Content uploaded by Federico Tramarin
All content in this area was uploaded by Federico Tramarin on Nov 18, 2022
Content may be subject to copyright.
Improved Rate Adaptation Strategies for
Real–Time Industrial IEEE 802.11n WLANs
Federico Tramarin and Stefano Vitturi
Institute of Electronics, Computer and Telecommunication Engineering
National Research Council of Italy, CNR–IEIIT
Via Gradenigo 6/B, 35131 Padova, Italy
Department of Information Engineering
University of Padova
Via Gradenigo 6/B, 35131 Padova, Italy
Abstract—The IEEE 802.11 standard, since its earliest ver-
sions, provides the multi–rate support feature typically exploited
by Rate Adaptation (RA) techniques to dynamically select the
most suitable transmission rate, based on an estimation of the
channel status. With the release of the IEEE 802.11n amendment,
several enhancements have been introduced to the standard,
notably the support for MIMO architectures, whose beneﬁts can
be eﬀectively combined with multi–rate support. In an industrial
communication scenario, the RA algorithms commonly available
for general purpose applications revealed ineﬀective. This led to
the deﬁnition of purposely designed algorithms, with the aim of
improving the real–time behavior of IEEE 802.11 networks. In
this paper we take into consideration these techniques, as well as
some general purpose RA strategies, and analyze their implemen-
tation on an IEEE 802.11n communication system deployed in an
industrial scenario. Furthermore, we propose an eﬀective param-
eters tuning for the considered RA algorithms, as well as some
enhancements conceived to enforce their timeliness. An exhaus-
tive assessment, carried out via numerical simulations, shows that
the improved techniques allow to achieve excellent performance.
The IEEE 802.11n amendment to the IEEE 802.11 wireless
LAN (WLAN) standard  is nowadays implemented in
commercially available devices, that are massively employed
for general purpose applications. Such an amendment
provides several enhancements to the previous versions of the
standard that reﬂect in a considerable advancement of network
performance. Among the new features, IEEE 802.11n allows
for more eﬃcient modulations and higher channel bandwidth
with respect to IEEE 802.11b/g systems, as well as it enables
multiple input multiple output (MIMO) capabilities. This latter
feature, in particular, allows to implement multi-antenna sys-
tems, which are extensively exploited in home/oﬃce scenarios
to increase system throughput and/or to improve reliability.
The aforementioned enhancements look particularly
appealing also for the industrial communication scenario,
since it is expected that the intrinsic physical (PHY)
layer redundancy can be exploited, possibly sacriﬁcing the
throughput, to bring several beneﬁts in terms of timeliness and
reliability, which represent critical issues in such a context.
However, the adoption of IEEE 802.11n for industrial
real–time communication has only recently started to be
considered by the scientiﬁc community and the currently
available contributions are, actually, rather limited. Some
papers, however, are worth to be mentioned. In both  and
 some examples related to the adoption of IEEE 802.11n for
real–time industrial multimedia traﬃc are discussed. Another
typical issue of industrial communication, namely real-time
frame scheduling, is addressed in both  and . Here the
authors investigate the use of frame aggregation, a speciﬁc
IEEE 802.11n feature, that unfortunately reveals somewhat un-
suitable for such a kind of applications. As a last contribution,
it is worth mentioning paper , in which we investigated the
behavior of IEEE 802.11n in typical industrial applications,
through the execution of some practical experiments.
In the context of IEEE 802.11 WLANs, an important feature
is represented by the multi–rate support functions deﬁned by
the standard for any compliant device provided with a suitable
set of available transmission rates which, in principle, allows
a station to select the most suitable rate with the objective
of improving performance. To this aim, rate adaptation
algorithms have been made available since quite a very long
time for diﬀerent WLAN versions , and the selection of the
transmission rate has been commonly based on an estimation
of the transmission channel status. Unfortunately most of
these strategies, such as Automatic Rate Fallback (ARF),
revealed unsuitable for industrial communication due to their
inherent randomness that introduces additional uncertainty on
packet delivery. Thus, in  some of the authors of this paper
introduced two new RA techniques speciﬁcally conceived for
the industrial scenario, namely Static retransmission rate ARF
(SARF) and Fast rate reduction ARF (FARF), that provide
better performance in an industrial communication scenario.
In this paper we focus on industrial real-time applications of
IEEE 802.11n and, in particular, we provide an investigation
of the performance achieved by some RA techniques
implemented on such a network. Speciﬁcally, we ﬁrstly
consider an IEEE 802.11n conﬁguration that proved to allow
a considerable performance improvement with respect to IEEE
802.11g . Then, we propose the implementation of the
ARF, SARF and FARF algorithms for such a conﬁguration,
and analyze, through numerical simulations, their behavior
within this new scenario. Also, we take into consideration
that in general purpose applications a wide range of systems
now implement, in place of ARF–based algorithms, a recent978-1-4673-7929-8/15/$31.00 c
RA technique called Minstrel . For this reason we also
handled its implementation in our simulator, and addressed the
performance of such a technique for industrial applications.
