137.59 184.66 137.59 184.66 .12 55.04
N17 2 287.73 334.82 287.73 334.82 .98 54.73
N91 143.87 167.42 .00029 22.48 .00029 31.18
N10 0 10.06 0 18.87 0 18.79
All 3903.3 5030.7 2752.3 3941.5 7.11 1350.2
than one, Plisten is the same as standard TSCH.
Regarding latency, Table IV shows that PRIL-M adoption
makes performance worse than conventional TSCH. As ex-
pected, there is a consistent increase of the average latency µd
and, more in general, of all the statistical indicators related to
latency. Besides statistics evaluated over the entire set of flows,
the table also reports results related to the two flows character-
ized by the minimum and maximum average latency, labeled
“Min” and “Max”, respectively. As can be seen, performance
indicators about latency experienced large variations. In Fig. 8
the very same results are depicted graphically as diagrams.
One of the problems of PRIL-M is that the packet with the
minimum period originated by Nref deactivates the receiving
This article has been accepted for publication in IEEE Internet of Things Journal. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2024.3406646
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
TSCH PRIL-F PRIL-M
0
10
20
30
40
50
TSCH PRIL-F PRIL-M
0
1
2
3
4
5
6
TSCH PRIL-F PRIL-M
0
10
20
30
40
50
TSCH PRIL-F PRIL-M
[mW]
P
Plisten
Star topology
(a) Power consumption
[s]
P
(b) Latency µd
[mW]
P
Plisten
Deep topology
(c) Power consumption
[s]
P
Deep topology (d) Latency µd
Fig. 8. Comparison among the proposed PRIL strategies (star and deep topologies): responsiveness is traded for power saving in PRIL-M.
TABLE IV
COMPARISON AMONG TSCH AND TH E PRO POS ED PRIL S TRATE GI ES
(STAR AND DEEP TOPOLOGIES): END -TO-E ND LAT ENC Y.
b) Star topology (3 layers)
Flow µdσddp99 dp99.9dp99.99 dmax
[s]
TSCH All 2.503 1.872 8.760 12.660 16.960 32.320
PRIL-F All 2.503 1.872 8.780 12.680 17.120 32.900
PRIL-M
Min 12.493 3.278 21.340 25.380 29.020 41.080
Max 114.79 50.05 215.88 232.10 240.12 246.50
All 45.392 34.271 158.22 205.80 227.72 246.50
c) Deep topology (6 layers)
Flow µdσddp99 dp99.9dp99.99 dmax
[s]
TSCH All 3.543 2.352 10.800 14.500 18.240 28.140
PRIL-F All 3.544 2.354 10.840 14.540 18.320 27.400
PRIL-M
Min 24.102 4.559 35.740 40.500 45.040 55.020
Max 70.374 24.551 121.66 130.12 136.10 147.08
All 47.658 23.847 112.96 125.18 132.40 147.08
Sporadic traffic
TSCH NS3.925 2.663 12.340 16.160 20.500 21.820
PRIL-M NS98.992 29.575 164.40 176.86 184.16 184.78
node (OFF state) even if the children nodes of the transmitting
node still have packets to be sent. This phenomenon causes
an increase in the latency on the other traffic flows, but also
on the subsequent transmissions of packets related to the flow
associated with Nref . This issue poses new research challenges
for improving PRIL-M.
Above set of experiments confirms that PRIL-M lowers en-
ergy consumption at the expense of latency and determinism.
This implies that it can be profitably applied to real-world
applications for which short latency is irrelevant, e.g., for
monitoring purposes (periodic logging of physical quantities
where timestamping is carried out by sensors). Conversely,
PRIL-F is a valid option in all those contexts with real-time
requirements, although energy saving is suboptimal.
C. Traffic characteristics
Additional experiments were finally carried out to analyze
how the characteristics of the traffic impact on performance.
In all cases the deep topology in Fig 6.c was considered.
