Experimental Evaluation of Techniques to Lower Spectrum Consumption in Wi-Red

Abstract

Seamless redundancy layered atop Wi-Fi has been shown able to tangibly increase communication quality, hence offering industry-grade reliability. However, it also implies much higher network traffic, which is often unbearable as the wireless spectrum is a shared and scarce resource. To deal with this drawback the Wi-Red proposal includes suitable duplication avoidance mechanisms, which reduce spectrum consumption by preventing transmission on air of inessential frame duplicates. In this paper, the ability of such mechanisms to save wireless bandwidth is experimentally evaluated. To this purpose, specific post-analysis techniques have been defined, which permit to carry out such an assessment on a simple testbed that relies on plain redundancy and do not require any changes to the adapters' firmware. As results show, spectrum consumption decreases noticeably without communication quality is impaired. Further saving can be obtained if a slight worsening is tolerated for latencies.

Full text

1
Experimental Evaluation of Techniques to Lower
Spectrum Consumption in Wi-Red
Gianluca Cena, Senior Member,IEEE, Stefano Scanzio, Member,IEEE, and
Adriano Valenzano, Senior Member,IEEE
Abstract—Seamless redundancy layered atop Wi-Fi has been
shown able to tangibly increase communication quality, hence
offering industry-grade reliability. However, it also implies much
higher network traffic, which is often unbearable as the wireless
spectrum is a shared and scarce resource. To deal with this
drawback the Wi-Red proposal includes suitable duplication
avoidance mechanisms, which reduce spectrum consumption by
preventing transmission on air of inessential frame duplicates.
In this paper, the ability of such mechanisms to save wireless
bandwidth is experimentally evaluated. To this purpose, specific
post-analysis techniques have been defined, which permit to carry
out such an assessment on a simple testbed that relies on plain
redundancy and do not require any changes to the adapters’
firmware. As results show, spectrum consumption decreases
noticeably without communication quality is impaired. Further
saving can be obtained if a slight worsening is tolerated for
latencies.
Index Terms—IEEE 802.11, Wi-Fi, seamless redundancy, PRP,
experimental evaluation, communication efficiency.
I. INTRODUCTION
Wireless communications are currently employed in many
different contexts. Besides personal connectivity and of-
fice/home environments, they are becoming increasingly pop-
ular also in other scenarios, like intelligent transportation
systems [1], Internet of Things (IoT) [2], environmental moni-
toring [3], and precision agriculture [4], to cite a few. Although
wireless networks have the potential to bring tangible benefits
also to time-sensitive applications, they are quite unreliable
and scarcely deterministic, and so are often deemed unsuitable
in contexts like factory automation. As a matter of fact, with
the exception of WirelessHART and ISA100.11a [5], real-time
distributed control systems based on wireless technologies are
currently quite unusual in industrial plants.
Unlike multimedia applications, where data can typically
be buffered on both sides of a network connection, control
applications (relying on, e.g., the request-response or producer-
consumer data exchange paradigms [6]) demand that every sin-
gle piece of information is delivered timely and reliably. This
can be hardly achieved with conventional wireless solutions,
since they are typically based on random access mechanisms,
which unavoidably lead to collisions. Behavior can be im-
proved by using deterministic access schemes, which prevent
This work was partially supported by Regione Piemonte and the Ministry
of Education, University, and Research of Italy in the POR FESR 2014/2020
framework, Call “Piattaforma tecnologica Fabbrica Intelligente”, Project “Hu-
man centered Manufacturing Systems” (application number 312-36). The
authors are with the National Research Council of Italy, Istituto di Elettronica
e di Ingegneria dell’Informazione e delle Telecomunicazioni (CNR-IEIIT),
I-10129 Torino, Italy (e-mail: {name.surname}@ieiit.cnr.it).
intra-network interference, e.g., Time Slotted Channel Hop-
ping (TSCH) for IEEE 802.15.4 [7] and Hybrid-coordination-
function Controlled Channel Access (HCCA) for IEEE 802.11
[8], not counting the many proposals found in literature, for
instance [9][10][11][12]. It is worth noting that most of these
mechanisms are unable to face electromagnetic disturbance
(including multipath fading effects) and, in several cases, not
even interference of nearby non-compliant wireless devices.
These phenomena are unpredictable, and can be effectively
counteracted by exploiting diversity, in time, frequency, space,
and so on. For instance, automatic retransmission mechanisms
are customarily included in any wireless protocol, frequency
hopping is adopted by TSCH and Bluetooth, while multiple
spatial streams are exploited in recent IEEE 802.11 versions.
In the following, we consider the use of IEEE 802.11,
also known as Wi-Fi, for connecting devices in time-sensitive
control applications. The reasons of this choice are threefold:
1) Wi-Fi is extremely popular, and consequently not expensive,
2) it features very high throughput, and 3) it achieves com-
plete interoperability with Ethernet. We specifically focused
on mechanisms based on time and frequency diversity (i.e.,
retransmissions and channel redundancy), both of which bring
improvements on communication quality at the cost of higher
spectrum consumption. Techniques based on message schedul-
ing, like HCCA or the like, were not taken into account as, to
the best of our knowledge, there is currently little availability
of commercial equipment offering proper support.
Seamless redundancy, as defined by the Parallel Redun-
dancy Protocol (PRP) [13], is meant to increase availability
of real-time Ethernet networks. Nonetheless, it can be also
used to improve communication quality of wireless links. In
[14], an arrangement was proposed, we denote PRP over
Wi-Fi (PoW), which layers end-to-end seamless redundancy
as per PRP atop IEEE 802.11, by using commercial Wi-Fi
equipment, like access points (AP) and wireless adapters
(WA), as well as specialized PRP devices (RedBoxes) [15].
Behavior, in terms of the fraction of frames delivered to
destination correctly and timely, was experimentally shown to
improve substantially. Similar approaches were proposed for
streaming phasor measurements in smart grids [16].
While relying on the same principles as PoW, the Wi-Fi
Redundancy (Wi-Red) approach [17] is a link-level solution
where PRP and the IEEE 802.11 Medium Access Control
(MAC) mechanism are intertwined in order to improve per-
formance further. In particular, specific duplication avoidance
(DA) mechanisms are introduced to prevent transmission on
air of identical copies of the same packet, when doing so
This is the author’s version of an article that has been published in this journal.
Changes were made to this version by the publisher prior to publication.
The final version of record is available at https://doi.org/10.1109/TWC.2018.2884914
Copyright (c) 2018 IEEE. Personal use is permitted.
For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
arXiv:2211.12198v1 [cs.NI] 22 Nov 2022

2
is useless, so as to reduce network load. Basically, Wi-Red
can be seen as an holistic approach aimed at optimizing
time and frequency diversity in Wi-Fi. In theory, the same
approach could be also used with wireless technologies other
than Wi-Fi, like those employed in wireless sensor networks
(WSN). However, this is typically pointless, as WSNs are
mainly aimed at ensuring low power consumption, and not
high performance. For example, solutions like TSCH offer
time-frequency diversity through channel hopping, without the
need to have a dual radio block, at the expense of larger
transmission latencies.
This paper, which grounds on the preliminary work pre-
sented in [18], experimentally evaluates the effects of DA
mechanisms on the overall spectrum consumption (combined
traffic on all physical channels) of a redundant link between
two wireless stations, and analyzes the trade-off between band-
width saving and communication quality for a basic proactive
heuristic. With respect to [18]: 1) communication efficiency
is evaluated considering the actual number of transmission
attempts on air, inclusive of retries; 2) runs include more
than twice samples and employ more realistic interfering
traffic patterns; 3) assumptions on virtual analysis of proactive
approaches are validated; and 4) both frame losses and more-
than-duplex redundancy are considered. The paper is organized
as follows: Section II summarizes Wi-Fi seamless redundancy
basics, while Section III describes our experimental testbed.
Quantitative evaluation of the ability of DA mechanisms to
reduce spectrum consumption, based on measurements carried
out on the testbed, is reported in Sections IV and V, which
concern two types of approaches, that are reactive and proac-
tive ones, respectively. Finally, some conclusions are drawn in
Section VI.
II. LINK-L EV EL WI-FISEAMLESS REDUNDANCY
In its simplest embodiment, Wi-Red corresponds to a thin
layer, located just above Wi-Fi adapters and implemented
by bringing software modifications to their device drivers.
However, to unleash all its potential, adapters’ firmware has
likely to be rewritten in part.
A. Redundant Network Architecture
An IEEE 802.11 Basic Service Set (BSS) is a group of
wireless stations (STA) located in the same place, operating
on the same channel, and enabled to exchange data among
each other. Similarly, a redundant BSS (RBSS) is defined as a
set of co-located redundant STAs (RSTA) communicating on
multiple physical channels according to seamless redundancy
principles. An RSTA comprises two or more sub-STAs, operat-
ing on distinct channels, and one link redundancy entity (LRE),
which coordinates their operations so as to achieve seamless
redundancy. Each sub-STA includes both a MAC and a radio
block, and behaves mostly the same as a conventional STA.
More details can be found in [17].
Generally speaking, effectiveness of seamless redundancy in
networks including multiple uncoordinated RSTAs is typically
not as good as on point-to-point links. In fact, intra-network
interference they generate on physical channels is clearly not
independent. The easiest way to face this situation is to exploit
for RSTAs a deterministic protocol overlay (e.g., HCCA
or, possibly, solutions specifically engineered for seamless
redundancy), which prevents them from transmitting on air
at the same time. This is not unfeasible, as these devices are
usually part of the same control system. Actual advantages
of redundancy in presence of many nodes operating with
random access is still an open issue, and is left as future work.
For such reason, the following analysis takes into account a
unidirectional redundant link, denoted L, between two RSTAs
(originating RSTA and recipient RSTA), and CL={A, B, ...}
is the set of the related physical channels. To keep the testbed
simple, only duplex redundancy has been considered in the
experiments, in which case Lis denoted A+B. We assume
that fragmentation, aggregation, and block acknowledgment
mechanisms are not exploited, since they are not meant to
meet the requirements of time-sensitive control applications.
B. Seamless Redundancy Basics
Although seamless redundancy also applies to unconfirmed
transmissions, this paper only focuses on acknowledged ex-
changes. Besides being practically mandatory to support reli-
able data transfers in Wi-Fi, the latter also enable effective
optimizations when paired with seamless redundancy. For
our purposes, the term packet denotes a unit of data to be
exchanged over the network, as seen above the redundancy
and data-link layers, and approximately corresponds to a MAC
Service Data Unit (MSDU). When a request is made in the
originating RSTA to transmit a packet over a redundant link
L, the related LRE issues concurrent requests on the adapters
of its sub-STAs for sending identical copies of the packet
on every channel C∈ CL. Sub-STAs’ operations are mostly
disjoint [19], [20], and the transmission of each packet copy
on the related adapter obeys MAC rules. This involves sending
specific frames on air, corresponding to MAC Protocol Data
Units (MPDU), as well as dealing with carrier sensing, in-
terframe spaces, and random backoff, issuing retransmissions
upon missing acknowledgment, and so on.
Transmission of each packet copy on the related channel
consists of one or more attempts: exactly one initial try plus,
upon unsuccessful delivery, a variable number of retries. To
avoid starvation, a retry limit is defined to bound the number
of transmission attempts for any packet copy. Each attempt
is made up of a DATA frame (carrying user data), sent
by the originator, immediately followed by an ACK frame
(acknowledgment), returned by the unique recipient after a
Short Interframe Space (SIFS). Failed attempts are detected
by the MAC in the originating sub-STA by setting a specific
ACKtimeout whenever a DATA frame is sent. If the ACK
frame is not received before the timeout expires, the attempt
is considered as failed.
On the recipient RSTA, the LRE takes care of removing
duplicates, reordering packets, and delivering them to the
upper layers. To do so, packet copies received on the physi-
cal channels of the redundant link are paired using specific
sequence numbers added on the originating side. For any
packet, the DATA frame (either the initial try or a retry) that is