As a further contribution, we address the tuning of the
original SARF and FARF strategies to enhance their perfor-
mance. Moreover, we show that an adequate optimization
of some Minstrel, parameters can considerably improve its
behavior making it suitable for some industrial applications.
In detail, the paper is organized as follows. Section II
introduces the features of IEEE 802.11n that are of interest
for industrial communication. Section III describes the char-
acteristics of the RA techniques that will be considered in this
paper. Section IV illustrates both the network and simulation
models that are used to investigate the behavior of the RA
techniques. Section V provides the results of the performance
assessment we carried out. Finally, Section VI concludes the
paper and points out some future research directions.
II. IEEE 802.11n for Industrial Wireless Communication
The IEEE 802.11n amendment was introduced several new
features at both the PHY and MAC layer with respect to the
previous versions of the IEEE 802.11 standard. However, only
some of those features reveal actually useful in the context
of industrial wireless networks, due to the diﬀerent types
of traﬃc and requirements typical of this scenario. Indeed,
we brieﬂy recall that industrial communication systems
are usually characterized by critical timing and reliability
constraints, with the exchange of small–size packets.
Among the enhancements introduced at the PHY layer,
mostly directed to increase data rate, the introduction of
40 MHz channels in place of the traditional 20 MHz ones
revealed very useful, allowing to roughly double the raw data
rate. The set of available modulations has been upgraded with
respect to those allowed by IEEE 802.11a/g, and also a higher
number of Orthogonal Frequency–Division Multiplexing
(OFDM) subcarriers is used for data transmission, both
yielding a further rate improvement.
The most important feature introduced by IEEE 802.11n,
however, is the support for MIMO capabilities, which comes
with a set of several conﬁguration options for the tuning of
its performance. As a matter of fact, in typical home/oﬃce
scenarios, the availability of a multi–antenna device is
exploited to enhance transmission rate by sending multiple
independent data streams in parallel, with a technique known
as Spatial Division Multiplexing (SDM). Conversely, in an
industrial communication context, it is more convenient to
use secondary antennas to improve communication reliability
through the adoption of Space–Time Block Coding (STBC)
techniques, actually implementing a sort of redundancy at the
PHY layer . Therefore, in this work we will address the
use of devices with two antennas that transmit a single spatial
stream and adopt STBC to improve system robustness.
The term Modulation and Coding Scheme (MCS) has been
introduced to indicate a speciﬁc PHY conﬁguration that yields
a certain rate. In this work we will employ the ﬁrst 8 MCSs
(from MCS 0 to MCS 7) that, with 40 MHz channels and
the aforementioned choice about the MIMO settings, provide
data rates ranging from 13.5 Mbps to 135 Mbps.
Finally, as far as the MAC layer is concerned, IEEE 802.11n
introduced some enhancements to increase its eﬃciency, such
as frame aggregation and block–ack. However, their adoption
does not bring any signiﬁcant performance improvement for
real–time communication, and for this reason has not been
further considered in the present work.
III. Rate adaptation
A. The ARF, SARF and FARF Techniques
The dynamic selection of the transmission rate carried out
by ARF is based on the number of consecutive failed or,
respectively, successful transmission attempts. In particular,
given a transmission rate set, a station decreases its rate to the
immediately lower one after Kconsecutive failed attempts.
Conversely, after Nconsecutive successful attempts, the rate is
increased to the immediately higher value. The default values
for parameters Kand Nare, respectively, 2 and 10. To enhance
the eﬀectiveness of ARF, two additional features have been
introduced. The ﬁrst one speciﬁes that, if the ﬁrst transmission
attempt after a rate increase fails, the rate is immediately
restored at the previous value. The second feature is meant
to avoid a station remains at a low rate for long time, and is
achieved through the use of a timer, started when the rate is
decreased, whose expiration triggers a rate increase, regardless
of the number of successful transmissions collected yet.
The analysis carried out in  for IEEE 802.11g showed
that ARF, particularly in the presence of fast varying channel
conditions, may introduce considerable randomness in packet
delivery as well as increase the packet error rate (PER). These
problems were solved, at least partially, with the deﬁnition
of two new techniques, namely SARF and FARF, speciﬁcally
conceived for industrial communication.
SARF speciﬁes that each retransmission attempt that takes
place after a failure is carried out at the lowest transmission
rate in the set supported by a station, e.g. 13.5 Mbps in our
system. Indeed, roughly speaking, since the success probability
at this rate is very high, the number of retransmissions (and
hence the randomness introduced by the backoﬀprocedures)
will be limited. However, the dynamic rate selection of SARF
is the same as ARF. In particular, successful retransmission
attempts at the lowest rate are not taken into consideration
to increase the rate. In practice, Kconsecutive failures at a
speciﬁc rate (of diﬀerent packets) trigger a rate decrease.
FARF, instead, is based on a modiﬁed rate selection mech-
anism with respect to ARF. Speciﬁcally, after a failed trans-
mission, the new rate is selected as the lowest one and then
it will be incremented at the immediately higher value in the
transmission set after Nconsecutive successful transmissions.