In the first set of experiments the generation period Tapp
of leaf nodes was varied from 20 s to 5 minutes with a step
of 20 s (always using coprime periods). The topmost plot
in Fig. 9 shows the power consumption versus Tapp. As
expected, PRIL-M exhibits tangibly lower energy consumption
than both PRIL-F and TSCH. When the generation period
is shortened, the behavior of the three approaches becomes
similar. In fact, the fraction of cells allocated to the link that
are actually exploited for frame transmissions progressively
increases, consequently decreasing the effects of idle listening.
At the limit, when no cells remain unused (which means that
the link operates in saturation conditions), idle listening cannot
take place, irrespective of PRIL usage. On the other hand,
when the period grows larger the power consumption of every
approach converges to a specific limit value, and differences
among them become more or less constant. Reasoning again
at the limit, when Tapp is enlarged consistently the majority of
the allocated cells remain unused, which implies that the most
part of energy is wasted by idle listening. Hence, PRIL ability
to lessen this phenomenon (especially the multi-hop version)
plays a crucial role in determining power consumption.
Concerning the average latency (lower part of Fig. 9), in
PRIL-M it increases linearly with the generation period Tapp.
This is an expected behavior: in fact, the time for which the
outgoing link in every intermediate relay node remains OFF
is equal to the minimum generation period Tmin it observes
among the incoming flows. In this experiment, all periods
Tapp were set (about) equal, therefore Tmin (and the sleep
time Tend) are approximately the same for the different relays.
0
2000
4000
6000
P [ W]
TSCH
PRIL-F
PRIL-M
012345
Tapp
[min]
0
50
100
150
d
[
s
]
012345
3.0
3.5
4.0 TSCH and PRIL-F Zoom
TSCH
PRIL-F
PRIL-M
Fig. 9. Effect of the variation of the generation periods (deep topology).
This article has been accepted for publication in IEEE Internet of Things Journal. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2024.3406646
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
As explained in Section III-D, for any given relay the frames
of flows other than the one originating from Nref may find
the link OFF. In this case, they will wait for the arrival
of a packet from Nref , after which all the queued frames
will be orderly sent as a batch. When frame arrivals are not
aligned with forwarding opportunities (this can be determined
by looking at the network topology, see red arrows in Fig. 6.c),
the forwarding delay introduced by every relay node is, on
average, Tmin/2. As a consequence, the network-wide mean
latency is proportional to Tapp.
As far as PRIL-F and TSCH are concerned (see the zoomed
in portion included in Fig. 9), they have the same latency,
which decreases as the period increases and tends asymptot-
ically to a limit value because packet queuing phenomena
within relays become progressively less likely. Conversely,
when the period gets shorter, the probability for a packet to
arrive to a relay when other packets are also enqueued grows
higher. This causes additional queuing delays, which explains
the reason why, when Tapp is small, mean latency increases up
to several times Tsfr (remaining nevertheless below PRIL-M).
A second set of experiments was performed to assess the
effect of PRIL-M on best effort sporadic traffic, whose results
are reported in the lower part of Table III. A new node NS
(shown with a dashed red circle) was added to this purpose
as a leaf to the network in Fig 6.c, and it was programmed to
generate sporadic traffic where packet intertimes are selected
randomly according to an exponential distribution with mean
value 1 hour. As expected, packets generated by NSsuffer
in PRIL-M from the largest latency increase with respect
to TSCH, because they are delayed more in every hop of
their path (the PRIL-M state machines running in relay nodes
privilege the flow with the minimum period, which never
coincides with the sporadic flow).
VI. CONCLUSIONS
Time-slotted networks like TSCH may benefit from the
adoption of the PRIL-F and PRIL-M techniques, which exploit
traffic periodicity to reduce power consumption dramatically
(periodic traffic patterns are commonplace in WSNs deployed
at the perception layer in the IoT pyramid). In particular, in all
the network configurations we analyzed in this work, the joint
use of PRIL-M and PRIL-F decreased power consumption to
less than 25% compared to standard TSCH, and to less than
50% compared to PRIL-F alone.