3
Recipient RSTA
Channel B
Channel A
Originating RSTA (PC)
WAA
WAB
Eth.
xiA(ACK)
miA
miA(DATA)
miB(DATA)
xiB(ACK)
miB
miB
miA
i
i
i
•Î•
mΕ
n…
Meas. task APA
APB
Switch
TM
For checking purposes only
Redundant link (A+B)
Fig. 1. Experimental PoW testbed.
correctly received first among all its copies is retained, while
the following ones are discarded.
C. Duplication Avoidance Mechanisms
In PoW, every packet is always sent on all channels. Since
channels can be assumed to be reasonably independent in
a well-configured RBSS [19], overall spectrum consumption
in PoW coincides with the sum of the traffics that would
be induced if non-redundant, conventional Wi-Fi were used
to send packets separately on each physical channel. The
main improvement Wi-Red [17] brings over PoW [14] are
duplication avoidance mechanisms. Basically, they are aimed
at reducing the amount of frames sent on air. This is a strict
requirement when smooth coexistence has to be ensured with
nearby networks, as the wireless spectrum is a scarce resource
not to be wasted, especially in the 2.4 GHz band and, to a
lesser (but increasing) extent, 5 GHz band.
Two kinds of DA mechanisms exist, namely reactive and
proactive, both of which operate on the originating side
of a redundant link. The former simply prevents duplicate
transmissions on air that are recognized to be useless, since
an acknowledgment has already been received from the re-
cipient, on any physical channel of the redundant link, for
the information they convey. Instead, the latter also includes
heuristics aimed at complementing the reactive approach, so
as to increase its effectiveness.
III. EXP ER IM EN TAL EVALUATION
The experimental setup used for assessing Wi-Red perfor-
mance resembles the one in [20]. However, the analysis carried
out in this paper does not focus on communication quality, but
rather on the ability of DA mechanisms to reduce spectrum
consumption (which is the main performance metric here). As
shown in Fig. 1, the testbed is made up of a PC running the
Linux OS (kernel v. 3.16.0) and provided with two identical
Wi-Fi adapters (dual-band TP-Link TL-WDN4800), which
emulates the behavior of the originating RSTA. Two identical
APs (NETGEAR WAC120), set to operate on completely
disjoint frequencies to avoid interference (channel 1in the
2.4 GHz band and 44 in the 5 GHz band), partially mimic the
behavior of the recipient RSTA. Each Wi-Fi adapter on the PC
is associated to a distinct AP, which, on correct reception of a
DATA frame, returns an ACK frame to the related adapter and
relays the packet to the PC Ethernet port through a Gigabit
switch.
TABLE I
SYMBOLS AND QUANTITIES USED TO CHARACTERIZE EXPERIMENTS.
Symbol Meaning
Cgeneric physical channel (either Aor B, in our testbed)
Lredundant link (A+B, in our testbed)
CLset of physical channels of L({A, B}, in our testbed)
TLRE LRE delay (from XACK to early termination)
TC
SIFS SIFS duration on channel C
TC
ACKto ACKtimeout duration on channel C
mii-th packet generated in the experiment
xifirst received ACK (XACK) for packet mi
Qiquickest channel, on which xiis received (either Aor B)
Qichannel(s) other than the quickest (either Bor A)
tX,i cross acknowledgment (XACK) time for packet mi
mC
icopy of packet misent on channel C
xC
iacknowledgment for packet copy mC
i
wC
inumber of transmission attempts of packet copy mC
i
tC
T,i transmission time of packet copy mC
i
tC
W,i starting time on air of the final attempt of packet copy mC
i
tC
R,i receive time of packet copy mC
i
tC
X,i acknowledgment time of packet copy mC
i
TC
D,i duration of the DATA frame in the final attempt of mC
i
TC
A,i duration of the ACK frame in the final attempt of mC
i
mC
i,ℓ ℓ-th transmission attempt of packet copy mC
i
tC
W,i,ℓ starting time on air of attempt mC
i,ℓ
TC
D,i,ℓ duration of the DATA frame in attempt mC
i,ℓ
TC
A,i,ℓ duration of the ACK frame in attempt mC
i,ℓ
On the PC, a measurement task produces a cyclic pattern
of packets. Generation period TMis set to 100 ms. To obtain
reliable statistics, each run lasted 24 hours and included
N= 864000 packets, which are sent on the redundant
link according to PoW rules (each packet is concurrently
fed to both Wi-Fi adapters). Timestamps are taken on every
packet transmission and reception. Since they refer to the
same time base, i.e., the CPU time stamp counter (TSC)
register of the PC, transmission latencies can be computed.
Following the approach in [12], receive times were not taken
on packet arrival to the Ethernet port. Instead, we modified
Wi-Fi adapters’ drivers so as to acquire a precise timestamp in
kernel space upon ACK frame reception or ACKtimeout expiry.
Doing so improves both accuracy and precision of measured
transmission latencies. In fact, although APs were, on average,
quite fast in forwarding packets, they also introduced non-
negligible jitters and delays. This is no surprise, since they
were not conceived for time-sensitive applications. Results
were checked against timestamps and losses obtained on the
Ethernet port.
A. Modeling Experimental Results
Each experimental run in the testbed is modeled as a
packet sequence M= (m1, ..., mN), where miis the i-th
packet generated by the measurement task. For every packet,
a transmission request is separately issued on all adapters.
The copy of misent on channel C∈ CL(either Aor B,
in our testbed) is denoted mC
i, and its transmission outcome
is described as a pair lC
i, dC
i. Boolean value lC
iis the loss
indication, either true if miwent lost on Cor false when
it was successfully delivered. In the following, Booleans are
treated as integers, i.e., true ≡1and false ≡0. The non-