B. The Minstrel Algorithm
The Minstrel rate control strategy  was ﬁrstly proposed
in 2005 as part of the MadWiﬁ driver, developed for Atheros
chipsets in Linux–based system. Since then, its popularity
has increased thanks to the good performance that it oﬀers
in general purpose wireless networks subjected to best
eﬀort traﬃc. Consequently, Minstrel has been included as
the default rate control algorithm in the Linux kernel, and
in many popular wireless drivers currently employed by
oﬀ–the–shelf devices, such as ath5k and ath9k.
With respect to other RA strategies, Minstrel can be
classiﬁed as a random sampling algorithm, which tries to
select, within the set of available transmission rates, the
optimal one(s), based on the statistics collected for each rate
during the past communication history.
In practice, a station keeps a table (retry chain) with four
elements of type (Ri,Ci),i=1, . . . , 4, where each entry
indicates a speciﬁc rate Riand a number of attempts Ci. When
a frame has to be transmitted, the station performs the ﬁrst C1
transmission attempts at rate R1, then C2attempts at rate R2,
and so on. The retry chain is built using a speciﬁc procedure:
R1and R2are chosen as the rates that guarantee the highest
throughput (deﬁned as the best transmission speed weighted
by success probability), R3is the rate that yields the highest
success probability and R4is the lowest available rate. The
number of attempts Ciis computed as the maximum number
of transmission attempts at rate Rithat can be performed in
a window of length Tmax for a generic frame with a payload
length Lre f , taking into account the delays due to exponential
backoﬀmechanism. The default values for those parameters
are Lre f =1200 Bytes and Tmax =6 ms respectively.
The retry chain is updated with a period Tu=100 ms
adopting an Exponential Weighted Moving Average (EWMA)
approach. Speciﬁcally, if pi(k−1) is the success probability
for rate Ribefore the update, and pu
iis the success probability
empirically computed on all frames transmitted at rate Ri
during the current window Tu, the new value of the success
probability for rate Riis updated as
where αis the weight given to the previous history, whose
default value is 0.75. After the transmission rates update, the
four new entries of the retry chain can be selected according
to the aforementioned rules.
Finally, Minstrel is a random sampling RA strategy in the
sense that, with a probability Ps, it performs a frame transmis-
sion at a random rate Rsto estimate the channel performance
for that rate. Indeed, this random sampling procedure is used
by Minstrel to gather statistics for all the available rates, in
order to be always able to select the one better suiting the
current channel conditions. From a practical point of view,
this sampling procedure is achieved by substituting the ﬁrst
two entries of the retry chain with the new one Rsand R1,
where the faster rate will be placed at the ﬁrst position. In
this way, the algorithm calculates a number of attempts Cs
that have to be carried out with the temporary rate Rs, and
hence is able to keep updated its table even for those rates
that are less utilized.
From the performance point of view, Minstrel has been eval-
uated so far in general purpose networks , where it showed
better performance than previous algorithms, for example
ARF–like ones. Conversely, an evaluation of this strategy in an
industrial communication scenario has never been addressed.
IV. Network and Simulation Models
A. Network model
The network setup we considered in this work is composed
by one central station (the controller) which is connected
to a set of Msensors/actuators (nodes), as described in
Fig. 1. Without loss of generality, the controller is placed
in the middle of the network area whereas all wireless
nodes are located in a circular area at a distance such that
communications are always possible. In the following, if a
distinction amongst controller and node is not necessary, we
will simply refer to them as stations.
Fig. 1. Sketch of the adopted simulation setup
The whole set of stations does neither change over time nor
move within the selected area. In addition, we assumed that
the network set up phase has been carried out in advance, that
is, nodes have been successfully associated with the controller,
before the start of the network operations. Moreover, all
stations are compliant with the IEEE 802.11n standard, operate
in the same ISM 2.4 GHz frequency band, and share the
same wireless channel in such a way that communications are
always possible between the controller and any of the nodes.
The conﬁguration for both PHY and MAC layers of IEEE
802.11n is derived from the discussion provided in Sec. II.
Data exchange in the proposed network is based on a polling
scheme, in which the controller node acts as a master, whereas
the connected nodes represent the slaves. Given this protocol,
the master issues a unicast polling–request frame (with pay-
load size Lreq ) toward a single node, which in turn responds
with its polling–response frame (with payload size Ldata). Both
frames can possibly carry some output or input data, respec-
tively, and their payloads are, as typically assumed in factory
communication systems, in the order of some to some tens of
Bytes. Moreover, since it is also supposed that the master is
connected to a backbone Real–Time Ethernet network (even
if this is not signiﬁcant for this work, and hence has not been
considered), the polling–request frames issued by the master
are encapsulated within Ethernet frames, thus having a payload
size of at least 46 Bytes (i.e. minimum size Ethernet frames).