This remarkable achievement permits PRIL-M to be prof-
itably adopted whenever IoT solutions based on WSNs have to
be set up according to the deploy-and-forget paradigm. In these
cases, reducing energy consumption to the bare minimum
is a prerequisite to reduce maintenance activities involved
in battery replacement, hence decreasing the total cost of
ownership. Another non-negligible advantage of PRIL-M is
that, it reduces the overall amount of exhausted batteries that
need to be disposed, which makes these solutions greener.
The only disadvantage of PRIL-M is a non-negligible in-
crease of end-to-end latency, which suggests that it is un-
suitable for applications with real-time constraints. In these
cases, PRIL-F should be employed alone. For this reason, a
consistent part of our future activities on the subject will be
devoted to reducing the impact of PRIL-M on communication
latency as much as possible, for instance by limiting the max-
imum time for which scheduled slots can remain in the OFF
state. Other potential research activities include the extension
of PRIL techniques to similar communication technologies for
WSNs, such as DSME.
REFERENCES
[1] I. Chukwuemeka Chimsom and M. K. Habib, “Design of a Two-
Tier WSN-based IoT Surveillance System with Cloud Integration,” in
International Conference on Research and Education in Mechatronics
(REM), May 2019, pp. 1–7.
[2] M. Erdelj, M. Kr´
ol, and E. Natalizio, “Wireless Sensor Networks
and Multi-UAV systems for natural disaster management,” Computer
Networks, vol. 124, pp. 72–86, 2017.
[3] S. Scanzio, L. Wisniewski, and P. Gaj, “Heterogeneous and dependable
networks in industry – A survey,” Computers in Industry, vol. 125, p.
103388, 2021.
[4] X. Yu, P. Wu, W. Han, and Z. Zhang, “A survey on wireless sensor net-
work infrastructure for agriculture,” Computer Standards & Interfaces,
vol. 35, no. 1, pp. 59–64, 2013.
[5] J. M. Batalla, G. Mastorakis, C. X. Mavromoustakis, and J. Zurek, “On
cohabitating networking technologies with common wireless access for
home automation system purposes,” IEEE Wireless Communications,
vol. 23, no. 5, pp. 76–83, 2016.
[6] IEEE, “IEEE Standard for Low-Rate Wireless Networks,” IEEE Std
802.15.4-2015 (Rev. of IEEE Std 802.15.4-2011), pp. 1–709, Apr 2016.
[7] I. A. Mantilla Gonzalez, F. Meyer, and V. Turau, “A
Comprehensive Performance Comparison of IEEE 802.15.4 DSME
and TSCH in a Realistic IoT Scenario for Industrial Applications,”
ACM Trans. Internet Things, vol. 4, no. 3, jun 2023. [Online].
Available: https://doi.org/10.1145/3595188
[8] X. Vilajosana, Q. Wang, F. Chraim, T. Watteyne, T. Chang, and K. S. J.
Pister, “A Realistic Energy Consumption Model for TSCH Networks,”
IEEE Sensors Journal, vol. 14, no. 2, pp. 482–489, 2014.
[9] S. Scanzio, G. Cena, and A. Valenzano, “Enhanced Energy-Saving
Mechanisms in TSCH Networks for the IIoT: The PRIL Approach,”
IEEE Transactions on Industrial Informatics, pp. 1–11, 2022.
[10] M. Mongelli and S. Scanzio, “A neural approach to synchronization
in wireless networks with heterogeneous sources of noise,” Ad Hoc
Networks, vol. 49, pp. 1–16, 2016.
[11] G. Cena, S. Scanzio, M. G. Vakili, C. G. Demartini, and A. Valenzano,
“Assessing the Effectiveness of Channel Hopping in IEEE 802.15.4
TSCH Networks,” IEEE Open Journal of the Industrial Electronics
Society, vol. 4, pp. 214–229, 2023.