4
negative value dC
iis the transmission latency of mion C,
measured from the time tC
T,i when the transmission request is
issued on the related adapter to the time tC
R,i when the packet
is correctly received by the recipient, that is, dC
i,tC
R,i −tC
T,i .
Latency is only defined for packets for which lC
i= 0, since
on the contrary tC
R,i does not exist.
Behavior of redundant link Lcan be analyzed starting from
its physical channels. Let tL
T,i be the time when packet miis
generated. With the exception of proactive mechanisms [17],
described in Section V, we can assume that, with very good
approximation, transmission times for all the copies of any
given packet coincide, that is tC
T,i ≈tL
T,i ,∀C∈ CL. Because
of delays introduced at the MAC layer by the shared medium
(due to, e.g., the channel sensed busy or retransmissions), the
same does not necessarily hold for receive times tC
R,i.
Links exploiting seamless redundancy satisfy two basic
properties: 1) a packet miis lost on L(i.e., it never arrives
to the recipient, not even after all allowed retries have been
carried out by all adapters) only if its copies are definitely lost
on every channel C∈ CL; 2) the receive time of any packet mi
correctly delivered on Lcorresponds to the minimum between
the receive times of its copies on all physical channels, i.e.,
tL
R,i = minC∈CL,lC
i=0 tC
R,i. This means that communication
quality on a redundant link is never worse than on any of its
physical channels.
B. Measurement Technique
From a practical viewpoint, two timestamps were acquired
for every packet mion each physical channel C:tC
T,i is taken
immediately before the transmission request is invoked, while
tC
X,i is taken in the device driver when transmission ends and
passed to the measurement task through a character device. In
case of successful delivery, tC
X,i is used to infer tC
R,i and hence
dC
i. Characterization of the redundant link Lis carried out a
posteriori by applying PRP rules, that is lL
i=QC∈CLlC
iand,
when lL
i= 0,dL
i,tL
R,i −tL
T,i ≈minC∈CL,lC
i=0 dC
i.
According to IEEE 802.11, retransmission of confirmed
frames is automatically performed by Wi-Fi adapters upon
missing acknowledgment. Transmission attempts for the same
packet on any channel C∈ CLare carried out consecutively,
spaced by random backoff periods. As soon as an ACK frame,
denoted xC
i, is correctly received on Cfor packet copy mC
i,
its transmission on that channel ends. Let mC
i,` denote the `-
th transmission attempt of packet mion C, where `= 1
indicates the initial try while 2≤`≤wC
iare retries (wC
iis
the overall number of attempts performed for mC
i). From the
originator viewpoint, attempts mC
i,` with ` < wC
iare certainly
unsuccessful. However, due to the retry limit, not necessarily
the final attempt (`=wC
i) succeeds. Attempts performed on
any physical channel Cin a run on the PoW testbed can be
modeled as a sequence





mC
1,1, ..., mC
1,wC
1
| {z }
packet m1
, ...,
initial
try
z}|{
mC
i,1,
1-st
retry
z}|{
mC
i,2, ...,
(`-1)-th
retry
z}|{
mC
i,` , ...,
final
attempt
z }| {
mC
i,wC
i,
| {z }
packet mi
...




(1)
where each sub-sequence refers to a specific packet and
includes all the related attempts.
The starting time on air of mC
i,`, denoted tC
W,i,`, corresponds
to the actual beginning of the `-th transmission attempt of mi
on C. Since carrier sensing is performed by Wi-Fi adapters
independently, starting times of the initial attempts for packet
mion the different channels not necessarily coincide. Even
more so this holds for retries. Attempts in (1) are globally
ordered according to their starting time on air, i.e., tC
W,i,` <
tC
W,i,`+1,1≤i≤N, 1≤` < wC
iand tC
W,i,wC
i
< tC
W,i+1,1,1≤
i<N. The overall duration of attempt mC
i,` includes three
contributions, TC
D,i,`,TC
SIFS, and TC
A,i,`, which correspond to
the DATA frame, SIFS, and ACK frame, respectively. Rate
adaptation techniques in the originator directly affect TC
D,i,`
and, indirectly, TC
A,i,` (by influencing the bit rate used by the
recipient to reply). Instead, TC
SIFS is fixed and only depends
on the specific Wi-Fi amendment, i.e., physical layer (PHY),
selected for channel C.
C. Assumptions on Channel Errors
Both DATA and ACK frames may be corrupted during
transmission on air. If an error affects the ACK frame, the
MAC on the recipient notifies the correct packet reception,
but additional retries are attempted by the originator, until it
either receives an ACK or exceeds the retry limit. Typically,
this causes the packet acknowledgment time on the originator
to be delayed to one of the following retries. If none of these
succeed, the packet is considered lost by the originator, hence
making its delivery state inconsistent with the recipient. This
issue is usually dealt with by recovery mechanisms in the
upper protocol layers.
In order to simplify the following analysis, we will assume
that errors always affect a transmission attempt as a whole, and
not the ACK frame alone. In the real world, this assumption
is typically verified because: 1) ACK frames are very short
and rely on robust modulation and coding schemes, and 2)
under the reasonable hypothesis that the BSS does not suffer
from the hidden node problem, they cannot incur in collisions
not involving also the related DATA frame. This means that,
when a DATA frame is correctly transferred, very likely the
related ACK frame also does. We checked this assumption a
posteriori on experimental samples, by comparing the set of
packets arrived on the Ethernet port against those for which an
ACK was received, and found that it actually holds in almost
all the cases. However, it is worth pointing out that the correct
behavior of both seamless redundancy and DA mechanisms
does not depend in any way on this assumption, which only
leads to slightly pessimistic results for communication quality.
Under the above hypothesis, the events raised upon ac-
knowledgment (on the originator) and reception (on the recip-
ient) of packet mion channel Crefer to the same (successful)
transmission attempt, which corresponds to mC
i,wC
i
.
D. Limitations of Commercial Equipment
Real Wi-Fi adapters, including those used in our testbed,
typically raise interrupts only at the end of packet transmission

5
and reception. Thus, it is not possible to obtain timestamps
tC
W,i,` for inner attempts, but only for the final ones, in
which case either an ACK frame has been correctly received
(success) or the ACKtimeout has expired and the retry limit
has been exceeded (failure). These conditions conclude the
transmission process for packet copy mC
i. To keep notation
simple, the starting time on air of the final attempt on Cis
denoted tC
W,i, while the end of transmission coincides with
tC
X,i. The durations of the related DATA and ACK frames
(indicated as TC
D,i and TC
A,i, respectively) can be obtained by
inspecting the driver.
In case of success (lC
i= 0) the starting time on air of the
final transmission attempt of mC
ican be calculated as
tC
W,i =tC
X,i −TC
D,i +TC
SIFS +TC
A,i(2)
and a reliable estimate for the receive time can be obtained as
tC
R,i =tC
X,i −TC
SIFS +TC
A,i,(3)
while in case of failure (lC
i= 1)
tC
W,i =tC
X,i −TC
D,i +TC
ACKto(4)
where TC
ACKto is the ACKtimeout duration on C.
For the purpose of the following analysis, transmission
of mion physical channel Cof the testbed is completely
described by a tuple τC
i,lC
i, tC
T,i , tC
X,i, wC
i, T C
D,i, T C
A,i.
Overall, every experimental run is modeled as T=
τA
i, τ B
ii=1...N .
IV. REACTIVE DUPLICATE AVOIDAN CE
Reactive Duplication Avoidance (RDA) is the basis on
which all DA mechanisms rely. Basically, the LRE of the
originating RSTA exploits information provided at runtime
by the MAC of its sub-STAs to prevent useless transmission
attempts on air. In particular, as soon as an ACK frame is
received for a certain packet on any sub-STA, a cross ACK
event (XACK) is generated in the LRE, which forces the
transmissions of the copies of that packet on all the other
sub-STAs to be canceled. Let TLRE be the time taken by the
LRE to force early termination. If RDA is implemented in
the adapters’ hardware/firmware, TLRE can be as low as a
few µs, whereas when it is performed in software (e.g., in the
drivers) TLRE is expected to be in the order of several tens
to hundreds µsand is noticeably less deterministic [21]. For
packets whose transmission has failed on every sub-STA (and
hence, on the redundant link), no XACK is generated, and
hence no optimization is possible by RDA.
Concerning a packet misuccessfully delivered on the re-
dundant link L(that is, for which an ACK frame is correctly
received on at least one of its physical channels), let Qibe the
quickest channel, i.e., the one on which a valid ACK frame
is received first (see the diagram at the bottom of Fig. 2). In
formulas, Qi= arg minC∈CL,lC
i=0 tC
X,i. That ACK, as well
as the XACK it triggers, are denoted xiand occur at time
tX,i =tQi
X,i. Channels other than Qiare collectively denoted
Qi, in formulas, Qi=CL\ {Qi}. For duplex redundancy (as
in our testbed) one such channel exists. With a slight abuse of
notation, it can be directly identified as Qi.
Fig. 2. Early termination on XACK events carried out by RDA.
As shown in the topmost diagram of Fig. 2 (case 1),
according to RDA rules the LRE can prevent the `-th trans-
mission attempt of mion C∈Qifrom starting provided
that XACK processing has completed before tC
W,i,`. On the
contrary (case 2), terminating the attempt midway has to be
avoided, since this forces an error condition in all the receiving
STAs (similarly to when a frame gets corrupted on air) and
worsens coexistence with nearby Wi-Fi devices. Moreover,
since frame duration is set in advance in the frame preamble
(encoded as data rate and length) and the channel is reserved
up to the end of the ACK frame by the Network Allocation
Vector (NAV), doing so would bring little benefits. In formulas,
attempt mC
i,` (and the following) can be prevented on any
channel C∈Qiwhen
tX,i +TLRE < tC
W,i,`.(5)
A. Performance Comparison between RDA and PoW
1) Communication Quality: When the RBSS is correctly
dimensioned, so that generated packets seldom suffer from
appreciable queuing delays in local buffers, RDA offers the
same performance as PoW. In fact, RDA acts after ACK
reception, i.e., after the related packet has been delivered to
the recipient. This means that, in absence of queuing, RDA
operations can not affect in any way neither latencies dL
i
nor packet losses lL
iof PoW. This is mostly the case of our
experiments, as packets produced by the measurement task
were intentionally spaced wide enough. When traffic increases
and the RBSS operates in saturated conditions, RDA sensibly
improves communication quality over PoW [17], because of
its ability to lower network load, which reduces collisions
and, hence, packet losses and mean latencies. Latencies also
shrink because transmission queues in the originating RSTA
are drained more quickly. Definitely, RDA adoption can never
worse communication quality.
2) Spectrum Consumption: Unless interfering traffic and
disturbance in the RBSS are very low (in which case redun-
dancy is useless), RDA typically offers tangible advantages
with respect to PoW, as our experiments indicate. The ability
to reduce wasted bandwidth improves further when network
traffic increases and packets are buffered, since in this case
they can also be removed from transmission queues. Actual