It is also assumed that input data coming from the nodes
are unicast frames, resembling a client–server relationship
between the master and the slaves. This clearly leads to
the further assumption that both the controller and the
nodes receive a feedback about the correctness of polling
frames delivery through the reception of an IEEE 802.11
Once the polling procedure of node iis completed, the
controller continues with the following node i+1 in the set,
until the whole sequence of Mnodes is completed. The time
to complete the polling of the entire sequence of nodes is
hence stored as a measure of the current cycle time, that
represents a signiﬁcant performance indicator for this study.
B. Simulation model
The simulation analysis has been carried out with the
goal of approximating as much as possible a real application
environment. In this light, we implemented a discrete–event
self–developed simulator in Matlab R
to model the application
protocol described above, as well as the IEEE 802.11n MAC
and PHY layer protocols. Moreover, the simulator implements
suitable channel models to emulate the behavior of a wireless
medium in the case of a multi–antenna system working in
a real application environment. A brief description of these
models is given in the next Section IV-C.
The simulations do not take into account external interfer-
ence eﬀects, and hence the source for erroneous receptions is
only found in the bad performance of the wireless channel. At
each received frame, the simulator calculates the perceived
SNR level sampling the channel status at the frame start.
This seems a justiﬁed assumption since the SNR ﬂuctuations
within the reception time of a single frame can be actually
considered limited, because of the small frame transmission
times (<100 µs) compared to the channel coherence time,
which is typically longer for the considered channel models
, as it will be clariﬁed in the next subsection. The SNR
level is used in the decision for the correctness of the reception,
based on PER curves which we have experimentally measured
in a real IEEE 802.11n system, reported in Fig. 2 for a payload
length of 50 Bytes. Details about the measurements we carried
out to obtain such curves can be found in . For the sake
of readability, the ﬁgure considers only a subset of the avail-
able modulation schemes. Nonetheless, MCS 0 and MCS 7
represent, respectively, the best and the worst behaviors of
the wireless channel. Also, MCS 6 has been added in order to
exemplify the behavior of an intermediate modulation scheme.
A polling trial fails if either a polling–request or a polling–
response frame is lost, within a maximum number of possible
transmission retries, which we have set to nmax.
Within the simulation sessions, a number of indicators
is collected to allow a statistical analysis of the network
behavior, for example the cycle time, the number of
transmission attempts needed to convey successfully a frame
to destination, the corresponding transmission rate, and the
real–time throughput deﬁned in .
Finally, singularities due to particular positions of the
stations as well as to particular network conﬁgurations have
been reduced by exploiting a random node placement strategy.
Speciﬁcally, with reference to Fig. 1, the circular area is split
in a number Mof contiguous sections. Within each section,
-10 -5 0 5 10 15 20 25
Fig. 2. Experimental PER vs SNR curves, payload size of 50 Bytes.
during the network setup phase each node niis placed in a
random position identiﬁed by (di, θi), which are two values
chosen randomly from a uniform distribution, and diranges
between dmin and dmax , the minimum and maximum allowed
distances, respectively, chosen in such a way that even the
farthest node will be able to communicate with the controller.
Furthermore, each simulation is repeated a number Nsof
times, and the node placement is updated at each repetition.
The full set of simulation parameters is reported in Tab. I.
Main simulation parameters
Parameter Description Value
MNumber of slave nodes 10
NsNumber of simulations 100
NcNetwork cycles for each simulations 10000
dmin Minimum distance between controller and node 3 m
dmax Maximum distance between controller and node 6 m
Lreq Payload size of polling–request frame 50 Bytes
Ldata Payload size of polling–response frame 10 Bytes
Ptx Transmit power 100 mW
nmax Maximum number of transmission attempts 7
CWmin Initial size of the contention window 15
C. Channel models
The developed simulator takes into account several eﬀects
relevant to an indoor wireless medium in the ISM band,
basically a variable transmit power, the log-distance path-loss
model, and a freely adjustable noise power at the receiver.
Moreover, since the communication scheme considered in
this paper is a multi–antenna system (i.e. MIMO), a suitable
model for small–scale fading is also provided.1
To model accurately the fading eﬀects on a MIMO channel,
we followed the proposals and the analysis originally carried
out by the Task Group n, in charge of the development of the
standard, which actually deﬁne six diﬀerent MIMO channel
models, often referenced as TGn model, identiﬁed with the
capital letters from A to F. The detailed description of these
models can be found in , along with the rationale that
1The common channel models usually referred in several studies about
wireless networks, such as AWGN channels and Gilbert–Elliott models, are
no longer suitable in the case of MIMO systems. For a in–depth analysis of
this topic, the interested reader can refer to, e.g., .
led to their deﬁnition. In the following, we limit to provide
a summary of the models main characteristics, with the aim
of targeting the industrial scenario.
To eﬀectively model the wireless channel in a multi–
antenna system, the ﬁrst eﬀect to take into account is that
the received signal at antenna jis the combined eﬀect of the
signals coming from all the transmitter antennas, each one
with its channel eﬀects. Secondly, considered any antennas
pair (transmitting iand receiving j), one has to account
for all the possible transmission paths, due for example to
reﬂections from the surrounding environment that lead to
diﬀerent propagation delays and received power.