[12] S. Scanzio, M. G. Vakili, G. Cena, C. G. Demartini, B. Montrucchio,
A. Valenzano, and C. Zunino, “Wireless Sensor Networks and TSCH:
A Compromise Between Reliability, Power Consumption, and Latency,”
IEEE Access, vol. 8, pp. 167 042–167 058, 2020.
[13] M. Mohamadi, B. Djamaa, and M. R. Senouci, “RAST: Rapid and
energy-efficient network formation in TSCH-based Industrial Internet
of Things,” Computer Communications, vol. 183, pp. 1–18, 2022.
[14] D. Z. Fawwaz and S.-H. Chung, “Adaptive Trickle Timer for Efficient
6TiSCH Network Formation Using Q-Learning,” IEEE Access, vol. 11,
pp. 37 931–37 943, 2023.
[15] S. Afshari, M. Nassiri, and R. Mohammadi, “SA-RPL: a scheduling-
aware forwarding mechanism in RPL/TSCH-operated networks,”
International Journal of Ad Hoc and Ubiquitous Computing, vol. 34,
no. 1, pp. 35–44, 2020.
[16] M. A. Sordi, O. K. Rayel, G. L. Moritz, and J. L. Rebelatto, “Towards
Improving TSCH Energy Efficiency: An Analytical Approach to a
Practical Implementation,” Sensors, vol. 20, no. 21, 2020.
[17] A. Mavromatis, G. Z. Papadopoulos, A. Elsts, N. Montavont,
R. Piechocki, T. Tryfonas, G. Oikonomou, and X. Fafoutis, “Adaptive
Guard Time for Energy-Efficient IEEE 802.15.4 TSCH Networks,” in
Wired/Wireless Internet Communications, M. Di Felice, E. Natalizio,
R. Bruno, and A. Kassler, Eds. Cham: Springer International Publish-
ing, 2019, pp. 15–26.
[18] J. Vera-P´
erez, D. Todol´
ı-Ferrandis, J. Silvestre-Blanes, and V. Sempere-
Pay´
a, “Bell-X, An Opportunistic Time Synchronization Mechanism for
Scheduled Wireless Sensor Networks,” Sensors, vol. 19, no. 19, 2019.
This article has been accepted for publication in IEEE Internet of Things Journal. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2024.3406646
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
[19] H. Farag, P. ¨
Osterberg, M. Gidlund, and S. Han, “RMA-RP: A Reliable
Mobility-Aware Routing Protocol for Industrial IoT Networks,” in IEEE
Global Conference on Internet of Things (GCIoT 2019), 2019, pp. 1–6.
[20] S. Murali and A. Jamalipour, “Mobility-Aware Energy-Efficient Parent
Selection Algorithm for Low Power and Lossy Networks,” IEEE Internet
of Things Journal, vol. 6, no. 2, pp. 2593–2601, 2019.
[21] Y. Shin and S. Seol, “Improvement of Power Consumption in RPL-
based Networks for Mobility Environment,” in International Conference
on Electronics, Information, and Communication (ICEIC 2020), 2020,
pp. 1–3.
[22] V. Kotsiou, G. Z. Papadopoulos, P. Chatzimisios, and F. Theoleyre, “LA-
BeL: Link-Based Adaptive BLacklisting Technique for 6TiSCH Wireless
Industrial Networks,” in Proceedings of the 20th ACM International
Conference on Modelling, Analysis and Simulation of Wireless and
Mobile Systems, ser. MSWiM. New York, NY, USA: Association for
Computing Machinery, 2017, pp. 25–33.
[23] G. Cena, S. Scanzio, and A. Valenzano, “Ultra-Low Power Wireless
Sensor Networks Based on Time Slotted Channel Hopping with
Probabilistic Blacklisting,” Electronics, vol. 11, no. 3, p. 304, Jan 2022.