6
Fig. 3. Virtual RDA experiment vs. real PoW experiment.
improvements in real scenarios heavily depend on the ability
of the LRE to quickly terminate ongoing transmissions upon
ACK reception. As will be shown, delays introduced by, e.g.,
software LRE implementations, impair RDA performance. In
order to reliably assess spectrum consumption on a real RDA
testbed, modifications to the firmware of Wi-Fi adapters are
likely required, which is unfeasible with most commercial de-
vices (including ours). For this reason, we evaluated bandwidth
saving in RDA using a different approach.
B. Virtual RDA Analysis on PoW Logs
PoW behavior closely resembles basic Wi-Red, i.e., seam-
less redundancy with no DA mechanisms. Once started, trans-
mission of each packet copy is carried out independently
on the related channel as per IEEE 802.11 MAC rules.
Interestingly, experimental data acquired in a run from the
PoW testbed can be used to analyze RDA performance as
well. In particular, the amount of transmission attempts that
can be saved by early termination in the very same operating
conditions (including environmental ones, which are almost
impossible for us to replicate) can be inferred. This provides
a reliable estimate of the improvements RDA achieves over
PoW in terms of spectrum consumption, without the need to
actually implement RDA.
A prerequisite for the following analysis is that, terminating
a packet transmission in advance does not affect the following
ones. This is mostly true if, on every physical channel C, adja-
cent packet transmissions do not overlap, i.e., if transmission
latency dC
inever exceeds the generation period TM. Under
such hypothesis, if a given PoW experiment was repeated, in
exactly the same operating and environmental conditions, us-
ing a real RDA testbed, the sequence of transmission attempts
on air would simply consist of a subset of the PoW run. In
fact, some PoW attempts are possibly saved because of early
termination. In other words, for any attempt mC
i,` carried out
when RDA is enabled there would be an equivalent attempt
(same starting time on air and duration) carried out by PoW,
but not vice-versa, because all attempts for which (5) holds
are prevented by RDA. For instance, in the example shown in
Fig. 3, RDA cancels the last two attempts carried out by PoW
on B∈Qi, that is, mB
i,3and mB
i,4.
In the above analysis there is no need to define new
quantities that explicitly refer to RDA, as they coincide with
PoW. The only difference lies in the number of transmission
attempts, which for RDA is possibly smaller than PoW. The
number of attempts for packet mion channel Cin the RDA
case is denoted w0C
i. For the quickest channel w0Qi
i=wQi
i
while, for any other channel C∈Qi,0≤w0C
i≤wC
i. In
case of early termination on C∈Qi, the lowest value of `for
which (5) is true coincides with w0C
i+1. In particular, w0C
i= 0
refers to the case where also the initial attempt (`= 1) was
prevented on C(which implies that miwas not sent there).
If (5) is not verified for any `≤wC
i, early termination is not
possible for mC
iand w0C
i=wC
i.
In our experiments on the PoW testbed, less than 0.1% of
the packets experienced queuing delays (see column ΥC
d>100ms
in Table II), i.e., more than 99.9% of the packet transmissions
did not overlap. Therefore, above virtual RDA analysis can be
reliably applied to the related real PoW logs.
C. Experimental Evaluation of RDA
Let Lbe a redundant link like the one implemented by
our PoW testbed, for which the hypotheses that permit to
apply virtual RDA analysis hold. In theory, early termination
could be completely analyzed from the related experimental
logs. However, since starting times tC
W,i,` are only available
for final attempts, exact w0C
ivalues can not be determined.
Nevertheless, an indication can be still obtained on whether,
for any given packet, some attempts can be saved. This
provides an upper bound on RDA spectrum consumption. As
shown in Fig. 3, if (5) holds for the final attempt mC
i,` on
channel C∈Qi, then at least one attempt is prevented for
sure on C. In the sample packet transmission depicted in the
figure, where wB
i= 4, this approach permits to infer that
w0B
i≤3(although, actually, w0B
i= 2).
Let eC
i(early termination condition) be a Boolean value
that is true if and only if, owing to XACK xi, RDA is able to
terminate the transmission of packet mion channel C∈Qiin
advance with respect to what is foreseen by conventional Wi-
Fi. As said before, this is only possible for packets correctly
delivered (and acknowledged) on L. From (5),
eC
i,lL
i= 0 ∧tX,i +TLRE < tC
W,i(6)
where tC
W,i is given by (2) and (4), for successful and failed
packet copy transmissions, respectively. For the quickest chan-
nel, (6) correctly provides eQi
i= 0, so it holds for every
C∈ CL. Above considerations identically apply when RDA
is coupled with some heuristics (PDA).
D. Metrics about Spectrum Consumption
Reliability through time and frequency diversity comes to
the price of a higher spectrum consumption. The holistic
approach provided by DA mechanisms offers non-negligible
advantages in this respect. The related benefits in terms of
bandwidth saving can be evaluated by analyzing the number
of transmission attempts actually performed on air during any
given experiment.
A number of simple metrics have been introduced to this
extent. In the following, when referring to redundant links,
the kind of DA mechanism will be specified, if relevant, as
superscript, in the place of the link. Term DA generically refers

7
to both RDA and PDA, while PoW denotes the lack of any
DA. In these cases, the link can be inferred from the context
(either L, in general reasoning, or A+B, when referring to
our experimental testbed).
1) Saved Bandwidth: The fraction of packets in an experi-
mental run whose transmission is terminated early on physical
channel C∈ CLis
eC,1
NX
i=1...N
eC
i.(7)
For any packet sent on the redundant link L, the average
number of copies that experience early termination, all chan-
nels considered, is
eL,1
NX
i=1...N X
C∈CL
eC
i=X
C∈CL
eC.(8)
When L=A+B, as in our duplex testbed, there is only one
channel other than the quickest, and eA+Bcoincides with the
fraction of packets terminated early on either Aor B.
Upon early termination of a packet copy, one or more
transmission attempts are canceled for sure on the related
channel. This implies that eLconstitutes a lower bound on
the average number of useless attempts that, overall, can be
prevented for each packet by RDA with respect to PoW, and
provides an estimate of the bandwidth saving achieved over
basic seamless redundancy. Packets that are dropped on all
physical channels but one are those for which the most savings
are expected. In fact, in the absence of early termination, all
attempts up to the retry limit are (unsuccessfully) carried out
on faulty channels.
Our testbed does not make TC
D,i available when transmission
on adapter Cfails, nor it can be easily inferred (because of rate
adaptation mechanisms), and so eC
ican not be evaluated for
the related packet copies. This is not a severe drawback since,
due to the effectiveness of MAC retransmissions, the fraction
of packets that were dropped on either of the physical channels
in our experiments was always very low (less than 0.0033%).
Hence, their contribution to eLcan be quietly neglected, by
assuming for them eC
i= 0. It is worth pointing out that
the case when the loss rate is not negligible only leads to a
pessimistic, smaller value for eL, which is acceptable as this
quantity represents a lower bound.
2) Network load: The average number of transmission
attempts per packet (including retries) carried out on any
channel Cduring an experiment on the PoW testbed is
wC,1
NX
i=1...N
wC
i.(9)
As in Wi-Fi every packet is sent on air at least once, wC≥1.
The value of wC
ifor successfully delivered packet copies
can be obtained by inspecting the device driver of the re-
lated adapter, whereas it is not available for failed copies.
In the latter case, using for wC
ithe default retry limit as
per Wi-Fi specifications [8] is not the best choice, because
of the rate adaptation mechanism (Minstrel) [22], which
changes transmission parameters dynamically according to
some channel statistics. Hence, for dropped packets (i.e., if
lC
i= 1) we assumed wC
ito be equal to the maximum
among all the values we measured in the experiments, that
is wC
i= maxi0∈1...N,C0∈CLwC0
i0= 21. Again, since the
loss ratio in our experiments was quite low, this pessimistic
approximation is perfectly acceptable.
Let wLbe the average number of transmission attempts
on air per packet, considering all channels. It depends on
the overhead caused by both retransmissions and seamless
redundancy. In PoW, every packet is always transmitted on
all channels, so
wPoW ,1
NX
i=1...N X
C∈CL
wC
i=X
C∈CL
wC=|CL| · w, (10)
where w=1
|CL|PC∈CLwCis the mean number of attempts
per packet and per channel, and describes, on average, the
behavior of a single conventional Wi-Fi channel. Due to the
different spectrum conditions of physical channels C∈ CL,
wC
ivalues usually do not coincide and the same holds for
wC. Since wC≥1,wPoW ≥ |CL|, where |CL|= 2 for duplex
links. Roughly speaking, spectrum consumption in our PoW
testbed is, on average, twice as much as Wi-Fi.
When a DA mechanism is in effect, equation (10) can be
rewritten as
wDA =1
NX
i=1...N

wQi
i+X
C∈Qi
w0C
i
.(11)
Unlike PoW, w0C
ican be equal to 0for C∈Qi, as the
only constraint is that every packet has to be sent on air at
least once on at least one channel. Therefore, like plain Wi-
Fi, wDA could be as low as 1. This happens when, for every
packet mi,wQi
i= 1 and w0Qi
i= 0. In theory, for some specific
runs, spectrum consumption for Lmay be even lower than any
of its physical channels when used as simplex links. In other
words, condition wDA <minC∈CLwCmight possibly hold.
For instance, when L=A+B, consider the case wQi
i< wQi
i
and w0Qi
i= 0 (i.e., the number of attempts that would be
performed on Qiin the absence of DA is higher than on Qi,
but all of them are prevented by DA), where Qi=Aif iis
even and Qi=Botherwise.
In general, eC
i= 0 →w0C
i=wC
i, while eC
i= 1 →
w0C
i< wC
i. This implies that w0C
i≤wC
i−eC
i. For a given
DA mechanism, from (8), (10), and (11), we have
wDA ≤wPoW −eDA.(12)
Hence, DA mechanisms never worsens spectrum consumption
of Wi-Red.
3) Communication efficiency: from the originator view-
point, it is defined as the reciprocal of the average number
of transmission attempts per packet. For physical channel C
it corresponds to ηC= 1/wC. While being related to the
goodput, they do not coincide, as ηCdoes not take into account
protocol overheads (frame preambles, MAC headers and trail-
ers, acknowledgments, inter-frame spaces, random backoff,
etc.), but only retransmissions. Definition for redundant link
Lis similar, ηL= 1/wL. If DA is used, a lower bound can
be found for ηDA, defined as
ˇηDA ,1
wPoW −eDA ≤ηDA.(13)