Additionally, a third signiﬁcant eﬀect results from motion
of parts of the system, and is referred in the models
to as the Doppler spread. Actually, even if the stations
positions are ﬁxed (as it is the case in our scenario), the
surrounding environment may practically present moving
obstacles that modify the multi–path eﬀect during time, such
as the movement of people or machineries. Nonetheless,
especially in an industrial scenario, the speed of motion can
be considered quite low, typically at most some meters per
second. This results in an eﬀect strictly related to channel
coherence time that can be considered always much longer
than the packet transmission time, as it comes from the
analysis presented in ,  for an IEEE 802.11n system.
This holds particularly true in the case of industrial traﬃc,
where the typical payload lengths are very short.
Several further eﬀects have also to be accounted for, such
as the angle–of–departure or arrival, antenna correlations,
etc. The superimposition of all the mentioned components
led to the deﬁnition of a complex tunable channel model,
representative of real application environments. The six
reference indoor scenarios deﬁned by the TGn work group
range from small environments (home/oﬃce) up to the “limit”
case of a very large building, with reﬂections from diﬀerent
(and distant) obstacles, encompassed by the TGn model F.
In our perspective of a wireless communication system
applied in a typical industrial environment, we assume that the
network of Fig. 1 is deployed to manage a single production
island. Hence, in this scenario, the TGn model F represents the
best choice to model fading eﬀects on the wireless medium
since, even if nodes are placed in a limited area around the
controller, reﬂections and multi–path eﬀects could result from
objects far away from the considered stations, in agreement
with the assumption of that channel model. Consequently,
all the simulation outcomes included in this paper have been
obtained assuming the aforementioned model.
V. Performance Analysis
In this section we ﬁrstly address the tuning of the diﬀerent
RA schemes discussed in the previous Section III, with the
aim of improving their performance in an industrial scenario,
then we present the outcomes of a thorough performance
comparison among the behavior of the legacy strategies and
the improved ones.
A. Tuning of Minstrel algorithm
The Minstrel algorithm described in Section III-B was
designed for oﬃce networks, where the traﬃc patterns and
desired behavior are typically very diﬀerent from those of
industrial communication scenario. As a consequence, its pa-
rameters conﬁguration can be optimized to better meet the re-
quirements and the expected performance of this new context.
A ﬁrst step toward an enhanced version of the Minstrel
algorithm is represented by the revision of the reference
payload size Lre f . Indeed, the default value of 1200 Bytes is
actually much greater than that typically adopted by industrial
networks. Therefore, in agreement with the parameters
reported in Table I, we set the reference payload size to
Lre f =Lr eq =50 Bytes, since this value is representative of
a large set of industrial applications, and is also appropriate
for frame coming from a wired Ethernet segment and
encapsulated within a wireless frame.
Such a lowering of the reference payload size has a direct
inﬂuence on the computation of the throughput associated
with each rate, and also strongly impacts on the number of
possible attempts Cithat the algorithm is able to perform for
each rate, thus reﬂecting on network performance. Indeed,
with a smaller reference payload size, a higher number of
transmission attempts can be accommodated in a window of
length Tmax =6 ms. In particular, we observed that within
the allowed transmission window the value of Cicalculated
by the algorithm is always equal to the maximum number of
transmission attempts for a frame, nmax , given the fact that with
a very small reference payload a frame takes much less time to
be transmitted. As a result, only the ﬁrst rate in the retry chain
happens to be used, which is clearly an undesired behavior.
It is evident that such an issue is raised up from the
choice for Tmax , whose default value of 6 ms was empirically
derived from measurements on a TCP–based oﬃce network,
as the authors of Minstrel reported . However, in industrial
applications with generally much lower service times, that
value revealed not suitable and also not well balanced with
the previous tuning of the value of Lre f .
Indeed, for the network conﬁguration proposed in the
previous Section, we show in Fig. 3 the Empirical Cumulative
Distribution Function (ECDF) of the obtained cycle time for
varying values of Tmax. In the ﬁgure, the default conﬁguration
is assumed as that with Lre f =1200 Bytes and Tmax =6 ms,
while the other conﬁgurations adopt the new reference
payload size Lre f =50 Bytes. It can be clearly observed
that with a reference payload size of 50 Bytes, the cycle
time behavior is much more deterministic for lower values
of Tmax , and the best results are obtained with a period
of Tmax =200 µs, which will be hence considered in the
following as the optimal value for this parameter.
Another issue we identiﬁed in the Minstrel algorithm is
the sampling probability Ps, which is set by default to 0.1,
meaning that one packet out of ten is sent at a random rate to
gather statistics. This can be regarded as a quite high value in
the considered context, since it introduces a very high degree
of randomness in the packet delivery time. An increased deter-
Cycle time [us]
2000 4000 6000 8000 10000
Fig. 3. Cycle time for Minstrel algorithm with diﬀerent values of Tmax
minism in rate selection can be actually achieved by decreasing
the sampling probability Ps, as highlighted by Fig. 4, which
reports the jitter on cycle time for diﬀerent values of Ps,
keeping ﬁxed all the other parameters, including Lre f and Tma x.