[Online]. Available: http://dx.doi.org/10.3390/electronics11030304
[24] W. Jerbi, O. Cheickhrouhou, A. Guermazi, and H. Trabelsi, “MSU-
TSCH: A Mobile scheduling updated algorithm for TSCH in the internet
of things,” IEEE Transactions on Industrial Informatics, pp. 1–8, 2022.
[25] N. T. Javan, M. Sabaei, and V. Hakami, “Adaptive Channel Hopping for
IEEE 802.15.4 TSCH-Based Networks: A Dynamic Bernoulli Bandit
Approach,” IEEE Sensors Journal, vol. 21, no. 20, pp. 23667–23 681,
2021.
[26] N. Accettura, E. Vogli, M. R. Palattella, L. A. Grieco, G. Boggia,
and M. Dohler, “Decentralized Traffic Aware Scheduling in 6TiSCH
Networks: Design and Experimental Evaluation,” IEEE Internet of
Things Journal, vol. 2, no. 6, pp. 455–470, 2015.
[27] M. Ojo, S. Giordano, G. Portaluri, D. Adami, and M. Pagano, “An
energy efficient centralized scheduling scheme in TSCH networks,” in
IEEE International Conference on Communications Workshops (ICC),
2017, pp. 570–575.
[28] T. Matsui and H. Nishi, “Time slotted channel hopping scheduling based
on the energy consumption of wireless sensor networks,” in IEEE 15th
International Workshop on Advanced Motion Control (AMC 2018),
2018, pp. 605–610.
[29] T. van der Lee, G. Exarchakos, and S. H. de Groot, “Swarm-Based
Energy Efficient Scheduling for Wireless Sensor Networks,” in IEEE
Conference on Standards for Communications and Networking (CSCN),
2019, pp. 1–6.
[30] B. ¨
Ozceylan, B. ¨
Unl¨
u, and B. Baykal, “An energy efficient optimum
shared cell scheduling for TSCH networks,” in IEEE 13th International
Conference on Wireless and Mobile Computing, Networking and
Communications (WiMob 2017), 2017, pp. 1–8.
[31] M. G. Gait´
an, D. Dujovne, J. Zu˜
niga, A. Figueroa, and L. Almeida,
“Multi-Gateway Designation for Real-Time TSCH Networks using
Spectral Clustering and Centrality,” IEEE Embedded Systems Letters,
pp. 1–1, 2022.
[32] F. Veisi, J. Montavont, and F. Theoleyre, “Enabling Centralized Schedul-
ing Using Software Defined Networking in Industrial Wireless Sensor
Networks,” IEEE Internet of Things Journal, pp. 1–1, 2023.
[33] S. Scanzio, G. Cena, A. Valenzano, and C. Zunino, “Energy Saving in
TSCH Networks by Means of Proactive Reduction of Idle Listening,”
in Ad-Hoc, Mobile, and Wireless Networks, L. A. Grieco, G. Boggia,
G. Piro, Y. Jararweh, and C. Campolo, Eds. Cham: Springer Interna-
tional Publishing, 2020, pp. 131–144.
[34] C. Vallati, S. Brienza, G. Anastasi, and S. K. Das,
“Improving Network Formation in 6TiSCH Networks,”
IEEE Transactions on Mobile Computing, vol. 18, no. 1, pp. 98–
110, 2019.
[35] E. Municio, G. Daneels, M. Vuˇ
cini´
c, S. Latr´
e, J. Famaey, Y. Tanaka,
K. Brun, K. Muraoka, X. Vilajosana, and T. Watteyne, “Simulating
6TiSCH networks,” Transactions on Emerging Telecommunications
Technologies, vol. 30, no. 3, p. e3494, 2019.
[36] A. Elsts, “TSCH-Sim: Scaling Up Simulations of TSCH and 6TiSCH
Networks,” Sensors, vol. 20, no. 19, 2020.
This article has been accepted for publication in IEEE Internet of Things Journal. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2024.3406646
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