8
4) Relative network load: The load on the link for a given
DA mechanism with respect to PoW is defined as
ϑDA ,wDA
wPoW =wDA
|CL| · w(14)
where 1/wPoW ≤ϑDA ≤1(the lower, the better). Similarly,
the relative load implied, on average, over plain, simplex Wi-Fi
links, is expressed as ΘDA ,wDA/w =|CL| · ϑDA.
Equation (14) can be rewritten as wDA =ϑDA · |CL| · w,
which shows that the average number of transmission attempts
on air per packet, including all channels, is given by the
product of three factors: the relative load of the specific DA
technique over PoW, the number of channels, and the mean
per-packet load in Wi-Fi due to retransmissions, obtained by
averaging all channels of the redundant link. Only the first
factor depends on DA effectiveness (which is affected, e.g.,
by TLRE). Thus, ϑDA is a valuable metric to assess the ability
of a certain DA mechanism to save bandwidth.
Due to the limitations of our testbed, what we can measure
is an upper bound on ϑDA, i.e.,
ˆ
ϑDA ,1−eDA
wPoW ≥ϑDA.(15)
Therefore, DA performance is never worse than what we found
using our approximate analysis.
5) Simplex transmissions on the redundant link: The last
quantity we consider is the fraction of packets which, as a
consequence of early termination, are only sent on Qi, that is,
those for which w0C
i= 0,∀C∈Qi. From the point of view of
nearby wireless STAs, transmission of these packets appears
to take place as on a conventional Wi-Fi (simplex) link, which
is not fixed but dynamically selected among channels C∈ CL.
Counting these packets exactly is not possible in our testbed.
However, condition
zC
i,eC
i∧wC
i= 1 (16)
implies that no transmission attempts at all are performed for
mion C. In fact, wC
i= 1 means that only the initial attempt
is performed by PoW, but it is canceled by DA since eC
i= 1.
For the quickest channel (16) correctly yields zQi
i= 0. Hence,
zC=1
NPi=1...N zC
iis a lower bound on the fraction of
packets whose transmission is fully prevented on C∈ CLby
DA.
For the redundant link L, Boolean condition zL
i,
QC∈QizC
iimplies that mitransmission is completely pre-
vented on every channel other than the quickest. We are
interested in the metric
zL=1
NX
i=1...N
zL
i,(17)
which represents a lower bound on the fraction of packets
that were sent on air on one channel only (the quickest), as
on a simplex link. Since in our duplex testbed there is only
one channel other than Qi,zA
iand zB
ican not be both true,
which means that zA+B
i=zA
i+zB
i. In this specific case,
zA+B=zA+zB. Given that zA+B
i→eA+B
i= 1, where
eA+B
i=eA
i+eB
i, it also follows that zA+B≤eA+B.
E. Experimental Results for RDA
To assess the improvements RDA can bring in the real
world, three experiments were run on the PoW testbed in
different operating conditions. Channel A, set in the 2.4 GHz
band, was quite crammed, due to the presence of many active
wireless devices in nearby premises, which were not under our
control. To try ensuring some repeatability, every experiment
lasted for a whole day (24 hours). Even so, results concerning
channel Ain the different runs can not be reliably compared
one another. Channel B, set in the 5 GHz band, had very light
load. Thus, random interfering traffic was purposely injected
on that channel during experiments. Interference consisted
in the repeated generation of packet bursts. Payload size of
interfering packets was set to 1500 bytes. The number of pack-
ets in each burst was not fixed, but followed an exponential
random distribution with mean value 300 and truncated to
1500. Within any burst, transmission requests for packets were
spaced by 400 µs. The duration of the gap between the end of
a burst and the beginning of the following one was randomly
selected, according to an exponential distribution with mean
value 200 ms and truncated to 20 s.
Runs were carried out with either 1,2, or 4interfering STAs
on B(1 Int.,2 Int., and 4 Int., respectively). Experimental re-
sults, including bounds calculated using virtual RDA analysis,
are reported in Table II, under the assumption that TLRE = 0
(i.e., when early termination is assumed to be implemented
in hardware in the RSTA). For each experiment, three rows
are included, for physical channels (A,B) and the redundant
link (A+B). The first set of columns includes statistics about
the transmission latency d: mean value, standard deviation,
median, 99.99 percentile, and maximum. In the second set of
columns, the deadline miss ratio ΥC
d>H (percentage of packets
whose latency exceeded H, where His set to either 10 or
100 ms) and the packet loss ratio ΥC
Lare shown. Although
this paper does not focus on communication quality, these
quantities provide a glimpse of the improvements seamless
redundancy achieves.
The most interesting results, reported in the sets of columns
on the right, concern RDA ability to prevent useless attempts.
Relevant measured quantities for physical channels Aand B
and the redundant link A+Binclude the fractions of early
terminated packets (eC) and simplex packets (zC), as well
as the mean number of transmission attempts performed per
packet (wC). Communication efficiency when no DA is in use,
both for physical channels (ηAand ηB) and the redundant link
(ηPoW), is given in column ηC. Finally, the rightmost columns
report metrics about RDA computed using virtual analysis, i.e.,
the lower bound on communication efficiency (ˇηRDA) and the
upper bounds on the relative network load with respect to
either PoW ( ˆ
ϑRDA) or plain Wi-Fi ( ˆ
ΘRDA).
Experimental results are quite valuable and show that, in
real operating conditions, RDA cuts redundancy overhead
tangibly. The lower bound ˇηRDA on communication efficiency
(and hence, ηRDA as well) was always noticeably higher than
ηPoW, irrespective of the interfering traffic, and the upper
bound ˆ
ϑRDA on the network load relative to PoW ranged from
64.54% to 79.19%. This means that all benefits on commu-