Jitter on cycle time [us]
Fig. 4. Jitter on cycle time for Minstrel algorithm with diﬀerent values of Ps
From the ﬁgure it could be inferred that Psmust be as low as
possible to improve real–time performance of Minstrel. How-
ever, it has to be noted that a very low value of Psalong with a
short update window Tumay result in an almost static behavior
of Minstrel algorithm. Indeed, if the window is short and
the sampling probability is low, very few information about
rates diﬀerent from the current ones will be gathered, and the
algorithm will be strongly dependent on the initial rates set
and poorly reactive to changes in the environment. As a con-
sequence, very low sampling probabilities should be avoided.
Tab. II shows the average and standard deviation of the
cycle time for diﬀerent values of Psand Tu(all other
parameters are set to default values). It is evident that a
decrease in Psshould be balanced by an increase of Tuin
order to reduce both mean value and jitter of the cycle time
under the same environmental conditions.
Cycle time for Minstrel algorithm with different values of Psand Tu
Tu=100 ms Tu=1 s
Ps=0.1 Ps=0.02 Ps=0.1 Ps=0.02
Average cycle time 8.37 ms 6.73 ms 6.41 ms 5.57 ms
Jitter on cycle time 3 ms 1.57 ms 2.23 ms 1.58 ms
Finally, based on the above considerations, the following
optimized parameters conﬁguration has been chosen for the
Minstrel algorithm in the proposed industrial communication
scenario: Lre f =50 Bytes, Tma x =200 µs, Tu=1 s,
Ps=0.02. The weight αhas been kept equal to 0.75 since
further simulations showed that it does not have any deﬁned
impact on the cycle time behavior.
B. Improving rate control strategies with smarter controllers
All the rate control algorithms discussed so far rely on the
past communication history to select the optimal transmission
rate. The assumption behind this behavior is that consecutive
samples of the communication channel are correlated. Nev-
ertheless, one should consider that the communication links
between each possible node pair in the network form a set
of diﬀerent realizations of the shared wireless medium, all of
them characterized by a diﬀerent behavior. However, when a
generic station Ahas to send a packet, on the basis of the
discussed rate control scheme it selects the next transmission
rate by looking at the past channel history regardless of the in-
tended receiver.This means that the rate used to send a packet
to a destination node Bis likely inﬂuenced by the rate(s) pre-
viously selected to send packets to another destination node,
say node C, despite the fact that the communication channel
between Aand Bis generally uncorrelated with that between
Aand C. As a consequence of this fact, the current RA algo-
rithm implementations can lead to poor system performance,
especially for stations that have packets for multiple receivers.
In the network described in Section IV, this issue is never
encountered by the slaves, since they only exchange data with
the controller. Conversely, it occurs for the controller and,
hence, it can signiﬁcantly aﬀect the performance. To overcome
this problem, we implemented smarter rate control strategies
in the controller, that has now to diﬀerentiate the previous
communication history according to the suitable destination
node, and hence has only to exploit the statistics relevant to
that destination node when choosing the transmission rate.
From an implementation point of view, in the case of SARF
and FARF strategies a node needs only to keep track of the
previous rate and the number of consecutive failed/successful
transmissions at this rate, whereas with the proposed
modiﬁcation, the controller needs to store those information
for each of the Mslave nodes in a suitable vector. Analogously,
in the case of the Minstrel algorithm, the controller needs to
store a diﬀerent retry chain and diﬀerentiated rate statistics
for each of the Mslave node. In both cases, the previous
communication history is accounted by looking only at the
result of the packet exchange with the single node of interest.
Clearly, this modiﬁcation requires a higher amount of memory
in the controller to be used by the RA algorithm, especially
for the case of the Minstrel algorithm. However, this should
not be an issue, since the controller is generally implemented
by devices with adequate resources and capabilities.
To prove the eﬀectiveness of the adoption of a smarter
controller, we provide in Fig. 5 the ECDF of the cycle time
for the two diﬀerent implementations of the FARF rate control
strategy: the standard one (FARF) and that with a smarter con-
troller (FARF 2). The other RA schemes show a similar behav-
ior, and have been omitted in the ﬁgure to avoid clutter. As can
be seen, the beneﬁts in terms of system timeliness are evident.
Cycle time [us]
2900 3000 3100 3200 3300 3400 3500
Fig. 5. Cycle time for FARF, and FARF with a smarter controller (FARF 2)
C. Comparison among diﬀerent rate control strategies
Fig. 6 reports the ECDF of the cycle time when the four
RA techniques mentioned in Section III, namely ARF, SARF,
FARF and Minstrel, are adopted in their standard (i.e. without
any tuning) implementation. It is shown here to provide a
clear picture of the expected performance of these techniques
when introduced in an industrial applications context.
Cycle time [us]
2000 4000 6000 8000 10000
Fig. 6. Cycle time with diﬀerent RA algorithms in their legacy implementation
This simulation also provides a conﬁrmation that the
Minstrel algorithm in its legacy implementation is totally
unsuitable for industrial communication, mainly because
of the high sampling probability Psand the fact that all
retransmissions are performed at the initial rate because of
the values assigned to Lre f and Tmax. Conversely, FARF,
SARF and ARF exhibit a much more deterministic behavior,
with FARF outperforming the other two techniques.
A subsequent evaluation has been carried out by considering
the proposed improved versions of the rate control algorithms.
Speciﬁcally, FARF 2 uses the discussed smart controller, while
in SARF 2 both the smart controller is introduced and, also,
Khas been reduced from 2 to 1 to further improve robustness.
Finally, Minstrel 2 has the parameter conﬁguration determined
in subsection V-A, namely Lre f =50 Bytes, Tmax =200 µs,
Tu=1 s and Ps=0.02, along with the smart controller. The
ECDF of the cycle time, is reported in Fig. 7.
As can be seen, the behavior of Minstrel is considerably
improved with respect to that of Fig. 6, but the enhanced
versions of both SARF and FARF still behave rather better,
Cycle time [us]
2500 3000 3500 4000 4500 5000 5500
Fig. 7. Cycle time with diﬀerent RA algorithms in their improved versions
providing both a lower average cycle time and a lower bound
on the maximum cycle time (almost 3.5 ms versus more than
5.5 ms). The improved SARF 2, with a smart controller and
K=1, is slightly faster than FARF 2, while this latter is more
conservative, yielding a reduced jitter.
To better understand the behavior of the diﬀerent RA
strategies, Fig. 8 reports the histogram of the MCSs adopted
during the whole simulation by all nodes and the AP, showing
the results for each of the three proposed rate control strategies.
Number of transmissions
Fig. 8. Histogram of adopted MCSs with diﬀerent rate control algorithms
in their improved versions
The diﬀerent approach of each strategy is clearly visible.
SARF 2 tends to explore all the rates with a prevalence of
MCS 3 and, in general, lower rates are preferred to higher
ones due to the conservative rate selection scheme and the
channel conditions. The same holds true for FARF 2 which,
however, adopts more often MCS 0, then MCS 1 and so on
in a descending order. This is due to the fact that, each time
there is an error, FARF restarts from the lowest MCS, and
hence lower MCSs are explored more often. Finally, with
Minstrel 2 two rates are mostly adopted, namely MCS 0,
which gives the best transmission success probability, and
MCS 3, which guarantees the highest throughput weighted
for success probability (in the simulated channel conditions).
As a ﬁnal evaluation, the performance of the proposed
rate adaptation techniques has been assessed for packets
with larger payload sizes, as it is the case of, for example,
real–time multimedia traﬃc .
The introduction of this new type of traﬃc in the
proposed industrial network can ultimately be modeled
by simply increasing the size of polling–response packets
sent by the nodes (Ldata ). Tab. III reports the simulation
outcomes obtained by the proposed rate control strategies for
Ldata =500 Bytes and compares them with those achieved
with Ldata =10 Bytes. The considered metrics are mean and
standard deviation of cycle time and real–time throughput.
Outcomes with the proposed RA techniques and different payload sizes
Ldata 10 B 500 B 10 B 500 B 10 B 500 B
SARF 2 3.05 109.46 0.09 3.77 1.42 0.36
FARF 2 3.10 106.92 0.05 3.45 1.40 0.37
Minstrel 2 3.59 3.65 0.52 0.68 1.22 11.13
It emerges that, with larger payloads, the optimized version
of Minstrel signiﬁcantly outperforms both the enhanced
versons of SARF and FARF, conversely to what happened in
the smaller payload case.
This derives, basically, from the increase of the packet error
rate. In this situation, both FARF and SARF tend to often
use the lowest rate (MCS 0), since they are very conservative
schemes with a sharp reaction to transmission errors. The
use of such a slow rate actually keeps jitter under control
but, on the other hand, implies a great increase in packet
delivery time, which reﬂects on the cycle time computation.
As a result, average cycle time results much more increased
(roughly from 3 ms to more than 100 ms) and real–time
throughput drops from 1.4 Mbps to roughly 0.35 Mbps.
Conversely, the Minstrel algorithm has a broader set of
parameters, among which Lre f that can be tuned according
to the expected type of traﬃc, and takes into account the
outcomes of a larger window of transmission attempts instead
of considering only the latest one. As a result, the increase
in payload size causes only a slight growth in cycle time
mean and jitter, while the real–time throughput is strongly
augmented (from roughly 1.2 Mbps to more than 11 Mbps),
since it is also a function of the payload size.
To conclude, it can be stated that the enhanced
implementations of both SARF and FARF techniques, provide
the best performance as far as traditional industrial traﬃc is
concerned, with the former yielding higher throughput and the
latter providing a smaller jitter on cyclic operations. However,
if new types of industrial traﬃc (e.g. multimedia) characterized
by larger frames are considered, the proposed tuned version
of the Minstrel algorithm can be regarded as the best choice,
in that it optimizes both real–time throughput and timeliness.