9
TABLE II
EXPERIMENTAL RESULTS FOR PHYSICAL CHANNELS AAND BAND TH E RE DUN DAN T LI NK A+BIN RDA (TLRE = 0 µs) .
Exp. Chan. dCσC
ddC
p50 dC
p99.99 dC
Max ΥC
d>10ms ΥC
d>100ms ΥC
LeCzCwCηCˇηRDA ˆ
ϑRDA ˆ
ΘRDA
(C) (ms) (‰) (%) (%)
1 Int.
A0.407 1.113 0.083 26.138 135.78 1.7385 0.0313 0.0278 0.1917 0.1760 1.0231 97.75 – – –
B0.234 0.331 0.086 4.821 52.242 0.0255 0.0000 0.0000 0.2382 0.2044 1.0422 95.95 – – –
A+B0.108 0.135 0.081 2.193 52.242 0.0012 0.0000 0.0000 0.4298 0.3804 2.0653 48.42 61.14 79.19 158.38
2 Int.
A0.560 1.564 0.084 48.406 209.41 2.7142 0.0359 0.0243 0.2215 0.2039 1.0293 97.15 – – –
B0.488 0.754 0.229 19.677 91.892 0.5046 0.0000 0.0000 0.4289 3439 1.1145 89.73 – – –
A+B0.162 0.275 0.079 4.081 91.892 0.0023 0.0000 0.0000 0.6503 0.5478 2.1438 46.65 66.96 69.67 139.34
4 Int.
A0.576 1.787 0.076 49.721 201.38 4.8914 0.0394 0.0324 0.1339 0.1227 1.0299 97.10 – – –
B2.532 6.753 1.021 126.92 219.09 41.683 0.7326 0.0000 0.7427 0.4636 1.4424 69.33 – – –
A+B0.281 0.590 0.076 12.516 40.726 0.2014 0.0000 0.0000 0.8766 0.5863 2.4723 40.45 62.67 64.54 129.08
nication quality brought by seamless redundancy are achieved
at a fraction of the bandwidth. Since spectrum conditions
may vary suddenly and unexpectedly, fair comparison between
redundant and simplex Wi-Fi links requires that the worst of
the physical channels is taken into account. As can be seen
in the table, in the 4 Int. case communication efficiency of
RDA is marginally worse than channel Balone (62.67% vs.
69.33%), but communication quality on A+Bis much better
than both Aand B.
An additional benefit of RDA is that, the higher the overall
network traffic, the better it works. See, e.g., when inter-
fering traffic on Bis varied. The measured fraction eRDA
of early terminated packet transmissions with 1,2, and 4
interferers was equal to 42.98% (1 Int.), 65.03% (2 Int.), and
87.66% (4 Int.), respectively. Concerning the fraction zRDA of
simplex packet transmissions, when traffic is not negligible
more than half of the packets (54.78% and 58.63%, for
2 Int. and 4 Int., respectively) were actually sent only on one
channel, hence causing no overhead on the other. This is
no surprise, as RDA effectiveness improves when variability
of transmission latencies on replicated channels gets higher.
Such behavior is highly desirable since, when overall network
traffic increases, the overhead due to redundancy consequently
shrinks, hence lowering the impact of transmitting RSTAs
on the network load. This counteracts congestion phenomena
affecting random access techniques, which worsen the quality
of communication (see, e.g., the latency on channel Bwhen
interfering traffic is increased). As a consequence, network
stability (overall traffic on air vs. total offered load) improves.
The effect of using sub-optimal (e.g., software) RDA solu-
tions, modeled by considering non-negligible LRE latencies,
is shown in Fig. 4. Three plots are reported, corresponding to
the above experimental runs, where TLRE in (6) is virtually
varied in the range [0 −1000] µs. As can be seen, delays due
to XACK processing always worsen spectrum consumption
as, on average, less transmission attempts can be terminated
in advance. For example, in the 1 Int. case, eRDA decreases by
a factor 4(from 42.98% to 10.18%) when TLRE is increased to
1 ms. For this reason, implementation of RDA in the hardware
(or firmware) of Wi-Fi adapters is highly suggested. This
also confirms the advantages brought by our approximate
virtual approach to RDA evaluation (based on the analysis
of PoW experimental logs) with respect to measurements
carried out directly on a non-properly optimized software RDA
implementation. Moreover, it is important to remark the ability
of the virtual approach to infer results for different techniques
and parameters starting from the same set of data, allowing
for a fair comparison between them.
V. PROACT IV E DUP LI CATION AVOI DANCE
Proactive Duplication Avoidance (PDA) is meant to improve
RDA behavior by further reducing the number of useless
duplicates sent on air. Unlike RDA, which has no counter-
indications, PDA may set some trade-off among performance
indices (e.g., the approach described below trades responsive-
ness for communication efficiency). Therefore, when employ-
ing these mechanisms, a compromise has to be necessarily
found, typically by tuning suitable configuration parameters.
PDA often relies on heuristics, which could be arbitrarily
complex and possibly intertwined with the IEEE 802.11 MAC.
The class of PDA mechanisms we considered in this paper
only concerns duplex redundant links, and operates by inten-
tionally displacing transmission requests on physical channels.
In this way, the likelihood that the LRE succeeds in canceling
the transmission of the deferred copy may increase tangibly,
hence reducing spectrum consumption.
A. Timed Duplicate Deferral
The PDA technique we analyze here is called Timed Dupli-
cate Deferral (TDD). A similar solution, not coupled with early
termination, was proposed in [23]. TDD is much simpler than
the Dynamic Duplicate Deferral (DDD) technique presented
in [17], as its operation does not rely on any indication
provided at runtime by the MAC layer. Thus, it can be easily
implemented using commercial Wi-Fi adapters.
TDD operates as follows: upon generation of packet mi,
the LRE immediately invokes its transmission on A(referred
to in the following as primary channel). At the same time, it
starts a countdown timer initialized to the deferral time TD,i.
When TD,i expires, and provided that the XACK from Ahas
not been raised in the meanwhile, transmission is also invoked
on B(secondary channel). In formulas
tB
T,i =tA
T,i +TD,i.(18)
A distinct TD,i value can be chosen for each packet mi
depending, e.g., on its payload size and the bit rate chosen
at runtime by the rate adaptation mechanism.

10
0
10
20
30
40
50
60
70
80
90
100
0 250 500 750 1000
1 Interferer
0
10
20
30
40
50
60
70
80
90
100
0 250 500 750 1000
2 Interferers
0
10
20
30
40
50
60
70
80
90
100
0 250 500 750 1000
4 Interferers
eRDA / zRDA /ˆ
ϑRDA [%]
TLRE [µs]
eRDA
zRDA
ˆ
ϑRDA
TLRE [µs]
eRDA
zRDA
ˆ
ϑRDA
TLRE [µs]
eRDA
zRDA
ˆ
ϑRDA
Fig. 4. Fractions of early terminated (eRDA ) and simplex (zRDA) transmissions, and relative load (ˆ
ϑRDA), vs. TLRE (for RDA).
B. Experimental Evaluation of TDD
TDD performance can be experimentally assessed in our
testbed using two different approaches, which will be shown
to provide very similar results. In both cases, bandwidth saving
is evaluated when varying the deferral time, by analyzing logs
in the same way as for RDA.
1) Real transmission deferral: The most straightforward
way to analyze TDD behavior is to rely on a slightly modified
PoW testbed, where the transmission request issued by the
measurement task on Bfor each packet miis actually delayed
by TD,i with respect to A. To make timing accuracy indepen-
dent from the implementation of the measurement task and
the underlying computing platform (including the operating
system), the exact displacement between the copies of every
packet mi(that in theory corresponds to TD,i) was evaluated
from measured timestamps as tB
T,i −tA
T,i .
2) Virtual transmission deferral: Approximate yet satis-
factory results can be obtained from the logs acquired on
the unmodified PoW testbed (i.e., without any displacement
between tA
T,i and tB
T,i ). Results in [19] show that, in a well-
designed redundant link, channel conditions can be considered
statistically independent. This means that deferring transmis-
sion requests on Bdoes not affect the behavior of A, and
vice-versa. Moreover, channel behavior in the short term is
mostly stationary, hence, statistics on experimental results do
not change upon displacement in time, as long as the offset is
short enough (in the order of 1 ms).
According to the above hypotheses, one can reasonably
assume that the sequence of samples obtained by: a) adding
the deferral time to every timestamp (tB
T,i and tB
X,i) in the
tuples τB
i, i = 1...N of a log acquired from channel Bin the
unmodified PoW testbed, and b) leaving all the other quantities
(lB
i,wB
i,TB
D,i, and TB
A,i) unmodified, could satisfactorily
mimic the sequence of tuples acquired on Bin a real setup
like the one described in Section V-B1, where transmissions of
copies are actually displaced. This approach permits to create
a bunch of virtual TDD experiments from a single real PoW
experimental run, where TD,i values (provided that they are
small, so as not to contradict stationary hypothesis) can be
decided during post-analysis.
In the following, timestamps with the double prime symbol
refer explicitly to virtual TDD experiments. For example, the
estimated ACK reception time for packet mion the secondary
channel Bin a virtual TDD experiment (characterized by the
sequence of TD,i values) is
t00B
X,i =tB
X,i +TD,i (19)
where tB
X,i is the timestamp taken in the unmodified PoW
testbed. Timestamps t00B
T,i and t00B
W,i are defined similarly. For the
primary channel A, virtual and measured timestamps coincide.
In presence of TDD deferral, the transmission latency on the
redundant link for packet miis dTDD
i= min dA
i, dB
i+TD,i.
This implies that, unlike RDA, dTDD
i≥dPoW
i.
The same approach used for RDA analysis can be applied to
estimate the effects of early termination in TDD, by replacing
real samples in the log with virtually deferred ones. Since
channel usage is no longer symmetric, (6) splits into
eB
i,hlA
i= 0 ∧t00 A
X,i +TLRE < t00B
W,ii(20)
eA
i,hlB
i= 0 ∧t00 B
X,i +TLRE < t00A
W,ii
and, from (19),
eB
i,lA
i= 0 ∧tA
X,i +TLRE < tB
W,i +TD,i(21)
eA
i,lB
i= 0 ∧tB
X,i +TD,i +TLRE < tA
W,i.
C. Validation of Virtual TDD Analysis
A preliminary experiment was carried out to assess if results
obtained through virtual deferral of samples performed offline
are comparable to actually displacing transmission requests
in the testbed at runtime. Since the payload size of packets
sent by the measurement task was fixed, all TD,i offsets in
experimental runs have to be set to the same value TD. Jitter
induced by the operating system on the actual displacement
between the transmission times of the two copies of each
packet in the modified PoW testbed described in Section V-B1
was found to be quite small. Therefore, we can practically
assume that the deferral time in any run is fixed and co-
incides with the mean value of the measured offset, that is
TD≈TD=1
NPi=1...N tB
T,i −tA
T,i .
We performed several runs, where TDwas varied in the
range [−250... + 250] µs. The nominal displacement used in
each run was selected in such a way that the related TD
approximately corresponds to one among 17 predefined values.
In the case TD= 0 (which corresponds to pure RDA),
transmissions were not displaced, as in the unmodified PoW
testbed. Negative TDvalues refer to experiments where Bwas
the primary channel, while Awas deferred by |TD|. These