VI. Conclusions and future directions of research
In this paper the behavior of some RA techniques has been
analyzed through numerical simulations in a reference real–
time industrial application, deployed over an IEEE 802.11n
WLAN. Speciﬁcally, we took into consideration some general
purpose RA algorithms (i.e. ARF and Minstrel), as well as
strategies purposely developed for industrial communication,
such as SARF and FARF. Enhanced SARF 2 and FARF 2
work better if a traditional industrial traﬃc is considered,
providing very low average cycle time with bounded jitter.
Conversely, the optimized version of the Minstrel algorithm
becomes deﬁnitely the best choice for multimedia industrial
traﬃc, characterized by the exchange of larger frames.
A straightforward extension of this work is the experimental
validation of the obtained simulation results. This can be
actually carried out exploiting the possibilities oﬀered by
recently developed wireless cards based on softMAC drivers,
that allow an accurate control of the MAC layer and,
speciﬁcally, of the rate selection mechanisms, through a
reprogramming of the suitable parts of the driver source code.
Finally, the insights provided by this work should hopefully
represent a further step towards the introduction of the IEEE
802.11n WLAN in the industrial communication scenario.
 IEEE Standard for Wireless LAN Medium Access Control (MAC)
and Physical Layer (PHY) Speciﬁcations: Enhancements for Higher
Throughput, IEEE Std., Oct. 2009.
 S. Santonja-Climent, D. Todoli-Ferrandis, T. Albero-Albero,
V. Sempere-Paya, J. Silvestre-Blanes, and J. Alcober, “Analysis
of control and multimedia real-time traﬃc over SIP and RTP on
802.11n wireless links for utilities networks,” in Emerging Technologies
and Factory Automation (ETFA), 2010 IEEE Conference on, Sept 2010.
 J. S. Blanes, J. Berenguer-Sebasti, V. Sempere-Paya, and D. T. Ferrandis,
“802.11n Performance analysis for a real multimedia industrial
application,” Computers in Industry, vol. 66, no. 0, pp. 31 – 40, 2015.
 E. Charﬁ, L. Chaari Fourati, and L. Kamoun, “QoS support of
voice/video services under IEEE 802.11n WLANs,” in Communication
Systems, Networks Digital Signal Processing (CSNDSP), 2014 9th
International Symposium on, July 2014, pp. 600–605.
 B. Maqhat, M. Baba, and R. Rahman, “A-MSDU real time traﬃc
scheduler for IEEE802.11n WLANs,” in Wireless Technology and Ap-
plications (ISWTA), 2012 IEEE Symposium on, Sept 2012, pp. 286–290.
 F. Tramarin, S. Vitturi, M. Luvisotto, and A. Zanella, “The IEEE
802.11n wireless LAN for real-time industrial communication,” in
Proc. of IEEE WFCS, Palma de Majorca, Spain, 2015.
 P. Kulkarni and S. Quadri, “Simple and Practical Rate Adaptation
Algorithms for Wireless Networks,” in World of Wireless, Mobile
and Multimedia Networks Workshops, 2009. WoWMoM 2009. IEEE
International Symposium on, June 2009, pp. 1–9.
 S. Vitturi, L. Seno, F. Tramarin, and M. Bertocco, “On the Rate
Adaptation Techniques of IEEE 802.11 Networks for Industrial
Applications,” Industrial Informatics, IEEE Transactions on, vol. 9,
no. 1, pp. 198–208, Feb 2013.
 (2015, January) Minstrel speciﬁcation. [Online]. Available:
 D. Xia, J. Hart, and Q. Fu, “Evaluation of the minstrel rate adaptation
algorithm in IEEE 802.11g WLANs,” in Communications (ICC), 2013
IEEE International Conference on, June 2013, pp. 2223–2228.
 T. Paul and T. Ogunfunmi, “Wireless LAN Comes of Age:
Understanding the IEEE 802.11n Amendment,” Circuits and Systems
Magazine, IEEE, vol. 8, no. 1, pp. 28–54, First 2008.
 IEC 61784: Digital data communications for measurement and control
– Part 2: Additional proﬁles for ISO/IEC 8802–3 based communication
networks in real–time applications, International Electrotechnical
Commission Std., November 2007.
 E. Perahia and R. Stacey, Next Generation Wireless LANs: 802.11n and
802.11ac. Cambridge university press, 2013.
 V. Erceg, L. Schumacher, P. Kyritsi, A. Molisch, D. S. Baum et al.,
“TGn channel models,” IEEE document 802.11-03/940r4, May 2004.
 J. Silvestre-Blanes, L. Almeida, R. Marau, and P. Pedreiras, “Online
QoS Management for Multimedia Real-Time Transmission in Industrial
Networks,” Industrial Electronics, IEEE Transactions on, vol. 58, no. 3,
pp. 1061–1071, March 2011.