11
60
70
80
90
100
−250 0 250
2 Interferers
0.15
0.2
0.25
−250 0 250
eTDD [%]
TD[µs]
Virtual
Real
dTDD [ms]
TD[µs]
Fig. 5. Comparison of eTDD and dTDD obtained with real and virtual
transmission deferrals.
cases are explicitly included because behavior of channels is
not symmetric.
To mitigate discrepancies due to slow variations of the
wireless spectrum conditions, runs for the different deferral
times were interleaved in a single experiment, which lasted
one day. Such a kind of assessment was lacking in [18].
Two interferers were activated on Bduring the experiment
(as in the 2 Int. case). Results, concerning both the fraction
of early terminated packets (eTDD) and mean transmission
latency (dTDD), are reported in Fig. 5 using “×” symbols.
The very same logs, restricted to the subset of samples for
which TD= 0 (i.e., when the testbed operated according
to pure PoW, with no deferrals), were then used to generate
virtual TDD runs according to Section V-B2, where TDwas
varied with fine granularity in the range [−300... + 300] µs.
Resulting plots are superposed to Fig. 5 using solid lines. As
can be seen, there is a very good agreement between the
results obtained with the two approaches. This means that
virtual deferral carried out during post-analysis provides a
reliable approximation of real deferral performed by TDD.
For this reason, the relation between spectrum consumption,
communication quality, and TD, will be analyzed in the
following by using the virtual approach.
D. Experimental Results for TDD
To provide coherent results, the very same experimental
PoW logs used in Section IV to analyze RDA (1 Int., 2 Int.,
and 4 Int., depending on the amount of interference on B) were
employed to carry out virtual TDD analysis. First, they were
processed to cope (virtually) with transmission deferral; then,
the same analysis as for RDA was performed to estimate the
effects of early terminations. In particular, we analyzed the
tradeoff between spectrum consumption and communication
quality when TDis varied between 0(that corresponds to pure
RDA) and 500 µs.
Results for the three runs are reported in Fig. 6. Plots
in the upper part of the figure show some metrics about
bandwidth consumption, i.e., the fraction of early terminated
packets (eTDD), the fraction of simplex packets (zTDD), and
the relative network load with respect to PoW (ˆ
ϑTDD). Instead,
those in the lower part concern the average latency on the
redundant link (dTDD).
As can be seen, in each run there are two thresholds for TD,
located around ±100 µs, within which spectrum consumption
does not vary appreciably, this meaning that TDD adoption
does not bring any improvement over pure RDA. However,
passed those thresholds, behavior improves suddenly, as wit-
nessed by, e.g., the noticeable increase of eTDD values. This
is because transmission requests on the secondary channel
are deferred enough by TDD, so that most of them can be
prevented by the LRE. TDD benefits are higher when channels
are not heavily loaded, as in the 1 Int. case, mostly because
saving due to RDA alone is less tangible. In this respect, TDD
adoption improves Wi-Red stability versus network traffic.
In theory, as long as queuing phenomena in Wi-Fi adapters
are negligible, parameter TDhas no effect on the amount
of lost packets in TDD, which is the same as RDA and
PoW. Instead, it affects transmission latencies, as they are
measured from the time the transmission request is issued
on the primary channel. As can be seen in the plots of
Fig. 6, mean latency dTDD progressively worsens as |TD|
increases (in either direction). This is quite obvious, since
TDD intentionally delays one of the two channels.
Plots about dTDD are not symmetric about the y-axis. Their
shape (but not the values) in the different runs is mostly
the same when TD>0. However, this is no longer true
when TD<0. This can be explained by remembering that,
in the latter case, transmissions are delayed on A, and so
a higher number of packets arrive first on Bwhen |TD|is
increased. As a consequence, communication quality on B
matters more. When interference on that channel is higher, as
in the 4 Int. case, the related latency is larger. This implies
that, for negative TDvalues, the slope of the plots concerning
the latency on the redundant link is steeper.
The optimal values for parameter TDare located just
above said thresholds, and approximately correspond to the
overall duration on air of the initial attempt for mion the
primary channel (including DATA and ACK frames, SIFS, and
TLRE), plus a safety margin. With this arrangement, all packet
transmissions that find such channel idle and immediately
succeed (a non-unusual condition in a well-dimensioned Wi-Fi
network, as experimentally shown in [24]) are not replicated
on the secondary channel. This particular choice provides sig-
nificant benefits concerning spectrum consumption, by keeping
the latency increase, on average, to the minimum.
For example, if TDis set to either −100 µsor +100 µs
(near to the optimal values for our experimental conditions),
the values in Table III are obtained. With these configurations,
eTDD was always higher than ∼90% for all the experiments
reported in the plots (and, in the 4 Int. case, it exceeded 97%
when TD= +100 µs). This means that less than 10% of
the packets did not incur in early termination. At the same
time, dTDD never grown more than 76 µswith respect to pure
RDA (see the 4 Int. case, TD=−100 µs). In the case adaptive
TDD algorithms were implemented, able to dynamically select

12
0
0.1
0.2
0.3
0.4
0.5
−500 −250 0 250 500 0
0.1
0.2
0.3
0.4
0.5
−500 −250 0 250 500 0
0.1
0.2
0.3
0.4
0.5
−500 −250 0 250 500
40
50
60
70
80
90
100
−500 −250 0 250 500
1 Interferer
40
50
60
70
80
90
100
−500 −250 0 250 500
2 Interferers
40
50
60
70
80
90
100
−500 −250 0 250 500
4 Interferers
dTDD [ms]
TD[µs]
dTDD
TD[µs]
dTDD
TD[µs]
dTDD
eTDD / zTDD /ˆ
ϑTDD [%]
TD[µs]
eTDD
zTDD
ˆ
ϑTDD
TD[µs]
eTDD zTDD ˆ
ϑTDD
TD[µs]
eTDD zTDD ˆ
ϑTDD
Fig. 6. Fractions of early terminated (eTDD ) and simplex (zTDD) transmissions, relative load ( ˆ
ϑTDD), and mean latency (dTDD ), vs. TD(for TDD).
TABLE III
EXPERIMENTAL RESULTS FOR TDD WITH OPTIMAL TDVALUE .
Exp. TDeTDD zTDD ˆ
ϑTDD ˆ
ΘTDD dTDD (∆d)dTDD
p99 dTDD
p99.99
(µs) (%) (ms)
1 Int.
−100 0.9061 0.8567 56.13 112.26 0.135 (+0.027) 0.740 2.217
0 0.4298 0.3804 79.19 158.36 0.108 ( 0) 0.727 2.193
+100 0.9322 0.8834 54.86 109.72 0.130 (+0.022) 0.813 2.260
2 Int.
−100 0.8938 0.7915 58.31 116.62 0.208 (+0.046) 1.345 4.094
0 0.6503 0.5478 69.67 139.34 0.162 ( 0) 1.320 4.081
+100 0.9444 0.8421 55.95 111.90 0.187 (+0.025) 1.394 4.135
4 Int.
−100 0.9202 0.6304 62.78 125.56 0.357 (+0.076) 2.766 12.539
0 0.8766 0.5863 64.54 129.08 0.281 ( 0) 2.709 12.516
+100 0.9725 0.6818 60.66 121.32 0.296 (+0.015) 2.755 12.519
the best primary channel, dTDD increase would likely remain
below 25 µs(see the 2 Int. case, TD= +100 µs).
VI. CONCLUSION
Seamless redundancy applied to IEEE 802.11 greatly im-
proves its reliability and makes it suitable for time-sensitive
control applications, preserving backward compatibility with
Wi-Fi. When PRP is layered directly above wireless adapters,
as in the PoW approach, communication quality improves
dramatically, but spectrum consumption becomes twice as
much as non-redundant Wi-Fi. The Wi-Red proposal exploits
synergies between the MAC retransmission mechanism and
channel redundancy in order to improve behavior further. In
particular, duplication avoidance mechanisms reduce spectrum
consumption by preventing unnecessary transmission attempts.
This paper specifically focuses on the ability of such
mechanisms to cut down the additional traffic implied by
seamless redundancy. Suitable metrics have been defined to
this purpose, which were evaluated experimentally on a real
testbed. In order to be effective, early termination of packet
transmissions has to be implemented in the hardware/firmware
of Wi-Fi adapters, which is currently out of our reach. For
this reason, an approximate method was devised, where ex-
perimental data are acquired on a real testbed that implements
PRP on commercial Wi-Fi devices and then processed under
the (verified) hypotheses of independence between subsequent
transmissions on air and slowly-varying spectrum conditions.
Although in this way we can only obtain a lower bound
on the amount of bandwidth saved by duplication avoidance
mechanisms (which means that the related results are pes-
simistic), improvements we measured are nevertheless note-
worthy. Reactive approaches, which cancel transmission of
duplicate copies upon reception of the related ACK on at least
one channel, provide all the benefits of PoW (and often behave
even better in terms of improved reliability and timeliness)
but using less bandwidth. In our experiments, RDA traffic
(or, better, its upper bound as given by ˆ
ΘRDA) was 29.08%
to 58.36% higher than what generated, on average, by non-
redundant Wi-Fi. For the proactive TDD approach, additional
spectrum consumption with respect to Wi-Fi, as per ˆ
ΘTDD,
lay in the range from 9.72% to 25.56%. In the TDD case, the
price to pay is a slight increase in the transmission latency
(some tens of µs).
In a world where the wireless spectrum is more and more
congested, adoption of seamless redundancy to increase Wi-Fi
reliability practically mandates the inclusion of duplication
avoidance mechanisms. As future work, we plan to investigate
adaptive PDA solutions, aimed at providing optimal perfor-
mance in spite of variations of the generated traffic pattern
and the wireless spectrum conditions, and to analyze networks
including multiple RSTAs.
REFERENCES
[1] “IEEE Guide for Wireless Access in Vehicular Environments (WAVE)
- Architecture,” IEEE Std 1609.0-2013, pp. 1–78, March 2014.
[2] T. Watteyne, M. Palattella, and L. Grieco, “Using IEEE 802.15.4e
Time-Slotted Channel Hopping (TSCH) in the Internet of Things
(IoT): Problem Statement,” Internet Requests for Comments, RFC
Editor, RFC 7554, May 2015, http://www.rfc-editor.org/rfc/rfc7554.txt.
[Online]. Available: http://www.rfc-editor.org/rfc/rfc7554.txt
[3] A. Boubrima, W. Bechkit, and H. Rivano, “Optimal WSN Deployment
Models for Air Pollution Monitoring,” IEEE Transactions on Wireless
Communications, vol. 16, no. 5, pp. 2723–2735, May 2017.

13
[4] C. T. Kone, A. Hafid, and M. Boushaba, “Performance Management
of IEEE 802.15.4 Wireless Sensor Network for Precision Agriculture,”
IEEE Sensors Journal, vol. 15, no. 10, pp. 5734–5747, Oct 2015.
[5] S. Petersen and S. Carlsen, “Wirelesshart versus isa100.11a: The format
war hits the factory floor,” IEEE Industrial Electronics Magazine, vol. 5,
no. 4, pp. 23–34, Dec. 2011.
[6] J. D. Decotignie, “Ethernet-Based Real-Time and Industrial Communi-
cations,” Proc. IEEE, vol. 93, no. 6, pp. 1102–1117, June 2005.
[7] “IEEE Standard for Low-Rate Wireless Networks,” IEEE Std 802.15.4-
2015 (Revision of IEEE Std 802.15.4-2011), pp. 1–709, April 2016.
[8] “IEEE Standard for Information technology–Telecommunications and
information exchange between systems Local and metropolitan area
networks–Specific requirements - Part 11: Wireless LAN Medium Ac-
cess Control (MAC) and Physical Layer (PHY) Specifications,” IEEE
Std 802.11-2016 (Revision of IEEE Std 802.11-2012), pp. 1–3534, Dec
2016.
[9] R. Moraes, F. Vasques, P. Portugal, and J. Fonseca, “VTP-CSMA: A
Virtual Token Passing Approach for Real-Time Communication in IEEE
802.11 Wireless Networks,” IEEE Trans. Ind. Informat., vol. 3, no. 3,
pp. 215–224, 2007.
[10] P. Wang and W. Zhuang, “A collision-free MAC scheme for multimedia
wireless mesh backbone,” IEEE Transactions on Wireless Communica-
tions, vol. 8, no. 7, pp. 3577–3589, July 2009.
[11] Y. H. Wei, Q. Leng, S. Han, A. K. Mok, W. Zhang, and M. Tomizuka,
“RT-WiFi: Real-Time High-Speed Communication Protocol for Wireless
Cyber-Physical Control Applications,” in 2013 IEEE 34th Real-Time
Systems Symposium, Dec 2013, pp. 140–149.
[12] L. Seno, G. Cena, S. Scanzio, A. Valenzano, and C. Zunino, “Enhancing
Communication Determinism in Wi-Fi Networks for Soft Real-Time
Industrial Applications,” IEEE Trans. on Ind. Informat., vol. 13, no. 2,
pp. 866–876, April 2017.
[13] IEC, “Industrial communication networks - High availability automation
networks - Part 3: Parallel Redundancy Protocol (PRP) and High-
availability Seamless Redundancy (HSR),” IEC Std 62439-3 ed 2.0, pp.
1–177, 2010.
[14] M. Rentschler and P. Laukemann, “Towards a reliable parallel redundant
WLAN black channel,” in IEEE Int. Workshop on Factory Communica-
tion Systems (WFCS), 2012, pp. 255–264.
[15] Belden Inc., “PRP - Parallel Redundancy Protocol,” [Online]. Available:
http://www.hirschmann.com/en/Hirschmann Produkte/Industrial
Ethernet/Technologies/PRP - Parallel Redundancy Protocol,
Accessed: 2016-01-22.
[16] M. Mohiuddin, M. Popovic, A. Giannakopoulos, and J. L. Boudec,
“Experimental validation of the usability of Wi-Fi over redundant paths
for streaming phasor data,” in IEEE International Conference on Smart
Grid Communications (SmartGridComm), Nov 2016, pp. 533–538.
[17] G. Cena, S. Scanzio, and A. Valenzano, “Seamless Link-Level Redun-
dancy to Improve Reliability of Industrial Wi-Fi Networks,” IEEE Trans.
Ind. Informat., vol. 12, no. 2, pp. 608–620, April 2016.
[18] ——, “Duplication avoidance mechanisms to reduce bandwidth usage in
redundant Wi-Fi networks,” in Proc. IEEE 13th Int. Workshop Factory
Communication Systems (WFCS), May 2017, pp. 1–10.
[19] ——, “Experimental Evaluation of Seamless Redundancy Applied to
Industrial Wi-Fi Networks,” IEEE Trans. on Ind. Informat., vol. 13,
no. 2, pp. 856–865, April 2017.
[20] ——, “Improving Effectiveness of Seamless Redundancy in Real Indus-
trial Wi-Fi Networks,” IEEE Trans. on Ind. Informat., vol. 14, no. 5, pp.
2095–2107, May 2018.
[21] ——, “A software-defined MAC architecture for Wi-Fi operating in user
space on conventional PCs,” in Proc. IEEE 13th Int. Workshop Factory
Communication Systems (WFCS), May 2017, pp. 1–10.
[22] D. Xia, J. Hart, and Q. Fu, “Evaluation of the Minstrel rate adaptation
algorithm in IEEE 802.11g WLANs,” in 2013 IEEE International
Conference on Communications (ICC), June 2013, pp. 2223–2228.
[23] M. T. Kassis, O. A. Mady, H. H. Halawa, M. Rentschler, R. M.
Daoud, H. H. Amer, and H. M. ElSayed, “Analysis of parallel redundant
WLAN with timing diversity,” in 2013 World Congress on Computer and
Information Technology (WCCIT), June 2013, pp. 1–6.
[24] G. Cena, I. C. Bertolotti, A. Valenzano, and C. Zunino, “Evaluation
of Response Times in Industrial WLANs,” IEEE Trans. Ind. Informat.,
vol. 3, no. 3, pp. 191–201, Aug 2007.
Gianluca Cena (SM’09) received the Laurea degree
in electronic engineering and the Ph.D. degree in in-
formation and system engineering from the Politec-
nico di Torino, Italy, in 1991 and 1996, respectively.
Since 2005 he has been a Director of Research with
the Institute of Electronics, Computer and Telecom-
munication Engineering, National Research Council
of Italy (CNR–IEIIT), Torino.
His research interests include wired and wireless
industrial communication systems, real-time proto-
cols, and automotive networks. In these areas he has
coauthored about 130 technical papers, three of which awarded as Best Papers
of the 2004, 2010, and 2017 editions of the IEEE Workshop on Factory
Communication Systems, and one as 2017 Best Paper for the IEEE TRA NS -
ACT ION S ON INDUSTRIAL INFORMATICS, plus one international patent.
Dr. Cena served as a Program Co-Chairman for the 2006 and 2008 editions
of the IEEE International Workshop on Factory Communication Systems, and
as a Track Co-Chairman in six editions of the IEEE International Conference
on Emerging Technologies and Factory Automation. Since 2009 he has
been an Associate Editor of the IE EE TRANSACTIONS ON INDUSTRIAL
INFORMATICS.
Stefano Scanzio (S’06-M’12) received the Lau-
rea and Ph.D. degrees in Computer Science from
Politecnico di Torino, Torino, Italy, in 2004 and
2008, respectively. He was with the Department of
Computer Engineering, Politecnico di Torino, from
2004 to 2009, where he was involved in research
on speech recognition and, in particular, he has
been active in classification methods and algorithms.
Since 2009, he has been with the National Research
Council of Italy (CNR), where he is a tenured Re-
searcher with the Institute of Electronics, Computer
and Telecommunication Engineering (IEIIT), Torino.
Dr. Scanzio served as a Work-in-Progress Co-Chairs in the 2018 edition
of the IEEE International Workshop on Factory Communication Systems
(WFCS 2018). He teaches several courses on Computer Science at Politecnico
di Torino. He has authored and co-authored of more than 50 papers in
international journals and conferences, in the area of industrial communication
systems, real-time networks, wireless networks and clock synchronization pro-
tocols. He received the 2017 Best Paper Award for the IEE E TRANSACTIONS
ON INDUSTRIAL INFORMATICS, and the Best Paper Awards for the papers he
presented at the 8th and 13th IEEE Workshops on Factory Communication
Systems (WFCS 2010 and WFCS 2017).
Adriano Valenzano (SM’09) received the Laurea
degree magna cum laude in electronic engineering
from Politecnico di Torino, Torino, Italy, in 1980. He
is Director of Research with the National Research
Council of Italy (CNR). He is currently with the
Institute of Electronics, Computer and Telecommu-
nication Engineering (IEIIT), Torino, Italy, where
he is responsible for research concerning distributed
computer systems, local area networks, and commu-
nication protocols. He has coauthored approximately
200 refereed journal and conference papers in the
area of computer engineering.
Dr. Valenzano is the recipient of the 2013 IEEE IES and ABB Lifetime
Contribution to Factory Automation Award. He was also awarded for the best
paper published in the IE EE TRANSACTIONS ON INDUSTRIAL INFORMATICS
during 2016, and received the Best Paper Awards for the papers presented at
the 5th, 8th and 13th IEEE Workshops on Factory Communication Systems
(WFCS 2004, WFCS 2010 and WFCS 2017).
Adriano Valenzano has served as a technical referee for several international
journals and conferences, also taking part in the program committees of
international events of primary importance. Since 2007, he has been serving
as an Associate Editor for the IE EE TRANSACTIONS ON INDUSTRIAL
INFORMATICS.